6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
243 *
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
253 */
254
255 #include <sys/spa.h>
256 #include <sys/zio.h>
257 #include <sys/spa_impl.h>
258 #include <sys/zio_compress.h>
259 #include <sys/zio_checksum.h>
260 #include <sys/zfs_context.h>
261 #include <sys/arc.h>
262 #include <sys/refcount.h>
263 #include <sys/vdev.h>
264 #include <sys/vdev_impl.h>
265 #include <sys/dsl_pool.h>
266 #include <sys/zio_checksum.h>
267 #include <sys/multilist.h>
268 #include <sys/abd.h>
269 #ifdef _KERNEL
270 #include <sys/vmsystm.h>
271 #include <vm/anon.h>
272 #include <sys/fs/swapnode.h>
273 #include <sys/dnlc.h>
274 #endif
275 #include <sys/callb.h>
276 #include <sys/kstat.h>
277 #include <zfs_fletcher.h>
278 #include <sys/aggsum.h>
279 #include <sys/cityhash.h>
280
281 #ifndef _KERNEL
282 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
283 boolean_t arc_watch = B_FALSE;
284 int arc_procfd;
285 #endif
286
287 static kmutex_t arc_reclaim_lock;
288 static kcondvar_t arc_reclaim_thread_cv;
289 static boolean_t arc_reclaim_thread_exit;
290 static kcondvar_t arc_reclaim_waiters_cv;
291
292 uint_t arc_reduce_dnlc_percent = 3;
293
294 /*
295 * The number of headers to evict in arc_evict_state_impl() before
296 * dropping the sublist lock and evicting from another sublist. A lower
297 * value means we're more likely to evict the "correct" header (i.e. the
298 * oldest header in the arc state), but comes with higher overhead
299 * (i.e. more invocations of arc_evict_state_impl()).
340
341 static int arc_dead;
342
343 /*
344 * The arc has filled available memory and has now warmed up.
345 */
346 static boolean_t arc_warm;
347
348 /*
349 * log2 fraction of the zio arena to keep free.
350 */
351 int arc_zio_arena_free_shift = 2;
352
353 /*
354 * These tunables are for performance analysis.
355 */
356 uint64_t zfs_arc_max;
357 uint64_t zfs_arc_min;
358 uint64_t zfs_arc_meta_limit = 0;
359 uint64_t zfs_arc_meta_min = 0;
360 int zfs_arc_grow_retry = 0;
361 int zfs_arc_shrink_shift = 0;
362 int zfs_arc_p_min_shift = 0;
363 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
364
365 boolean_t zfs_compressed_arc_enabled = B_TRUE;
366
367 /*
368 * Note that buffers can be in one of 6 states:
369 * ARC_anon - anonymous (discussed below)
370 * ARC_mru - recently used, currently cached
371 * ARC_mru_ghost - recentely used, no longer in cache
372 * ARC_mfu - frequently used, currently cached
373 * ARC_mfu_ghost - frequently used, no longer in cache
374 * ARC_l2c_only - exists in L2ARC but not other states
375 * When there are no active references to the buffer, they are
376 * are linked onto a list in one of these arc states. These are
377 * the only buffers that can be evicted or deleted. Within each
378 * state there are multiple lists, one for meta-data and one for
379 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
380 * etc.) is tracked separately so that it can be managed more
381 * explicitly: favored over data, limited explicitly.
382 *
383 * Anonymous buffers are buffers that are not associated with
384 * a DVA. These are buffers that hold dirty block copies
390 * The ARC_l2c_only state is for buffers that are in the second
391 * level ARC but no longer in any of the ARC_m* lists. The second
392 * level ARC itself may also contain buffers that are in any of
393 * the ARC_m* states - meaning that a buffer can exist in two
394 * places. The reason for the ARC_l2c_only state is to keep the
395 * buffer header in the hash table, so that reads that hit the
396 * second level ARC benefit from these fast lookups.
397 */
398
399 typedef struct arc_state {
400 /*
401 * list of evictable buffers
402 */
403 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
404 /*
405 * total amount of evictable data in this state
406 */
407 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
408 /*
409 * total amount of data in this state; this includes: evictable,
410 * non-evictable, ARC_BUFC_DATA, and ARC_BUFC_METADATA.
411 */
412 refcount_t arcs_size;
413 } arc_state_t;
414
415 /* The 6 states: */
416 static arc_state_t ARC_anon;
417 static arc_state_t ARC_mru;
418 static arc_state_t ARC_mru_ghost;
419 static arc_state_t ARC_mfu;
420 static arc_state_t ARC_mfu_ghost;
421 static arc_state_t ARC_l2c_only;
422
423 typedef struct arc_stats {
424 kstat_named_t arcstat_hits;
425 kstat_named_t arcstat_misses;
426 kstat_named_t arcstat_demand_data_hits;
427 kstat_named_t arcstat_demand_data_misses;
428 kstat_named_t arcstat_demand_metadata_hits;
429 kstat_named_t arcstat_demand_metadata_misses;
430 kstat_named_t arcstat_prefetch_data_hits;
431 kstat_named_t arcstat_prefetch_data_misses;
432 kstat_named_t arcstat_prefetch_metadata_hits;
433 kstat_named_t arcstat_prefetch_metadata_misses;
434 kstat_named_t arcstat_mru_hits;
435 kstat_named_t arcstat_mru_ghost_hits;
436 kstat_named_t arcstat_mfu_hits;
437 kstat_named_t arcstat_mfu_ghost_hits;
438 kstat_named_t arcstat_deleted;
439 /*
440 * Number of buffers that could not be evicted because the hash lock
441 * was held by another thread. The lock may not necessarily be held
442 * by something using the same buffer, since hash locks are shared
443 * by multiple buffers.
444 */
445 kstat_named_t arcstat_mutex_miss;
446 /*
447 * Number of buffers skipped because they have I/O in progress, are
448 * indrect prefetch buffers that have not lived long enough, or are
449 * not from the spa we're trying to evict from.
450 */
451 kstat_named_t arcstat_evict_skip;
452 /*
453 * Number of times arc_evict_state() was unable to evict enough
454 * buffers to reach it's target amount.
455 */
456 kstat_named_t arcstat_evict_not_enough;
457 kstat_named_t arcstat_evict_l2_cached;
458 kstat_named_t arcstat_evict_l2_eligible;
459 kstat_named_t arcstat_evict_l2_ineligible;
460 kstat_named_t arcstat_evict_l2_skip;
461 kstat_named_t arcstat_hash_elements;
462 kstat_named_t arcstat_hash_elements_max;
463 kstat_named_t arcstat_hash_collisions;
464 kstat_named_t arcstat_hash_chains;
465 kstat_named_t arcstat_hash_chain_max;
466 kstat_named_t arcstat_p;
467 kstat_named_t arcstat_c;
468 kstat_named_t arcstat_c_min;
469 kstat_named_t arcstat_c_max;
470 /* Not updated directly; only synced in arc_kstat_update. */
471 kstat_named_t arcstat_size;
472 /*
473 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
474 * Note that the compressed bytes may match the uncompressed bytes
475 * if the block is either not compressed or compressed arc is disabled.
476 */
477 kstat_named_t arcstat_compressed_size;
478 /*
479 * Uncompressed size of the data stored in b_pabd. If compressed
480 * arc is disabled then this value will be identical to the stat
481 * above.
482 */
483 kstat_named_t arcstat_uncompressed_size;
484 /*
485 * Number of bytes stored in all the arc_buf_t's. This is classified
486 * as "overhead" since this data is typically short-lived and will
487 * be evicted from the arc when it becomes unreferenced unless the
488 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
489 * values have been set (see comment in dbuf.c for more information).
490 */
491 kstat_named_t arcstat_overhead_size;
492 /*
493 * Number of bytes consumed by internal ARC structures necessary
494 * for tracking purposes; these structures are not actually
495 * backed by ARC buffers. This includes arc_buf_hdr_t structures
496 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
497 * caches), and arc_buf_t structures (allocated via arc_buf_t
498 * cache).
499 * Not updated directly; only synced in arc_kstat_update.
500 */
501 kstat_named_t arcstat_hdr_size;
502 /*
503 * Number of bytes consumed by ARC buffers of type equal to
504 * ARC_BUFC_DATA. This is generally consumed by buffers backing
505 * on disk user data (e.g. plain file contents).
506 * Not updated directly; only synced in arc_kstat_update.
507 */
508 kstat_named_t arcstat_data_size;
509 /*
510 * Number of bytes consumed by ARC buffers of type equal to
511 * ARC_BUFC_METADATA. This is generally consumed by buffers
512 * backing on disk data that is used for internal ZFS
513 * structures (e.g. ZAP, dnode, indirect blocks, etc).
514 * Not updated directly; only synced in arc_kstat_update.
515 */
516 kstat_named_t arcstat_metadata_size;
517 /*
518 * Number of bytes consumed by various buffers and structures
519 * not actually backed with ARC buffers. This includes bonus
520 * buffers (allocated directly via zio_buf_* functions),
521 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
522 * cache), and dnode_t structures (allocated via dnode_t cache).
523 * Not updated directly; only synced in arc_kstat_update.
524 */
525 kstat_named_t arcstat_other_size;
526 /*
527 * Total number of bytes consumed by ARC buffers residing in the
528 * arc_anon state. This includes *all* buffers in the arc_anon
529 * state; e.g. data, metadata, evictable, and unevictable buffers
530 * are all included in this value.
531 * Not updated directly; only synced in arc_kstat_update.
532 */
533 kstat_named_t arcstat_anon_size;
534 /*
535 * Number of bytes consumed by ARC buffers that meet the
536 * following criteria: backing buffers of type ARC_BUFC_DATA,
537 * residing in the arc_anon state, and are eligible for eviction
538 * (e.g. have no outstanding holds on the buffer).
539 * Not updated directly; only synced in arc_kstat_update.
540 */
541 kstat_named_t arcstat_anon_evictable_data;
542 /*
543 * Number of bytes consumed by ARC buffers that meet the
544 * following criteria: backing buffers of type ARC_BUFC_METADATA,
545 * residing in the arc_anon state, and are eligible for eviction
546 * (e.g. have no outstanding holds on the buffer).
547 * Not updated directly; only synced in arc_kstat_update.
548 */
549 kstat_named_t arcstat_anon_evictable_metadata;
550 /*
551 * Total number of bytes consumed by ARC buffers residing in the
552 * arc_mru state. This includes *all* buffers in the arc_mru
553 * state; e.g. data, metadata, evictable, and unevictable buffers
554 * are all included in this value.
555 * Not updated directly; only synced in arc_kstat_update.
556 */
557 kstat_named_t arcstat_mru_size;
558 /*
559 * Number of bytes consumed by ARC buffers that meet the
560 * following criteria: backing buffers of type ARC_BUFC_DATA,
561 * residing in the arc_mru state, and are eligible for eviction
562 * (e.g. have no outstanding holds on the buffer).
563 * Not updated directly; only synced in arc_kstat_update.
564 */
565 kstat_named_t arcstat_mru_evictable_data;
566 /*
567 * Number of bytes consumed by ARC buffers that meet the
568 * following criteria: backing buffers of type ARC_BUFC_METADATA,
569 * residing in the arc_mru state, and are eligible for eviction
570 * (e.g. have no outstanding holds on the buffer).
571 * Not updated directly; only synced in arc_kstat_update.
572 */
573 kstat_named_t arcstat_mru_evictable_metadata;
574 /*
575 * Total number of bytes that *would have been* consumed by ARC
576 * buffers in the arc_mru_ghost state. The key thing to note
577 * here, is the fact that this size doesn't actually indicate
578 * RAM consumption. The ghost lists only consist of headers and
579 * don't actually have ARC buffers linked off of these headers.
580 * Thus, *if* the headers had associated ARC buffers, these
581 * buffers *would have* consumed this number of bytes.
582 * Not updated directly; only synced in arc_kstat_update.
583 */
584 kstat_named_t arcstat_mru_ghost_size;
585 /*
586 * Number of bytes that *would have been* consumed by ARC
587 * buffers that are eligible for eviction, of type
588 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
589 * Not updated directly; only synced in arc_kstat_update.
590 */
591 kstat_named_t arcstat_mru_ghost_evictable_data;
592 /*
593 * Number of bytes that *would have been* consumed by ARC
594 * buffers that are eligible for eviction, of type
595 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
596 * Not updated directly; only synced in arc_kstat_update.
597 */
598 kstat_named_t arcstat_mru_ghost_evictable_metadata;
599 /*
600 * Total number of bytes consumed by ARC buffers residing in the
601 * arc_mfu state. This includes *all* buffers in the arc_mfu
602 * state; e.g. data, metadata, evictable, and unevictable buffers
603 * are all included in this value.
604 * Not updated directly; only synced in arc_kstat_update.
605 */
606 kstat_named_t arcstat_mfu_size;
607 /*
608 * Number of bytes consumed by ARC buffers that are eligible for
609 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
610 * state.
611 * Not updated directly; only synced in arc_kstat_update.
612 */
613 kstat_named_t arcstat_mfu_evictable_data;
614 /*
615 * Number of bytes consumed by ARC buffers that are eligible for
616 * eviction, of type ARC_BUFC_METADATA, and reside in the
617 * arc_mfu state.
618 * Not updated directly; only synced in arc_kstat_update.
619 */
620 kstat_named_t arcstat_mfu_evictable_metadata;
621 /*
622 * Total number of bytes that *would have been* consumed by ARC
623 * buffers in the arc_mfu_ghost state. See the comment above
624 * arcstat_mru_ghost_size for more details.
625 * Not updated directly; only synced in arc_kstat_update.
626 */
627 kstat_named_t arcstat_mfu_ghost_size;
628 /*
629 * Number of bytes that *would have been* consumed by ARC
630 * buffers that are eligible for eviction, of type
631 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
632 * Not updated directly; only synced in arc_kstat_update.
633 */
634 kstat_named_t arcstat_mfu_ghost_evictable_data;
635 /*
636 * Number of bytes that *would have been* consumed by ARC
637 * buffers that are eligible for eviction, of type
638 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
639 * Not updated directly; only synced in arc_kstat_update.
640 */
641 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
642 kstat_named_t arcstat_l2_hits;
643 kstat_named_t arcstat_l2_misses;
644 kstat_named_t arcstat_l2_feeds;
645 kstat_named_t arcstat_l2_rw_clash;
646 kstat_named_t arcstat_l2_read_bytes;
647 kstat_named_t arcstat_l2_write_bytes;
648 kstat_named_t arcstat_l2_writes_sent;
649 kstat_named_t arcstat_l2_writes_done;
650 kstat_named_t arcstat_l2_writes_error;
651 kstat_named_t arcstat_l2_writes_lock_retry;
652 kstat_named_t arcstat_l2_evict_lock_retry;
653 kstat_named_t arcstat_l2_evict_reading;
654 kstat_named_t arcstat_l2_evict_l1cached;
655 kstat_named_t arcstat_l2_free_on_write;
656 kstat_named_t arcstat_l2_abort_lowmem;
657 kstat_named_t arcstat_l2_cksum_bad;
658 kstat_named_t arcstat_l2_io_error;
659 kstat_named_t arcstat_l2_lsize;
660 kstat_named_t arcstat_l2_psize;
661 /* Not updated directly; only synced in arc_kstat_update. */
662 kstat_named_t arcstat_l2_hdr_size;
663 kstat_named_t arcstat_memory_throttle_count;
664 /* Not updated directly; only synced in arc_kstat_update. */
665 kstat_named_t arcstat_meta_used;
666 kstat_named_t arcstat_meta_limit;
667 kstat_named_t arcstat_meta_max;
668 kstat_named_t arcstat_meta_min;
669 kstat_named_t arcstat_sync_wait_for_async;
670 kstat_named_t arcstat_demand_hit_predictive_prefetch;
671 } arc_stats_t;
672
673 static arc_stats_t arc_stats = {
674 { "hits", KSTAT_DATA_UINT64 },
675 { "misses", KSTAT_DATA_UINT64 },
676 { "demand_data_hits", KSTAT_DATA_UINT64 },
677 { "demand_data_misses", KSTAT_DATA_UINT64 },
678 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
679 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
680 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
681 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
682 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
683 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
684 { "mru_hits", KSTAT_DATA_UINT64 },
685 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
686 { "mfu_hits", KSTAT_DATA_UINT64 },
687 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
688 { "deleted", KSTAT_DATA_UINT64 },
689 { "mutex_miss", KSTAT_DATA_UINT64 },
690 { "evict_skip", KSTAT_DATA_UINT64 },
691 { "evict_not_enough", KSTAT_DATA_UINT64 },
692 { "evict_l2_cached", KSTAT_DATA_UINT64 },
693 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
694 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
695 { "evict_l2_skip", KSTAT_DATA_UINT64 },
696 { "hash_elements", KSTAT_DATA_UINT64 },
697 { "hash_elements_max", KSTAT_DATA_UINT64 },
698 { "hash_collisions", KSTAT_DATA_UINT64 },
699 { "hash_chains", KSTAT_DATA_UINT64 },
700 { "hash_chain_max", KSTAT_DATA_UINT64 },
701 { "p", KSTAT_DATA_UINT64 },
702 { "c", KSTAT_DATA_UINT64 },
703 { "c_min", KSTAT_DATA_UINT64 },
704 { "c_max", KSTAT_DATA_UINT64 },
705 { "size", KSTAT_DATA_UINT64 },
706 { "compressed_size", KSTAT_DATA_UINT64 },
707 { "uncompressed_size", KSTAT_DATA_UINT64 },
708 { "overhead_size", KSTAT_DATA_UINT64 },
709 { "hdr_size", KSTAT_DATA_UINT64 },
710 { "data_size", KSTAT_DATA_UINT64 },
711 { "metadata_size", KSTAT_DATA_UINT64 },
712 { "other_size", KSTAT_DATA_UINT64 },
713 { "anon_size", KSTAT_DATA_UINT64 },
714 { "anon_evictable_data", KSTAT_DATA_UINT64 },
715 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
716 { "mru_size", KSTAT_DATA_UINT64 },
717 { "mru_evictable_data", KSTAT_DATA_UINT64 },
718 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
719 { "mru_ghost_size", KSTAT_DATA_UINT64 },
720 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
721 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
722 { "mfu_size", KSTAT_DATA_UINT64 },
723 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
724 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
725 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
726 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
727 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
728 { "l2_hits", KSTAT_DATA_UINT64 },
729 { "l2_misses", KSTAT_DATA_UINT64 },
730 { "l2_feeds", KSTAT_DATA_UINT64 },
731 { "l2_rw_clash", KSTAT_DATA_UINT64 },
732 { "l2_read_bytes", KSTAT_DATA_UINT64 },
733 { "l2_write_bytes", KSTAT_DATA_UINT64 },
734 { "l2_writes_sent", KSTAT_DATA_UINT64 },
735 { "l2_writes_done", KSTAT_DATA_UINT64 },
736 { "l2_writes_error", KSTAT_DATA_UINT64 },
737 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
738 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
739 { "l2_evict_reading", KSTAT_DATA_UINT64 },
740 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
741 { "l2_free_on_write", KSTAT_DATA_UINT64 },
742 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
743 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
744 { "l2_io_error", KSTAT_DATA_UINT64 },
745 { "l2_size", KSTAT_DATA_UINT64 },
746 { "l2_asize", KSTAT_DATA_UINT64 },
747 { "l2_hdr_size", KSTAT_DATA_UINT64 },
748 { "memory_throttle_count", KSTAT_DATA_UINT64 },
749 { "arc_meta_used", KSTAT_DATA_UINT64 },
750 { "arc_meta_limit", KSTAT_DATA_UINT64 },
751 { "arc_meta_max", KSTAT_DATA_UINT64 },
752 { "arc_meta_min", KSTAT_DATA_UINT64 },
753 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
754 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
755 };
756
757 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
758
759 #define ARCSTAT_INCR(stat, val) \
760 atomic_add_64(&arc_stats.stat.value.ui64, (val))
761
762 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
763 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
764
765 #define ARCSTAT_MAX(stat, val) { \
766 uint64_t m; \
767 while ((val) > (m = arc_stats.stat.value.ui64) && \
768 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
769 continue; \
770 }
771
772 #define ARCSTAT_MAXSTAT(stat) \
773 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
774
775 /*
776 * We define a macro to allow ARC hits/misses to be easily broken down by
777 * two separate conditions, giving a total of four different subtypes for
778 * each of hits and misses (so eight statistics total).
779 */
780 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
781 if (cond1) { \
782 if (cond2) { \
783 ARCSTAT_BUMP(arcstat_##stat1##_##stat2##_##stat); \
784 } else { \
785 ARCSTAT_BUMP(arcstat_##stat1##_##notstat2##_##stat); \
786 } \
787 } else { \
788 if (cond2) { \
789 ARCSTAT_BUMP(arcstat_##notstat1##_##stat2##_##stat); \
790 } else { \
791 ARCSTAT_BUMP(arcstat_##notstat1##_##notstat2##_##stat);\
792 } \
793 }
794
795 kstat_t *arc_ksp;
796 static arc_state_t *arc_anon;
797 static arc_state_t *arc_mru;
798 static arc_state_t *arc_mru_ghost;
799 static arc_state_t *arc_mfu;
800 static arc_state_t *arc_mfu_ghost;
801 static arc_state_t *arc_l2c_only;
802
803 /*
804 * There are several ARC variables that are critical to export as kstats --
805 * but we don't want to have to grovel around in the kstat whenever we wish to
806 * manipulate them. For these variables, we therefore define them to be in
807 * terms of the statistic variable. This assures that we are not introducing
808 * the possibility of inconsistency by having shadow copies of the variables,
809 * while still allowing the code to be readable.
810 */
811 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
812 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
813 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
814 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
815 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
816 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
817 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
818
819 /* compressed size of entire arc */
820 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
821 /* uncompressed size of entire arc */
822 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
823 /* number of bytes in the arc from arc_buf_t's */
824 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
825
826 /*
827 * There are also some ARC variables that we want to export, but that are
828 * updated so often that having the canonical representation be the statistic
829 * variable causes a performance bottleneck. We want to use aggsum_t's for these
830 * instead, but still be able to export the kstat in the same way as before.
831 * The solution is to always use the aggsum version, except in the kstat update
832 * callback.
833 */
834 aggsum_t arc_size;
835 aggsum_t arc_meta_used;
836 aggsum_t astat_data_size;
837 aggsum_t astat_metadata_size;
838 aggsum_t astat_hdr_size;
839 aggsum_t astat_other_size;
840 aggsum_t astat_l2_hdr_size;
841
842 static int arc_no_grow; /* Don't try to grow cache size */
843 static uint64_t arc_tempreserve;
844 static uint64_t arc_loaned_bytes;
845
846 typedef struct arc_callback arc_callback_t;
847
848 struct arc_callback {
849 void *acb_private;
850 arc_done_func_t *acb_done;
851 arc_buf_t *acb_buf;
852 boolean_t acb_compressed;
853 zio_t *acb_zio_dummy;
854 arc_callback_t *acb_next;
855 };
856
857 typedef struct arc_write_callback arc_write_callback_t;
858
859 struct arc_write_callback {
860 void *awcb_private;
881 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
882 * | (undefined if L1-only) | | |
883 * +------------------------+ +------------------------+
884 * | l1arc_buf_hdr_t |
885 * | |
886 * | |
887 * | |
888 * | |
889 * +------------------------+
890 *
891 * Because it's possible for the L2ARC to become extremely large, we can wind
892 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
893 * is minimized by only allocating the fields necessary for an L1-cached buffer
894 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
895 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
896 * words in pointers. arc_hdr_realloc() is used to switch a header between
897 * these two allocation states.
898 */
899 typedef struct l1arc_buf_hdr {
900 kmutex_t b_freeze_lock;
901 zio_cksum_t *b_freeze_cksum;
902 #ifdef ZFS_DEBUG
903 /*
904 * Used for debugging with kmem_flags - by allocating and freeing
905 * b_thawed when the buffer is thawed, we get a record of the stack
906 * trace that thawed it.
907 */
908 void *b_thawed;
909 #endif
910
911 arc_buf_t *b_buf;
912 uint32_t b_bufcnt;
913 /* for waiting on writes to complete */
914 kcondvar_t b_cv;
915 uint8_t b_byteswap;
916
917 /* protected by arc state mutex */
918 arc_state_t *b_state;
919 multilist_node_t b_arc_node;
920
921 /* updated atomically */
922 clock_t b_arc_access;
923
924 /* self protecting */
925 refcount_t b_refcnt;
926
927 arc_callback_t *b_acb;
928 abd_t *b_pabd;
929 } l1arc_buf_hdr_t;
930
931 typedef struct l2arc_dev l2arc_dev_t;
932
933 typedef struct l2arc_buf_hdr {
934 /* protected by arc_buf_hdr mutex */
935 l2arc_dev_t *b_dev; /* L2ARC device */
936 uint64_t b_daddr; /* disk address, offset byte */
937
938 list_node_t b_l2node;
939 } l2arc_buf_hdr_t;
940
941 struct arc_buf_hdr {
942 /* protected by hash lock */
943 dva_t b_dva;
944 uint64_t b_birth;
945
946 arc_buf_contents_t b_type;
947 arc_buf_hdr_t *b_hash_next;
948 arc_flags_t b_flags;
949
950 /*
951 * This field stores the size of the data buffer after
952 * compression, and is set in the arc's zio completion handlers.
953 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
954 *
955 * While the block pointers can store up to 32MB in their psize
956 * field, we can only store up to 32MB minus 512B. This is due
957 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
958 * a field of zeros represents 512B in the bp). We can't use a
959 * bias of 1 since we need to reserve a psize of zero, here, to
960 * represent holes and embedded blocks.
961 *
962 * This isn't a problem in practice, since the maximum size of a
963 * buffer is limited to 16MB, so we never need to store 32MB in
964 * this field. Even in the upstream illumos code base, the
965 * maximum size of a buffer is limited to 16MB.
983 #define GHOST_STATE(state) \
984 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
985 (state) == arc_l2c_only)
986
987 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
988 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
989 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
990 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
991 #define HDR_COMPRESSION_ENABLED(hdr) \
992 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
993
994 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
995 #define HDR_L2_READING(hdr) \
996 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
997 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
998 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
999 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1000 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1001 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1002
1003 #define HDR_ISTYPE_METADATA(hdr) \
1004 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1005 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr))
1006
1007 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1008 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1009
1010 /* For storing compression mode in b_flags */
1011 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1012
1013 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1014 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1015 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1016 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1017
1018 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1019 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1020 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1021
1022 /*
1023 * Other sizes
1024 */
1025
1026 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1027 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1028
1029 /*
1030 * Hash table routines
1031 */
1032
1033 #define HT_LOCK_PAD 64
1034
1035 struct ht_lock {
1036 kmutex_t ht_lock;
1037 #ifdef _KERNEL
1038 unsigned char pad[(HT_LOCK_PAD - sizeof (kmutex_t))];
1039 #endif
1040 };
1041
1042 #define BUF_LOCKS 256
1043 typedef struct buf_hash_table {
1044 uint64_t ht_mask;
1045 arc_buf_hdr_t **ht_table;
1046 struct ht_lock ht_locks[BUF_LOCKS];
1047 } buf_hash_table_t;
1048
1049 static buf_hash_table_t buf_hash_table;
1050
1051 #define BUF_HASH_INDEX(spa, dva, birth) \
1052 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1053 #define BUF_HASH_LOCK_NTRY(idx) (buf_hash_table.ht_locks[idx & (BUF_LOCKS-1)])
1054 #define BUF_HASH_LOCK(idx) (&(BUF_HASH_LOCK_NTRY(idx).ht_lock))
1055 #define HDR_LOCK(hdr) \
1056 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1057
1058 uint64_t zfs_crc64_table[256];
1059
1060 /*
1061 * Level 2 ARC
1062 */
1063
1064 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1065 #define L2ARC_HEADROOM 2 /* num of writes */
1066 /*
1067 * If we discover during ARC scan any buffers to be compressed, we boost
1068 * our headroom for the next scanning cycle by this percentage multiple.
1069 */
1070 #define L2ARC_HEADROOM_BOOST 200
1071 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1072 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1073
1074 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1075 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1076
1077 /* L2ARC Performance Tunables */
1078 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1079 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1080 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1081 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1082 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1083 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1084 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1085 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1086 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1087
1088 /*
1089 * L2ARC Internals
1090 */
1091 struct l2arc_dev {
1092 vdev_t *l2ad_vdev; /* vdev */
1093 spa_t *l2ad_spa; /* spa */
1094 uint64_t l2ad_hand; /* next write location */
1095 uint64_t l2ad_start; /* first addr on device */
1096 uint64_t l2ad_end; /* last addr on device */
1097 boolean_t l2ad_first; /* first sweep through */
1098 boolean_t l2ad_writing; /* currently writing */
1099 kmutex_t l2ad_mtx; /* lock for buffer list */
1100 list_t l2ad_buflist; /* buffer list */
1101 list_node_t l2ad_node; /* device list node */
1102 refcount_t l2ad_alloc; /* allocated bytes */
1103 };
1104
1105 static list_t L2ARC_dev_list; /* device list */
1106 static list_t *l2arc_dev_list; /* device list pointer */
1107 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1108 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1109 static list_t L2ARC_free_on_write; /* free after write buf list */
1110 static list_t *l2arc_free_on_write; /* free after write list ptr */
1111 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1112 static uint64_t l2arc_ndev; /* number of devices */
1113
1114 typedef struct l2arc_read_callback {
1115 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1116 blkptr_t l2rcb_bp; /* original blkptr */
1117 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1118 int l2rcb_flags; /* original flags */
1119 abd_t *l2rcb_abd; /* temporary buffer */
1120 } l2arc_read_callback_t;
1121
1122 typedef struct l2arc_write_callback {
1123 l2arc_dev_t *l2wcb_dev; /* device info */
1124 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1125 } l2arc_write_callback_t;
1126
1127 typedef struct l2arc_data_free {
1128 /* protected by l2arc_free_on_write_mtx */
1129 abd_t *l2df_abd;
1130 size_t l2df_size;
1131 arc_buf_contents_t l2df_type;
1132 list_node_t l2df_list_node;
1133 } l2arc_data_free_t;
1134
1135 static kmutex_t l2arc_feed_thr_lock;
1136 static kcondvar_t l2arc_feed_thr_cv;
1137 static uint8_t l2arc_thread_exit;
1138
1139 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1140 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1141 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1142 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1143 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1144 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1145 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1146 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1147 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1148 static boolean_t arc_is_overflowing();
1149 static void arc_buf_watch(arc_buf_t *);
1150
1151 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1152 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1153 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1154 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1155
1156 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1157 static void l2arc_read_done(zio_t *);
1158
1159
1160 /*
1161 * We use Cityhash for this. It's fast, and has good hash properties without
1162 * requiring any large static buffers.
1163 */
1164 static uint64_t
1165 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1166 {
1167 return (cityhash4(spa, dva->dva_word[0], dva->dva_word[1], birth));
1168 }
1169
1170 #define HDR_EMPTY(hdr) \
1171 ((hdr)->b_dva.dva_word[0] == 0 && \
1172 (hdr)->b_dva.dva_word[1] == 0)
1173
1174 #define HDR_EQUAL(spa, dva, birth, hdr) \
1175 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1176 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1177 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1178
1179 static void
1180 buf_discard_identity(arc_buf_hdr_t *hdr)
1181 {
1182 hdr->b_dva.dva_word[0] = 0;
1183 hdr->b_dva.dva_word[1] = 0;
1184 hdr->b_birth = 0;
1185 }
1186
1187 static arc_buf_hdr_t *
1188 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1189 {
1190 const dva_t *dva = BP_IDENTITY(bp);
1191 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1192 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1193 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1194 arc_buf_hdr_t *hdr;
1195
1196 mutex_enter(hash_lock);
1197 for (hdr = buf_hash_table.ht_table[idx]; hdr != NULL;
1198 hdr = hdr->b_hash_next) {
1199 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1200 *lockp = hash_lock;
1201 return (hdr);
1202 }
1203 }
1204 mutex_exit(hash_lock);
1205 *lockp = NULL;
1206 return (NULL);
1207 }
1208
1209 /*
1210 * Insert an entry into the hash table. If there is already an element
1211 * equal to elem in the hash table, then the already existing element
1212 * will be returned and the new element will not be inserted.
1213 * Otherwise returns NULL.
1214 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1215 */
1216 static arc_buf_hdr_t *
1217 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1218 {
1219 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1220 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1221 arc_buf_hdr_t *fhdr;
1222 uint32_t i;
1223
1224 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1225 ASSERT(hdr->b_birth != 0);
1226 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1227
1228 if (lockp != NULL) {
1229 *lockp = hash_lock;
1230 mutex_enter(hash_lock);
1231 } else {
1232 ASSERT(MUTEX_HELD(hash_lock));
1233 }
1234
1235 for (fhdr = buf_hash_table.ht_table[idx], i = 0; fhdr != NULL;
1236 fhdr = fhdr->b_hash_next, i++) {
1237 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1238 return (fhdr);
1239 }
1240
1241 hdr->b_hash_next = buf_hash_table.ht_table[idx];
1242 buf_hash_table.ht_table[idx] = hdr;
1243 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1244
1245 /* collect some hash table performance data */
1246 if (i > 0) {
1247 ARCSTAT_BUMP(arcstat_hash_collisions);
1248 if (i == 1)
1249 ARCSTAT_BUMP(arcstat_hash_chains);
1250
1251 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1252 }
1253
1254 ARCSTAT_BUMP(arcstat_hash_elements);
1255 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1256
1257 return (NULL);
1258 }
1259
1260 static void
1261 buf_hash_remove(arc_buf_hdr_t *hdr)
1262 {
1263 arc_buf_hdr_t *fhdr, **hdrp;
1264 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1265
1266 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1267 ASSERT(HDR_IN_HASH_TABLE(hdr));
1268
1269 hdrp = &buf_hash_table.ht_table[idx];
1270 while ((fhdr = *hdrp) != hdr) {
1271 ASSERT3P(fhdr, !=, NULL);
1272 hdrp = &fhdr->b_hash_next;
1273 }
1274 *hdrp = hdr->b_hash_next;
1275 hdr->b_hash_next = NULL;
1276 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1277
1278 /* collect some hash table performance data */
1279 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1280
1281 if (buf_hash_table.ht_table[idx] &&
1282 buf_hash_table.ht_table[idx]->b_hash_next == NULL)
1283 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1284 }
1285
1286 /*
1287 * Global data structures and functions for the buf kmem cache.
1288 */
1289 static kmem_cache_t *hdr_full_cache;
1290 static kmem_cache_t *hdr_l2only_cache;
1291 static kmem_cache_t *buf_cache;
1292
1293 static void
1294 buf_fini(void)
1295 {
1296 int i;
1297
1298 kmem_free(buf_hash_table.ht_table,
1299 (buf_hash_table.ht_mask + 1) * sizeof (void *));
1300 for (i = 0; i < BUF_LOCKS; i++)
1301 mutex_destroy(&buf_hash_table.ht_locks[i].ht_lock);
1302 kmem_cache_destroy(hdr_full_cache);
1303 kmem_cache_destroy(hdr_l2only_cache);
1304 kmem_cache_destroy(buf_cache);
1305 }
1306
1307 /*
1308 * Constructor callback - called when the cache is empty
1309 * and a new buf is requested.
1310 */
1311 /* ARGSUSED */
1312 static int
1313 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1314 {
1315 arc_buf_hdr_t *hdr = vbuf;
1316
1317 bzero(hdr, HDR_FULL_SIZE);
1318 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1319 refcount_create(&hdr->b_l1hdr.b_refcnt);
1320 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1321 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1404 }
1405
1406 static void
1407 buf_init(void)
1408 {
1409 uint64_t *ct;
1410 uint64_t hsize = 1ULL << 12;
1411 int i, j;
1412
1413 /*
1414 * The hash table is big enough to fill all of physical memory
1415 * with an average block size of zfs_arc_average_blocksize (default 8K).
1416 * By default, the table will take up
1417 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1418 */
1419 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1420 hsize <<= 1;
1421 retry:
1422 buf_hash_table.ht_mask = hsize - 1;
1423 buf_hash_table.ht_table =
1424 kmem_zalloc(hsize * sizeof (void*), KM_NOSLEEP);
1425 if (buf_hash_table.ht_table == NULL) {
1426 ASSERT(hsize > (1ULL << 8));
1427 hsize >>= 1;
1428 goto retry;
1429 }
1430
1431 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1432 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1433 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1434 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1435 NULL, NULL, 0);
1436 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1437 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1438
1439 for (i = 0; i < 256; i++)
1440 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1441 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1442
1443 for (i = 0; i < BUF_LOCKS; i++) {
1444 mutex_init(&buf_hash_table.ht_locks[i].ht_lock,
1445 NULL, MUTEX_DEFAULT, NULL);
1446 }
1447 }
1448
1449 /*
1450 * This is the size that the buf occupies in memory. If the buf is compressed,
1451 * it will correspond to the compressed size. You should use this method of
1452 * getting the buf size unless you explicitly need the logical size.
1453 */
1454 int32_t
1455 arc_buf_size(arc_buf_t *buf)
1456 {
1457 return (ARC_BUF_COMPRESSED(buf) ?
1458 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1459 }
1460
1461 int32_t
1462 arc_buf_lsize(arc_buf_t *buf)
1463 {
1464 return (HDR_GET_LSIZE(buf->b_hdr));
1465 }
1466
1467 enum zio_compress
1468 arc_get_compression(arc_buf_t *buf)
1484 IMPLY(shared, ARC_BUF_SHARED(buf));
1485 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1486
1487 /*
1488 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1489 * already being shared" requirement prevents us from doing that.
1490 */
1491
1492 return (shared);
1493 }
1494
1495 /*
1496 * Free the checksum associated with this header. If there is no checksum, this
1497 * is a no-op.
1498 */
1499 static inline void
1500 arc_cksum_free(arc_buf_hdr_t *hdr)
1501 {
1502 ASSERT(HDR_HAS_L1HDR(hdr));
1503 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1504 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1505 kmem_free(hdr->b_l1hdr.b_freeze_cksum, sizeof (zio_cksum_t));
1506 hdr->b_l1hdr.b_freeze_cksum = NULL;
1507 }
1508 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1509 }
1510
1511 /*
1512 * Return true iff at least one of the bufs on hdr is not compressed.
1513 */
1514 static boolean_t
1515 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1516 {
1517 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1518 if (!ARC_BUF_COMPRESSED(b)) {
1519 return (B_TRUE);
1520 }
1521 }
1522 return (B_FALSE);
1523 }
1524
1525 /*
1526 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1527 * matches the checksum that is stored in the hdr. If there is no checksum,
1528 * or if the buf is compressed, this is a no-op.
1529 */
1530 static void
1531 arc_cksum_verify(arc_buf_t *buf)
1532 {
1533 arc_buf_hdr_t *hdr = buf->b_hdr;
1534 zio_cksum_t zc;
1535
1536 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1537 return;
1538
1539 if (ARC_BUF_COMPRESSED(buf)) {
1540 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1541 arc_hdr_has_uncompressed_buf(hdr));
1542 return;
1543 }
1544
1545 ASSERT(HDR_HAS_L1HDR(hdr));
1546
1547 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1548 if (hdr->b_l1hdr.b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1549 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1550 return;
1551 }
1552
1553 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1554 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_l1hdr.b_freeze_cksum, zc))
1555 panic("buffer modified while frozen!");
1556 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1557 }
1558
1559 static boolean_t
1560 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1561 {
1562 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1563 boolean_t valid_cksum;
1564
1565 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1566 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1567
1568 /*
1569 * We rely on the blkptr's checksum to determine if the block
1570 * is valid or not. When compressed arc is enabled, the l2arc
1571 * writes the block to the l2arc just as it appears in the pool.
1572 * This allows us to use the blkptr's checksum to validate the
1573 * data that we just read off of the l2arc without having to store
1574 * a separate checksum in the arc_buf_hdr_t. However, if compressed
1575 * arc is disabled, then the data written to the l2arc is always
1576 * uncompressed and won't match the block as it exists in the main
1577 * pool. When this is the case, we must first compress it if it is
1578 * compressed on the main pool before we can validate the checksum.
1579 */
1580 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
1581 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
1582 uint64_t lsize = HDR_GET_LSIZE(hdr);
1583 uint64_t csize;
1584
1585 abd_t *cdata = abd_alloc_linear(HDR_GET_PSIZE(hdr), B_TRUE);
1586 csize = zio_compress_data(compress, zio->io_abd,
1587 abd_to_buf(cdata), lsize);
1588
1589 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
1590 if (csize < HDR_GET_PSIZE(hdr)) {
1591 /*
1592 * Compressed blocks are always a multiple of the
1593 * smallest ashift in the pool. Ideally, we would
1594 * like to round up the csize to the next
1595 * spa_min_ashift but that value may have changed
1596 * since the block was last written. Instead,
1597 * we rely on the fact that the hdr's psize
1598 * was set to the psize of the block when it was
1599 * last written. We set the csize to that value
1600 * and zero out any part that should not contain
1601 * data.
1602 */
1603 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
1604 csize = HDR_GET_PSIZE(hdr);
1605 }
1606 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
1607 }
1626 return (valid_cksum);
1627 }
1628
1629 /*
1630 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
1631 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
1632 * isn't modified later on. If buf is compressed or there is already a checksum
1633 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
1634 */
1635 static void
1636 arc_cksum_compute(arc_buf_t *buf)
1637 {
1638 arc_buf_hdr_t *hdr = buf->b_hdr;
1639
1640 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1641 return;
1642
1643 ASSERT(HDR_HAS_L1HDR(hdr));
1644
1645 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
1646 if (hdr->b_l1hdr.b_freeze_cksum != NULL) {
1647 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
1648 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1649 return;
1650 } else if (ARC_BUF_COMPRESSED(buf)) {
1651 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1652 return;
1653 }
1654
1655 ASSERT(!ARC_BUF_COMPRESSED(buf));
1656 hdr->b_l1hdr.b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
1657 KM_SLEEP);
1658 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
1659 hdr->b_l1hdr.b_freeze_cksum);
1660 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1661 arc_buf_watch(buf);
1662 }
1663
1664 #ifndef _KERNEL
1665 typedef struct procctl {
1666 long cmd;
1667 prwatch_t prwatch;
1668 } procctl_t;
1669 #endif
1670
1671 /* ARGSUSED */
1672 static void
1673 arc_buf_unwatch(arc_buf_t *buf)
1674 {
1675 #ifndef _KERNEL
1676 if (arc_watch) {
1677 int result;
1678 procctl_t ctl;
1679 ctl.cmd = PCWATCH;
1691 arc_buf_watch(arc_buf_t *buf)
1692 {
1693 #ifndef _KERNEL
1694 if (arc_watch) {
1695 int result;
1696 procctl_t ctl;
1697 ctl.cmd = PCWATCH;
1698 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
1699 ctl.prwatch.pr_size = arc_buf_size(buf);
1700 ctl.prwatch.pr_wflags = WA_WRITE;
1701 result = write(arc_procfd, &ctl, sizeof (ctl));
1702 ASSERT3U(result, ==, sizeof (ctl));
1703 }
1704 #endif
1705 }
1706
1707 static arc_buf_contents_t
1708 arc_buf_type(arc_buf_hdr_t *hdr)
1709 {
1710 arc_buf_contents_t type;
1711 if (HDR_ISTYPE_METADATA(hdr)) {
1712 type = ARC_BUFC_METADATA;
1713 } else {
1714 type = ARC_BUFC_DATA;
1715 }
1716 VERIFY3U(hdr->b_type, ==, type);
1717 return (type);
1718 }
1719
1720 boolean_t
1721 arc_is_metadata(arc_buf_t *buf)
1722 {
1723 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
1724 }
1725
1726 static uint32_t
1727 arc_bufc_to_flags(arc_buf_contents_t type)
1728 {
1729 switch (type) {
1730 case ARC_BUFC_DATA:
1731 /* metadata field is 0 if buffer contains normal data */
1732 return (0);
1733 case ARC_BUFC_METADATA:
1734 return (ARC_FLAG_BUFC_METADATA);
1735 default:
1736 break;
1737 }
1738 panic("undefined ARC buffer type!");
1739 return ((uint32_t)-1);
1740 }
1741
1742 void
1743 arc_buf_thaw(arc_buf_t *buf)
1744 {
1745 arc_buf_hdr_t *hdr = buf->b_hdr;
1746
1747 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
1748 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
1749
1750 arc_cksum_verify(buf);
1751
1752 /*
1753 * Compressed buffers do not manipulate the b_freeze_cksum or
1754 * allocate b_thawed.
1755 */
1756 if (ARC_BUF_COMPRESSED(buf)) {
1757 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1758 arc_hdr_has_uncompressed_buf(hdr));
1759 return;
1760 }
1761
1762 ASSERT(HDR_HAS_L1HDR(hdr));
1763 arc_cksum_free(hdr);
1764
1765 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1766 #ifdef ZFS_DEBUG
1767 if (zfs_flags & ZFS_DEBUG_MODIFY) {
1768 if (hdr->b_l1hdr.b_thawed != NULL)
1769 kmem_free(hdr->b_l1hdr.b_thawed, 1);
1770 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
1771 }
1772 #endif
1773
1774 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1775
1776 arc_buf_unwatch(buf);
1777 }
1778
1779 void
1780 arc_buf_freeze(arc_buf_t *buf)
1781 {
1782 arc_buf_hdr_t *hdr = buf->b_hdr;
1783 kmutex_t *hash_lock;
1784
1785 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1786 return;
1787
1788 if (ARC_BUF_COMPRESSED(buf)) {
1789 ASSERT(hdr->b_l1hdr.b_freeze_cksum == NULL ||
1790 arc_hdr_has_uncompressed_buf(hdr));
1791 return;
1792 }
1793
1794 hash_lock = HDR_LOCK(hdr);
1795 mutex_enter(hash_lock);
1796
1797 ASSERT(HDR_HAS_L1HDR(hdr));
1798 ASSERT(hdr->b_l1hdr.b_freeze_cksum != NULL ||
1799 hdr->b_l1hdr.b_state == arc_anon);
1800 arc_cksum_compute(buf);
1801 mutex_exit(hash_lock);
1802 }
1803
1804 /*
1805 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
1806 * the following functions should be used to ensure that the flags are
1807 * updated in a thread-safe way. When manipulating the flags either
1808 * the hash_lock must be held or the hdr must be undiscoverable. This
1809 * ensures that we're not racing with any other threads when updating
1810 * the flags.
1811 */
1812 static inline void
1813 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
1814 {
1815 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
1816 hdr->b_flags |= flags;
1817 }
1818
1870 ASSERT(!ARC_BUF_COMPRESSED(buf));
1871
1872 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
1873 from = from->b_next) {
1874 /* can't use our own data buffer */
1875 if (from == buf) {
1876 continue;
1877 }
1878
1879 if (!ARC_BUF_COMPRESSED(from)) {
1880 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
1881 copied = B_TRUE;
1882 break;
1883 }
1884 }
1885
1886 /*
1887 * There were no decompressed bufs, so there should not be a
1888 * checksum on the hdr either.
1889 */
1890 EQUIV(!copied, hdr->b_l1hdr.b_freeze_cksum == NULL);
1891
1892 return (copied);
1893 }
1894
1895 /*
1896 * Given a buf that has a data buffer attached to it, this function will
1897 * efficiently fill the buf with data of the specified compression setting from
1898 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
1899 * are already sharing a data buf, no copy is performed.
1900 *
1901 * If the buf is marked as compressed but uncompressed data was requested, this
1902 * will allocate a new data buffer for the buf, remove that flag, and fill the
1903 * buf with uncompressed data. You can't request a compressed buf on a hdr with
1904 * uncompressed data, and (since we haven't added support for it yet) if you
1905 * want compressed data your buf must already be marked as compressed and have
1906 * the correct-sized data buffer.
1907 */
1908 static int
1909 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
1910 {
1949 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
1950
1951 /* We increased the size of b_data; update overhead */
1952 ARCSTAT_INCR(arcstat_overhead_size,
1953 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
1954 }
1955
1956 /*
1957 * Regardless of the buf's previous compression settings, it
1958 * should not be compressed at the end of this function.
1959 */
1960 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
1961
1962 /*
1963 * Try copying the data from another buf which already has a
1964 * decompressed version. If that's not possible, it's time to
1965 * bite the bullet and decompress the data from the hdr.
1966 */
1967 if (arc_buf_try_copy_decompressed_data(buf)) {
1968 /* Skip byteswapping and checksumming (already done) */
1969 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, !=, NULL);
1970 return (0);
1971 } else {
1972 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
1973 hdr->b_l1hdr.b_pabd, buf->b_data,
1974 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1975
1976 /*
1977 * Absent hardware errors or software bugs, this should
1978 * be impossible, but log it anyway so we can debug it.
1979 */
1980 if (error != 0) {
1981 zfs_dbgmsg(
1982 "hdr %p, compress %d, psize %d, lsize %d",
1983 hdr, HDR_GET_COMPRESS(hdr),
1984 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
1985 return (SET_ERROR(EIO));
1986 }
1987 }
1988 }
1989
2212
2213 /*
2214 * An L1 header always exists here, since if we're
2215 * moving to some L1-cached state (i.e. not l2c_only or
2216 * anonymous), we realloc the header to add an L1hdr
2217 * beforehand.
2218 */
2219 ASSERT(HDR_HAS_L1HDR(hdr));
2220 multilist_insert(new_state->arcs_list[buftype], hdr);
2221
2222 if (GHOST_STATE(new_state)) {
2223 ASSERT0(bufcnt);
2224 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2225 update_new = B_TRUE;
2226 }
2227 arc_evictable_space_increment(hdr, new_state);
2228 }
2229 }
2230
2231 ASSERT(!HDR_EMPTY(hdr));
2232 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr))
2233 buf_hash_remove(hdr);
2234
2235 /* adjust state sizes (ignore arc_l2c_only) */
2236
2237 if (update_new && new_state != arc_l2c_only) {
2238 ASSERT(HDR_HAS_L1HDR(hdr));
2239 if (GHOST_STATE(new_state)) {
2240 ASSERT0(bufcnt);
2241
2242 /*
2243 * When moving a header to a ghost state, we first
2244 * remove all arc buffers. Thus, we'll have a
2245 * bufcnt of zero, and no arc buffer to use for
2246 * the reference. As a result, we use the arc
2247 * header pointer for the reference.
2248 */
2249 (void) refcount_add_many(&new_state->arcs_size,
2250 HDR_GET_LSIZE(hdr), hdr);
2251 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2252 } else {
2253 uint32_t buffers = 0;
2326 continue;
2327
2328 (void) refcount_remove_many(
2329 &old_state->arcs_size, arc_buf_size(buf),
2330 buf);
2331 }
2332 ASSERT3U(bufcnt, ==, buffers);
2333 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2334 (void) refcount_remove_many(
2335 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2336 }
2337 }
2338
2339 if (HDR_HAS_L1HDR(hdr))
2340 hdr->b_l1hdr.b_state = new_state;
2341
2342 /*
2343 * L2 headers should never be on the L2 state list since they don't
2344 * have L1 headers allocated.
2345 */
2346 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]) &&
2347 multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2348 }
2349
2350 void
2351 arc_space_consume(uint64_t space, arc_space_type_t type)
2352 {
2353 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2354
2355 switch (type) {
2356 case ARC_SPACE_DATA:
2357 aggsum_add(&astat_data_size, space);
2358 break;
2359 case ARC_SPACE_META:
2360 aggsum_add(&astat_metadata_size, space);
2361 break;
2362 case ARC_SPACE_OTHER:
2363 aggsum_add(&astat_other_size, space);
2364 break;
2365 case ARC_SPACE_HDRS:
2366 aggsum_add(&astat_hdr_size, space);
2367 break;
2368 case ARC_SPACE_L2HDRS:
2369 aggsum_add(&astat_l2_hdr_size, space);
2370 break;
2371 }
2372
2373 if (type != ARC_SPACE_DATA)
2374 aggsum_add(&arc_meta_used, space);
2375
2376 aggsum_add(&arc_size, space);
2377 }
2378
2379 void
2380 arc_space_return(uint64_t space, arc_space_type_t type)
2381 {
2382 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2383
2384 switch (type) {
2385 case ARC_SPACE_DATA:
2386 aggsum_add(&astat_data_size, -space);
2387 break;
2388 case ARC_SPACE_META:
2389 aggsum_add(&astat_metadata_size, -space);
2390 break;
2391 case ARC_SPACE_OTHER:
2392 aggsum_add(&astat_other_size, -space);
2393 break;
2394 case ARC_SPACE_HDRS:
2395 aggsum_add(&astat_hdr_size, -space);
2396 break;
2397 case ARC_SPACE_L2HDRS:
2398 aggsum_add(&astat_l2_hdr_size, -space);
2399 break;
2400 }
2401
2402 if (type != ARC_SPACE_DATA) {
2403 ASSERT(aggsum_compare(&arc_meta_used, space) >= 0);
2404 /*
2405 * We use the upper bound here rather than the precise value
2406 * because the arc_meta_max value doesn't need to be
2407 * precise. It's only consumed by humans via arcstats.
2408 */
2409 if (arc_meta_max < aggsum_upper_bound(&arc_meta_used))
2410 arc_meta_max = aggsum_upper_bound(&arc_meta_used);
2411 aggsum_add(&arc_meta_used, -space);
2412 }
2413
2414 ASSERT(aggsum_compare(&arc_size, space) >= 0);
2415 aggsum_add(&arc_size, -space);
2416 }
2417
2418 /*
2419 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2420 * with the hdr's b_pabd.
2421 */
2422 static boolean_t
2423 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2424 {
2425 /*
2426 * The criteria for sharing a hdr's data are:
2427 * 1. the hdr's compression matches the buf's compression
2428 * 2. the hdr doesn't need to be byteswapped
2429 * 3. the hdr isn't already being shared
2430 * 4. the buf is either compressed or it is the last buf in the hdr list
2431 *
2432 * Criterion #4 maintains the invariant that shared uncompressed
2433 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2434 * might ask, "if a compressed buf is allocated first, won't that be the
2435 * last thing in the list?", but in that case it's impossible to create
2449 return (buf_compressed == hdr_compressed &&
2450 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2451 !HDR_SHARED_DATA(hdr) &&
2452 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2453 }
2454
2455 /*
2456 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2457 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2458 * copy was made successfully, or an error code otherwise.
2459 */
2460 static int
2461 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2462 boolean_t fill, arc_buf_t **ret)
2463 {
2464 arc_buf_t *buf;
2465
2466 ASSERT(HDR_HAS_L1HDR(hdr));
2467 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2468 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2469 hdr->b_type == ARC_BUFC_METADATA);
2470 ASSERT3P(ret, !=, NULL);
2471 ASSERT3P(*ret, ==, NULL);
2472
2473 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2474 buf->b_hdr = hdr;
2475 buf->b_data = NULL;
2476 buf->b_next = hdr->b_l1hdr.b_buf;
2477 buf->b_flags = 0;
2478
2479 add_reference(hdr, tag);
2480
2481 /*
2482 * We're about to change the hdr's b_flags. We must either
2483 * hold the hash_lock or be undiscoverable.
2484 */
2485 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2486
2487 /*
2488 * Only honor requests for compressed bufs if the hdr is actually
2489 * compressed.
2529 */
2530 if (fill) {
2531 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2532 }
2533
2534 return (0);
2535 }
2536
2537 static char *arc_onloan_tag = "onloan";
2538
2539 static inline void
2540 arc_loaned_bytes_update(int64_t delta)
2541 {
2542 atomic_add_64(&arc_loaned_bytes, delta);
2543
2544 /* assert that it did not wrap around */
2545 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2546 }
2547
2548 /*
2549 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
2550 * flight data by arc_tempreserve_space() until they are "returned". Loaned
2551 * buffers must be returned to the arc before they can be used by the DMU or
2552 * freed.
2553 */
2554 arc_buf_t *
2555 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
2556 {
2557 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
2558 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
2559
2560 arc_loaned_bytes_update(size);
2561
2562 return (buf);
2563 }
2564
2565 arc_buf_t *
2566 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
2567 enum zio_compress compression_type)
2568 {
2617 list_insert_head(l2arc_free_on_write, df);
2618 mutex_exit(&l2arc_free_on_write_mtx);
2619 }
2620
2621 static void
2622 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
2623 {
2624 arc_state_t *state = hdr->b_l1hdr.b_state;
2625 arc_buf_contents_t type = arc_buf_type(hdr);
2626 uint64_t size = arc_hdr_size(hdr);
2627
2628 /* protected by hash lock, if in the hash table */
2629 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
2630 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
2631 ASSERT(state != arc_anon && state != arc_l2c_only);
2632
2633 (void) refcount_remove_many(&state->arcs_esize[type],
2634 size, hdr);
2635 }
2636 (void) refcount_remove_many(&state->arcs_size, size, hdr);
2637 if (type == ARC_BUFC_METADATA) {
2638 arc_space_return(size, ARC_SPACE_META);
2639 } else {
2640 ASSERT(type == ARC_BUFC_DATA);
2641 arc_space_return(size, ARC_SPACE_DATA);
2642 }
2643
2644 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
2645 }
2646
2647 /*
2648 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
2649 * data buffer, we transfer the refcount ownership to the hdr and update
2650 * the appropriate kstats.
2651 */
2652 static void
2653 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2654 {
2655 arc_state_t *state = hdr->b_l1hdr.b_state;
2656
2657 ASSERT(arc_can_share(hdr, buf));
2658 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2659 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2660
2661 /*
2662 * Start sharing the data buffer. We transfer the
2663 * refcount ownership to the hdr since it always owns
2664 * the refcount whenever an arc_buf_t is shared.
2665 */
2666 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
2667 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
2668 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
2669 HDR_ISTYPE_METADATA(hdr));
2670 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
2671 buf->b_flags |= ARC_BUF_FLAG_SHARED;
2672
2673 /*
2674 * Since we've transferred ownership to the hdr we need
2675 * to increment its compressed and uncompressed kstats and
2676 * decrement the overhead size.
2677 */
2678 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2679 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2680 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
2681 }
2682
2683 static void
2684 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2685 {
2686 arc_state_t *state = hdr->b_l1hdr.b_state;
2687
2688 ASSERT(arc_buf_is_shared(buf));
2689 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2842
2843 /* clean up the buf */
2844 buf->b_hdr = NULL;
2845 kmem_cache_free(buf_cache, buf);
2846 }
2847
2848 static void
2849 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
2850 {
2851 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2852 ASSERT(HDR_HAS_L1HDR(hdr));
2853 ASSERT(!HDR_SHARED_DATA(hdr));
2854
2855 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2856 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
2857 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2858 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2859
2860 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
2861 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
2862 }
2863
2864 static void
2865 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
2866 {
2867 ASSERT(HDR_HAS_L1HDR(hdr));
2868 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2869
2870 /*
2871 * If the hdr is currently being written to the l2arc then
2872 * we defer freeing the data by adding it to the l2arc_free_on_write
2873 * list. The l2arc will free the data once it's finished
2874 * writing it to the l2arc device.
2875 */
2876 if (HDR_L2_WRITING(hdr)) {
2877 arc_hdr_free_on_write(hdr);
2878 ARCSTAT_BUMP(arcstat_l2_free_on_write);
2879 } else {
2880 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
2881 arc_hdr_size(hdr), hdr);
2882 }
2883 hdr->b_l1hdr.b_pabd = NULL;
2884 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
2885
2886 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
2887 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
2888 }
2889
2890 static arc_buf_hdr_t *
2891 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
2892 enum zio_compress compression_type, arc_buf_contents_t type)
2893 {
2894 arc_buf_hdr_t *hdr;
2895
2896 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA);
2897
2898 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
2899 ASSERT(HDR_EMPTY(hdr));
2900 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2901 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
2902 HDR_SET_PSIZE(hdr, psize);
2903 HDR_SET_LSIZE(hdr, lsize);
2904 hdr->b_spa = spa;
2905 hdr->b_type = type;
2906 hdr->b_flags = 0;
2907 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
2908 arc_hdr_set_compress(hdr, compression_type);
2909
2910 hdr->b_l1hdr.b_state = arc_anon;
2911 hdr->b_l1hdr.b_arc_access = 0;
2912 hdr->b_l1hdr.b_bufcnt = 0;
2913 hdr->b_l1hdr.b_buf = NULL;
2914
2915 /*
2916 * Allocate the hdr's buffer. This will contain either
2917 * the compressed or uncompressed data depending on the block
2918 * it references and compressed arc enablement.
2919 */
2920 arc_hdr_alloc_pabd(hdr);
2945
2946 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
2947 buf_hash_remove(hdr);
2948
2949 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
2950
2951 if (new == hdr_full_cache) {
2952 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
2953 /*
2954 * arc_access and arc_change_state need to be aware that a
2955 * header has just come out of L2ARC, so we set its state to
2956 * l2c_only even though it's about to change.
2957 */
2958 nhdr->b_l1hdr.b_state = arc_l2c_only;
2959
2960 /* Verify previous threads set to NULL before freeing */
2961 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
2962 } else {
2963 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2964 ASSERT0(hdr->b_l1hdr.b_bufcnt);
2965 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
2966
2967 /*
2968 * If we've reached here, We must have been called from
2969 * arc_evict_hdr(), as such we should have already been
2970 * removed from any ghost list we were previously on
2971 * (which protects us from racing with arc_evict_state),
2972 * thus no locking is needed during this check.
2973 */
2974 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
2975
2976 /*
2977 * A buffer must not be moved into the arc_l2c_only
2978 * state if it's not finished being written out to the
2979 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
2980 * might try to be accessed, even though it was removed.
2981 */
2982 VERIFY(!HDR_L2_WRITING(hdr));
2983 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2984
2985 #ifdef ZFS_DEBUG
3050 /*
3051 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3052 * for bufs containing metadata.
3053 */
3054 arc_buf_t *
3055 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3056 enum zio_compress compression_type)
3057 {
3058 ASSERT3U(lsize, >, 0);
3059 ASSERT3U(lsize, >=, psize);
3060 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3061 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3062
3063 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3064 compression_type, ARC_BUFC_DATA);
3065 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3066
3067 arc_buf_t *buf = NULL;
3068 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3069 arc_buf_thaw(buf);
3070 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
3071
3072 if (!arc_buf_is_shared(buf)) {
3073 /*
3074 * To ensure that the hdr has the correct data in it if we call
3075 * arc_decompress() on this buf before it's been written to
3076 * disk, it's easiest if we just set up sharing between the
3077 * buf and the hdr.
3078 */
3079 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3080 arc_hdr_free_pabd(hdr);
3081 arc_share_buf(hdr, buf);
3082 }
3083
3084 return (buf);
3085 }
3086
3087 static void
3088 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3089 {
3090 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3091 l2arc_dev_t *dev = l2hdr->b_dev;
3092 uint64_t psize = arc_hdr_size(hdr);
3093
3094 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3095 ASSERT(HDR_HAS_L2HDR(hdr));
3096
3097 list_remove(&dev->l2ad_buflist, hdr);
3098
3099 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3100 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3101
3102 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3103
3104 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3105 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3106 }
3107
3108 static void
3109 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3110 {
3111 if (HDR_HAS_L1HDR(hdr)) {
3112 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3113 hdr->b_l1hdr.b_bufcnt > 0);
3114 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3115 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3116 }
3117 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3118 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3119
3120 if (!HDR_EMPTY(hdr))
3121 buf_discard_identity(hdr);
3122
3123 if (HDR_HAS_L2HDR(hdr)) {
3124 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3125 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3126
3127 if (!buflist_held)
3128 mutex_enter(&dev->l2ad_mtx);
3129
3130 /*
3131 * Even though we checked this conditional above, we
3132 * need to check this again now that we have the
3133 * l2ad_mtx. This is because we could be racing with
3134 * another thread calling l2arc_evict() which might have
3135 * destroyed this header's L2 portion as we were waiting
3136 * to acquire the l2ad_mtx. If that happens, we don't
3137 * want to re-destroy the header's L2 portion.
3138 */
3139 if (HDR_HAS_L2HDR(hdr))
3140 arc_hdr_l2hdr_destroy(hdr);
3141
3142 if (!buflist_held)
3143 mutex_exit(&dev->l2ad_mtx);
3144 }
3145
3146 if (HDR_HAS_L1HDR(hdr)) {
3147 arc_cksum_free(hdr);
3148
3149 while (hdr->b_l1hdr.b_buf != NULL)
3150 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3151
3152 #ifdef ZFS_DEBUG
3153 if (hdr->b_l1hdr.b_thawed != NULL) {
3154 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3155 hdr->b_l1hdr.b_thawed = NULL;
3156 }
3157 #endif
3158
3159 if (hdr->b_l1hdr.b_pabd != NULL) {
3160 arc_hdr_free_pabd(hdr);
3161 }
3162 }
3163
3164 ASSERT3P(hdr->b_hash_next, ==, NULL);
3165 if (HDR_HAS_L1HDR(hdr)) {
3201 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3202 * state of the header is dependent on it's state prior to entering this
3203 * function. The following transitions are possible:
3204 *
3205 * - arc_mru -> arc_mru_ghost
3206 * - arc_mfu -> arc_mfu_ghost
3207 * - arc_mru_ghost -> arc_l2c_only
3208 * - arc_mru_ghost -> deleted
3209 * - arc_mfu_ghost -> arc_l2c_only
3210 * - arc_mfu_ghost -> deleted
3211 */
3212 static int64_t
3213 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3214 {
3215 arc_state_t *evicted_state, *state;
3216 int64_t bytes_evicted = 0;
3217
3218 ASSERT(MUTEX_HELD(hash_lock));
3219 ASSERT(HDR_HAS_L1HDR(hdr));
3220
3221 state = hdr->b_l1hdr.b_state;
3222 if (GHOST_STATE(state)) {
3223 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3224 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3225
3226 /*
3227 * l2arc_write_buffers() relies on a header's L1 portion
3228 * (i.e. its b_pabd field) during it's write phase.
3229 * Thus, we cannot push a header onto the arc_l2c_only
3230 * state (removing it's L1 piece) until the header is
3231 * done being written to the l2arc.
3232 */
3233 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3234 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3235 return (bytes_evicted);
3236 }
3237
3238 ARCSTAT_BUMP(arcstat_deleted);
3239 bytes_evicted += HDR_GET_LSIZE(hdr);
3240
3589 * prevents us from trying to evict more from a state's list than
3590 * is "evictable", and to skip evicting altogether when passed a
3591 * negative value for "bytes". In contrast, arc_evict_state() will
3592 * evict everything it can, when passed a negative value for "bytes".
3593 */
3594 static uint64_t
3595 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
3596 arc_buf_contents_t type)
3597 {
3598 int64_t delta;
3599
3600 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
3601 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
3602 return (arc_evict_state(state, spa, delta, type));
3603 }
3604
3605 return (0);
3606 }
3607
3608 /*
3609 * Evict metadata buffers from the cache, such that arc_meta_used is
3610 * capped by the arc_meta_limit tunable.
3611 */
3612 static uint64_t
3613 arc_adjust_meta(uint64_t meta_used)
3614 {
3615 uint64_t total_evicted = 0;
3616 int64_t target;
3617
3618 /*
3619 * If we're over the meta limit, we want to evict enough
3620 * metadata to get back under the meta limit. We don't want to
3621 * evict so much that we drop the MRU below arc_p, though. If
3622 * we're over the meta limit more than we're over arc_p, we
3623 * evict some from the MRU here, and some from the MFU below.
3624 */
3625 target = MIN((int64_t)(meta_used - arc_meta_limit),
3626 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3627 refcount_count(&arc_mru->arcs_size) - arc_p));
3628
3629 total_evicted += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3630
3631 /*
3632 * Similar to the above, we want to evict enough bytes to get us
3633 * below the meta limit, but not so much as to drop us below the
3634 * space allotted to the MFU (which is defined as arc_c - arc_p).
3635 */
3636 target = MIN((int64_t)(meta_used - arc_meta_limit),
3637 (int64_t)(refcount_count(&arc_mfu->arcs_size) -
3638 (arc_c - arc_p)));
3639
3640 total_evicted += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3641
3642 return (total_evicted);
3643 }
3644
3645 /*
3646 * Return the type of the oldest buffer in the given arc state
3647 *
3648 * This function will select a random sublist of type ARC_BUFC_DATA and
3649 * a random sublist of type ARC_BUFC_METADATA. The tail of each sublist
3650 * is compared, and the type which contains the "older" buffer will be
3651 * returned.
3652 */
3653 static arc_buf_contents_t
3654 arc_adjust_type(arc_state_t *state)
3655 {
3656 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
3657 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
3658 int data_idx = multilist_get_random_index(data_ml);
3659 int meta_idx = multilist_get_random_index(meta_ml);
3660 multilist_sublist_t *data_mls;
3661 multilist_sublist_t *meta_mls;
3662 arc_buf_contents_t type;
3663 arc_buf_hdr_t *data_hdr;
3664 arc_buf_hdr_t *meta_hdr;
3665
3666 /*
3667 * We keep the sublist lock until we're finished, to prevent
3668 * the headers from being destroyed via arc_evict_state().
3669 */
3670 data_mls = multilist_sublist_lock(data_ml, data_idx);
3671 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
3672
3673 /*
3674 * These two loops are to ensure we skip any markers that
3675 * might be at the tail of the lists due to arc_evict_state().
3676 */
3677
3678 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
3679 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
3680 if (data_hdr->b_spa != 0)
3681 break;
3682 }
3683
3684 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
3685 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
3686 if (meta_hdr->b_spa != 0)
3687 break;
3688 }
3689
3690 if (data_hdr == NULL && meta_hdr == NULL) {
3691 type = ARC_BUFC_DATA;
3692 } else if (data_hdr == NULL) {
3693 ASSERT3P(meta_hdr, !=, NULL);
3694 type = ARC_BUFC_METADATA;
3695 } else if (meta_hdr == NULL) {
3696 ASSERT3P(data_hdr, !=, NULL);
3697 type = ARC_BUFC_DATA;
3698 } else {
3699 ASSERT3P(data_hdr, !=, NULL);
3700 ASSERT3P(meta_hdr, !=, NULL);
3701
3702 /* The headers can't be on the sublist without an L1 header */
3703 ASSERT(HDR_HAS_L1HDR(data_hdr));
3704 ASSERT(HDR_HAS_L1HDR(meta_hdr));
3705
3706 if (data_hdr->b_l1hdr.b_arc_access <
3707 meta_hdr->b_l1hdr.b_arc_access) {
3708 type = ARC_BUFC_DATA;
3709 } else {
3710 type = ARC_BUFC_METADATA;
3711 }
3712 }
3713
3714 multilist_sublist_unlock(meta_mls);
3715 multilist_sublist_unlock(data_mls);
3716
3717 return (type);
3718 }
3719
3720 /*
3721 * Evict buffers from the cache, such that arc_size is capped by arc_c.
3722 */
3723 static uint64_t
3724 arc_adjust(void)
3725 {
3726 uint64_t total_evicted = 0;
3727 uint64_t bytes;
3728 int64_t target;
3729 uint64_t asize = aggsum_value(&arc_size);
3730 uint64_t ameta = aggsum_value(&arc_meta_used);
3731
3732 /*
3733 * If we're over arc_meta_limit, we want to correct that before
3734 * potentially evicting data buffers below.
3735 */
3736 total_evicted += arc_adjust_meta(ameta);
3737
3738 /*
3739 * Adjust MRU size
3740 *
3741 * If we're over the target cache size, we want to evict enough
3742 * from the list to get back to our target size. We don't want
3743 * to evict too much from the MRU, such that it drops below
3744 * arc_p. So, if we're over our target cache size more than
3745 * the MRU is over arc_p, we'll evict enough to get back to
3746 * arc_p here, and then evict more from the MFU below.
3747 */
3748 target = MIN((int64_t)(asize - arc_c),
3749 (int64_t)(refcount_count(&arc_anon->arcs_size) +
3750 refcount_count(&arc_mru->arcs_size) + ameta - arc_p));
3751
3752 /*
3753 * If we're below arc_meta_min, always prefer to evict data.
3754 * Otherwise, try to satisfy the requested number of bytes to
3755 * evict from the type which contains older buffers; in an
3756 * effort to keep newer buffers in the cache regardless of their
3757 * type. If we cannot satisfy the number of bytes from this
3758 * type, spill over into the next type.
3759 */
3760 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
3761 ameta > arc_meta_min) {
3762 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3763 total_evicted += bytes;
3764
3765 /*
3766 * If we couldn't evict our target number of bytes from
3767 * metadata, we try to get the rest from data.
3768 */
3769 target -= bytes;
3770
3771 total_evicted +=
3772 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3773 } else {
3774 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
3775 total_evicted += bytes;
3776
3777 /*
3778 * If we couldn't evict our target number of bytes from
3779 * data, we try to get the rest from metadata.
3780 */
3781 target -= bytes;
3782
3783 total_evicted +=
3784 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
3785 }
3786
3787 /*
3788 * Adjust MFU size
3789 *
3790 * Now that we've tried to evict enough from the MRU to get its
3791 * size back to arc_p, if we're still above the target cache
3792 * size, we evict the rest from the MFU.
3793 */
3794 target = asize - arc_c;
3795
3796 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
3797 ameta > arc_meta_min) {
3798 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3799 total_evicted += bytes;
3800
3801 /*
3802 * If we couldn't evict our target number of bytes from
3803 * metadata, we try to get the rest from data.
3804 */
3805 target -= bytes;
3806
3807 total_evicted +=
3808 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3809 } else {
3810 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
3811 total_evicted += bytes;
3812
3813 /*
3814 * If we couldn't evict our target number of bytes from
3815 * data, we try to get the rest from data.
3816 */
3817 target -= bytes;
3818
3819 total_evicted +=
3820 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
3821 }
3822
3823 /*
3824 * Adjust ghost lists
3825 *
3826 * In addition to the above, the ARC also defines target values
3827 * for the ghost lists. The sum of the mru list and mru ghost
3828 * list should never exceed the target size of the cache, and
3829 * the sum of the mru list, mfu list, mru ghost list, and mfu
3830 * ghost list should never exceed twice the target size of the
3831 * cache. The following logic enforces these limits on the ghost
3832 * caches, and evicts from them as needed.
3833 */
3834 target = refcount_count(&arc_mru->arcs_size) +
3835 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
3836
3837 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
3838 total_evicted += bytes;
3839
3840 target -= bytes;
3841
3842 total_evicted +=
3843 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
3844
3845 /*
3846 * We assume the sum of the mru list and mfu list is less than
3847 * or equal to arc_c (we enforced this above), which means we
3848 * can use the simpler of the two equations below:
3849 *
3850 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
3851 * mru ghost + mfu ghost <= arc_c
3852 */
3853 target = refcount_count(&arc_mru_ghost->arcs_size) +
3854 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
3855
3856 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
3857 total_evicted += bytes;
3858
3859 target -= bytes;
3860
3861 total_evicted +=
3862 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
3863
3864 return (total_evicted);
3865 }
3866
3867 void
3868 arc_flush(spa_t *spa, boolean_t retry)
3869 {
3870 uint64_t guid = 0;
3871
3872 /*
3873 * If retry is B_TRUE, a spa must not be specified since we have
3874 * no good way to determine if all of a spa's buffers have been
3875 * evicted from an arc state.
3876 */
3877 ASSERT(!retry || spa == 0);
3878
3879 if (spa != NULL)
3880 guid = spa_load_guid(spa);
3881
3882 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_DATA, retry);
3883 (void) arc_flush_state(arc_mru, guid, ARC_BUFC_METADATA, retry);
3884
3885 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_DATA, retry);
3886 (void) arc_flush_state(arc_mfu, guid, ARC_BUFC_METADATA, retry);
3887
3888 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_DATA, retry);
3889 (void) arc_flush_state(arc_mru_ghost, guid, ARC_BUFC_METADATA, retry);
3890
3891 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
3892 (void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
3893 }
3894
3895 void
3896 arc_shrink(int64_t to_free)
3897 {
3898 uint64_t asize = aggsum_value(&arc_size);
3899 if (arc_c > arc_c_min) {
3900
3901 if (arc_c > arc_c_min + to_free)
3902 atomic_add_64(&arc_c, -to_free);
3903 else
3904 arc_c = arc_c_min;
3905
3906 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
3907 if (asize < arc_c)
3908 arc_c = MAX(asize, arc_c_min);
3909 if (arc_p > arc_c)
3910 arc_p = (arc_c >> 1);
3911 ASSERT(arc_c >= arc_c_min);
3912 ASSERT((int64_t)arc_p >= 0);
3913 }
3914
3915 if (asize > arc_c)
3916 (void) arc_adjust();
3917 }
3918
3919 typedef enum free_memory_reason_t {
3920 FMR_UNKNOWN,
3921 FMR_NEEDFREE,
3922 FMR_LOTSFREE,
3923 FMR_SWAPFS_MINFREE,
3924 FMR_PAGES_PP_MAXIMUM,
3925 FMR_HEAP_ARENA,
3926 FMR_ZIO_ARENA,
3927 } free_memory_reason_t;
3928
3929 int64_t last_free_memory;
3930 free_memory_reason_t last_free_reason;
3931
3932 /*
3933 * Additional reserve of pages for pp_reserve.
3934 */
3935 int64_t arc_pages_pp_reserve = 64;
4059 * is under memory pressure and that the arc should adjust accordingly.
4060 */
4061 static boolean_t
4062 arc_reclaim_needed(void)
4063 {
4064 return (arc_available_memory() < 0);
4065 }
4066
4067 static void
4068 arc_kmem_reap_now(void)
4069 {
4070 size_t i;
4071 kmem_cache_t *prev_cache = NULL;
4072 kmem_cache_t *prev_data_cache = NULL;
4073 extern kmem_cache_t *zio_buf_cache[];
4074 extern kmem_cache_t *zio_data_buf_cache[];
4075 extern kmem_cache_t *range_seg_cache;
4076 extern kmem_cache_t *abd_chunk_cache;
4077
4078 #ifdef _KERNEL
4079 if (aggsum_compare(&arc_meta_used, arc_meta_limit) >= 0) {
4080 /*
4081 * We are exceeding our meta-data cache limit.
4082 * Purge some DNLC entries to release holds on meta-data.
4083 */
4084 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4085 }
4086 #if defined(__i386)
4087 /*
4088 * Reclaim unused memory from all kmem caches.
4089 */
4090 kmem_reap();
4091 #endif
4092 #endif
4093
4094 /*
4095 * If a kmem reap is already active, don't schedule more. We must
4096 * check for this because kmem_cache_reap_soon() won't actually
4097 * block on the cache being reaped (this is to prevent callers from
4098 * becoming implicitly blocked by a system-wide kmem reap -- which,
4099 * on a system with many, many full magazines, can take minutes).
4100 */
4101 if (kmem_cache_reap_active())
4102 return;
4218 #endif
4219 arc_shrink(to_free);
4220 }
4221 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4222 arc_no_grow = B_TRUE;
4223 } else if (gethrtime() >= growtime) {
4224 arc_no_grow = B_FALSE;
4225 }
4226
4227 mutex_enter(&arc_reclaim_lock);
4228
4229 /*
4230 * If evicted is zero, we couldn't evict anything via
4231 * arc_adjust(). This could be due to hash lock
4232 * collisions, but more likely due to the majority of
4233 * arc buffers being unevictable. Therefore, even if
4234 * arc_size is above arc_c, another pass is unlikely to
4235 * be helpful and could potentially cause us to enter an
4236 * infinite loop.
4237 */
4238 if (aggsum_compare(&arc_size, arc_c) <= 0|| evicted == 0) {
4239 /*
4240 * We're either no longer overflowing, or we
4241 * can't evict anything more, so we should wake
4242 * up any threads before we go to sleep.
4243 */
4244 cv_broadcast(&arc_reclaim_waiters_cv);
4245
4246 /*
4247 * Block until signaled, or after one second (we
4248 * might need to perform arc_kmem_reap_now()
4249 * even if we aren't being signalled)
4250 */
4251 CALLB_CPR_SAFE_BEGIN(&cpr);
4252 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4253 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4254 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4255 }
4256 }
4257
4258 arc_reclaim_thread_exit = B_FALSE;
4300 delta = MIN(bytes * mult, arc_p);
4301 arc_p = MAX(arc_p_min, arc_p - delta);
4302 }
4303 ASSERT((int64_t)arc_p >= 0);
4304
4305 if (arc_reclaim_needed()) {
4306 cv_signal(&arc_reclaim_thread_cv);
4307 return;
4308 }
4309
4310 if (arc_no_grow)
4311 return;
4312
4313 if (arc_c >= arc_c_max)
4314 return;
4315
4316 /*
4317 * If we're within (2 * maxblocksize) bytes of the target
4318 * cache size, increment the target cache size
4319 */
4320 if (aggsum_compare(&arc_size, arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) >
4321 0) {
4322 atomic_add_64(&arc_c, (int64_t)bytes);
4323 if (arc_c > arc_c_max)
4324 arc_c = arc_c_max;
4325 else if (state == arc_anon)
4326 atomic_add_64(&arc_p, (int64_t)bytes);
4327 if (arc_p > arc_c)
4328 arc_p = arc_c;
4329 }
4330 ASSERT((int64_t)arc_p >= 0);
4331 }
4332
4333 /*
4334 * Check if arc_size has grown past our upper threshold, determined by
4335 * zfs_arc_overflow_shift.
4336 */
4337 static boolean_t
4338 arc_is_overflowing(void)
4339 {
4340 /* Always allow at least one block of overflow */
4341 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
4342 arc_c >> zfs_arc_overflow_shift);
4343
4344 /*
4345 * We just compare the lower bound here for performance reasons. Our
4346 * primary goals are to make sure that the arc never grows without
4347 * bound, and that it can reach its maximum size. This check
4348 * accomplishes both goals. The maximum amount we could run over by is
4349 * 2 * aggsum_borrow_multiplier * NUM_CPUS * the average size of a block
4350 * in the ARC. In practice, that's in the tens of MB, which is low
4351 * enough to be safe.
4352 */
4353 return (aggsum_lower_bound(&arc_size) >= arc_c + overflow);
4354 }
4355
4356 static abd_t *
4357 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4358 {
4359 arc_buf_contents_t type = arc_buf_type(hdr);
4360
4361 arc_get_data_impl(hdr, size, tag);
4362 if (type == ARC_BUFC_METADATA) {
4363 return (abd_alloc(size, B_TRUE));
4364 } else {
4365 ASSERT(type == ARC_BUFC_DATA);
4366 return (abd_alloc(size, B_FALSE));
4367 }
4368 }
4369
4370 static void *
4371 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4372 {
4373 arc_buf_contents_t type = arc_buf_type(hdr);
4374
4375 arc_get_data_impl(hdr, size, tag);
4376 if (type == ARC_BUFC_METADATA) {
4377 return (zio_buf_alloc(size));
4378 } else {
4379 ASSERT(type == ARC_BUFC_DATA);
4380 return (zio_data_buf_alloc(size));
4381 }
4382 }
4383
4384 /*
4385 * Allocate a block and return it to the caller. If we are hitting the
4386 * hard limit for the cache size, we must sleep, waiting for the eviction
4387 * thread to catch up. If we're past the target size but below the hard
4388 * limit, we'll only signal the reclaim thread and continue on.
4389 */
4390 static void
4391 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4392 {
4393 arc_state_t *state = hdr->b_l1hdr.b_state;
4394 arc_buf_contents_t type = arc_buf_type(hdr);
4395
4396 arc_adapt(size, state);
4415 /*
4416 * Now that we've acquired the lock, we may no longer be
4417 * over the overflow limit, lets check.
4418 *
4419 * We're ignoring the case of spurious wake ups. If that
4420 * were to happen, it'd let this thread consume an ARC
4421 * buffer before it should have (i.e. before we're under
4422 * the overflow limit and were signalled by the reclaim
4423 * thread). As long as that is a rare occurrence, it
4424 * shouldn't cause any harm.
4425 */
4426 if (arc_is_overflowing()) {
4427 cv_signal(&arc_reclaim_thread_cv);
4428 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
4429 }
4430
4431 mutex_exit(&arc_reclaim_lock);
4432 }
4433
4434 VERIFY3U(hdr->b_type, ==, type);
4435 if (type == ARC_BUFC_METADATA) {
4436 arc_space_consume(size, ARC_SPACE_META);
4437 } else {
4438 arc_space_consume(size, ARC_SPACE_DATA);
4439 }
4440
4441 /*
4442 * Update the state size. Note that ghost states have a
4443 * "ghost size" and so don't need to be updated.
4444 */
4445 if (!GHOST_STATE(state)) {
4446
4447 (void) refcount_add_many(&state->arcs_size, size, tag);
4448
4449 /*
4450 * If this is reached via arc_read, the link is
4451 * protected by the hash lock. If reached via
4452 * arc_buf_alloc, the header should not be accessed by
4453 * any other thread. And, if reached via arc_read_done,
4454 * the hash lock will protect it if it's found in the
4455 * hash table; otherwise no other thread should be
4456 * trying to [add|remove]_reference it.
4457 */
4458 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4459 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4460 (void) refcount_add_many(&state->arcs_esize[type],
4461 size, tag);
4462 }
4463
4464 /*
4465 * If we are growing the cache, and we are adding anonymous
4466 * data, and we have outgrown arc_p, update arc_p
4467 */
4468 if (aggsum_compare(&arc_size, arc_c) < 0 &&
4469 hdr->b_l1hdr.b_state == arc_anon &&
4470 (refcount_count(&arc_anon->arcs_size) +
4471 refcount_count(&arc_mru->arcs_size) > arc_p))
4472 arc_p = MIN(arc_c, arc_p + size);
4473 }
4474 }
4475
4476 static void
4477 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
4478 {
4479 arc_free_data_impl(hdr, size, tag);
4480 abd_free(abd);
4481 }
4482
4483 static void
4484 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
4485 {
4486 arc_buf_contents_t type = arc_buf_type(hdr);
4487
4488 arc_free_data_impl(hdr, size, tag);
4489 if (type == ARC_BUFC_METADATA) {
4490 zio_buf_free(buf, size);
4491 } else {
4492 ASSERT(type == ARC_BUFC_DATA);
4493 zio_data_buf_free(buf, size);
4494 }
4495 }
4496
4497 /*
4498 * Free the arc data buffer.
4499 */
4500 static void
4501 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
4502 {
4503 arc_state_t *state = hdr->b_l1hdr.b_state;
4504 arc_buf_contents_t type = arc_buf_type(hdr);
4505
4506 /* protected by hash lock, if in the hash table */
4507 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
4508 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
4509 ASSERT(state != arc_anon && state != arc_l2c_only);
4510
4511 (void) refcount_remove_many(&state->arcs_esize[type],
4512 size, tag);
4513 }
4514 (void) refcount_remove_many(&state->arcs_size, size, tag);
4515
4516 VERIFY3U(hdr->b_type, ==, type);
4517 if (type == ARC_BUFC_METADATA) {
4518 arc_space_return(size, ARC_SPACE_META);
4519 } else {
4520 ASSERT(type == ARC_BUFC_DATA);
4521 arc_space_return(size, ARC_SPACE_DATA);
4522 }
4523 }
4524
4525 /*
4526 * This routine is called whenever a buffer is accessed.
4527 * NOTE: the hash lock is dropped in this function.
4528 */
4529 static void
4530 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
4531 {
4532 clock_t now;
4533
4534 ASSERT(MUTEX_HELD(hash_lock));
4535 ASSERT(HDR_HAS_L1HDR(hdr));
4536
4537 if (hdr->b_l1hdr.b_state == arc_anon) {
4643 }
4644
4645 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4646 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4647 arc_change_state(new_state, hdr, hash_lock);
4648
4649 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
4650 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
4651 /*
4652 * This buffer is on the 2nd Level ARC.
4653 */
4654
4655 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
4656 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
4657 arc_change_state(arc_mfu, hdr, hash_lock);
4658 } else {
4659 ASSERT(!"invalid arc state");
4660 }
4661 }
4662
4663 /* a generic arc_done_func_t which you can use */
4664 /* ARGSUSED */
4665 void
4666 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
4667 {
4668 if (zio == NULL || zio->io_error == 0)
4669 bcopy(buf->b_data, arg, arc_buf_size(buf));
4670 arc_buf_destroy(buf, arg);
4671 }
4672
4673 /* a generic arc_done_func_t */
4674 void
4675 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
4676 {
4677 arc_buf_t **bufp = arg;
4678 if (zio && zio->io_error) {
4679 arc_buf_destroy(buf, arg);
4680 *bufp = NULL;
4681 } else {
4682 *bufp = buf;
4785 zio->io_error = error;
4786 }
4787 }
4788 hdr->b_l1hdr.b_acb = NULL;
4789 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
4790 if (callback_cnt == 0) {
4791 ASSERT(HDR_PREFETCH(hdr));
4792 ASSERT0(hdr->b_l1hdr.b_bufcnt);
4793 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
4794 }
4795
4796 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
4797 callback_list != NULL);
4798
4799 if (no_zio_error) {
4800 arc_hdr_verify(hdr, zio->io_bp);
4801 } else {
4802 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
4803 if (hdr->b_l1hdr.b_state != arc_anon)
4804 arc_change_state(arc_anon, hdr, hash_lock);
4805 if (HDR_IN_HASH_TABLE(hdr))
4806 buf_hash_remove(hdr);
4807 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
4808 }
4809
4810 /*
4811 * Broadcast before we drop the hash_lock to avoid the possibility
4812 * that the hdr (and hence the cv) might be freed before we get to
4813 * the cv_broadcast().
4814 */
4815 cv_broadcast(&hdr->b_l1hdr.b_cv);
4816
4817 if (hash_lock != NULL) {
4818 mutex_exit(hash_lock);
4819 } else {
4820 /*
4821 * This block was freed while we waited for the read to
4822 * complete. It has been removed from the hash table and
4823 * moved to the anonymous state (so that it won't show up
4824 * in the cache).
4825 */
4826 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
4829
4830 /* execute each callback and free its structure */
4831 while ((acb = callback_list) != NULL) {
4832 if (acb->acb_done)
4833 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
4834
4835 if (acb->acb_zio_dummy != NULL) {
4836 acb->acb_zio_dummy->io_error = zio->io_error;
4837 zio_nowait(acb->acb_zio_dummy);
4838 }
4839
4840 callback_list = acb->acb_next;
4841 kmem_free(acb, sizeof (arc_callback_t));
4842 }
4843
4844 if (freeable)
4845 arc_hdr_destroy(hdr);
4846 }
4847
4848 /*
4849 * "Read" the block at the specified DVA (in bp) via the
4850 * cache. If the block is found in the cache, invoke the provided
4851 * callback immediately and return. Note that the `zio' parameter
4852 * in the callback will be NULL in this case, since no IO was
4853 * required. If the block is not in the cache pass the read request
4854 * on to the spa with a substitute callback function, so that the
4855 * requested block will be added to the cache.
4856 *
4857 * If a read request arrives for a block that has a read in-progress,
4858 * either wait for the in-progress read to complete (and return the
4859 * results); or, if this is a read with a "done" func, add a record
4860 * to the read to invoke the "done" func when the read completes,
4861 * and return; or just return.
4862 *
4863 * arc_read_done() will invoke all the requested "done" functions
4864 * for readers of this block.
4865 */
4866 int
4867 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
4868 void *private, zio_priority_t priority, int zio_flags,
4968 ARCSTAT_BUMP(
4969 arcstat_demand_hit_predictive_prefetch);
4970 arc_hdr_clear_flags(hdr,
4971 ARC_FLAG_PREDICTIVE_PREFETCH);
4972 }
4973 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
4974
4975 /* Get a buf with the desired data in it. */
4976 VERIFY0(arc_buf_alloc_impl(hdr, private,
4977 compressed_read, B_TRUE, &buf));
4978 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
4979 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
4980 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
4981 }
4982 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
4983 arc_access(hdr, hash_lock);
4984 if (*arc_flags & ARC_FLAG_L2CACHE)
4985 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
4986 mutex_exit(hash_lock);
4987 ARCSTAT_BUMP(arcstat_hits);
4988 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
4989 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
4990 data, metadata, hits);
4991
4992 if (done)
4993 done(NULL, buf, private);
4994 } else {
4995 uint64_t lsize = BP_GET_LSIZE(bp);
4996 uint64_t psize = BP_GET_PSIZE(bp);
4997 arc_callback_t *acb;
4998 vdev_t *vd = NULL;
4999 uint64_t addr = 0;
5000 boolean_t devw = B_FALSE;
5001 uint64_t size;
5002
5003 if (hdr == NULL) {
5004 /* this block is not in the cache */
5005 arc_buf_hdr_t *exists = NULL;
5006 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5007 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5008 BP_GET_COMPRESS(bp), type);
5009
5010 if (!BP_IS_EMBEDDED(bp)) {
5011 hdr->b_dva = *BP_IDENTITY(bp);
5012 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5013 exists = buf_hash_insert(hdr, &hash_lock);
5014 }
5015 if (exists != NULL) {
5016 /* somebody beat us to the hash insert */
5017 mutex_exit(hash_lock);
5018 buf_discard_identity(hdr);
5019 arc_hdr_destroy(hdr);
5020 goto top; /* restart the IO request */
5021 }
5022 } else {
5023 /*
5024 * This block is in the ghost cache. If it was L2-only
5025 * (and thus didn't have an L1 hdr), we realloc the
5026 * header to add an L1 hdr.
5027 */
5028 if (!HDR_HAS_L1HDR(hdr)) {
5029 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5030 hdr_full_cache);
5031 }
5032 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5033 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5034 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5035 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5036 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5037 ASSERT3P(hdr->b_l1hdr.b_freeze_cksum, ==, NULL);
5038
5039 /*
5040 * This is a delicate dance that we play here.
5041 * This hdr is in the ghost list so we access it
5042 * to move it out of the ghost list before we
5043 * initiate the read. If it's a prefetch then
5044 * it won't have a callback so we'll remove the
5045 * reference that arc_buf_alloc_impl() created. We
5046 * do this after we've called arc_access() to
5047 * avoid hitting an assert in remove_reference().
5048 */
5049 arc_access(hdr, hash_lock);
5050 arc_hdr_alloc_pabd(hdr);
5051 }
5052 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5053 size = arc_hdr_size(hdr);
5054
5055 /*
5056 * If compression is enabled on the hdr, then will do
5057 * RAW I/O and will store the compressed data in the hdr's
5069 if (BP_GET_LEVEL(bp) > 0)
5070 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5071 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5072 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5073 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5074
5075 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5076 acb->acb_done = done;
5077 acb->acb_private = private;
5078 acb->acb_compressed = compressed_read;
5079
5080 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5081 hdr->b_l1hdr.b_acb = acb;
5082 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5083
5084 if (HDR_HAS_L2HDR(hdr) &&
5085 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5086 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5087 addr = hdr->b_l2hdr.b_daddr;
5088 /*
5089 * Lock out L2ARC device removal.
5090 */
5091 if (vdev_is_dead(vd) ||
5092 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5093 vd = NULL;
5094 }
5095
5096 if (priority == ZIO_PRIORITY_ASYNC_READ)
5097 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5098 else
5099 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5100
5101 if (hash_lock != NULL)
5102 mutex_exit(hash_lock);
5103
5104 /*
5105 * At this point, we have a level 1 cache miss. Try again in
5106 * L2ARC if possible.
5107 */
5108 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5109
5110 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5111 uint64_t, lsize, zbookmark_phys_t *, zb);
5112 ARCSTAT_BUMP(arcstat_misses);
5113 ARCSTAT_CONDSTAT(!HDR_PREFETCH(hdr),
5114 demand, prefetch, !HDR_ISTYPE_METADATA(hdr),
5115 data, metadata, misses);
5116
5117 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5118 /*
5119 * Read from the L2ARC if the following are true:
5120 * 1. The L2ARC vdev was previously cached.
5121 * 2. This buffer still has L2ARC metadata.
5122 * 3. This buffer isn't currently writing to the L2ARC.
5123 * 4. The L2ARC entry wasn't evicted, which may
5124 * also have invalidated the vdev.
5125 * 5. This isn't prefetch and l2arc_noprefetch is set.
5126 */
5127 if (HDR_HAS_L2HDR(hdr) &&
5128 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5129 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5130 l2arc_read_callback_t *cb;
5131 abd_t *abd;
5132 uint64_t asize;
5133
5134 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5135 ARCSTAT_BUMP(arcstat_l2_hits);
5136
5137 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5138 KM_SLEEP);
5139 cb->l2rcb_hdr = hdr;
5140 cb->l2rcb_bp = *bp;
5141 cb->l2rcb_zb = *zb;
5142 cb->l2rcb_flags = zio_flags;
5143
5144 asize = vdev_psize_to_asize(vd, size);
5145 if (asize != size) {
5146 abd = abd_alloc_for_io(asize,
5147 HDR_ISTYPE_METADATA(hdr));
5148 cb->l2rcb_abd = abd;
5149 } else {
5150 abd = hdr->b_l1hdr.b_pabd;
5151 }
5152
5153 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5154 addr + asize <= vd->vdev_psize -
5155 VDEV_LABEL_END_SIZE);
5156
5157 /*
5158 * l2arc read. The SCL_L2ARC lock will be
5159 * released by l2arc_read_done().
5160 * Issue a null zio if the underlying buffer
5161 * was squashed to zero size by compression.
5162 */
5163 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5164 ZIO_COMPRESS_EMPTY);
5165 rzio = zio_read_phys(pio, vd, addr,
5166 asize, abd,
5167 ZIO_CHECKSUM_OFF,
5168 l2arc_read_done, cb, priority,
5169 zio_flags | ZIO_FLAG_DONT_CACHE |
5170 ZIO_FLAG_CANFAIL |
5171 ZIO_FLAG_DONT_PROPAGATE |
5172 ZIO_FLAG_DONT_RETRY, B_FALSE);
5173 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5174 zio_t *, rzio);
5175 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5176
5177 if (*arc_flags & ARC_FLAG_NOWAIT) {
5178 zio_nowait(rzio);
5179 return (0);
5180 }
5181
5182 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5183 if (zio_wait(rzio) == 0)
5184 return (0);
5185
5186 /* l2arc read error; goto zio_read() */
5187 } else {
5188 DTRACE_PROBE1(l2arc__miss,
5189 arc_buf_hdr_t *, hdr);
5190 ARCSTAT_BUMP(arcstat_l2_misses);
5191 if (HDR_L2_WRITING(hdr))
5192 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5193 spa_config_exit(spa, SCL_L2ARC, vd);
5194 }
5195 } else {
5427 hdr->b_l1hdr.b_bufcnt -= 1;
5428 arc_cksum_verify(buf);
5429 arc_buf_unwatch(buf);
5430
5431 mutex_exit(hash_lock);
5432
5433 /*
5434 * Allocate a new hdr. The new hdr will contain a b_pabd
5435 * buffer which will be freed in arc_write().
5436 */
5437 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
5438 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
5439 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
5440 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
5441 VERIFY3U(nhdr->b_type, ==, type);
5442 ASSERT(!HDR_SHARED_DATA(nhdr));
5443
5444 nhdr->b_l1hdr.b_buf = buf;
5445 nhdr->b_l1hdr.b_bufcnt = 1;
5446 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
5447 buf->b_hdr = nhdr;
5448
5449 mutex_exit(&buf->b_evict_lock);
5450 (void) refcount_add_many(&arc_anon->arcs_size,
5451 arc_buf_size(buf), buf);
5452 } else {
5453 mutex_exit(&buf->b_evict_lock);
5454 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
5455 /* protected by hash lock, or hdr is on arc_anon */
5456 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
5457 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5458 arc_change_state(arc_anon, hdr, hash_lock);
5459 hdr->b_l1hdr.b_arc_access = 0;
5460 mutex_exit(hash_lock);
5461
5462 buf_discard_identity(hdr);
5463 arc_buf_thaw(buf);
5464 }
5465 }
5466
5637 kmutex_t *hash_lock;
5638
5639 ASSERT3U(zio->io_error, ==, 0);
5640
5641 arc_cksum_verify(buf);
5642
5643 exists = buf_hash_insert(hdr, &hash_lock);
5644 if (exists != NULL) {
5645 /*
5646 * This can only happen if we overwrite for
5647 * sync-to-convergence, because we remove
5648 * buffers from the hash table when we arc_free().
5649 */
5650 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
5651 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5652 panic("bad overwrite, hdr=%p exists=%p",
5653 (void *)hdr, (void *)exists);
5654 ASSERT(refcount_is_zero(
5655 &exists->b_l1hdr.b_refcnt));
5656 arc_change_state(arc_anon, exists, hash_lock);
5657 mutex_exit(hash_lock);
5658 arc_hdr_destroy(exists);
5659 exists = buf_hash_insert(hdr, &hash_lock);
5660 ASSERT3P(exists, ==, NULL);
5661 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
5662 /* nopwrite */
5663 ASSERT(zio->io_prop.zp_nopwrite);
5664 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
5665 panic("bad nopwrite, hdr=%p exists=%p",
5666 (void *)hdr, (void *)exists);
5667 } else {
5668 /* Dedup */
5669 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
5670 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
5671 ASSERT(BP_GET_DEDUP(zio->io_bp));
5672 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
5673 }
5674 }
5675 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5676 /* if it's not anon, we are doing a scrub */
5677 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
5678 arc_access(hdr, hash_lock);
5679 mutex_exit(hash_lock);
5680 } else {
5681 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5682 }
5683
5684 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5685 callback->awcb_done(zio, buf, callback->awcb_private);
5686
5687 abd_put(zio->io_abd);
5688 kmem_free(callback, sizeof (arc_write_callback_t));
5689 }
5690
5691 zio_t *
5692 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
5693 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
5694 arc_done_func_t *children_ready, arc_done_func_t *physdone,
5695 arc_done_func_t *done, void *private, zio_priority_t priority,
5696 int zio_flags, const zbookmark_phys_t *zb)
5697 {
5698 arc_buf_hdr_t *hdr = buf->b_hdr;
5699 arc_write_callback_t *callback;
5700 zio_t *zio;
5701 zio_prop_t localprop = *zp;
5702
5703 ASSERT3P(ready, !=, NULL);
5704 ASSERT3P(done, !=, NULL);
5705 ASSERT(!HDR_IO_ERROR(hdr));
5706 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5707 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5708 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
5709 if (l2arc)
5710 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5711 if (ARC_BUF_COMPRESSED(buf)) {
5712 /*
5713 * We're writing a pre-compressed buffer. Make the
5714 * compression algorithm requested by the zio_prop_t match
5715 * the pre-compressed buffer's compression algorithm.
5716 */
5737 * the hdr then we need to break that relationship here.
5738 * The hdr will remain with a NULL data pointer and the
5739 * buf will take sole ownership of the block.
5740 */
5741 if (arc_buf_is_shared(buf)) {
5742 arc_unshare_buf(hdr, buf);
5743 } else {
5744 arc_hdr_free_pabd(hdr);
5745 }
5746 VERIFY3P(buf->b_data, !=, NULL);
5747 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
5748 }
5749 ASSERT(!arc_buf_is_shared(buf));
5750 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5751
5752 zio = zio_write(pio, spa, txg, bp,
5753 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
5754 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
5755 (children_ready != NULL) ? arc_write_children_ready : NULL,
5756 arc_write_physdone, arc_write_done, callback,
5757 priority, zio_flags, zb);
5758
5759 return (zio);
5760 }
5761
5762 static int
5763 arc_memory_throttle(uint64_t reserve, uint64_t txg)
5764 {
5765 #ifdef _KERNEL
5766 uint64_t available_memory = ptob(freemem);
5767 static uint64_t page_load = 0;
5768 static uint64_t last_txg = 0;
5769
5770 #if defined(__i386)
5771 available_memory =
5772 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
5773 #endif
5774
5775 if (freemem > physmem * arc_lotsfree_percent / 100)
5776 return (0);
5777
5829
5830 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
5831 arc_loaned_bytes), 0);
5832
5833 /*
5834 * Writes will, almost always, require additional memory allocations
5835 * in order to compress/encrypt/etc the data. We therefore need to
5836 * make sure that there is sufficient available memory for this.
5837 */
5838 error = arc_memory_throttle(reserve, txg);
5839 if (error != 0)
5840 return (error);
5841
5842 /*
5843 * Throttle writes when the amount of dirty data in the cache
5844 * gets too large. We try to keep the cache less than half full
5845 * of dirty blocks so that our sync times don't grow too large.
5846 * Note: if two requests come in concurrently, we might let them
5847 * both succeed, when one of them should fail. Not a huge deal.
5848 */
5849
5850 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
5851 anon_size > arc_c / 4) {
5852 uint64_t meta_esize =
5853 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
5854 uint64_t data_esize =
5855 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
5856 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
5857 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
5858 arc_tempreserve >> 10, meta_esize >> 10,
5859 data_esize >> 10, reserve >> 10, arc_c >> 10);
5860 return (SET_ERROR(ERESTART));
5861 }
5862 atomic_add_64(&arc_tempreserve, reserve);
5863 return (0);
5864 }
5865
5866 static void
5867 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
5868 kstat_named_t *evict_data, kstat_named_t *evict_metadata)
5869 {
5870 size->value.ui64 = refcount_count(&state->arcs_size);
5871 evict_data->value.ui64 =
5872 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
5873 evict_metadata->value.ui64 =
5874 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
5875 }
5876
5877 static int
5878 arc_kstat_update(kstat_t *ksp, int rw)
5879 {
5880 arc_stats_t *as = ksp->ks_data;
5881
5882 if (rw == KSTAT_WRITE) {
5883 return (EACCES);
5884 } else {
5885 arc_kstat_update_state(arc_anon,
5886 &as->arcstat_anon_size,
5887 &as->arcstat_anon_evictable_data,
5888 &as->arcstat_anon_evictable_metadata);
5889 arc_kstat_update_state(arc_mru,
5890 &as->arcstat_mru_size,
5891 &as->arcstat_mru_evictable_data,
5892 &as->arcstat_mru_evictable_metadata);
5893 arc_kstat_update_state(arc_mru_ghost,
5894 &as->arcstat_mru_ghost_size,
5895 &as->arcstat_mru_ghost_evictable_data,
5896 &as->arcstat_mru_ghost_evictable_metadata);
5897 arc_kstat_update_state(arc_mfu,
5898 &as->arcstat_mfu_size,
5899 &as->arcstat_mfu_evictable_data,
5900 &as->arcstat_mfu_evictable_metadata);
5901 arc_kstat_update_state(arc_mfu_ghost,
5902 &as->arcstat_mfu_ghost_size,
5903 &as->arcstat_mfu_ghost_evictable_data,
5904 &as->arcstat_mfu_ghost_evictable_metadata);
5905
5906 ARCSTAT(arcstat_size) = aggsum_value(&arc_size);
5907 ARCSTAT(arcstat_meta_used) = aggsum_value(&arc_meta_used);
5908 ARCSTAT(arcstat_data_size) = aggsum_value(&astat_data_size);
5909 ARCSTAT(arcstat_metadata_size) =
5910 aggsum_value(&astat_metadata_size);
5911 ARCSTAT(arcstat_hdr_size) = aggsum_value(&astat_hdr_size);
5912 ARCSTAT(arcstat_other_size) = aggsum_value(&astat_other_size);
5913 ARCSTAT(arcstat_l2_hdr_size) = aggsum_value(&astat_l2_hdr_size);
5914 }
5915
5916 return (0);
5917 }
5918
5919 /*
5920 * This function *must* return indices evenly distributed between all
5921 * sublists of the multilist. This is needed due to how the ARC eviction
5922 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
5923 * distributed between all sublists and uses this assumption when
5924 * deciding which sublist to evict from and how much to evict from it.
5925 */
5926 unsigned int
5927 arc_state_multilist_index_func(multilist_t *ml, void *obj)
5928 {
5929 arc_buf_hdr_t *hdr = obj;
5930
5931 /*
5932 * We rely on b_dva to generate evenly distributed index
5933 * numbers using buf_hash below. So, as an added precaution,
5943 * on insertion, as this index can be recalculated on removal.
5944 *
5945 * Also, the low order bits of the hash value are thought to be
5946 * distributed evenly. Otherwise, in the case that the multilist
5947 * has a power of two number of sublists, each sublists' usage
5948 * would not be evenly distributed.
5949 */
5950 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
5951 multilist_get_num_sublists(ml));
5952 }
5953
5954 static void
5955 arc_state_init(void)
5956 {
5957 arc_anon = &ARC_anon;
5958 arc_mru = &ARC_mru;
5959 arc_mru_ghost = &ARC_mru_ghost;
5960 arc_mfu = &ARC_mfu;
5961 arc_mfu_ghost = &ARC_mfu_ghost;
5962 arc_l2c_only = &ARC_l2c_only;
5963
5964 arc_mru->arcs_list[ARC_BUFC_METADATA] =
5965 multilist_create(sizeof (arc_buf_hdr_t),
5966 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5967 arc_state_multilist_index_func);
5968 arc_mru->arcs_list[ARC_BUFC_DATA] =
5969 multilist_create(sizeof (arc_buf_hdr_t),
5970 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5971 arc_state_multilist_index_func);
5972 arc_mru_ghost->arcs_list[ARC_BUFC_METADATA] =
5973 multilist_create(sizeof (arc_buf_hdr_t),
5974 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5975 arc_state_multilist_index_func);
5976 arc_mru_ghost->arcs_list[ARC_BUFC_DATA] =
5977 multilist_create(sizeof (arc_buf_hdr_t),
5978 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5979 arc_state_multilist_index_func);
5980 arc_mfu->arcs_list[ARC_BUFC_METADATA] =
5981 multilist_create(sizeof (arc_buf_hdr_t),
5982 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5983 arc_state_multilist_index_func);
5984 arc_mfu->arcs_list[ARC_BUFC_DATA] =
5985 multilist_create(sizeof (arc_buf_hdr_t),
5986 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5987 arc_state_multilist_index_func);
5988 arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA] =
5989 multilist_create(sizeof (arc_buf_hdr_t),
5990 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5991 arc_state_multilist_index_func);
5992 arc_mfu_ghost->arcs_list[ARC_BUFC_DATA] =
5993 multilist_create(sizeof (arc_buf_hdr_t),
5994 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5995 arc_state_multilist_index_func);
5996 arc_l2c_only->arcs_list[ARC_BUFC_METADATA] =
5997 multilist_create(sizeof (arc_buf_hdr_t),
5998 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
5999 arc_state_multilist_index_func);
6000 arc_l2c_only->arcs_list[ARC_BUFC_DATA] =
6001 multilist_create(sizeof (arc_buf_hdr_t),
6002 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6003 arc_state_multilist_index_func);
6004
6005 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6006 refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6007 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6008 refcount_create(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6009 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6010 refcount_create(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6011 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6012 refcount_create(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6013 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6014 refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6015 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6016 refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6017
6018 refcount_create(&arc_anon->arcs_size);
6019 refcount_create(&arc_mru->arcs_size);
6020 refcount_create(&arc_mru_ghost->arcs_size);
6021 refcount_create(&arc_mfu->arcs_size);
6022 refcount_create(&arc_mfu_ghost->arcs_size);
6023 refcount_create(&arc_l2c_only->arcs_size);
6024
6025 aggsum_init(&arc_meta_used, 0);
6026 aggsum_init(&arc_size, 0);
6027 aggsum_init(&astat_data_size, 0);
6028 aggsum_init(&astat_metadata_size, 0);
6029 aggsum_init(&astat_hdr_size, 0);
6030 aggsum_init(&astat_other_size, 0);
6031 aggsum_init(&astat_l2_hdr_size, 0);
6032 }
6033
6034 static void
6035 arc_state_fini(void)
6036 {
6037 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6038 refcount_destroy(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6039 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_METADATA]);
6040 refcount_destroy(&arc_mru->arcs_esize[ARC_BUFC_DATA]);
6041 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_METADATA]);
6042 refcount_destroy(&arc_mru_ghost->arcs_esize[ARC_BUFC_DATA]);
6043 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
6044 refcount_destroy(&arc_mfu->arcs_esize[ARC_BUFC_DATA]);
6045 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_METADATA]);
6046 refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
6047 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
6048 refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
6049
6050 refcount_destroy(&arc_anon->arcs_size);
6051 refcount_destroy(&arc_mru->arcs_size);
6052 refcount_destroy(&arc_mru_ghost->arcs_size);
6053 refcount_destroy(&arc_mfu->arcs_size);
6054 refcount_destroy(&arc_mfu_ghost->arcs_size);
6055 refcount_destroy(&arc_l2c_only->arcs_size);
6056
6057 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_METADATA]);
6058 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
6059 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_METADATA]);
6060 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA]);
6061 multilist_destroy(arc_mru->arcs_list[ARC_BUFC_DATA]);
6062 multilist_destroy(arc_mru_ghost->arcs_list[ARC_BUFC_DATA]);
6063 multilist_destroy(arc_mfu->arcs_list[ARC_BUFC_DATA]);
6064 multilist_destroy(arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
6065 }
6066
6067 uint64_t
6068 arc_max_bytes(void)
6069 {
6070 return (arc_c_max);
6071 }
6072
6073 void
6074 arc_init(void)
6075 {
6076 /*
6077 * allmem is "all memory that we could possibly use".
6078 */
6079 #ifdef _KERNEL
6080 uint64_t allmem = ptob(physmem - swapfs_minfree);
6081 #else
6082 uint64_t allmem = (physmem * PAGESIZE) / 2;
6083 #endif
6084
6104 * small, because it can cause transactions to be larger than
6105 * arc_c, causing arc_tempreserve_space() to fail.
6106 */
6107 #ifndef _KERNEL
6108 arc_c_min = arc_c_max / 2;
6109 #endif
6110
6111 /*
6112 * Allow the tunables to override our calculations if they are
6113 * reasonable (ie. over 64MB)
6114 */
6115 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6116 arc_c_max = zfs_arc_max;
6117 arc_c_min = MIN(arc_c_min, arc_c_max);
6118 }
6119 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6120 arc_c_min = zfs_arc_min;
6121
6122 arc_c = arc_c_max;
6123 arc_p = (arc_c >> 1);
6124
6125 /* limit meta-data to 1/4 of the arc capacity */
6126 arc_meta_limit = arc_c_max / 4;
6127
6128 #ifdef _KERNEL
6129 /*
6130 * Metadata is stored in the kernel's heap. Don't let us
6131 * use more than half the heap for the ARC.
6132 */
6133 arc_meta_limit = MIN(arc_meta_limit,
6134 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6135 #endif
6136
6137 /* Allow the tunable to override if it is reasonable */
6138 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6139 arc_meta_limit = zfs_arc_meta_limit;
6140
6141 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6142 arc_c_min = arc_meta_limit / 2;
6143
6144 if (zfs_arc_meta_min > 0) {
6145 arc_meta_min = zfs_arc_meta_min;
6146 } else {
6147 arc_meta_min = arc_c_min / 2;
6148 }
6149
6150 if (zfs_arc_grow_retry > 0)
6151 arc_grow_retry = zfs_arc_grow_retry;
6152
6153 if (zfs_arc_shrink_shift > 0)
6154 arc_shrink_shift = zfs_arc_shrink_shift;
6155
6156 /*
6157 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6212 /*
6213 * The reclaim thread will set arc_reclaim_thread_exit back to
6214 * B_FALSE when it is finished exiting; we're waiting for that.
6215 */
6216 while (arc_reclaim_thread_exit) {
6217 cv_signal(&arc_reclaim_thread_cv);
6218 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6219 }
6220 mutex_exit(&arc_reclaim_lock);
6221
6222 /* Use B_TRUE to ensure *all* buffers are evicted */
6223 arc_flush(NULL, B_TRUE);
6224
6225 arc_dead = B_TRUE;
6226
6227 if (arc_ksp != NULL) {
6228 kstat_delete(arc_ksp);
6229 arc_ksp = NULL;
6230 }
6231
6232 mutex_destroy(&arc_reclaim_lock);
6233 cv_destroy(&arc_reclaim_thread_cv);
6234 cv_destroy(&arc_reclaim_waiters_cv);
6235
6236 arc_state_fini();
6237 buf_fini();
6238
6239 ASSERT0(arc_loaned_bytes);
6240 }
6241
6242 /*
6243 * Level 2 ARC
6244 *
6245 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
6246 * It uses dedicated storage devices to hold cached data, which are populated
6247 * using large infrequent writes. The main role of this cache is to boost
6248 * the performance of random read workloads. The intended L2ARC devices
6249 * include short-stroked disks, solid state disks, and other media with
6250 * substantially faster read latency than disk.
6251 *
6365 * l2arc_noprefetch skip caching prefetched buffers
6366 * l2arc_headroom number of max device writes to precache
6367 * l2arc_headroom_boost when we find compressed buffers during ARC
6368 * scanning, we multiply headroom by this
6369 * percentage factor for the next scan cycle,
6370 * since more compressed buffers are likely to
6371 * be present
6372 * l2arc_feed_secs seconds between L2ARC writing
6373 *
6374 * Tunables may be removed or added as future performance improvements are
6375 * integrated, and also may become zpool properties.
6376 *
6377 * There are three key functions that control how the L2ARC warms up:
6378 *
6379 * l2arc_write_eligible() check if a buffer is eligible to cache
6380 * l2arc_write_size() calculate how much to write
6381 * l2arc_write_interval() calculate sleep delay between writes
6382 *
6383 * These three functions determine what to write, how much, and how quickly
6384 * to send writes.
6385 */
6386
6387 static boolean_t
6388 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
6389 {
6390 /*
6391 * A buffer is *not* eligible for the L2ARC if it:
6392 * 1. belongs to a different spa.
6393 * 2. is already cached on the L2ARC.
6394 * 3. has an I/O in progress (it may be an incomplete read).
6395 * 4. is flagged not eligible (zfs property).
6396 */
6397 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
6398 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
6399 return (B_FALSE);
6400
6401 return (B_TRUE);
6402 }
6403
6404 static uint64_t
6430 {
6431 clock_t interval, next, now;
6432
6433 /*
6434 * If the ARC lists are busy, increase our write rate; if the
6435 * lists are stale, idle back. This is achieved by checking
6436 * how much we previously wrote - if it was more than half of
6437 * what we wanted, schedule the next write much sooner.
6438 */
6439 if (l2arc_feed_again && wrote > (wanted / 2))
6440 interval = (hz * l2arc_feed_min_ms) / 1000;
6441 else
6442 interval = hz * l2arc_feed_secs;
6443
6444 now = ddi_get_lbolt();
6445 next = MAX(now, MIN(now + interval, began + interval));
6446
6447 return (next);
6448 }
6449
6450 /*
6451 * Cycle through L2ARC devices. This is how L2ARC load balances.
6452 * If a device is returned, this also returns holding the spa config lock.
6453 */
6454 static l2arc_dev_t *
6455 l2arc_dev_get_next(void)
6456 {
6457 l2arc_dev_t *first, *next = NULL;
6458
6459 /*
6460 * Lock out the removal of spas (spa_namespace_lock), then removal
6461 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
6462 * both locks will be dropped and a spa config lock held instead.
6463 */
6464 mutex_enter(&spa_namespace_lock);
6465 mutex_enter(&l2arc_dev_mtx);
6466
6467 /* if there are no vdevs, there is nothing to do */
6468 if (l2arc_ndev == 0)
6469 goto out;
6470
6471 first = NULL;
6472 next = l2arc_dev_last;
6473 do {
6474 /* loop around the list looking for a non-faulted vdev */
6475 if (next == NULL) {
6476 next = list_head(l2arc_dev_list);
6477 } else {
6478 next = list_next(l2arc_dev_list, next);
6479 if (next == NULL)
6480 next = list_head(l2arc_dev_list);
6481 }
6482
6483 /* if we have come back to the start, bail out */
6484 if (first == NULL)
6485 first = next;
6486 else if (next == first)
6487 break;
6488
6489 } while (vdev_is_dead(next->l2ad_vdev));
6490
6491 /* if we were unable to find any usable vdevs, return NULL */
6492 if (vdev_is_dead(next->l2ad_vdev))
6493 next = NULL;
6494
6495 l2arc_dev_last = next;
6496
6497 out:
6498 mutex_exit(&l2arc_dev_mtx);
6499
6500 /*
6501 * Grab the config lock to prevent the 'next' device from being
6502 * removed while we are writing to it.
6503 */
6504 if (next != NULL)
6505 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
6506 mutex_exit(&spa_namespace_lock);
6507
6508 return (next);
6509 }
6510
6511 /*
6512 * Free buffers that were tagged for destruction.
6513 */
6514 static void
6527 list_remove(buflist, df);
6528 kmem_free(df, sizeof (l2arc_data_free_t));
6529 }
6530
6531 mutex_exit(&l2arc_free_on_write_mtx);
6532 }
6533
6534 /*
6535 * A write to a cache device has completed. Update all headers to allow
6536 * reads from these buffers to begin.
6537 */
6538 static void
6539 l2arc_write_done(zio_t *zio)
6540 {
6541 l2arc_write_callback_t *cb;
6542 l2arc_dev_t *dev;
6543 list_t *buflist;
6544 arc_buf_hdr_t *head, *hdr, *hdr_prev;
6545 kmutex_t *hash_lock;
6546 int64_t bytes_dropped = 0;
6547
6548 cb = zio->io_private;
6549 ASSERT3P(cb, !=, NULL);
6550 dev = cb->l2wcb_dev;
6551 ASSERT3P(dev, !=, NULL);
6552 head = cb->l2wcb_head;
6553 ASSERT3P(head, !=, NULL);
6554 buflist = &dev->l2ad_buflist;
6555 ASSERT3P(buflist, !=, NULL);
6556 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
6557 l2arc_write_callback_t *, cb);
6558
6559 if (zio->io_error != 0)
6560 ARCSTAT_BUMP(arcstat_l2_writes_error);
6561
6562 /*
6563 * All writes completed, or an error was hit.
6564 */
6565 top:
6566 mutex_enter(&dev->l2ad_mtx);
6623 bytes_dropped += arc_hdr_size(hdr);
6624 (void) refcount_remove_many(&dev->l2ad_alloc,
6625 arc_hdr_size(hdr), hdr);
6626 }
6627
6628 /*
6629 * Allow ARC to begin reads and ghost list evictions to
6630 * this L2ARC entry.
6631 */
6632 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
6633
6634 mutex_exit(hash_lock);
6635 }
6636
6637 atomic_inc_64(&l2arc_writes_done);
6638 list_remove(buflist, head);
6639 ASSERT(!HDR_HAS_L1HDR(head));
6640 kmem_cache_free(hdr_l2only_cache, head);
6641 mutex_exit(&dev->l2ad_mtx);
6642
6643 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
6644
6645 l2arc_do_free_on_write();
6646
6647 kmem_free(cb, sizeof (l2arc_write_callback_t));
6648 }
6649
6650 /*
6651 * A read to a cache device completed. Validate buffer contents before
6652 * handing over to the regular ARC routines.
6653 */
6654 static void
6655 l2arc_read_done(zio_t *zio)
6656 {
6657 l2arc_read_callback_t *cb;
6658 arc_buf_hdr_t *hdr;
6659 kmutex_t *hash_lock;
6660 boolean_t valid_cksum;
6661
6662 ASSERT3P(zio->io_vd, !=, NULL);
6663 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
6664
6665 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
6666
6733 * storage now. If there *is* a waiter, the caller must
6734 * issue the i/o in a context where it's OK to block.
6735 */
6736 if (zio->io_waiter == NULL) {
6737 zio_t *pio = zio_unique_parent(zio);
6738
6739 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
6740
6741 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
6742 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
6743 hdr, zio->io_priority, cb->l2rcb_flags,
6744 &cb->l2rcb_zb));
6745 }
6746 }
6747
6748 kmem_free(cb, sizeof (l2arc_read_callback_t));
6749 }
6750
6751 /*
6752 * This is the list priority from which the L2ARC will search for pages to
6753 * cache. This is used within loops (0..3) to cycle through lists in the
6754 * desired order. This order can have a significant effect on cache
6755 * performance.
6756 *
6757 * Currently the metadata lists are hit first, MFU then MRU, followed by
6758 * the data lists. This function returns a locked list, and also returns
6759 * the lock pointer.
6760 */
6761 static multilist_sublist_t *
6762 l2arc_sublist_lock(int list_num)
6763 {
6764 multilist_t *ml = NULL;
6765 unsigned int idx;
6766
6767 ASSERT(list_num >= 0 && list_num <= 3);
6768
6769 switch (list_num) {
6770 case 0:
6771 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
6772 break;
6773 case 1:
6774 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
6775 break;
6776 case 2:
6777 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
6778 break;
6779 case 3:
6780 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
6781 break;
6782 }
6783
6784 /*
6785 * Return a randomly-selected sublist. This is acceptable
6786 * because the caller feeds only a little bit of data for each
6787 * call (8MB). Subsequent calls will result in different
6788 * sublists being selected.
6789 */
6790 idx = multilist_get_random_index(ml);
6791 return (multilist_sublist_lock(ml, idx));
6792 }
6793
6794 /*
6795 * Evict buffers from the device write hand to the distance specified in
6796 * bytes. This distance may span populated buffers, it may span nothing.
6797 * This is clearing a region on the L2ARC device ready for writing.
6798 * If the 'all' boolean is set, every buffer is evicted.
6799 */
6800 static void
6801 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
6802 {
6803 list_t *buflist;
6804 arc_buf_hdr_t *hdr, *hdr_prev;
6805 kmutex_t *hash_lock;
6806 uint64_t taddr;
6807
6808 buflist = &dev->l2ad_buflist;
6809
6810 if (!all && dev->l2ad_first) {
6811 /*
6812 * This is the first sweep through the device. There is
6813 * nothing to evict.
6814 */
6815 return;
6816 }
6817
6818 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
6819 /*
6820 * When nearing the end of the device, evict to the end
6821 * before the device write hand jumps to the start.
6822 */
6823 taddr = dev->l2ad_end;
6824 } else {
6825 taddr = dev->l2ad_hand + distance;
6826 }
6827 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
6828 uint64_t, taddr, boolean_t, all);
6829
6830 top:
6831 mutex_enter(&dev->l2ad_mtx);
6832 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
6833 hdr_prev = list_prev(buflist, hdr);
6834
6835 hash_lock = HDR_LOCK(hdr);
6836
6837 /*
6881 } else {
6882 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
6883 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
6884 /*
6885 * Invalidate issued or about to be issued
6886 * reads, since we may be about to write
6887 * over this location.
6888 */
6889 if (HDR_L2_READING(hdr)) {
6890 ARCSTAT_BUMP(arcstat_l2_evict_reading);
6891 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
6892 }
6893
6894 arc_hdr_l2hdr_destroy(hdr);
6895 }
6896 mutex_exit(hash_lock);
6897 }
6898 mutex_exit(&dev->l2ad_mtx);
6899 }
6900
6901 /*
6902 * Find and write ARC buffers to the L2ARC device.
6903 *
6904 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
6905 * for reading until they have completed writing.
6906 * The headroom_boost is an in-out parameter used to maintain headroom boost
6907 * state between calls to this function.
6908 *
6909 * Returns the number of bytes actually written (which may be smaller than
6910 * the delta by which the device hand has changed due to alignment).
6911 */
6912 static uint64_t
6913 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz)
6914 {
6915 arc_buf_hdr_t *hdr, *hdr_prev, *head;
6916 uint64_t write_asize, write_psize, write_lsize, headroom;
6917 boolean_t full;
6918 l2arc_write_callback_t *cb;
6919 zio_t *pio, *wzio;
6920 uint64_t guid = spa_load_guid(spa);
6921
6922 ASSERT3P(dev->l2ad_vdev, !=, NULL);
6923
6924 pio = NULL;
6925 write_lsize = write_asize = write_psize = 0;
6926 full = B_FALSE;
6927 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
6928 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
6929
6930 /*
6931 * Copy buffers for L2ARC writing.
6932 */
6933 for (int try = 0; try <= 3; try++) {
6934 multilist_sublist_t *mls = l2arc_sublist_lock(try);
6935 uint64_t passed_sz = 0;
6936
6937 /*
6938 * L2ARC fast warmup.
6939 *
6940 * Until the ARC is warm and starts to evict, read from the
6941 * head of the ARC lists rather than the tail.
6942 */
6943 if (arc_warm == B_FALSE)
6944 hdr = multilist_sublist_head(mls);
6945 else
6946 hdr = multilist_sublist_tail(mls);
6947
6948 headroom = target_sz * l2arc_headroom;
6949 if (zfs_compressed_arc_enabled)
6950 headroom = (headroom * l2arc_headroom_boost) / 100;
6951
6952 for (; hdr; hdr = hdr_prev) {
6953 kmutex_t *hash_lock;
6983 * We rely on the L1 portion of the header below, so
6984 * it's invalid for this header to have been evicted out
6985 * of the ghost cache, prior to being written out. The
6986 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
6987 */
6988 ASSERT(HDR_HAS_L1HDR(hdr));
6989
6990 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
6991 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
6992 ASSERT3U(arc_hdr_size(hdr), >, 0);
6993 uint64_t psize = arc_hdr_size(hdr);
6994 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
6995 psize);
6996
6997 if ((write_asize + asize) > target_sz) {
6998 full = B_TRUE;
6999 mutex_exit(hash_lock);
7000 break;
7001 }
7002
7003 if (pio == NULL) {
7004 /*
7005 * Insert a dummy header on the buflist so
7006 * l2arc_write_done() can find where the
7007 * write buffers begin without searching.
7008 */
7009 mutex_enter(&dev->l2ad_mtx);
7010 list_insert_head(&dev->l2ad_buflist, head);
7011 mutex_exit(&dev->l2ad_mtx);
7012
7013 cb = kmem_alloc(
7014 sizeof (l2arc_write_callback_t), KM_SLEEP);
7015 cb->l2wcb_dev = dev;
7016 cb->l2wcb_head = head;
7017 pio = zio_root(spa, l2arc_write_done, cb,
7018 ZIO_FLAG_CANFAIL);
7019 }
7020
7021 hdr->b_l2hdr.b_dev = dev;
7022 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
7023 arc_hdr_set_flags(hdr,
7024 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
7025
7026 mutex_enter(&dev->l2ad_mtx);
7027 list_insert_head(&dev->l2ad_buflist, hdr);
7028 mutex_exit(&dev->l2ad_mtx);
7029
7030 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
7031
7032 /*
7033 * Normally the L2ARC can use the hdr's data, but if
7034 * we're sharing data between the hdr and one of its
7035 * bufs, L2ARC needs its own copy of the data so that
7036 * the ZIO below can't race with the buf consumer.
7037 * Another case where we need to create a copy of the
7038 * data is when the buffer size is not device-aligned
7039 * and we need to pad the block to make it such.
7040 * That also keeps the clock hand suitably aligned.
7041 *
7042 * To ensure that the copy will be available for the
7043 * lifetime of the ZIO and be cleaned up afterwards, we
7044 * add it to the l2arc_free_on_write queue.
7045 */
7046 abd_t *to_write;
7047 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
7048 to_write = hdr->b_l1hdr.b_pabd;
7049 } else {
7050 to_write = abd_alloc_for_io(asize,
7051 HDR_ISTYPE_METADATA(hdr));
7052 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
7053 if (asize != psize) {
7054 abd_zero_off(to_write, psize,
7055 asize - psize);
7056 }
7057 l2arc_free_abd_on_write(to_write, asize,
7058 arc_buf_type(hdr));
7059 }
7060 wzio = zio_write_phys(pio, dev->l2ad_vdev,
7061 hdr->b_l2hdr.b_daddr, asize, to_write,
7062 ZIO_CHECKSUM_OFF, NULL, hdr,
7063 ZIO_PRIORITY_ASYNC_WRITE,
7064 ZIO_FLAG_CANFAIL, B_FALSE);
7065
7066 write_lsize += HDR_GET_LSIZE(hdr);
7067 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
7068 zio_t *, wzio);
7069
7070 write_psize += psize;
7071 write_asize += asize;
7072 dev->l2ad_hand += asize;
7073
7074 mutex_exit(hash_lock);
7075
7076 (void) zio_nowait(wzio);
7077 }
7078
7079 multilist_sublist_unlock(mls);
7080
7081 if (full == B_TRUE)
7082 break;
7083 }
7084
7085 /* No buffers selected for writing? */
7086 if (pio == NULL) {
7087 ASSERT0(write_lsize);
7088 ASSERT(!HDR_HAS_L1HDR(head));
7089 kmem_cache_free(hdr_l2only_cache, head);
7090 return (0);
7091 }
7092
7093 ASSERT3U(write_asize, <=, target_sz);
7094 ARCSTAT_BUMP(arcstat_l2_writes_sent);
7095 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
7096 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
7097 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
7098 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
7099
7100 /*
7101 * Bump device hand to the device start if it is approaching the end.
7102 * l2arc_evict() will already have evicted ahead for this case.
7103 */
7104 if (dev->l2ad_hand >= (dev->l2ad_end - target_sz)) {
7105 dev->l2ad_hand = dev->l2ad_start;
7106 dev->l2ad_first = B_FALSE;
7107 }
7108
7109 dev->l2ad_writing = B_TRUE;
7110 (void) zio_wait(pio);
7111 dev->l2ad_writing = B_FALSE;
7112
7113 return (write_asize);
7114 }
7115
7116 /*
7117 * This thread feeds the L2ARC at regular intervals. This is the beating
7118 * heart of the L2ARC.
7119 */
7120 /* ARGSUSED */
7121 static void
7122 l2arc_feed_thread(void *unused)
7123 {
7124 callb_cpr_t cpr;
7125 l2arc_dev_t *dev;
7126 spa_t *spa;
7127 uint64_t size, wrote;
7128 clock_t begin, next = ddi_get_lbolt();
7129
7130 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
7131
7132 mutex_enter(&l2arc_feed_thr_lock);
7133
7134 while (l2arc_thread_exit == 0) {
7135 CALLB_CPR_SAFE_BEGIN(&cpr);
7136 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
7137 next);
7138 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
7139 next = ddi_get_lbolt() + hz;
7140
7141 /*
7142 * Quick check for L2ARC devices.
7143 */
7144 mutex_enter(&l2arc_dev_mtx);
7145 if (l2arc_ndev == 0) {
7146 mutex_exit(&l2arc_dev_mtx);
7147 continue;
7148 }
7149 mutex_exit(&l2arc_dev_mtx);
7150 begin = ddi_get_lbolt();
7151
7152 /*
7153 * This selects the next l2arc device to write to, and in
7154 * doing so the next spa to feed from: dev->l2ad_spa. This
7155 * will return NULL if there are now no l2arc devices or if
7156 * they are all faulted.
7157 *
7158 * If a device is returned, its spa's config lock is also
7159 * held to prevent device removal. l2arc_dev_get_next()
7160 * will grab and release l2arc_dev_mtx.
7161 */
7162 if ((dev = l2arc_dev_get_next()) == NULL)
7163 continue;
7164
7165 spa = dev->l2ad_spa;
7166 ASSERT3P(spa, !=, NULL);
7167
7168 /*
7169 * If the pool is read-only then force the feed thread to
7170 * sleep a little longer.
7171 */
7172 if (!spa_writeable(spa)) {
7173 next = ddi_get_lbolt() + 5 * l2arc_feed_secs * hz;
7174 spa_config_exit(spa, SCL_L2ARC, dev);
7175 continue;
7176 }
7177
7178 /*
7179 * Avoid contributing to memory pressure.
7180 */
7181 if (arc_reclaim_needed()) {
7182 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
7183 spa_config_exit(spa, SCL_L2ARC, dev);
7184 continue;
7185 }
7186
7187 ARCSTAT_BUMP(arcstat_l2_feeds);
7188
7189 size = l2arc_write_size();
7190
7191 /*
7192 * Evict L2ARC buffers that will be overwritten.
7193 */
7194 l2arc_evict(dev, size, B_FALSE);
7195
7196 /*
7197 * Write ARC buffers.
7198 */
7199 wrote = l2arc_write_buffers(spa, dev, size);
7200
7201 /*
7202 * Calculate interval between writes.
7203 */
7204 next = l2arc_write_interval(begin, size, wrote);
7205 spa_config_exit(spa, SCL_L2ARC, dev);
7206 }
7207
7208 l2arc_thread_exit = 0;
7209 cv_broadcast(&l2arc_feed_thr_cv);
7210 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
7211 thread_exit();
7212 }
7213
7214 boolean_t
7215 l2arc_vdev_present(vdev_t *vd)
7216 {
7217 l2arc_dev_t *dev;
7218
7219 mutex_enter(&l2arc_dev_mtx);
7220 for (dev = list_head(l2arc_dev_list); dev != NULL;
7221 dev = list_next(l2arc_dev_list, dev)) {
7222 if (dev->l2ad_vdev == vd)
7223 break;
7224 }
7225 mutex_exit(&l2arc_dev_mtx);
7226
7227 return (dev != NULL);
7228 }
7229
7230 /*
7231 * Add a vdev for use by the L2ARC. By this point the spa has already
7232 * validated the vdev and opened it.
7233 */
7234 void
7235 l2arc_add_vdev(spa_t *spa, vdev_t *vd)
7236 {
7237 l2arc_dev_t *adddev;
7238
7239 ASSERT(!l2arc_vdev_present(vd));
7240
7241 /*
7242 * Create a new l2arc device entry.
7243 */
7244 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
7245 adddev->l2ad_spa = spa;
7246 adddev->l2ad_vdev = vd;
7247 adddev->l2ad_start = VDEV_LABEL_START_SIZE;
7248 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
7249 adddev->l2ad_hand = adddev->l2ad_start;
7250 adddev->l2ad_first = B_TRUE;
7251 adddev->l2ad_writing = B_FALSE;
7252
7253 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
7254 /*
7255 * This is a list of all ARC buffers that are still valid on the
7256 * device.
7257 */
7258 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
7259 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
7260
7261 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
7262 refcount_create(&adddev->l2ad_alloc);
7263
7264 /*
7265 * Add device to global list
7266 */
7267 mutex_enter(&l2arc_dev_mtx);
7268 list_insert_head(l2arc_dev_list, adddev);
7269 atomic_inc_64(&l2arc_ndev);
7270 mutex_exit(&l2arc_dev_mtx);
7271 }
7272
7273 /*
7274 * Remove a vdev from the L2ARC.
7275 */
7276 void
7277 l2arc_remove_vdev(vdev_t *vd)
7278 {
7279 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
7280
7281 /*
7282 * Find the device by vdev
7283 */
7284 mutex_enter(&l2arc_dev_mtx);
7285 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
7286 nextdev = list_next(l2arc_dev_list, dev);
7287 if (vd == dev->l2ad_vdev) {
7288 remdev = dev;
7289 break;
7290 }
7291 }
7292 ASSERT3P(remdev, !=, NULL);
7293
7294 /*
7295 * Remove device from global list
7296 */
7297 list_remove(l2arc_dev_list, remdev);
7298 l2arc_dev_last = NULL; /* may have been invalidated */
7299 atomic_dec_64(&l2arc_ndev);
7300 mutex_exit(&l2arc_dev_mtx);
7301
7302 /*
7303 * Clear all buflists and ARC references. L2ARC device flush.
7304 */
7305 l2arc_evict(remdev, 0, B_TRUE);
7306 list_destroy(&remdev->l2ad_buflist);
7307 mutex_destroy(&remdev->l2ad_mtx);
7308 refcount_destroy(&remdev->l2ad_alloc);
7309 kmem_free(remdev, sizeof (l2arc_dev_t));
7310 }
7311
7312 void
7313 l2arc_init(void)
7314 {
7315 l2arc_thread_exit = 0;
7316 l2arc_ndev = 0;
7317 l2arc_writes_sent = 0;
7318 l2arc_writes_done = 0;
7319
7320 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
7321 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
7322 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
7323 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
7324
7325 l2arc_dev_list = &L2ARC_dev_list;
7326 l2arc_free_on_write = &L2ARC_free_on_write;
7327 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
7328 offsetof(l2arc_dev_t, l2ad_node));
7329 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
7355 {
7356 if (!(spa_mode_global & FWRITE))
7357 return;
7358
7359 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
7360 TS_RUN, minclsyspri);
7361 }
7362
7363 void
7364 l2arc_stop(void)
7365 {
7366 if (!(spa_mode_global & FWRITE))
7367 return;
7368
7369 mutex_enter(&l2arc_feed_thr_lock);
7370 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
7371 l2arc_thread_exit = 1;
7372 while (l2arc_thread_exit != 0)
7373 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
7374 mutex_exit(&l2arc_feed_thr_lock);
7375 }
|
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2018, Joyent, Inc.
24 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26 * Copyright 2019 Nexenta Systems, Inc. All rights reserved.
27 */
28
29 /*
30 * DVA-based Adjustable Replacement Cache
31 *
32 * While much of the theory of operation used here is
33 * based on the self-tuning, low overhead replacement cache
34 * presented by Megiddo and Modha at FAST 2003, there are some
35 * significant differences:
36 *
37 * 1. The Megiddo and Modha model assumes any page is evictable.
38 * Pages in its cache cannot be "locked" into memory. This makes
39 * the eviction algorithm simple: evict the last page in the list.
40 * This also make the performance characteristics easy to reason
41 * about. Our cache is not so simple. At any given moment, some
42 * subset of the blocks in the cache are un-evictable because we
43 * have handed out a reference to them. Blocks are only evictable
44 * when there are no external references active. This makes
45 * eviction far more problematic: we choose to evict the evictable
46 * blocks that are the "lowest" in the list.
236 * it may compress the data before writing it to disk. The ARC will be called
237 * with the transformed data and will bcopy the transformed on-disk block into
238 * a newly allocated b_pabd. Writes are always done into buffers which have
239 * either been loaned (and hence are new and don't have other readers) or
240 * buffers which have been released (and hence have their own hdr, if there
241 * were originally other readers of the buf's original hdr). This ensures that
242 * the ARC only needs to update a single buf and its hdr after a write occurs.
243 *
244 * When the L2ARC is in use, it will also take advantage of the b_pabd. The
245 * L2ARC will always write the contents of b_pabd to the L2ARC. This means
246 * that when compressed ARC is enabled that the L2ARC blocks are identical
247 * to the on-disk block in the main data pool. This provides a significant
248 * advantage since the ARC can leverage the bp's checksum when reading from the
249 * L2ARC to determine if the contents are valid. However, if the compressed
250 * ARC is disabled, then the L2ARC's block must be transformed to look
251 * like the physical block in the main data pool before comparing the
252 * checksum and determining its validity.
253 */
254
255 #include <sys/spa.h>
256 #include <sys/spa_impl.h>
257 #include <sys/zio.h>
258 #include <sys/spa_impl.h>
259 #include <sys/zio_compress.h>
260 #include <sys/zio_checksum.h>
261 #include <sys/zfs_context.h>
262 #include <sys/arc.h>
263 #include <sys/refcount.h>
264 #include <sys/vdev.h>
265 #include <sys/vdev_impl.h>
266 #include <sys/dsl_pool.h>
267 #include <sys/zio_checksum.h>
268 #include <sys/multilist.h>
269 #include <sys/abd.h>
270 #ifdef _KERNEL
271 #include <sys/vmsystm.h>
272 #include <vm/anon.h>
273 #include <sys/fs/swapnode.h>
274 #include <sys/dnlc.h>
275 #endif
276 #include <sys/callb.h>
277 #include <sys/kstat.h>
278 #include <zfs_fletcher.h>
279 #include <sys/byteorder.h>
280 #include <sys/spa_impl.h>
281
282 #ifndef _KERNEL
283 /* set with ZFS_DEBUG=watch, to enable watchpoints on frozen buffers */
284 boolean_t arc_watch = B_FALSE;
285 int arc_procfd;
286 #endif
287
288 static kmutex_t arc_reclaim_lock;
289 static kcondvar_t arc_reclaim_thread_cv;
290 static boolean_t arc_reclaim_thread_exit;
291 static kcondvar_t arc_reclaim_waiters_cv;
292
293 uint_t arc_reduce_dnlc_percent = 3;
294
295 /*
296 * The number of headers to evict in arc_evict_state_impl() before
297 * dropping the sublist lock and evicting from another sublist. A lower
298 * value means we're more likely to evict the "correct" header (i.e. the
299 * oldest header in the arc state), but comes with higher overhead
300 * (i.e. more invocations of arc_evict_state_impl()).
341
342 static int arc_dead;
343
344 /*
345 * The arc has filled available memory and has now warmed up.
346 */
347 static boolean_t arc_warm;
348
349 /*
350 * log2 fraction of the zio arena to keep free.
351 */
352 int arc_zio_arena_free_shift = 2;
353
354 /*
355 * These tunables are for performance analysis.
356 */
357 uint64_t zfs_arc_max;
358 uint64_t zfs_arc_min;
359 uint64_t zfs_arc_meta_limit = 0;
360 uint64_t zfs_arc_meta_min = 0;
361 uint64_t zfs_arc_ddt_limit = 0;
362 /*
363 * Tunable to control "dedup ceiling"
364 * Possible values:
365 * DDT_NO_LIMIT - default behaviour, ie no ceiling
366 * DDT_LIMIT_TO_ARC - stop DDT growth if DDT is bigger than it's "ARC space"
367 * DDT_LIMIT_TO_L2ARC - stop DDT growth when DDT size is bigger than the
368 * L2ARC DDT dev(s) for that pool
369 */
370 zfs_ddt_limit_t zfs_ddt_limit_type = DDT_LIMIT_TO_ARC;
371 /*
372 * Alternative to the above way of controlling "dedup ceiling":
373 * Stop DDT growth when in core DDTs size is above the below tunable.
374 * This tunable overrides the zfs_ddt_limit_type tunable.
375 */
376 uint64_t zfs_ddt_byte_ceiling = 0;
377 boolean_t zfs_arc_segregate_ddt = B_TRUE;
378 int zfs_arc_grow_retry = 0;
379 int zfs_arc_shrink_shift = 0;
380 int zfs_arc_p_min_shift = 0;
381 int zfs_arc_average_blocksize = 8 * 1024; /* 8KB */
382
383 /* Tuneable, default is 64, which is essentially arbitrary */
384 int zfs_flush_ntasks = 64;
385
386 boolean_t zfs_compressed_arc_enabled = B_TRUE;
387
388 /*
389 * Note that buffers can be in one of 6 states:
390 * ARC_anon - anonymous (discussed below)
391 * ARC_mru - recently used, currently cached
392 * ARC_mru_ghost - recentely used, no longer in cache
393 * ARC_mfu - frequently used, currently cached
394 * ARC_mfu_ghost - frequently used, no longer in cache
395 * ARC_l2c_only - exists in L2ARC but not other states
396 * When there are no active references to the buffer, they are
397 * are linked onto a list in one of these arc states. These are
398 * the only buffers that can be evicted or deleted. Within each
399 * state there are multiple lists, one for meta-data and one for
400 * non-meta-data. Meta-data (indirect blocks, blocks of dnodes,
401 * etc.) is tracked separately so that it can be managed more
402 * explicitly: favored over data, limited explicitly.
403 *
404 * Anonymous buffers are buffers that are not associated with
405 * a DVA. These are buffers that hold dirty block copies
411 * The ARC_l2c_only state is for buffers that are in the second
412 * level ARC but no longer in any of the ARC_m* lists. The second
413 * level ARC itself may also contain buffers that are in any of
414 * the ARC_m* states - meaning that a buffer can exist in two
415 * places. The reason for the ARC_l2c_only state is to keep the
416 * buffer header in the hash table, so that reads that hit the
417 * second level ARC benefit from these fast lookups.
418 */
419
420 typedef struct arc_state {
421 /*
422 * list of evictable buffers
423 */
424 multilist_t *arcs_list[ARC_BUFC_NUMTYPES];
425 /*
426 * total amount of evictable data in this state
427 */
428 refcount_t arcs_esize[ARC_BUFC_NUMTYPES];
429 /*
430 * total amount of data in this state; this includes: evictable,
431 * non-evictable, ARC_BUFC_DATA, ARC_BUFC_METADATA and ARC_BUFC_DDT.
432 * ARC_BUFC_DDT list is only populated when zfs_arc_segregate_ddt is
433 * true.
434 */
435 refcount_t arcs_size;
436 } arc_state_t;
437
438 /*
439 * We loop through these in l2arc_write_buffers() starting from
440 * PRIORITY_MFU_DDT until we reach PRIORITY_NUMTYPES or the buffer that we
441 * will be writing to L2ARC dev gets full.
442 */
443 enum l2arc_priorities {
444 PRIORITY_MFU_DDT,
445 PRIORITY_MRU_DDT,
446 PRIORITY_MFU_META,
447 PRIORITY_MRU_META,
448 PRIORITY_MFU_DATA,
449 PRIORITY_MRU_DATA,
450 PRIORITY_NUMTYPES,
451 };
452
453 /* The 6 states: */
454 static arc_state_t ARC_anon;
455 static arc_state_t ARC_mru;
456 static arc_state_t ARC_mru_ghost;
457 static arc_state_t ARC_mfu;
458 static arc_state_t ARC_mfu_ghost;
459 static arc_state_t ARC_l2c_only;
460
461 typedef struct arc_stats {
462 kstat_named_t arcstat_hits;
463 kstat_named_t arcstat_ddt_hits;
464 kstat_named_t arcstat_misses;
465 kstat_named_t arcstat_demand_data_hits;
466 kstat_named_t arcstat_demand_data_misses;
467 kstat_named_t arcstat_demand_metadata_hits;
468 kstat_named_t arcstat_demand_metadata_misses;
469 kstat_named_t arcstat_demand_ddt_hits;
470 kstat_named_t arcstat_demand_ddt_misses;
471 kstat_named_t arcstat_prefetch_data_hits;
472 kstat_named_t arcstat_prefetch_data_misses;
473 kstat_named_t arcstat_prefetch_metadata_hits;
474 kstat_named_t arcstat_prefetch_metadata_misses;
475 kstat_named_t arcstat_prefetch_ddt_hits;
476 kstat_named_t arcstat_prefetch_ddt_misses;
477 kstat_named_t arcstat_mru_hits;
478 kstat_named_t arcstat_mru_ghost_hits;
479 kstat_named_t arcstat_mfu_hits;
480 kstat_named_t arcstat_mfu_ghost_hits;
481 kstat_named_t arcstat_deleted;
482 /*
483 * Number of buffers that could not be evicted because the hash lock
484 * was held by another thread. The lock may not necessarily be held
485 * by something using the same buffer, since hash locks are shared
486 * by multiple buffers.
487 */
488 kstat_named_t arcstat_mutex_miss;
489 /*
490 * Number of buffers skipped when updating the access state due to the
491 * header having already been released after acquiring the hash lock.
492 */
493 kstat_named_t arcstat_access_skip;
494 /*
495 * Number of buffers skipped because they have I/O in progress, are
496 * indirect prefetch buffers that have not lived long enough, or are
497 * not from the spa we're trying to evict from.
498 */
499 kstat_named_t arcstat_evict_skip;
500 /*
501 * Number of times arc_evict_state() was unable to evict enough
502 * buffers to reach it's target amount.
503 */
504 kstat_named_t arcstat_evict_not_enough;
505 kstat_named_t arcstat_evict_l2_cached;
506 kstat_named_t arcstat_evict_l2_eligible;
507 kstat_named_t arcstat_evict_l2_ineligible;
508 kstat_named_t arcstat_evict_l2_skip;
509 kstat_named_t arcstat_hash_elements;
510 kstat_named_t arcstat_hash_elements_max;
511 kstat_named_t arcstat_hash_collisions;
512 kstat_named_t arcstat_hash_chains;
513 kstat_named_t arcstat_hash_chain_max;
514 kstat_named_t arcstat_p;
515 kstat_named_t arcstat_c;
516 kstat_named_t arcstat_c_min;
517 kstat_named_t arcstat_c_max;
518 kstat_named_t arcstat_size;
519 /*
520 * Number of compressed bytes stored in the arc_buf_hdr_t's b_pabd.
521 * Note that the compressed bytes may match the uncompressed bytes
522 * if the block is either not compressed or compressed arc is disabled.
523 */
524 kstat_named_t arcstat_compressed_size;
525 /*
526 * Uncompressed size of the data stored in b_pabd. If compressed
527 * arc is disabled then this value will be identical to the stat
528 * above.
529 */
530 kstat_named_t arcstat_uncompressed_size;
531 /*
532 * Number of bytes stored in all the arc_buf_t's. This is classified
533 * as "overhead" since this data is typically short-lived and will
534 * be evicted from the arc when it becomes unreferenced unless the
535 * zfs_keep_uncompressed_metadata or zfs_keep_uncompressed_level
536 * values have been set (see comment in dbuf.c for more information).
537 */
538 kstat_named_t arcstat_overhead_size;
539 /*
540 * Number of bytes consumed by internal ARC structures necessary
541 * for tracking purposes; these structures are not actually
542 * backed by ARC buffers. This includes arc_buf_hdr_t structures
543 * (allocated via arc_buf_hdr_t_full and arc_buf_hdr_t_l2only
544 * caches), and arc_buf_t structures (allocated via arc_buf_t
545 * cache).
546 */
547 kstat_named_t arcstat_hdr_size;
548 /*
549 * Number of bytes consumed by ARC buffers of type equal to
550 * ARC_BUFC_DATA. This is generally consumed by buffers backing
551 * on disk user data (e.g. plain file contents).
552 */
553 kstat_named_t arcstat_data_size;
554 /*
555 * Number of bytes consumed by ARC buffers of type equal to
556 * ARC_BUFC_METADATA. This is generally consumed by buffers
557 * backing on disk data that is used for internal ZFS
558 * structures (e.g. ZAP, dnode, indirect blocks, etc).
559 */
560 kstat_named_t arcstat_metadata_size;
561 /*
562 * Number of bytes consumed by ARC buffers of type equal to
563 * ARC_BUFC_DDT. This is consumed by buffers backing on disk data
564 * that is used to store DDT (ZAP, ddt stats).
565 * Only used if zfs_arc_segregate_ddt is true.
566 */
567 kstat_named_t arcstat_ddt_size;
568 /*
569 * Number of bytes consumed by various buffers and structures
570 * not actually backed with ARC buffers. This includes bonus
571 * buffers (allocated directly via zio_buf_* functions),
572 * dmu_buf_impl_t structures (allocated via dmu_buf_impl_t
573 * cache), and dnode_t structures (allocated via dnode_t cache).
574 */
575 kstat_named_t arcstat_other_size;
576 /*
577 * Total number of bytes consumed by ARC buffers residing in the
578 * arc_anon state. This includes *all* buffers in the arc_anon
579 * state; e.g. data, metadata, evictable, and unevictable buffers
580 * are all included in this value.
581 */
582 kstat_named_t arcstat_anon_size;
583 /*
584 * Number of bytes consumed by ARC buffers that meet the
585 * following criteria: backing buffers of type ARC_BUFC_DATA,
586 * residing in the arc_anon state, and are eligible for eviction
587 * (e.g. have no outstanding holds on the buffer).
588 */
589 kstat_named_t arcstat_anon_evictable_data;
590 /*
591 * Number of bytes consumed by ARC buffers that meet the
592 * following criteria: backing buffers of type ARC_BUFC_METADATA,
593 * residing in the arc_anon state, and are eligible for eviction
594 * (e.g. have no outstanding holds on the buffer).
595 */
596 kstat_named_t arcstat_anon_evictable_metadata;
597 /*
598 * Number of bytes consumed by ARC buffers that meet the
599 * following criteria: backing buffers of type ARC_BUFC_DDT,
600 * residing in the arc_anon state, and are eligible for eviction
601 * Only used if zfs_arc_segregate_ddt is true.
602 */
603 kstat_named_t arcstat_anon_evictable_ddt;
604 /*
605 * Total number of bytes consumed by ARC buffers residing in the
606 * arc_mru state. This includes *all* buffers in the arc_mru
607 * state; e.g. data, metadata, evictable, and unevictable buffers
608 * are all included in this value.
609 */
610 kstat_named_t arcstat_mru_size;
611 /*
612 * Number of bytes consumed by ARC buffers that meet the
613 * following criteria: backing buffers of type ARC_BUFC_DATA,
614 * residing in the arc_mru state, and are eligible for eviction
615 * (e.g. have no outstanding holds on the buffer).
616 */
617 kstat_named_t arcstat_mru_evictable_data;
618 /*
619 * Number of bytes consumed by ARC buffers that meet the
620 * following criteria: backing buffers of type ARC_BUFC_METADATA,
621 * residing in the arc_mru state, and are eligible for eviction
622 * (e.g. have no outstanding holds on the buffer).
623 */
624 kstat_named_t arcstat_mru_evictable_metadata;
625 /*
626 * Number of bytes consumed by ARC buffers that meet the
627 * following criteria: backing buffers of type ARC_BUFC_DDT,
628 * residing in the arc_mru state, and are eligible for eviction
629 * (e.g. have no outstanding holds on the buffer).
630 * Only used if zfs_arc_segregate_ddt is true.
631 */
632 kstat_named_t arcstat_mru_evictable_ddt;
633 /*
634 * Total number of bytes that *would have been* consumed by ARC
635 * buffers in the arc_mru_ghost state. The key thing to note
636 * here, is the fact that this size doesn't actually indicate
637 * RAM consumption. The ghost lists only consist of headers and
638 * don't actually have ARC buffers linked off of these headers.
639 * Thus, *if* the headers had associated ARC buffers, these
640 * buffers *would have* consumed this number of bytes.
641 */
642 kstat_named_t arcstat_mru_ghost_size;
643 /*
644 * Number of bytes that *would have been* consumed by ARC
645 * buffers that are eligible for eviction, of type
646 * ARC_BUFC_DATA, and linked off the arc_mru_ghost state.
647 */
648 kstat_named_t arcstat_mru_ghost_evictable_data;
649 /*
650 * Number of bytes that *would have been* consumed by ARC
651 * buffers that are eligible for eviction, of type
652 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
653 */
654 kstat_named_t arcstat_mru_ghost_evictable_metadata;
655 /*
656 * Number of bytes that *would have been* consumed by ARC
657 * buffers that are eligible for eviction, of type
658 * ARC_BUFC_DDT, and linked off the arc_mru_ghost state.
659 * Only used if zfs_arc_segregate_ddt is true.
660 */
661 kstat_named_t arcstat_mru_ghost_evictable_ddt;
662 /*
663 * Total number of bytes consumed by ARC buffers residing in the
664 * arc_mfu state. This includes *all* buffers in the arc_mfu
665 * state; e.g. data, metadata, evictable, and unevictable buffers
666 * are all included in this value.
667 */
668 kstat_named_t arcstat_mfu_size;
669 /*
670 * Number of bytes consumed by ARC buffers that are eligible for
671 * eviction, of type ARC_BUFC_DATA, and reside in the arc_mfu
672 * state.
673 */
674 kstat_named_t arcstat_mfu_evictable_data;
675 /*
676 * Number of bytes consumed by ARC buffers that are eligible for
677 * eviction, of type ARC_BUFC_METADATA, and reside in the
678 * arc_mfu state.
679 */
680 kstat_named_t arcstat_mfu_evictable_metadata;
681 /*
682 * Number of bytes consumed by ARC buffers that are eligible for
683 * eviction, of type ARC_BUFC_DDT, and reside in the
684 * arc_mfu state.
685 * Only used if zfs_arc_segregate_ddt is true.
686 */
687 kstat_named_t arcstat_mfu_evictable_ddt;
688 /*
689 * Total number of bytes that *would have been* consumed by ARC
690 * buffers in the arc_mfu_ghost state. See the comment above
691 * arcstat_mru_ghost_size for more details.
692 */
693 kstat_named_t arcstat_mfu_ghost_size;
694 /*
695 * Number of bytes that *would have been* consumed by ARC
696 * buffers that are eligible for eviction, of type
697 * ARC_BUFC_DATA, and linked off the arc_mfu_ghost state.
698 */
699 kstat_named_t arcstat_mfu_ghost_evictable_data;
700 /*
701 * Number of bytes that *would have been* consumed by ARC
702 * buffers that are eligible for eviction, of type
703 * ARC_BUFC_METADATA, and linked off the arc_mru_ghost state.
704 */
705 kstat_named_t arcstat_mfu_ghost_evictable_metadata;
706 /*
707 * Number of bytes that *would have been* consumed by ARC
708 * buffers that are eligible for eviction, of type
709 * ARC_BUFC_DDT, and linked off the arc_mru_ghost state.
710 * Only used if zfs_arc_segregate_ddt is true.
711 */
712 kstat_named_t arcstat_mfu_ghost_evictable_ddt;
713 kstat_named_t arcstat_l2_hits;
714 kstat_named_t arcstat_l2_ddt_hits;
715 kstat_named_t arcstat_l2_misses;
716 kstat_named_t arcstat_l2_feeds;
717 kstat_named_t arcstat_l2_rw_clash;
718 kstat_named_t arcstat_l2_read_bytes;
719 kstat_named_t arcstat_l2_ddt_read_bytes;
720 kstat_named_t arcstat_l2_write_bytes;
721 kstat_named_t arcstat_l2_ddt_write_bytes;
722 kstat_named_t arcstat_l2_writes_sent;
723 kstat_named_t arcstat_l2_writes_done;
724 kstat_named_t arcstat_l2_writes_error;
725 kstat_named_t arcstat_l2_writes_lock_retry;
726 kstat_named_t arcstat_l2_evict_lock_retry;
727 kstat_named_t arcstat_l2_evict_reading;
728 kstat_named_t arcstat_l2_evict_l1cached;
729 kstat_named_t arcstat_l2_free_on_write;
730 kstat_named_t arcstat_l2_abort_lowmem;
731 kstat_named_t arcstat_l2_cksum_bad;
732 kstat_named_t arcstat_l2_io_error;
733 kstat_named_t arcstat_l2_lsize;
734 kstat_named_t arcstat_l2_psize;
735 kstat_named_t arcstat_l2_hdr_size;
736 kstat_named_t arcstat_l2_log_blk_writes;
737 kstat_named_t arcstat_l2_log_blk_avg_size;
738 kstat_named_t arcstat_l2_data_to_meta_ratio;
739 kstat_named_t arcstat_l2_rebuild_successes;
740 kstat_named_t arcstat_l2_rebuild_abort_unsupported;
741 kstat_named_t arcstat_l2_rebuild_abort_io_errors;
742 kstat_named_t arcstat_l2_rebuild_abort_cksum_errors;
743 kstat_named_t arcstat_l2_rebuild_abort_loop_errors;
744 kstat_named_t arcstat_l2_rebuild_abort_lowmem;
745 kstat_named_t arcstat_l2_rebuild_size;
746 kstat_named_t arcstat_l2_rebuild_bufs;
747 kstat_named_t arcstat_l2_rebuild_bufs_precached;
748 kstat_named_t arcstat_l2_rebuild_psize;
749 kstat_named_t arcstat_l2_rebuild_log_blks;
750 kstat_named_t arcstat_memory_throttle_count;
751 kstat_named_t arcstat_meta_used;
752 kstat_named_t arcstat_meta_limit;
753 kstat_named_t arcstat_meta_max;
754 kstat_named_t arcstat_meta_min;
755 kstat_named_t arcstat_ddt_limit;
756 kstat_named_t arcstat_sync_wait_for_async;
757 kstat_named_t arcstat_demand_hit_predictive_prefetch;
758 } arc_stats_t;
759
760 static arc_stats_t arc_stats = {
761 { "hits", KSTAT_DATA_UINT64 },
762 { "ddt_hits", KSTAT_DATA_UINT64 },
763 { "misses", KSTAT_DATA_UINT64 },
764 { "demand_data_hits", KSTAT_DATA_UINT64 },
765 { "demand_data_misses", KSTAT_DATA_UINT64 },
766 { "demand_metadata_hits", KSTAT_DATA_UINT64 },
767 { "demand_metadata_misses", KSTAT_DATA_UINT64 },
768 { "demand_ddt_hits", KSTAT_DATA_UINT64 },
769 { "demand_ddt_misses", KSTAT_DATA_UINT64 },
770 { "prefetch_data_hits", KSTAT_DATA_UINT64 },
771 { "prefetch_data_misses", KSTAT_DATA_UINT64 },
772 { "prefetch_metadata_hits", KSTAT_DATA_UINT64 },
773 { "prefetch_metadata_misses", KSTAT_DATA_UINT64 },
774 { "prefetch_ddt_hits", KSTAT_DATA_UINT64 },
775 { "prefetch_ddt_misses", KSTAT_DATA_UINT64 },
776 { "mru_hits", KSTAT_DATA_UINT64 },
777 { "mru_ghost_hits", KSTAT_DATA_UINT64 },
778 { "mfu_hits", KSTAT_DATA_UINT64 },
779 { "mfu_ghost_hits", KSTAT_DATA_UINT64 },
780 { "deleted", KSTAT_DATA_UINT64 },
781 { "mutex_miss", KSTAT_DATA_UINT64 },
782 { "access_skip", KSTAT_DATA_UINT64 },
783 { "evict_skip", KSTAT_DATA_UINT64 },
784 { "evict_not_enough", KSTAT_DATA_UINT64 },
785 { "evict_l2_cached", KSTAT_DATA_UINT64 },
786 { "evict_l2_eligible", KSTAT_DATA_UINT64 },
787 { "evict_l2_ineligible", KSTAT_DATA_UINT64 },
788 { "evict_l2_skip", KSTAT_DATA_UINT64 },
789 { "hash_elements", KSTAT_DATA_UINT64 },
790 { "hash_elements_max", KSTAT_DATA_UINT64 },
791 { "hash_collisions", KSTAT_DATA_UINT64 },
792 { "hash_chains", KSTAT_DATA_UINT64 },
793 { "hash_chain_max", KSTAT_DATA_UINT64 },
794 { "p", KSTAT_DATA_UINT64 },
795 { "c", KSTAT_DATA_UINT64 },
796 { "c_min", KSTAT_DATA_UINT64 },
797 { "c_max", KSTAT_DATA_UINT64 },
798 { "size", KSTAT_DATA_UINT64 },
799 { "compressed_size", KSTAT_DATA_UINT64 },
800 { "uncompressed_size", KSTAT_DATA_UINT64 },
801 { "overhead_size", KSTAT_DATA_UINT64 },
802 { "hdr_size", KSTAT_DATA_UINT64 },
803 { "data_size", KSTAT_DATA_UINT64 },
804 { "metadata_size", KSTAT_DATA_UINT64 },
805 { "ddt_size", KSTAT_DATA_UINT64 },
806 { "other_size", KSTAT_DATA_UINT64 },
807 { "anon_size", KSTAT_DATA_UINT64 },
808 { "anon_evictable_data", KSTAT_DATA_UINT64 },
809 { "anon_evictable_metadata", KSTAT_DATA_UINT64 },
810 { "anon_evictable_ddt", KSTAT_DATA_UINT64 },
811 { "mru_size", KSTAT_DATA_UINT64 },
812 { "mru_evictable_data", KSTAT_DATA_UINT64 },
813 { "mru_evictable_metadata", KSTAT_DATA_UINT64 },
814 { "mru_evictable_ddt", KSTAT_DATA_UINT64 },
815 { "mru_ghost_size", KSTAT_DATA_UINT64 },
816 { "mru_ghost_evictable_data", KSTAT_DATA_UINT64 },
817 { "mru_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
818 { "mru_ghost_evictable_ddt", KSTAT_DATA_UINT64 },
819 { "mfu_size", KSTAT_DATA_UINT64 },
820 { "mfu_evictable_data", KSTAT_DATA_UINT64 },
821 { "mfu_evictable_metadata", KSTAT_DATA_UINT64 },
822 { "mfu_evictable_ddt", KSTAT_DATA_UINT64 },
823 { "mfu_ghost_size", KSTAT_DATA_UINT64 },
824 { "mfu_ghost_evictable_data", KSTAT_DATA_UINT64 },
825 { "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
826 { "mfu_ghost_evictable_ddt", KSTAT_DATA_UINT64 },
827 { "l2_hits", KSTAT_DATA_UINT64 },
828 { "l2_ddt_hits", KSTAT_DATA_UINT64 },
829 { "l2_misses", KSTAT_DATA_UINT64 },
830 { "l2_feeds", KSTAT_DATA_UINT64 },
831 { "l2_rw_clash", KSTAT_DATA_UINT64 },
832 { "l2_read_bytes", KSTAT_DATA_UINT64 },
833 { "l2_ddt_read_bytes", KSTAT_DATA_UINT64 },
834 { "l2_write_bytes", KSTAT_DATA_UINT64 },
835 { "l2_ddt_write_bytes", KSTAT_DATA_UINT64 },
836 { "l2_writes_sent", KSTAT_DATA_UINT64 },
837 { "l2_writes_done", KSTAT_DATA_UINT64 },
838 { "l2_writes_error", KSTAT_DATA_UINT64 },
839 { "l2_writes_lock_retry", KSTAT_DATA_UINT64 },
840 { "l2_evict_lock_retry", KSTAT_DATA_UINT64 },
841 { "l2_evict_reading", KSTAT_DATA_UINT64 },
842 { "l2_evict_l1cached", KSTAT_DATA_UINT64 },
843 { "l2_free_on_write", KSTAT_DATA_UINT64 },
844 { "l2_abort_lowmem", KSTAT_DATA_UINT64 },
845 { "l2_cksum_bad", KSTAT_DATA_UINT64 },
846 { "l2_io_error", KSTAT_DATA_UINT64 },
847 { "l2_size", KSTAT_DATA_UINT64 },
848 { "l2_asize", KSTAT_DATA_UINT64 },
849 { "l2_hdr_size", KSTAT_DATA_UINT64 },
850 { "l2_log_blk_writes", KSTAT_DATA_UINT64 },
851 { "l2_log_blk_avg_size", KSTAT_DATA_UINT64 },
852 { "l2_data_to_meta_ratio", KSTAT_DATA_UINT64 },
853 { "l2_rebuild_successes", KSTAT_DATA_UINT64 },
854 { "l2_rebuild_unsupported", KSTAT_DATA_UINT64 },
855 { "l2_rebuild_io_errors", KSTAT_DATA_UINT64 },
856 { "l2_rebuild_cksum_errors", KSTAT_DATA_UINT64 },
857 { "l2_rebuild_loop_errors", KSTAT_DATA_UINT64 },
858 { "l2_rebuild_lowmem", KSTAT_DATA_UINT64 },
859 { "l2_rebuild_size", KSTAT_DATA_UINT64 },
860 { "l2_rebuild_bufs", KSTAT_DATA_UINT64 },
861 { "l2_rebuild_bufs_precached", KSTAT_DATA_UINT64 },
862 { "l2_rebuild_psize", KSTAT_DATA_UINT64 },
863 { "l2_rebuild_log_blks", KSTAT_DATA_UINT64 },
864 { "memory_throttle_count", KSTAT_DATA_UINT64 },
865 { "arc_meta_used", KSTAT_DATA_UINT64 },
866 { "arc_meta_limit", KSTAT_DATA_UINT64 },
867 { "arc_meta_max", KSTAT_DATA_UINT64 },
868 { "arc_meta_min", KSTAT_DATA_UINT64 },
869 { "arc_ddt_limit", KSTAT_DATA_UINT64 },
870 { "sync_wait_for_async", KSTAT_DATA_UINT64 },
871 { "demand_hit_predictive_prefetch", KSTAT_DATA_UINT64 },
872 };
873
874 #define ARCSTAT(stat) (arc_stats.stat.value.ui64)
875
876 #define ARCSTAT_INCR(stat, val) \
877 atomic_add_64(&arc_stats.stat.value.ui64, (val))
878
879 #define ARCSTAT_BUMP(stat) ARCSTAT_INCR(stat, 1)
880 #define ARCSTAT_BUMPDOWN(stat) ARCSTAT_INCR(stat, -1)
881
882 #define ARCSTAT_MAX(stat, val) { \
883 uint64_t m; \
884 while ((val) > (m = arc_stats.stat.value.ui64) && \
885 (m != atomic_cas_64(&arc_stats.stat.value.ui64, m, (val)))) \
886 continue; \
887 }
888
889 #define ARCSTAT_MAXSTAT(stat) \
890 ARCSTAT_MAX(stat##_max, arc_stats.stat.value.ui64)
891
892 /*
893 * We define a macro to allow ARC hits/misses to be easily broken down by
894 * two separate conditions, giving a total of four different subtypes for
895 * each of hits and misses (so eight statistics total).
896 */
897 #define ARCSTAT_CONDSTAT(cond1, stat1, notstat1, cond2, stat2, notstat2, stat) \
898 if (cond1) { \
899 if (cond2) { \
900 ARCSTAT_BUMP(arcstat_##stat1##_##stat##_##stat2); \
901 } else { \
902 ARCSTAT_BUMP(arcstat_##stat1##_##stat##_##notstat2); \
903 } \
904 } else { \
905 if (cond2) { \
906 ARCSTAT_BUMP(arcstat_##notstat1##_##stat##_##stat2); \
907 } else { \
908 ARCSTAT_BUMP(arcstat_##notstat1##_##stat##_##notstat2);\
909 } \
910 }
911
912 /*
913 * This macro allows us to use kstats as floating averages. Each time we
914 * update this kstat, we first factor it and the update value by
915 * ARCSTAT_AVG_FACTOR to shrink the new value's contribution to the overall
916 * average. This macro assumes that integer loads and stores are atomic, but
917 * is not safe for multiple writers updating the kstat in parallel (only the
918 * last writer's update will remain).
919 */
920 #define ARCSTAT_F_AVG_FACTOR 3
921 #define ARCSTAT_F_AVG(stat, value) \
922 do { \
923 uint64_t x = ARCSTAT(stat); \
924 x = x - x / ARCSTAT_F_AVG_FACTOR + \
925 (value) / ARCSTAT_F_AVG_FACTOR; \
926 ARCSTAT(stat) = x; \
927 _NOTE(CONSTCOND) \
928 } while (0)
929
930 kstat_t *arc_ksp;
931 static arc_state_t *arc_anon;
932 static arc_state_t *arc_mru;
933 static arc_state_t *arc_mru_ghost;
934 static arc_state_t *arc_mfu;
935 static arc_state_t *arc_mfu_ghost;
936 static arc_state_t *arc_l2c_only;
937
938 /*
939 * There are several ARC variables that are critical to export as kstats --
940 * but we don't want to have to grovel around in the kstat whenever we wish to
941 * manipulate them. For these variables, we therefore define them to be in
942 * terms of the statistic variable. This assures that we are not introducing
943 * the possibility of inconsistency by having shadow copies of the variables,
944 * while still allowing the code to be readable.
945 */
946 #define arc_size ARCSTAT(arcstat_size) /* actual total arc size */
947 #define arc_p ARCSTAT(arcstat_p) /* target size of MRU */
948 #define arc_c ARCSTAT(arcstat_c) /* target size of cache */
949 #define arc_c_min ARCSTAT(arcstat_c_min) /* min target cache size */
950 #define arc_c_max ARCSTAT(arcstat_c_max) /* max target cache size */
951 #define arc_meta_limit ARCSTAT(arcstat_meta_limit) /* max size for metadata */
952 #define arc_meta_min ARCSTAT(arcstat_meta_min) /* min size for metadata */
953 #define arc_meta_used ARCSTAT(arcstat_meta_used) /* size of metadata */
954 #define arc_meta_max ARCSTAT(arcstat_meta_max) /* max size of metadata */
955 #define arc_ddt_size ARCSTAT(arcstat_ddt_size) /* ddt size in arc */
956 #define arc_ddt_limit ARCSTAT(arcstat_ddt_limit) /* ddt in arc size limit */
957
958 /*
959 * Used int zio.c to optionally keep DDT cached in ARC
960 */
961 uint64_t const *arc_ddt_evict_threshold;
962
963 /* compressed size of entire arc */
964 #define arc_compressed_size ARCSTAT(arcstat_compressed_size)
965 /* uncompressed size of entire arc */
966 #define arc_uncompressed_size ARCSTAT(arcstat_uncompressed_size)
967 /* number of bytes in the arc from arc_buf_t's */
968 #define arc_overhead_size ARCSTAT(arcstat_overhead_size)
969
970
971 static int arc_no_grow; /* Don't try to grow cache size */
972 static uint64_t arc_tempreserve;
973 static uint64_t arc_loaned_bytes;
974
975 typedef struct arc_callback arc_callback_t;
976
977 struct arc_callback {
978 void *acb_private;
979 arc_done_func_t *acb_done;
980 arc_buf_t *acb_buf;
981 boolean_t acb_compressed;
982 zio_t *acb_zio_dummy;
983 arc_callback_t *acb_next;
984 };
985
986 typedef struct arc_write_callback arc_write_callback_t;
987
988 struct arc_write_callback {
989 void *awcb_private;
1010 * | l2arc_buf_hdr_t | | l2arc_buf_hdr_t |
1011 * | (undefined if L1-only) | | |
1012 * +------------------------+ +------------------------+
1013 * | l1arc_buf_hdr_t |
1014 * | |
1015 * | |
1016 * | |
1017 * | |
1018 * +------------------------+
1019 *
1020 * Because it's possible for the L2ARC to become extremely large, we can wind
1021 * up eating a lot of memory in L2ARC buffer headers, so the size of a header
1022 * is minimized by only allocating the fields necessary for an L1-cached buffer
1023 * when a header is actually in the L1 cache. The sub-headers (l1arc_buf_hdr and
1024 * l2arc_buf_hdr) are embedded rather than allocated separately to save a couple
1025 * words in pointers. arc_hdr_realloc() is used to switch a header between
1026 * these two allocation states.
1027 */
1028 typedef struct l1arc_buf_hdr {
1029 kmutex_t b_freeze_lock;
1030 #ifdef ZFS_DEBUG
1031 /*
1032 * Used for debugging with kmem_flags - by allocating and freeing
1033 * b_thawed when the buffer is thawed, we get a record of the stack
1034 * trace that thawed it.
1035 */
1036 void *b_thawed;
1037 #endif
1038
1039 /* number of krrp tasks using this buffer */
1040 uint64_t b_krrp;
1041
1042 arc_buf_t *b_buf;
1043 uint32_t b_bufcnt;
1044 /* for waiting on writes to complete */
1045 kcondvar_t b_cv;
1046 uint8_t b_byteswap;
1047
1048 /* protected by arc state mutex */
1049 arc_state_t *b_state;
1050 multilist_node_t b_arc_node;
1051
1052 /* updated atomically */
1053 clock_t b_arc_access;
1054
1055 /* self protecting */
1056 refcount_t b_refcnt;
1057
1058 arc_callback_t *b_acb;
1059 abd_t *b_pabd;
1060 } l1arc_buf_hdr_t;
1061
1062 typedef struct l2arc_dev l2arc_dev_t;
1063
1064 typedef struct l2arc_buf_hdr {
1065 /* protected by arc_buf_hdr mutex */
1066 l2arc_dev_t *b_dev; /* L2ARC device */
1067 uint64_t b_daddr; /* disk address, offset byte */
1068
1069 list_node_t b_l2node;
1070 } l2arc_buf_hdr_t;
1071
1072 struct arc_buf_hdr {
1073 /* protected by hash lock */
1074 dva_t b_dva;
1075 uint64_t b_birth;
1076
1077 /*
1078 * Even though this checksum is only set/verified when a buffer is in
1079 * the L1 cache, it needs to be in the set of common fields because it
1080 * must be preserved from the time before a buffer is written out to
1081 * L2ARC until after it is read back in.
1082 */
1083 zio_cksum_t *b_freeze_cksum;
1084
1085 arc_buf_contents_t b_type;
1086 arc_buf_hdr_t *b_hash_next;
1087 arc_flags_t b_flags;
1088
1089 /*
1090 * This field stores the size of the data buffer after
1091 * compression, and is set in the arc's zio completion handlers.
1092 * It is in units of SPA_MINBLOCKSIZE (e.g. 1 == 512 bytes).
1093 *
1094 * While the block pointers can store up to 32MB in their psize
1095 * field, we can only store up to 32MB minus 512B. This is due
1096 * to the bp using a bias of 1, whereas we use a bias of 0 (i.e.
1097 * a field of zeros represents 512B in the bp). We can't use a
1098 * bias of 1 since we need to reserve a psize of zero, here, to
1099 * represent holes and embedded blocks.
1100 *
1101 * This isn't a problem in practice, since the maximum size of a
1102 * buffer is limited to 16MB, so we never need to store 32MB in
1103 * this field. Even in the upstream illumos code base, the
1104 * maximum size of a buffer is limited to 16MB.
1122 #define GHOST_STATE(state) \
1123 ((state) == arc_mru_ghost || (state) == arc_mfu_ghost || \
1124 (state) == arc_l2c_only)
1125
1126 #define HDR_IN_HASH_TABLE(hdr) ((hdr)->b_flags & ARC_FLAG_IN_HASH_TABLE)
1127 #define HDR_IO_IN_PROGRESS(hdr) ((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS)
1128 #define HDR_IO_ERROR(hdr) ((hdr)->b_flags & ARC_FLAG_IO_ERROR)
1129 #define HDR_PREFETCH(hdr) ((hdr)->b_flags & ARC_FLAG_PREFETCH)
1130 #define HDR_COMPRESSION_ENABLED(hdr) \
1131 ((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
1132
1133 #define HDR_L2CACHE(hdr) ((hdr)->b_flags & ARC_FLAG_L2CACHE)
1134 #define HDR_L2_READING(hdr) \
1135 (((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) && \
1136 ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
1137 #define HDR_L2_WRITING(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITING)
1138 #define HDR_L2_EVICTED(hdr) ((hdr)->b_flags & ARC_FLAG_L2_EVICTED)
1139 #define HDR_L2_WRITE_HEAD(hdr) ((hdr)->b_flags & ARC_FLAG_L2_WRITE_HEAD)
1140 #define HDR_SHARED_DATA(hdr) ((hdr)->b_flags & ARC_FLAG_SHARED_DATA)
1141
1142 #define HDR_ISTYPE_DDT(hdr) \
1143 ((hdr)->b_flags & ARC_FLAG_BUFC_DDT)
1144 #define HDR_ISTYPE_METADATA(hdr) \
1145 ((hdr)->b_flags & ARC_FLAG_BUFC_METADATA)
1146 #define HDR_ISTYPE_DATA(hdr) (!HDR_ISTYPE_METADATA(hdr) && \
1147 !HDR_ISTYPE_DDT(hdr))
1148
1149 #define HDR_HAS_L1HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L1HDR)
1150 #define HDR_HAS_L2HDR(hdr) ((hdr)->b_flags & ARC_FLAG_HAS_L2HDR)
1151
1152 /* For storing compression mode in b_flags */
1153 #define HDR_COMPRESS_OFFSET (highbit64(ARC_FLAG_COMPRESS_0) - 1)
1154
1155 #define HDR_GET_COMPRESS(hdr) ((enum zio_compress)BF32_GET((hdr)->b_flags, \
1156 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS))
1157 #define HDR_SET_COMPRESS(hdr, cmp) BF32_SET((hdr)->b_flags, \
1158 HDR_COMPRESS_OFFSET, SPA_COMPRESSBITS, (cmp));
1159
1160 #define ARC_BUF_LAST(buf) ((buf)->b_next == NULL)
1161 #define ARC_BUF_SHARED(buf) ((buf)->b_flags & ARC_BUF_FLAG_SHARED)
1162 #define ARC_BUF_COMPRESSED(buf) ((buf)->b_flags & ARC_BUF_FLAG_COMPRESSED)
1163
1164 /*
1165 * Other sizes
1166 */
1167
1168 #define HDR_FULL_SIZE ((int64_t)sizeof (arc_buf_hdr_t))
1169 #define HDR_L2ONLY_SIZE ((int64_t)offsetof(arc_buf_hdr_t, b_l1hdr))
1170
1171 /*
1172 * Hash table routines
1173 */
1174
1175 struct ht_table {
1176 arc_buf_hdr_t *hdr;
1177 kmutex_t lock;
1178 };
1179
1180 typedef struct buf_hash_table {
1181 uint64_t ht_mask;
1182 struct ht_table *ht_table;
1183 } buf_hash_table_t;
1184
1185 #pragma align 64(buf_hash_table)
1186 static buf_hash_table_t buf_hash_table;
1187
1188 #define BUF_HASH_INDEX(spa, dva, birth) \
1189 (buf_hash(spa, dva, birth) & buf_hash_table.ht_mask)
1190 #define BUF_HASH_LOCK(idx) (&buf_hash_table.ht_table[idx].lock)
1191 #define HDR_LOCK(hdr) \
1192 (BUF_HASH_LOCK(BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth)))
1193
1194 uint64_t zfs_crc64_table[256];
1195
1196 /*
1197 * Level 2 ARC
1198 */
1199
1200 #define L2ARC_WRITE_SIZE (8 * 1024 * 1024) /* initial write max */
1201 #define L2ARC_HEADROOM 2 /* num of writes */
1202 /*
1203 * If we discover during ARC scan any buffers to be compressed, we boost
1204 * our headroom for the next scanning cycle by this percentage multiple.
1205 */
1206 #define L2ARC_HEADROOM_BOOST 200
1207 #define L2ARC_FEED_SECS 1 /* caching interval secs */
1208 #define L2ARC_FEED_MIN_MS 200 /* min caching interval ms */
1209
1210 #define l2arc_writes_sent ARCSTAT(arcstat_l2_writes_sent)
1211 #define l2arc_writes_done ARCSTAT(arcstat_l2_writes_done)
1212
1213 /* L2ARC Performance Tunables */
1214 uint64_t l2arc_write_max = L2ARC_WRITE_SIZE; /* default max write size */
1215 uint64_t l2arc_write_boost = L2ARC_WRITE_SIZE; /* extra write during warmup */
1216 uint64_t l2arc_headroom = L2ARC_HEADROOM; /* number of dev writes */
1217 uint64_t l2arc_headroom_boost = L2ARC_HEADROOM_BOOST;
1218 uint64_t l2arc_feed_secs = L2ARC_FEED_SECS; /* interval seconds */
1219 uint64_t l2arc_feed_min_ms = L2ARC_FEED_MIN_MS; /* min interval milliseconds */
1220 boolean_t l2arc_noprefetch = B_TRUE; /* don't cache prefetch bufs */
1221 boolean_t l2arc_feed_again = B_TRUE; /* turbo warmup */
1222 boolean_t l2arc_norw = B_TRUE; /* no reads during writes */
1223
1224 static list_t L2ARC_dev_list; /* device list */
1225 static list_t *l2arc_dev_list; /* device list pointer */
1226 static kmutex_t l2arc_dev_mtx; /* device list mutex */
1227 static l2arc_dev_t *l2arc_dev_last; /* last device used */
1228 static l2arc_dev_t *l2arc_ddt_dev_last; /* last DDT device used */
1229 static list_t L2ARC_free_on_write; /* free after write buf list */
1230 static list_t *l2arc_free_on_write; /* free after write list ptr */
1231 static kmutex_t l2arc_free_on_write_mtx; /* mutex for list */
1232 static uint64_t l2arc_ndev; /* number of devices */
1233
1234 typedef struct l2arc_read_callback {
1235 arc_buf_hdr_t *l2rcb_hdr; /* read header */
1236 blkptr_t l2rcb_bp; /* original blkptr */
1237 zbookmark_phys_t l2rcb_zb; /* original bookmark */
1238 int l2rcb_flags; /* original flags */
1239 abd_t *l2rcb_abd; /* temporary buffer */
1240 } l2arc_read_callback_t;
1241
1242 typedef struct l2arc_write_callback {
1243 l2arc_dev_t *l2wcb_dev; /* device info */
1244 arc_buf_hdr_t *l2wcb_head; /* head of write buflist */
1245 list_t l2wcb_log_blk_buflist; /* in-flight log blocks */
1246 } l2arc_write_callback_t;
1247
1248 typedef struct l2arc_data_free {
1249 /* protected by l2arc_free_on_write_mtx */
1250 abd_t *l2df_abd;
1251 size_t l2df_size;
1252 arc_buf_contents_t l2df_type;
1253 list_node_t l2df_list_node;
1254 } l2arc_data_free_t;
1255
1256 static kmutex_t l2arc_feed_thr_lock;
1257 static kcondvar_t l2arc_feed_thr_cv;
1258 static uint8_t l2arc_thread_exit;
1259
1260 static abd_t *arc_get_data_abd(arc_buf_hdr_t *, uint64_t, void *);
1261 static void *arc_get_data_buf(arc_buf_hdr_t *, uint64_t, void *);
1262 static void arc_get_data_impl(arc_buf_hdr_t *, uint64_t, void *);
1263 static void arc_free_data_abd(arc_buf_hdr_t *, abd_t *, uint64_t, void *);
1264 static void arc_free_data_buf(arc_buf_hdr_t *, void *, uint64_t, void *);
1265 static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag);
1266 static void arc_hdr_free_pabd(arc_buf_hdr_t *);
1267 static void arc_hdr_alloc_pabd(arc_buf_hdr_t *);
1268 static void arc_access(arc_buf_hdr_t *, kmutex_t *);
1269 static boolean_t arc_is_overflowing();
1270 static void arc_buf_watch(arc_buf_t *);
1271 static l2arc_dev_t *l2arc_vdev_get(vdev_t *vd);
1272
1273 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
1274 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
1275 static arc_buf_contents_t arc_flags_to_bufc(uint32_t);
1276 static inline void arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1277 static inline void arc_hdr_clear_flags(arc_buf_hdr_t *hdr, arc_flags_t flags);
1278
1279 static boolean_t l2arc_write_eligible(uint64_t, arc_buf_hdr_t *);
1280 static void l2arc_read_done(zio_t *);
1281
1282 static void
1283 arc_update_hit_stat(arc_buf_hdr_t *hdr, boolean_t hit)
1284 {
1285 boolean_t pf = !HDR_PREFETCH(hdr);
1286 switch (arc_buf_type(hdr)) {
1287 case ARC_BUFC_DATA:
1288 ARCSTAT_CONDSTAT(pf, demand, prefetch, hit, hits, misses, data);
1289 break;
1290 case ARC_BUFC_METADATA:
1291 ARCSTAT_CONDSTAT(pf, demand, prefetch, hit, hits, misses,
1292 metadata);
1293 break;
1294 case ARC_BUFC_DDT:
1295 ARCSTAT_CONDSTAT(pf, demand, prefetch, hit, hits, misses, ddt);
1296 break;
1297 default:
1298 break;
1299 }
1300 }
1301
1302 enum {
1303 L2ARC_DEV_HDR_EVICT_FIRST = (1 << 0) /* mirror of l2ad_first */
1304 };
1305
1306 /*
1307 * Pointer used in persistent L2ARC (for pointing to log blocks & ARC buffers).
1308 */
1309 typedef struct l2arc_log_blkptr {
1310 uint64_t lbp_daddr; /* device address of log */
1311 /*
1312 * lbp_prop is the same format as the blk_prop in blkptr_t:
1313 * * logical size (in sectors)
1314 * * physical size (in sectors)
1315 * * checksum algorithm (used for lbp_cksum)
1316 * * object type & level (unused for now)
1317 */
1318 uint64_t lbp_prop;
1319 zio_cksum_t lbp_cksum; /* fletcher4 of log */
1320 } l2arc_log_blkptr_t;
1321
1322 /*
1323 * The persistent L2ARC device header.
1324 * Byte order of magic determines whether 64-bit bswap of fields is necessary.
1325 */
1326 typedef struct l2arc_dev_hdr_phys {
1327 uint64_t dh_magic; /* L2ARC_DEV_HDR_MAGIC_Vx */
1328 zio_cksum_t dh_self_cksum; /* fletcher4 of fields below */
1329
1330 /*
1331 * Global L2ARC device state and metadata.
1332 */
1333 uint64_t dh_spa_guid;
1334 uint64_t dh_alloc_space; /* vdev space alloc status */
1335 uint64_t dh_flags; /* l2arc_dev_hdr_flags_t */
1336
1337 /*
1338 * Start of log block chain. [0] -> newest log, [1] -> one older (used
1339 * for initiating prefetch).
1340 */
1341 l2arc_log_blkptr_t dh_start_lbps[2];
1342
1343 const uint64_t dh_pad[44]; /* pad to 512 bytes */
1344 } l2arc_dev_hdr_phys_t;
1345 CTASSERT(sizeof (l2arc_dev_hdr_phys_t) == SPA_MINBLOCKSIZE);
1346
1347 /*
1348 * A single ARC buffer header entry in a l2arc_log_blk_phys_t.
1349 */
1350 typedef struct l2arc_log_ent_phys {
1351 dva_t le_dva; /* dva of buffer */
1352 uint64_t le_birth; /* birth txg of buffer */
1353 zio_cksum_t le_freeze_cksum;
1354 /*
1355 * le_prop is the same format as the blk_prop in blkptr_t:
1356 * * logical size (in sectors)
1357 * * physical size (in sectors)
1358 * * checksum algorithm (used for b_freeze_cksum)
1359 * * object type & level (used to restore arc_buf_contents_t)
1360 */
1361 uint64_t le_prop;
1362 uint64_t le_daddr; /* buf location on l2dev */
1363 const uint64_t le_pad[7]; /* resv'd for future use */
1364 } l2arc_log_ent_phys_t;
1365
1366 /*
1367 * These design limits give us the following metadata overhead (before
1368 * compression):
1369 * avg_blk_sz overhead
1370 * 1k 12.51 %
1371 * 2k 6.26 %
1372 * 4k 3.13 %
1373 * 8k 1.56 %
1374 * 16k 0.78 %
1375 * 32k 0.39 %
1376 * 64k 0.20 %
1377 * 128k 0.10 %
1378 * Compression should be able to sequeeze these down by about a factor of 2x.
1379 */
1380 #define L2ARC_LOG_BLK_SIZE (128 * 1024) /* 128k */
1381 #define L2ARC_LOG_BLK_HEADER_LEN (128)
1382 #define L2ARC_LOG_BLK_ENTRIES /* 1023 entries */ \
1383 ((L2ARC_LOG_BLK_SIZE - L2ARC_LOG_BLK_HEADER_LEN) / \
1384 sizeof (l2arc_log_ent_phys_t))
1385 /*
1386 * Maximum amount of data in an l2arc log block (used to terminate rebuilding
1387 * before we hit the write head and restore potentially corrupted blocks).
1388 */
1389 #define L2ARC_LOG_BLK_MAX_PAYLOAD_SIZE \
1390 (SPA_MAXBLOCKSIZE * L2ARC_LOG_BLK_ENTRIES)
1391 /*
1392 * For the persistency and rebuild algorithms to operate reliably we need
1393 * the L2ARC device to at least be able to hold 3 full log blocks (otherwise
1394 * excessive log block looping might confuse the log chain end detection).
1395 * Under normal circumstances this is not a problem, since this is somewhere
1396 * around only 400 MB.
1397 */
1398 #define L2ARC_PERSIST_MIN_SIZE (3 * L2ARC_LOG_BLK_MAX_PAYLOAD_SIZE)
1399
1400 /*
1401 * A log block of up to 1023 ARC buffer log entries, chained into the
1402 * persistent L2ARC metadata linked list. Byte order of magic determines
1403 * whether 64-bit bswap of fields is necessary.
1404 */
1405 typedef struct l2arc_log_blk_phys {
1406 /* Header - see L2ARC_LOG_BLK_HEADER_LEN above */
1407 uint64_t lb_magic; /* L2ARC_LOG_BLK_MAGIC */
1408 l2arc_log_blkptr_t lb_back2_lbp; /* back 2 steps in chain */
1409 uint64_t lb_pad[9]; /* resv'd for future use */
1410 /* Payload */
1411 l2arc_log_ent_phys_t lb_entries[L2ARC_LOG_BLK_ENTRIES];
1412 } l2arc_log_blk_phys_t;
1413
1414 CTASSERT(sizeof (l2arc_log_blk_phys_t) == L2ARC_LOG_BLK_SIZE);
1415 CTASSERT(offsetof(l2arc_log_blk_phys_t, lb_entries) -
1416 offsetof(l2arc_log_blk_phys_t, lb_magic) == L2ARC_LOG_BLK_HEADER_LEN);
1417
1418 /*
1419 * These structures hold in-flight l2arc_log_blk_phys_t's as they're being
1420 * written to the L2ARC device. They may be compressed, hence the uint8_t[].
1421 */
1422 typedef struct l2arc_log_blk_buf {
1423 uint8_t lbb_log_blk[sizeof (l2arc_log_blk_phys_t)];
1424 list_node_t lbb_node;
1425 } l2arc_log_blk_buf_t;
1426
1427 /* Macros for the manipulation fields in the blk_prop format of blkptr_t */
1428 #define BLKPROP_GET_LSIZE(_obj, _field) \
1429 BF64_GET_SB((_obj)->_field, 0, 16, SPA_MINBLOCKSHIFT, 1)
1430 #define BLKPROP_SET_LSIZE(_obj, _field, x) \
1431 BF64_SET_SB((_obj)->_field, 0, 16, SPA_MINBLOCKSHIFT, 1, x)
1432 #define BLKPROP_GET_PSIZE(_obj, _field) \
1433 BF64_GET_SB((_obj)->_field, 16, 16, SPA_MINBLOCKSHIFT, 0)
1434 #define BLKPROP_SET_PSIZE(_obj, _field, x) \
1435 BF64_SET_SB((_obj)->_field, 16, 16, SPA_MINBLOCKSHIFT, 0, x)
1436 #define BLKPROP_GET_COMPRESS(_obj, _field) \
1437 BF64_GET((_obj)->_field, 32, 7)
1438 #define BLKPROP_SET_COMPRESS(_obj, _field, x) \
1439 BF64_SET((_obj)->_field, 32, 7, x)
1440 #define BLKPROP_GET_ARC_COMPRESS(_obj, _field) \
1441 BF64_GET((_obj)->_field, 39, 1)
1442 #define BLKPROP_SET_ARC_COMPRESS(_obj, _field, x) \
1443 BF64_SET((_obj)->_field, 39, 1, x)
1444 #define BLKPROP_GET_CHECKSUM(_obj, _field) \
1445 BF64_GET((_obj)->_field, 40, 8)
1446 #define BLKPROP_SET_CHECKSUM(_obj, _field, x) \
1447 BF64_SET((_obj)->_field, 40, 8, x)
1448 #define BLKPROP_GET_TYPE(_obj, _field) \
1449 BF64_GET((_obj)->_field, 48, 8)
1450 #define BLKPROP_SET_TYPE(_obj, _field, x) \
1451 BF64_SET((_obj)->_field, 48, 8, x)
1452
1453 /* Macros for manipulating a l2arc_log_blkptr_t->lbp_prop field */
1454 #define LBP_GET_LSIZE(_add) BLKPROP_GET_LSIZE(_add, lbp_prop)
1455 #define LBP_SET_LSIZE(_add, x) BLKPROP_SET_LSIZE(_add, lbp_prop, x)
1456 #define LBP_GET_PSIZE(_add) BLKPROP_GET_PSIZE(_add, lbp_prop)
1457 #define LBP_SET_PSIZE(_add, x) BLKPROP_SET_PSIZE(_add, lbp_prop, x)
1458 #define LBP_GET_COMPRESS(_add) BLKPROP_GET_COMPRESS(_add, lbp_prop)
1459 #define LBP_SET_COMPRESS(_add, x) BLKPROP_SET_COMPRESS(_add, lbp_prop, x)
1460 #define LBP_GET_CHECKSUM(_add) BLKPROP_GET_CHECKSUM(_add, lbp_prop)
1461 #define LBP_SET_CHECKSUM(_add, x) BLKPROP_SET_CHECKSUM(_add, lbp_prop, x)
1462 #define LBP_GET_TYPE(_add) BLKPROP_GET_TYPE(_add, lbp_prop)
1463 #define LBP_SET_TYPE(_add, x) BLKPROP_SET_TYPE(_add, lbp_prop, x)
1464
1465 /* Macros for manipulating a l2arc_log_ent_phys_t->le_prop field */
1466 #define LE_GET_LSIZE(_le) BLKPROP_GET_LSIZE(_le, le_prop)
1467 #define LE_SET_LSIZE(_le, x) BLKPROP_SET_LSIZE(_le, le_prop, x)
1468 #define LE_GET_PSIZE(_le) BLKPROP_GET_PSIZE(_le, le_prop)
1469 #define LE_SET_PSIZE(_le, x) BLKPROP_SET_PSIZE(_le, le_prop, x)
1470 #define LE_GET_COMPRESS(_le) BLKPROP_GET_COMPRESS(_le, le_prop)
1471 #define LE_SET_COMPRESS(_le, x) BLKPROP_SET_COMPRESS(_le, le_prop, x)
1472 #define LE_GET_ARC_COMPRESS(_le) BLKPROP_GET_ARC_COMPRESS(_le, le_prop)
1473 #define LE_SET_ARC_COMPRESS(_le, x) BLKPROP_SET_ARC_COMPRESS(_le, le_prop, x)
1474 #define LE_GET_CHECKSUM(_le) BLKPROP_GET_CHECKSUM(_le, le_prop)
1475 #define LE_SET_CHECKSUM(_le, x) BLKPROP_SET_CHECKSUM(_le, le_prop, x)
1476 #define LE_GET_TYPE(_le) BLKPROP_GET_TYPE(_le, le_prop)
1477 #define LE_SET_TYPE(_le, x) BLKPROP_SET_TYPE(_le, le_prop, x)
1478
1479 #define PTR_SWAP(x, y) \
1480 do { \
1481 void *tmp = (x);\
1482 x = y; \
1483 y = tmp; \
1484 _NOTE(CONSTCOND)\
1485 } while (0)
1486
1487 /*
1488 * Sadly, after compressed ARC integration older kernels would panic
1489 * when trying to rebuild persistent L2ARC created by the new code.
1490 */
1491 #define L2ARC_DEV_HDR_MAGIC_V1 0x4c32415243763031LLU /* ASCII: "L2ARCv01" */
1492 #define L2ARC_LOG_BLK_MAGIC 0x4c4f47424c4b4844LLU /* ASCII: "LOGBLKHD" */
1493
1494 /*
1495 * Performance tuning of L2ARC persistency:
1496 *
1497 * l2arc_rebuild_enabled : Controls whether L2ARC device adds (either at
1498 * pool import or when adding one manually later) will attempt
1499 * to rebuild L2ARC buffer contents. In special circumstances,
1500 * the administrator may want to set this to B_FALSE, if they
1501 * are having trouble importing a pool or attaching an L2ARC
1502 * device (e.g. the L2ARC device is slow to read in stored log
1503 * metadata, or the metadata has become somehow
1504 * fragmented/unusable).
1505 */
1506 boolean_t l2arc_rebuild_enabled = B_TRUE;
1507
1508 /* L2ARC persistency rebuild control routines. */
1509 static void l2arc_dev_rebuild_start(l2arc_dev_t *dev);
1510 static int l2arc_rebuild(l2arc_dev_t *dev);
1511
1512 /* L2ARC persistency read I/O routines. */
1513 static int l2arc_dev_hdr_read(l2arc_dev_t *dev);
1514 static int l2arc_log_blk_read(l2arc_dev_t *dev,
1515 const l2arc_log_blkptr_t *this_lp, const l2arc_log_blkptr_t *next_lp,
1516 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
1517 uint8_t *this_lb_buf, uint8_t *next_lb_buf,
1518 zio_t *this_io, zio_t **next_io);
1519 static zio_t *l2arc_log_blk_prefetch(vdev_t *vd,
1520 const l2arc_log_blkptr_t *lp, uint8_t *lb_buf);
1521 static void l2arc_log_blk_prefetch_abort(zio_t *zio);
1522
1523 /* L2ARC persistency block restoration routines. */
1524 static void l2arc_log_blk_restore(l2arc_dev_t *dev, uint64_t load_guid,
1525 const l2arc_log_blk_phys_t *lb, uint64_t lb_psize);
1526 static void l2arc_hdr_restore(const l2arc_log_ent_phys_t *le,
1527 l2arc_dev_t *dev, uint64_t guid);
1528
1529 /* L2ARC persistency write I/O routines. */
1530 static void l2arc_dev_hdr_update(l2arc_dev_t *dev, zio_t *pio);
1531 static void l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
1532 l2arc_write_callback_t *cb);
1533
1534 /* L2ARC persistency auxilliary routines. */
1535 static boolean_t l2arc_log_blkptr_valid(l2arc_dev_t *dev,
1536 const l2arc_log_blkptr_t *lp);
1537 static void l2arc_dev_hdr_checksum(const l2arc_dev_hdr_phys_t *hdr,
1538 zio_cksum_t *cksum);
1539 static boolean_t l2arc_log_blk_insert(l2arc_dev_t *dev,
1540 const arc_buf_hdr_t *ab);
1541 static inline boolean_t l2arc_range_check_overlap(uint64_t bottom,
1542 uint64_t top, uint64_t check);
1543
1544 /*
1545 * L2ARC Internals
1546 */
1547 struct l2arc_dev {
1548 vdev_t *l2ad_vdev; /* vdev */
1549 spa_t *l2ad_spa; /* spa */
1550 uint64_t l2ad_hand; /* next write location */
1551 uint64_t l2ad_start; /* first addr on device */
1552 uint64_t l2ad_end; /* last addr on device */
1553 boolean_t l2ad_first; /* first sweep through */
1554 boolean_t l2ad_writing; /* currently writing */
1555 kmutex_t l2ad_mtx; /* lock for buffer list */
1556 list_t l2ad_buflist; /* buffer list */
1557 list_node_t l2ad_node; /* device list node */
1558 refcount_t l2ad_alloc; /* allocated bytes */
1559 l2arc_dev_hdr_phys_t *l2ad_dev_hdr; /* persistent device header */
1560 uint64_t l2ad_dev_hdr_asize; /* aligned hdr size */
1561 l2arc_log_blk_phys_t l2ad_log_blk; /* currently open log block */
1562 int l2ad_log_ent_idx; /* index into cur log blk */
1563 /* number of bytes in current log block's payload */
1564 uint64_t l2ad_log_blk_payload_asize;
1565 /* flag indicating whether a rebuild is scheduled or is going on */
1566 boolean_t l2ad_rebuild;
1567 boolean_t l2ad_rebuild_cancel;
1568 kt_did_t l2ad_rebuild_did;
1569 };
1570
1571 static inline uint64_t
1572 buf_hash(uint64_t spa, const dva_t *dva, uint64_t birth)
1573 {
1574 uint8_t *vdva = (uint8_t *)dva;
1575 uint64_t crc = -1ULL;
1576 int i;
1577
1578 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
1579
1580 for (i = 0; i < sizeof (dva_t); i++)
1581 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ vdva[i]) & 0xFF];
1582
1583 crc ^= (spa>>8) ^ birth;
1584
1585 return (crc);
1586 }
1587
1588 #define HDR_EMPTY(hdr) \
1589 ((hdr)->b_dva.dva_word[0] == 0 && \
1590 (hdr)->b_dva.dva_word[1] == 0)
1591
1592 #define HDR_EQUAL(spa, dva, birth, hdr) \
1593 ((hdr)->b_dva.dva_word[0] == (dva)->dva_word[0]) && \
1594 ((hdr)->b_dva.dva_word[1] == (dva)->dva_word[1]) && \
1595 ((hdr)->b_birth == birth) && ((hdr)->b_spa == spa)
1596
1597 static void
1598 buf_discard_identity(arc_buf_hdr_t *hdr)
1599 {
1600 hdr->b_dva.dva_word[0] = 0;
1601 hdr->b_dva.dva_word[1] = 0;
1602 hdr->b_birth = 0;
1603 }
1604
1605 static arc_buf_hdr_t *
1606 buf_hash_find(uint64_t spa, const blkptr_t *bp, kmutex_t **lockp)
1607 {
1608 const dva_t *dva = BP_IDENTITY(bp);
1609 uint64_t birth = BP_PHYSICAL_BIRTH(bp);
1610 uint64_t idx = BUF_HASH_INDEX(spa, dva, birth);
1611 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1612 arc_buf_hdr_t *hdr;
1613
1614 mutex_enter(hash_lock);
1615 for (hdr = buf_hash_table.ht_table[idx].hdr; hdr != NULL;
1616 hdr = hdr->b_hash_next) {
1617 if (HDR_EQUAL(spa, dva, birth, hdr)) {
1618 *lockp = hash_lock;
1619 return (hdr);
1620 }
1621 }
1622 mutex_exit(hash_lock);
1623 *lockp = NULL;
1624 return (NULL);
1625 }
1626
1627 /*
1628 * Insert an entry into the hash table. If there is already an element
1629 * equal to elem in the hash table, then the already existing element
1630 * will be returned and the new element will not be inserted.
1631 * Otherwise returns NULL.
1632 * If lockp == NULL, the caller is assumed to already hold the hash lock.
1633 */
1634 static arc_buf_hdr_t *
1635 buf_hash_insert(arc_buf_hdr_t *hdr, kmutex_t **lockp)
1636 {
1637 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1638 kmutex_t *hash_lock = BUF_HASH_LOCK(idx);
1639 arc_buf_hdr_t *fhdr;
1640 uint32_t i;
1641
1642 ASSERT(!DVA_IS_EMPTY(&hdr->b_dva));
1643 ASSERT(hdr->b_birth != 0);
1644 ASSERT(!HDR_IN_HASH_TABLE(hdr));
1645
1646 if (lockp != NULL) {
1647 *lockp = hash_lock;
1648 mutex_enter(hash_lock);
1649 } else {
1650 ASSERT(MUTEX_HELD(hash_lock));
1651 }
1652
1653 for (fhdr = buf_hash_table.ht_table[idx].hdr, i = 0; fhdr != NULL;
1654 fhdr = fhdr->b_hash_next, i++) {
1655 if (HDR_EQUAL(hdr->b_spa, &hdr->b_dva, hdr->b_birth, fhdr))
1656 return (fhdr);
1657 }
1658
1659 hdr->b_hash_next = buf_hash_table.ht_table[idx].hdr;
1660 buf_hash_table.ht_table[idx].hdr = hdr;
1661 arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1662
1663 /* collect some hash table performance data */
1664 if (i > 0) {
1665 ARCSTAT_BUMP(arcstat_hash_collisions);
1666 if (i == 1)
1667 ARCSTAT_BUMP(arcstat_hash_chains);
1668
1669 ARCSTAT_MAX(arcstat_hash_chain_max, i);
1670 }
1671
1672 ARCSTAT_BUMP(arcstat_hash_elements);
1673 ARCSTAT_MAXSTAT(arcstat_hash_elements);
1674
1675 return (NULL);
1676 }
1677
1678 static void
1679 buf_hash_remove(arc_buf_hdr_t *hdr)
1680 {
1681 arc_buf_hdr_t *fhdr, **hdrp;
1682 uint64_t idx = BUF_HASH_INDEX(hdr->b_spa, &hdr->b_dva, hdr->b_birth);
1683
1684 ASSERT(MUTEX_HELD(BUF_HASH_LOCK(idx)));
1685 ASSERT(HDR_IN_HASH_TABLE(hdr));
1686
1687 hdrp = &buf_hash_table.ht_table[idx].hdr;
1688 while ((fhdr = *hdrp) != hdr) {
1689 ASSERT3P(fhdr, !=, NULL);
1690 hdrp = &fhdr->b_hash_next;
1691 }
1692 *hdrp = hdr->b_hash_next;
1693 hdr->b_hash_next = NULL;
1694 arc_hdr_clear_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
1695
1696 /* collect some hash table performance data */
1697 ARCSTAT_BUMPDOWN(arcstat_hash_elements);
1698
1699 if (buf_hash_table.ht_table[idx].hdr &&
1700 buf_hash_table.ht_table[idx].hdr->b_hash_next == NULL)
1701 ARCSTAT_BUMPDOWN(arcstat_hash_chains);
1702 }
1703
1704 /*
1705 * Global data structures and functions for the buf kmem cache.
1706 */
1707 static kmem_cache_t *hdr_full_cache;
1708 static kmem_cache_t *hdr_l2only_cache;
1709 static kmem_cache_t *buf_cache;
1710
1711 static void
1712 buf_fini(void)
1713 {
1714 int i;
1715
1716 for (i = 0; i < buf_hash_table.ht_mask + 1; i++)
1717 mutex_destroy(&buf_hash_table.ht_table[i].lock);
1718 kmem_free(buf_hash_table.ht_table,
1719 (buf_hash_table.ht_mask + 1) * sizeof (struct ht_table));
1720 kmem_cache_destroy(hdr_full_cache);
1721 kmem_cache_destroy(hdr_l2only_cache);
1722 kmem_cache_destroy(buf_cache);
1723 }
1724
1725 /*
1726 * Constructor callback - called when the cache is empty
1727 * and a new buf is requested.
1728 */
1729 /* ARGSUSED */
1730 static int
1731 hdr_full_cons(void *vbuf, void *unused, int kmflag)
1732 {
1733 arc_buf_hdr_t *hdr = vbuf;
1734
1735 bzero(hdr, HDR_FULL_SIZE);
1736 cv_init(&hdr->b_l1hdr.b_cv, NULL, CV_DEFAULT, NULL);
1737 refcount_create(&hdr->b_l1hdr.b_refcnt);
1738 mutex_init(&hdr->b_l1hdr.b_freeze_lock, NULL, MUTEX_DEFAULT, NULL);
1739 multilist_link_init(&hdr->b_l1hdr.b_arc_node);
1822 }
1823
1824 static void
1825 buf_init(void)
1826 {
1827 uint64_t *ct;
1828 uint64_t hsize = 1ULL << 12;
1829 int i, j;
1830
1831 /*
1832 * The hash table is big enough to fill all of physical memory
1833 * with an average block size of zfs_arc_average_blocksize (default 8K).
1834 * By default, the table will take up
1835 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
1836 */
1837 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
1838 hsize <<= 1;
1839 retry:
1840 buf_hash_table.ht_mask = hsize - 1;
1841 buf_hash_table.ht_table =
1842 kmem_zalloc(hsize * sizeof (struct ht_table), KM_NOSLEEP);
1843 if (buf_hash_table.ht_table == NULL) {
1844 ASSERT(hsize > (1ULL << 8));
1845 hsize >>= 1;
1846 goto retry;
1847 }
1848
1849 hdr_full_cache = kmem_cache_create("arc_buf_hdr_t_full", HDR_FULL_SIZE,
1850 0, hdr_full_cons, hdr_full_dest, hdr_recl, NULL, NULL, 0);
1851 hdr_l2only_cache = kmem_cache_create("arc_buf_hdr_t_l2only",
1852 HDR_L2ONLY_SIZE, 0, hdr_l2only_cons, hdr_l2only_dest, hdr_recl,
1853 NULL, NULL, 0);
1854 buf_cache = kmem_cache_create("arc_buf_t", sizeof (arc_buf_t),
1855 0, buf_cons, buf_dest, NULL, NULL, NULL, 0);
1856
1857 for (i = 0; i < 256; i++)
1858 for (ct = zfs_crc64_table + i, *ct = i, j = 8; j > 0; j--)
1859 *ct = (*ct >> 1) ^ (-(*ct & 1) & ZFS_CRC64_POLY);
1860
1861 for (i = 0; i < hsize; i++) {
1862 mutex_init(&buf_hash_table.ht_table[i].lock,
1863 NULL, MUTEX_DEFAULT, NULL);
1864 }
1865 }
1866
1867 /* wait until krrp releases the buffer */
1868 static inline void
1869 arc_wait_for_krrp(arc_buf_hdr_t *hdr)
1870 {
1871 while (HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_krrp != 0)
1872 cv_wait(&hdr->b_l1hdr.b_cv, HDR_LOCK(hdr));
1873 }
1874
1875 /*
1876 * This is the size that the buf occupies in memory. If the buf is compressed,
1877 * it will correspond to the compressed size. You should use this method of
1878 * getting the buf size unless you explicitly need the logical size.
1879 */
1880 int32_t
1881 arc_buf_size(arc_buf_t *buf)
1882 {
1883 return (ARC_BUF_COMPRESSED(buf) ?
1884 HDR_GET_PSIZE(buf->b_hdr) : HDR_GET_LSIZE(buf->b_hdr));
1885 }
1886
1887 int32_t
1888 arc_buf_lsize(arc_buf_t *buf)
1889 {
1890 return (HDR_GET_LSIZE(buf->b_hdr));
1891 }
1892
1893 enum zio_compress
1894 arc_get_compression(arc_buf_t *buf)
1910 IMPLY(shared, ARC_BUF_SHARED(buf));
1911 IMPLY(shared, ARC_BUF_COMPRESSED(buf) || ARC_BUF_LAST(buf));
1912
1913 /*
1914 * It would be nice to assert arc_can_share() too, but the "hdr isn't
1915 * already being shared" requirement prevents us from doing that.
1916 */
1917
1918 return (shared);
1919 }
1920
1921 /*
1922 * Free the checksum associated with this header. If there is no checksum, this
1923 * is a no-op.
1924 */
1925 static inline void
1926 arc_cksum_free(arc_buf_hdr_t *hdr)
1927 {
1928 ASSERT(HDR_HAS_L1HDR(hdr));
1929 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1930 if (hdr->b_freeze_cksum != NULL) {
1931 kmem_free(hdr->b_freeze_cksum, sizeof (zio_cksum_t));
1932 hdr->b_freeze_cksum = NULL;
1933 }
1934 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1935 }
1936
1937 /*
1938 * Return true iff at least one of the bufs on hdr is not compressed.
1939 */
1940 static boolean_t
1941 arc_hdr_has_uncompressed_buf(arc_buf_hdr_t *hdr)
1942 {
1943 for (arc_buf_t *b = hdr->b_l1hdr.b_buf; b != NULL; b = b->b_next) {
1944 if (!ARC_BUF_COMPRESSED(b)) {
1945 return (B_TRUE);
1946 }
1947 }
1948 return (B_FALSE);
1949 }
1950
1951 /*
1952 * If we've turned on the ZFS_DEBUG_MODIFY flag, verify that the buf's data
1953 * matches the checksum that is stored in the hdr. If there is no checksum,
1954 * or if the buf is compressed, this is a no-op.
1955 */
1956 static void
1957 arc_cksum_verify(arc_buf_t *buf)
1958 {
1959 arc_buf_hdr_t *hdr = buf->b_hdr;
1960 zio_cksum_t zc;
1961
1962 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
1963 return;
1964
1965 if (ARC_BUF_COMPRESSED(buf)) {
1966 ASSERT(hdr->b_freeze_cksum == NULL ||
1967 arc_hdr_has_uncompressed_buf(hdr));
1968 return;
1969 }
1970
1971 ASSERT(HDR_HAS_L1HDR(hdr));
1972
1973 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
1974 if (hdr->b_freeze_cksum == NULL || HDR_IO_ERROR(hdr)) {
1975 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1976 return;
1977 }
1978
1979 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL, &zc);
1980 if (!ZIO_CHECKSUM_EQUAL(*hdr->b_freeze_cksum, zc))
1981 panic("buffer modified while frozen!");
1982 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
1983 }
1984
1985 static boolean_t
1986 arc_cksum_is_equal(arc_buf_hdr_t *hdr, zio_t *zio)
1987 {
1988 enum zio_compress compress = BP_GET_COMPRESS(zio->io_bp);
1989 boolean_t valid_cksum;
1990
1991 ASSERT(!BP_IS_EMBEDDED(zio->io_bp));
1992 VERIFY3U(BP_GET_PSIZE(zio->io_bp), ==, HDR_GET_PSIZE(hdr));
1993
1994 /*
1995 * We rely on the blkptr's checksum to determine if the block
1996 * is valid or not. When compressed arc is enabled, the l2arc
1997 * writes the block to the l2arc just as it appears in the pool.
1998 * This allows us to use the blkptr's checksum to validate the
1999 * data that we just read off of the l2arc without having to store
2000 * a separate checksum in the arc_buf_hdr_t. However, if compressed
2001 * arc is disabled, then the data written to the l2arc is always
2002 * uncompressed and won't match the block as it exists in the main
2003 * pool. When this is the case, we must first compress it if it is
2004 * compressed on the main pool before we can validate the checksum.
2005 */
2006 if (!HDR_COMPRESSION_ENABLED(hdr) && compress != ZIO_COMPRESS_OFF) {
2007 ASSERT3U(HDR_GET_COMPRESS(hdr), ==, ZIO_COMPRESS_OFF);
2008 uint64_t lsize = HDR_GET_LSIZE(hdr);
2009 uint64_t csize;
2010
2011 void *cbuf = zio_buf_alloc(HDR_GET_PSIZE(hdr));
2012 csize = zio_compress_data(compress, zio->io_abd, cbuf, lsize);
2013 abd_t *cdata = abd_get_from_buf(cbuf, HDR_GET_PSIZE(hdr));
2014 abd_take_ownership_of_buf(cdata, B_TRUE);
2015
2016 ASSERT3U(csize, <=, HDR_GET_PSIZE(hdr));
2017 if (csize < HDR_GET_PSIZE(hdr)) {
2018 /*
2019 * Compressed blocks are always a multiple of the
2020 * smallest ashift in the pool. Ideally, we would
2021 * like to round up the csize to the next
2022 * spa_min_ashift but that value may have changed
2023 * since the block was last written. Instead,
2024 * we rely on the fact that the hdr's psize
2025 * was set to the psize of the block when it was
2026 * last written. We set the csize to that value
2027 * and zero out any part that should not contain
2028 * data.
2029 */
2030 abd_zero_off(cdata, csize, HDR_GET_PSIZE(hdr) - csize);
2031 csize = HDR_GET_PSIZE(hdr);
2032 }
2033 zio_push_transform(zio, cdata, csize, HDR_GET_PSIZE(hdr), NULL);
2034 }
2053 return (valid_cksum);
2054 }
2055
2056 /*
2057 * Given a buf full of data, if ZFS_DEBUG_MODIFY is enabled this computes a
2058 * checksum and attaches it to the buf's hdr so that we can ensure that the buf
2059 * isn't modified later on. If buf is compressed or there is already a checksum
2060 * on the hdr, this is a no-op (we only checksum uncompressed bufs).
2061 */
2062 static void
2063 arc_cksum_compute(arc_buf_t *buf)
2064 {
2065 arc_buf_hdr_t *hdr = buf->b_hdr;
2066
2067 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2068 return;
2069
2070 ASSERT(HDR_HAS_L1HDR(hdr));
2071
2072 mutex_enter(&buf->b_hdr->b_l1hdr.b_freeze_lock);
2073 if (hdr->b_freeze_cksum != NULL) {
2074 ASSERT(arc_hdr_has_uncompressed_buf(hdr));
2075 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2076 return;
2077 } else if (ARC_BUF_COMPRESSED(buf)) {
2078 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2079 return;
2080 }
2081
2082 ASSERT(!ARC_BUF_COMPRESSED(buf));
2083 hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t),
2084 KM_SLEEP);
2085 fletcher_2_native(buf->b_data, arc_buf_size(buf), NULL,
2086 hdr->b_freeze_cksum);
2087 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2088 arc_buf_watch(buf);
2089 }
2090
2091 #ifndef _KERNEL
2092 typedef struct procctl {
2093 long cmd;
2094 prwatch_t prwatch;
2095 } procctl_t;
2096 #endif
2097
2098 /* ARGSUSED */
2099 static void
2100 arc_buf_unwatch(arc_buf_t *buf)
2101 {
2102 #ifndef _KERNEL
2103 if (arc_watch) {
2104 int result;
2105 procctl_t ctl;
2106 ctl.cmd = PCWATCH;
2118 arc_buf_watch(arc_buf_t *buf)
2119 {
2120 #ifndef _KERNEL
2121 if (arc_watch) {
2122 int result;
2123 procctl_t ctl;
2124 ctl.cmd = PCWATCH;
2125 ctl.prwatch.pr_vaddr = (uintptr_t)buf->b_data;
2126 ctl.prwatch.pr_size = arc_buf_size(buf);
2127 ctl.prwatch.pr_wflags = WA_WRITE;
2128 result = write(arc_procfd, &ctl, sizeof (ctl));
2129 ASSERT3U(result, ==, sizeof (ctl));
2130 }
2131 #endif
2132 }
2133
2134 static arc_buf_contents_t
2135 arc_buf_type(arc_buf_hdr_t *hdr)
2136 {
2137 arc_buf_contents_t type;
2138
2139 if (HDR_ISTYPE_METADATA(hdr)) {
2140 type = ARC_BUFC_METADATA;
2141 } else if (HDR_ISTYPE_DDT(hdr)) {
2142 type = ARC_BUFC_DDT;
2143 } else {
2144 type = ARC_BUFC_DATA;
2145 }
2146 VERIFY3U(hdr->b_type, ==, type);
2147 return (type);
2148 }
2149
2150 boolean_t
2151 arc_is_metadata(arc_buf_t *buf)
2152 {
2153 return (HDR_ISTYPE_METADATA(buf->b_hdr) != 0);
2154 }
2155
2156 static uint32_t
2157 arc_bufc_to_flags(arc_buf_contents_t type)
2158 {
2159 switch (type) {
2160 case ARC_BUFC_DATA:
2161 /* metadata field is 0 if buffer contains normal data */
2162 return (0);
2163 case ARC_BUFC_METADATA:
2164 return (ARC_FLAG_BUFC_METADATA);
2165 case ARC_BUFC_DDT:
2166 return (ARC_FLAG_BUFC_DDT);
2167 default:
2168 break;
2169 }
2170 panic("undefined ARC buffer type!");
2171 return ((uint32_t)-1);
2172 }
2173
2174 static arc_buf_contents_t
2175 arc_flags_to_bufc(uint32_t flags)
2176 {
2177 if (flags & ARC_FLAG_BUFC_DDT)
2178 return (ARC_BUFC_DDT);
2179 if (flags & ARC_FLAG_BUFC_METADATA)
2180 return (ARC_BUFC_METADATA);
2181 return (ARC_BUFC_DATA);
2182 }
2183
2184 void
2185 arc_buf_thaw(arc_buf_t *buf)
2186 {
2187 arc_buf_hdr_t *hdr = buf->b_hdr;
2188
2189 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
2190 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
2191
2192 arc_cksum_verify(buf);
2193
2194 /*
2195 * Compressed buffers do not manipulate the b_freeze_cksum or
2196 * allocate b_thawed.
2197 */
2198 if (ARC_BUF_COMPRESSED(buf)) {
2199 ASSERT(hdr->b_freeze_cksum == NULL ||
2200 arc_hdr_has_uncompressed_buf(hdr));
2201 return;
2202 }
2203
2204 ASSERT(HDR_HAS_L1HDR(hdr));
2205 arc_cksum_free(hdr);
2206
2207 mutex_enter(&hdr->b_l1hdr.b_freeze_lock);
2208 #ifdef ZFS_DEBUG
2209 if (zfs_flags & ZFS_DEBUG_MODIFY) {
2210 if (hdr->b_l1hdr.b_thawed != NULL)
2211 kmem_free(hdr->b_l1hdr.b_thawed, 1);
2212 hdr->b_l1hdr.b_thawed = kmem_alloc(1, KM_SLEEP);
2213 }
2214 #endif
2215
2216 mutex_exit(&hdr->b_l1hdr.b_freeze_lock);
2217
2218 arc_buf_unwatch(buf);
2219 }
2220
2221 void
2222 arc_buf_freeze(arc_buf_t *buf)
2223 {
2224 arc_buf_hdr_t *hdr = buf->b_hdr;
2225 kmutex_t *hash_lock;
2226
2227 if (!(zfs_flags & ZFS_DEBUG_MODIFY))
2228 return;
2229
2230 if (ARC_BUF_COMPRESSED(buf)) {
2231 ASSERT(hdr->b_freeze_cksum == NULL ||
2232 arc_hdr_has_uncompressed_buf(hdr));
2233 return;
2234 }
2235
2236 hash_lock = HDR_LOCK(hdr);
2237 mutex_enter(hash_lock);
2238
2239 ASSERT(HDR_HAS_L1HDR(hdr));
2240 ASSERT(hdr->b_freeze_cksum != NULL ||
2241 hdr->b_l1hdr.b_state == arc_anon);
2242 arc_cksum_compute(buf);
2243 mutex_exit(hash_lock);
2244 }
2245
2246 /*
2247 * The arc_buf_hdr_t's b_flags should never be modified directly. Instead,
2248 * the following functions should be used to ensure that the flags are
2249 * updated in a thread-safe way. When manipulating the flags either
2250 * the hash_lock must be held or the hdr must be undiscoverable. This
2251 * ensures that we're not racing with any other threads when updating
2252 * the flags.
2253 */
2254 static inline void
2255 arc_hdr_set_flags(arc_buf_hdr_t *hdr, arc_flags_t flags)
2256 {
2257 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2258 hdr->b_flags |= flags;
2259 }
2260
2312 ASSERT(!ARC_BUF_COMPRESSED(buf));
2313
2314 for (arc_buf_t *from = hdr->b_l1hdr.b_buf; from != NULL;
2315 from = from->b_next) {
2316 /* can't use our own data buffer */
2317 if (from == buf) {
2318 continue;
2319 }
2320
2321 if (!ARC_BUF_COMPRESSED(from)) {
2322 bcopy(from->b_data, buf->b_data, arc_buf_size(buf));
2323 copied = B_TRUE;
2324 break;
2325 }
2326 }
2327
2328 /*
2329 * There were no decompressed bufs, so there should not be a
2330 * checksum on the hdr either.
2331 */
2332 EQUIV(!copied, hdr->b_freeze_cksum == NULL);
2333
2334 return (copied);
2335 }
2336
2337 /*
2338 * Given a buf that has a data buffer attached to it, this function will
2339 * efficiently fill the buf with data of the specified compression setting from
2340 * the hdr and update the hdr's b_freeze_cksum if necessary. If the buf and hdr
2341 * are already sharing a data buf, no copy is performed.
2342 *
2343 * If the buf is marked as compressed but uncompressed data was requested, this
2344 * will allocate a new data buffer for the buf, remove that flag, and fill the
2345 * buf with uncompressed data. You can't request a compressed buf on a hdr with
2346 * uncompressed data, and (since we haven't added support for it yet) if you
2347 * want compressed data your buf must already be marked as compressed and have
2348 * the correct-sized data buffer.
2349 */
2350 static int
2351 arc_buf_fill(arc_buf_t *buf, boolean_t compressed)
2352 {
2391 arc_get_data_buf(hdr, HDR_GET_LSIZE(hdr), buf);
2392
2393 /* We increased the size of b_data; update overhead */
2394 ARCSTAT_INCR(arcstat_overhead_size,
2395 HDR_GET_LSIZE(hdr) - HDR_GET_PSIZE(hdr));
2396 }
2397
2398 /*
2399 * Regardless of the buf's previous compression settings, it
2400 * should not be compressed at the end of this function.
2401 */
2402 buf->b_flags &= ~ARC_BUF_FLAG_COMPRESSED;
2403
2404 /*
2405 * Try copying the data from another buf which already has a
2406 * decompressed version. If that's not possible, it's time to
2407 * bite the bullet and decompress the data from the hdr.
2408 */
2409 if (arc_buf_try_copy_decompressed_data(buf)) {
2410 /* Skip byteswapping and checksumming (already done) */
2411 ASSERT3P(hdr->b_freeze_cksum, !=, NULL);
2412 return (0);
2413 } else {
2414 int error = zio_decompress_data(HDR_GET_COMPRESS(hdr),
2415 hdr->b_l1hdr.b_pabd, buf->b_data,
2416 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2417
2418 /*
2419 * Absent hardware errors or software bugs, this should
2420 * be impossible, but log it anyway so we can debug it.
2421 */
2422 if (error != 0) {
2423 zfs_dbgmsg(
2424 "hdr %p, compress %d, psize %d, lsize %d",
2425 hdr, HDR_GET_COMPRESS(hdr),
2426 HDR_GET_PSIZE(hdr), HDR_GET_LSIZE(hdr));
2427 return (SET_ERROR(EIO));
2428 }
2429 }
2430 }
2431
2654
2655 /*
2656 * An L1 header always exists here, since if we're
2657 * moving to some L1-cached state (i.e. not l2c_only or
2658 * anonymous), we realloc the header to add an L1hdr
2659 * beforehand.
2660 */
2661 ASSERT(HDR_HAS_L1HDR(hdr));
2662 multilist_insert(new_state->arcs_list[buftype], hdr);
2663
2664 if (GHOST_STATE(new_state)) {
2665 ASSERT0(bufcnt);
2666 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
2667 update_new = B_TRUE;
2668 }
2669 arc_evictable_space_increment(hdr, new_state);
2670 }
2671 }
2672
2673 ASSERT(!HDR_EMPTY(hdr));
2674 if (new_state == arc_anon && HDR_IN_HASH_TABLE(hdr)) {
2675 arc_wait_for_krrp(hdr);
2676 buf_hash_remove(hdr);
2677 }
2678
2679 /* adjust state sizes (ignore arc_l2c_only) */
2680
2681 if (update_new && new_state != arc_l2c_only) {
2682 ASSERT(HDR_HAS_L1HDR(hdr));
2683 if (GHOST_STATE(new_state)) {
2684 ASSERT0(bufcnt);
2685
2686 /*
2687 * When moving a header to a ghost state, we first
2688 * remove all arc buffers. Thus, we'll have a
2689 * bufcnt of zero, and no arc buffer to use for
2690 * the reference. As a result, we use the arc
2691 * header pointer for the reference.
2692 */
2693 (void) refcount_add_many(&new_state->arcs_size,
2694 HDR_GET_LSIZE(hdr), hdr);
2695 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
2696 } else {
2697 uint32_t buffers = 0;
2770 continue;
2771
2772 (void) refcount_remove_many(
2773 &old_state->arcs_size, arc_buf_size(buf),
2774 buf);
2775 }
2776 ASSERT3U(bufcnt, ==, buffers);
2777 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
2778 (void) refcount_remove_many(
2779 &old_state->arcs_size, arc_hdr_size(hdr), hdr);
2780 }
2781 }
2782
2783 if (HDR_HAS_L1HDR(hdr))
2784 hdr->b_l1hdr.b_state = new_state;
2785
2786 /*
2787 * L2 headers should never be on the L2 state list since they don't
2788 * have L1 headers allocated.
2789 */
2790 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DATA]));
2791 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_METADATA]));
2792 ASSERT(multilist_is_empty(arc_l2c_only->arcs_list[ARC_BUFC_DDT]));
2793 }
2794
2795 void
2796 arc_space_consume(uint64_t space, arc_space_type_t type)
2797 {
2798 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2799
2800 switch (type) {
2801 case ARC_SPACE_DATA:
2802 ARCSTAT_INCR(arcstat_data_size, space);
2803 break;
2804 case ARC_SPACE_META:
2805 ARCSTAT_INCR(arcstat_metadata_size, space);
2806 break;
2807 case ARC_SPACE_DDT:
2808 ARCSTAT_INCR(arcstat_ddt_size, space);
2809 break;
2810 case ARC_SPACE_OTHER:
2811 ARCSTAT_INCR(arcstat_other_size, space);
2812 break;
2813 case ARC_SPACE_HDRS:
2814 ARCSTAT_INCR(arcstat_hdr_size, space);
2815 break;
2816 case ARC_SPACE_L2HDRS:
2817 ARCSTAT_INCR(arcstat_l2_hdr_size, space);
2818 break;
2819 }
2820
2821 if (type != ARC_SPACE_DATA && type != ARC_SPACE_DDT)
2822 ARCSTAT_INCR(arcstat_meta_used, space);
2823
2824 atomic_add_64(&arc_size, space);
2825 }
2826
2827 void
2828 arc_space_return(uint64_t space, arc_space_type_t type)
2829 {
2830 ASSERT(type >= 0 && type < ARC_SPACE_NUMTYPES);
2831
2832 switch (type) {
2833 case ARC_SPACE_DATA:
2834 ARCSTAT_INCR(arcstat_data_size, -space);
2835 break;
2836 case ARC_SPACE_META:
2837 ARCSTAT_INCR(arcstat_metadata_size, -space);
2838 break;
2839 case ARC_SPACE_DDT:
2840 ARCSTAT_INCR(arcstat_ddt_size, -space);
2841 break;
2842 case ARC_SPACE_OTHER:
2843 ARCSTAT_INCR(arcstat_other_size, -space);
2844 break;
2845 case ARC_SPACE_HDRS:
2846 ARCSTAT_INCR(arcstat_hdr_size, -space);
2847 break;
2848 case ARC_SPACE_L2HDRS:
2849 ARCSTAT_INCR(arcstat_l2_hdr_size, -space);
2850 break;
2851 }
2852
2853 if (type != ARC_SPACE_DATA && type != ARC_SPACE_DDT) {
2854 ASSERT(arc_meta_used >= space);
2855 if (arc_meta_max < arc_meta_used)
2856 arc_meta_max = arc_meta_used;
2857 ARCSTAT_INCR(arcstat_meta_used, -space);
2858 }
2859
2860 ASSERT(arc_size >= space);
2861 atomic_add_64(&arc_size, -space);
2862 }
2863
2864 /*
2865 * Given a hdr and a buf, returns whether that buf can share its b_data buffer
2866 * with the hdr's b_pabd.
2867 */
2868 static boolean_t
2869 arc_can_share(arc_buf_hdr_t *hdr, arc_buf_t *buf)
2870 {
2871 /*
2872 * The criteria for sharing a hdr's data are:
2873 * 1. the hdr's compression matches the buf's compression
2874 * 2. the hdr doesn't need to be byteswapped
2875 * 3. the hdr isn't already being shared
2876 * 4. the buf is either compressed or it is the last buf in the hdr list
2877 *
2878 * Criterion #4 maintains the invariant that shared uncompressed
2879 * bufs must be the final buf in the hdr's b_buf list. Reading this, you
2880 * might ask, "if a compressed buf is allocated first, won't that be the
2881 * last thing in the list?", but in that case it's impossible to create
2895 return (buf_compressed == hdr_compressed &&
2896 hdr->b_l1hdr.b_byteswap == DMU_BSWAP_NUMFUNCS &&
2897 !HDR_SHARED_DATA(hdr) &&
2898 (ARC_BUF_LAST(buf) || ARC_BUF_COMPRESSED(buf)));
2899 }
2900
2901 /*
2902 * Allocate a buf for this hdr. If you care about the data that's in the hdr,
2903 * or if you want a compressed buffer, pass those flags in. Returns 0 if the
2904 * copy was made successfully, or an error code otherwise.
2905 */
2906 static int
2907 arc_buf_alloc_impl(arc_buf_hdr_t *hdr, void *tag, boolean_t compressed,
2908 boolean_t fill, arc_buf_t **ret)
2909 {
2910 arc_buf_t *buf;
2911
2912 ASSERT(HDR_HAS_L1HDR(hdr));
2913 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
2914 VERIFY(hdr->b_type == ARC_BUFC_DATA ||
2915 hdr->b_type == ARC_BUFC_METADATA ||
2916 hdr->b_type == ARC_BUFC_DDT);
2917 ASSERT3P(ret, !=, NULL);
2918 ASSERT3P(*ret, ==, NULL);
2919
2920 buf = *ret = kmem_cache_alloc(buf_cache, KM_PUSHPAGE);
2921 buf->b_hdr = hdr;
2922 buf->b_data = NULL;
2923 buf->b_next = hdr->b_l1hdr.b_buf;
2924 buf->b_flags = 0;
2925
2926 add_reference(hdr, tag);
2927
2928 /*
2929 * We're about to change the hdr's b_flags. We must either
2930 * hold the hash_lock or be undiscoverable.
2931 */
2932 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
2933
2934 /*
2935 * Only honor requests for compressed bufs if the hdr is actually
2936 * compressed.
2976 */
2977 if (fill) {
2978 return (arc_buf_fill(buf, ARC_BUF_COMPRESSED(buf) != 0));
2979 }
2980
2981 return (0);
2982 }
2983
2984 static char *arc_onloan_tag = "onloan";
2985
2986 static inline void
2987 arc_loaned_bytes_update(int64_t delta)
2988 {
2989 atomic_add_64(&arc_loaned_bytes, delta);
2990
2991 /* assert that it did not wrap around */
2992 ASSERT3S(atomic_add_64_nv(&arc_loaned_bytes, 0), >=, 0);
2993 }
2994
2995 /*
2996 * Allocates an ARC buf header that's in an evicted & L2-cached state.
2997 * This is used during l2arc reconstruction to make empty ARC buffers
2998 * which circumvent the regular disk->arc->l2arc path and instead come
2999 * into being in the reverse order, i.e. l2arc->arc.
3000 */
3001 static arc_buf_hdr_t *
3002 arc_buf_alloc_l2only(uint64_t load_guid, arc_buf_contents_t type,
3003 l2arc_dev_t *dev, dva_t dva, uint64_t daddr, uint64_t lsize,
3004 uint64_t psize, uint64_t birth, zio_cksum_t cksum, int checksum_type,
3005 enum zio_compress compress, boolean_t arc_compress)
3006 {
3007 arc_buf_hdr_t *hdr;
3008
3009 if (type == ARC_BUFC_DDT && !zfs_arc_segregate_ddt)
3010 type = ARC_BUFC_METADATA;
3011
3012 ASSERT(lsize != 0);
3013 hdr = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
3014 ASSERT(HDR_EMPTY(hdr));
3015 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
3016
3017 hdr->b_spa = load_guid;
3018 hdr->b_type = type;
3019 hdr->b_flags = 0;
3020
3021 if (arc_compress)
3022 arc_hdr_set_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
3023 else
3024 arc_hdr_clear_flags(hdr, ARC_FLAG_COMPRESSED_ARC);
3025
3026 HDR_SET_COMPRESS(hdr, compress);
3027
3028 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L2HDR);
3029 hdr->b_dva = dva;
3030 hdr->b_birth = birth;
3031 if (checksum_type != ZIO_CHECKSUM_OFF) {
3032 hdr->b_freeze_cksum = kmem_alloc(sizeof (zio_cksum_t), KM_SLEEP);
3033 bcopy(&cksum, hdr->b_freeze_cksum, sizeof (cksum));
3034 }
3035
3036 HDR_SET_PSIZE(hdr, psize);
3037 HDR_SET_LSIZE(hdr, lsize);
3038
3039 hdr->b_l2hdr.b_dev = dev;
3040 hdr->b_l2hdr.b_daddr = daddr;
3041
3042 return (hdr);
3043 }
3044
3045 /*
3046 * Loan out an anonymous arc buffer. Loaned buffers are not counted as in
3047 * flight data by arc_tempreserve_space() until they are "returned". Loaned
3048 * buffers must be returned to the arc before they can be used by the DMU or
3049 * freed.
3050 */
3051 arc_buf_t *
3052 arc_loan_buf(spa_t *spa, boolean_t is_metadata, int size)
3053 {
3054 arc_buf_t *buf = arc_alloc_buf(spa, arc_onloan_tag,
3055 is_metadata ? ARC_BUFC_METADATA : ARC_BUFC_DATA, size);
3056
3057 arc_loaned_bytes_update(size);
3058
3059 return (buf);
3060 }
3061
3062 arc_buf_t *
3063 arc_loan_compressed_buf(spa_t *spa, uint64_t psize, uint64_t lsize,
3064 enum zio_compress compression_type)
3065 {
3114 list_insert_head(l2arc_free_on_write, df);
3115 mutex_exit(&l2arc_free_on_write_mtx);
3116 }
3117
3118 static void
3119 arc_hdr_free_on_write(arc_buf_hdr_t *hdr)
3120 {
3121 arc_state_t *state = hdr->b_l1hdr.b_state;
3122 arc_buf_contents_t type = arc_buf_type(hdr);
3123 uint64_t size = arc_hdr_size(hdr);
3124
3125 /* protected by hash lock, if in the hash table */
3126 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
3127 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3128 ASSERT(state != arc_anon && state != arc_l2c_only);
3129
3130 (void) refcount_remove_many(&state->arcs_esize[type],
3131 size, hdr);
3132 }
3133 (void) refcount_remove_many(&state->arcs_size, size, hdr);
3134 if (type == ARC_BUFC_DDT) {
3135 arc_space_return(size, ARC_SPACE_DDT);
3136 } else if (type == ARC_BUFC_METADATA) {
3137 arc_space_return(size, ARC_SPACE_META);
3138 } else {
3139 ASSERT(type == ARC_BUFC_DATA);
3140 arc_space_return(size, ARC_SPACE_DATA);
3141 }
3142
3143 l2arc_free_abd_on_write(hdr->b_l1hdr.b_pabd, size, type);
3144 }
3145
3146 /*
3147 * Share the arc_buf_t's data with the hdr. Whenever we are sharing the
3148 * data buffer, we transfer the refcount ownership to the hdr and update
3149 * the appropriate kstats.
3150 */
3151 static void
3152 arc_share_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3153 {
3154 arc_state_t *state = hdr->b_l1hdr.b_state;
3155
3156 ASSERT(arc_can_share(hdr, buf));
3157 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3158 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)) || HDR_EMPTY(hdr));
3159
3160 /*
3161 * Start sharing the data buffer. We transfer the
3162 * refcount ownership to the hdr since it always owns
3163 * the refcount whenever an arc_buf_t is shared.
3164 */
3165 refcount_transfer_ownership(&state->arcs_size, buf, hdr);
3166 hdr->b_l1hdr.b_pabd = abd_get_from_buf(buf->b_data, arc_buf_size(buf));
3167 abd_take_ownership_of_buf(hdr->b_l1hdr.b_pabd,
3168 !HDR_ISTYPE_DATA(hdr));
3169 arc_hdr_set_flags(hdr, ARC_FLAG_SHARED_DATA);
3170 buf->b_flags |= ARC_BUF_FLAG_SHARED;
3171
3172 /*
3173 * Since we've transferred ownership to the hdr we need
3174 * to increment its compressed and uncompressed kstats and
3175 * decrement the overhead size.
3176 */
3177 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3178 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3179 ARCSTAT_INCR(arcstat_overhead_size, -arc_buf_size(buf));
3180 }
3181
3182 static void
3183 arc_unshare_buf(arc_buf_hdr_t *hdr, arc_buf_t *buf)
3184 {
3185 arc_state_t *state = hdr->b_l1hdr.b_state;
3186
3187 ASSERT(arc_buf_is_shared(buf));
3188 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3341
3342 /* clean up the buf */
3343 buf->b_hdr = NULL;
3344 kmem_cache_free(buf_cache, buf);
3345 }
3346
3347 static void
3348 arc_hdr_alloc_pabd(arc_buf_hdr_t *hdr)
3349 {
3350 ASSERT3U(HDR_GET_LSIZE(hdr), >, 0);
3351 ASSERT(HDR_HAS_L1HDR(hdr));
3352 ASSERT(!HDR_SHARED_DATA(hdr));
3353
3354 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3355 hdr->b_l1hdr.b_pabd = arc_get_data_abd(hdr, arc_hdr_size(hdr), hdr);
3356 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3357 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3358
3359 ARCSTAT_INCR(arcstat_compressed_size, arc_hdr_size(hdr));
3360 ARCSTAT_INCR(arcstat_uncompressed_size, HDR_GET_LSIZE(hdr));
3361 arc_update_hit_stat(hdr, B_TRUE);
3362 }
3363
3364 static void
3365 arc_hdr_free_pabd(arc_buf_hdr_t *hdr)
3366 {
3367 ASSERT(HDR_HAS_L1HDR(hdr));
3368 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
3369
3370 /*
3371 * If the hdr is currently being written to the l2arc then
3372 * we defer freeing the data by adding it to the l2arc_free_on_write
3373 * list. The l2arc will free the data once it's finished
3374 * writing it to the l2arc device.
3375 */
3376 if (HDR_L2_WRITING(hdr)) {
3377 arc_hdr_free_on_write(hdr);
3378 ARCSTAT_BUMP(arcstat_l2_free_on_write);
3379 } else {
3380 arc_free_data_abd(hdr, hdr->b_l1hdr.b_pabd,
3381 arc_hdr_size(hdr), hdr);
3382 }
3383 hdr->b_l1hdr.b_pabd = NULL;
3384 hdr->b_l1hdr.b_byteswap = DMU_BSWAP_NUMFUNCS;
3385
3386 ARCSTAT_INCR(arcstat_compressed_size, -arc_hdr_size(hdr));
3387 ARCSTAT_INCR(arcstat_uncompressed_size, -HDR_GET_LSIZE(hdr));
3388 }
3389
3390 static arc_buf_hdr_t *
3391 arc_hdr_alloc(uint64_t spa, int32_t psize, int32_t lsize,
3392 enum zio_compress compression_type, arc_buf_contents_t type)
3393 {
3394 arc_buf_hdr_t *hdr;
3395
3396 ASSERT3U(lsize, >, 0);
3397
3398 if (type == ARC_BUFC_DDT && !zfs_arc_segregate_ddt)
3399 type = ARC_BUFC_METADATA;
3400 VERIFY(type == ARC_BUFC_DATA || type == ARC_BUFC_METADATA ||
3401 type == ARC_BUFC_DDT);
3402
3403 hdr = kmem_cache_alloc(hdr_full_cache, KM_PUSHPAGE);
3404 ASSERT(HDR_EMPTY(hdr));
3405 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
3406 ASSERT3P(hdr->b_l1hdr.b_thawed, ==, NULL);
3407 HDR_SET_PSIZE(hdr, psize);
3408 HDR_SET_LSIZE(hdr, lsize);
3409 hdr->b_spa = spa;
3410 hdr->b_type = type;
3411 hdr->b_flags = 0;
3412 arc_hdr_set_flags(hdr, arc_bufc_to_flags(type) | ARC_FLAG_HAS_L1HDR);
3413 arc_hdr_set_compress(hdr, compression_type);
3414
3415 hdr->b_l1hdr.b_state = arc_anon;
3416 hdr->b_l1hdr.b_arc_access = 0;
3417 hdr->b_l1hdr.b_bufcnt = 0;
3418 hdr->b_l1hdr.b_buf = NULL;
3419
3420 /*
3421 * Allocate the hdr's buffer. This will contain either
3422 * the compressed or uncompressed data depending on the block
3423 * it references and compressed arc enablement.
3424 */
3425 arc_hdr_alloc_pabd(hdr);
3450
3451 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3452 buf_hash_remove(hdr);
3453
3454 bcopy(hdr, nhdr, HDR_L2ONLY_SIZE);
3455
3456 if (new == hdr_full_cache) {
3457 arc_hdr_set_flags(nhdr, ARC_FLAG_HAS_L1HDR);
3458 /*
3459 * arc_access and arc_change_state need to be aware that a
3460 * header has just come out of L2ARC, so we set its state to
3461 * l2c_only even though it's about to change.
3462 */
3463 nhdr->b_l1hdr.b_state = arc_l2c_only;
3464
3465 /* Verify previous threads set to NULL before freeing */
3466 ASSERT3P(nhdr->b_l1hdr.b_pabd, ==, NULL);
3467 } else {
3468 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3469 ASSERT0(hdr->b_l1hdr.b_bufcnt);
3470 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
3471
3472 /*
3473 * If we've reached here, We must have been called from
3474 * arc_evict_hdr(), as such we should have already been
3475 * removed from any ghost list we were previously on
3476 * (which protects us from racing with arc_evict_state),
3477 * thus no locking is needed during this check.
3478 */
3479 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
3480
3481 /*
3482 * A buffer must not be moved into the arc_l2c_only
3483 * state if it's not finished being written out to the
3484 * l2arc device. Otherwise, the b_l1hdr.b_pabd field
3485 * might try to be accessed, even though it was removed.
3486 */
3487 VERIFY(!HDR_L2_WRITING(hdr));
3488 VERIFY3P(hdr->b_l1hdr.b_pabd, ==, NULL);
3489
3490 #ifdef ZFS_DEBUG
3555 /*
3556 * Allocate a compressed buf in the same manner as arc_alloc_buf. Don't use this
3557 * for bufs containing metadata.
3558 */
3559 arc_buf_t *
3560 arc_alloc_compressed_buf(spa_t *spa, void *tag, uint64_t psize, uint64_t lsize,
3561 enum zio_compress compression_type)
3562 {
3563 ASSERT3U(lsize, >, 0);
3564 ASSERT3U(lsize, >=, psize);
3565 ASSERT(compression_type > ZIO_COMPRESS_OFF);
3566 ASSERT(compression_type < ZIO_COMPRESS_FUNCTIONS);
3567
3568 arc_buf_hdr_t *hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
3569 compression_type, ARC_BUFC_DATA);
3570 ASSERT(!MUTEX_HELD(HDR_LOCK(hdr)));
3571
3572 arc_buf_t *buf = NULL;
3573 VERIFY0(arc_buf_alloc_impl(hdr, tag, B_TRUE, B_FALSE, &buf));
3574 arc_buf_thaw(buf);
3575 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
3576
3577 if (!arc_buf_is_shared(buf)) {
3578 /*
3579 * To ensure that the hdr has the correct data in it if we call
3580 * arc_decompress() on this buf before it's been written to
3581 * disk, it's easiest if we just set up sharing between the
3582 * buf and the hdr.
3583 */
3584 ASSERT(!abd_is_linear(hdr->b_l1hdr.b_pabd));
3585 arc_hdr_free_pabd(hdr);
3586 arc_share_buf(hdr, buf);
3587 }
3588
3589 return (buf);
3590 }
3591
3592 static void
3593 arc_hdr_l2hdr_destroy(arc_buf_hdr_t *hdr)
3594 {
3595 l2arc_buf_hdr_t *l2hdr = &hdr->b_l2hdr;
3596 l2arc_dev_t *dev = l2hdr->b_dev;
3597 uint64_t psize = arc_hdr_size(hdr);
3598
3599 ASSERT(MUTEX_HELD(&dev->l2ad_mtx));
3600 ASSERT(HDR_HAS_L2HDR(hdr));
3601
3602 list_remove(&dev->l2ad_buflist, hdr);
3603
3604 ARCSTAT_INCR(arcstat_l2_psize, -psize);
3605 ARCSTAT_INCR(arcstat_l2_lsize, -HDR_GET_LSIZE(hdr));
3606
3607 /*
3608 * l2ad_vdev can be NULL here if we async evicted it
3609 */
3610 if (dev->l2ad_vdev != NULL)
3611 vdev_space_update(dev->l2ad_vdev, -psize, 0, 0);
3612
3613 (void) refcount_remove_many(&dev->l2ad_alloc, psize, hdr);
3614 arc_hdr_clear_flags(hdr, ARC_FLAG_HAS_L2HDR);
3615 }
3616
3617 static void
3618 arc_hdr_destroy(arc_buf_hdr_t *hdr)
3619 {
3620 if (HDR_HAS_L1HDR(hdr)) {
3621 ASSERT(hdr->b_l1hdr.b_buf == NULL ||
3622 hdr->b_l1hdr.b_bufcnt > 0);
3623 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
3624 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
3625 }
3626 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3627 ASSERT(!HDR_IN_HASH_TABLE(hdr));
3628
3629 if (HDR_HAS_L2HDR(hdr)) {
3630 l2arc_dev_t *dev = hdr->b_l2hdr.b_dev;
3631 boolean_t buflist_held = MUTEX_HELD(&dev->l2ad_mtx);
3632
3633 /* To avoid racing with L2ARC the header needs to be locked */
3634 ASSERT(MUTEX_HELD(HDR_LOCK(hdr)));
3635
3636 if (!buflist_held)
3637 mutex_enter(&dev->l2ad_mtx);
3638
3639 /*
3640 * L2ARC buflist has been held, so we can safety discard
3641 * identity, otherwise L2ARC can lock incorrect mutex
3642 * for the hdr, that will cause a panic. That is possible,
3643 * because a mutex is selected according to identity.
3644 */
3645 if (!HDR_EMPTY(hdr))
3646 buf_discard_identity(hdr);
3647
3648 /*
3649 * Even though we checked this conditional above, we
3650 * need to check this again now that we have the
3651 * l2ad_mtx. This is because we could be racing with
3652 * another thread calling l2arc_evict() which might have
3653 * destroyed this header's L2 portion as we were waiting
3654 * to acquire the l2ad_mtx. If that happens, we don't
3655 * want to re-destroy the header's L2 portion.
3656 */
3657 if (HDR_HAS_L2HDR(hdr))
3658 arc_hdr_l2hdr_destroy(hdr);
3659
3660 if (!buflist_held)
3661 mutex_exit(&dev->l2ad_mtx);
3662 }
3663
3664 if (!HDR_EMPTY(hdr))
3665 buf_discard_identity(hdr);
3666
3667 if (HDR_HAS_L1HDR(hdr)) {
3668 arc_cksum_free(hdr);
3669
3670 while (hdr->b_l1hdr.b_buf != NULL)
3671 arc_buf_destroy_impl(hdr->b_l1hdr.b_buf);
3672
3673 #ifdef ZFS_DEBUG
3674 if (hdr->b_l1hdr.b_thawed != NULL) {
3675 kmem_free(hdr->b_l1hdr.b_thawed, 1);
3676 hdr->b_l1hdr.b_thawed = NULL;
3677 }
3678 #endif
3679
3680 if (hdr->b_l1hdr.b_pabd != NULL) {
3681 arc_hdr_free_pabd(hdr);
3682 }
3683 }
3684
3685 ASSERT3P(hdr->b_hash_next, ==, NULL);
3686 if (HDR_HAS_L1HDR(hdr)) {
3722 * Evict the arc_buf_hdr that is provided as a parameter. The resultant
3723 * state of the header is dependent on it's state prior to entering this
3724 * function. The following transitions are possible:
3725 *
3726 * - arc_mru -> arc_mru_ghost
3727 * - arc_mfu -> arc_mfu_ghost
3728 * - arc_mru_ghost -> arc_l2c_only
3729 * - arc_mru_ghost -> deleted
3730 * - arc_mfu_ghost -> arc_l2c_only
3731 * - arc_mfu_ghost -> deleted
3732 */
3733 static int64_t
3734 arc_evict_hdr(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
3735 {
3736 arc_state_t *evicted_state, *state;
3737 int64_t bytes_evicted = 0;
3738
3739 ASSERT(MUTEX_HELD(hash_lock));
3740 ASSERT(HDR_HAS_L1HDR(hdr));
3741
3742 arc_wait_for_krrp(hdr);
3743
3744 state = hdr->b_l1hdr.b_state;
3745 if (GHOST_STATE(state)) {
3746 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
3747 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
3748
3749 /*
3750 * l2arc_write_buffers() relies on a header's L1 portion
3751 * (i.e. its b_pabd field) during it's write phase.
3752 * Thus, we cannot push a header onto the arc_l2c_only
3753 * state (removing it's L1 piece) until the header is
3754 * done being written to the l2arc.
3755 */
3756 if (HDR_HAS_L2HDR(hdr) && HDR_L2_WRITING(hdr)) {
3757 ARCSTAT_BUMP(arcstat_evict_l2_skip);
3758 return (bytes_evicted);
3759 }
3760
3761 ARCSTAT_BUMP(arcstat_deleted);
3762 bytes_evicted += HDR_GET_LSIZE(hdr);
3763
4112 * prevents us from trying to evict more from a state's list than
4113 * is "evictable", and to skip evicting altogether when passed a
4114 * negative value for "bytes". In contrast, arc_evict_state() will
4115 * evict everything it can, when passed a negative value for "bytes".
4116 */
4117 static uint64_t
4118 arc_adjust_impl(arc_state_t *state, uint64_t spa, int64_t bytes,
4119 arc_buf_contents_t type)
4120 {
4121 int64_t delta;
4122
4123 if (bytes > 0 && refcount_count(&state->arcs_esize[type]) > 0) {
4124 delta = MIN(refcount_count(&state->arcs_esize[type]), bytes);
4125 return (arc_evict_state(state, spa, delta, type));
4126 }
4127
4128 return (0);
4129 }
4130
4131 /*
4132 * Depending on the value of adjust_ddt arg evict either DDT (B_TRUE)
4133 * or metadata (B_TRUE) buffers.
4134 * Evict metadata or DDT buffers from the cache, such that arc_meta_used or
4135 * arc_ddt_size is capped by the arc_meta_limit or arc_ddt_limit tunable.
4136 */
4137 static uint64_t
4138 arc_adjust_meta_or_ddt(boolean_t adjust_ddt)
4139 {
4140 uint64_t total_evicted = 0;
4141 int64_t target, over_limit;
4142 arc_buf_contents_t type;
4143
4144 if (adjust_ddt) {
4145 over_limit = arc_ddt_size - arc_ddt_limit;
4146 type = ARC_BUFC_DDT;
4147 } else {
4148 over_limit = arc_meta_used - arc_meta_limit;
4149 type = ARC_BUFC_METADATA;
4150 }
4151
4152 /*
4153 * If we're over the limit, we want to evict enough
4154 * to get back under the limit. We don't want to
4155 * evict so much that we drop the MRU below arc_p, though. If
4156 * we're over the meta limit more than we're over arc_p, we
4157 * evict some from the MRU here, and some from the MFU below.
4158 */
4159 target = MIN(over_limit,
4160 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4161 refcount_count(&arc_mru->arcs_size) - arc_p));
4162
4163 total_evicted += arc_adjust_impl(arc_mru, 0, target, type);
4164
4165 over_limit = adjust_ddt ? arc_ddt_size - arc_ddt_limit :
4166 arc_meta_used - arc_meta_limit;
4167
4168 /*
4169 * Similar to the above, we want to evict enough bytes to get us
4170 * below the meta limit, but not so much as to drop us below the
4171 * space allotted to the MFU (which is defined as arc_c - arc_p).
4172 */
4173 target = MIN(over_limit,
4174 (int64_t)(refcount_count(&arc_mfu->arcs_size) - (arc_c - arc_p)));
4175
4176 total_evicted += arc_adjust_impl(arc_mfu, 0, target, type);
4177
4178 return (total_evicted);
4179 }
4180
4181 /*
4182 * Return the type of the oldest buffer in the given arc state
4183 *
4184 * This function will select a random sublists of type ARC_BUFC_DATA,
4185 * ARC_BUFC_METADATA, and ARC_BUFC_DDT. The tail of each sublist
4186 * is compared, and the type which contains the "older" buffer will be
4187 * returned.
4188 */
4189 static arc_buf_contents_t
4190 arc_adjust_type(arc_state_t *state)
4191 {
4192 multilist_t *data_ml = state->arcs_list[ARC_BUFC_DATA];
4193 multilist_t *meta_ml = state->arcs_list[ARC_BUFC_METADATA];
4194 multilist_t *ddt_ml = state->arcs_list[ARC_BUFC_DDT];
4195 int data_idx = multilist_get_random_index(data_ml);
4196 int meta_idx = multilist_get_random_index(meta_ml);
4197 int ddt_idx = multilist_get_random_index(ddt_ml);
4198 multilist_sublist_t *data_mls;
4199 multilist_sublist_t *meta_mls;
4200 multilist_sublist_t *ddt_mls;
4201 arc_buf_contents_t type = ARC_BUFC_DATA; /* silence compiler warning */
4202 arc_buf_hdr_t *data_hdr;
4203 arc_buf_hdr_t *meta_hdr;
4204 arc_buf_hdr_t *ddt_hdr;
4205 clock_t oldest;
4206
4207 /*
4208 * We keep the sublist lock until we're finished, to prevent
4209 * the headers from being destroyed via arc_evict_state().
4210 */
4211 data_mls = multilist_sublist_lock(data_ml, data_idx);
4212 meta_mls = multilist_sublist_lock(meta_ml, meta_idx);
4213 ddt_mls = multilist_sublist_lock(ddt_ml, ddt_idx);
4214
4215 /*
4216 * These two loops are to ensure we skip any markers that
4217 * might be at the tail of the lists due to arc_evict_state().
4218 */
4219
4220 for (data_hdr = multilist_sublist_tail(data_mls); data_hdr != NULL;
4221 data_hdr = multilist_sublist_prev(data_mls, data_hdr)) {
4222 if (data_hdr->b_spa != 0)
4223 break;
4224 }
4225
4226 for (meta_hdr = multilist_sublist_tail(meta_mls); meta_hdr != NULL;
4227 meta_hdr = multilist_sublist_prev(meta_mls, meta_hdr)) {
4228 if (meta_hdr->b_spa != 0)
4229 break;
4230 }
4231
4232 for (ddt_hdr = multilist_sublist_tail(ddt_mls); ddt_hdr != NULL;
4233 ddt_hdr = multilist_sublist_prev(ddt_mls, ddt_hdr)) {
4234 if (ddt_hdr->b_spa != 0)
4235 break;
4236 }
4237
4238 if (data_hdr == NULL && meta_hdr == NULL && ddt_hdr == NULL) {
4239 type = ARC_BUFC_DATA;
4240 } else if (data_hdr != NULL && meta_hdr != NULL && ddt_hdr != NULL) {
4241 /* The headers can't be on the sublist without an L1 header */
4242 ASSERT(HDR_HAS_L1HDR(data_hdr));
4243 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4244 ASSERT(HDR_HAS_L1HDR(ddt_hdr));
4245
4246 oldest = data_hdr->b_l1hdr.b_arc_access;
4247 type = ARC_BUFC_DATA;
4248 if (oldest > meta_hdr->b_l1hdr.b_arc_access) {
4249 oldest = meta_hdr->b_l1hdr.b_arc_access;
4250 type = ARC_BUFC_METADATA;
4251 }
4252 if (oldest > ddt_hdr->b_l1hdr.b_arc_access) {
4253 type = ARC_BUFC_DDT;
4254 }
4255 } else if (data_hdr == NULL && ddt_hdr == NULL) {
4256 ASSERT3P(meta_hdr, !=, NULL);
4257 type = ARC_BUFC_METADATA;
4258 } else if (meta_hdr == NULL && ddt_hdr == NULL) {
4259 ASSERT3P(data_hdr, !=, NULL);
4260 type = ARC_BUFC_DATA;
4261 } else if (meta_hdr == NULL && data_hdr == NULL) {
4262 ASSERT3P(ddt_hdr, !=, NULL);
4263 type = ARC_BUFC_DDT;
4264 } else if (data_hdr != NULL && ddt_hdr != NULL) {
4265 ASSERT3P(meta_hdr, ==, NULL);
4266
4267 /* The headers can't be on the sublist without an L1 header */
4268 ASSERT(HDR_HAS_L1HDR(data_hdr));
4269 ASSERT(HDR_HAS_L1HDR(ddt_hdr));
4270
4271 if (data_hdr->b_l1hdr.b_arc_access <
4272 ddt_hdr->b_l1hdr.b_arc_access) {
4273 type = ARC_BUFC_DATA;
4274 } else {
4275 type = ARC_BUFC_DDT;
4276 }
4277 } else if (meta_hdr != NULL && ddt_hdr != NULL) {
4278 ASSERT3P(data_hdr, ==, NULL);
4279
4280 /* The headers can't be on the sublist without an L1 header */
4281 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4282 ASSERT(HDR_HAS_L1HDR(ddt_hdr));
4283
4284 if (meta_hdr->b_l1hdr.b_arc_access <
4285 ddt_hdr->b_l1hdr.b_arc_access) {
4286 type = ARC_BUFC_METADATA;
4287 } else {
4288 type = ARC_BUFC_DDT;
4289 }
4290 } else if (meta_hdr != NULL && data_hdr != NULL) {
4291 ASSERT3P(ddt_hdr, ==, NULL);
4292
4293 /* The headers can't be on the sublist without an L1 header */
4294 ASSERT(HDR_HAS_L1HDR(data_hdr));
4295 ASSERT(HDR_HAS_L1HDR(meta_hdr));
4296
4297 if (data_hdr->b_l1hdr.b_arc_access <
4298 meta_hdr->b_l1hdr.b_arc_access) {
4299 type = ARC_BUFC_DATA;
4300 } else {
4301 type = ARC_BUFC_METADATA;
4302 }
4303 } else {
4304 /* should never get here */
4305 ASSERT(0);
4306 }
4307
4308 multilist_sublist_unlock(ddt_mls);
4309 multilist_sublist_unlock(meta_mls);
4310 multilist_sublist_unlock(data_mls);
4311
4312 return (type);
4313 }
4314
4315 /*
4316 * Evict buffers from the cache, such that arc_size is capped by arc_c.
4317 */
4318 static uint64_t
4319 arc_adjust(void)
4320 {
4321 uint64_t total_evicted = 0;
4322 uint64_t bytes;
4323 int64_t target;
4324
4325 /*
4326 * If we're over arc_meta_limit, we want to correct that before
4327 * potentially evicting data buffers below.
4328 */
4329 total_evicted += arc_adjust_meta_or_ddt(B_FALSE);
4330
4331 /*
4332 * If we're over arc_ddt_limit, we want to correct that before
4333 * potentially evicting data buffers below.
4334 */
4335 total_evicted += arc_adjust_meta_or_ddt(B_TRUE);
4336
4337 /*
4338 * Adjust MRU size
4339 *
4340 * If we're over the target cache size, we want to evict enough
4341 * from the list to get back to our target size. We don't want
4342 * to evict too much from the MRU, such that it drops below
4343 * arc_p. So, if we're over our target cache size more than
4344 * the MRU is over arc_p, we'll evict enough to get back to
4345 * arc_p here, and then evict more from the MFU below.
4346 */
4347 target = MIN((int64_t)(arc_size - arc_c),
4348 (int64_t)(refcount_count(&arc_anon->arcs_size) +
4349 refcount_count(&arc_mru->arcs_size) + arc_meta_used - arc_p));
4350
4351 /*
4352 * If we're below arc_meta_min, always prefer to evict data.
4353 * Otherwise, try to satisfy the requested number of bytes to
4354 * evict from the type which contains older buffers; in an
4355 * effort to keep newer buffers in the cache regardless of their
4356 * type. If we cannot satisfy the number of bytes from this
4357 * type, spill over into the next type.
4358 */
4359 if (arc_adjust_type(arc_mru) == ARC_BUFC_METADATA &&
4360 arc_meta_used > arc_meta_min) {
4361 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4362 total_evicted += bytes;
4363
4364 /*
4365 * If we couldn't evict our target number of bytes from
4366 * metadata, we try to get the rest from data.
4367 */
4368 target -= bytes;
4369
4370 bytes += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4371 total_evicted += bytes;
4372 } else {
4373 bytes = arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DATA);
4374 total_evicted += bytes;
4375
4376 /*
4377 * If we couldn't evict our target number of bytes from
4378 * data, we try to get the rest from metadata.
4379 */
4380 target -= bytes;
4381
4382 bytes += arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_METADATA);
4383 total_evicted += bytes;
4384 }
4385
4386 /*
4387 * If we couldn't evict our target number of bytes from
4388 * data and metadata, we try to get the rest from ddt.
4389 */
4390 target -= bytes;
4391 total_evicted +=
4392 arc_adjust_impl(arc_mru, 0, target, ARC_BUFC_DDT);
4393
4394 /*
4395 * Adjust MFU size
4396 *
4397 * Now that we've tried to evict enough from the MRU to get its
4398 * size back to arc_p, if we're still above the target cache
4399 * size, we evict the rest from the MFU.
4400 */
4401 target = arc_size - arc_c;
4402
4403 if (arc_adjust_type(arc_mfu) == ARC_BUFC_METADATA &&
4404 arc_meta_used > arc_meta_min) {
4405 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4406 total_evicted += bytes;
4407
4408 /*
4409 * If we couldn't evict our target number of bytes from
4410 * metadata, we try to get the rest from data.
4411 */
4412 target -= bytes;
4413
4414 bytes += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4415 total_evicted += bytes;
4416 } else {
4417 bytes = arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DATA);
4418 total_evicted += bytes;
4419
4420 /*
4421 * If we couldn't evict our target number of bytes from
4422 * data, we try to get the rest from data.
4423 */
4424 target -= bytes;
4425
4426 bytes += arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_METADATA);
4427 total_evicted += bytes;
4428 }
4429
4430 /*
4431 * If we couldn't evict our target number of bytes from
4432 * data and metadata, we try to get the rest from ddt.
4433 */
4434 target -= bytes;
4435 total_evicted +=
4436 arc_adjust_impl(arc_mfu, 0, target, ARC_BUFC_DDT);
4437
4438 /*
4439 * Adjust ghost lists
4440 *
4441 * In addition to the above, the ARC also defines target values
4442 * for the ghost lists. The sum of the mru list and mru ghost
4443 * list should never exceed the target size of the cache, and
4444 * the sum of the mru list, mfu list, mru ghost list, and mfu
4445 * ghost list should never exceed twice the target size of the
4446 * cache. The following logic enforces these limits on the ghost
4447 * caches, and evicts from them as needed.
4448 */
4449 target = refcount_count(&arc_mru->arcs_size) +
4450 refcount_count(&arc_mru_ghost->arcs_size) - arc_c;
4451
4452 bytes = arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DATA);
4453 total_evicted += bytes;
4454
4455 target -= bytes;
4456
4457 bytes += arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_METADATA);
4458 total_evicted += bytes;
4459
4460 target -= bytes;
4461
4462 total_evicted +=
4463 arc_adjust_impl(arc_mru_ghost, 0, target, ARC_BUFC_DDT);
4464
4465 /*
4466 * We assume the sum of the mru list and mfu list is less than
4467 * or equal to arc_c (we enforced this above), which means we
4468 * can use the simpler of the two equations below:
4469 *
4470 * mru + mfu + mru ghost + mfu ghost <= 2 * arc_c
4471 * mru ghost + mfu ghost <= arc_c
4472 */
4473 target = refcount_count(&arc_mru_ghost->arcs_size) +
4474 refcount_count(&arc_mfu_ghost->arcs_size) - arc_c;
4475
4476 bytes = arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DATA);
4477 total_evicted += bytes;
4478
4479 target -= bytes;
4480
4481 bytes += arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_METADATA);
4482 total_evicted += bytes;
4483
4484 target -= bytes;
4485
4486 total_evicted +=
4487 arc_adjust_impl(arc_mfu_ghost, 0, target, ARC_BUFC_DDT);
4488
4489 return (total_evicted);
4490 }
4491
4492 typedef struct arc_async_flush_data {
4493 uint64_t aaf_guid;
4494 boolean_t aaf_retry;
4495 } arc_async_flush_data_t;
4496
4497 static taskq_t *arc_flush_taskq;
4498
4499 static void
4500 arc_flush_impl(uint64_t guid, boolean_t retry)
4501 {
4502 arc_buf_contents_t arcs;
4503
4504 for (arcs = ARC_BUFC_DATA; arcs < ARC_BUFC_NUMTYPES; ++arcs) {
4505 (void) arc_flush_state(arc_mru, guid, arcs, retry);
4506 (void) arc_flush_state(arc_mfu, guid, arcs, retry);
4507 (void) arc_flush_state(arc_mru_ghost, guid, arcs, retry);
4508 (void) arc_flush_state(arc_mfu_ghost, guid, arcs, retry);
4509 }
4510 }
4511
4512 static void
4513 arc_flush_task(void *arg)
4514 {
4515 arc_async_flush_data_t *aaf = (arc_async_flush_data_t *)arg;
4516 arc_flush_impl(aaf->aaf_guid, aaf->aaf_retry);
4517 kmem_free(aaf, sizeof (arc_async_flush_data_t));
4518 }
4519
4520 boolean_t zfs_fastflush = B_TRUE;
4521
4522 void
4523 arc_flush(spa_t *spa, boolean_t retry)
4524 {
4525 uint64_t guid = 0;
4526 boolean_t async_flush = (spa != NULL ? zfs_fastflush : FALSE);
4527 arc_async_flush_data_t *aaf = NULL;
4528
4529 /*
4530 * If retry is B_TRUE, a spa must not be specified since we have
4531 * no good way to determine if all of a spa's buffers have been
4532 * evicted from an arc state.
4533 */
4534 ASSERT(!retry || spa == NULL);
4535
4536 if (spa != NULL) {
4537 guid = spa_load_guid(spa);
4538 if (async_flush) {
4539 aaf = kmem_alloc(sizeof (arc_async_flush_data_t),
4540 KM_SLEEP);
4541 aaf->aaf_guid = guid;
4542 aaf->aaf_retry = retry;
4543 }
4544 }
4545
4546 /*
4547 * Try to flush per-spa remaining ARC ghost buffers asynchronously
4548 * while a pool is being closed.
4549 * An ARC buffer is bound to spa only by guid, so buffer can
4550 * exist even when pool has already gone. If asynchronous flushing
4551 * fails we fall back to regular (synchronous) one.
4552 * NOTE: If asynchronous flushing had not yet finished when the pool
4553 * was imported again it wouldn't be a problem, even when guids before
4554 * and after export/import are the same. We can evict only unreferenced
4555 * buffers, other are skipped.
4556 */
4557 if (!async_flush || (taskq_dispatch(arc_flush_taskq, arc_flush_task,
4558 aaf, TQ_NOSLEEP) == NULL)) {
4559 arc_flush_impl(guid, retry);
4560 if (async_flush)
4561 kmem_free(aaf, sizeof (arc_async_flush_data_t));
4562 }
4563 }
4564
4565 void
4566 arc_shrink(int64_t to_free)
4567 {
4568 if (arc_c > arc_c_min) {
4569
4570 if (arc_c > arc_c_min + to_free)
4571 atomic_add_64(&arc_c, -to_free);
4572 else
4573 arc_c = arc_c_min;
4574
4575 atomic_add_64(&arc_p, -(arc_p >> arc_shrink_shift));
4576 if (arc_c > arc_size)
4577 arc_c = MAX(arc_size, arc_c_min);
4578 if (arc_p > arc_c)
4579 arc_p = (arc_c >> 1);
4580 ASSERT(arc_c >= arc_c_min);
4581 ASSERT((int64_t)arc_p >= 0);
4582 }
4583
4584 if (arc_size > arc_c)
4585 (void) arc_adjust();
4586 }
4587
4588 typedef enum free_memory_reason_t {
4589 FMR_UNKNOWN,
4590 FMR_NEEDFREE,
4591 FMR_LOTSFREE,
4592 FMR_SWAPFS_MINFREE,
4593 FMR_PAGES_PP_MAXIMUM,
4594 FMR_HEAP_ARENA,
4595 FMR_ZIO_ARENA,
4596 } free_memory_reason_t;
4597
4598 int64_t last_free_memory;
4599 free_memory_reason_t last_free_reason;
4600
4601 /*
4602 * Additional reserve of pages for pp_reserve.
4603 */
4604 int64_t arc_pages_pp_reserve = 64;
4728 * is under memory pressure and that the arc should adjust accordingly.
4729 */
4730 static boolean_t
4731 arc_reclaim_needed(void)
4732 {
4733 return (arc_available_memory() < 0);
4734 }
4735
4736 static void
4737 arc_kmem_reap_now(void)
4738 {
4739 size_t i;
4740 kmem_cache_t *prev_cache = NULL;
4741 kmem_cache_t *prev_data_cache = NULL;
4742 extern kmem_cache_t *zio_buf_cache[];
4743 extern kmem_cache_t *zio_data_buf_cache[];
4744 extern kmem_cache_t *range_seg_cache;
4745 extern kmem_cache_t *abd_chunk_cache;
4746
4747 #ifdef _KERNEL
4748 if (arc_meta_used >= arc_meta_limit || arc_ddt_size >= arc_ddt_limit) {
4749 /*
4750 * We are exceeding our meta-data or DDT cache limit.
4751 * Purge some DNLC entries to release holds on meta-data/DDT.
4752 */
4753 dnlc_reduce_cache((void *)(uintptr_t)arc_reduce_dnlc_percent);
4754 }
4755 #if defined(__i386)
4756 /*
4757 * Reclaim unused memory from all kmem caches.
4758 */
4759 kmem_reap();
4760 #endif
4761 #endif
4762
4763 /*
4764 * If a kmem reap is already active, don't schedule more. We must
4765 * check for this because kmem_cache_reap_soon() won't actually
4766 * block on the cache being reaped (this is to prevent callers from
4767 * becoming implicitly blocked by a system-wide kmem reap -- which,
4768 * on a system with many, many full magazines, can take minutes).
4769 */
4770 if (kmem_cache_reap_active())
4771 return;
4887 #endif
4888 arc_shrink(to_free);
4889 }
4890 } else if (free_memory < arc_c >> arc_no_grow_shift) {
4891 arc_no_grow = B_TRUE;
4892 } else if (gethrtime() >= growtime) {
4893 arc_no_grow = B_FALSE;
4894 }
4895
4896 mutex_enter(&arc_reclaim_lock);
4897
4898 /*
4899 * If evicted is zero, we couldn't evict anything via
4900 * arc_adjust(). This could be due to hash lock
4901 * collisions, but more likely due to the majority of
4902 * arc buffers being unevictable. Therefore, even if
4903 * arc_size is above arc_c, another pass is unlikely to
4904 * be helpful and could potentially cause us to enter an
4905 * infinite loop.
4906 */
4907 if (arc_size <= arc_c || evicted == 0) {
4908 /*
4909 * We're either no longer overflowing, or we
4910 * can't evict anything more, so we should wake
4911 * up any threads before we go to sleep.
4912 */
4913 cv_broadcast(&arc_reclaim_waiters_cv);
4914
4915 /*
4916 * Block until signaled, or after one second (we
4917 * might need to perform arc_kmem_reap_now()
4918 * even if we aren't being signalled)
4919 */
4920 CALLB_CPR_SAFE_BEGIN(&cpr);
4921 (void) cv_timedwait_hires(&arc_reclaim_thread_cv,
4922 &arc_reclaim_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
4923 CALLB_CPR_SAFE_END(&cpr, &arc_reclaim_lock);
4924 }
4925 }
4926
4927 arc_reclaim_thread_exit = B_FALSE;
4969 delta = MIN(bytes * mult, arc_p);
4970 arc_p = MAX(arc_p_min, arc_p - delta);
4971 }
4972 ASSERT((int64_t)arc_p >= 0);
4973
4974 if (arc_reclaim_needed()) {
4975 cv_signal(&arc_reclaim_thread_cv);
4976 return;
4977 }
4978
4979 if (arc_no_grow)
4980 return;
4981
4982 if (arc_c >= arc_c_max)
4983 return;
4984
4985 /*
4986 * If we're within (2 * maxblocksize) bytes of the target
4987 * cache size, increment the target cache size
4988 */
4989 if (arc_size > arc_c - (2ULL << SPA_MAXBLOCKSHIFT)) {
4990 atomic_add_64(&arc_c, (int64_t)bytes);
4991 if (arc_c > arc_c_max)
4992 arc_c = arc_c_max;
4993 else if (state == arc_anon)
4994 atomic_add_64(&arc_p, (int64_t)bytes);
4995 if (arc_p > arc_c)
4996 arc_p = arc_c;
4997 }
4998 ASSERT((int64_t)arc_p >= 0);
4999 }
5000
5001 /*
5002 * Check if arc_size has grown past our upper threshold, determined by
5003 * zfs_arc_overflow_shift.
5004 */
5005 static boolean_t
5006 arc_is_overflowing(void)
5007 {
5008 /* Always allow at least one block of overflow */
5009 uint64_t overflow = MAX(SPA_MAXBLOCKSIZE,
5010 arc_c >> zfs_arc_overflow_shift);
5011
5012 return (arc_size >= arc_c + overflow);
5013 }
5014
5015 static abd_t *
5016 arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5017 {
5018 arc_buf_contents_t type = arc_buf_type(hdr);
5019
5020 arc_get_data_impl(hdr, size, tag);
5021 if (type == ARC_BUFC_METADATA || type == ARC_BUFC_DDT) {
5022 return (abd_alloc(size, B_TRUE));
5023 } else {
5024 ASSERT(type == ARC_BUFC_DATA);
5025 return (abd_alloc(size, B_FALSE));
5026 }
5027 }
5028
5029 static void *
5030 arc_get_data_buf(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5031 {
5032 arc_buf_contents_t type = arc_buf_type(hdr);
5033
5034 arc_get_data_impl(hdr, size, tag);
5035 if (type == ARC_BUFC_METADATA || type == ARC_BUFC_DDT) {
5036 return (zio_buf_alloc(size));
5037 } else {
5038 ASSERT(type == ARC_BUFC_DATA);
5039 return (zio_data_buf_alloc(size));
5040 }
5041 }
5042
5043 /*
5044 * Allocate a block and return it to the caller. If we are hitting the
5045 * hard limit for the cache size, we must sleep, waiting for the eviction
5046 * thread to catch up. If we're past the target size but below the hard
5047 * limit, we'll only signal the reclaim thread and continue on.
5048 */
5049 static void
5050 arc_get_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5051 {
5052 arc_state_t *state = hdr->b_l1hdr.b_state;
5053 arc_buf_contents_t type = arc_buf_type(hdr);
5054
5055 arc_adapt(size, state);
5074 /*
5075 * Now that we've acquired the lock, we may no longer be
5076 * over the overflow limit, lets check.
5077 *
5078 * We're ignoring the case of spurious wake ups. If that
5079 * were to happen, it'd let this thread consume an ARC
5080 * buffer before it should have (i.e. before we're under
5081 * the overflow limit and were signalled by the reclaim
5082 * thread). As long as that is a rare occurrence, it
5083 * shouldn't cause any harm.
5084 */
5085 if (arc_is_overflowing()) {
5086 cv_signal(&arc_reclaim_thread_cv);
5087 cv_wait(&arc_reclaim_waiters_cv, &arc_reclaim_lock);
5088 }
5089
5090 mutex_exit(&arc_reclaim_lock);
5091 }
5092
5093 VERIFY3U(hdr->b_type, ==, type);
5094 if (type == ARC_BUFC_DDT) {
5095 arc_space_consume(size, ARC_SPACE_DDT);
5096 } else if (type == ARC_BUFC_METADATA) {
5097 arc_space_consume(size, ARC_SPACE_META);
5098 } else {
5099 arc_space_consume(size, ARC_SPACE_DATA);
5100 }
5101
5102 /*
5103 * Update the state size. Note that ghost states have a
5104 * "ghost size" and so don't need to be updated.
5105 */
5106 if (!GHOST_STATE(state)) {
5107
5108 (void) refcount_add_many(&state->arcs_size, size, tag);
5109
5110 /*
5111 * If this is reached via arc_read, the link is
5112 * protected by the hash lock. If reached via
5113 * arc_buf_alloc, the header should not be accessed by
5114 * any other thread. And, if reached via arc_read_done,
5115 * the hash lock will protect it if it's found in the
5116 * hash table; otherwise no other thread should be
5117 * trying to [add|remove]_reference it.
5118 */
5119 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5120 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5121 (void) refcount_add_many(&state->arcs_esize[type],
5122 size, tag);
5123 }
5124
5125 /*
5126 * If we are growing the cache, and we are adding anonymous
5127 * data, and we have outgrown arc_p, update arc_p
5128 */
5129 if (arc_size < arc_c && hdr->b_l1hdr.b_state == arc_anon &&
5130 (refcount_count(&arc_anon->arcs_size) +
5131 refcount_count(&arc_mru->arcs_size) > arc_p))
5132 arc_p = MIN(arc_c, arc_p + size);
5133 }
5134 }
5135
5136 static void
5137 arc_free_data_abd(arc_buf_hdr_t *hdr, abd_t *abd, uint64_t size, void *tag)
5138 {
5139 arc_free_data_impl(hdr, size, tag);
5140 abd_free(abd);
5141 }
5142
5143 static void
5144 arc_free_data_buf(arc_buf_hdr_t *hdr, void *buf, uint64_t size, void *tag)
5145 {
5146 arc_buf_contents_t type = arc_buf_type(hdr);
5147
5148 arc_free_data_impl(hdr, size, tag);
5149 if (type == ARC_BUFC_METADATA || type == ARC_BUFC_DDT) {
5150 zio_buf_free(buf, size);
5151 } else {
5152 ASSERT(type == ARC_BUFC_DATA);
5153 zio_data_buf_free(buf, size);
5154 }
5155 }
5156
5157 /*
5158 * Free the arc data buffer.
5159 */
5160 static void
5161 arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size, void *tag)
5162 {
5163 arc_state_t *state = hdr->b_l1hdr.b_state;
5164 arc_buf_contents_t type = arc_buf_type(hdr);
5165
5166 /* protected by hash lock, if in the hash table */
5167 if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
5168 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5169 ASSERT(state != arc_anon && state != arc_l2c_only);
5170
5171 (void) refcount_remove_many(&state->arcs_esize[type],
5172 size, tag);
5173 }
5174 (void) refcount_remove_many(&state->arcs_size, size, tag);
5175
5176 VERIFY3U(hdr->b_type, ==, type);
5177 if (type == ARC_BUFC_DDT) {
5178 arc_space_return(size, ARC_SPACE_DDT);
5179 } else if (type == ARC_BUFC_METADATA) {
5180 arc_space_return(size, ARC_SPACE_META);
5181 } else {
5182 ASSERT(type == ARC_BUFC_DATA);
5183 arc_space_return(size, ARC_SPACE_DATA);
5184 }
5185 }
5186
5187 /*
5188 * This routine is called whenever a buffer is accessed.
5189 * NOTE: the hash lock is dropped in this function.
5190 */
5191 static void
5192 arc_access(arc_buf_hdr_t *hdr, kmutex_t *hash_lock)
5193 {
5194 clock_t now;
5195
5196 ASSERT(MUTEX_HELD(hash_lock));
5197 ASSERT(HDR_HAS_L1HDR(hdr));
5198
5199 if (hdr->b_l1hdr.b_state == arc_anon) {
5305 }
5306
5307 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5308 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5309 arc_change_state(new_state, hdr, hash_lock);
5310
5311 ARCSTAT_BUMP(arcstat_mfu_ghost_hits);
5312 } else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
5313 /*
5314 * This buffer is on the 2nd Level ARC.
5315 */
5316
5317 hdr->b_l1hdr.b_arc_access = ddi_get_lbolt();
5318 DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
5319 arc_change_state(arc_mfu, hdr, hash_lock);
5320 } else {
5321 ASSERT(!"invalid arc state");
5322 }
5323 }
5324
5325 /*
5326 * This routine is called by dbuf_hold() to update the arc_access() state
5327 * which otherwise would be skipped for entries in the dbuf cache.
5328 */
5329 void
5330 arc_buf_access(arc_buf_t *buf)
5331 {
5332 mutex_enter(&buf->b_evict_lock);
5333 arc_buf_hdr_t *hdr = buf->b_hdr;
5334
5335 /*
5336 * Avoid taking the hash_lock when possible as an optimization.
5337 * The header must be checked again under the hash_lock in order
5338 * to handle the case where it is concurrently being released.
5339 */
5340 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5341 mutex_exit(&buf->b_evict_lock);
5342 return;
5343 }
5344
5345 kmutex_t *hash_lock = HDR_LOCK(hdr);
5346 mutex_enter(hash_lock);
5347
5348 if (hdr->b_l1hdr.b_state == arc_anon || HDR_EMPTY(hdr)) {
5349 mutex_exit(hash_lock);
5350 mutex_exit(&buf->b_evict_lock);
5351 ARCSTAT_BUMP(arcstat_access_skip);
5352 return;
5353 }
5354
5355 mutex_exit(&buf->b_evict_lock);
5356
5357 ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
5358 hdr->b_l1hdr.b_state == arc_mfu);
5359
5360 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5361 arc_access(hdr, hash_lock);
5362 mutex_exit(hash_lock);
5363
5364 ARCSTAT_BUMP(arcstat_hits);
5365 /*
5366 * Upstream used the ARCSTAT_CONDSTAT macro here, but they changed
5367 * the argument format for that macro, which would requie that we
5368 * go and modify all other uses of it. So it's easier to just expand
5369 * this one invocation of the macro to do the right thing.
5370 */
5371 if (!HDR_PREFETCH(hdr)) {
5372 if (!HDR_ISTYPE_METADATA(hdr))
5373 ARCSTAT_BUMP(arcstat_demand_data_hits);
5374 else
5375 ARCSTAT_BUMP(arcstat_demand_metadata_hits);
5376 } else {
5377 if (!HDR_ISTYPE_METADATA(hdr))
5378 ARCSTAT_BUMP(arcstat_prefetch_data_hits);
5379 else
5380 ARCSTAT_BUMP(arcstat_prefetch_metadata_hits);
5381 }
5382 }
5383
5384 /* a generic arc_done_func_t which you can use */
5385 /* ARGSUSED */
5386 void
5387 arc_bcopy_func(zio_t *zio, arc_buf_t *buf, void *arg)
5388 {
5389 if (zio == NULL || zio->io_error == 0)
5390 bcopy(buf->b_data, arg, arc_buf_size(buf));
5391 arc_buf_destroy(buf, arg);
5392 }
5393
5394 /* a generic arc_done_func_t */
5395 void
5396 arc_getbuf_func(zio_t *zio, arc_buf_t *buf, void *arg)
5397 {
5398 arc_buf_t **bufp = arg;
5399 if (zio && zio->io_error) {
5400 arc_buf_destroy(buf, arg);
5401 *bufp = NULL;
5402 } else {
5403 *bufp = buf;
5506 zio->io_error = error;
5507 }
5508 }
5509 hdr->b_l1hdr.b_acb = NULL;
5510 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5511 if (callback_cnt == 0) {
5512 ASSERT(HDR_PREFETCH(hdr));
5513 ASSERT0(hdr->b_l1hdr.b_bufcnt);
5514 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5515 }
5516
5517 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt) ||
5518 callback_list != NULL);
5519
5520 if (no_zio_error) {
5521 arc_hdr_verify(hdr, zio->io_bp);
5522 } else {
5523 arc_hdr_set_flags(hdr, ARC_FLAG_IO_ERROR);
5524 if (hdr->b_l1hdr.b_state != arc_anon)
5525 arc_change_state(arc_anon, hdr, hash_lock);
5526 if (HDR_IN_HASH_TABLE(hdr)) {
5527 if (hash_lock)
5528 arc_wait_for_krrp(hdr);
5529 buf_hash_remove(hdr);
5530 }
5531 freeable = refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
5532 }
5533
5534 /*
5535 * Broadcast before we drop the hash_lock to avoid the possibility
5536 * that the hdr (and hence the cv) might be freed before we get to
5537 * the cv_broadcast().
5538 */
5539 cv_broadcast(&hdr->b_l1hdr.b_cv);
5540
5541 if (hash_lock != NULL) {
5542 mutex_exit(hash_lock);
5543 } else {
5544 /*
5545 * This block was freed while we waited for the read to
5546 * complete. It has been removed from the hash table and
5547 * moved to the anonymous state (so that it won't show up
5548 * in the cache).
5549 */
5550 ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
5553
5554 /* execute each callback and free its structure */
5555 while ((acb = callback_list) != NULL) {
5556 if (acb->acb_done)
5557 acb->acb_done(zio, acb->acb_buf, acb->acb_private);
5558
5559 if (acb->acb_zio_dummy != NULL) {
5560 acb->acb_zio_dummy->io_error = zio->io_error;
5561 zio_nowait(acb->acb_zio_dummy);
5562 }
5563
5564 callback_list = acb->acb_next;
5565 kmem_free(acb, sizeof (arc_callback_t));
5566 }
5567
5568 if (freeable)
5569 arc_hdr_destroy(hdr);
5570 }
5571
5572 /*
5573 * The function to process data from arc by a callback
5574 * The main purpose is to directly copy data from arc to a target buffer
5575 */
5576 int
5577 arc_io_bypass(spa_t *spa, const blkptr_t *bp,
5578 arc_bypass_io_func func, void *arg)
5579 {
5580 arc_buf_hdr_t *hdr;
5581 kmutex_t *hash_lock = NULL;
5582 int error = 0;
5583 uint64_t guid = spa_load_guid(spa);
5584
5585 top:
5586 hdr = buf_hash_find(guid, bp, &hash_lock);
5587 if (hdr && HDR_HAS_L1HDR(hdr) && hdr->b_l1hdr.b_bufcnt > 0 &&
5588 hdr->b_l1hdr.b_buf->b_data) {
5589 if (HDR_IO_IN_PROGRESS(hdr)) {
5590 cv_wait(&hdr->b_l1hdr.b_cv, hash_lock);
5591 mutex_exit(hash_lock);
5592 DTRACE_PROBE(arc_bypass_wait);
5593 goto top;
5594 }
5595
5596 /*
5597 * As the func is an arbitrary callback, which can block, lock
5598 * should be released not to block other threads from
5599 * performing. A counter is used to hold a reference to block
5600 * which are held by krrp.
5601 */
5602
5603 hdr->b_l1hdr.b_krrp++;
5604 mutex_exit(hash_lock);
5605
5606 error = func(hdr->b_l1hdr.b_buf->b_data, hdr->b_lsize, arg);
5607
5608 mutex_enter(hash_lock);
5609 hdr->b_l1hdr.b_krrp--;
5610 cv_broadcast(&hdr->b_l1hdr.b_cv);
5611 mutex_exit(hash_lock);
5612
5613 return (error);
5614 } else {
5615 if (hash_lock)
5616 mutex_exit(hash_lock);
5617 return (ENODATA);
5618 }
5619 }
5620
5621 /*
5622 * "Read" the block at the specified DVA (in bp) via the
5623 * cache. If the block is found in the cache, invoke the provided
5624 * callback immediately and return. Note that the `zio' parameter
5625 * in the callback will be NULL in this case, since no IO was
5626 * required. If the block is not in the cache pass the read request
5627 * on to the spa with a substitute callback function, so that the
5628 * requested block will be added to the cache.
5629 *
5630 * If a read request arrives for a block that has a read in-progress,
5631 * either wait for the in-progress read to complete (and return the
5632 * results); or, if this is a read with a "done" func, add a record
5633 * to the read to invoke the "done" func when the read completes,
5634 * and return; or just return.
5635 *
5636 * arc_read_done() will invoke all the requested "done" functions
5637 * for readers of this block.
5638 */
5639 int
5640 arc_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, arc_done_func_t *done,
5641 void *private, zio_priority_t priority, int zio_flags,
5741 ARCSTAT_BUMP(
5742 arcstat_demand_hit_predictive_prefetch);
5743 arc_hdr_clear_flags(hdr,
5744 ARC_FLAG_PREDICTIVE_PREFETCH);
5745 }
5746 ASSERT(!BP_IS_EMBEDDED(bp) || !BP_IS_HOLE(bp));
5747
5748 /* Get a buf with the desired data in it. */
5749 VERIFY0(arc_buf_alloc_impl(hdr, private,
5750 compressed_read, B_TRUE, &buf));
5751 } else if (*arc_flags & ARC_FLAG_PREFETCH &&
5752 refcount_count(&hdr->b_l1hdr.b_refcnt) == 0) {
5753 arc_hdr_set_flags(hdr, ARC_FLAG_PREFETCH);
5754 }
5755 DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
5756 arc_access(hdr, hash_lock);
5757 if (*arc_flags & ARC_FLAG_L2CACHE)
5758 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
5759 mutex_exit(hash_lock);
5760 ARCSTAT_BUMP(arcstat_hits);
5761 if (HDR_ISTYPE_DDT(hdr))
5762 ARCSTAT_BUMP(arcstat_ddt_hits);
5763 arc_update_hit_stat(hdr, B_TRUE);
5764
5765 if (done)
5766 done(NULL, buf, private);
5767 } else {
5768 uint64_t lsize = BP_GET_LSIZE(bp);
5769 uint64_t psize = BP_GET_PSIZE(bp);
5770 arc_callback_t *acb;
5771 vdev_t *vd = NULL;
5772 uint64_t addr = 0;
5773 boolean_t devw = B_FALSE;
5774 uint64_t size;
5775
5776 if (hdr == NULL) {
5777 /* this block is not in the cache */
5778 arc_buf_hdr_t *exists = NULL;
5779 arc_buf_contents_t type = BP_GET_BUFC_TYPE(bp);
5780 hdr = arc_hdr_alloc(spa_load_guid(spa), psize, lsize,
5781 BP_GET_COMPRESS(bp), type);
5782
5783 if (!BP_IS_EMBEDDED(bp)) {
5784 hdr->b_dva = *BP_IDENTITY(bp);
5785 hdr->b_birth = BP_PHYSICAL_BIRTH(bp);
5786 exists = buf_hash_insert(hdr, &hash_lock);
5787 }
5788 if (exists != NULL) {
5789 /* somebody beat us to the hash insert */
5790 arc_hdr_destroy(hdr);
5791 mutex_exit(hash_lock);
5792 goto top; /* restart the IO request */
5793 }
5794 } else {
5795 /*
5796 * This block is in the ghost cache. If it was L2-only
5797 * (and thus didn't have an L1 hdr), we realloc the
5798 * header to add an L1 hdr.
5799 */
5800 if (!HDR_HAS_L1HDR(hdr)) {
5801 hdr = arc_hdr_realloc(hdr, hdr_l2only_cache,
5802 hdr_full_cache);
5803 }
5804 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
5805 ASSERT(GHOST_STATE(hdr->b_l1hdr.b_state));
5806 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
5807 ASSERT(refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
5808 ASSERT3P(hdr->b_l1hdr.b_buf, ==, NULL);
5809 ASSERT3P(hdr->b_freeze_cksum, ==, NULL);
5810
5811 /*
5812 * This is a delicate dance that we play here.
5813 * This hdr is in the ghost list so we access it
5814 * to move it out of the ghost list before we
5815 * initiate the read. If it's a prefetch then
5816 * it won't have a callback so we'll remove the
5817 * reference that arc_buf_alloc_impl() created. We
5818 * do this after we've called arc_access() to
5819 * avoid hitting an assert in remove_reference().
5820 */
5821 arc_access(hdr, hash_lock);
5822 arc_hdr_alloc_pabd(hdr);
5823 }
5824 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
5825 size = arc_hdr_size(hdr);
5826
5827 /*
5828 * If compression is enabled on the hdr, then will do
5829 * RAW I/O and will store the compressed data in the hdr's
5841 if (BP_GET_LEVEL(bp) > 0)
5842 arc_hdr_set_flags(hdr, ARC_FLAG_INDIRECT);
5843 if (*arc_flags & ARC_FLAG_PREDICTIVE_PREFETCH)
5844 arc_hdr_set_flags(hdr, ARC_FLAG_PREDICTIVE_PREFETCH);
5845 ASSERT(!GHOST_STATE(hdr->b_l1hdr.b_state));
5846
5847 acb = kmem_zalloc(sizeof (arc_callback_t), KM_SLEEP);
5848 acb->acb_done = done;
5849 acb->acb_private = private;
5850 acb->acb_compressed = compressed_read;
5851
5852 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
5853 hdr->b_l1hdr.b_acb = acb;
5854 arc_hdr_set_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
5855
5856 if (HDR_HAS_L2HDR(hdr) &&
5857 (vd = hdr->b_l2hdr.b_dev->l2ad_vdev) != NULL) {
5858 devw = hdr->b_l2hdr.b_dev->l2ad_writing;
5859 addr = hdr->b_l2hdr.b_daddr;
5860 /*
5861 * Lock out device removal.
5862 */
5863 if (vdev_is_dead(vd) ||
5864 !spa_config_tryenter(spa, SCL_L2ARC, vd, RW_READER))
5865 vd = NULL;
5866 }
5867
5868 if (priority == ZIO_PRIORITY_ASYNC_READ)
5869 arc_hdr_set_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5870 else
5871 arc_hdr_clear_flags(hdr, ARC_FLAG_PRIO_ASYNC_READ);
5872
5873 if (hash_lock != NULL)
5874 mutex_exit(hash_lock);
5875
5876 /*
5877 * At this point, we have a level 1 cache miss. Try again in
5878 * L2ARC if possible.
5879 */
5880 ASSERT3U(HDR_GET_LSIZE(hdr), ==, lsize);
5881
5882 DTRACE_PROBE4(arc__miss, arc_buf_hdr_t *, hdr, blkptr_t *, bp,
5883 uint64_t, lsize, zbookmark_phys_t *, zb);
5884 ARCSTAT_BUMP(arcstat_misses);
5885 arc_update_hit_stat(hdr, B_FALSE);
5886
5887 if (vd != NULL && l2arc_ndev != 0 && !(l2arc_norw && devw)) {
5888 /*
5889 * Read from the L2ARC if the following are true:
5890 * 1. The L2ARC vdev was previously cached.
5891 * 2. This buffer still has L2ARC metadata.
5892 * 3. This buffer isn't currently writing to the L2ARC.
5893 * 4. The L2ARC entry wasn't evicted, which may
5894 * also have invalidated the vdev.
5895 * 5. This isn't prefetch and l2arc_noprefetch is set.
5896 */
5897 if (HDR_HAS_L2HDR(hdr) &&
5898 !HDR_L2_WRITING(hdr) && !HDR_L2_EVICTED(hdr) &&
5899 !(l2arc_noprefetch && HDR_PREFETCH(hdr))) {
5900 l2arc_read_callback_t *cb;
5901 abd_t *abd;
5902 uint64_t asize;
5903
5904 DTRACE_PROBE1(l2arc__hit, arc_buf_hdr_t *, hdr);
5905 ARCSTAT_BUMP(arcstat_l2_hits);
5906 if (vdev_type_is_ddt(vd))
5907 ARCSTAT_BUMP(arcstat_l2_ddt_hits);
5908
5909 cb = kmem_zalloc(sizeof (l2arc_read_callback_t),
5910 KM_SLEEP);
5911 cb->l2rcb_hdr = hdr;
5912 cb->l2rcb_bp = *bp;
5913 cb->l2rcb_zb = *zb;
5914 cb->l2rcb_flags = zio_flags;
5915
5916 asize = vdev_psize_to_asize(vd, size);
5917 if (asize != size) {
5918 abd = abd_alloc_for_io(asize,
5919 !HDR_ISTYPE_DATA(hdr));
5920 cb->l2rcb_abd = abd;
5921 } else {
5922 abd = hdr->b_l1hdr.b_pabd;
5923 }
5924
5925 ASSERT(addr >= VDEV_LABEL_START_SIZE &&
5926 addr + asize <= vd->vdev_psize -
5927 VDEV_LABEL_END_SIZE);
5928
5929 /*
5930 * l2arc read. The SCL_L2ARC lock will be
5931 * released by l2arc_read_done().
5932 * Issue a null zio if the underlying buffer
5933 * was squashed to zero size by compression.
5934 */
5935 ASSERT3U(HDR_GET_COMPRESS(hdr), !=,
5936 ZIO_COMPRESS_EMPTY);
5937 rzio = zio_read_phys(pio, vd, addr,
5938 asize, abd,
5939 ZIO_CHECKSUM_OFF,
5940 l2arc_read_done, cb, priority,
5941 zio_flags | ZIO_FLAG_DONT_CACHE |
5942 ZIO_FLAG_CANFAIL |
5943 ZIO_FLAG_DONT_PROPAGATE |
5944 ZIO_FLAG_DONT_RETRY, B_FALSE);
5945 DTRACE_PROBE2(l2arc__read, vdev_t *, vd,
5946 zio_t *, rzio);
5947
5948 ARCSTAT_INCR(arcstat_l2_read_bytes, size);
5949 if (vdev_type_is_ddt(vd))
5950 ARCSTAT_INCR(arcstat_l2_ddt_read_bytes,
5951 size);
5952
5953 if (*arc_flags & ARC_FLAG_NOWAIT) {
5954 zio_nowait(rzio);
5955 return (0);
5956 }
5957
5958 ASSERT(*arc_flags & ARC_FLAG_WAIT);
5959 if (zio_wait(rzio) == 0)
5960 return (0);
5961
5962 /* l2arc read error; goto zio_read() */
5963 } else {
5964 DTRACE_PROBE1(l2arc__miss,
5965 arc_buf_hdr_t *, hdr);
5966 ARCSTAT_BUMP(arcstat_l2_misses);
5967 if (HDR_L2_WRITING(hdr))
5968 ARCSTAT_BUMP(arcstat_l2_rw_clash);
5969 spa_config_exit(spa, SCL_L2ARC, vd);
5970 }
5971 } else {
6203 hdr->b_l1hdr.b_bufcnt -= 1;
6204 arc_cksum_verify(buf);
6205 arc_buf_unwatch(buf);
6206
6207 mutex_exit(hash_lock);
6208
6209 /*
6210 * Allocate a new hdr. The new hdr will contain a b_pabd
6211 * buffer which will be freed in arc_write().
6212 */
6213 nhdr = arc_hdr_alloc(spa, psize, lsize, compress, type);
6214 ASSERT3P(nhdr->b_l1hdr.b_buf, ==, NULL);
6215 ASSERT0(nhdr->b_l1hdr.b_bufcnt);
6216 ASSERT0(refcount_count(&nhdr->b_l1hdr.b_refcnt));
6217 VERIFY3U(nhdr->b_type, ==, type);
6218 ASSERT(!HDR_SHARED_DATA(nhdr));
6219
6220 nhdr->b_l1hdr.b_buf = buf;
6221 nhdr->b_l1hdr.b_bufcnt = 1;
6222 (void) refcount_add(&nhdr->b_l1hdr.b_refcnt, tag);
6223 nhdr->b_l1hdr.b_krrp = 0;
6224
6225 buf->b_hdr = nhdr;
6226
6227 mutex_exit(&buf->b_evict_lock);
6228 (void) refcount_add_many(&arc_anon->arcs_size,
6229 arc_buf_size(buf), buf);
6230 } else {
6231 mutex_exit(&buf->b_evict_lock);
6232 ASSERT(refcount_count(&hdr->b_l1hdr.b_refcnt) == 1);
6233 /* protected by hash lock, or hdr is on arc_anon */
6234 ASSERT(!multilist_link_active(&hdr->b_l1hdr.b_arc_node));
6235 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6236 arc_change_state(arc_anon, hdr, hash_lock);
6237 hdr->b_l1hdr.b_arc_access = 0;
6238 mutex_exit(hash_lock);
6239
6240 buf_discard_identity(hdr);
6241 arc_buf_thaw(buf);
6242 }
6243 }
6244
6415 kmutex_t *hash_lock;
6416
6417 ASSERT3U(zio->io_error, ==, 0);
6418
6419 arc_cksum_verify(buf);
6420
6421 exists = buf_hash_insert(hdr, &hash_lock);
6422 if (exists != NULL) {
6423 /*
6424 * This can only happen if we overwrite for
6425 * sync-to-convergence, because we remove
6426 * buffers from the hash table when we arc_free().
6427 */
6428 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
6429 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6430 panic("bad overwrite, hdr=%p exists=%p",
6431 (void *)hdr, (void *)exists);
6432 ASSERT(refcount_is_zero(
6433 &exists->b_l1hdr.b_refcnt));
6434 arc_change_state(arc_anon, exists, hash_lock);
6435 arc_wait_for_krrp(exists);
6436 arc_hdr_destroy(exists);
6437 mutex_exit(hash_lock);
6438 exists = buf_hash_insert(hdr, &hash_lock);
6439 ASSERT3P(exists, ==, NULL);
6440 } else if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
6441 /* nopwrite */
6442 ASSERT(zio->io_prop.zp_nopwrite);
6443 if (!BP_EQUAL(&zio->io_bp_orig, zio->io_bp))
6444 panic("bad nopwrite, hdr=%p exists=%p",
6445 (void *)hdr, (void *)exists);
6446 } else {
6447 /* Dedup */
6448 ASSERT(hdr->b_l1hdr.b_bufcnt == 1);
6449 ASSERT(hdr->b_l1hdr.b_state == arc_anon);
6450 ASSERT(BP_GET_DEDUP(zio->io_bp));
6451 ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
6452 }
6453 }
6454 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6455 /* if it's not anon, we are doing a scrub */
6456 if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
6457 arc_access(hdr, hash_lock);
6458 mutex_exit(hash_lock);
6459 } else {
6460 arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
6461 }
6462
6463 ASSERT(!refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
6464 callback->awcb_done(zio, buf, callback->awcb_private);
6465
6466 abd_put(zio->io_abd);
6467 kmem_free(callback, sizeof (arc_write_callback_t));
6468 }
6469
6470 zio_t *
6471 arc_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, arc_buf_t *buf,
6472 boolean_t l2arc, const zio_prop_t *zp, arc_done_func_t *ready,
6473 arc_done_func_t *children_ready, arc_done_func_t *physdone,
6474 arc_done_func_t *done, void *private, zio_priority_t priority,
6475 int zio_flags, const zbookmark_phys_t *zb,
6476 const zio_smartcomp_info_t *smartcomp)
6477 {
6478 arc_buf_hdr_t *hdr = buf->b_hdr;
6479 arc_write_callback_t *callback;
6480 zio_t *zio;
6481 zio_prop_t localprop = *zp;
6482
6483 ASSERT3P(ready, !=, NULL);
6484 ASSERT3P(done, !=, NULL);
6485 ASSERT(!HDR_IO_ERROR(hdr));
6486 ASSERT(!HDR_IO_IN_PROGRESS(hdr));
6487 ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
6488 ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
6489 if (l2arc)
6490 arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
6491 if (ARC_BUF_COMPRESSED(buf)) {
6492 /*
6493 * We're writing a pre-compressed buffer. Make the
6494 * compression algorithm requested by the zio_prop_t match
6495 * the pre-compressed buffer's compression algorithm.
6496 */
6517 * the hdr then we need to break that relationship here.
6518 * The hdr will remain with a NULL data pointer and the
6519 * buf will take sole ownership of the block.
6520 */
6521 if (arc_buf_is_shared(buf)) {
6522 arc_unshare_buf(hdr, buf);
6523 } else {
6524 arc_hdr_free_pabd(hdr);
6525 }
6526 VERIFY3P(buf->b_data, !=, NULL);
6527 arc_hdr_set_compress(hdr, ZIO_COMPRESS_OFF);
6528 }
6529 ASSERT(!arc_buf_is_shared(buf));
6530 ASSERT3P(hdr->b_l1hdr.b_pabd, ==, NULL);
6531
6532 zio = zio_write(pio, spa, txg, bp,
6533 abd_get_from_buf(buf->b_data, HDR_GET_LSIZE(hdr)),
6534 HDR_GET_LSIZE(hdr), arc_buf_size(buf), &localprop, arc_write_ready,
6535 (children_ready != NULL) ? arc_write_children_ready : NULL,
6536 arc_write_physdone, arc_write_done, callback,
6537 priority, zio_flags, zb, smartcomp);
6538
6539 return (zio);
6540 }
6541
6542 static int
6543 arc_memory_throttle(uint64_t reserve, uint64_t txg)
6544 {
6545 #ifdef _KERNEL
6546 uint64_t available_memory = ptob(freemem);
6547 static uint64_t page_load = 0;
6548 static uint64_t last_txg = 0;
6549
6550 #if defined(__i386)
6551 available_memory =
6552 MIN(available_memory, vmem_size(heap_arena, VMEM_FREE));
6553 #endif
6554
6555 if (freemem > physmem * arc_lotsfree_percent / 100)
6556 return (0);
6557
6609
6610 anon_size = MAX((int64_t)(refcount_count(&arc_anon->arcs_size) -
6611 arc_loaned_bytes), 0);
6612
6613 /*
6614 * Writes will, almost always, require additional memory allocations
6615 * in order to compress/encrypt/etc the data. We therefore need to
6616 * make sure that there is sufficient available memory for this.
6617 */
6618 error = arc_memory_throttle(reserve, txg);
6619 if (error != 0)
6620 return (error);
6621
6622 /*
6623 * Throttle writes when the amount of dirty data in the cache
6624 * gets too large. We try to keep the cache less than half full
6625 * of dirty blocks so that our sync times don't grow too large.
6626 * Note: if two requests come in concurrently, we might let them
6627 * both succeed, when one of them should fail. Not a huge deal.
6628 */
6629 if (reserve + arc_tempreserve + anon_size > arc_c / 2 &&
6630 anon_size > arc_c / 4) {
6631 DTRACE_PROBE4(arc__tempreserve__space__throttle, uint64_t,
6632 arc_tempreserve, arc_state_t *, arc_anon, uint64_t,
6633 reserve, uint64_t, arc_c);
6634
6635 uint64_t meta_esize =
6636 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
6637 uint64_t data_esize =
6638 refcount_count(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
6639 dprintf("failing, arc_tempreserve=%lluK anon_meta=%lluK "
6640 "anon_data=%lluK tempreserve=%lluK arc_c=%lluK\n",
6641 arc_tempreserve >> 10, meta_esize >> 10,
6642 data_esize >> 10, reserve >> 10, arc_c >> 10);
6643 return (SET_ERROR(ERESTART));
6644 }
6645 atomic_add_64(&arc_tempreserve, reserve);
6646 return (0);
6647 }
6648
6649 static void
6650 arc_kstat_update_state(arc_state_t *state, kstat_named_t *size,
6651 kstat_named_t *evict_data, kstat_named_t *evict_metadata,
6652 kstat_named_t *evict_ddt)
6653 {
6654 size->value.ui64 = refcount_count(&state->arcs_size);
6655 evict_data->value.ui64 =
6656 refcount_count(&state->arcs_esize[ARC_BUFC_DATA]);
6657 evict_metadata->value.ui64 =
6658 refcount_count(&state->arcs_esize[ARC_BUFC_METADATA]);
6659 evict_ddt->value.ui64 =
6660 refcount_count(&state->arcs_esize[ARC_BUFC_DDT]);
6661 }
6662
6663 static int
6664 arc_kstat_update(kstat_t *ksp, int rw)
6665 {
6666 arc_stats_t *as = ksp->ks_data;
6667
6668 if (rw == KSTAT_WRITE) {
6669 return (EACCES);
6670 } else {
6671 arc_kstat_update_state(arc_anon,
6672 &as->arcstat_anon_size,
6673 &as->arcstat_anon_evictable_data,
6674 &as->arcstat_anon_evictable_metadata,
6675 &as->arcstat_anon_evictable_ddt);
6676 arc_kstat_update_state(arc_mru,
6677 &as->arcstat_mru_size,
6678 &as->arcstat_mru_evictable_data,
6679 &as->arcstat_mru_evictable_metadata,
6680 &as->arcstat_mru_evictable_ddt);
6681 arc_kstat_update_state(arc_mru_ghost,
6682 &as->arcstat_mru_ghost_size,
6683 &as->arcstat_mru_ghost_evictable_data,
6684 &as->arcstat_mru_ghost_evictable_metadata,
6685 &as->arcstat_mru_ghost_evictable_ddt);
6686 arc_kstat_update_state(arc_mfu,
6687 &as->arcstat_mfu_size,
6688 &as->arcstat_mfu_evictable_data,
6689 &as->arcstat_mfu_evictable_metadata,
6690 &as->arcstat_mfu_evictable_ddt);
6691 arc_kstat_update_state(arc_mfu_ghost,
6692 &as->arcstat_mfu_ghost_size,
6693 &as->arcstat_mfu_ghost_evictable_data,
6694 &as->arcstat_mfu_ghost_evictable_metadata,
6695 &as->arcstat_mfu_ghost_evictable_ddt);
6696 }
6697
6698 return (0);
6699 }
6700
6701 /*
6702 * This function *must* return indices evenly distributed between all
6703 * sublists of the multilist. This is needed due to how the ARC eviction
6704 * code is laid out; arc_evict_state() assumes ARC buffers are evenly
6705 * distributed between all sublists and uses this assumption when
6706 * deciding which sublist to evict from and how much to evict from it.
6707 */
6708 unsigned int
6709 arc_state_multilist_index_func(multilist_t *ml, void *obj)
6710 {
6711 arc_buf_hdr_t *hdr = obj;
6712
6713 /*
6714 * We rely on b_dva to generate evenly distributed index
6715 * numbers using buf_hash below. So, as an added precaution,
6725 * on insertion, as this index can be recalculated on removal.
6726 *
6727 * Also, the low order bits of the hash value are thought to be
6728 * distributed evenly. Otherwise, in the case that the multilist
6729 * has a power of two number of sublists, each sublists' usage
6730 * would not be evenly distributed.
6731 */
6732 return (buf_hash(hdr->b_spa, &hdr->b_dva, hdr->b_birth) %
6733 multilist_get_num_sublists(ml));
6734 }
6735
6736 static void
6737 arc_state_init(void)
6738 {
6739 arc_anon = &ARC_anon;
6740 arc_mru = &ARC_mru;
6741 arc_mru_ghost = &ARC_mru_ghost;
6742 arc_mfu = &ARC_mfu;
6743 arc_mfu_ghost = &ARC_mfu_ghost;
6744 arc_l2c_only = &ARC_l2c_only;
6745 arc_buf_contents_t arcs;
6746
6747 for (arcs = ARC_BUFC_DATA; arcs < ARC_BUFC_NUMTYPES; ++arcs) {
6748 arc_mru->arcs_list[arcs] =
6749 multilist_create(sizeof (arc_buf_hdr_t),
6750 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6751 arc_state_multilist_index_func);
6752 arc_mru_ghost->arcs_list[arcs] =
6753 multilist_create(sizeof (arc_buf_hdr_t),
6754 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6755 arc_state_multilist_index_func);
6756 arc_mfu->arcs_list[arcs] =
6757 multilist_create(sizeof (arc_buf_hdr_t),
6758 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6759 arc_state_multilist_index_func);
6760 arc_mfu_ghost->arcs_list[arcs] =
6761 multilist_create(sizeof (arc_buf_hdr_t),
6762 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6763 arc_state_multilist_index_func);
6764 arc_l2c_only->arcs_list[arcs] =
6765 multilist_create(sizeof (arc_buf_hdr_t),
6766 offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
6767 arc_state_multilist_index_func);
6768
6769 refcount_create(&arc_anon->arcs_esize[arcs]);
6770 refcount_create(&arc_mru->arcs_esize[arcs]);
6771 refcount_create(&arc_mru_ghost->arcs_esize[arcs]);
6772 refcount_create(&arc_mfu->arcs_esize[arcs]);
6773 refcount_create(&arc_mfu_ghost->arcs_esize[arcs]);
6774 refcount_create(&arc_l2c_only->arcs_esize[arcs]);
6775 }
6776
6777 arc_flush_taskq = taskq_create("arc_flush_tq",
6778 max_ncpus, minclsyspri, 1, zfs_flush_ntasks, TASKQ_DYNAMIC);
6779
6780 refcount_create(&arc_anon->arcs_size);
6781 refcount_create(&arc_mru->arcs_size);
6782 refcount_create(&arc_mru_ghost->arcs_size);
6783 refcount_create(&arc_mfu->arcs_size);
6784 refcount_create(&arc_mfu_ghost->arcs_size);
6785 refcount_create(&arc_l2c_only->arcs_size);
6786 }
6787
6788 static void
6789 arc_state_fini(void)
6790 {
6791 arc_buf_contents_t arcs;
6792
6793 refcount_destroy(&arc_anon->arcs_size);
6794 refcount_destroy(&arc_mru->arcs_size);
6795 refcount_destroy(&arc_mru_ghost->arcs_size);
6796 refcount_destroy(&arc_mfu->arcs_size);
6797 refcount_destroy(&arc_mfu_ghost->arcs_size);
6798 refcount_destroy(&arc_l2c_only->arcs_size);
6799
6800 for (arcs = ARC_BUFC_DATA; arcs < ARC_BUFC_NUMTYPES; ++arcs) {
6801 multilist_destroy(arc_mru->arcs_list[arcs]);
6802 multilist_destroy(arc_mru_ghost->arcs_list[arcs]);
6803 multilist_destroy(arc_mfu->arcs_list[arcs]);
6804 multilist_destroy(arc_mfu_ghost->arcs_list[arcs]);
6805 multilist_destroy(arc_l2c_only->arcs_list[arcs]);
6806
6807 refcount_destroy(&arc_anon->arcs_esize[arcs]);
6808 refcount_destroy(&arc_mru->arcs_esize[arcs]);
6809 refcount_destroy(&arc_mru_ghost->arcs_esize[arcs]);
6810 refcount_destroy(&arc_mfu->arcs_esize[arcs]);
6811 refcount_destroy(&arc_mfu_ghost->arcs_esize[arcs]);
6812 refcount_destroy(&arc_l2c_only->arcs_esize[arcs]);
6813 }
6814 }
6815
6816 uint64_t
6817 arc_max_bytes(void)
6818 {
6819 return (arc_c_max);
6820 }
6821
6822 void
6823 arc_init(void)
6824 {
6825 /*
6826 * allmem is "all memory that we could possibly use".
6827 */
6828 #ifdef _KERNEL
6829 uint64_t allmem = ptob(physmem - swapfs_minfree);
6830 #else
6831 uint64_t allmem = (physmem * PAGESIZE) / 2;
6832 #endif
6833
6853 * small, because it can cause transactions to be larger than
6854 * arc_c, causing arc_tempreserve_space() to fail.
6855 */
6856 #ifndef _KERNEL
6857 arc_c_min = arc_c_max / 2;
6858 #endif
6859
6860 /*
6861 * Allow the tunables to override our calculations if they are
6862 * reasonable (ie. over 64MB)
6863 */
6864 if (zfs_arc_max > 64 << 20 && zfs_arc_max < allmem) {
6865 arc_c_max = zfs_arc_max;
6866 arc_c_min = MIN(arc_c_min, arc_c_max);
6867 }
6868 if (zfs_arc_min > 64 << 20 && zfs_arc_min <= arc_c_max)
6869 arc_c_min = zfs_arc_min;
6870
6871 arc_c = arc_c_max;
6872 arc_p = (arc_c >> 1);
6873 arc_size = 0;
6874
6875 /* limit ddt meta-data to 1/4 of the arc capacity */
6876 arc_ddt_limit = arc_c_max / 4;
6877 /* limit meta-data to 1/4 of the arc capacity */
6878 arc_meta_limit = arc_c_max / 4;
6879
6880 #ifdef _KERNEL
6881 /*
6882 * Metadata is stored in the kernel's heap. Don't let us
6883 * use more than half the heap for the ARC.
6884 */
6885 arc_meta_limit = MIN(arc_meta_limit,
6886 vmem_size(heap_arena, VMEM_ALLOC | VMEM_FREE) / 2);
6887 #endif
6888
6889 /* Allow the tunable to override if it is reasonable */
6890 if (zfs_arc_ddt_limit > 0 && zfs_arc_ddt_limit <= arc_c_max)
6891 arc_ddt_limit = zfs_arc_ddt_limit;
6892 arc_ddt_evict_threshold =
6893 zfs_arc_segregate_ddt ? &arc_ddt_limit : &arc_meta_limit;
6894
6895 /* Allow the tunable to override if it is reasonable */
6896 if (zfs_arc_meta_limit > 0 && zfs_arc_meta_limit <= arc_c_max)
6897 arc_meta_limit = zfs_arc_meta_limit;
6898
6899 if (arc_c_min < arc_meta_limit / 2 && zfs_arc_min == 0)
6900 arc_c_min = arc_meta_limit / 2;
6901
6902 if (zfs_arc_meta_min > 0) {
6903 arc_meta_min = zfs_arc_meta_min;
6904 } else {
6905 arc_meta_min = arc_c_min / 2;
6906 }
6907
6908 if (zfs_arc_grow_retry > 0)
6909 arc_grow_retry = zfs_arc_grow_retry;
6910
6911 if (zfs_arc_shrink_shift > 0)
6912 arc_shrink_shift = zfs_arc_shrink_shift;
6913
6914 /*
6915 * Ensure that arc_no_grow_shift is less than arc_shrink_shift.
6970 /*
6971 * The reclaim thread will set arc_reclaim_thread_exit back to
6972 * B_FALSE when it is finished exiting; we're waiting for that.
6973 */
6974 while (arc_reclaim_thread_exit) {
6975 cv_signal(&arc_reclaim_thread_cv);
6976 cv_wait(&arc_reclaim_thread_cv, &arc_reclaim_lock);
6977 }
6978 mutex_exit(&arc_reclaim_lock);
6979
6980 /* Use B_TRUE to ensure *all* buffers are evicted */
6981 arc_flush(NULL, B_TRUE);
6982
6983 arc_dead = B_TRUE;
6984
6985 if (arc_ksp != NULL) {
6986 kstat_delete(arc_ksp);
6987 arc_ksp = NULL;
6988 }
6989
6990 taskq_destroy(arc_flush_taskq);
6991
6992 mutex_destroy(&arc_reclaim_lock);
6993 cv_destroy(&arc_reclaim_thread_cv);
6994 cv_destroy(&arc_reclaim_waiters_cv);
6995
6996 arc_state_fini();
6997 buf_fini();
6998
6999 ASSERT0(arc_loaned_bytes);
7000 }
7001
7002 /*
7003 * Level 2 ARC
7004 *
7005 * The level 2 ARC (L2ARC) is a cache layer in-between main memory and disk.
7006 * It uses dedicated storage devices to hold cached data, which are populated
7007 * using large infrequent writes. The main role of this cache is to boost
7008 * the performance of random read workloads. The intended L2ARC devices
7009 * include short-stroked disks, solid state disks, and other media with
7010 * substantially faster read latency than disk.
7011 *
7125 * l2arc_noprefetch skip caching prefetched buffers
7126 * l2arc_headroom number of max device writes to precache
7127 * l2arc_headroom_boost when we find compressed buffers during ARC
7128 * scanning, we multiply headroom by this
7129 * percentage factor for the next scan cycle,
7130 * since more compressed buffers are likely to
7131 * be present
7132 * l2arc_feed_secs seconds between L2ARC writing
7133 *
7134 * Tunables may be removed or added as future performance improvements are
7135 * integrated, and also may become zpool properties.
7136 *
7137 * There are three key functions that control how the L2ARC warms up:
7138 *
7139 * l2arc_write_eligible() check if a buffer is eligible to cache
7140 * l2arc_write_size() calculate how much to write
7141 * l2arc_write_interval() calculate sleep delay between writes
7142 *
7143 * These three functions determine what to write, how much, and how quickly
7144 * to send writes.
7145 *
7146 * L2ARC persistency:
7147 *
7148 * When writing buffers to L2ARC, we periodically add some metadata to
7149 * make sure we can pick them up after reboot, thus dramatically reducing
7150 * the impact that any downtime has on the performance of storage systems
7151 * with large caches.
7152 *
7153 * The implementation works fairly simply by integrating the following two
7154 * modifications:
7155 *
7156 * *) Every now and then we mix in a piece of metadata (called a log block)
7157 * into the L2ARC write. This allows us to understand what's been written,
7158 * so that we can rebuild the arc_buf_hdr_t structures of the main ARC
7159 * buffers. The log block also includes a "2-back-reference" pointer to
7160 * he second-to-previous block, forming a back-linked list of blocks on
7161 * the L2ARC device.
7162 *
7163 * *) We reserve SPA_MINBLOCKSIZE of space at the start of each L2ARC device
7164 * for our header bookkeeping purposes. This contains a device header,
7165 * which contains our top-level reference structures. We update it each
7166 * time we write a new log block, so that we're able to locate it in the
7167 * L2ARC device. If this write results in an inconsistent device header
7168 * (e.g. due to power failure), we detect this by verifying the header's
7169 * checksum and simply drop the entries from L2ARC.
7170 *
7171 * Implementation diagram:
7172 *
7173 * +=== L2ARC device (not to scale) ======================================+
7174 * | ___two newest log block pointers__.__________ |
7175 * | / \1 back \latest |
7176 * |.____/_. V V |
7177 * ||L2 dev|....|lb |bufs |lb |bufs |lb |bufs |lb |bufs |lb |---(empty)---|
7178 * || hdr| ^ /^ /^ / / |
7179 * |+------+ ...--\-------/ \-----/--\------/ / |
7180 * | \--------------/ \--------------/ |
7181 * +======================================================================+
7182 *
7183 * As can be seen on the diagram, rather than using a simple linked list,
7184 * we use a pair of linked lists with alternating elements. This is a
7185 * performance enhancement due to the fact that we only find out of the
7186 * address of the next log block access once the current block has been
7187 * completely read in. Obviously, this hurts performance, because we'd be
7188 * keeping the device's I/O queue at only a 1 operation deep, thus
7189 * incurring a large amount of I/O round-trip latency. Having two lists
7190 * allows us to "prefetch" two log blocks ahead of where we are currently
7191 * rebuilding L2ARC buffers.
7192 *
7193 * On-device data structures:
7194 *
7195 * L2ARC device header: l2arc_dev_hdr_phys_t
7196 * L2ARC log block: l2arc_log_blk_phys_t
7197 *
7198 * L2ARC reconstruction:
7199 *
7200 * When writing data, we simply write in the standard rotary fashion,
7201 * evicting buffers as we go and simply writing new data over them (writing
7202 * a new log block every now and then). This obviously means that once we
7203 * loop around the end of the device, we will start cutting into an already
7204 * committed log block (and its referenced data buffers), like so:
7205 *
7206 * current write head__ __old tail
7207 * \ /
7208 * V V
7209 * <--|bufs |lb |bufs |lb | |bufs |lb |bufs |lb |-->
7210 * ^ ^^^^^^^^^___________________________________
7211 * | \
7212 * <<nextwrite>> may overwrite this blk and/or its bufs --'
7213 *
7214 * When importing the pool, we detect this situation and use it to stop
7215 * our scanning process (see l2arc_rebuild).
7216 *
7217 * There is one significant caveat to consider when rebuilding ARC contents
7218 * from an L2ARC device: what about invalidated buffers? Given the above
7219 * construction, we cannot update blocks which we've already written to amend
7220 * them to remove buffers which were invalidated. Thus, during reconstruction,
7221 * we might be populating the cache with buffers for data that's not on the
7222 * main pool anymore, or may have been overwritten!
7223 *
7224 * As it turns out, this isn't a problem. Every arc_read request includes
7225 * both the DVA and, crucially, the birth TXG of the BP the caller is
7226 * looking for. So even if the cache were populated by completely rotten
7227 * blocks for data that had been long deleted and/or overwritten, we'll
7228 * never actually return bad data from the cache, since the DVA with the
7229 * birth TXG uniquely identify a block in space and time - once created,
7230 * a block is immutable on disk. The worst thing we have done is wasted
7231 * some time and memory at l2arc rebuild to reconstruct outdated ARC
7232 * entries that will get dropped from the l2arc as it is being updated
7233 * with new blocks.
7234 */
7235
7236 static boolean_t
7237 l2arc_write_eligible(uint64_t spa_guid, arc_buf_hdr_t *hdr)
7238 {
7239 /*
7240 * A buffer is *not* eligible for the L2ARC if it:
7241 * 1. belongs to a different spa.
7242 * 2. is already cached on the L2ARC.
7243 * 3. has an I/O in progress (it may be an incomplete read).
7244 * 4. is flagged not eligible (zfs property).
7245 */
7246 if (hdr->b_spa != spa_guid || HDR_HAS_L2HDR(hdr) ||
7247 HDR_IO_IN_PROGRESS(hdr) || !HDR_L2CACHE(hdr))
7248 return (B_FALSE);
7249
7250 return (B_TRUE);
7251 }
7252
7253 static uint64_t
7279 {
7280 clock_t interval, next, now;
7281
7282 /*
7283 * If the ARC lists are busy, increase our write rate; if the
7284 * lists are stale, idle back. This is achieved by checking
7285 * how much we previously wrote - if it was more than half of
7286 * what we wanted, schedule the next write much sooner.
7287 */
7288 if (l2arc_feed_again && wrote > (wanted / 2))
7289 interval = (hz * l2arc_feed_min_ms) / 1000;
7290 else
7291 interval = hz * l2arc_feed_secs;
7292
7293 now = ddi_get_lbolt();
7294 next = MAX(now, MIN(now + interval, began + interval));
7295
7296 return (next);
7297 }
7298
7299 typedef enum l2ad_feed {
7300 L2ARC_FEED_ALL = 1,
7301 L2ARC_FEED_DDT_DEV,
7302 L2ARC_FEED_NON_DDT_DEV,
7303 } l2ad_feed_t;
7304
7305 /*
7306 * Cycle through L2ARC devices. This is how L2ARC load balances.
7307 * If a device is returned, this also returns holding the spa config lock.
7308 */
7309 static l2arc_dev_t *
7310 l2arc_dev_get_next(l2ad_feed_t feed_type)
7311 {
7312 l2arc_dev_t *start = NULL, *next = NULL;
7313
7314 /*
7315 * Lock out the removal of spas (spa_namespace_lock), then removal
7316 * of cache devices (l2arc_dev_mtx). Once a device has been selected,
7317 * both locks will be dropped and a spa config lock held instead.
7318 */
7319 mutex_enter(&spa_namespace_lock);
7320 mutex_enter(&l2arc_dev_mtx);
7321
7322 /* if there are no vdevs, there is nothing to do */
7323 if (l2arc_ndev == 0)
7324 goto out;
7325
7326 if (feed_type == L2ARC_FEED_DDT_DEV)
7327 next = l2arc_ddt_dev_last;
7328 else
7329 next = l2arc_dev_last;
7330
7331 /* figure out what the next device we look at should be */
7332 if (next == NULL)
7333 next = list_head(l2arc_dev_list);
7334 else if (list_next(l2arc_dev_list, next) == NULL)
7335 next = list_head(l2arc_dev_list);
7336 else
7337 next = list_next(l2arc_dev_list, next);
7338 ASSERT(next);
7339
7340 /* loop through L2ARC devs looking for the one we need */
7341 /* LINTED(E_CONSTANT_CONDITION) */
7342 while (1) {
7343 if (next == NULL) /* reached list end, start from beginning */
7344 next = list_head(l2arc_dev_list);
7345
7346 if (start == NULL) { /* save starting dev */
7347 start = next;
7348 } else if (start == next) { /* full loop completed - stop now */
7349 next = NULL;
7350 if (feed_type == L2ARC_FEED_DDT_DEV) {
7351 l2arc_ddt_dev_last = NULL;
7352 goto out;
7353 } else {
7354 break;
7355 }
7356 }
7357
7358 if (!vdev_is_dead(next->l2ad_vdev) && !next->l2ad_rebuild) {
7359 if (feed_type == L2ARC_FEED_DDT_DEV) {
7360 if (vdev_type_is_ddt(next->l2ad_vdev)) {
7361 l2arc_ddt_dev_last = next;
7362 goto out;
7363 }
7364 } else if (feed_type == L2ARC_FEED_NON_DDT_DEV) {
7365 if (!vdev_type_is_ddt(next->l2ad_vdev)) {
7366 break;
7367 }
7368 } else {
7369 ASSERT(feed_type == L2ARC_FEED_ALL);
7370 break;
7371 }
7372 }
7373 next = list_next(l2arc_dev_list, next);
7374 }
7375 l2arc_dev_last = next;
7376
7377 out:
7378 mutex_exit(&l2arc_dev_mtx);
7379
7380 /*
7381 * Grab the config lock to prevent the 'next' device from being
7382 * removed while we are writing to it.
7383 */
7384 if (next != NULL)
7385 spa_config_enter(next->l2ad_spa, SCL_L2ARC, next, RW_READER);
7386 mutex_exit(&spa_namespace_lock);
7387
7388 return (next);
7389 }
7390
7391 /*
7392 * Free buffers that were tagged for destruction.
7393 */
7394 static void
7407 list_remove(buflist, df);
7408 kmem_free(df, sizeof (l2arc_data_free_t));
7409 }
7410
7411 mutex_exit(&l2arc_free_on_write_mtx);
7412 }
7413
7414 /*
7415 * A write to a cache device has completed. Update all headers to allow
7416 * reads from these buffers to begin.
7417 */
7418 static void
7419 l2arc_write_done(zio_t *zio)
7420 {
7421 l2arc_write_callback_t *cb;
7422 l2arc_dev_t *dev;
7423 list_t *buflist;
7424 arc_buf_hdr_t *head, *hdr, *hdr_prev;
7425 kmutex_t *hash_lock;
7426 int64_t bytes_dropped = 0;
7427 l2arc_log_blk_buf_t *lb_buf;
7428
7429 cb = zio->io_private;
7430 ASSERT3P(cb, !=, NULL);
7431 dev = cb->l2wcb_dev;
7432 ASSERT3P(dev, !=, NULL);
7433 head = cb->l2wcb_head;
7434 ASSERT3P(head, !=, NULL);
7435 buflist = &dev->l2ad_buflist;
7436 ASSERT3P(buflist, !=, NULL);
7437 DTRACE_PROBE2(l2arc__iodone, zio_t *, zio,
7438 l2arc_write_callback_t *, cb);
7439
7440 if (zio->io_error != 0)
7441 ARCSTAT_BUMP(arcstat_l2_writes_error);
7442
7443 /*
7444 * All writes completed, or an error was hit.
7445 */
7446 top:
7447 mutex_enter(&dev->l2ad_mtx);
7504 bytes_dropped += arc_hdr_size(hdr);
7505 (void) refcount_remove_many(&dev->l2ad_alloc,
7506 arc_hdr_size(hdr), hdr);
7507 }
7508
7509 /*
7510 * Allow ARC to begin reads and ghost list evictions to
7511 * this L2ARC entry.
7512 */
7513 arc_hdr_clear_flags(hdr, ARC_FLAG_L2_WRITING);
7514
7515 mutex_exit(hash_lock);
7516 }
7517
7518 atomic_inc_64(&l2arc_writes_done);
7519 list_remove(buflist, head);
7520 ASSERT(!HDR_HAS_L1HDR(head));
7521 kmem_cache_free(hdr_l2only_cache, head);
7522 mutex_exit(&dev->l2ad_mtx);
7523
7524 ASSERT(dev->l2ad_vdev != NULL);
7525 vdev_space_update(dev->l2ad_vdev, -bytes_dropped, 0, 0);
7526
7527 l2arc_do_free_on_write();
7528
7529 while ((lb_buf = list_remove_tail(&cb->l2wcb_log_blk_buflist)) != NULL)
7530 kmem_free(lb_buf, sizeof (*lb_buf));
7531 list_destroy(&cb->l2wcb_log_blk_buflist);
7532 kmem_free(cb, sizeof (l2arc_write_callback_t));
7533 }
7534
7535 /*
7536 * A read to a cache device completed. Validate buffer contents before
7537 * handing over to the regular ARC routines.
7538 */
7539 static void
7540 l2arc_read_done(zio_t *zio)
7541 {
7542 l2arc_read_callback_t *cb;
7543 arc_buf_hdr_t *hdr;
7544 kmutex_t *hash_lock;
7545 boolean_t valid_cksum;
7546
7547 ASSERT3P(zio->io_vd, !=, NULL);
7548 ASSERT(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE);
7549
7550 spa_config_exit(zio->io_spa, SCL_L2ARC, zio->io_vd);
7551
7618 * storage now. If there *is* a waiter, the caller must
7619 * issue the i/o in a context where it's OK to block.
7620 */
7621 if (zio->io_waiter == NULL) {
7622 zio_t *pio = zio_unique_parent(zio);
7623
7624 ASSERT(!pio || pio->io_child_type == ZIO_CHILD_LOGICAL);
7625
7626 zio_nowait(zio_read(pio, zio->io_spa, zio->io_bp,
7627 hdr->b_l1hdr.b_pabd, zio->io_size, arc_read_done,
7628 hdr, zio->io_priority, cb->l2rcb_flags,
7629 &cb->l2rcb_zb));
7630 }
7631 }
7632
7633 kmem_free(cb, sizeof (l2arc_read_callback_t));
7634 }
7635
7636 /*
7637 * This is the list priority from which the L2ARC will search for pages to
7638 * cache. This is used within loops to cycle through lists in the
7639 * desired order. This order can have a significant effect on cache
7640 * performance.
7641 *
7642 * Currently the ddt lists are hit first (MFU then MRU),
7643 * followed by metadata then by the data lists.
7644 * This function returns a locked list, and also returns the lock pointer.
7645 */
7646 static multilist_sublist_t *
7647 l2arc_sublist_lock(enum l2arc_priorities prio)
7648 {
7649 multilist_t *ml = NULL;
7650 unsigned int idx;
7651
7652 ASSERT(prio >= PRIORITY_MFU_DDT);
7653 ASSERT(prio < PRIORITY_NUMTYPES);
7654
7655 switch (prio) {
7656 case PRIORITY_MFU_DDT:
7657 ml = arc_mfu->arcs_list[ARC_BUFC_DDT];
7658 break;
7659 case PRIORITY_MRU_DDT:
7660 ml = arc_mru->arcs_list[ARC_BUFC_DDT];
7661 break;
7662 case PRIORITY_MFU_META:
7663 ml = arc_mfu->arcs_list[ARC_BUFC_METADATA];
7664 break;
7665 case PRIORITY_MRU_META:
7666 ml = arc_mru->arcs_list[ARC_BUFC_METADATA];
7667 break;
7668 case PRIORITY_MFU_DATA:
7669 ml = arc_mfu->arcs_list[ARC_BUFC_DATA];
7670 break;
7671 case PRIORITY_MRU_DATA:
7672 ml = arc_mru->arcs_list[ARC_BUFC_DATA];
7673 break;
7674 }
7675
7676 /*
7677 * Return a randomly-selected sublist. This is acceptable
7678 * because the caller feeds only a little bit of data for each
7679 * call (8MB). Subsequent calls will result in different
7680 * sublists being selected.
7681 */
7682 idx = multilist_get_random_index(ml);
7683 return (multilist_sublist_lock(ml, idx));
7684 }
7685
7686 /*
7687 * Calculates the maximum overhead of L2ARC metadata log blocks for a given
7688 * L2ARC write size. l2arc_evict and l2arc_write_buffers need to include this
7689 * overhead in processing to make sure there is enough headroom available
7690 * when writing buffers.
7691 */
7692 static inline uint64_t
7693 l2arc_log_blk_overhead(uint64_t write_sz)
7694 {
7695 return ((write_sz / SPA_MINBLOCKSIZE / L2ARC_LOG_BLK_ENTRIES) + 1) *
7696 L2ARC_LOG_BLK_SIZE;
7697 }
7698
7699 /*
7700 * Evict buffers from the device write hand to the distance specified in
7701 * bytes. This distance may span populated buffers, it may span nothing.
7702 * This is clearing a region on the L2ARC device ready for writing.
7703 * If the 'all' boolean is set, every buffer is evicted.
7704 */
7705 static void
7706 l2arc_evict_impl(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7707 {
7708 list_t *buflist;
7709 arc_buf_hdr_t *hdr, *hdr_prev;
7710 kmutex_t *hash_lock;
7711 uint64_t taddr;
7712
7713 buflist = &dev->l2ad_buflist;
7714
7715 if (!all && dev->l2ad_first) {
7716 /*
7717 * This is the first sweep through the device. There is
7718 * nothing to evict.
7719 */
7720 return;
7721 }
7722
7723 /*
7724 * We need to add in the worst case scenario of log block overhead.
7725 */
7726 distance += l2arc_log_blk_overhead(distance);
7727 if (dev->l2ad_hand >= (dev->l2ad_end - (2 * distance))) {
7728 /*
7729 * When nearing the end of the device, evict to the end
7730 * before the device write hand jumps to the start.
7731 */
7732 taddr = dev->l2ad_end;
7733 } else {
7734 taddr = dev->l2ad_hand + distance;
7735 }
7736 DTRACE_PROBE4(l2arc__evict, l2arc_dev_t *, dev, list_t *, buflist,
7737 uint64_t, taddr, boolean_t, all);
7738
7739 top:
7740 mutex_enter(&dev->l2ad_mtx);
7741 for (hdr = list_tail(buflist); hdr; hdr = hdr_prev) {
7742 hdr_prev = list_prev(buflist, hdr);
7743
7744 hash_lock = HDR_LOCK(hdr);
7745
7746 /*
7790 } else {
7791 ASSERT(hdr->b_l1hdr.b_state != arc_l2c_only);
7792 ARCSTAT_BUMP(arcstat_l2_evict_l1cached);
7793 /*
7794 * Invalidate issued or about to be issued
7795 * reads, since we may be about to write
7796 * over this location.
7797 */
7798 if (HDR_L2_READING(hdr)) {
7799 ARCSTAT_BUMP(arcstat_l2_evict_reading);
7800 arc_hdr_set_flags(hdr, ARC_FLAG_L2_EVICTED);
7801 }
7802
7803 arc_hdr_l2hdr_destroy(hdr);
7804 }
7805 mutex_exit(hash_lock);
7806 }
7807 mutex_exit(&dev->l2ad_mtx);
7808 }
7809
7810 static void
7811 l2arc_evict_task(void *arg)
7812 {
7813 l2arc_dev_t *dev = arg;
7814 ASSERT(dev);
7815
7816 /*
7817 * Evict l2arc buffers asynchronously; we need to keep the device
7818 * around until we are sure there aren't any buffers referencing it.
7819 * We do not need to hold any config locks, etc. because at this point,
7820 * we are the only ones who knows about this device (the in-core
7821 * structure), so no new buffers can be created (e.g. if the pool is
7822 * re-imported while the asynchronous eviction is in progress) that
7823 * reference this same in-core structure. Also remove the vdev link
7824 * since further use of it as l2arc device is prohibited.
7825 */
7826 dev->l2ad_vdev = NULL;
7827 l2arc_evict_impl(dev, 0LL, B_TRUE);
7828
7829 /* Same cleanup as in the synchronous path */
7830 list_destroy(&dev->l2ad_buflist);
7831 mutex_destroy(&dev->l2ad_mtx);
7832 refcount_destroy(&dev->l2ad_alloc);
7833 kmem_free(dev->l2ad_dev_hdr, dev->l2ad_dev_hdr_asize);
7834 kmem_free(dev, sizeof (l2arc_dev_t));
7835 }
7836
7837 boolean_t zfs_l2arc_async_evict = B_TRUE;
7838
7839 /*
7840 * Perform l2arc eviction for buffers associated with this device
7841 * If evicting all buffers (done at pool export time), try to evict
7842 * asynchronously, and fall back to synchronous eviction in case of error
7843 * Tell the caller whether to cleanup the device:
7844 * - B_TRUE means "asynchronous eviction, do not cleanup"
7845 * - B_FALSE means "synchronous eviction, done, please cleanup"
7846 */
7847 static boolean_t
7848 l2arc_evict(l2arc_dev_t *dev, uint64_t distance, boolean_t all)
7849 {
7850 /*
7851 * If we are evicting all the buffers for this device, which happens
7852 * at pool export time, schedule asynchronous task
7853 */
7854 if (all && zfs_l2arc_async_evict) {
7855 if ((taskq_dispatch(arc_flush_taskq, l2arc_evict_task,
7856 dev, TQ_NOSLEEP) == NULL)) {
7857 /*
7858 * Failed to dispatch asynchronous task
7859 * cleanup, evict synchronously
7860 */
7861 l2arc_evict_impl(dev, distance, all);
7862 } else {
7863 /*
7864 * Successful dispatch, vdev space updated
7865 */
7866 return (B_TRUE);
7867 }
7868 } else {
7869 /* Evict synchronously */
7870 l2arc_evict_impl(dev, distance, all);
7871 }
7872
7873 return (B_FALSE);
7874 }
7875
7876 /*
7877 * Find and write ARC buffers to the L2ARC device.
7878 *
7879 * An ARC_FLAG_L2_WRITING flag is set so that the L2ARC buffers are not valid
7880 * for reading until they have completed writing.
7881 * The headroom_boost is an in-out parameter used to maintain headroom boost
7882 * state between calls to this function.
7883 *
7884 * Returns the number of bytes actually written (which may be smaller than
7885 * the delta by which the device hand has changed due to alignment).
7886 */
7887 static uint64_t
7888 l2arc_write_buffers(spa_t *spa, l2arc_dev_t *dev, uint64_t target_sz,
7889 l2ad_feed_t feed_type)
7890 {
7891 arc_buf_hdr_t *hdr, *hdr_prev, *head;
7892 /*
7893 * We must carefully track the space we deal with here:
7894 * - write_size: sum of the size of all buffers to be written
7895 * without compression or inter-buffer alignment applied.
7896 * This size is added to arcstat_l2_size, because subsequent
7897 * eviction of buffers decrements this kstat by only the
7898 * buffer's b_lsize (which doesn't take alignment into account).
7899 * - write_asize: sum of the size of all buffers to be written
7900 * with inter-buffer alignment applied.
7901 * This size is used to estimate the maximum number of bytes
7902 * we could take up on the device and is thus used to gauge how
7903 * close we are to hitting target_sz.
7904 */
7905 uint64_t write_asize, write_psize, write_lsize, headroom;
7906 boolean_t full;
7907 l2arc_write_callback_t *cb;
7908 zio_t *pio, *wzio;
7909 enum l2arc_priorities try;
7910 uint64_t guid = spa_load_guid(spa);
7911 boolean_t dev_hdr_update = B_FALSE;
7912
7913 ASSERT3P(dev->l2ad_vdev, !=, NULL);
7914
7915 pio = NULL;
7916 cb = NULL;
7917 write_lsize = write_asize = write_psize = 0;
7918 full = B_FALSE;
7919 head = kmem_cache_alloc(hdr_l2only_cache, KM_PUSHPAGE);
7920 arc_hdr_set_flags(head, ARC_FLAG_L2_WRITE_HEAD | ARC_FLAG_HAS_L2HDR);
7921
7922 /*
7923 * Copy buffers for L2ARC writing.
7924 */
7925 for (try = PRIORITY_MFU_DDT; try < PRIORITY_NUMTYPES; try++) {
7926 multilist_sublist_t *mls = l2arc_sublist_lock(try);
7927 uint64_t passed_sz = 0;
7928
7929 /*
7930 * L2ARC fast warmup.
7931 *
7932 * Until the ARC is warm and starts to evict, read from the
7933 * head of the ARC lists rather than the tail.
7934 */
7935 if (arc_warm == B_FALSE)
7936 hdr = multilist_sublist_head(mls);
7937 else
7938 hdr = multilist_sublist_tail(mls);
7939
7940 headroom = target_sz * l2arc_headroom;
7941 if (zfs_compressed_arc_enabled)
7942 headroom = (headroom * l2arc_headroom_boost) / 100;
7943
7944 for (; hdr; hdr = hdr_prev) {
7945 kmutex_t *hash_lock;
7975 * We rely on the L1 portion of the header below, so
7976 * it's invalid for this header to have been evicted out
7977 * of the ghost cache, prior to being written out. The
7978 * ARC_FLAG_L2_WRITING bit ensures this won't happen.
7979 */
7980 ASSERT(HDR_HAS_L1HDR(hdr));
7981
7982 ASSERT3U(HDR_GET_PSIZE(hdr), >, 0);
7983 ASSERT3P(hdr->b_l1hdr.b_pabd, !=, NULL);
7984 ASSERT3U(arc_hdr_size(hdr), >, 0);
7985 uint64_t psize = arc_hdr_size(hdr);
7986 uint64_t asize = vdev_psize_to_asize(dev->l2ad_vdev,
7987 psize);
7988
7989 if ((write_asize + asize) > target_sz) {
7990 full = B_TRUE;
7991 mutex_exit(hash_lock);
7992 break;
7993 }
7994
7995 /* make sure buf we select corresponds to feed_type */
7996 if ((feed_type == L2ARC_FEED_DDT_DEV &&
7997 arc_buf_type(hdr) != ARC_BUFC_DDT) ||
7998 (feed_type == L2ARC_FEED_NON_DDT_DEV &&
7999 arc_buf_type(hdr) == ARC_BUFC_DDT)) {
8000 mutex_exit(hash_lock);
8001 continue;
8002 }
8003
8004 if (pio == NULL) {
8005 /*
8006 * Insert a dummy header on the buflist so
8007 * l2arc_write_done() can find where the
8008 * write buffers begin without searching.
8009 */
8010 mutex_enter(&dev->l2ad_mtx);
8011 list_insert_head(&dev->l2ad_buflist, head);
8012 mutex_exit(&dev->l2ad_mtx);
8013
8014 cb = kmem_zalloc(
8015 sizeof (l2arc_write_callback_t), KM_SLEEP);
8016 cb->l2wcb_dev = dev;
8017 cb->l2wcb_head = head;
8018 list_create(&cb->l2wcb_log_blk_buflist,
8019 sizeof (l2arc_log_blk_buf_t),
8020 offsetof(l2arc_log_blk_buf_t, lbb_node));
8021 pio = zio_root(spa, l2arc_write_done, cb,
8022 ZIO_FLAG_CANFAIL);
8023 }
8024
8025 hdr->b_l2hdr.b_dev = dev;
8026 hdr->b_l2hdr.b_daddr = dev->l2ad_hand;
8027 arc_hdr_set_flags(hdr,
8028 ARC_FLAG_L2_WRITING | ARC_FLAG_HAS_L2HDR);
8029
8030 mutex_enter(&dev->l2ad_mtx);
8031 list_insert_head(&dev->l2ad_buflist, hdr);
8032 mutex_exit(&dev->l2ad_mtx);
8033
8034 (void) refcount_add_many(&dev->l2ad_alloc, psize, hdr);
8035
8036 /*
8037 * Normally the L2ARC can use the hdr's data, but if
8038 * we're sharing data between the hdr and one of its
8039 * bufs, L2ARC needs its own copy of the data so that
8040 * the ZIO below can't race with the buf consumer.
8041 * Another case where we need to create a copy of the
8042 * data is when the buffer size is not device-aligned
8043 * and we need to pad the block to make it such.
8044 * That also keeps the clock hand suitably aligned.
8045 *
8046 * To ensure that the copy will be available for the
8047 * lifetime of the ZIO and be cleaned up afterwards, we
8048 * add it to the l2arc_free_on_write queue.
8049 */
8050 abd_t *to_write;
8051 if (!HDR_SHARED_DATA(hdr) && psize == asize) {
8052 to_write = hdr->b_l1hdr.b_pabd;
8053 } else {
8054 to_write = abd_alloc_for_io(asize,
8055 !HDR_ISTYPE_DATA(hdr));
8056 abd_copy(to_write, hdr->b_l1hdr.b_pabd, psize);
8057 if (asize != psize) {
8058 abd_zero_off(to_write, psize,
8059 asize - psize);
8060 }
8061 l2arc_free_abd_on_write(to_write, asize,
8062 arc_buf_type(hdr));
8063 }
8064 wzio = zio_write_phys(pio, dev->l2ad_vdev,
8065 hdr->b_l2hdr.b_daddr, asize, to_write,
8066 ZIO_CHECKSUM_OFF, NULL, hdr,
8067 ZIO_PRIORITY_ASYNC_WRITE,
8068 ZIO_FLAG_CANFAIL, B_FALSE);
8069
8070 write_lsize += HDR_GET_LSIZE(hdr);
8071 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev,
8072 zio_t *, wzio);
8073
8074 write_psize += psize;
8075 write_asize += asize;
8076 dev->l2ad_hand += asize;
8077
8078 mutex_exit(hash_lock);
8079
8080 (void) zio_nowait(wzio);
8081
8082 /*
8083 * Append buf info to current log and commit if full.
8084 * arcstat_l2_{size,asize} kstats are updated internally.
8085 */
8086 if (l2arc_log_blk_insert(dev, hdr)) {
8087 l2arc_log_blk_commit(dev, pio, cb);
8088 dev_hdr_update = B_TRUE;
8089 }
8090 }
8091
8092 multilist_sublist_unlock(mls);
8093
8094 if (full == B_TRUE)
8095 break;
8096 }
8097
8098 /* No buffers selected for writing? */
8099 if (pio == NULL) {
8100 ASSERT0(write_lsize);
8101 ASSERT(!HDR_HAS_L1HDR(head));
8102 kmem_cache_free(hdr_l2only_cache, head);
8103 return (0);
8104 }
8105
8106 /*
8107 * If we wrote any logs as part of this write, update dev hdr
8108 * to point to it.
8109 */
8110 if (dev_hdr_update)
8111 l2arc_dev_hdr_update(dev, pio);
8112
8113 ASSERT3U(write_asize, <=, target_sz);
8114 ARCSTAT_BUMP(arcstat_l2_writes_sent);
8115 ARCSTAT_INCR(arcstat_l2_write_bytes, write_psize);
8116 if (feed_type == L2ARC_FEED_DDT_DEV)
8117 ARCSTAT_INCR(arcstat_l2_ddt_write_bytes, write_psize);
8118 ARCSTAT_INCR(arcstat_l2_lsize, write_lsize);
8119 ARCSTAT_INCR(arcstat_l2_psize, write_psize);
8120 vdev_space_update(dev->l2ad_vdev, write_psize, 0, 0);
8121
8122 /*
8123 * Bump device hand to the device start if it is approaching the end.
8124 * l2arc_evict() will already have evicted ahead for this case.
8125 */
8126 if (dev->l2ad_hand + target_sz + l2arc_log_blk_overhead(target_sz) >=
8127 dev->l2ad_end) {
8128 dev->l2ad_hand = dev->l2ad_start;
8129 dev->l2ad_first = B_FALSE;
8130 }
8131
8132 dev->l2ad_writing = B_TRUE;
8133 (void) zio_wait(pio);
8134 dev->l2ad_writing = B_FALSE;
8135
8136 return (write_asize);
8137 }
8138
8139 static boolean_t
8140 l2arc_feed_dev(l2ad_feed_t feed_type, uint64_t *wrote)
8141 {
8142 spa_t *spa;
8143 l2arc_dev_t *dev;
8144 uint64_t size;
8145
8146 /*
8147 * This selects the next l2arc device to write to, and in
8148 * doing so the next spa to feed from: dev->l2ad_spa. This
8149 * will return NULL if there are now no l2arc devices or if
8150 * they are all faulted.
8151 *
8152 * If a device is returned, its spa's config lock is also
8153 * held to prevent device removal. l2arc_dev_get_next()
8154 * will grab and release l2arc_dev_mtx.
8155 */
8156 if ((dev = l2arc_dev_get_next(feed_type)) == NULL)
8157 return (B_FALSE);
8158
8159 spa = dev->l2ad_spa;
8160 ASSERT(spa != NULL);
8161
8162 /*
8163 * If the pool is read-only - skip it
8164 */
8165 if (!spa_writeable(spa)) {
8166 spa_config_exit(spa, SCL_L2ARC, dev);
8167 return (B_FALSE);
8168 }
8169
8170 ARCSTAT_BUMP(arcstat_l2_feeds);
8171 size = l2arc_write_size();
8172
8173 /*
8174 * Evict L2ARC buffers that will be overwritten.
8175 * B_FALSE guarantees synchronous eviction.
8176 */
8177 (void) l2arc_evict(dev, size, B_FALSE);
8178
8179 /*
8180 * Write ARC buffers.
8181 */
8182 *wrote = l2arc_write_buffers(spa, dev, size, feed_type);
8183
8184 spa_config_exit(spa, SCL_L2ARC, dev);
8185
8186 return (B_TRUE);
8187 }
8188
8189 /*
8190 * This thread feeds the L2ARC at regular intervals. This is the beating
8191 * heart of the L2ARC.
8192 */
8193 /* ARGSUSED */
8194 static void
8195 l2arc_feed_thread(void *unused)
8196 {
8197 callb_cpr_t cpr;
8198 uint64_t size, total_written = 0;
8199 clock_t begin, next = ddi_get_lbolt();
8200 l2ad_feed_t feed_type = L2ARC_FEED_ALL;
8201
8202 CALLB_CPR_INIT(&cpr, &l2arc_feed_thr_lock, callb_generic_cpr, FTAG);
8203
8204 mutex_enter(&l2arc_feed_thr_lock);
8205
8206 while (l2arc_thread_exit == 0) {
8207 CALLB_CPR_SAFE_BEGIN(&cpr);
8208 (void) cv_timedwait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock,
8209 next);
8210 CALLB_CPR_SAFE_END(&cpr, &l2arc_feed_thr_lock);
8211 next = ddi_get_lbolt() + hz;
8212
8213 /*
8214 * Quick check for L2ARC devices.
8215 */
8216 mutex_enter(&l2arc_dev_mtx);
8217 if (l2arc_ndev == 0) {
8218 mutex_exit(&l2arc_dev_mtx);
8219 continue;
8220 }
8221 mutex_exit(&l2arc_dev_mtx);
8222 begin = ddi_get_lbolt();
8223
8224 /*
8225 * Avoid contributing to memory pressure.
8226 */
8227 if (arc_reclaim_needed()) {
8228 ARCSTAT_BUMP(arcstat_l2_abort_lowmem);
8229 continue;
8230 }
8231
8232 /* try to write to DDT L2ARC device if any */
8233 if (l2arc_feed_dev(L2ARC_FEED_DDT_DEV, &size)) {
8234 total_written += size;
8235 feed_type = L2ARC_FEED_NON_DDT_DEV;
8236 }
8237
8238 /* try to write to the regular L2ARC device if any */
8239 if (l2arc_feed_dev(feed_type, &size)) {
8240 total_written += size;
8241 if (feed_type == L2ARC_FEED_NON_DDT_DEV)
8242 total_written /= 2; /* avg written per device */
8243 }
8244
8245 /*
8246 * Calculate interval between writes.
8247 */
8248 next = l2arc_write_interval(begin, l2arc_write_size(),
8249 total_written);
8250
8251 total_written = 0;
8252 }
8253
8254 l2arc_thread_exit = 0;
8255 cv_broadcast(&l2arc_feed_thr_cv);
8256 CALLB_CPR_EXIT(&cpr); /* drops l2arc_feed_thr_lock */
8257 thread_exit();
8258 }
8259
8260 boolean_t
8261 l2arc_vdev_present(vdev_t *vd)
8262 {
8263 return (l2arc_vdev_get(vd) != NULL);
8264 }
8265
8266 /*
8267 * Returns the l2arc_dev_t associated with a particular vdev_t or NULL if
8268 * the vdev_t isn't an L2ARC device.
8269 */
8270 static l2arc_dev_t *
8271 l2arc_vdev_get(vdev_t *vd)
8272 {
8273 l2arc_dev_t *dev;
8274 boolean_t held = MUTEX_HELD(&l2arc_dev_mtx);
8275
8276 if (!held)
8277 mutex_enter(&l2arc_dev_mtx);
8278 for (dev = list_head(l2arc_dev_list); dev != NULL;
8279 dev = list_next(l2arc_dev_list, dev)) {
8280 if (dev->l2ad_vdev == vd)
8281 break;
8282 }
8283 if (!held)
8284 mutex_exit(&l2arc_dev_mtx);
8285
8286 return (dev);
8287 }
8288
8289 /*
8290 * Add a vdev for use by the L2ARC. By this point the spa has already
8291 * validated the vdev and opened it. The `rebuild' flag indicates whether
8292 * we should attempt an L2ARC persistency rebuild.
8293 */
8294 void
8295 l2arc_add_vdev(spa_t *spa, vdev_t *vd, boolean_t rebuild)
8296 {
8297 l2arc_dev_t *adddev;
8298
8299 ASSERT(!l2arc_vdev_present(vd));
8300
8301 /*
8302 * Create a new l2arc device entry.
8303 */
8304 adddev = kmem_zalloc(sizeof (l2arc_dev_t), KM_SLEEP);
8305 adddev->l2ad_spa = spa;
8306 adddev->l2ad_vdev = vd;
8307 /* leave extra size for an l2arc device header */
8308 adddev->l2ad_dev_hdr_asize = MAX(sizeof (*adddev->l2ad_dev_hdr),
8309 1 << vd->vdev_ashift);
8310 adddev->l2ad_start = VDEV_LABEL_START_SIZE + adddev->l2ad_dev_hdr_asize;
8311 adddev->l2ad_end = VDEV_LABEL_START_SIZE + vdev_get_min_asize(vd);
8312 ASSERT3U(adddev->l2ad_start, <, adddev->l2ad_end);
8313 adddev->l2ad_hand = adddev->l2ad_start;
8314 adddev->l2ad_first = B_TRUE;
8315 adddev->l2ad_writing = B_FALSE;
8316 adddev->l2ad_dev_hdr = kmem_zalloc(adddev->l2ad_dev_hdr_asize,
8317 KM_SLEEP);
8318
8319 mutex_init(&adddev->l2ad_mtx, NULL, MUTEX_DEFAULT, NULL);
8320 /*
8321 * This is a list of all ARC buffers that are still valid on the
8322 * device.
8323 */
8324 list_create(&adddev->l2ad_buflist, sizeof (arc_buf_hdr_t),
8325 offsetof(arc_buf_hdr_t, b_l2hdr.b_l2node));
8326
8327 vdev_space_update(vd, 0, 0, adddev->l2ad_end - adddev->l2ad_hand);
8328 refcount_create(&adddev->l2ad_alloc);
8329
8330 /*
8331 * Add device to global list
8332 */
8333 mutex_enter(&l2arc_dev_mtx);
8334 list_insert_head(l2arc_dev_list, adddev);
8335 atomic_inc_64(&l2arc_ndev);
8336 if (rebuild && l2arc_rebuild_enabled &&
8337 adddev->l2ad_end - adddev->l2ad_start > L2ARC_PERSIST_MIN_SIZE) {
8338 /*
8339 * Just mark the device as pending for a rebuild. We won't
8340 * be starting a rebuild in line here as it would block pool
8341 * import. Instead spa_load_impl will hand that off to an
8342 * async task which will call l2arc_spa_rebuild_start.
8343 */
8344 adddev->l2ad_rebuild = B_TRUE;
8345 }
8346 mutex_exit(&l2arc_dev_mtx);
8347 }
8348
8349 /*
8350 * Remove a vdev from the L2ARC.
8351 */
8352 void
8353 l2arc_remove_vdev(vdev_t *vd)
8354 {
8355 l2arc_dev_t *dev, *nextdev, *remdev = NULL;
8356
8357 /*
8358 * Find the device by vdev
8359 */
8360 mutex_enter(&l2arc_dev_mtx);
8361 for (dev = list_head(l2arc_dev_list); dev; dev = nextdev) {
8362 nextdev = list_next(l2arc_dev_list, dev);
8363 if (vd == dev->l2ad_vdev) {
8364 remdev = dev;
8365 break;
8366 }
8367 }
8368 ASSERT3P(remdev, !=, NULL);
8369
8370 /*
8371 * Cancel any ongoing or scheduled rebuild (race protection with
8372 * l2arc_spa_rebuild_start provided via l2arc_dev_mtx).
8373 */
8374 remdev->l2ad_rebuild_cancel = B_TRUE;
8375 if (remdev->l2ad_rebuild_did != 0) {
8376 /*
8377 * N.B. it should be safe to thread_join with the rebuild
8378 * thread while holding l2arc_dev_mtx because it is not
8379 * accessed from anywhere in the l2arc rebuild code below
8380 * (except for l2arc_spa_rebuild_start, which is ok).
8381 */
8382 thread_join(remdev->l2ad_rebuild_did);
8383 }
8384
8385 /*
8386 * Remove device from global list
8387 */
8388 list_remove(l2arc_dev_list, remdev);
8389 l2arc_dev_last = NULL; /* may have been invalidated */
8390 l2arc_ddt_dev_last = NULL; /* may have been invalidated */
8391 atomic_dec_64(&l2arc_ndev);
8392 mutex_exit(&l2arc_dev_mtx);
8393
8394 if (vdev_type_is_ddt(remdev->l2ad_vdev))
8395 atomic_add_64(&remdev->l2ad_spa->spa_l2arc_ddt_devs_size,
8396 -(vdev_get_min_asize(remdev->l2ad_vdev)));
8397
8398 /*
8399 * Clear all buflists and ARC references. L2ARC device flush.
8400 */
8401 if (l2arc_evict(remdev, 0, B_TRUE) == B_FALSE) {
8402 /*
8403 * The eviction was done synchronously, cleanup here
8404 * Otherwise, the asynchronous task will cleanup
8405 */
8406 list_destroy(&remdev->l2ad_buflist);
8407 mutex_destroy(&remdev->l2ad_mtx);
8408 kmem_free(remdev->l2ad_dev_hdr, remdev->l2ad_dev_hdr_asize);
8409 kmem_free(remdev, sizeof (l2arc_dev_t));
8410 }
8411 }
8412
8413 void
8414 l2arc_init(void)
8415 {
8416 l2arc_thread_exit = 0;
8417 l2arc_ndev = 0;
8418 l2arc_writes_sent = 0;
8419 l2arc_writes_done = 0;
8420
8421 mutex_init(&l2arc_feed_thr_lock, NULL, MUTEX_DEFAULT, NULL);
8422 cv_init(&l2arc_feed_thr_cv, NULL, CV_DEFAULT, NULL);
8423 mutex_init(&l2arc_dev_mtx, NULL, MUTEX_DEFAULT, NULL);
8424 mutex_init(&l2arc_free_on_write_mtx, NULL, MUTEX_DEFAULT, NULL);
8425
8426 l2arc_dev_list = &L2ARC_dev_list;
8427 l2arc_free_on_write = &L2ARC_free_on_write;
8428 list_create(l2arc_dev_list, sizeof (l2arc_dev_t),
8429 offsetof(l2arc_dev_t, l2ad_node));
8430 list_create(l2arc_free_on_write, sizeof (l2arc_data_free_t),
8456 {
8457 if (!(spa_mode_global & FWRITE))
8458 return;
8459
8460 (void) thread_create(NULL, 0, l2arc_feed_thread, NULL, 0, &p0,
8461 TS_RUN, minclsyspri);
8462 }
8463
8464 void
8465 l2arc_stop(void)
8466 {
8467 if (!(spa_mode_global & FWRITE))
8468 return;
8469
8470 mutex_enter(&l2arc_feed_thr_lock);
8471 cv_signal(&l2arc_feed_thr_cv); /* kick thread out of startup */
8472 l2arc_thread_exit = 1;
8473 while (l2arc_thread_exit != 0)
8474 cv_wait(&l2arc_feed_thr_cv, &l2arc_feed_thr_lock);
8475 mutex_exit(&l2arc_feed_thr_lock);
8476 }
8477
8478 /*
8479 * Punches out rebuild threads for the L2ARC devices in a spa. This should
8480 * be called after pool import from the spa async thread, since starting
8481 * these threads directly from spa_import() will make them part of the
8482 * "zpool import" context and delay process exit (and thus pool import).
8483 */
8484 void
8485 l2arc_spa_rebuild_start(spa_t *spa)
8486 {
8487 /*
8488 * Locate the spa's l2arc devices and kick off rebuild threads.
8489 */
8490 mutex_enter(&l2arc_dev_mtx);
8491 for (int i = 0; i < spa->spa_l2cache.sav_count; i++) {
8492 l2arc_dev_t *dev =
8493 l2arc_vdev_get(spa->spa_l2cache.sav_vdevs[i]);
8494 if (dev == NULL) {
8495 /* Don't attempt a rebuild if the vdev is UNAVAIL */
8496 continue;
8497 }
8498 if (dev->l2ad_rebuild && !dev->l2ad_rebuild_cancel) {
8499 VERIFY3U(dev->l2ad_rebuild_did, ==, 0);
8500 #ifdef _KERNEL
8501 dev->l2ad_rebuild_did = thread_create(NULL, 0,
8502 l2arc_dev_rebuild_start, dev, 0, &p0, TS_RUN,
8503 minclsyspri)->t_did;
8504 #endif
8505 }
8506 }
8507 mutex_exit(&l2arc_dev_mtx);
8508 }
8509
8510 /*
8511 * Main entry point for L2ARC rebuilding.
8512 */
8513 static void
8514 l2arc_dev_rebuild_start(l2arc_dev_t *dev)
8515 {
8516 if (!dev->l2ad_rebuild_cancel) {
8517 VERIFY(dev->l2ad_rebuild);
8518 (void) l2arc_rebuild(dev);
8519 dev->l2ad_rebuild = B_FALSE;
8520 }
8521 }
8522
8523 /*
8524 * This function implements the actual L2ARC metadata rebuild. It:
8525 *
8526 * 1) reads the device's header
8527 * 2) if a good device header is found, starts reading the log block chain
8528 * 3) restores each block's contents to memory (reconstructing arc_buf_hdr_t's)
8529 *
8530 * Operation stops under any of the following conditions:
8531 *
8532 * 1) We reach the end of the log blk chain (the back-reference in the blk is
8533 * invalid or loops over our starting point).
8534 * 2) We encounter *any* error condition (cksum errors, io errors, looped
8535 * blocks, etc.).
8536 */
8537 static int
8538 l2arc_rebuild(l2arc_dev_t *dev)
8539 {
8540 vdev_t *vd = dev->l2ad_vdev;
8541 spa_t *spa = vd->vdev_spa;
8542 int err;
8543 l2arc_log_blk_phys_t *this_lb, *next_lb;
8544 uint8_t *this_lb_buf, *next_lb_buf;
8545 zio_t *this_io = NULL, *next_io = NULL;
8546 l2arc_log_blkptr_t lb_ptrs[2];
8547 boolean_t first_pass, lock_held;
8548 uint64_t load_guid;
8549
8550 this_lb = kmem_zalloc(sizeof (*this_lb), KM_SLEEP);
8551 next_lb = kmem_zalloc(sizeof (*next_lb), KM_SLEEP);
8552 this_lb_buf = kmem_zalloc(sizeof (l2arc_log_blk_phys_t), KM_SLEEP);
8553 next_lb_buf = kmem_zalloc(sizeof (l2arc_log_blk_phys_t), KM_SLEEP);
8554
8555 /*
8556 * We prevent device removal while issuing reads to the device,
8557 * then during the rebuilding phases we drop this lock again so
8558 * that a spa_unload or device remove can be initiated - this is
8559 * safe, because the spa will signal us to stop before removing
8560 * our device and wait for us to stop.
8561 */
8562 spa_config_enter(spa, SCL_L2ARC, vd, RW_READER);
8563 lock_held = B_TRUE;
8564
8565 load_guid = spa_load_guid(dev->l2ad_vdev->vdev_spa);
8566 /*
8567 * Device header processing phase.
8568 */
8569 if ((err = l2arc_dev_hdr_read(dev)) != 0) {
8570 /* device header corrupted, start a new one */
8571 bzero(dev->l2ad_dev_hdr, dev->l2ad_dev_hdr_asize);
8572 goto out;
8573 }
8574
8575 /* Retrieve the persistent L2ARC device state */
8576 dev->l2ad_hand = vdev_psize_to_asize(dev->l2ad_vdev,
8577 dev->l2ad_dev_hdr->dh_start_lbps[0].lbp_daddr +
8578 LBP_GET_PSIZE(&dev->l2ad_dev_hdr->dh_start_lbps[0]));
8579 dev->l2ad_first = !!(dev->l2ad_dev_hdr->dh_flags &
8580 L2ARC_DEV_HDR_EVICT_FIRST);
8581
8582 /* Prepare the rebuild processing state */
8583 bcopy(dev->l2ad_dev_hdr->dh_start_lbps, lb_ptrs, sizeof (lb_ptrs));
8584 first_pass = B_TRUE;
8585
8586 /* Start the rebuild process */
8587 for (;;) {
8588 if (!l2arc_log_blkptr_valid(dev, &lb_ptrs[0]))
8589 /* We hit an invalid block address, end the rebuild. */
8590 break;
8591
8592 if ((err = l2arc_log_blk_read(dev, &lb_ptrs[0], &lb_ptrs[1],
8593 this_lb, next_lb, this_lb_buf, next_lb_buf,
8594 this_io, &next_io)) != 0)
8595 break;
8596
8597 spa_config_exit(spa, SCL_L2ARC, vd);
8598 lock_held = B_FALSE;
8599
8600 /* Protection against infinite loops of log blocks. */
8601 if (l2arc_range_check_overlap(lb_ptrs[1].lbp_daddr,
8602 lb_ptrs[0].lbp_daddr,
8603 dev->l2ad_dev_hdr->dh_start_lbps[0].lbp_daddr) &&
8604 !first_pass) {
8605 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_loop_errors);
8606 err = SET_ERROR(ELOOP);
8607 break;
8608 }
8609
8610 /*
8611 * Our memory pressure valve. If the system is running low
8612 * on memory, rather than swamping memory with new ARC buf
8613 * hdrs, we opt not to rebuild the L2ARC. At this point,
8614 * however, we have already set up our L2ARC dev to chain in
8615 * new metadata log blk, so the user may choose to re-add the
8616 * L2ARC dev at a later time to reconstruct it (when there's
8617 * less memory pressure).
8618 */
8619 if (arc_reclaim_needed()) {
8620 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_lowmem);
8621 cmn_err(CE_NOTE, "System running low on memory, "
8622 "aborting L2ARC rebuild.");
8623 err = SET_ERROR(ENOMEM);
8624 break;
8625 }
8626
8627 /*
8628 * Now that we know that the next_lb checks out alright, we
8629 * can start reconstruction from this lb - we can be sure
8630 * that the L2ARC write hand has not yet reached any of our
8631 * buffers.
8632 */
8633 l2arc_log_blk_restore(dev, load_guid, this_lb,
8634 LBP_GET_PSIZE(&lb_ptrs[0]));
8635
8636 /*
8637 * End of list detection. We can look ahead two steps in the
8638 * blk chain and if the 2nd blk from this_lb dips below the
8639 * initial chain starting point, then we know two things:
8640 * 1) it can't be valid, and
8641 * 2) the next_lb's ARC entries might have already been
8642 * partially overwritten and so we should stop before
8643 * we restore it
8644 */
8645 if (l2arc_range_check_overlap(
8646 this_lb->lb_back2_lbp.lbp_daddr, lb_ptrs[0].lbp_daddr,
8647 dev->l2ad_dev_hdr->dh_start_lbps[0].lbp_daddr) &&
8648 !first_pass)
8649 break;
8650
8651 /* log blk restored, continue with next one in the list */
8652 lb_ptrs[0] = lb_ptrs[1];
8653 lb_ptrs[1] = this_lb->lb_back2_lbp;
8654 PTR_SWAP(this_lb, next_lb);
8655 PTR_SWAP(this_lb_buf, next_lb_buf);
8656 this_io = next_io;
8657 next_io = NULL;
8658 first_pass = B_FALSE;
8659
8660 for (;;) {
8661 if (dev->l2ad_rebuild_cancel) {
8662 err = SET_ERROR(ECANCELED);
8663 goto out;
8664 }
8665 if (spa_config_tryenter(spa, SCL_L2ARC, vd,
8666 RW_READER)) {
8667 lock_held = B_TRUE;
8668 break;
8669 }
8670 /*
8671 * L2ARC config lock held by somebody in writer,
8672 * possibly due to them trying to remove us. They'll
8673 * likely to want us to shut down, so after a little
8674 * delay, we check l2ad_rebuild_cancel and retry
8675 * the lock again.
8676 */
8677 delay(1);
8678 }
8679 }
8680 out:
8681 if (next_io != NULL)
8682 l2arc_log_blk_prefetch_abort(next_io);
8683 kmem_free(this_lb, sizeof (*this_lb));
8684 kmem_free(next_lb, sizeof (*next_lb));
8685 kmem_free(this_lb_buf, sizeof (l2arc_log_blk_phys_t));
8686 kmem_free(next_lb_buf, sizeof (l2arc_log_blk_phys_t));
8687 if (err == 0)
8688 ARCSTAT_BUMP(arcstat_l2_rebuild_successes);
8689
8690 if (lock_held)
8691 spa_config_exit(spa, SCL_L2ARC, vd);
8692
8693 return (err);
8694 }
8695
8696 /*
8697 * Attempts to read the device header on the provided L2ARC device and writes
8698 * it to `hdr'. On success, this function returns 0, otherwise the appropriate
8699 * error code is returned.
8700 */
8701 static int
8702 l2arc_dev_hdr_read(l2arc_dev_t *dev)
8703 {
8704 int err;
8705 uint64_t guid;
8706 zio_cksum_t cksum;
8707 l2arc_dev_hdr_phys_t *hdr = dev->l2ad_dev_hdr;
8708 const uint64_t hdr_asize = dev->l2ad_dev_hdr_asize;
8709 abd_t *abd;
8710
8711 guid = spa_guid(dev->l2ad_vdev->vdev_spa);
8712
8713 abd = abd_get_from_buf(hdr, hdr_asize);
8714 err = zio_wait(zio_read_phys(NULL, dev->l2ad_vdev,
8715 VDEV_LABEL_START_SIZE, hdr_asize, abd,
8716 ZIO_CHECKSUM_OFF, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
8717 ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
8718 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
8719 abd_put(abd);
8720 if (err != 0) {
8721 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
8722 return (err);
8723 }
8724
8725 if (hdr->dh_magic == BSWAP_64(L2ARC_DEV_HDR_MAGIC_V1))
8726 byteswap_uint64_array(hdr, sizeof (*hdr));
8727
8728 if (hdr->dh_magic != L2ARC_DEV_HDR_MAGIC_V1 ||
8729 hdr->dh_spa_guid != guid) {
8730 /*
8731 * Attempt to rebuild a device containing no actual dev hdr
8732 * or containing a header from some other pool.
8733 */
8734 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_unsupported);
8735 return (SET_ERROR(ENOTSUP));
8736 }
8737
8738 l2arc_dev_hdr_checksum(hdr, &cksum);
8739 if (!ZIO_CHECKSUM_EQUAL(hdr->dh_self_cksum, cksum)) {
8740 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_errors);
8741 return (SET_ERROR(EINVAL));
8742 }
8743
8744 return (0);
8745 }
8746
8747 /*
8748 * Reads L2ARC log blocks from storage and validates their contents.
8749 *
8750 * This function implements a simple prefetcher to make sure that while
8751 * we're processing one buffer the L2ARC is already prefetching the next
8752 * one in the chain.
8753 *
8754 * The arguments this_lp and next_lp point to the current and next log blk
8755 * address in the block chain. Similarly, this_lb and next_lb hold the
8756 * l2arc_log_blk_phys_t's of the current and next L2ARC blk. The this_lb_buf
8757 * and next_lb_buf must be buffers of appropriate to hold a raw
8758 * l2arc_log_blk_phys_t (they are used as catch buffers for read ops prior
8759 * to buffer decompression).
8760 *
8761 * The `this_io' and `next_io' arguments are used for block prefetching.
8762 * When issuing the first blk IO during rebuild, you should pass NULL for
8763 * `this_io'. This function will then issue a sync IO to read the block and
8764 * also issue an async IO to fetch the next block in the block chain. The
8765 * prefetch IO is returned in `next_io'. On subsequent calls to this
8766 * function, pass the value returned in `next_io' from the previous call
8767 * as `this_io' and a fresh `next_io' pointer to hold the next prefetch IO.
8768 * Prior to the call, you should initialize your `next_io' pointer to be
8769 * NULL. If no prefetch IO was issued, the pointer is left set at NULL.
8770 *
8771 * On success, this function returns 0, otherwise it returns an appropriate
8772 * error code. On error the prefetching IO is aborted and cleared before
8773 * returning from this function. Therefore, if we return `success', the
8774 * caller can assume that we have taken care of cleanup of prefetch IOs.
8775 */
8776 static int
8777 l2arc_log_blk_read(l2arc_dev_t *dev,
8778 const l2arc_log_blkptr_t *this_lbp, const l2arc_log_blkptr_t *next_lbp,
8779 l2arc_log_blk_phys_t *this_lb, l2arc_log_blk_phys_t *next_lb,
8780 uint8_t *this_lb_buf, uint8_t *next_lb_buf,
8781 zio_t *this_io, zio_t **next_io)
8782 {
8783 int err = 0;
8784 zio_cksum_t cksum;
8785
8786 ASSERT(this_lbp != NULL && next_lbp != NULL);
8787 ASSERT(this_lb != NULL && next_lb != NULL);
8788 ASSERT(this_lb_buf != NULL && next_lb_buf != NULL);
8789 ASSERT(next_io != NULL && *next_io == NULL);
8790 ASSERT(l2arc_log_blkptr_valid(dev, this_lbp));
8791
8792 /*
8793 * Check to see if we have issued the IO for this log blk in a
8794 * previous run. If not, this is the first call, so issue it now.
8795 */
8796 if (this_io == NULL) {
8797 this_io = l2arc_log_blk_prefetch(dev->l2ad_vdev, this_lbp,
8798 this_lb_buf);
8799 }
8800
8801 /*
8802 * Peek to see if we can start issuing the next IO immediately.
8803 */
8804 if (l2arc_log_blkptr_valid(dev, next_lbp)) {
8805 /*
8806 * Start issuing IO for the next log blk early - this
8807 * should help keep the L2ARC device busy while we
8808 * decompress and restore this log blk.
8809 */
8810 *next_io = l2arc_log_blk_prefetch(dev->l2ad_vdev, next_lbp,
8811 next_lb_buf);
8812 }
8813
8814 /* Wait for the IO to read this log block to complete */
8815 if ((err = zio_wait(this_io)) != 0) {
8816 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_io_errors);
8817 goto cleanup;
8818 }
8819
8820 /* Make sure the buffer checks out */
8821 fletcher_4_native(this_lb_buf, LBP_GET_PSIZE(this_lbp), NULL, &cksum);
8822 if (!ZIO_CHECKSUM_EQUAL(cksum, this_lbp->lbp_cksum)) {
8823 ARCSTAT_BUMP(arcstat_l2_rebuild_abort_cksum_errors);
8824 err = SET_ERROR(EINVAL);
8825 goto cleanup;
8826 }
8827
8828 /* Now we can take our time decoding this buffer */
8829 switch (LBP_GET_COMPRESS(this_lbp)) {
8830 case ZIO_COMPRESS_OFF:
8831 bcopy(this_lb_buf, this_lb, sizeof (*this_lb));
8832 break;
8833 case ZIO_COMPRESS_LZ4:
8834 err = zio_decompress_data_buf(LBP_GET_COMPRESS(this_lbp),
8835 this_lb_buf, this_lb, LBP_GET_PSIZE(this_lbp),
8836 sizeof (*this_lb));
8837 if (err != 0) {
8838 err = SET_ERROR(EINVAL);
8839 goto cleanup;
8840 }
8841
8842 break;
8843 default:
8844 err = SET_ERROR(EINVAL);
8845 break;
8846 }
8847
8848 if (this_lb->lb_magic == BSWAP_64(L2ARC_LOG_BLK_MAGIC))
8849 byteswap_uint64_array(this_lb, sizeof (*this_lb));
8850
8851 if (this_lb->lb_magic != L2ARC_LOG_BLK_MAGIC) {
8852 err = SET_ERROR(EINVAL);
8853 goto cleanup;
8854 }
8855
8856 cleanup:
8857 /* Abort an in-flight prefetch I/O in case of error */
8858 if (err != 0 && *next_io != NULL) {
8859 l2arc_log_blk_prefetch_abort(*next_io);
8860 *next_io = NULL;
8861 }
8862 return (err);
8863 }
8864
8865 /*
8866 * Restores the payload of a log blk to ARC. This creates empty ARC hdr
8867 * entries which only contain an l2arc hdr, essentially restoring the
8868 * buffers to their L2ARC evicted state. This function also updates space
8869 * usage on the L2ARC vdev to make sure it tracks restored buffers.
8870 */
8871 static void
8872 l2arc_log_blk_restore(l2arc_dev_t *dev, uint64_t load_guid,
8873 const l2arc_log_blk_phys_t *lb, uint64_t lb_psize)
8874 {
8875 uint64_t size = 0, psize = 0;
8876
8877 for (int i = L2ARC_LOG_BLK_ENTRIES - 1; i >= 0; i--) {
8878 /*
8879 * Restore goes in the reverse temporal direction to preserve
8880 * correct temporal ordering of buffers in the l2ad_buflist.
8881 * l2arc_hdr_restore also does a list_insert_tail instead of
8882 * list_insert_head on the l2ad_buflist:
8883 *
8884 * LIST l2ad_buflist LIST
8885 * HEAD <------ (time) ------ TAIL
8886 * direction +-----+-----+-----+-----+-----+ direction
8887 * of l2arc <== | buf | buf | buf | buf | buf | ===> of rebuild
8888 * fill +-----+-----+-----+-----+-----+
8889 * ^ ^
8890 * | |
8891 * | |
8892 * l2arc_fill_thread l2arc_rebuild
8893 * places new bufs here restores bufs here
8894 *
8895 * This also works when the restored bufs get evicted at any
8896 * point during the rebuild.
8897 */
8898 l2arc_hdr_restore(&lb->lb_entries[i], dev, load_guid);
8899 size += LE_GET_LSIZE(&lb->lb_entries[i]);
8900 psize += LE_GET_PSIZE(&lb->lb_entries[i]);
8901 }
8902
8903 /*
8904 * Record rebuild stats:
8905 * size In-memory size of restored buffer data in ARC
8906 * psize Physical size of restored buffers in the L2ARC
8907 * bufs # of ARC buffer headers restored
8908 * log_blks # of L2ARC log entries processed during restore
8909 */
8910 ARCSTAT_INCR(arcstat_l2_rebuild_size, size);
8911 ARCSTAT_INCR(arcstat_l2_rebuild_psize, psize);
8912 ARCSTAT_INCR(arcstat_l2_rebuild_bufs, L2ARC_LOG_BLK_ENTRIES);
8913 ARCSTAT_BUMP(arcstat_l2_rebuild_log_blks);
8914 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_size, lb_psize);
8915 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio, psize / lb_psize);
8916 vdev_space_update(dev->l2ad_vdev, psize, 0, 0);
8917 }
8918
8919 /*
8920 * Restores a single ARC buf hdr from a log block. The ARC buffer is put
8921 * into a state indicating that it has been evicted to L2ARC.
8922 */
8923 static void
8924 l2arc_hdr_restore(const l2arc_log_ent_phys_t *le, l2arc_dev_t *dev,
8925 uint64_t load_guid)
8926 {
8927 arc_buf_hdr_t *hdr, *exists;
8928 kmutex_t *hash_lock;
8929 arc_buf_contents_t type = LE_GET_TYPE(le);
8930
8931 /*
8932 * Do all the allocation before grabbing any locks, this lets us
8933 * sleep if memory is full and we don't have to deal with failed
8934 * allocations.
8935 */
8936 hdr = arc_buf_alloc_l2only(load_guid, type, dev, le->le_dva,
8937 le->le_daddr, LE_GET_LSIZE(le), LE_GET_PSIZE(le),
8938 le->le_birth, le->le_freeze_cksum, LE_GET_CHECKSUM(le),
8939 LE_GET_COMPRESS(le), LE_GET_ARC_COMPRESS(le));
8940
8941 ARCSTAT_INCR(arcstat_l2_lsize, HDR_GET_LSIZE(hdr));
8942 ARCSTAT_INCR(arcstat_l2_psize, arc_hdr_size(hdr));
8943
8944 mutex_enter(&dev->l2ad_mtx);
8945 /*
8946 * We connect the l2hdr to the hdr only after the hdr is in the hash
8947 * table, otherwise the rest of the arc hdr manipulation machinery
8948 * might get confused.
8949 */
8950 list_insert_tail(&dev->l2ad_buflist, hdr);
8951 (void) refcount_add_many(&dev->l2ad_alloc, arc_hdr_size(hdr), hdr);
8952 mutex_exit(&dev->l2ad_mtx);
8953
8954 exists = buf_hash_insert(hdr, &hash_lock);
8955 if (exists) {
8956 /* Buffer was already cached, no need to restore it. */
8957 arc_hdr_destroy(hdr);
8958 mutex_exit(hash_lock);
8959 ARCSTAT_BUMP(arcstat_l2_rebuild_bufs_precached);
8960 return;
8961 }
8962
8963 mutex_exit(hash_lock);
8964 }
8965
8966 /*
8967 * Used by PL2ARC related functions that do
8968 * async read/write
8969 */
8970 static void
8971 pl2arc_io_done(zio_t *zio)
8972 {
8973 abd_put(zio->io_private);
8974 zio->io_private = NULL;
8975 }
8976
8977 /*
8978 * Starts an asynchronous read IO to read a log block. This is used in log
8979 * block reconstruction to start reading the next block before we are done
8980 * decoding and reconstructing the current block, to keep the l2arc device
8981 * nice and hot with read IO to process.
8982 * The returned zio will contain a newly allocated memory buffers for the IO
8983 * data which should then be freed by the caller once the zio is no longer
8984 * needed (i.e. due to it having completed). If you wish to abort this
8985 * zio, you should do so using l2arc_log_blk_prefetch_abort, which takes
8986 * care of disposing of the allocated buffers correctly.
8987 */
8988 static zio_t *
8989 l2arc_log_blk_prefetch(vdev_t *vd, const l2arc_log_blkptr_t *lbp,
8990 uint8_t *lb_buf)
8991 {
8992 uint32_t psize;
8993 zio_t *pio;
8994 abd_t *abd;
8995
8996 psize = LBP_GET_PSIZE(lbp);
8997 ASSERT(psize <= sizeof (l2arc_log_blk_phys_t));
8998 pio = zio_root(vd->vdev_spa, NULL, NULL, ZIO_FLAG_DONT_CACHE |
8999 ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE |
9000 ZIO_FLAG_DONT_RETRY);
9001 abd = abd_get_from_buf(lb_buf, psize);
9002 (void) zio_nowait(zio_read_phys(pio, vd, lbp->lbp_daddr, psize,
9003 abd, ZIO_CHECKSUM_OFF, pl2arc_io_done, abd,
9004 ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_DONT_CACHE | ZIO_FLAG_CANFAIL |
9005 ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY, B_FALSE));
9006
9007 return (pio);
9008 }
9009
9010 /*
9011 * Aborts a zio returned from l2arc_log_blk_prefetch and frees the data
9012 * buffers allocated for it.
9013 */
9014 static void
9015 l2arc_log_blk_prefetch_abort(zio_t *zio)
9016 {
9017 (void) zio_wait(zio);
9018 }
9019
9020 /*
9021 * Creates a zio to update the device header on an l2arc device. The zio is
9022 * initiated as a child of `pio'.
9023 */
9024 static void
9025 l2arc_dev_hdr_update(l2arc_dev_t *dev, zio_t *pio)
9026 {
9027 zio_t *wzio;
9028 abd_t *abd;
9029 l2arc_dev_hdr_phys_t *hdr = dev->l2ad_dev_hdr;
9030 const uint64_t hdr_asize = dev->l2ad_dev_hdr_asize;
9031
9032 hdr->dh_magic = L2ARC_DEV_HDR_MAGIC_V1;
9033 hdr->dh_spa_guid = spa_guid(dev->l2ad_vdev->vdev_spa);
9034 hdr->dh_alloc_space = refcount_count(&dev->l2ad_alloc);
9035 hdr->dh_flags = 0;
9036 if (dev->l2ad_first)
9037 hdr->dh_flags |= L2ARC_DEV_HDR_EVICT_FIRST;
9038
9039 /* checksum operation goes last */
9040 l2arc_dev_hdr_checksum(hdr, &hdr->dh_self_cksum);
9041
9042 abd = abd_get_from_buf(hdr, hdr_asize);
9043 wzio = zio_write_phys(pio, dev->l2ad_vdev, VDEV_LABEL_START_SIZE,
9044 hdr_asize, abd, ZIO_CHECKSUM_OFF, pl2arc_io_done, abd,
9045 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
9046 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
9047 (void) zio_nowait(wzio);
9048 }
9049
9050 /*
9051 * Commits a log block to the L2ARC device. This routine is invoked from
9052 * l2arc_write_buffers when the log block fills up.
9053 * This function allocates some memory to temporarily hold the serialized
9054 * buffer to be written. This is then released in l2arc_write_done.
9055 */
9056 static void
9057 l2arc_log_blk_commit(l2arc_dev_t *dev, zio_t *pio,
9058 l2arc_write_callback_t *cb)
9059 {
9060 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
9061 uint64_t psize, asize;
9062 l2arc_log_blk_buf_t *lb_buf;
9063 abd_t *abd;
9064 zio_t *wzio;
9065
9066 VERIFY(dev->l2ad_log_ent_idx == L2ARC_LOG_BLK_ENTRIES);
9067
9068 /* link the buffer into the block chain */
9069 lb->lb_back2_lbp = dev->l2ad_dev_hdr->dh_start_lbps[1];
9070 lb->lb_magic = L2ARC_LOG_BLK_MAGIC;
9071
9072 /* try to compress the buffer */
9073 lb_buf = kmem_zalloc(sizeof (*lb_buf), KM_SLEEP);
9074 list_insert_tail(&cb->l2wcb_log_blk_buflist, lb_buf);
9075 abd = abd_get_from_buf(lb, sizeof (*lb));
9076 psize = zio_compress_data(ZIO_COMPRESS_LZ4, abd, lb_buf->lbb_log_blk,
9077 sizeof (*lb));
9078 abd_put(abd);
9079 /* a log block is never entirely zero */
9080 ASSERT(psize != 0);
9081 asize = vdev_psize_to_asize(dev->l2ad_vdev, psize);
9082 ASSERT(asize <= sizeof (lb_buf->lbb_log_blk));
9083
9084 /*
9085 * Update the start log blk pointer in the device header to point
9086 * to the log block we're about to write.
9087 */
9088 dev->l2ad_dev_hdr->dh_start_lbps[1] =
9089 dev->l2ad_dev_hdr->dh_start_lbps[0];
9090 dev->l2ad_dev_hdr->dh_start_lbps[0].lbp_daddr = dev->l2ad_hand;
9091 _NOTE(CONSTCOND)
9092 LBP_SET_LSIZE(&dev->l2ad_dev_hdr->dh_start_lbps[0], sizeof (*lb));
9093 LBP_SET_PSIZE(&dev->l2ad_dev_hdr->dh_start_lbps[0], asize);
9094 LBP_SET_CHECKSUM(&dev->l2ad_dev_hdr->dh_start_lbps[0],
9095 ZIO_CHECKSUM_FLETCHER_4);
9096 LBP_SET_TYPE(&dev->l2ad_dev_hdr->dh_start_lbps[0], 0);
9097
9098 if (asize < sizeof (*lb)) {
9099 /* compression succeeded */
9100 bzero(lb_buf->lbb_log_blk + psize, asize - psize);
9101 LBP_SET_COMPRESS(&dev->l2ad_dev_hdr->dh_start_lbps[0],
9102 ZIO_COMPRESS_LZ4);
9103 } else {
9104 /* compression failed */
9105 bcopy(lb, lb_buf->lbb_log_blk, sizeof (*lb));
9106 LBP_SET_COMPRESS(&dev->l2ad_dev_hdr->dh_start_lbps[0],
9107 ZIO_COMPRESS_OFF);
9108 }
9109
9110 /* checksum what we're about to write */
9111 fletcher_4_native(lb_buf->lbb_log_blk, asize,
9112 NULL, &dev->l2ad_dev_hdr->dh_start_lbps[0].lbp_cksum);
9113
9114 /* perform the write itself */
9115 CTASSERT(L2ARC_LOG_BLK_SIZE >= SPA_MINBLOCKSIZE &&
9116 L2ARC_LOG_BLK_SIZE <= SPA_MAXBLOCKSIZE);
9117 abd = abd_get_from_buf(lb_buf->lbb_log_blk, asize);
9118 wzio = zio_write_phys(pio, dev->l2ad_vdev, dev->l2ad_hand,
9119 asize, abd, ZIO_CHECKSUM_OFF, pl2arc_io_done, abd,
9120 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_CANFAIL, B_FALSE);
9121 DTRACE_PROBE2(l2arc__write, vdev_t *, dev->l2ad_vdev, zio_t *, wzio);
9122 (void) zio_nowait(wzio);
9123
9124 dev->l2ad_hand += asize;
9125 vdev_space_update(dev->l2ad_vdev, asize, 0, 0);
9126
9127 /* bump the kstats */
9128 ARCSTAT_INCR(arcstat_l2_write_bytes, asize);
9129 ARCSTAT_BUMP(arcstat_l2_log_blk_writes);
9130 ARCSTAT_F_AVG(arcstat_l2_log_blk_avg_size, asize);
9131 ARCSTAT_F_AVG(arcstat_l2_data_to_meta_ratio,
9132 dev->l2ad_log_blk_payload_asize / asize);
9133
9134 /* start a new log block */
9135 dev->l2ad_log_ent_idx = 0;
9136 dev->l2ad_log_blk_payload_asize = 0;
9137 }
9138
9139 /*
9140 * Validates an L2ARC log blk address to make sure that it can be read
9141 * from the provided L2ARC device. Returns B_TRUE if the address is
9142 * within the device's bounds, or B_FALSE if not.
9143 */
9144 static boolean_t
9145 l2arc_log_blkptr_valid(l2arc_dev_t *dev, const l2arc_log_blkptr_t *lbp)
9146 {
9147 uint64_t psize = LBP_GET_PSIZE(lbp);
9148 uint64_t end = lbp->lbp_daddr + psize;
9149
9150 /*
9151 * A log block is valid if all of the following conditions are true:
9152 * - it fits entirely between l2ad_start and l2ad_end
9153 * - it has a valid size
9154 */
9155 return (lbp->lbp_daddr >= dev->l2ad_start && end <= dev->l2ad_end &&
9156 psize > 0 && psize <= sizeof (l2arc_log_blk_phys_t));
9157 }
9158
9159 /*
9160 * Computes the checksum of `hdr' and stores it in `cksum'.
9161 */
9162 static void
9163 l2arc_dev_hdr_checksum(const l2arc_dev_hdr_phys_t *hdr, zio_cksum_t *cksum)
9164 {
9165 fletcher_4_native((uint8_t *)hdr +
9166 offsetof(l2arc_dev_hdr_phys_t, dh_spa_guid),
9167 sizeof (*hdr) - offsetof(l2arc_dev_hdr_phys_t, dh_spa_guid),
9168 NULL, cksum);
9169 }
9170
9171 /*
9172 * Inserts ARC buffer `ab' into the current L2ARC log blk on the device.
9173 * The buffer being inserted must be present in L2ARC.
9174 * Returns B_TRUE if the L2ARC log blk is full and needs to be committed
9175 * to L2ARC, or B_FALSE if it still has room for more ARC buffers.
9176 */
9177 static boolean_t
9178 l2arc_log_blk_insert(l2arc_dev_t *dev, const arc_buf_hdr_t *ab)
9179 {
9180 l2arc_log_blk_phys_t *lb = &dev->l2ad_log_blk;
9181 l2arc_log_ent_phys_t *le;
9182 int index = dev->l2ad_log_ent_idx++;
9183
9184 ASSERT(index < L2ARC_LOG_BLK_ENTRIES);
9185
9186 le = &lb->lb_entries[index];
9187 bzero(le, sizeof (*le));
9188 le->le_dva = ab->b_dva;
9189 le->le_birth = ab->b_birth;
9190 le->le_daddr = ab->b_l2hdr.b_daddr;
9191 LE_SET_LSIZE(le, HDR_GET_LSIZE(ab));
9192 LE_SET_PSIZE(le, HDR_GET_PSIZE(ab));
9193
9194 if ((ab->b_flags & ARC_FLAG_COMPRESSED_ARC) != 0) {
9195 LE_SET_ARC_COMPRESS(le, 1);
9196 LE_SET_COMPRESS(le, HDR_GET_COMPRESS(ab));
9197 } else {
9198 ASSERT3U(HDR_GET_COMPRESS(ab), ==, ZIO_COMPRESS_OFF);
9199 LE_SET_ARC_COMPRESS(le, 0);
9200 LE_SET_COMPRESS(le, ZIO_COMPRESS_OFF);
9201 }
9202
9203 if (ab->b_freeze_cksum != NULL) {
9204 le->le_freeze_cksum = *ab->b_freeze_cksum;
9205 LE_SET_CHECKSUM(le, ZIO_CHECKSUM_FLETCHER_2);
9206 } else {
9207 LE_SET_CHECKSUM(le, ZIO_CHECKSUM_OFF);
9208 }
9209
9210 LE_SET_TYPE(le, arc_flags_to_bufc(ab->b_flags));
9211 dev->l2ad_log_blk_payload_asize += arc_hdr_size((arc_buf_hdr_t *)ab);
9212
9213 return (dev->l2ad_log_ent_idx == L2ARC_LOG_BLK_ENTRIES);
9214 }
9215
9216 /*
9217 * Checks whether a given L2ARC device address sits in a time-sequential
9218 * range. The trick here is that the L2ARC is a rotary buffer, so we can't
9219 * just do a range comparison, we need to handle the situation in which the
9220 * range wraps around the end of the L2ARC device. Arguments:
9221 * bottom Lower end of the range to check (written to earlier).
9222 * top Upper end of the range to check (written to later).
9223 * check The address for which we want to determine if it sits in
9224 * between the top and bottom.
9225 *
9226 * The 3-way conditional below represents the following cases:
9227 *
9228 * bottom < top : Sequentially ordered case:
9229 * <check>--------+-------------------+
9230 * | (overlap here?) |
9231 * L2ARC dev V V
9232 * |---------------<bottom>============<top>--------------|
9233 *
9234 * bottom > top: Looped-around case:
9235 * <check>--------+------------------+
9236 * | (overlap here?) |
9237 * L2ARC dev V V
9238 * |===============<top>---------------<bottom>===========|
9239 * ^ ^
9240 * | (or here?) |
9241 * +---------------+---------<check>
9242 *
9243 * top == bottom : Just a single address comparison.
9244 */
9245 static inline boolean_t
9246 l2arc_range_check_overlap(uint64_t bottom, uint64_t top, uint64_t check)
9247 {
9248 if (bottom < top)
9249 return (bottom <= check && check <= top);
9250 else if (bottom > top)
9251 return (check <= top || bottom <= check);
9252 else
9253 return (check == top);
9254 }
|