Print this page
5056 ZFS deadlock on db_mtx and dn_holds
Reviewed by: Will Andrews <willa@spectralogic.com>
Reviewed by: Matt Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Approved by: Dan McDonald <danmcd@omniti.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/sys/dmu.h
+++ new/usr/src/uts/common/fs/zfs/sys/dmu.h
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
|
↓ open down ↓ |
18 lines elided |
↑ open up ↑ |
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 25 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
26 26 * Copyright (c) 2012, Joyent, Inc. All rights reserved.
27 27 * Copyright 2013 DEY Storage Systems, Inc.
28 28 * Copyright 2014 HybridCluster. All rights reserved.
29 + * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 30 */
30 31
31 32 /* Portions Copyright 2010 Robert Milkowski */
32 33
33 34 #ifndef _SYS_DMU_H
34 35 #define _SYS_DMU_H
35 36
36 37 /*
37 38 * This file describes the interface that the DMU provides for its
38 39 * consumers.
39 40 *
40 41 * The DMU also interacts with the SPA. That interface is described in
41 42 * dmu_spa.h.
42 43 */
43 44
45 +#include <sys/zfs_context.h>
44 46 #include <sys/inttypes.h>
45 -#include <sys/types.h>
46 -#include <sys/param.h>
47 47 #include <sys/cred.h>
48 -#include <sys/time.h>
49 48 #include <sys/fs/zfs.h>
50 49
51 50 #ifdef __cplusplus
52 51 extern "C" {
53 52 #endif
54 53
55 54 struct uio;
56 55 struct xuio;
57 56 struct page;
58 57 struct vnode;
59 58 struct spa;
60 59 struct zilog;
61 60 struct zio;
62 61 struct blkptr;
63 62 struct zap_cursor;
64 63 struct dsl_dataset;
65 64 struct dsl_pool;
66 65 struct dnode;
67 66 struct drr_begin;
68 67 struct drr_end;
69 68 struct zbookmark_phys;
70 69 struct spa;
71 70 struct nvlist;
72 71 struct arc_buf;
73 72 struct zio_prop;
74 73 struct sa_handle;
75 74
76 75 typedef struct objset objset_t;
77 76 typedef struct dmu_tx dmu_tx_t;
78 77 typedef struct dsl_dir dsl_dir_t;
79 78
80 79 typedef enum dmu_object_byteswap {
81 80 DMU_BSWAP_UINT8,
82 81 DMU_BSWAP_UINT16,
83 82 DMU_BSWAP_UINT32,
84 83 DMU_BSWAP_UINT64,
85 84 DMU_BSWAP_ZAP,
86 85 DMU_BSWAP_DNODE,
87 86 DMU_BSWAP_OBJSET,
88 87 DMU_BSWAP_ZNODE,
89 88 DMU_BSWAP_OLDACL,
90 89 DMU_BSWAP_ACL,
91 90 /*
92 91 * Allocating a new byteswap type number makes the on-disk format
93 92 * incompatible with any other format that uses the same number.
94 93 *
95 94 * Data can usually be structured to work with one of the
96 95 * DMU_BSWAP_UINT* or DMU_BSWAP_ZAP types.
97 96 */
98 97 DMU_BSWAP_NUMFUNCS
99 98 } dmu_object_byteswap_t;
100 99
101 100 #define DMU_OT_NEWTYPE 0x80
102 101 #define DMU_OT_METADATA 0x40
103 102 #define DMU_OT_BYTESWAP_MASK 0x3f
104 103
105 104 /*
106 105 * Defines a uint8_t object type. Object types specify if the data
107 106 * in the object is metadata (boolean) and how to byteswap the data
108 107 * (dmu_object_byteswap_t).
109 108 */
110 109 #define DMU_OT(byteswap, metadata) \
111 110 (DMU_OT_NEWTYPE | \
112 111 ((metadata) ? DMU_OT_METADATA : 0) | \
113 112 ((byteswap) & DMU_OT_BYTESWAP_MASK))
114 113
115 114 #define DMU_OT_IS_VALID(ot) (((ot) & DMU_OT_NEWTYPE) ? \
116 115 ((ot) & DMU_OT_BYTESWAP_MASK) < DMU_BSWAP_NUMFUNCS : \
117 116 (ot) < DMU_OT_NUMTYPES)
118 117
119 118 #define DMU_OT_IS_METADATA(ot) (((ot) & DMU_OT_NEWTYPE) ? \
120 119 ((ot) & DMU_OT_METADATA) : \
121 120 dmu_ot[(ot)].ot_metadata)
122 121
123 122 /*
124 123 * These object types use bp_fill != 1 for their L0 bp's. Therefore they can't
125 124 * have their data embedded (i.e. use a BP_IS_EMBEDDED() bp), because bp_fill
126 125 * is repurposed for embedded BPs.
127 126 */
128 127 #define DMU_OT_HAS_FILL(ot) \
129 128 ((ot) == DMU_OT_DNODE || (ot) == DMU_OT_OBJSET)
130 129
131 130 #define DMU_OT_BYTESWAP(ot) (((ot) & DMU_OT_NEWTYPE) ? \
132 131 ((ot) & DMU_OT_BYTESWAP_MASK) : \
133 132 dmu_ot[(ot)].ot_byteswap)
134 133
135 134 typedef enum dmu_object_type {
136 135 DMU_OT_NONE,
137 136 /* general: */
138 137 DMU_OT_OBJECT_DIRECTORY, /* ZAP */
139 138 DMU_OT_OBJECT_ARRAY, /* UINT64 */
140 139 DMU_OT_PACKED_NVLIST, /* UINT8 (XDR by nvlist_pack/unpack) */
141 140 DMU_OT_PACKED_NVLIST_SIZE, /* UINT64 */
142 141 DMU_OT_BPOBJ, /* UINT64 */
143 142 DMU_OT_BPOBJ_HDR, /* UINT64 */
144 143 /* spa: */
145 144 DMU_OT_SPACE_MAP_HEADER, /* UINT64 */
146 145 DMU_OT_SPACE_MAP, /* UINT64 */
147 146 /* zil: */
148 147 DMU_OT_INTENT_LOG, /* UINT64 */
149 148 /* dmu: */
150 149 DMU_OT_DNODE, /* DNODE */
151 150 DMU_OT_OBJSET, /* OBJSET */
152 151 /* dsl: */
153 152 DMU_OT_DSL_DIR, /* UINT64 */
154 153 DMU_OT_DSL_DIR_CHILD_MAP, /* ZAP */
155 154 DMU_OT_DSL_DS_SNAP_MAP, /* ZAP */
156 155 DMU_OT_DSL_PROPS, /* ZAP */
157 156 DMU_OT_DSL_DATASET, /* UINT64 */
158 157 /* zpl: */
159 158 DMU_OT_ZNODE, /* ZNODE */
160 159 DMU_OT_OLDACL, /* Old ACL */
161 160 DMU_OT_PLAIN_FILE_CONTENTS, /* UINT8 */
162 161 DMU_OT_DIRECTORY_CONTENTS, /* ZAP */
163 162 DMU_OT_MASTER_NODE, /* ZAP */
164 163 DMU_OT_UNLINKED_SET, /* ZAP */
165 164 /* zvol: */
166 165 DMU_OT_ZVOL, /* UINT8 */
167 166 DMU_OT_ZVOL_PROP, /* ZAP */
168 167 /* other; for testing only! */
169 168 DMU_OT_PLAIN_OTHER, /* UINT8 */
170 169 DMU_OT_UINT64_OTHER, /* UINT64 */
171 170 DMU_OT_ZAP_OTHER, /* ZAP */
172 171 /* new object types: */
173 172 DMU_OT_ERROR_LOG, /* ZAP */
174 173 DMU_OT_SPA_HISTORY, /* UINT8 */
175 174 DMU_OT_SPA_HISTORY_OFFSETS, /* spa_his_phys_t */
176 175 DMU_OT_POOL_PROPS, /* ZAP */
177 176 DMU_OT_DSL_PERMS, /* ZAP */
178 177 DMU_OT_ACL, /* ACL */
179 178 DMU_OT_SYSACL, /* SYSACL */
180 179 DMU_OT_FUID, /* FUID table (Packed NVLIST UINT8) */
181 180 DMU_OT_FUID_SIZE, /* FUID table size UINT64 */
182 181 DMU_OT_NEXT_CLONES, /* ZAP */
183 182 DMU_OT_SCAN_QUEUE, /* ZAP */
184 183 DMU_OT_USERGROUP_USED, /* ZAP */
185 184 DMU_OT_USERGROUP_QUOTA, /* ZAP */
186 185 DMU_OT_USERREFS, /* ZAP */
187 186 DMU_OT_DDT_ZAP, /* ZAP */
188 187 DMU_OT_DDT_STATS, /* ZAP */
189 188 DMU_OT_SA, /* System attr */
190 189 DMU_OT_SA_MASTER_NODE, /* ZAP */
191 190 DMU_OT_SA_ATTR_REGISTRATION, /* ZAP */
192 191 DMU_OT_SA_ATTR_LAYOUTS, /* ZAP */
193 192 DMU_OT_SCAN_XLATE, /* ZAP */
194 193 DMU_OT_DEDUP, /* fake dedup BP from ddt_bp_create() */
195 194 DMU_OT_DEADLIST, /* ZAP */
196 195 DMU_OT_DEADLIST_HDR, /* UINT64 */
197 196 DMU_OT_DSL_CLONES, /* ZAP */
198 197 DMU_OT_BPOBJ_SUBOBJ, /* UINT64 */
199 198 /*
200 199 * Do not allocate new object types here. Doing so makes the on-disk
201 200 * format incompatible with any other format that uses the same object
202 201 * type number.
203 202 *
204 203 * When creating an object which does not have one of the above types
205 204 * use the DMU_OTN_* type with the correct byteswap and metadata
206 205 * values.
207 206 *
208 207 * The DMU_OTN_* types do not have entries in the dmu_ot table,
209 208 * use the DMU_OT_IS_METDATA() and DMU_OT_BYTESWAP() macros instead
210 209 * of indexing into dmu_ot directly (this works for both DMU_OT_* types
211 210 * and DMU_OTN_* types).
212 211 */
213 212 DMU_OT_NUMTYPES,
214 213
215 214 /*
216 215 * Names for valid types declared with DMU_OT().
217 216 */
218 217 DMU_OTN_UINT8_DATA = DMU_OT(DMU_BSWAP_UINT8, B_FALSE),
219 218 DMU_OTN_UINT8_METADATA = DMU_OT(DMU_BSWAP_UINT8, B_TRUE),
220 219 DMU_OTN_UINT16_DATA = DMU_OT(DMU_BSWAP_UINT16, B_FALSE),
221 220 DMU_OTN_UINT16_METADATA = DMU_OT(DMU_BSWAP_UINT16, B_TRUE),
222 221 DMU_OTN_UINT32_DATA = DMU_OT(DMU_BSWAP_UINT32, B_FALSE),
223 222 DMU_OTN_UINT32_METADATA = DMU_OT(DMU_BSWAP_UINT32, B_TRUE),
224 223 DMU_OTN_UINT64_DATA = DMU_OT(DMU_BSWAP_UINT64, B_FALSE),
225 224 DMU_OTN_UINT64_METADATA = DMU_OT(DMU_BSWAP_UINT64, B_TRUE),
226 225 DMU_OTN_ZAP_DATA = DMU_OT(DMU_BSWAP_ZAP, B_FALSE),
227 226 DMU_OTN_ZAP_METADATA = DMU_OT(DMU_BSWAP_ZAP, B_TRUE),
228 227 } dmu_object_type_t;
229 228
230 229 typedef enum txg_how {
231 230 TXG_WAIT = 1,
232 231 TXG_NOWAIT,
233 232 TXG_WAITED,
234 233 } txg_how_t;
235 234
236 235 void byteswap_uint64_array(void *buf, size_t size);
237 236 void byteswap_uint32_array(void *buf, size_t size);
238 237 void byteswap_uint16_array(void *buf, size_t size);
239 238 void byteswap_uint8_array(void *buf, size_t size);
240 239 void zap_byteswap(void *buf, size_t size);
241 240 void zfs_oldacl_byteswap(void *buf, size_t size);
242 241 void zfs_acl_byteswap(void *buf, size_t size);
243 242 void zfs_znode_byteswap(void *buf, size_t size);
244 243
245 244 #define DS_FIND_SNAPSHOTS (1<<0)
246 245 #define DS_FIND_CHILDREN (1<<1)
247 246
248 247 /*
249 248 * The maximum number of bytes that can be accessed as part of one
250 249 * operation, including metadata.
251 250 */
252 251 #define DMU_MAX_ACCESS (32 * 1024 * 1024) /* 32MB */
253 252 #define DMU_MAX_DELETEBLKCNT (20480) /* ~5MB of indirect blocks */
254 253
255 254 #define DMU_USERUSED_OBJECT (-1ULL)
256 255 #define DMU_GROUPUSED_OBJECT (-2ULL)
257 256
258 257 /*
259 258 * artificial blkids for bonus buffer and spill blocks
260 259 */
261 260 #define DMU_BONUS_BLKID (-1ULL)
262 261 #define DMU_SPILL_BLKID (-2ULL)
263 262 /*
264 263 * Public routines to create, destroy, open, and close objsets.
265 264 */
266 265 int dmu_objset_hold(const char *name, void *tag, objset_t **osp);
267 266 int dmu_objset_own(const char *name, dmu_objset_type_t type,
268 267 boolean_t readonly, void *tag, objset_t **osp);
269 268 void dmu_objset_rele(objset_t *os, void *tag);
270 269 void dmu_objset_disown(objset_t *os, void *tag);
271 270 int dmu_objset_open_ds(struct dsl_dataset *ds, objset_t **osp);
272 271
273 272 void dmu_objset_evict_dbufs(objset_t *os);
274 273 int dmu_objset_create(const char *name, dmu_objset_type_t type, uint64_t flags,
275 274 void (*func)(objset_t *os, void *arg, cred_t *cr, dmu_tx_t *tx), void *arg);
276 275 int dmu_objset_clone(const char *name, const char *origin);
277 276 int dsl_destroy_snapshots_nvl(struct nvlist *snaps, boolean_t defer,
278 277 struct nvlist *errlist);
279 278 int dmu_objset_snapshot_one(const char *fsname, const char *snapname);
280 279 int dmu_objset_snapshot_tmp(const char *, const char *, int);
281 280 int dmu_objset_find(char *name, int func(const char *, void *), void *arg,
282 281 int flags);
283 282 void dmu_objset_byteswap(void *buf, size_t size);
|
↓ open down ↓ |
225 lines elided |
↑ open up ↑ |
284 283 int dsl_dataset_rename_snapshot(const char *fsname,
285 284 const char *oldsnapname, const char *newsnapname, boolean_t recursive);
286 285
287 286 typedef struct dmu_buf {
288 287 uint64_t db_object; /* object that this buffer is part of */
289 288 uint64_t db_offset; /* byte offset in this object */
290 289 uint64_t db_size; /* size of buffer in bytes */
291 290 void *db_data; /* data in buffer */
292 291 } dmu_buf_t;
293 292
294 -typedef void dmu_buf_evict_func_t(struct dmu_buf *db, void *user_ptr);
295 -
296 293 /*
297 294 * The names of zap entries in the DIRECTORY_OBJECT of the MOS.
298 295 */
299 296 #define DMU_POOL_DIRECTORY_OBJECT 1
300 297 #define DMU_POOL_CONFIG "config"
301 298 #define DMU_POOL_FEATURES_FOR_WRITE "features_for_write"
302 299 #define DMU_POOL_FEATURES_FOR_READ "features_for_read"
303 300 #define DMU_POOL_FEATURE_DESCRIPTIONS "feature_descriptions"
304 301 #define DMU_POOL_FEATURE_ENABLED_TXG "feature_enabled_txg"
305 302 #define DMU_POOL_ROOT_DATASET "root_dataset"
306 303 #define DMU_POOL_SYNC_BPOBJ "sync_bplist"
307 304 #define DMU_POOL_ERRLOG_SCRUB "errlog_scrub"
308 305 #define DMU_POOL_ERRLOG_LAST "errlog_last"
309 306 #define DMU_POOL_SPARES "spares"
310 307 #define DMU_POOL_DEFLATE "deflate"
311 308 #define DMU_POOL_HISTORY "history"
312 309 #define DMU_POOL_PROPS "pool_props"
313 310 #define DMU_POOL_L2CACHE "l2cache"
314 311 #define DMU_POOL_TMP_USERREFS "tmp_userrefs"
315 312 #define DMU_POOL_DDT "DDT-%s-%s-%s"
316 313 #define DMU_POOL_DDT_STATS "DDT-statistics"
317 314 #define DMU_POOL_CREATION_VERSION "creation_version"
318 315 #define DMU_POOL_SCAN "scan"
319 316 #define DMU_POOL_FREE_BPOBJ "free_bpobj"
320 317 #define DMU_POOL_BPTREE_OBJ "bptree_obj"
321 318 #define DMU_POOL_EMPTY_BPOBJ "empty_bpobj"
322 319
323 320 /*
324 321 * Allocate an object from this objset. The range of object numbers
325 322 * available is (0, DN_MAX_OBJECT). Object 0 is the meta-dnode.
326 323 *
327 324 * The transaction must be assigned to a txg. The newly allocated
328 325 * object will be "held" in the transaction (ie. you can modify the
329 326 * newly allocated object in this transaction).
330 327 *
331 328 * dmu_object_alloc() chooses an object and returns it in *objectp.
332 329 *
333 330 * dmu_object_claim() allocates a specific object number. If that
334 331 * number is already allocated, it fails and returns EEXIST.
335 332 *
336 333 * Return 0 on success, or ENOSPC or EEXIST as specified above.
337 334 */
338 335 uint64_t dmu_object_alloc(objset_t *os, dmu_object_type_t ot,
339 336 int blocksize, dmu_object_type_t bonus_type, int bonus_len, dmu_tx_t *tx);
340 337 int dmu_object_claim(objset_t *os, uint64_t object, dmu_object_type_t ot,
341 338 int blocksize, dmu_object_type_t bonus_type, int bonus_len, dmu_tx_t *tx);
342 339 int dmu_object_reclaim(objset_t *os, uint64_t object, dmu_object_type_t ot,
343 340 int blocksize, dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *txp);
344 341
345 342 /*
346 343 * Free an object from this objset.
347 344 *
348 345 * The object's data will be freed as well (ie. you don't need to call
349 346 * dmu_free(object, 0, -1, tx)).
350 347 *
351 348 * The object need not be held in the transaction.
352 349 *
353 350 * If there are any holds on this object's buffers (via dmu_buf_hold()),
354 351 * or tx holds on the object (via dmu_tx_hold_object()), you can not
355 352 * free it; it fails and returns EBUSY.
356 353 *
357 354 * If the object is not allocated, it fails and returns ENOENT.
358 355 *
359 356 * Return 0 on success, or EBUSY or ENOENT as specified above.
360 357 */
361 358 int dmu_object_free(objset_t *os, uint64_t object, dmu_tx_t *tx);
362 359
363 360 /*
364 361 * Find the next allocated or free object.
365 362 *
366 363 * The objectp parameter is in-out. It will be updated to be the next
367 364 * object which is allocated. Ignore objects which have not been
368 365 * modified since txg.
369 366 *
370 367 * XXX Can only be called on a objset with no dirty data.
371 368 *
372 369 * Returns 0 on success, or ENOENT if there are no more objects.
373 370 */
374 371 int dmu_object_next(objset_t *os, uint64_t *objectp,
375 372 boolean_t hole, uint64_t txg);
376 373
377 374 /*
378 375 * Set the data blocksize for an object.
379 376 *
380 377 * The object cannot have any blocks allcated beyond the first. If
381 378 * the first block is allocated already, the new size must be greater
382 379 * than the current block size. If these conditions are not met,
383 380 * ENOTSUP will be returned.
384 381 *
385 382 * Returns 0 on success, or EBUSY if there are any holds on the object
386 383 * contents, or ENOTSUP as described above.
387 384 */
388 385 int dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size,
389 386 int ibs, dmu_tx_t *tx);
390 387
391 388 /*
392 389 * Set the checksum property on a dnode. The new checksum algorithm will
393 390 * apply to all newly written blocks; existing blocks will not be affected.
394 391 */
395 392 void dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
396 393 dmu_tx_t *tx);
397 394
398 395 /*
399 396 * Set the compress property on a dnode. The new compression algorithm will
400 397 * apply to all newly written blocks; existing blocks will not be affected.
401 398 */
402 399 void dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
403 400 dmu_tx_t *tx);
404 401
405 402 void
406 403 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
407 404 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
408 405 int compressed_size, int byteorder, dmu_tx_t *tx);
409 406
410 407 /*
411 408 * Decide how to write a block: checksum, compression, number of copies, etc.
412 409 */
413 410 #define WP_NOFILL 0x1
414 411 #define WP_DMU_SYNC 0x2
415 412 #define WP_SPILL 0x4
416 413
417 414 void dmu_write_policy(objset_t *os, struct dnode *dn, int level, int wp,
418 415 struct zio_prop *zp);
419 416 /*
420 417 * The bonus data is accessed more or less like a regular buffer.
421 418 * You must dmu_bonus_hold() to get the buffer, which will give you a
422 419 * dmu_buf_t with db_offset==-1ULL, and db_size = the size of the bonus
423 420 * data. As with any normal buffer, you must call dmu_buf_read() to
424 421 * read db_data, dmu_buf_will_dirty() before modifying it, and the
425 422 * object must be held in an assigned transaction before calling
426 423 * dmu_buf_will_dirty. You may use dmu_buf_set_user() on the bonus
427 424 * buffer as well. You must release your hold with dmu_buf_rele().
428 425 *
429 426 * Returns ENOENT, EIO, or 0.
430 427 */
431 428 int dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **);
432 429 int dmu_bonus_max(void);
433 430 int dmu_set_bonus(dmu_buf_t *, int, dmu_tx_t *);
434 431 int dmu_set_bonustype(dmu_buf_t *, dmu_object_type_t, dmu_tx_t *);
435 432 dmu_object_type_t dmu_get_bonustype(dmu_buf_t *);
436 433 int dmu_rm_spill(objset_t *, uint64_t, dmu_tx_t *);
437 434
438 435 /*
439 436 * Special spill buffer support used by "SA" framework
440 437 */
441 438
442 439 int dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp);
443 440 int dmu_spill_hold_by_dnode(struct dnode *dn, uint32_t flags,
444 441 void *tag, dmu_buf_t **dbp);
445 442 int dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp);
446 443
447 444 /*
448 445 * Obtain the DMU buffer from the specified object which contains the
449 446 * specified offset. dmu_buf_hold() puts a "hold" on the buffer, so
450 447 * that it will remain in memory. You must release the hold with
451 448 * dmu_buf_rele(). You musn't access the dmu_buf_t after releasing your
452 449 * hold. You must have a hold on any dmu_buf_t* you pass to the DMU.
453 450 *
454 451 * You must call dmu_buf_read, dmu_buf_will_dirty, or dmu_buf_will_fill
455 452 * on the returned buffer before reading or writing the buffer's
456 453 * db_data. The comments for those routines describe what particular
457 454 * operations are valid after calling them.
458 455 *
459 456 * The object number must be a valid, allocated object number.
460 457 */
461 458 int dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
462 459 void *tag, dmu_buf_t **, int flags);
463 460 void dmu_buf_add_ref(dmu_buf_t *db, void* tag);
464 461 void dmu_buf_rele(dmu_buf_t *db, void *tag);
465 462 uint64_t dmu_buf_refcount(dmu_buf_t *db);
466 463
467 464 /*
468 465 * dmu_buf_hold_array holds the DMU buffers which contain all bytes in a
469 466 * range of an object. A pointer to an array of dmu_buf_t*'s is
470 467 * returned (in *dbpp).
|
↓ open down ↓ |
165 lines elided |
↑ open up ↑ |
471 468 *
472 469 * dmu_buf_rele_array releases the hold on an array of dmu_buf_t*'s, and
473 470 * frees the array. The hold on the array of buffers MUST be released
474 471 * with dmu_buf_rele_array. You can NOT release the hold on each buffer
475 472 * individually with dmu_buf_rele.
476 473 */
477 474 int dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
478 475 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp);
479 476 void dmu_buf_rele_array(dmu_buf_t **, int numbufs, void *tag);
480 477
478 +typedef void dmu_buf_evict_func_t(void *user_ptr);
479 +
481 480 /*
482 - * Returns NULL on success, or the existing user ptr if it's already
483 - * been set.
481 + * A DMU buffer user object may be associated with a dbuf for the
482 + * duration of its lifetime. This allows the user of a dbuf (client)
483 + * to attach private data to a dbuf (e.g. in-core only data such as a
484 + * dnode_children_t, zap_t, or zap_leaf_t) and be optionally notified
485 + * when that dbuf has been evicted. Clients typically respond to the
486 + * eviction notification by freeing their private data, thus ensuring
487 + * the same lifetime for both dbuf and private data.
484 488 *
485 - * user_ptr is for use by the user and can be obtained via dmu_buf_get_user().
489 + * The mapping from a dmu_buf_user_t to any client private data is the
490 + * client's responsibility. All current consumers of the API with private
491 + * data embed a dmu_buf_user_t as the first member of the structure for
492 + * their private data. This allows conversions between the two types
493 + * with a simple cast. Since the DMU buf user API never needs access
494 + * to the private data, other strategies can be employed if necessary
495 + * or convenient for the client (e.g. using container_of() to do the
496 + * conversion for private data that cannot have the dmu_buf_user_t as
497 + * its first member).
486 498 *
487 - * If non-NULL, pageout func will be called when this buffer is being
488 - * excised from the cache, so that you can clean up the data structure
489 - * pointed to by user_ptr.
499 + * Eviction callbacks are executed without the dbuf mutex held or any
500 + * other type of mechanism to guarantee that the dbuf is still available.
501 + * For this reason, users must assume the dbuf has already been freed
502 + * and not reference the dbuf from the callback context.
490 503 *
491 - * dmu_evict_user() will call the pageout func for all buffers in a
492 - * objset with a given pageout func.
504 + * Users requesting "immediate eviction" are notified as soon as the dbuf
505 + * is only referenced by dirty records (dirties == holds). Otherwise the
506 + * notification occurs after eviction processing for the dbuf begins.
493 507 */
494 -void *dmu_buf_set_user(dmu_buf_t *db, void *user_ptr,
495 - dmu_buf_evict_func_t *pageout_func);
508 +typedef struct dmu_buf_user {
509 + /*
510 + * Asynchronous user eviction callback state.
511 + */
512 + taskq_ent_t dbu_tqent;
513 +
514 + /* This instance's eviction function pointer. */
515 + dmu_buf_evict_func_t *dbu_evict_func;
516 +#ifdef ZFS_DEBUG
517 + /*
518 + * Pointer to user's dbuf pointer. NULL for clients that do
519 + * not associate a dbuf with their user data.
520 + *
521 + * The dbuf pointer is cleared upon eviction so as to catch
522 + * use-after-evict bugs in clients.
523 + */
524 + dmu_buf_t **dbu_clear_on_evict_dbufp;
525 +#endif
526 +} dmu_buf_user_t;
527 +
496 528 /*
497 - * set_user_ie is the same as set_user, but request immediate eviction
498 - * when hold count goes to zero.
529 + * Initialize the given dmu_buf_user_t instance with the eviction function
530 + * evict_func, to be called when the user is evicted.
531 + *
532 + * NOTE: This function should only be called once on a given dmu_buf_user_t.
533 + * To allow enforcement of this, dbu must already be zeroed on entry.
499 534 */
500 -void *dmu_buf_set_user_ie(dmu_buf_t *db, void *user_ptr,
501 - dmu_buf_evict_func_t *pageout_func);
502 -void *dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr,
503 - void *user_ptr, dmu_buf_evict_func_t *pageout_func);
504 -void dmu_evict_user(objset_t *os, dmu_buf_evict_func_t *func);
535 +#ifdef __lint
536 +/* Very ugly, but it beats issuing suppression directives in many Makefiles. */
537 +extern void
538 +dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func,
539 + dmu_buf_t **clear_on_evict_dbufp);
540 +#else /* __lint */
541 +inline void
542 +dmu_buf_init_user(dmu_buf_user_t *dbu, dmu_buf_evict_func_t *evict_func,
543 + dmu_buf_t **clear_on_evict_dbufp)
544 +{
545 + ASSERT(dbu->dbu_evict_func == NULL);
546 + ASSERT(evict_func != NULL);
547 + dbu->dbu_evict_func = evict_func;
548 +#ifdef ZFS_DEBUG
549 + dbu->dbu_clear_on_evict_dbufp = clear_on_evict_dbufp;
550 +#endif
551 +}
552 +#endif /* __lint */
505 553
506 554 /*
507 - * Returns the user_ptr set with dmu_buf_set_user(), or NULL if not set.
555 + * Attach user data to a dbuf and mark it for normal (when the dbuf's
556 + * data is cleared or its reference count goes to zero) eviction processing.
557 + *
558 + * Returns NULL on success, or the existing user if another user currently
559 + * owns the buffer.
508 560 */
561 +void *dmu_buf_set_user(dmu_buf_t *db, dmu_buf_user_t *user);
562 +
563 +/*
564 + * Attach user data to a dbuf and mark it for immediate (its dirty and
565 + * reference counts are equal) eviction processing.
566 + *
567 + * Returns NULL on success, or the existing user if another user currently
568 + * owns the buffer.
569 + */
570 +void *dmu_buf_set_user_ie(dmu_buf_t *db, dmu_buf_user_t *user);
571 +
572 +/*
573 + * Replace the current user of a dbuf.
574 + *
575 + * If given the current user of a dbuf, replaces the dbuf's user with
576 + * "new_user" and returns the user data pointer that was replaced.
577 + * Otherwise returns the current, and unmodified, dbuf user pointer.
578 + */
579 +void *dmu_buf_replace_user(dmu_buf_t *db,
580 + dmu_buf_user_t *old_user, dmu_buf_user_t *new_user);
581 +
582 +/*
583 + * Remove the specified user data for a DMU buffer.
584 + *
585 + * Returns the user that was removed on success, or the current user if
586 + * another user currently owns the buffer.
587 + */
588 +void *dmu_buf_remove_user(dmu_buf_t *db, dmu_buf_user_t *user);
589 +
590 +/*
591 + * Returns the user data (dmu_buf_user_t *) associated with this dbuf.
592 + */
509 593 void *dmu_buf_get_user(dmu_buf_t *db);
510 594
595 +/* Block until any in-progress dmu buf user evictions complete. */
596 +void dmu_buf_user_evict_wait(void);
597 +
511 598 /*
512 599 * Returns the blkptr associated with this dbuf, or NULL if not set.
513 600 */
514 601 struct blkptr *dmu_buf_get_blkptr(dmu_buf_t *db);
515 602
516 603 /*
517 604 * Indicate that you are going to modify the buffer's data (db_data).
518 605 *
519 606 * The transaction (tx) must be assigned to a txg (ie. you've called
520 607 * dmu_tx_assign()). The buffer's object must be held in the tx
521 608 * (ie. you've called dmu_tx_hold_object(tx, db->db_object)).
522 609 */
523 610 void dmu_buf_will_dirty(dmu_buf_t *db, dmu_tx_t *tx);
524 611
525 612 /*
526 613 * Tells if the given dbuf is freeable.
527 614 */
528 615 boolean_t dmu_buf_freeable(dmu_buf_t *);
529 616
530 617 /*
531 618 * You must create a transaction, then hold the objects which you will
532 619 * (or might) modify as part of this transaction. Then you must assign
533 620 * the transaction to a transaction group. Once the transaction has
534 621 * been assigned, you can modify buffers which belong to held objects as
535 622 * part of this transaction. You can't modify buffers before the
536 623 * transaction has been assigned; you can't modify buffers which don't
537 624 * belong to objects which this transaction holds; you can't hold
538 625 * objects once the transaction has been assigned. You may hold an
539 626 * object which you are going to free (with dmu_object_free()), but you
540 627 * don't have to.
541 628 *
542 629 * You can abort the transaction before it has been assigned.
543 630 *
544 631 * Note that you may hold buffers (with dmu_buf_hold) at any time,
545 632 * regardless of transaction state.
546 633 */
547 634
548 635 #define DMU_NEW_OBJECT (-1ULL)
549 636 #define DMU_OBJECT_END (-1ULL)
550 637
551 638 dmu_tx_t *dmu_tx_create(objset_t *os);
552 639 void dmu_tx_hold_write(dmu_tx_t *tx, uint64_t object, uint64_t off, int len);
553 640 void dmu_tx_hold_free(dmu_tx_t *tx, uint64_t object, uint64_t off,
554 641 uint64_t len);
555 642 void dmu_tx_hold_zap(dmu_tx_t *tx, uint64_t object, int add, const char *name);
556 643 void dmu_tx_hold_bonus(dmu_tx_t *tx, uint64_t object);
557 644 void dmu_tx_hold_spill(dmu_tx_t *tx, uint64_t object);
558 645 void dmu_tx_hold_sa(dmu_tx_t *tx, struct sa_handle *hdl, boolean_t may_grow);
559 646 void dmu_tx_hold_sa_create(dmu_tx_t *tx, int total_size);
560 647 void dmu_tx_abort(dmu_tx_t *tx);
561 648 int dmu_tx_assign(dmu_tx_t *tx, enum txg_how txg_how);
562 649 void dmu_tx_wait(dmu_tx_t *tx);
563 650 void dmu_tx_commit(dmu_tx_t *tx);
564 651 void dmu_tx_mark_netfree(dmu_tx_t *tx);
565 652
566 653 /*
567 654 * To register a commit callback, dmu_tx_callback_register() must be called.
568 655 *
569 656 * dcb_data is a pointer to caller private data that is passed on as a
570 657 * callback parameter. The caller is responsible for properly allocating and
571 658 * freeing it.
572 659 *
573 660 * When registering a callback, the transaction must be already created, but
574 661 * it cannot be committed or aborted. It can be assigned to a txg or not.
575 662 *
576 663 * The callback will be called after the transaction has been safely written
577 664 * to stable storage and will also be called if the dmu_tx is aborted.
578 665 * If there is any error which prevents the transaction from being committed to
579 666 * disk, the callback will be called with a value of error != 0.
580 667 */
581 668 typedef void dmu_tx_callback_func_t(void *dcb_data, int error);
582 669
583 670 void dmu_tx_callback_register(dmu_tx_t *tx, dmu_tx_callback_func_t *dcb_func,
584 671 void *dcb_data);
585 672
586 673 /*
587 674 * Free up the data blocks for a defined range of a file. If size is
588 675 * -1, the range from offset to end-of-file is freed.
589 676 */
590 677 int dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
591 678 uint64_t size, dmu_tx_t *tx);
592 679 int dmu_free_long_range(objset_t *os, uint64_t object, uint64_t offset,
593 680 uint64_t size);
594 681 int dmu_free_long_object(objset_t *os, uint64_t object);
595 682
596 683 /*
597 684 * Convenience functions.
598 685 *
599 686 * Canfail routines will return 0 on success, or an errno if there is a
600 687 * nonrecoverable I/O error.
601 688 */
602 689 #define DMU_READ_PREFETCH 0 /* prefetch */
603 690 #define DMU_READ_NO_PREFETCH 1 /* don't prefetch */
604 691 int dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
605 692 void *buf, uint32_t flags);
606 693 void dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
607 694 const void *buf, dmu_tx_t *tx);
608 695 void dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
609 696 dmu_tx_t *tx);
610 697 int dmu_read_uio(objset_t *os, uint64_t object, struct uio *uio, uint64_t size);
611 698 int dmu_read_uio_dbuf(dmu_buf_t *zdb, struct uio *uio, uint64_t size);
612 699 int dmu_write_uio(objset_t *os, uint64_t object, struct uio *uio, uint64_t size,
613 700 dmu_tx_t *tx);
614 701 int dmu_write_uio_dbuf(dmu_buf_t *zdb, struct uio *uio, uint64_t size,
615 702 dmu_tx_t *tx);
616 703 int dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset,
617 704 uint64_t size, struct page *pp, dmu_tx_t *tx);
618 705 struct arc_buf *dmu_request_arcbuf(dmu_buf_t *handle, int size);
619 706 void dmu_return_arcbuf(struct arc_buf *buf);
620 707 void dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, struct arc_buf *buf,
621 708 dmu_tx_t *tx);
622 709 int dmu_xuio_init(struct xuio *uio, int niov);
623 710 void dmu_xuio_fini(struct xuio *uio);
624 711 int dmu_xuio_add(struct xuio *uio, struct arc_buf *abuf, offset_t off,
625 712 size_t n);
626 713 int dmu_xuio_cnt(struct xuio *uio);
627 714 struct arc_buf *dmu_xuio_arcbuf(struct xuio *uio, int i);
628 715 void dmu_xuio_clear(struct xuio *uio, int i);
629 716 void xuio_stat_wbuf_copied();
630 717 void xuio_stat_wbuf_nocopy();
631 718
632 719 extern int zfs_prefetch_disable;
633 720 extern int zfs_max_recordsize;
634 721
635 722 /*
636 723 * Asynchronously try to read in the data.
637 724 */
638 725 void dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset,
639 726 uint64_t len);
640 727
641 728 typedef struct dmu_object_info {
642 729 /* All sizes are in bytes unless otherwise indicated. */
643 730 uint32_t doi_data_block_size;
644 731 uint32_t doi_metadata_block_size;
645 732 dmu_object_type_t doi_type;
646 733 dmu_object_type_t doi_bonus_type;
647 734 uint64_t doi_bonus_size;
648 735 uint8_t doi_indirection; /* 2 = dnode->indirect->data */
649 736 uint8_t doi_checksum;
650 737 uint8_t doi_compress;
651 738 uint8_t doi_nblkptr;
652 739 uint8_t doi_pad[4];
653 740 uint64_t doi_physical_blocks_512; /* data + metadata, 512b blks */
654 741 uint64_t doi_max_offset;
655 742 uint64_t doi_fill_count; /* number of non-empty blocks */
656 743 } dmu_object_info_t;
657 744
658 745 typedef void arc_byteswap_func_t(void *buf, size_t size);
659 746
660 747 typedef struct dmu_object_type_info {
661 748 dmu_object_byteswap_t ot_byteswap;
662 749 boolean_t ot_metadata;
663 750 char *ot_name;
664 751 } dmu_object_type_info_t;
665 752
666 753 typedef struct dmu_object_byteswap_info {
667 754 arc_byteswap_func_t *ob_func;
668 755 char *ob_name;
669 756 } dmu_object_byteswap_info_t;
670 757
671 758 extern const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES];
672 759 extern const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS];
673 760
674 761 /*
675 762 * Get information on a DMU object.
676 763 *
677 764 * Return 0 on success or ENOENT if object is not allocated.
678 765 *
679 766 * If doi is NULL, just indicates whether the object exists.
680 767 */
681 768 int dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi);
682 769 /* Like dmu_object_info, but faster if you have a held dnode in hand. */
683 770 void dmu_object_info_from_dnode(struct dnode *dn, dmu_object_info_t *doi);
684 771 /* Like dmu_object_info, but faster if you have a held dbuf in hand. */
685 772 void dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi);
686 773 /*
687 774 * Like dmu_object_info_from_db, but faster still when you only care about
688 775 * the size. This is specifically optimized for zfs_getattr().
689 776 */
690 777 void dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize,
691 778 u_longlong_t *nblk512);
692 779
693 780 typedef struct dmu_objset_stats {
694 781 uint64_t dds_num_clones; /* number of clones of this */
695 782 uint64_t dds_creation_txg;
696 783 uint64_t dds_guid;
697 784 dmu_objset_type_t dds_type;
698 785 uint8_t dds_is_snapshot;
699 786 uint8_t dds_inconsistent;
700 787 char dds_origin[MAXNAMELEN];
701 788 } dmu_objset_stats_t;
702 789
703 790 /*
704 791 * Get stats on a dataset.
705 792 */
706 793 void dmu_objset_fast_stat(objset_t *os, dmu_objset_stats_t *stat);
707 794
708 795 /*
709 796 * Add entries to the nvlist for all the objset's properties. See
710 797 * zfs_prop_table[] and zfs(1m) for details on the properties.
711 798 */
712 799 void dmu_objset_stats(objset_t *os, struct nvlist *nv);
713 800
714 801 /*
715 802 * Get the space usage statistics for statvfs().
716 803 *
717 804 * refdbytes is the amount of space "referenced" by this objset.
718 805 * availbytes is the amount of space available to this objset, taking
719 806 * into account quotas & reservations, assuming that no other objsets
720 807 * use the space first. These values correspond to the 'referenced' and
721 808 * 'available' properties, described in the zfs(1m) manpage.
722 809 *
723 810 * usedobjs and availobjs are the number of objects currently allocated,
724 811 * and available.
725 812 */
726 813 void dmu_objset_space(objset_t *os, uint64_t *refdbytesp, uint64_t *availbytesp,
727 814 uint64_t *usedobjsp, uint64_t *availobjsp);
728 815
729 816 /*
730 817 * The fsid_guid is a 56-bit ID that can change to avoid collisions.
731 818 * (Contrast with the ds_guid which is a 64-bit ID that will never
732 819 * change, so there is a small probability that it will collide.)
733 820 */
734 821 uint64_t dmu_objset_fsid_guid(objset_t *os);
735 822
736 823 /*
737 824 * Get the [cm]time for an objset's snapshot dir
738 825 */
739 826 timestruc_t dmu_objset_snap_cmtime(objset_t *os);
740 827
741 828 int dmu_objset_is_snapshot(objset_t *os);
742 829
743 830 extern struct spa *dmu_objset_spa(objset_t *os);
744 831 extern struct zilog *dmu_objset_zil(objset_t *os);
745 832 extern struct dsl_pool *dmu_objset_pool(objset_t *os);
746 833 extern struct dsl_dataset *dmu_objset_ds(objset_t *os);
747 834 extern void dmu_objset_name(objset_t *os, char *buf);
748 835 extern dmu_objset_type_t dmu_objset_type(objset_t *os);
749 836 extern uint64_t dmu_objset_id(objset_t *os);
750 837 extern zfs_sync_type_t dmu_objset_syncprop(objset_t *os);
751 838 extern zfs_logbias_op_t dmu_objset_logbias(objset_t *os);
752 839 extern int dmu_snapshot_list_next(objset_t *os, int namelen, char *name,
753 840 uint64_t *id, uint64_t *offp, boolean_t *case_conflict);
754 841 extern int dmu_snapshot_realname(objset_t *os, char *name, char *real,
755 842 int maxlen, boolean_t *conflict);
756 843 extern int dmu_dir_list_next(objset_t *os, int namelen, char *name,
757 844 uint64_t *idp, uint64_t *offp);
758 845
759 846 typedef int objset_used_cb_t(dmu_object_type_t bonustype,
760 847 void *bonus, uint64_t *userp, uint64_t *groupp);
761 848 extern void dmu_objset_register_type(dmu_objset_type_t ost,
762 849 objset_used_cb_t *cb);
763 850 extern void dmu_objset_set_user(objset_t *os, void *user_ptr);
764 851 extern void *dmu_objset_get_user(objset_t *os);
765 852
766 853 /*
767 854 * Return the txg number for the given assigned transaction.
768 855 */
769 856 uint64_t dmu_tx_get_txg(dmu_tx_t *tx);
770 857
771 858 /*
772 859 * Synchronous write.
773 860 * If a parent zio is provided this function initiates a write on the
774 861 * provided buffer as a child of the parent zio.
775 862 * In the absence of a parent zio, the write is completed synchronously.
776 863 * At write completion, blk is filled with the bp of the written block.
777 864 * Note that while the data covered by this function will be on stable
778 865 * storage when the write completes this new data does not become a
779 866 * permanent part of the file until the associated transaction commits.
780 867 */
781 868
782 869 /*
783 870 * {zfs,zvol,ztest}_get_done() args
784 871 */
785 872 typedef struct zgd {
786 873 struct zilog *zgd_zilog;
787 874 struct blkptr *zgd_bp;
788 875 dmu_buf_t *zgd_db;
789 876 struct rl *zgd_rl;
790 877 void *zgd_private;
791 878 } zgd_t;
792 879
793 880 typedef void dmu_sync_cb_t(zgd_t *arg, int error);
794 881 int dmu_sync(struct zio *zio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd);
795 882
796 883 /*
797 884 * Find the next hole or data block in file starting at *off
798 885 * Return found offset in *off. Return ESRCH for end of file.
799 886 */
800 887 int dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole,
801 888 uint64_t *off);
802 889
803 890 /*
804 891 * Initial setup and final teardown.
805 892 */
806 893 extern void dmu_init(void);
807 894 extern void dmu_fini(void);
808 895
809 896 typedef void (*dmu_traverse_cb_t)(objset_t *os, void *arg, struct blkptr *bp,
810 897 uint64_t object, uint64_t offset, int len);
811 898 void dmu_traverse_objset(objset_t *os, uint64_t txg_start,
812 899 dmu_traverse_cb_t cb, void *arg);
813 900
814 901 int dmu_diff(const char *tosnap_name, const char *fromsnap_name,
815 902 struct vnode *vp, offset_t *offp);
816 903
817 904 /* CRC64 table */
818 905 #define ZFS_CRC64_POLY 0xC96C5795D7870F42ULL /* ECMA-182, reflected form */
819 906 extern uint64_t zfs_crc64_table[256];
820 907
821 908 extern int zfs_mdcomp_disable;
822 909
823 910 #ifdef __cplusplus
824 911 }
825 912 #endif
826 913
827 914 #endif /* _SYS_DMU_H */
|
↓ open down ↓ |
307 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX