1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
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) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
25 * Copyright 2018, Joyent, Inc.
26 */
27
28 /*
29 * Kernel memory allocator, as described in the following two papers and a
30 * statement about the consolidator:
31 *
32 * Jeff Bonwick,
33 * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
34 * Proceedings of the Summer 1994 Usenix Conference.
35 * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
36 *
37 * Jeff Bonwick and Jonathan Adams,
38 * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
39 * Arbitrary Resources.
40 * Proceedings of the 2001 Usenix Conference.
41 * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
42 *
43 * kmem Slab Consolidator Big Theory Statement:
44 *
45 * 1. Motivation
46 *
47 * As stated in Bonwick94, slabs provide the following advantages over other
48 * allocation structures in terms of memory fragmentation:
49 *
50 * - Internal fragmentation (per-buffer wasted space) is minimal.
51 * - Severe external fragmentation (unused buffers on the free list) is
52 * unlikely.
53 *
54 * Segregating objects by size eliminates one source of external fragmentation,
55 * and according to Bonwick:
56 *
57 * The other reason that slabs reduce external fragmentation is that all
58 * objects in a slab are of the same type, so they have the same lifetime
59 * distribution. The resulting segregation of short-lived and long-lived
60 * objects at slab granularity reduces the likelihood of an entire page being
61 * held hostage due to a single long-lived allocation [Barrett93, Hanson90].
62 *
63 * While unlikely, severe external fragmentation remains possible. Clients that
64 * allocate both short- and long-lived objects from the same cache cannot
65 * anticipate the distribution of long-lived objects within the allocator's slab
66 * implementation. Even a small percentage of long-lived objects distributed
67 * randomly across many slabs can lead to a worst case scenario where the client
68 * frees the majority of its objects and the system gets back almost none of the
69 * slabs. Despite the client doing what it reasonably can to help the system
70 * reclaim memory, the allocator cannot shake free enough slabs because of
71 * lonely allocations stubbornly hanging on. Although the allocator is in a
72 * position to diagnose the fragmentation, there is nothing that the allocator
73 * by itself can do about it. It only takes a single allocated object to prevent
74 * an entire slab from being reclaimed, and any object handed out by
75 * kmem_cache_alloc() is by definition in the client's control. Conversely,
76 * although the client is in a position to move a long-lived object, it has no
77 * way of knowing if the object is causing fragmentation, and if so, where to
78 * move it. A solution necessarily requires further cooperation between the
79 * allocator and the client.
80 *
81 * 2. Move Callback
82 *
83 * The kmem slab consolidator therefore adds a move callback to the
84 * allocator/client interface, improving worst-case external fragmentation in
85 * kmem caches that supply a function to move objects from one memory location
86 * to another. In a situation of low memory kmem attempts to consolidate all of
87 * a cache's slabs at once; otherwise it works slowly to bring external
88 * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
89 * thereby helping to avoid a low memory situation in the future.
90 *
91 * The callback has the following signature:
92 *
93 * kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
94 *
95 * It supplies the kmem client with two addresses: the allocated object that
96 * kmem wants to move and a buffer selected by kmem for the client to use as the
97 * copy destination. The callback is kmem's way of saying "Please get off of
98 * this buffer and use this one instead." kmem knows where it wants to move the
99 * object in order to best reduce fragmentation. All the client needs to know
100 * about the second argument (void *new) is that it is an allocated, constructed
101 * object ready to take the contents of the old object. When the move function
102 * is called, the system is likely to be low on memory, and the new object
103 * spares the client from having to worry about allocating memory for the
104 * requested move. The third argument supplies the size of the object, in case a
105 * single move function handles multiple caches whose objects differ only in
106 * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
107 * user argument passed to the constructor, destructor, and reclaim functions is
108 * also passed to the move callback.
109 *
110 * 2.1 Setting the Move Callback
111 *
112 * The client sets the move callback after creating the cache and before
113 * allocating from it:
114 *
115 * object_cache = kmem_cache_create(...);
116 * kmem_cache_set_move(object_cache, object_move);
117 *
118 * 2.2 Move Callback Return Values
119 *
120 * Only the client knows about its own data and when is a good time to move it.
121 * The client is cooperating with kmem to return unused memory to the system,
122 * and kmem respectfully accepts this help at the client's convenience. When
123 * asked to move an object, the client can respond with any of the following:
124 *
125 * typedef enum kmem_cbrc {
126 * KMEM_CBRC_YES,
127 * KMEM_CBRC_NO,
128 * KMEM_CBRC_LATER,
129 * KMEM_CBRC_DONT_NEED,
130 * KMEM_CBRC_DONT_KNOW
131 * } kmem_cbrc_t;
132 *
133 * The client must not explicitly kmem_cache_free() either of the objects passed
134 * to the callback, since kmem wants to free them directly to the slab layer
135 * (bypassing the per-CPU magazine layer). The response tells kmem which of the
136 * objects to free:
137 *
138 * YES: (Did it) The client moved the object, so kmem frees the old one.
139 * NO: (Never) The client refused, so kmem frees the new object (the
140 * unused copy destination). kmem also marks the slab of the old
141 * object so as not to bother the client with further callbacks for
142 * that object as long as the slab remains on the partial slab list.
143 * (The system won't be getting the slab back as long as the
144 * immovable object holds it hostage, so there's no point in moving
145 * any of its objects.)
146 * LATER: The client is using the object and cannot move it now, so kmem
147 * frees the new object (the unused copy destination). kmem still
148 * attempts to move other objects off the slab, since it expects to
149 * succeed in clearing the slab in a later callback. The client
150 * should use LATER instead of NO if the object is likely to become
151 * movable very soon.
152 * DONT_NEED: The client no longer needs the object, so kmem frees the old along
153 * with the new object (the unused copy destination). This response
154 * is the client's opportunity to be a model citizen and give back as
155 * much as it can.
156 * DONT_KNOW: The client does not know about the object because
157 * a) the client has just allocated the object and not yet put it
158 * wherever it expects to find known objects
159 * b) the client has removed the object from wherever it expects to
160 * find known objects and is about to free it, or
161 * c) the client has freed the object.
162 * In all these cases (a, b, and c) kmem frees the new object (the
163 * unused copy destination). In the first case, the object is in
164 * use and the correct action is that for LATER; in the latter two
165 * cases, we know that the object is either freed or about to be
166 * freed, in which case it is either already in a magazine or about
167 * to be in one. In these cases, we know that the object will either
168 * be reallocated and reused, or it will end up in a full magazine
169 * that will be reaped (thereby liberating the slab). Because it
170 * is prohibitively expensive to differentiate these cases, and
171 * because the defrag code is executed when we're low on memory
172 * (thereby biasing the system to reclaim full magazines) we treat
173 * all DONT_KNOW cases as LATER and rely on cache reaping to
174 * generally clean up full magazines. While we take the same action
175 * for these cases, we maintain their semantic distinction: if
176 * defragmentation is not occurring, it is useful to know if this
177 * is due to objects in use (LATER) or objects in an unknown state
178 * of transition (DONT_KNOW).
179 *
180 * 2.3 Object States
181 *
182 * Neither kmem nor the client can be assumed to know the object's whereabouts
183 * at the time of the callback. An object belonging to a kmem cache may be in
184 * any of the following states:
185 *
186 * 1. Uninitialized on the slab
187 * 2. Allocated from the slab but not constructed (still uninitialized)
188 * 3. Allocated from the slab, constructed, but not yet ready for business
189 * (not in a valid state for the move callback)
190 * 4. In use (valid and known to the client)
191 * 5. About to be freed (no longer in a valid state for the move callback)
192 * 6. Freed to a magazine (still constructed)
193 * 7. Allocated from a magazine, not yet ready for business (not in a valid
194 * state for the move callback), and about to return to state #4
195 * 8. Deconstructed on a magazine that is about to be freed
196 * 9. Freed to the slab
197 *
198 * Since the move callback may be called at any time while the object is in any
199 * of the above states (except state #1), the client needs a safe way to
200 * determine whether or not it knows about the object. Specifically, the client
201 * needs to know whether or not the object is in state #4, the only state in
202 * which a move is valid. If the object is in any other state, the client should
203 * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
204 * the object's fields.
205 *
206 * Note that although an object may be in state #4 when kmem initiates the move
207 * request, the object may no longer be in that state by the time kmem actually
208 * calls the move function. Not only does the client free objects
209 * asynchronously, kmem itself puts move requests on a queue where thay are
210 * pending until kmem processes them from another context. Also, objects freed
211 * to a magazine appear allocated from the point of view of the slab layer, so
212 * kmem may even initiate requests for objects in a state other than state #4.
213 *
214 * 2.3.1 Magazine Layer
215 *
216 * An important insight revealed by the states listed above is that the magazine
217 * layer is populated only by kmem_cache_free(). Magazines of constructed
218 * objects are never populated directly from the slab layer (which contains raw,
219 * unconstructed objects). Whenever an allocation request cannot be satisfied
220 * from the magazine layer, the magazines are bypassed and the request is
221 * satisfied from the slab layer (creating a new slab if necessary). kmem calls
222 * the object constructor only when allocating from the slab layer, and only in
223 * response to kmem_cache_alloc() or to prepare the destination buffer passed in
224 * the move callback. kmem does not preconstruct objects in anticipation of
225 * kmem_cache_alloc().
226 *
227 * 2.3.2 Object Constructor and Destructor
228 *
229 * If the client supplies a destructor, it must be valid to call the destructor
230 * on a newly created object (immediately after the constructor).
231 *
232 * 2.4 Recognizing Known Objects
233 *
234 * There is a simple test to determine safely whether or not the client knows
235 * about a given object in the move callback. It relies on the fact that kmem
236 * guarantees that the object of the move callback has only been touched by the
237 * client itself or else by kmem. kmem does this by ensuring that none of the
238 * cache's slabs are freed to the virtual memory (VM) subsystem while a move
239 * callback is pending. When the last object on a slab is freed, if there is a
240 * pending move, kmem puts the slab on a per-cache dead list and defers freeing
241 * slabs on that list until all pending callbacks are completed. That way,
242 * clients can be certain that the object of a move callback is in one of the
243 * states listed above, making it possible to distinguish known objects (in
244 * state #4) using the two low order bits of any pointer member (with the
245 * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
246 * platforms).
247 *
248 * The test works as long as the client always transitions objects from state #4
249 * (known, in use) to state #5 (about to be freed, invalid) by setting the low
250 * order bit of the client-designated pointer member. Since kmem only writes
251 * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
252 * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
253 * guaranteed to set at least one of the two low order bits. Therefore, given an
254 * object with a back pointer to a 'container_t *o_container', the client can
255 * test
256 *
257 * container_t *container = object->o_container;
258 * if ((uintptr_t)container & 0x3) {
259 * return (KMEM_CBRC_DONT_KNOW);
260 * }
261 *
262 * Typically, an object will have a pointer to some structure with a list or
263 * hash where objects from the cache are kept while in use. Assuming that the
264 * client has some way of knowing that the container structure is valid and will
265 * not go away during the move, and assuming that the structure includes a lock
266 * to protect whatever collection is used, then the client would continue as
267 * follows:
268 *
269 * // Ensure that the container structure does not go away.
270 * if (container_hold(container) == 0) {
271 * return (KMEM_CBRC_DONT_KNOW);
272 * }
273 * mutex_enter(&container->c_objects_lock);
274 * if (container != object->o_container) {
275 * mutex_exit(&container->c_objects_lock);
276 * container_rele(container);
277 * return (KMEM_CBRC_DONT_KNOW);
278 * }
279 *
280 * At this point the client knows that the object cannot be freed as long as
281 * c_objects_lock is held. Note that after acquiring the lock, the client must
282 * recheck the o_container pointer in case the object was removed just before
283 * acquiring the lock.
284 *
285 * When the client is about to free an object, it must first remove that object
286 * from the list, hash, or other structure where it is kept. At that time, to
287 * mark the object so it can be distinguished from the remaining, known objects,
288 * the client sets the designated low order bit:
289 *
290 * mutex_enter(&container->c_objects_lock);
291 * object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
292 * list_remove(&container->c_objects, object);
293 * mutex_exit(&container->c_objects_lock);
294 *
295 * In the common case, the object is freed to the magazine layer, where it may
296 * be reused on a subsequent allocation without the overhead of calling the
297 * constructor. While in the magazine it appears allocated from the point of
298 * view of the slab layer, making it a candidate for the move callback. Most
299 * objects unrecognized by the client in the move callback fall into this
300 * category and are cheaply distinguished from known objects by the test
301 * described earlier. Because searching magazines is prohibitively expensive
302 * for kmem, clients that do not mark freed objects (and therefore return
303 * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
304 * efficacy reduced.
305 *
306 * Invalidating the designated pointer member before freeing the object marks
307 * the object to be avoided in the callback, and conversely, assigning a valid
308 * value to the designated pointer member after allocating the object makes the
309 * object fair game for the callback:
310 *
311 * ... allocate object ...
312 * ... set any initial state not set by the constructor ...
313 *
314 * mutex_enter(&container->c_objects_lock);
315 * list_insert_tail(&container->c_objects, object);
316 * membar_producer();
317 * object->o_container = container;
318 * mutex_exit(&container->c_objects_lock);
319 *
320 * Note that everything else must be valid before setting o_container makes the
321 * object fair game for the move callback. The membar_producer() call ensures
322 * that all the object's state is written to memory before setting the pointer
323 * that transitions the object from state #3 or #7 (allocated, constructed, not
324 * yet in use) to state #4 (in use, valid). That's important because the move
325 * function has to check the validity of the pointer before it can safely
326 * acquire the lock protecting the collection where it expects to find known
327 * objects.
328 *
329 * This method of distinguishing known objects observes the usual symmetry:
330 * invalidating the designated pointer is the first thing the client does before
331 * freeing the object, and setting the designated pointer is the last thing the
332 * client does after allocating the object. Of course, the client is not
333 * required to use this method. Fundamentally, how the client recognizes known
334 * objects is completely up to the client, but this method is recommended as an
335 * efficient and safe way to take advantage of the guarantees made by kmem. If
336 * the entire object is arbitrary data without any markable bits from a suitable
337 * pointer member, then the client must find some other method, such as
338 * searching a hash table of known objects.
339 *
340 * 2.5 Preventing Objects From Moving
341 *
342 * Besides a way to distinguish known objects, the other thing that the client
343 * needs is a strategy to ensure that an object will not move while the client
344 * is actively using it. The details of satisfying this requirement tend to be
345 * highly cache-specific. It might seem that the same rules that let a client
346 * remove an object safely should also decide when an object can be moved
347 * safely. However, any object state that makes a removal attempt invalid is
348 * likely to be long-lasting for objects that the client does not expect to
349 * remove. kmem knows nothing about the object state and is equally likely (from
350 * the client's point of view) to request a move for any object in the cache,
351 * whether prepared for removal or not. Even a low percentage of objects stuck
352 * in place by unremovability will defeat the consolidator if the stuck objects
353 * are the same long-lived allocations likely to hold slabs hostage.
354 * Fundamentally, the consolidator is not aimed at common cases. Severe external
355 * fragmentation is a worst case scenario manifested as sparsely allocated
356 * slabs, by definition a low percentage of the cache's objects. When deciding
357 * what makes an object movable, keep in mind the goal of the consolidator: to
358 * bring worst-case external fragmentation within the limits guaranteed for
359 * internal fragmentation. Removability is a poor criterion if it is likely to
360 * exclude more than an insignificant percentage of objects for long periods of
361 * time.
362 *
363 * A tricky general solution exists, and it has the advantage of letting you
364 * move any object at almost any moment, practically eliminating the likelihood
365 * that an object can hold a slab hostage. However, if there is a cache-specific
366 * way to ensure that an object is not actively in use in the vast majority of
367 * cases, a simpler solution that leverages this cache-specific knowledge is
368 * preferred.
369 *
370 * 2.5.1 Cache-Specific Solution
371 *
372 * As an example of a cache-specific solution, the ZFS znode cache takes
373 * advantage of the fact that the vast majority of znodes are only being
374 * referenced from the DNLC. (A typical case might be a few hundred in active
375 * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
376 * client has established that it recognizes the znode and can access its fields
377 * safely (using the method described earlier), it then tests whether the znode
378 * is referenced by anything other than the DNLC. If so, it assumes that the
379 * znode may be in active use and is unsafe to move, so it drops its locks and
380 * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
381 * else znodes are used, no change is needed to protect against the possibility
382 * of the znode moving. The disadvantage is that it remains possible for an
383 * application to hold a znode slab hostage with an open file descriptor.
384 * However, this case ought to be rare and the consolidator has a way to deal
385 * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
386 * object, kmem eventually stops believing it and treats the slab as if the
387 * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
388 * then focus on getting it off of the partial slab list by allocating rather
389 * than freeing all of its objects. (Either way of getting a slab off the
390 * free list reduces fragmentation.)
391 *
392 * 2.5.2 General Solution
393 *
394 * The general solution, on the other hand, requires an explicit hold everywhere
395 * the object is used to prevent it from moving. To keep the client locking
396 * strategy as uncomplicated as possible, kmem guarantees the simplifying
397 * assumption that move callbacks are sequential, even across multiple caches.
398 * Internally, a global queue processed by a single thread supports all caches
399 * implementing the callback function. No matter how many caches supply a move
400 * function, the consolidator never moves more than one object at a time, so the
401 * client does not have to worry about tricky lock ordering involving several
402 * related objects from different kmem caches.
403 *
404 * The general solution implements the explicit hold as a read-write lock, which
405 * allows multiple readers to access an object from the cache simultaneously
406 * while a single writer is excluded from moving it. A single rwlock for the
407 * entire cache would lock out all threads from using any of the cache's objects
408 * even though only a single object is being moved, so to reduce contention,
409 * the client can fan out the single rwlock into an array of rwlocks hashed by
410 * the object address, making it probable that moving one object will not
411 * prevent other threads from using a different object. The rwlock cannot be a
412 * member of the object itself, because the possibility of the object moving
413 * makes it unsafe to access any of the object's fields until the lock is
414 * acquired.
415 *
416 * Assuming a small, fixed number of locks, it's possible that multiple objects
417 * will hash to the same lock. A thread that needs to use multiple objects in
418 * the same function may acquire the same lock multiple times. Since rwlocks are
419 * reentrant for readers, and since there is never more than a single writer at
420 * a time (assuming that the client acquires the lock as a writer only when
421 * moving an object inside the callback), there would seem to be no problem.
422 * However, a client locking multiple objects in the same function must handle
423 * one case of potential deadlock: Assume that thread A needs to prevent both
424 * object 1 and object 2 from moving, and thread B, the callback, meanwhile
425 * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
426 * same lock, that thread A will acquire the lock for object 1 as a reader
427 * before thread B sets the lock's write-wanted bit, preventing thread A from
428 * reacquiring the lock for object 2 as a reader. Unable to make forward
429 * progress, thread A will never release the lock for object 1, resulting in
430 * deadlock.
431 *
432 * There are two ways of avoiding the deadlock just described. The first is to
433 * use rw_tryenter() rather than rw_enter() in the callback function when
434 * attempting to acquire the lock as a writer. If tryenter discovers that the
435 * same object (or another object hashed to the same lock) is already in use, it
436 * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
437 * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
438 * since it allows a thread to acquire the lock as a reader in spite of a
439 * waiting writer. This second approach insists on moving the object now, no
440 * matter how many readers the move function must wait for in order to do so,
441 * and could delay the completion of the callback indefinitely (blocking
442 * callbacks to other clients). In practice, a less insistent callback using
443 * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
444 * little reason to use anything else.
445 *
446 * Avoiding deadlock is not the only problem that an implementation using an
447 * explicit hold needs to solve. Locking the object in the first place (to
448 * prevent it from moving) remains a problem, since the object could move
449 * between the time you obtain a pointer to the object and the time you acquire
450 * the rwlock hashed to that pointer value. Therefore the client needs to
451 * recheck the value of the pointer after acquiring the lock, drop the lock if
452 * the value has changed, and try again. This requires a level of indirection:
453 * something that points to the object rather than the object itself, that the
454 * client can access safely while attempting to acquire the lock. (The object
455 * itself cannot be referenced safely because it can move at any time.)
456 * The following lock-acquisition function takes whatever is safe to reference
457 * (arg), follows its pointer to the object (using function f), and tries as
458 * often as necessary to acquire the hashed lock and verify that the object
459 * still has not moved:
460 *
461 * object_t *
462 * object_hold(object_f f, void *arg)
463 * {
464 * object_t *op;
465 *
466 * op = f(arg);
467 * if (op == NULL) {
468 * return (NULL);
469 * }
470 *
471 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
472 * while (op != f(arg)) {
473 * rw_exit(OBJECT_RWLOCK(op));
474 * op = f(arg);
475 * if (op == NULL) {
476 * break;
477 * }
478 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
479 * }
480 *
481 * return (op);
482 * }
483 *
484 * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
485 * lock reacquisition loop, while necessary, almost never executes. The function
486 * pointer f (used to obtain the object pointer from arg) has the following type
487 * definition:
488 *
489 * typedef object_t *(*object_f)(void *arg);
490 *
491 * An object_f implementation is likely to be as simple as accessing a structure
492 * member:
493 *
494 * object_t *
495 * s_object(void *arg)
496 * {
497 * something_t *sp = arg;
498 * return (sp->s_object);
499 * }
500 *
501 * The flexibility of a function pointer allows the path to the object to be
502 * arbitrarily complex and also supports the notion that depending on where you
503 * are using the object, you may need to get it from someplace different.
504 *
505 * The function that releases the explicit hold is simpler because it does not
506 * have to worry about the object moving:
507 *
508 * void
509 * object_rele(object_t *op)
510 * {
511 * rw_exit(OBJECT_RWLOCK(op));
512 * }
513 *
514 * The caller is spared these details so that obtaining and releasing an
515 * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
516 * of object_hold() only needs to know that the returned object pointer is valid
517 * if not NULL and that the object will not move until released.
518 *
519 * Although object_hold() prevents an object from moving, it does not prevent it
520 * from being freed. The caller must take measures before calling object_hold()
521 * (afterwards is too late) to ensure that the held object cannot be freed. The
522 * caller must do so without accessing the unsafe object reference, so any lock
523 * or reference count used to ensure the continued existence of the object must
524 * live outside the object itself.
525 *
526 * Obtaining a new object is a special case where an explicit hold is impossible
527 * for the caller. Any function that returns a newly allocated object (either as
528 * a return value, or as an in-out paramter) must return it already held; after
529 * the caller gets it is too late, since the object cannot be safely accessed
530 * without the level of indirection described earlier. The following
531 * object_alloc() example uses the same code shown earlier to transition a new
532 * object into the state of being recognized (by the client) as a known object.
533 * The function must acquire the hold (rw_enter) before that state transition
534 * makes the object movable:
535 *
536 * static object_t *
537 * object_alloc(container_t *container)
538 * {
539 * object_t *object = kmem_cache_alloc(object_cache, 0);
540 * ... set any initial state not set by the constructor ...
541 * rw_enter(OBJECT_RWLOCK(object), RW_READER);
542 * mutex_enter(&container->c_objects_lock);
543 * list_insert_tail(&container->c_objects, object);
544 * membar_producer();
545 * object->o_container = container;
546 * mutex_exit(&container->c_objects_lock);
547 * return (object);
548 * }
549 *
550 * Functions that implicitly acquire an object hold (any function that calls
551 * object_alloc() to supply an object for the caller) need to be carefully noted
552 * so that the matching object_rele() is not neglected. Otherwise, leaked holds
553 * prevent all objects hashed to the affected rwlocks from ever being moved.
554 *
555 * The pointer to a held object can be hashed to the holding rwlock even after
556 * the object has been freed. Although it is possible to release the hold
557 * after freeing the object, you may decide to release the hold implicitly in
558 * whatever function frees the object, so as to release the hold as soon as
559 * possible, and for the sake of symmetry with the function that implicitly
560 * acquires the hold when it allocates the object. Here, object_free() releases
561 * the hold acquired by object_alloc(). Its implicit object_rele() forms a
562 * matching pair with object_hold():
563 *
564 * void
565 * object_free(object_t *object)
566 * {
567 * container_t *container;
568 *
569 * ASSERT(object_held(object));
570 * container = object->o_container;
571 * mutex_enter(&container->c_objects_lock);
572 * object->o_container =
573 * (void *)((uintptr_t)object->o_container | 0x1);
574 * list_remove(&container->c_objects, object);
575 * mutex_exit(&container->c_objects_lock);
576 * object_rele(object);
577 * kmem_cache_free(object_cache, object);
578 * }
579 *
580 * Note that object_free() cannot safely accept an object pointer as an argument
581 * unless the object is already held. Any function that calls object_free()
582 * needs to be carefully noted since it similarly forms a matching pair with
583 * object_hold().
584 *
585 * To complete the picture, the following callback function implements the
586 * general solution by moving objects only if they are currently unheld:
587 *
588 * static kmem_cbrc_t
589 * object_move(void *buf, void *newbuf, size_t size, void *arg)
590 * {
591 * object_t *op = buf, *np = newbuf;
592 * container_t *container;
593 *
594 * container = op->o_container;
595 * if ((uintptr_t)container & 0x3) {
596 * return (KMEM_CBRC_DONT_KNOW);
597 * }
598 *
599 * // Ensure that the container structure does not go away.
600 * if (container_hold(container) == 0) {
601 * return (KMEM_CBRC_DONT_KNOW);
602 * }
603 *
604 * mutex_enter(&container->c_objects_lock);
605 * if (container != op->o_container) {
606 * mutex_exit(&container->c_objects_lock);
607 * container_rele(container);
608 * return (KMEM_CBRC_DONT_KNOW);
609 * }
610 *
611 * if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
612 * mutex_exit(&container->c_objects_lock);
613 * container_rele(container);
614 * return (KMEM_CBRC_LATER);
615 * }
616 *
617 * object_move_impl(op, np); // critical section
618 * rw_exit(OBJECT_RWLOCK(op));
619 *
620 * op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
621 * list_link_replace(&op->o_link_node, &np->o_link_node);
622 * mutex_exit(&container->c_objects_lock);
623 * container_rele(container);
624 * return (KMEM_CBRC_YES);
625 * }
626 *
627 * Note that object_move() must invalidate the designated o_container pointer of
628 * the old object in the same way that object_free() does, since kmem will free
629 * the object in response to the KMEM_CBRC_YES return value.
630 *
631 * The lock order in object_move() differs from object_alloc(), which locks
632 * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
633 * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
634 * not a problem. Holding the lock on the object list in the example above
635 * through the entire callback not only prevents the object from going away, it
636 * also allows you to lock the list elsewhere and know that none of its elements
637 * will move during iteration.
638 *
639 * Adding an explicit hold everywhere an object from the cache is used is tricky
640 * and involves much more change to client code than a cache-specific solution
641 * that leverages existing state to decide whether or not an object is
642 * movable. However, this approach has the advantage that no object remains
643 * immovable for any significant length of time, making it extremely unlikely
644 * that long-lived allocations can continue holding slabs hostage; and it works
645 * for any cache.
646 *
647 * 3. Consolidator Implementation
648 *
649 * Once the client supplies a move function that a) recognizes known objects and
650 * b) avoids moving objects that are actively in use, the remaining work is up
651 * to the consolidator to decide which objects to move and when to issue
652 * callbacks.
653 *
654 * The consolidator relies on the fact that a cache's slabs are ordered by
655 * usage. Each slab has a fixed number of objects. Depending on the slab's
656 * "color" (the offset of the first object from the beginning of the slab;
657 * offsets are staggered to mitigate false sharing of cache lines) it is either
658 * the maximum number of objects per slab determined at cache creation time or
659 * else the number closest to the maximum that fits within the space remaining
660 * after the initial offset. A completely allocated slab may contribute some
661 * internal fragmentation (per-slab overhead) but no external fragmentation, so
662 * it is of no interest to the consolidator. At the other extreme, slabs whose
663 * objects have all been freed to the slab are released to the virtual memory
664 * (VM) subsystem (objects freed to magazines are still allocated as far as the
665 * slab is concerned). External fragmentation exists when there are slabs
666 * somewhere between these extremes. A partial slab has at least one but not all
667 * of its objects allocated. The more partial slabs, and the fewer allocated
668 * objects on each of them, the higher the fragmentation. Hence the
669 * consolidator's overall strategy is to reduce the number of partial slabs by
670 * moving allocated objects from the least allocated slabs to the most allocated
671 * slabs.
672 *
673 * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
674 * slabs are kept separately in an unordered list. Since the majority of slabs
675 * tend to be completely allocated (a typical unfragmented cache may have
676 * thousands of complete slabs and only a single partial slab), separating
677 * complete slabs improves the efficiency of partial slab ordering, since the
678 * complete slabs do not affect the depth or balance of the AVL tree. This
679 * ordered sequence of partial slabs acts as a "free list" supplying objects for
680 * allocation requests.
681 *
682 * Objects are always allocated from the first partial slab in the free list,
683 * where the allocation is most likely to eliminate a partial slab (by
684 * completely allocating it). Conversely, when a single object from a completely
685 * allocated slab is freed to the slab, that slab is added to the front of the
686 * free list. Since most free list activity involves highly allocated slabs
687 * coming and going at the front of the list, slabs tend naturally toward the
688 * ideal order: highly allocated at the front, sparsely allocated at the back.
689 * Slabs with few allocated objects are likely to become completely free if they
690 * keep a safe distance away from the front of the free list. Slab misorders
691 * interfere with the natural tendency of slabs to become completely free or
692 * completely allocated. For example, a slab with a single allocated object
693 * needs only a single free to escape the cache; its natural desire is
694 * frustrated when it finds itself at the front of the list where a second
695 * allocation happens just before the free could have released it. Another slab
696 * with all but one object allocated might have supplied the buffer instead, so
697 * that both (as opposed to neither) of the slabs would have been taken off the
698 * free list.
699 *
700 * Although slabs tend naturally toward the ideal order, misorders allowed by a
701 * simple list implementation defeat the consolidator's strategy of merging
702 * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
703 * needs another way to fix misorders to optimize its callback strategy. One
704 * approach is to periodically scan a limited number of slabs, advancing a
705 * marker to hold the current scan position, and to move extreme misorders to
706 * the front or back of the free list and to the front or back of the current
707 * scan range. By making consecutive scan ranges overlap by one slab, the least
708 * allocated slab in the current range can be carried along from the end of one
709 * scan to the start of the next.
710 *
711 * Maintaining partial slabs in an AVL tree relieves kmem of this additional
712 * task, however. Since most of the cache's activity is in the magazine layer,
713 * and allocations from the slab layer represent only a startup cost, the
714 * overhead of maintaining a balanced tree is not a significant concern compared
715 * to the opportunity of reducing complexity by eliminating the partial slab
716 * scanner just described. The overhead of an AVL tree is minimized by
717 * maintaining only partial slabs in the tree and keeping completely allocated
718 * slabs separately in a list. To avoid increasing the size of the slab
719 * structure the AVL linkage pointers are reused for the slab's list linkage,
720 * since the slab will always be either partial or complete, never stored both
721 * ways at the same time. To further minimize the overhead of the AVL tree the
722 * compare function that orders partial slabs by usage divides the range of
723 * allocated object counts into bins such that counts within the same bin are
724 * considered equal. Binning partial slabs makes it less likely that allocating
725 * or freeing a single object will change the slab's order, requiring a tree
726 * reinsertion (an avl_remove() followed by an avl_add(), both potentially
727 * requiring some rebalancing of the tree). Allocation counts closest to
728 * completely free and completely allocated are left unbinned (finely sorted) to
729 * better support the consolidator's strategy of merging slabs at either
730 * extreme.
731 *
732 * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
733 *
734 * The consolidator piggybacks on the kmem maintenance thread and is called on
735 * the same interval as kmem_cache_update(), once per cache every fifteen
736 * seconds. kmem maintains a running count of unallocated objects in the slab
737 * layer (cache_bufslab). The consolidator checks whether that number exceeds
738 * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
739 * there is a significant number of slabs in the cache (arbitrarily a minimum
740 * 101 total slabs). Unused objects that have fallen out of the magazine layer's
741 * working set are included in the assessment, and magazines in the depot are
742 * reaped if those objects would lift cache_bufslab above the fragmentation
743 * threshold. Once the consolidator decides that a cache is fragmented, it looks
744 * for a candidate slab to reclaim, starting at the end of the partial slab free
745 * list and scanning backwards. At first the consolidator is choosy: only a slab
746 * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
747 * single allocated object, regardless of percentage). If there is difficulty
748 * finding a candidate slab, kmem raises the allocation threshold incrementally,
749 * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
750 * external fragmentation (unused objects on the free list) below 12.5% (1/8),
751 * even in the worst case of every slab in the cache being almost 7/8 allocated.
752 * The threshold can also be lowered incrementally when candidate slabs are easy
753 * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
754 * is no longer fragmented.
755 *
756 * 3.2 Generating Callbacks
757 *
758 * Once an eligible slab is chosen, a callback is generated for every allocated
759 * object on the slab, in the hope that the client will move everything off the
760 * slab and make it reclaimable. Objects selected as move destinations are
761 * chosen from slabs at the front of the free list. Assuming slabs in the ideal
762 * order (most allocated at the front, least allocated at the back) and a
763 * cooperative client, the consolidator will succeed in removing slabs from both
764 * ends of the free list, completely allocating on the one hand and completely
765 * freeing on the other. Objects selected as move destinations are allocated in
766 * the kmem maintenance thread where move requests are enqueued. A separate
767 * callback thread removes pending callbacks from the queue and calls the
768 * client. The separate thread ensures that client code (the move function) does
769 * not interfere with internal kmem maintenance tasks. A map of pending
770 * callbacks keyed by object address (the object to be moved) is checked to
771 * ensure that duplicate callbacks are not generated for the same object.
772 * Allocating the move destination (the object to move to) prevents subsequent
773 * callbacks from selecting the same destination as an earlier pending callback.
774 *
775 * Move requests can also be generated by kmem_cache_reap() when the system is
776 * desperate for memory and by kmem_cache_move_notify(), called by the client to
777 * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
778 * The map of pending callbacks is protected by the same lock that protects the
779 * slab layer.
780 *
781 * When the system is desperate for memory, kmem does not bother to determine
782 * whether or not the cache exceeds the fragmentation threshold, but tries to
783 * consolidate as many slabs as possible. Normally, the consolidator chews
784 * slowly, one sparsely allocated slab at a time during each maintenance
785 * interval that the cache is fragmented. When desperate, the consolidator
786 * starts at the last partial slab and enqueues callbacks for every allocated
787 * object on every partial slab, working backwards until it reaches the first
788 * partial slab. The first partial slab, meanwhile, advances in pace with the
789 * consolidator as allocations to supply move destinations for the enqueued
790 * callbacks use up the highly allocated slabs at the front of the free list.
791 * Ideally, the overgrown free list collapses like an accordion, starting at
792 * both ends and ending at the center with a single partial slab.
793 *
794 * 3.3 Client Responses
795 *
796 * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
797 * marks the slab that supplied the stuck object non-reclaimable and moves it to
798 * front of the free list. The slab remains marked as long as it remains on the
799 * free list, and it appears more allocated to the partial slab compare function
800 * than any unmarked slab, no matter how many of its objects are allocated.
801 * Since even one immovable object ties up the entire slab, the goal is to
802 * completely allocate any slab that cannot be completely freed. kmem does not
803 * bother generating callbacks to move objects from a marked slab unless the
804 * system is desperate.
805 *
806 * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
807 * slab. If the client responds LATER too many times, kmem disbelieves and
808 * treats the response as a NO. The count is cleared when the slab is taken off
809 * the partial slab list or when the client moves one of the slab's objects.
810 *
811 * 4. Observability
812 *
813 * A kmem cache's external fragmentation is best observed with 'mdb -k' using
814 * the ::kmem_slabs dcmd. For a complete description of the command, enter
815 * '::help kmem_slabs' at the mdb prompt.
816 */
817
818 #include <sys/kmem_impl.h>
819 #include <sys/vmem_impl.h>
820 #include <sys/param.h>
821 #include <sys/sysmacros.h>
822 #include <sys/vm.h>
823 #include <sys/proc.h>
824 #include <sys/tuneable.h>
825 #include <sys/systm.h>
826 #include <sys/cmn_err.h>
827 #include <sys/debug.h>
828 #include <sys/sdt.h>
829 #include <sys/mutex.h>
830 #include <sys/bitmap.h>
831 #include <sys/atomic.h>
832 #include <sys/kobj.h>
833 #include <sys/disp.h>
834 #include <vm/seg_kmem.h>
835 #include <sys/log.h>
836 #include <sys/callb.h>
837 #include <sys/taskq.h>
838 #include <sys/modctl.h>
839 #include <sys/reboot.h>
840 #include <sys/id32.h>
841 #include <sys/zone.h>
842 #include <sys/netstack.h>
843 #ifdef DEBUG
844 #include <sys/random.h>
845 #endif
846
847 extern void streams_msg_init(void);
848 extern int segkp_fromheap;
849 extern void segkp_cache_free(void);
850 extern int callout_init_done;
851
852 struct kmem_cache_kstat {
853 kstat_named_t kmc_buf_size;
854 kstat_named_t kmc_align;
855 kstat_named_t kmc_chunk_size;
856 kstat_named_t kmc_slab_size;
857 kstat_named_t kmc_alloc;
858 kstat_named_t kmc_alloc_fail;
859 kstat_named_t kmc_free;
860 kstat_named_t kmc_depot_alloc;
861 kstat_named_t kmc_depot_free;
862 kstat_named_t kmc_depot_contention;
863 kstat_named_t kmc_slab_alloc;
864 kstat_named_t kmc_slab_free;
865 kstat_named_t kmc_buf_constructed;
866 kstat_named_t kmc_buf_avail;
867 kstat_named_t kmc_buf_inuse;
868 kstat_named_t kmc_buf_total;
869 kstat_named_t kmc_buf_max;
870 kstat_named_t kmc_slab_create;
871 kstat_named_t kmc_slab_destroy;
872 kstat_named_t kmc_vmem_source;
873 kstat_named_t kmc_hash_size;
874 kstat_named_t kmc_hash_lookup_depth;
875 kstat_named_t kmc_hash_rescale;
876 kstat_named_t kmc_full_magazines;
877 kstat_named_t kmc_empty_magazines;
878 kstat_named_t kmc_magazine_size;
879 kstat_named_t kmc_reap; /* number of kmem_cache_reap() calls */
880 kstat_named_t kmc_defrag; /* attempts to defrag all partial slabs */
881 kstat_named_t kmc_scan; /* attempts to defrag one partial slab */
882 kstat_named_t kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
883 kstat_named_t kmc_move_yes;
884 kstat_named_t kmc_move_no;
885 kstat_named_t kmc_move_later;
886 kstat_named_t kmc_move_dont_need;
887 kstat_named_t kmc_move_dont_know; /* obj unrecognized by client ... */
888 kstat_named_t kmc_move_hunt_found; /* ... but found in mag layer */
889 kstat_named_t kmc_move_slabs_freed; /* slabs freed by consolidator */
890 kstat_named_t kmc_move_reclaimable; /* buffers, if consolidator ran */
891 } kmem_cache_kstat = {
892 { "buf_size", KSTAT_DATA_UINT64 },
893 { "align", KSTAT_DATA_UINT64 },
894 { "chunk_size", KSTAT_DATA_UINT64 },
895 { "slab_size", KSTAT_DATA_UINT64 },
896 { "alloc", KSTAT_DATA_UINT64 },
897 { "alloc_fail", KSTAT_DATA_UINT64 },
898 { "free", KSTAT_DATA_UINT64 },
899 { "depot_alloc", KSTAT_DATA_UINT64 },
900 { "depot_free", KSTAT_DATA_UINT64 },
901 { "depot_contention", KSTAT_DATA_UINT64 },
902 { "slab_alloc", KSTAT_DATA_UINT64 },
903 { "slab_free", KSTAT_DATA_UINT64 },
904 { "buf_constructed", KSTAT_DATA_UINT64 },
905 { "buf_avail", KSTAT_DATA_UINT64 },
906 { "buf_inuse", KSTAT_DATA_UINT64 },
907 { "buf_total", KSTAT_DATA_UINT64 },
908 { "buf_max", KSTAT_DATA_UINT64 },
909 { "slab_create", KSTAT_DATA_UINT64 },
910 { "slab_destroy", KSTAT_DATA_UINT64 },
911 { "vmem_source", KSTAT_DATA_UINT64 },
912 { "hash_size", KSTAT_DATA_UINT64 },
913 { "hash_lookup_depth", KSTAT_DATA_UINT64 },
914 { "hash_rescale", KSTAT_DATA_UINT64 },
915 { "full_magazines", KSTAT_DATA_UINT64 },
916 { "empty_magazines", KSTAT_DATA_UINT64 },
917 { "magazine_size", KSTAT_DATA_UINT64 },
918 { "reap", KSTAT_DATA_UINT64 },
919 { "defrag", KSTAT_DATA_UINT64 },
920 { "scan", KSTAT_DATA_UINT64 },
921 { "move_callbacks", KSTAT_DATA_UINT64 },
922 { "move_yes", KSTAT_DATA_UINT64 },
923 { "move_no", KSTAT_DATA_UINT64 },
924 { "move_later", KSTAT_DATA_UINT64 },
925 { "move_dont_need", KSTAT_DATA_UINT64 },
926 { "move_dont_know", KSTAT_DATA_UINT64 },
927 { "move_hunt_found", KSTAT_DATA_UINT64 },
928 { "move_slabs_freed", KSTAT_DATA_UINT64 },
929 { "move_reclaimable", KSTAT_DATA_UINT64 },
930 };
931
932 static kmutex_t kmem_cache_kstat_lock;
933
934 /*
935 * The default set of caches to back kmem_alloc().
936 * These sizes should be reevaluated periodically.
937 *
938 * We want allocations that are multiples of the coherency granularity
939 * (64 bytes) to be satisfied from a cache which is a multiple of 64
940 * bytes, so that it will be 64-byte aligned. For all multiples of 64,
941 * the next kmem_cache_size greater than or equal to it must be a
942 * multiple of 64.
943 *
944 * We split the table into two sections: size <= 4k and size > 4k. This
945 * saves a lot of space and cache footprint in our cache tables.
946 */
947 static const int kmem_alloc_sizes[] = {
948 1 * 8,
949 2 * 8,
950 3 * 8,
951 4 * 8, 5 * 8, 6 * 8, 7 * 8,
952 4 * 16, 5 * 16, 6 * 16, 7 * 16,
953 4 * 32, 5 * 32, 6 * 32, 7 * 32,
954 4 * 64, 5 * 64, 6 * 64, 7 * 64,
955 4 * 128, 5 * 128, 6 * 128, 7 * 128,
956 P2ALIGN(8192 / 7, 64),
957 P2ALIGN(8192 / 6, 64),
958 P2ALIGN(8192 / 5, 64),
959 P2ALIGN(8192 / 4, 64),
960 P2ALIGN(8192 / 3, 64),
961 P2ALIGN(8192 / 2, 64),
962 };
963
964 static const int kmem_big_alloc_sizes[] = {
965 2 * 4096, 3 * 4096,
966 2 * 8192, 3 * 8192,
967 4 * 8192, 5 * 8192, 6 * 8192, 7 * 8192,
968 8 * 8192, 9 * 8192, 10 * 8192, 11 * 8192,
969 12 * 8192, 13 * 8192, 14 * 8192, 15 * 8192,
970 16 * 8192
971 };
972
973 #define KMEM_MAXBUF 4096
974 #define KMEM_BIG_MAXBUF_32BIT 32768
975 #define KMEM_BIG_MAXBUF 131072
976
977 #define KMEM_BIG_MULTIPLE 4096 /* big_alloc_sizes must be a multiple */
978 #define KMEM_BIG_SHIFT 12 /* lg(KMEM_BIG_MULTIPLE) */
979
980 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
981 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
982
983 #define KMEM_ALLOC_TABLE_MAX (KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
984 static size_t kmem_big_alloc_table_max = 0; /* # of filled elements */
985
986 static kmem_magtype_t kmem_magtype[] = {
987 { 1, 8, 3200, 65536 },
988 { 3, 16, 256, 32768 },
989 { 7, 32, 64, 16384 },
990 { 15, 64, 0, 8192 },
991 { 31, 64, 0, 4096 },
992 { 47, 64, 0, 2048 },
993 { 63, 64, 0, 1024 },
994 { 95, 64, 0, 512 },
995 { 143, 64, 0, 0 },
996 };
997
998 static uint32_t kmem_reaping;
999 static uint32_t kmem_reaping_idspace;
1000
1001 /*
1002 * kmem tunables
1003 */
1004 clock_t kmem_reap_interval; /* cache reaping rate [15 * HZ ticks] */
1005 int kmem_depot_contention = 3; /* max failed tryenters per real interval */
1006 pgcnt_t kmem_reapahead = 0; /* start reaping N pages before pageout */
1007 int kmem_panic = 1; /* whether to panic on error */
1008 int kmem_logging = 1; /* kmem_log_enter() override */
1009 uint32_t kmem_mtbf = 0; /* mean time between failures [default: off] */
1010 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
1011 size_t kmem_content_log_size; /* content log size [2% of memory] */
1012 size_t kmem_failure_log_size; /* failure log [4 pages per CPU] */
1013 size_t kmem_slab_log_size; /* slab create log [4 pages per CPU] */
1014 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1015 size_t kmem_lite_minsize = 0; /* minimum buffer size for KMF_LITE */
1016 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1017 int kmem_lite_pcs = 4; /* number of PCs to store in KMF_LITE mode */
1018 size_t kmem_maxverify; /* maximum bytes to inspect in debug routines */
1019 size_t kmem_minfirewall; /* hardware-enforced redzone threshold */
1020
1021 #ifdef _LP64
1022 size_t kmem_max_cached = KMEM_BIG_MAXBUF; /* maximum kmem_alloc cache */
1023 #else
1024 size_t kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1025 #endif
1026
1027 #ifdef DEBUG
1028 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1029 #else
1030 int kmem_flags = 0;
1031 #endif
1032 int kmem_ready;
1033
1034 static kmem_cache_t *kmem_slab_cache;
1035 static kmem_cache_t *kmem_bufctl_cache;
1036 static kmem_cache_t *kmem_bufctl_audit_cache;
1037
1038 static kmutex_t kmem_cache_lock; /* inter-cache linkage only */
1039 static list_t kmem_caches;
1040
1041 static taskq_t *kmem_taskq;
1042 static kmutex_t kmem_flags_lock;
1043 static vmem_t *kmem_metadata_arena;
1044 static vmem_t *kmem_msb_arena; /* arena for metadata caches */
1045 static vmem_t *kmem_cache_arena;
1046 static vmem_t *kmem_hash_arena;
1047 static vmem_t *kmem_log_arena;
1048 static vmem_t *kmem_oversize_arena;
1049 static vmem_t *kmem_va_arena;
1050 static vmem_t *kmem_default_arena;
1051 static vmem_t *kmem_firewall_va_arena;
1052 static vmem_t *kmem_firewall_arena;
1053
1054 /*
1055 * kmem slab consolidator thresholds (tunables)
1056 */
1057 size_t kmem_frag_minslabs = 101; /* minimum total slabs */
1058 size_t kmem_frag_numer = 1; /* free buffers (numerator) */
1059 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1060 /*
1061 * Maximum number of slabs from which to move buffers during a single
1062 * maintenance interval while the system is not low on memory.
1063 */
1064 size_t kmem_reclaim_max_slabs = 1;
1065 /*
1066 * Number of slabs to scan backwards from the end of the partial slab list
1067 * when searching for buffers to relocate.
1068 */
1069 size_t kmem_reclaim_scan_range = 12;
1070
1071 /* consolidator knobs */
1072 boolean_t kmem_move_noreap;
1073 boolean_t kmem_move_blocked;
1074 boolean_t kmem_move_fulltilt;
1075 boolean_t kmem_move_any_partial;
1076
1077 #ifdef DEBUG
1078 /*
1079 * kmem consolidator debug tunables:
1080 * Ensure code coverage by occasionally running the consolidator even when the
1081 * caches are not fragmented (they may never be). These intervals are mean time
1082 * in cache maintenance intervals (kmem_cache_update).
1083 */
1084 uint32_t kmem_mtb_move = 60; /* defrag 1 slab (~15min) */
1085 uint32_t kmem_mtb_reap = 1800; /* defrag all slabs (~7.5hrs) */
1086 #endif /* DEBUG */
1087
1088 static kmem_cache_t *kmem_defrag_cache;
1089 static kmem_cache_t *kmem_move_cache;
1090 static taskq_t *kmem_move_taskq;
1091
1092 static void kmem_cache_scan(kmem_cache_t *);
1093 static void kmem_cache_defrag(kmem_cache_t *);
1094 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1095
1096
1097 kmem_log_header_t *kmem_transaction_log;
1098 kmem_log_header_t *kmem_content_log;
1099 kmem_log_header_t *kmem_failure_log;
1100 kmem_log_header_t *kmem_slab_log;
1101
1102 static int kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1103
1104 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller) \
1105 if ((count) > 0) { \
1106 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1107 pc_t *_e; \
1108 /* memmove() the old entries down one notch */ \
1109 for (_e = &_s[(count) - 1]; _e > _s; _e--) \
1110 *_e = *(_e - 1); \
1111 *_s = (uintptr_t)(caller); \
1112 }
1113
1114 #define KMERR_MODIFIED 0 /* buffer modified while on freelist */
1115 #define KMERR_REDZONE 1 /* redzone violation (write past end of buf) */
1116 #define KMERR_DUPFREE 2 /* freed a buffer twice */
1117 #define KMERR_BADADDR 3 /* freed a bad (unallocated) address */
1118 #define KMERR_BADBUFTAG 4 /* buftag corrupted */
1119 #define KMERR_BADBUFCTL 5 /* bufctl corrupted */
1120 #define KMERR_BADCACHE 6 /* freed a buffer to the wrong cache */
1121 #define KMERR_BADSIZE 7 /* alloc size != free size */
1122 #define KMERR_BADBASE 8 /* buffer base address wrong */
1123
1124 struct {
1125 hrtime_t kmp_timestamp; /* timestamp of panic */
1126 int kmp_error; /* type of kmem error */
1127 void *kmp_buffer; /* buffer that induced panic */
1128 void *kmp_realbuf; /* real start address for buffer */
1129 kmem_cache_t *kmp_cache; /* buffer's cache according to client */
1130 kmem_cache_t *kmp_realcache; /* actual cache containing buffer */
1131 kmem_slab_t *kmp_slab; /* slab accoring to kmem_findslab() */
1132 kmem_bufctl_t *kmp_bufctl; /* bufctl */
1133 } kmem_panic_info;
1134
1135
1136 static void
1137 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1138 {
1139 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1140 uint64_t *buf = buf_arg;
1141
1142 while (buf < bufend)
1143 *buf++ = pattern;
1144 }
1145
1146 static void *
1147 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1148 {
1149 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1150 uint64_t *buf;
1151
1152 for (buf = buf_arg; buf < bufend; buf++)
1153 if (*buf != pattern)
1154 return (buf);
1155 return (NULL);
1156 }
1157
1158 static void *
1159 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1160 {
1161 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1162 uint64_t *buf;
1163
1164 for (buf = buf_arg; buf < bufend; buf++) {
1165 if (*buf != old) {
1166 copy_pattern(old, buf_arg,
1167 (char *)buf - (char *)buf_arg);
1168 return (buf);
1169 }
1170 *buf = new;
1171 }
1172
1173 return (NULL);
1174 }
1175
1176 static void
1177 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1178 {
1179 kmem_cache_t *cp;
1180
1181 mutex_enter(&kmem_cache_lock);
1182 for (cp = list_head(&kmem_caches); cp != NULL;
1183 cp = list_next(&kmem_caches, cp))
1184 if (tq != NULL)
1185 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1186 tqflag);
1187 else
1188 func(cp);
1189 mutex_exit(&kmem_cache_lock);
1190 }
1191
1192 static void
1193 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1194 {
1195 kmem_cache_t *cp;
1196
1197 mutex_enter(&kmem_cache_lock);
1198 for (cp = list_head(&kmem_caches); cp != NULL;
1199 cp = list_next(&kmem_caches, cp)) {
1200 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1201 continue;
1202 if (tq != NULL)
1203 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1204 tqflag);
1205 else
1206 func(cp);
1207 }
1208 mutex_exit(&kmem_cache_lock);
1209 }
1210
1211 /*
1212 * Debugging support. Given a buffer address, find its slab.
1213 */
1214 static kmem_slab_t *
1215 kmem_findslab(kmem_cache_t *cp, void *buf)
1216 {
1217 kmem_slab_t *sp;
1218
1219 mutex_enter(&cp->cache_lock);
1220 for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1221 sp = list_next(&cp->cache_complete_slabs, sp)) {
1222 if (KMEM_SLAB_MEMBER(sp, buf)) {
1223 mutex_exit(&cp->cache_lock);
1224 return (sp);
1225 }
1226 }
1227 for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1228 sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1229 if (KMEM_SLAB_MEMBER(sp, buf)) {
1230 mutex_exit(&cp->cache_lock);
1231 return (sp);
1232 }
1233 }
1234 mutex_exit(&cp->cache_lock);
1235
1236 return (NULL);
1237 }
1238
1239 static void
1240 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1241 {
1242 kmem_buftag_t *btp = NULL;
1243 kmem_bufctl_t *bcp = NULL;
1244 kmem_cache_t *cp = cparg;
1245 kmem_slab_t *sp;
1246 uint64_t *off;
1247 void *buf = bufarg;
1248
1249 kmem_logging = 0; /* stop logging when a bad thing happens */
1250
1251 kmem_panic_info.kmp_timestamp = gethrtime();
1252
1253 sp = kmem_findslab(cp, buf);
1254 if (sp == NULL) {
1255 for (cp = list_tail(&kmem_caches); cp != NULL;
1256 cp = list_prev(&kmem_caches, cp)) {
1257 if ((sp = kmem_findslab(cp, buf)) != NULL)
1258 break;
1259 }
1260 }
1261
1262 if (sp == NULL) {
1263 cp = NULL;
1264 error = KMERR_BADADDR;
1265 } else {
1266 if (cp != cparg)
1267 error = KMERR_BADCACHE;
1268 else
1269 buf = (char *)bufarg - ((uintptr_t)bufarg -
1270 (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1271 if (buf != bufarg)
1272 error = KMERR_BADBASE;
1273 if (cp->cache_flags & KMF_BUFTAG)
1274 btp = KMEM_BUFTAG(cp, buf);
1275 if (cp->cache_flags & KMF_HASH) {
1276 mutex_enter(&cp->cache_lock);
1277 for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1278 if (bcp->bc_addr == buf)
1279 break;
1280 mutex_exit(&cp->cache_lock);
1281 if (bcp == NULL && btp != NULL)
1282 bcp = btp->bt_bufctl;
1283 if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1284 NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1285 bcp->bc_addr != buf) {
1286 error = KMERR_BADBUFCTL;
1287 bcp = NULL;
1288 }
1289 }
1290 }
1291
1292 kmem_panic_info.kmp_error = error;
1293 kmem_panic_info.kmp_buffer = bufarg;
1294 kmem_panic_info.kmp_realbuf = buf;
1295 kmem_panic_info.kmp_cache = cparg;
1296 kmem_panic_info.kmp_realcache = cp;
1297 kmem_panic_info.kmp_slab = sp;
1298 kmem_panic_info.kmp_bufctl = bcp;
1299
1300 printf("kernel memory allocator: ");
1301
1302 switch (error) {
1303
1304 case KMERR_MODIFIED:
1305 printf("buffer modified after being freed\n");
1306 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1307 if (off == NULL) /* shouldn't happen */
1308 off = buf;
1309 printf("modification occurred at offset 0x%lx "
1310 "(0x%llx replaced by 0x%llx)\n",
1311 (uintptr_t)off - (uintptr_t)buf,
1312 (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1313 break;
1314
1315 case KMERR_REDZONE:
1316 printf("redzone violation: write past end of buffer\n");
1317 break;
1318
1319 case KMERR_BADADDR:
1320 printf("invalid free: buffer not in cache\n");
1321 break;
1322
1323 case KMERR_DUPFREE:
1324 printf("duplicate free: buffer freed twice\n");
1325 break;
1326
1327 case KMERR_BADBUFTAG:
1328 printf("boundary tag corrupted\n");
1329 printf("bcp ^ bxstat = %lx, should be %lx\n",
1330 (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1331 KMEM_BUFTAG_FREE);
1332 break;
1333
1334 case KMERR_BADBUFCTL:
1335 printf("bufctl corrupted\n");
1336 break;
1337
1338 case KMERR_BADCACHE:
1339 printf("buffer freed to wrong cache\n");
1340 printf("buffer was allocated from %s,\n", cp->cache_name);
1341 printf("caller attempting free to %s.\n", cparg->cache_name);
1342 break;
1343
1344 case KMERR_BADSIZE:
1345 printf("bad free: free size (%u) != alloc size (%u)\n",
1346 KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1347 KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1348 break;
1349
1350 case KMERR_BADBASE:
1351 printf("bad free: free address (%p) != alloc address (%p)\n",
1352 bufarg, buf);
1353 break;
1354 }
1355
1356 printf("buffer=%p bufctl=%p cache: %s\n",
1357 bufarg, (void *)bcp, cparg->cache_name);
1358
1359 if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1360 error != KMERR_BADBUFCTL) {
1361 int d;
1362 timestruc_t ts;
1363 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1364
1365 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1366 printf("previous transaction on buffer %p:\n", buf);
1367 printf("thread=%p time=T-%ld.%09ld slab=%p cache: %s\n",
1368 (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1369 (void *)sp, cp->cache_name);
1370 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1371 ulong_t off;
1372 char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1373 printf("%s+%lx\n", sym ? sym : "?", off);
1374 }
1375 }
1376 if (kmem_panic > 0)
1377 panic("kernel heap corruption detected");
1378 if (kmem_panic == 0)
1379 debug_enter(NULL);
1380 kmem_logging = 1; /* resume logging */
1381 }
1382
1383 static kmem_log_header_t *
1384 kmem_log_init(size_t logsize)
1385 {
1386 kmem_log_header_t *lhp;
1387 int nchunks = 4 * max_ncpus;
1388 size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1389 int i;
1390
1391 /*
1392 * Make sure that lhp->lh_cpu[] is nicely aligned
1393 * to prevent false sharing of cache lines.
1394 */
1395 lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1396 lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1397 NULL, NULL, VM_SLEEP);
1398 bzero(lhp, lhsize);
1399
1400 mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1401 lhp->lh_nchunks = nchunks;
1402 lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1403 lhp->lh_base = vmem_alloc(kmem_log_arena,
1404 lhp->lh_chunksize * nchunks, VM_SLEEP);
1405 lhp->lh_free = vmem_alloc(kmem_log_arena,
1406 nchunks * sizeof (int), VM_SLEEP);
1407 bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1408
1409 for (i = 0; i < max_ncpus; i++) {
1410 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1411 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1412 clhp->clh_chunk = i;
1413 }
1414
1415 for (i = max_ncpus; i < nchunks; i++)
1416 lhp->lh_free[i] = i;
1417
1418 lhp->lh_head = max_ncpus;
1419 lhp->lh_tail = 0;
1420
1421 return (lhp);
1422 }
1423
1424 static void *
1425 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1426 {
1427 void *logspace;
1428 kmem_cpu_log_header_t *clhp;
1429
1430 if (lhp == NULL || kmem_logging == 0 || panicstr)
1431 return (NULL);
1432
1433 clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1434
1435 mutex_enter(&clhp->clh_lock);
1436 clhp->clh_hits++;
1437 if (size > clhp->clh_avail) {
1438 mutex_enter(&lhp->lh_lock);
1439 lhp->lh_hits++;
1440 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1441 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1442 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1443 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1444 clhp->clh_current = lhp->lh_base +
1445 clhp->clh_chunk * lhp->lh_chunksize;
1446 clhp->clh_avail = lhp->lh_chunksize;
1447 if (size > lhp->lh_chunksize)
1448 size = lhp->lh_chunksize;
1449 mutex_exit(&lhp->lh_lock);
1450 }
1451 logspace = clhp->clh_current;
1452 clhp->clh_current += size;
1453 clhp->clh_avail -= size;
1454 bcopy(data, logspace, size);
1455 mutex_exit(&clhp->clh_lock);
1456 return (logspace);
1457 }
1458
1459 #define KMEM_AUDIT(lp, cp, bcp) \
1460 { \
1461 kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp); \
1462 _bcp->bc_timestamp = gethrtime(); \
1463 _bcp->bc_thread = curthread; \
1464 _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH); \
1465 _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp)); \
1466 }
1467
1468 static void
1469 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1470 kmem_slab_t *sp, void *addr)
1471 {
1472 kmem_bufctl_audit_t bca;
1473
1474 bzero(&bca, sizeof (kmem_bufctl_audit_t));
1475 bca.bc_addr = addr;
1476 bca.bc_slab = sp;
1477 bca.bc_cache = cp;
1478 KMEM_AUDIT(lp, cp, &bca);
1479 }
1480
1481 /*
1482 * Create a new slab for cache cp.
1483 */
1484 static kmem_slab_t *
1485 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1486 {
1487 size_t slabsize = cp->cache_slabsize;
1488 size_t chunksize = cp->cache_chunksize;
1489 int cache_flags = cp->cache_flags;
1490 size_t color, chunks;
1491 char *buf, *slab;
1492 kmem_slab_t *sp;
1493 kmem_bufctl_t *bcp;
1494 vmem_t *vmp = cp->cache_arena;
1495
1496 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1497
1498 color = cp->cache_color + cp->cache_align;
1499 if (color > cp->cache_maxcolor)
1500 color = cp->cache_mincolor;
1501 cp->cache_color = color;
1502
1503 slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1504
1505 if (slab == NULL)
1506 goto vmem_alloc_failure;
1507
1508 ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1509
1510 /*
1511 * Reverify what was already checked in kmem_cache_set_move(), since the
1512 * consolidator depends (for correctness) on slabs being initialized
1513 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1514 * clients to distinguish uninitialized memory from known objects).
1515 */
1516 ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1517 if (!(cp->cache_cflags & KMC_NOTOUCH))
1518 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1519
1520 if (cache_flags & KMF_HASH) {
1521 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1522 goto slab_alloc_failure;
1523 chunks = (slabsize - color) / chunksize;
1524 } else {
1525 sp = KMEM_SLAB(cp, slab);
1526 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1527 }
1528
1529 sp->slab_cache = cp;
1530 sp->slab_head = NULL;
1531 sp->slab_refcnt = 0;
1532 sp->slab_base = buf = slab + color;
1533 sp->slab_chunks = chunks;
1534 sp->slab_stuck_offset = (uint32_t)-1;
1535 sp->slab_later_count = 0;
1536 sp->slab_flags = 0;
1537
1538 ASSERT(chunks > 0);
1539 while (chunks-- != 0) {
1540 if (cache_flags & KMF_HASH) {
1541 bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1542 if (bcp == NULL)
1543 goto bufctl_alloc_failure;
1544 if (cache_flags & KMF_AUDIT) {
1545 kmem_bufctl_audit_t *bcap =
1546 (kmem_bufctl_audit_t *)bcp;
1547 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1548 bcap->bc_cache = cp;
1549 }
1550 bcp->bc_addr = buf;
1551 bcp->bc_slab = sp;
1552 } else {
1553 bcp = KMEM_BUFCTL(cp, buf);
1554 }
1555 if (cache_flags & KMF_BUFTAG) {
1556 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1557 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1558 btp->bt_bufctl = bcp;
1559 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1560 if (cache_flags & KMF_DEADBEEF) {
1561 copy_pattern(KMEM_FREE_PATTERN, buf,
1562 cp->cache_verify);
1563 }
1564 }
1565 bcp->bc_next = sp->slab_head;
1566 sp->slab_head = bcp;
1567 buf += chunksize;
1568 }
1569
1570 kmem_log_event(kmem_slab_log, cp, sp, slab);
1571
1572 return (sp);
1573
1574 bufctl_alloc_failure:
1575
1576 while ((bcp = sp->slab_head) != NULL) {
1577 sp->slab_head = bcp->bc_next;
1578 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1579 }
1580 kmem_cache_free(kmem_slab_cache, sp);
1581
1582 slab_alloc_failure:
1583
1584 vmem_free(vmp, slab, slabsize);
1585
1586 vmem_alloc_failure:
1587
1588 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1589 atomic_inc_64(&cp->cache_alloc_fail);
1590
1591 return (NULL);
1592 }
1593
1594 /*
1595 * Destroy a slab.
1596 */
1597 static void
1598 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1599 {
1600 vmem_t *vmp = cp->cache_arena;
1601 void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1602
1603 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1604 ASSERT(sp->slab_refcnt == 0);
1605
1606 if (cp->cache_flags & KMF_HASH) {
1607 kmem_bufctl_t *bcp;
1608 while ((bcp = sp->slab_head) != NULL) {
1609 sp->slab_head = bcp->bc_next;
1610 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1611 }
1612 kmem_cache_free(kmem_slab_cache, sp);
1613 }
1614 vmem_free(vmp, slab, cp->cache_slabsize);
1615 }
1616
1617 static void *
1618 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1619 {
1620 kmem_bufctl_t *bcp, **hash_bucket;
1621 void *buf;
1622 boolean_t new_slab = (sp->slab_refcnt == 0);
1623
1624 ASSERT(MUTEX_HELD(&cp->cache_lock));
1625 /*
1626 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1627 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1628 * slab is newly created.
1629 */
1630 ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1631 (sp == avl_first(&cp->cache_partial_slabs))));
1632 ASSERT(sp->slab_cache == cp);
1633
1634 cp->cache_slab_alloc++;
1635 cp->cache_bufslab--;
1636 sp->slab_refcnt++;
1637
1638 bcp = sp->slab_head;
1639 sp->slab_head = bcp->bc_next;
1640
1641 if (cp->cache_flags & KMF_HASH) {
1642 /*
1643 * Add buffer to allocated-address hash table.
1644 */
1645 buf = bcp->bc_addr;
1646 hash_bucket = KMEM_HASH(cp, buf);
1647 bcp->bc_next = *hash_bucket;
1648 *hash_bucket = bcp;
1649 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1650 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1651 }
1652 } else {
1653 buf = KMEM_BUF(cp, bcp);
1654 }
1655
1656 ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1657
1658 if (sp->slab_head == NULL) {
1659 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1660 if (new_slab) {
1661 ASSERT(sp->slab_chunks == 1);
1662 } else {
1663 ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1664 avl_remove(&cp->cache_partial_slabs, sp);
1665 sp->slab_later_count = 0; /* clear history */
1666 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1667 sp->slab_stuck_offset = (uint32_t)-1;
1668 }
1669 list_insert_head(&cp->cache_complete_slabs, sp);
1670 cp->cache_complete_slab_count++;
1671 return (buf);
1672 }
1673
1674 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1675 /*
1676 * Peek to see if the magazine layer is enabled before
1677 * we prefill. We're not holding the cpu cache lock,
1678 * so the peek could be wrong, but there's no harm in it.
1679 */
1680 if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1681 (KMEM_CPU_CACHE(cp)->cc_magsize != 0)) {
1682 kmem_slab_prefill(cp, sp);
1683 return (buf);
1684 }
1685
1686 if (new_slab) {
1687 avl_add(&cp->cache_partial_slabs, sp);
1688 return (buf);
1689 }
1690
1691 /*
1692 * The slab is now more allocated than it was, so the
1693 * order remains unchanged.
1694 */
1695 ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1696 return (buf);
1697 }
1698
1699 /*
1700 * Allocate a raw (unconstructed) buffer from cp's slab layer.
1701 */
1702 static void *
1703 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1704 {
1705 kmem_slab_t *sp;
1706 void *buf;
1707 boolean_t test_destructor;
1708
1709 mutex_enter(&cp->cache_lock);
1710 test_destructor = (cp->cache_slab_alloc == 0);
1711 sp = avl_first(&cp->cache_partial_slabs);
1712 if (sp == NULL) {
1713 ASSERT(cp->cache_bufslab == 0);
1714
1715 /*
1716 * The freelist is empty. Create a new slab.
1717 */
1718 mutex_exit(&cp->cache_lock);
1719 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1720 return (NULL);
1721 }
1722 mutex_enter(&cp->cache_lock);
1723 cp->cache_slab_create++;
1724 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1725 cp->cache_bufmax = cp->cache_buftotal;
1726 cp->cache_bufslab += sp->slab_chunks;
1727 }
1728
1729 buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1730 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1731 (cp->cache_complete_slab_count +
1732 avl_numnodes(&cp->cache_partial_slabs) +
1733 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1734 mutex_exit(&cp->cache_lock);
1735
1736 if (test_destructor && cp->cache_destructor != NULL) {
1737 /*
1738 * On the first kmem_slab_alloc(), assert that it is valid to
1739 * call the destructor on a newly constructed object without any
1740 * client involvement.
1741 */
1742 if ((cp->cache_constructor == NULL) ||
1743 cp->cache_constructor(buf, cp->cache_private,
1744 kmflag) == 0) {
1745 cp->cache_destructor(buf, cp->cache_private);
1746 }
1747 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1748 cp->cache_bufsize);
1749 if (cp->cache_flags & KMF_DEADBEEF) {
1750 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1751 }
1752 }
1753
1754 return (buf);
1755 }
1756
1757 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1758
1759 /*
1760 * Free a raw (unconstructed) buffer to cp's slab layer.
1761 */
1762 static void
1763 kmem_slab_free(kmem_cache_t *cp, void *buf)
1764 {
1765 kmem_slab_t *sp;
1766 kmem_bufctl_t *bcp, **prev_bcpp;
1767
1768 ASSERT(buf != NULL);
1769
1770 mutex_enter(&cp->cache_lock);
1771 cp->cache_slab_free++;
1772
1773 if (cp->cache_flags & KMF_HASH) {
1774 /*
1775 * Look up buffer in allocated-address hash table.
1776 */
1777 prev_bcpp = KMEM_HASH(cp, buf);
1778 while ((bcp = *prev_bcpp) != NULL) {
1779 if (bcp->bc_addr == buf) {
1780 *prev_bcpp = bcp->bc_next;
1781 sp = bcp->bc_slab;
1782 break;
1783 }
1784 cp->cache_lookup_depth++;
1785 prev_bcpp = &bcp->bc_next;
1786 }
1787 } else {
1788 bcp = KMEM_BUFCTL(cp, buf);
1789 sp = KMEM_SLAB(cp, buf);
1790 }
1791
1792 if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1793 mutex_exit(&cp->cache_lock);
1794 kmem_error(KMERR_BADADDR, cp, buf);
1795 return;
1796 }
1797
1798 if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1799 /*
1800 * If this is the buffer that prevented the consolidator from
1801 * clearing the slab, we can reset the slab flags now that the
1802 * buffer is freed. (It makes sense to do this in
1803 * kmem_cache_free(), where the client gives up ownership of the
1804 * buffer, but on the hot path the test is too expensive.)
1805 */
1806 kmem_slab_move_yes(cp, sp, buf);
1807 }
1808
1809 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1810 if (cp->cache_flags & KMF_CONTENTS)
1811 ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1812 kmem_log_enter(kmem_content_log, buf,
1813 cp->cache_contents);
1814 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1815 }
1816
1817 bcp->bc_next = sp->slab_head;
1818 sp->slab_head = bcp;
1819
1820 cp->cache_bufslab++;
1821 ASSERT(sp->slab_refcnt >= 1);
1822
1823 if (--sp->slab_refcnt == 0) {
1824 /*
1825 * There are no outstanding allocations from this slab,
1826 * so we can reclaim the memory.
1827 */
1828 if (sp->slab_chunks == 1) {
1829 list_remove(&cp->cache_complete_slabs, sp);
1830 cp->cache_complete_slab_count--;
1831 } else {
1832 avl_remove(&cp->cache_partial_slabs, sp);
1833 }
1834
1835 cp->cache_buftotal -= sp->slab_chunks;
1836 cp->cache_bufslab -= sp->slab_chunks;
1837 /*
1838 * Defer releasing the slab to the virtual memory subsystem
1839 * while there is a pending move callback, since we guarantee
1840 * that buffers passed to the move callback have only been
1841 * touched by kmem or by the client itself. Since the memory
1842 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1843 * set at least one of the two lowest order bits, the client can
1844 * test those bits in the move callback to determine whether or
1845 * not it knows about the buffer (assuming that the client also
1846 * sets one of those low order bits whenever it frees a buffer).
1847 */
1848 if (cp->cache_defrag == NULL ||
1849 (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1850 !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1851 cp->cache_slab_destroy++;
1852 mutex_exit(&cp->cache_lock);
1853 kmem_slab_destroy(cp, sp);
1854 } else {
1855 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1856 /*
1857 * Slabs are inserted at both ends of the deadlist to
1858 * distinguish between slabs freed while move callbacks
1859 * are pending (list head) and a slab freed while the
1860 * lock is dropped in kmem_move_buffers() (list tail) so
1861 * that in both cases slab_destroy() is called from the
1862 * right context.
1863 */
1864 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1865 list_insert_tail(deadlist, sp);
1866 } else {
1867 list_insert_head(deadlist, sp);
1868 }
1869 cp->cache_defrag->kmd_deadcount++;
1870 mutex_exit(&cp->cache_lock);
1871 }
1872 return;
1873 }
1874
1875 if (bcp->bc_next == NULL) {
1876 /* Transition the slab from completely allocated to partial. */
1877 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1878 ASSERT(sp->slab_chunks > 1);
1879 list_remove(&cp->cache_complete_slabs, sp);
1880 cp->cache_complete_slab_count--;
1881 avl_add(&cp->cache_partial_slabs, sp);
1882 } else {
1883 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1884 }
1885
1886 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1887 (cp->cache_complete_slab_count +
1888 avl_numnodes(&cp->cache_partial_slabs) +
1889 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1890 mutex_exit(&cp->cache_lock);
1891 }
1892
1893 /*
1894 * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1895 */
1896 static int
1897 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1898 caddr_t caller)
1899 {
1900 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1901 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1902 uint32_t mtbf;
1903
1904 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1905 kmem_error(KMERR_BADBUFTAG, cp, buf);
1906 return (-1);
1907 }
1908
1909 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1910
1911 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1912 kmem_error(KMERR_BADBUFCTL, cp, buf);
1913 return (-1);
1914 }
1915
1916 if (cp->cache_flags & KMF_DEADBEEF) {
1917 if (!construct && (cp->cache_flags & KMF_LITE)) {
1918 if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1919 kmem_error(KMERR_MODIFIED, cp, buf);
1920 return (-1);
1921 }
1922 if (cp->cache_constructor != NULL)
1923 *(uint64_t *)buf = btp->bt_redzone;
1924 else
1925 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1926 } else {
1927 construct = 1;
1928 if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1929 KMEM_UNINITIALIZED_PATTERN, buf,
1930 cp->cache_verify)) {
1931 kmem_error(KMERR_MODIFIED, cp, buf);
1932 return (-1);
1933 }
1934 }
1935 }
1936 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1937
1938 if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1939 gethrtime() % mtbf == 0 &&
1940 (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1941 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1942 if (!construct && cp->cache_destructor != NULL)
1943 cp->cache_destructor(buf, cp->cache_private);
1944 } else {
1945 mtbf = 0;
1946 }
1947
1948 if (mtbf || (construct && cp->cache_constructor != NULL &&
1949 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1950 atomic_inc_64(&cp->cache_alloc_fail);
1951 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1952 if (cp->cache_flags & KMF_DEADBEEF)
1953 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1954 kmem_slab_free(cp, buf);
1955 return (1);
1956 }
1957
1958 if (cp->cache_flags & KMF_AUDIT) {
1959 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1960 }
1961
1962 if ((cp->cache_flags & KMF_LITE) &&
1963 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1964 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1965 }
1966
1967 return (0);
1968 }
1969
1970 static int
1971 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1972 {
1973 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1974 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1975 kmem_slab_t *sp;
1976
1977 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1978 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1979 kmem_error(KMERR_DUPFREE, cp, buf);
1980 return (-1);
1981 }
1982 sp = kmem_findslab(cp, buf);
1983 if (sp == NULL || sp->slab_cache != cp)
1984 kmem_error(KMERR_BADADDR, cp, buf);
1985 else
1986 kmem_error(KMERR_REDZONE, cp, buf);
1987 return (-1);
1988 }
1989
1990 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1991
1992 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1993 kmem_error(KMERR_BADBUFCTL, cp, buf);
1994 return (-1);
1995 }
1996
1997 if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
1998 kmem_error(KMERR_REDZONE, cp, buf);
1999 return (-1);
2000 }
2001
2002 if (cp->cache_flags & KMF_AUDIT) {
2003 if (cp->cache_flags & KMF_CONTENTS)
2004 bcp->bc_contents = kmem_log_enter(kmem_content_log,
2005 buf, cp->cache_contents);
2006 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2007 }
2008
2009 if ((cp->cache_flags & KMF_LITE) &&
2010 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2011 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2012 }
2013
2014 if (cp->cache_flags & KMF_DEADBEEF) {
2015 if (cp->cache_flags & KMF_LITE)
2016 btp->bt_redzone = *(uint64_t *)buf;
2017 else if (cp->cache_destructor != NULL)
2018 cp->cache_destructor(buf, cp->cache_private);
2019
2020 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2021 }
2022
2023 return (0);
2024 }
2025
2026 /*
2027 * Free each object in magazine mp to cp's slab layer, and free mp itself.
2028 */
2029 static void
2030 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2031 {
2032 int round;
2033
2034 ASSERT(!list_link_active(&cp->cache_link) ||
2035 taskq_member(kmem_taskq, curthread));
2036
2037 for (round = 0; round < nrounds; round++) {
2038 void *buf = mp->mag_round[round];
2039
2040 if (cp->cache_flags & KMF_DEADBEEF) {
2041 if (verify_pattern(KMEM_FREE_PATTERN, buf,
2042 cp->cache_verify) != NULL) {
2043 kmem_error(KMERR_MODIFIED, cp, buf);
2044 continue;
2045 }
2046 if ((cp->cache_flags & KMF_LITE) &&
2047 cp->cache_destructor != NULL) {
2048 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2049 *(uint64_t *)buf = btp->bt_redzone;
2050 cp->cache_destructor(buf, cp->cache_private);
2051 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2052 }
2053 } else if (cp->cache_destructor != NULL) {
2054 cp->cache_destructor(buf, cp->cache_private);
2055 }
2056
2057 kmem_slab_free(cp, buf);
2058 }
2059 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2060 kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2061 }
2062
2063 /*
2064 * Allocate a magazine from the depot.
2065 */
2066 static kmem_magazine_t *
2067 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2068 {
2069 kmem_magazine_t *mp;
2070
2071 /*
2072 * If we can't get the depot lock without contention,
2073 * update our contention count. We use the depot
2074 * contention rate to determine whether we need to
2075 * increase the magazine size for better scalability.
2076 */
2077 if (!mutex_tryenter(&cp->cache_depot_lock)) {
2078 mutex_enter(&cp->cache_depot_lock);
2079 cp->cache_depot_contention++;
2080 }
2081
2082 if ((mp = mlp->ml_list) != NULL) {
2083 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2084 mlp->ml_list = mp->mag_next;
2085 if (--mlp->ml_total < mlp->ml_min)
2086 mlp->ml_min = mlp->ml_total;
2087 mlp->ml_alloc++;
2088 }
2089
2090 mutex_exit(&cp->cache_depot_lock);
2091
2092 return (mp);
2093 }
2094
2095 /*
2096 * Free a magazine to the depot.
2097 */
2098 static void
2099 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2100 {
2101 mutex_enter(&cp->cache_depot_lock);
2102 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2103 mp->mag_next = mlp->ml_list;
2104 mlp->ml_list = mp;
2105 mlp->ml_total++;
2106 mutex_exit(&cp->cache_depot_lock);
2107 }
2108
2109 /*
2110 * Update the working set statistics for cp's depot.
2111 */
2112 static void
2113 kmem_depot_ws_update(kmem_cache_t *cp)
2114 {
2115 mutex_enter(&cp->cache_depot_lock);
2116 cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2117 cp->cache_full.ml_min = cp->cache_full.ml_total;
2118 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2119 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2120 mutex_exit(&cp->cache_depot_lock);
2121 }
2122
2123 /*
2124 * Set the working set statistics for cp's depot to zero. (Everything is
2125 * eligible for reaping.)
2126 */
2127 static void
2128 kmem_depot_ws_zero(kmem_cache_t *cp)
2129 {
2130 mutex_enter(&cp->cache_depot_lock);
2131 cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2132 cp->cache_full.ml_min = cp->cache_full.ml_total;
2133 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2134 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2135 mutex_exit(&cp->cache_depot_lock);
2136 }
2137
2138 /*
2139 * The number of bytes to reap before we call kpreempt(). The default (1MB)
2140 * causes us to preempt reaping up to hundreds of times per second. Using a
2141 * larger value (1GB) causes this to have virtually no effect.
2142 */
2143 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2144
2145 /*
2146 * Reap all magazines that have fallen out of the depot's working set.
2147 */
2148 static void
2149 kmem_depot_ws_reap(kmem_cache_t *cp)
2150 {
2151 size_t bytes = 0;
2152 long reap;
2153 kmem_magazine_t *mp;
2154
2155 ASSERT(!list_link_active(&cp->cache_link) ||
2156 taskq_member(kmem_taskq, curthread));
2157
2158 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2159 while (reap-- &&
2160 (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2161 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2162 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2163 if (bytes > kmem_reap_preempt_bytes) {
2164 kpreempt(KPREEMPT_SYNC);
2165 bytes = 0;
2166 }
2167 }
2168
2169 reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2170 while (reap-- &&
2171 (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2172 kmem_magazine_destroy(cp, mp, 0);
2173 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2174 if (bytes > kmem_reap_preempt_bytes) {
2175 kpreempt(KPREEMPT_SYNC);
2176 bytes = 0;
2177 }
2178 }
2179 }
2180
2181 static void
2182 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2183 {
2184 ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2185 (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2186 ASSERT(ccp->cc_magsize > 0);
2187
2188 ccp->cc_ploaded = ccp->cc_loaded;
2189 ccp->cc_prounds = ccp->cc_rounds;
2190 ccp->cc_loaded = mp;
2191 ccp->cc_rounds = rounds;
2192 }
2193
2194 /*
2195 * Intercept kmem alloc/free calls during crash dump in order to avoid
2196 * changing kmem state while memory is being saved to the dump device.
2197 * Otherwise, ::kmem_verify will report "corrupt buffers". Note that
2198 * there are no locks because only one CPU calls kmem during a crash
2199 * dump. To enable this feature, first create the associated vmem
2200 * arena with VMC_DUMPSAFE.
2201 */
2202 static void *kmem_dump_start; /* start of pre-reserved heap */
2203 static void *kmem_dump_end; /* end of heap area */
2204 static void *kmem_dump_curr; /* current free heap pointer */
2205 static size_t kmem_dump_size; /* size of heap area */
2206
2207 /* append to each buf created in the pre-reserved heap */
2208 typedef struct kmem_dumpctl {
2209 void *kdc_next; /* cache dump free list linkage */
2210 } kmem_dumpctl_t;
2211
2212 #define KMEM_DUMPCTL(cp, buf) \
2213 ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2214 sizeof (void *)))
2215
2216 /* set non zero for full report */
2217 uint_t kmem_dump_verbose = 0;
2218
2219 /* stats for overize heap */
2220 uint_t kmem_dump_oversize_allocs = 0;
2221 uint_t kmem_dump_oversize_max = 0;
2222
2223 static void
2224 kmem_dumppr(char **pp, char *e, const char *format, ...)
2225 {
2226 char *p = *pp;
2227
2228 if (p < e) {
2229 int n;
2230 va_list ap;
2231
2232 va_start(ap, format);
2233 n = vsnprintf(p, e - p, format, ap);
2234 va_end(ap);
2235 *pp = p + n;
2236 }
2237 }
2238
2239 /*
2240 * Called when dumpadm(1M) configures dump parameters.
2241 */
2242 void
2243 kmem_dump_init(size_t size)
2244 {
2245 /* Our caller ensures size is always set. */
2246 ASSERT3U(size, >, 0);
2247
2248 if (kmem_dump_start != NULL)
2249 kmem_free(kmem_dump_start, kmem_dump_size);
2250
2251 kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2252 kmem_dump_size = size;
2253 kmem_dump_curr = kmem_dump_start;
2254 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2255 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2256 }
2257
2258 /*
2259 * Set flag for each kmem_cache_t if is safe to use alternate dump
2260 * memory. Called just before panic crash dump starts. Set the flag
2261 * for the calling CPU.
2262 */
2263 void
2264 kmem_dump_begin(void)
2265 {
2266 kmem_cache_t *cp;
2267
2268 ASSERT(panicstr != NULL);
2269
2270 for (cp = list_head(&kmem_caches); cp != NULL;
2271 cp = list_next(&kmem_caches, cp)) {
2272 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2273
2274 if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2275 cp->cache_flags |= KMF_DUMPDIVERT;
2276 ccp->cc_flags |= KMF_DUMPDIVERT;
2277 ccp->cc_dump_rounds = ccp->cc_rounds;
2278 ccp->cc_dump_prounds = ccp->cc_prounds;
2279 ccp->cc_rounds = ccp->cc_prounds = -1;
2280 } else {
2281 cp->cache_flags |= KMF_DUMPUNSAFE;
2282 ccp->cc_flags |= KMF_DUMPUNSAFE;
2283 }
2284 }
2285 }
2286
2287 /*
2288 * finished dump intercept
2289 * print any warnings on the console
2290 * return verbose information to dumpsys() in the given buffer
2291 */
2292 size_t
2293 kmem_dump_finish(char *buf, size_t size)
2294 {
2295 int percent = 0;
2296 size_t used;
2297 char *e = buf + size;
2298 char *p = buf;
2299
2300 if (kmem_dump_curr == kmem_dump_end) {
2301 cmn_err(CE_WARN, "exceeded kmem_dump space of %lu "
2302 "bytes: kmem state in dump may be inconsistent",
2303 kmem_dump_size);
2304 }
2305
2306 if (kmem_dump_verbose == 0)
2307 return (0);
2308
2309 used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2310 percent = (used * 100) / kmem_dump_size;
2311
2312 kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2313 kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2314 kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2315 kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2316 kmem_dump_oversize_allocs);
2317 kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2318 kmem_dump_oversize_max);
2319
2320 /* return buffer size used */
2321 if (p < e)
2322 bzero(p, e - p);
2323 return (p - buf);
2324 }
2325
2326 /*
2327 * Allocate a constructed object from alternate dump memory.
2328 */
2329 void *
2330 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2331 {
2332 void *buf;
2333 void *curr;
2334 char *bufend;
2335
2336 /* return a constructed object */
2337 if ((buf = cp->cache_dump.kd_freelist) != NULL) {
2338 cp->cache_dump.kd_freelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2339 return (buf);
2340 }
2341
2342 /* create a new constructed object */
2343 curr = kmem_dump_curr;
2344 buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2345 bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2346
2347 /* hat layer objects cannot cross a page boundary */
2348 if (cp->cache_align < PAGESIZE) {
2349 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2350 if (bufend > page) {
2351 bufend += page - (char *)buf;
2352 buf = (void *)page;
2353 }
2354 }
2355
2356 /* fall back to normal alloc if reserved area is used up */
2357 if (bufend > (char *)kmem_dump_end) {
2358 kmem_dump_curr = kmem_dump_end;
2359 cp->cache_dump.kd_alloc_fails++;
2360 return (NULL);
2361 }
2362
2363 /*
2364 * Must advance curr pointer before calling a constructor that
2365 * may also allocate memory.
2366 */
2367 kmem_dump_curr = bufend;
2368
2369 /* run constructor */
2370 if (cp->cache_constructor != NULL &&
2371 cp->cache_constructor(buf, cp->cache_private, kmflag)
2372 != 0) {
2373 #ifdef DEBUG
2374 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2375 cp->cache_name, (void *)cp);
2376 #endif
2377 /* reset curr pointer iff no allocs were done */
2378 if (kmem_dump_curr == bufend)
2379 kmem_dump_curr = curr;
2380
2381 cp->cache_dump.kd_alloc_fails++;
2382 /* fall back to normal alloc if the constructor fails */
2383 return (NULL);
2384 }
2385
2386 return (buf);
2387 }
2388
2389 /*
2390 * Free a constructed object in alternate dump memory.
2391 */
2392 int
2393 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2394 {
2395 /* save constructed buffers for next time */
2396 if ((char *)buf >= (char *)kmem_dump_start &&
2397 (char *)buf < (char *)kmem_dump_end) {
2398 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dump.kd_freelist;
2399 cp->cache_dump.kd_freelist = buf;
2400 return (0);
2401 }
2402
2403 /* just drop buffers that were allocated before dump started */
2404 if (kmem_dump_curr < kmem_dump_end)
2405 return (0);
2406
2407 /* fall back to normal free if reserved area is used up */
2408 return (1);
2409 }
2410
2411 /*
2412 * Allocate a constructed object from cache cp.
2413 */
2414 void *
2415 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2416 {
2417 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2418 kmem_magazine_t *fmp;
2419 void *buf;
2420
2421 mutex_enter(&ccp->cc_lock);
2422 for (;;) {
2423 /*
2424 * If there's an object available in the current CPU's
2425 * loaded magazine, just take it and return.
2426 */
2427 if (ccp->cc_rounds > 0) {
2428 buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2429 ccp->cc_alloc++;
2430 mutex_exit(&ccp->cc_lock);
2431 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2432 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2433 ASSERT(!(ccp->cc_flags &
2434 KMF_DUMPDIVERT));
2435 cp->cache_dump.kd_unsafe++;
2436 }
2437 if ((ccp->cc_flags & KMF_BUFTAG) &&
2438 kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2439 caller()) != 0) {
2440 if (kmflag & KM_NOSLEEP)
2441 return (NULL);
2442 mutex_enter(&ccp->cc_lock);
2443 continue;
2444 }
2445 }
2446 return (buf);
2447 }
2448
2449 /*
2450 * The loaded magazine is empty. If the previously loaded
2451 * magazine was full, exchange them and try again.
2452 */
2453 if (ccp->cc_prounds > 0) {
2454 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2455 continue;
2456 }
2457
2458 /*
2459 * Return an alternate buffer at dump time to preserve
2460 * the heap.
2461 */
2462 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2463 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2464 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2465 /* log it so that we can warn about it */
2466 cp->cache_dump.kd_unsafe++;
2467 } else {
2468 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2469 NULL) {
2470 mutex_exit(&ccp->cc_lock);
2471 return (buf);
2472 }
2473 break; /* fall back to slab layer */
2474 }
2475 }
2476
2477 /*
2478 * If the magazine layer is disabled, break out now.
2479 */
2480 if (ccp->cc_magsize == 0)
2481 break;
2482
2483 /*
2484 * Try to get a full magazine from the depot.
2485 */
2486 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2487 if (fmp != NULL) {
2488 if (ccp->cc_ploaded != NULL)
2489 kmem_depot_free(cp, &cp->cache_empty,
2490 ccp->cc_ploaded);
2491 kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2492 continue;
2493 }
2494
2495 /*
2496 * There are no full magazines in the depot,
2497 * so fall through to the slab layer.
2498 */
2499 break;
2500 }
2501 mutex_exit(&ccp->cc_lock);
2502
2503 /*
2504 * We couldn't allocate a constructed object from the magazine layer,
2505 * so get a raw buffer from the slab layer and apply its constructor.
2506 */
2507 buf = kmem_slab_alloc(cp, kmflag);
2508
2509 if (buf == NULL)
2510 return (NULL);
2511
2512 if (cp->cache_flags & KMF_BUFTAG) {
2513 /*
2514 * Make kmem_cache_alloc_debug() apply the constructor for us.
2515 */
2516 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2517 if (rc != 0) {
2518 if (kmflag & KM_NOSLEEP)
2519 return (NULL);
2520 /*
2521 * kmem_cache_alloc_debug() detected corruption
2522 * but didn't panic (kmem_panic <= 0). We should not be
2523 * here because the constructor failed (indicated by a
2524 * return code of 1). Try again.
2525 */
2526 ASSERT(rc == -1);
2527 return (kmem_cache_alloc(cp, kmflag));
2528 }
2529 return (buf);
2530 }
2531
2532 if (cp->cache_constructor != NULL &&
2533 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2534 atomic_inc_64(&cp->cache_alloc_fail);
2535 kmem_slab_free(cp, buf);
2536 return (NULL);
2537 }
2538
2539 return (buf);
2540 }
2541
2542 /*
2543 * The freed argument tells whether or not kmem_cache_free_debug() has already
2544 * been called so that we can avoid the duplicate free error. For example, a
2545 * buffer on a magazine has already been freed by the client but is still
2546 * constructed.
2547 */
2548 static void
2549 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2550 {
2551 if (!freed && (cp->cache_flags & KMF_BUFTAG))
2552 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2553 return;
2554
2555 /*
2556 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2557 * kmem_cache_free_debug() will have already applied the destructor.
2558 */
2559 if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2560 cp->cache_destructor != NULL) {
2561 if (cp->cache_flags & KMF_DEADBEEF) { /* KMF_LITE implied */
2562 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2563 *(uint64_t *)buf = btp->bt_redzone;
2564 cp->cache_destructor(buf, cp->cache_private);
2565 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2566 } else {
2567 cp->cache_destructor(buf, cp->cache_private);
2568 }
2569 }
2570
2571 kmem_slab_free(cp, buf);
2572 }
2573
2574 /*
2575 * Used when there's no room to free a buffer to the per-CPU cache.
2576 * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2577 * caller should try freeing to the per-CPU cache again.
2578 * Note that we don't directly install the magazine in the cpu cache,
2579 * since its state may have changed wildly while the lock was dropped.
2580 */
2581 static int
2582 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2583 {
2584 kmem_magazine_t *emp;
2585 kmem_magtype_t *mtp;
2586
2587 ASSERT(MUTEX_HELD(&ccp->cc_lock));
2588 ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2589 ((uint_t)ccp->cc_rounds == -1)) &&
2590 ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2591 ((uint_t)ccp->cc_prounds == -1)));
2592
2593 emp = kmem_depot_alloc(cp, &cp->cache_empty);
2594 if (emp != NULL) {
2595 if (ccp->cc_ploaded != NULL)
2596 kmem_depot_free(cp, &cp->cache_full,
2597 ccp->cc_ploaded);
2598 kmem_cpu_reload(ccp, emp, 0);
2599 return (1);
2600 }
2601 /*
2602 * There are no empty magazines in the depot,
2603 * so try to allocate a new one. We must drop all locks
2604 * across kmem_cache_alloc() because lower layers may
2605 * attempt to allocate from this cache.
2606 */
2607 mtp = cp->cache_magtype;
2608 mutex_exit(&ccp->cc_lock);
2609 emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2610 mutex_enter(&ccp->cc_lock);
2611
2612 if (emp != NULL) {
2613 /*
2614 * We successfully allocated an empty magazine.
2615 * However, we had to drop ccp->cc_lock to do it,
2616 * so the cache's magazine size may have changed.
2617 * If so, free the magazine and try again.
2618 */
2619 if (ccp->cc_magsize != mtp->mt_magsize) {
2620 mutex_exit(&ccp->cc_lock);
2621 kmem_cache_free(mtp->mt_cache, emp);
2622 mutex_enter(&ccp->cc_lock);
2623 return (1);
2624 }
2625
2626 /*
2627 * We got a magazine of the right size. Add it to
2628 * the depot and try the whole dance again.
2629 */
2630 kmem_depot_free(cp, &cp->cache_empty, emp);
2631 return (1);
2632 }
2633
2634 /*
2635 * We couldn't allocate an empty magazine,
2636 * so fall through to the slab layer.
2637 */
2638 return (0);
2639 }
2640
2641 /*
2642 * Free a constructed object to cache cp.
2643 */
2644 void
2645 kmem_cache_free(kmem_cache_t *cp, void *buf)
2646 {
2647 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2648
2649 /*
2650 * The client must not free either of the buffers passed to the move
2651 * callback function.
2652 */
2653 ASSERT(cp->cache_defrag == NULL ||
2654 cp->cache_defrag->kmd_thread != curthread ||
2655 (buf != cp->cache_defrag->kmd_from_buf &&
2656 buf != cp->cache_defrag->kmd_to_buf));
2657
2658 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2659 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2660 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2661 /* log it so that we can warn about it */
2662 cp->cache_dump.kd_unsafe++;
2663 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2664 return;
2665 }
2666 if (ccp->cc_flags & KMF_BUFTAG) {
2667 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2668 return;
2669 }
2670 }
2671
2672 mutex_enter(&ccp->cc_lock);
2673 /*
2674 * Any changes to this logic should be reflected in kmem_slab_prefill()
2675 */
2676 for (;;) {
2677 /*
2678 * If there's a slot available in the current CPU's
2679 * loaded magazine, just put the object there and return.
2680 */
2681 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2682 ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2683 ccp->cc_free++;
2684 mutex_exit(&ccp->cc_lock);
2685 return;
2686 }
2687
2688 /*
2689 * The loaded magazine is full. If the previously loaded
2690 * magazine was empty, exchange them and try again.
2691 */
2692 if (ccp->cc_prounds == 0) {
2693 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2694 continue;
2695 }
2696
2697 /*
2698 * If the magazine layer is disabled, break out now.
2699 */
2700 if (ccp->cc_magsize == 0)
2701 break;
2702
2703 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2704 /*
2705 * We couldn't free our constructed object to the
2706 * magazine layer, so apply its destructor and free it
2707 * to the slab layer.
2708 */
2709 break;
2710 }
2711 }
2712 mutex_exit(&ccp->cc_lock);
2713 kmem_slab_free_constructed(cp, buf, B_TRUE);
2714 }
2715
2716 static void
2717 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2718 {
2719 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2720 int cache_flags = cp->cache_flags;
2721
2722 kmem_bufctl_t *next, *head;
2723 size_t nbufs;
2724
2725 /*
2726 * Completely allocate the newly created slab and put the pre-allocated
2727 * buffers in magazines. Any of the buffers that cannot be put in
2728 * magazines must be returned to the slab.
2729 */
2730 ASSERT(MUTEX_HELD(&cp->cache_lock));
2731 ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2732 ASSERT(cp->cache_constructor == NULL);
2733 ASSERT(sp->slab_cache == cp);
2734 ASSERT(sp->slab_refcnt == 1);
2735 ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2736 ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2737
2738 head = sp->slab_head;
2739 nbufs = (sp->slab_chunks - sp->slab_refcnt);
2740 sp->slab_head = NULL;
2741 sp->slab_refcnt += nbufs;
2742 cp->cache_bufslab -= nbufs;
2743 cp->cache_slab_alloc += nbufs;
2744 list_insert_head(&cp->cache_complete_slabs, sp);
2745 cp->cache_complete_slab_count++;
2746 mutex_exit(&cp->cache_lock);
2747 mutex_enter(&ccp->cc_lock);
2748
2749 while (head != NULL) {
2750 void *buf = KMEM_BUF(cp, head);
2751 /*
2752 * If there's a slot available in the current CPU's
2753 * loaded magazine, just put the object there and
2754 * continue.
2755 */
2756 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2757 ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2758 buf;
2759 ccp->cc_free++;
2760 nbufs--;
2761 head = head->bc_next;
2762 continue;
2763 }
2764
2765 /*
2766 * The loaded magazine is full. If the previously
2767 * loaded magazine was empty, exchange them and try
2768 * again.
2769 */
2770 if (ccp->cc_prounds == 0) {
2771 kmem_cpu_reload(ccp, ccp->cc_ploaded,
2772 ccp->cc_prounds);
2773 continue;
2774 }
2775
2776 /*
2777 * If the magazine layer is disabled, break out now.
2778 */
2779
2780 if (ccp->cc_magsize == 0) {
2781 break;
2782 }
2783
2784 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2785 break;
2786 }
2787 mutex_exit(&ccp->cc_lock);
2788 if (nbufs != 0) {
2789 ASSERT(head != NULL);
2790
2791 /*
2792 * If there was a failure, return remaining objects to
2793 * the slab
2794 */
2795 while (head != NULL) {
2796 ASSERT(nbufs != 0);
2797 next = head->bc_next;
2798 head->bc_next = NULL;
2799 kmem_slab_free(cp, KMEM_BUF(cp, head));
2800 head = next;
2801 nbufs--;
2802 }
2803 }
2804 ASSERT(head == NULL);
2805 ASSERT(nbufs == 0);
2806 mutex_enter(&cp->cache_lock);
2807 }
2808
2809 void *
2810 kmem_zalloc(size_t size, int kmflag)
2811 {
2812 size_t index;
2813 void *buf;
2814
2815 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2816 kmem_cache_t *cp = kmem_alloc_table[index];
2817 buf = kmem_cache_alloc(cp, kmflag);
2818 if (buf != NULL) {
2819 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2820 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2821 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2822 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2823
2824 if (cp->cache_flags & KMF_LITE) {
2825 KMEM_BUFTAG_LITE_ENTER(btp,
2826 kmem_lite_count, caller());
2827 }
2828 }
2829 bzero(buf, size);
2830 }
2831 } else {
2832 buf = kmem_alloc(size, kmflag);
2833 if (buf != NULL)
2834 bzero(buf, size);
2835 }
2836 return (buf);
2837 }
2838
2839 void *
2840 kmem_alloc(size_t size, int kmflag)
2841 {
2842 size_t index;
2843 kmem_cache_t *cp;
2844 void *buf;
2845
2846 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2847 cp = kmem_alloc_table[index];
2848 /* fall through to kmem_cache_alloc() */
2849
2850 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2851 kmem_big_alloc_table_max) {
2852 cp = kmem_big_alloc_table[index];
2853 /* fall through to kmem_cache_alloc() */
2854
2855 } else {
2856 if (size == 0)
2857 return (NULL);
2858
2859 buf = vmem_alloc(kmem_oversize_arena, size,
2860 kmflag & KM_VMFLAGS);
2861 if (buf == NULL)
2862 kmem_log_event(kmem_failure_log, NULL, NULL,
2863 (void *)size);
2864 else if (KMEM_DUMP(kmem_slab_cache)) {
2865 /* stats for dump intercept */
2866 kmem_dump_oversize_allocs++;
2867 if (size > kmem_dump_oversize_max)
2868 kmem_dump_oversize_max = size;
2869 }
2870 return (buf);
2871 }
2872
2873 buf = kmem_cache_alloc(cp, kmflag);
2874 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2875 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2876 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2877 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2878
2879 if (cp->cache_flags & KMF_LITE) {
2880 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2881 }
2882 }
2883 return (buf);
2884 }
2885
2886 void
2887 kmem_free(void *buf, size_t size)
2888 {
2889 size_t index;
2890 kmem_cache_t *cp;
2891
2892 if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2893 cp = kmem_alloc_table[index];
2894 /* fall through to kmem_cache_free() */
2895
2896 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2897 kmem_big_alloc_table_max) {
2898 cp = kmem_big_alloc_table[index];
2899 /* fall through to kmem_cache_free() */
2900
2901 } else {
2902 EQUIV(buf == NULL, size == 0);
2903 if (buf == NULL && size == 0)
2904 return;
2905 vmem_free(kmem_oversize_arena, buf, size);
2906 return;
2907 }
2908
2909 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2910 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2911 uint32_t *ip = (uint32_t *)btp;
2912 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2913 if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2914 kmem_error(KMERR_DUPFREE, cp, buf);
2915 return;
2916 }
2917 if (KMEM_SIZE_VALID(ip[1])) {
2918 ip[0] = KMEM_SIZE_ENCODE(size);
2919 kmem_error(KMERR_BADSIZE, cp, buf);
2920 } else {
2921 kmem_error(KMERR_REDZONE, cp, buf);
2922 }
2923 return;
2924 }
2925 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2926 kmem_error(KMERR_REDZONE, cp, buf);
2927 return;
2928 }
2929 btp->bt_redzone = KMEM_REDZONE_PATTERN;
2930 if (cp->cache_flags & KMF_LITE) {
2931 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2932 caller());
2933 }
2934 }
2935 kmem_cache_free(cp, buf);
2936 }
2937
2938 void *
2939 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2940 {
2941 size_t realsize = size + vmp->vm_quantum;
2942 void *addr;
2943
2944 /*
2945 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2946 * vm_quantum will cause integer wraparound. Check for this, and
2947 * blow off the firewall page in this case. Note that such a
2948 * giant allocation (the entire kernel address space) can never
2949 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
2950 * or sleep forever (VM_SLEEP). Thus, there is no need for a
2951 * corresponding check in kmem_firewall_va_free().
2952 */
2953 if (realsize < size)
2954 realsize = size;
2955
2956 /*
2957 * While boot still owns resource management, make sure that this
2958 * redzone virtual address allocation is properly accounted for in
2959 * OBPs "virtual-memory" "available" lists because we're
2960 * effectively claiming them for a red zone. If we don't do this,
2961 * the available lists become too fragmented and too large for the
2962 * current boot/kernel memory list interface.
2963 */
2964 addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
2965
2966 if (addr != NULL && kvseg.s_base == NULL && realsize != size)
2967 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
2968
2969 return (addr);
2970 }
2971
2972 void
2973 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2974 {
2975 ASSERT((kvseg.s_base == NULL ?
2976 va_to_pfn((char *)addr + size) :
2977 hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
2978
2979 vmem_free(vmp, addr, size + vmp->vm_quantum);
2980 }
2981
2982 /*
2983 * Try to allocate at least `size' bytes of memory without sleeping or
2984 * panicking. Return actual allocated size in `asize'. If allocation failed,
2985 * try final allocation with sleep or panic allowed.
2986 */
2987 void *
2988 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
2989 {
2990 void *p;
2991
2992 *asize = P2ROUNDUP(size, KMEM_ALIGN);
2993 do {
2994 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
2995 if (p != NULL)
2996 return (p);
2997 *asize += KMEM_ALIGN;
2998 } while (*asize <= PAGESIZE);
2999
3000 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3001 return (kmem_alloc(*asize, kmflag));
3002 }
3003
3004 /*
3005 * Reclaim all unused memory from a cache.
3006 */
3007 static void
3008 kmem_cache_reap(kmem_cache_t *cp)
3009 {
3010 ASSERT(taskq_member(kmem_taskq, curthread));
3011 cp->cache_reap++;
3012
3013 /*
3014 * Ask the cache's owner to free some memory if possible.
3015 * The idea is to handle things like the inode cache, which
3016 * typically sits on a bunch of memory that it doesn't truly
3017 * *need*. Reclaim policy is entirely up to the owner; this
3018 * callback is just an advisory plea for help.
3019 */
3020 if (cp->cache_reclaim != NULL) {
3021 long delta;
3022
3023 /*
3024 * Reclaimed memory should be reapable (not included in the
3025 * depot's working set).
3026 */
3027 delta = cp->cache_full.ml_total;
3028 cp->cache_reclaim(cp->cache_private);
3029 delta = cp->cache_full.ml_total - delta;
3030 if (delta > 0) {
3031 mutex_enter(&cp->cache_depot_lock);
3032 cp->cache_full.ml_reaplimit += delta;
3033 cp->cache_full.ml_min += delta;
3034 mutex_exit(&cp->cache_depot_lock);
3035 }
3036 }
3037
3038 kmem_depot_ws_reap(cp);
3039
3040 if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3041 kmem_cache_defrag(cp);
3042 }
3043 }
3044
3045 static void
3046 kmem_reap_timeout(void *flag_arg)
3047 {
3048 uint32_t *flag = (uint32_t *)flag_arg;
3049
3050 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3051 *flag = 0;
3052 }
3053
3054 static void
3055 kmem_reap_done(void *flag)
3056 {
3057 if (!callout_init_done) {
3058 /* can't schedule a timeout at this point */
3059 kmem_reap_timeout(flag);
3060 } else {
3061 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3062 }
3063 }
3064
3065 static void
3066 kmem_reap_start(void *flag)
3067 {
3068 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3069
3070 if (flag == &kmem_reaping) {
3071 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3072 /*
3073 * if we have segkp under heap, reap segkp cache.
3074 */
3075 if (segkp_fromheap)
3076 segkp_cache_free();
3077 }
3078 else
3079 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3080
3081 /*
3082 * We use taskq_dispatch() to schedule a timeout to clear
3083 * the flag so that kmem_reap() becomes self-throttling:
3084 * we won't reap again until the current reap completes *and*
3085 * at least kmem_reap_interval ticks have elapsed.
3086 */
3087 if (taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP) ==
3088 TASKQID_INVALID)
3089 kmem_reap_done(flag);
3090 }
3091
3092 static void
3093 kmem_reap_common(void *flag_arg)
3094 {
3095 uint32_t *flag = (uint32_t *)flag_arg;
3096
3097 if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3098 atomic_cas_32(flag, 0, 1) != 0)
3099 return;
3100
3101 /*
3102 * It may not be kosher to do memory allocation when a reap is called
3103 * (for example, if vmem_populate() is in the call chain). So we
3104 * start the reap going with a TQ_NOALLOC dispatch. If the dispatch
3105 * fails, we reset the flag, and the next reap will try again.
3106 */
3107 if (taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC) ==
3108 TASKQID_INVALID)
3109 *flag = 0;
3110 }
3111
3112 /*
3113 * Reclaim all unused memory from all caches. Called from the VM system
3114 * when memory gets tight.
3115 */
3116 void
3117 kmem_reap(void)
3118 {
3119 kmem_reap_common(&kmem_reaping);
3120 }
3121
3122 /*
3123 * Reclaim all unused memory from identifier arenas, called when a vmem
3124 * arena not back by memory is exhausted. Since reaping memory-backed caches
3125 * cannot help with identifier exhaustion, we avoid both a large amount of
3126 * work and unwanted side-effects from reclaim callbacks.
3127 */
3128 void
3129 kmem_reap_idspace(void)
3130 {
3131 kmem_reap_common(&kmem_reaping_idspace);
3132 }
3133
3134 /*
3135 * Purge all magazines from a cache and set its magazine limit to zero.
3136 * All calls are serialized by the kmem_taskq lock, except for the final
3137 * call from kmem_cache_destroy().
3138 */
3139 static void
3140 kmem_cache_magazine_purge(kmem_cache_t *cp)
3141 {
3142 kmem_cpu_cache_t *ccp;
3143 kmem_magazine_t *mp, *pmp;
3144 int rounds, prounds, cpu_seqid;
3145
3146 ASSERT(!list_link_active(&cp->cache_link) ||
3147 taskq_member(kmem_taskq, curthread));
3148 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3149
3150 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3151 ccp = &cp->cache_cpu[cpu_seqid];
3152
3153 mutex_enter(&ccp->cc_lock);
3154 mp = ccp->cc_loaded;
3155 pmp = ccp->cc_ploaded;
3156 rounds = ccp->cc_rounds;
3157 prounds = ccp->cc_prounds;
3158 ccp->cc_loaded = NULL;
3159 ccp->cc_ploaded = NULL;
3160 ccp->cc_rounds = -1;
3161 ccp->cc_prounds = -1;
3162 ccp->cc_magsize = 0;
3163 mutex_exit(&ccp->cc_lock);
3164
3165 if (mp)
3166 kmem_magazine_destroy(cp, mp, rounds);
3167 if (pmp)
3168 kmem_magazine_destroy(cp, pmp, prounds);
3169 }
3170
3171 kmem_depot_ws_zero(cp);
3172 kmem_depot_ws_reap(cp);
3173 }
3174
3175 /*
3176 * Enable per-cpu magazines on a cache.
3177 */
3178 static void
3179 kmem_cache_magazine_enable(kmem_cache_t *cp)
3180 {
3181 int cpu_seqid;
3182
3183 if (cp->cache_flags & KMF_NOMAGAZINE)
3184 return;
3185
3186 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3187 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3188 mutex_enter(&ccp->cc_lock);
3189 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3190 mutex_exit(&ccp->cc_lock);
3191 }
3192
3193 }
3194
3195 /*
3196 * Allow our caller to determine if there are running reaps.
3197 *
3198 * This call is very conservative and may return B_TRUE even when
3199 * reaping activity isn't active. If it returns B_FALSE, then reaping
3200 * activity is definitely inactive.
3201 */
3202 boolean_t
3203 kmem_cache_reap_active(void)
3204 {
3205 return (!taskq_empty(kmem_taskq));
3206 }
3207
3208 /*
3209 * Reap (almost) everything soon.
3210 *
3211 * Note: this does not wait for the reap-tasks to complete. Caller
3212 * should use kmem_cache_reap_active() (above) and/or moderation to
3213 * avoid scheduling too many reap-tasks.
3214 */
3215 void
3216 kmem_cache_reap_soon(kmem_cache_t *cp)
3217 {
3218 ASSERT(list_link_active(&cp->cache_link));
3219
3220 kmem_depot_ws_zero(cp);
3221
3222 (void) taskq_dispatch(kmem_taskq,
3223 (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3224 }
3225
3226 /*
3227 * Recompute a cache's magazine size. The trade-off is that larger magazines
3228 * provide a higher transfer rate with the depot, while smaller magazines
3229 * reduce memory consumption. Magazine resizing is an expensive operation;
3230 * it should not be done frequently.
3231 *
3232 * Changes to the magazine size are serialized by the kmem_taskq lock.
3233 *
3234 * Note: at present this only grows the magazine size. It might be useful
3235 * to allow shrinkage too.
3236 */
3237 static void
3238 kmem_cache_magazine_resize(kmem_cache_t *cp)
3239 {
3240 kmem_magtype_t *mtp = cp->cache_magtype;
3241
3242 ASSERT(taskq_member(kmem_taskq, curthread));
3243
3244 if (cp->cache_chunksize < mtp->mt_maxbuf) {
3245 kmem_cache_magazine_purge(cp);
3246 mutex_enter(&cp->cache_depot_lock);
3247 cp->cache_magtype = ++mtp;
3248 cp->cache_depot_contention_prev =
3249 cp->cache_depot_contention + INT_MAX;
3250 mutex_exit(&cp->cache_depot_lock);
3251 kmem_cache_magazine_enable(cp);
3252 }
3253 }
3254
3255 /*
3256 * Rescale a cache's hash table, so that the table size is roughly the
3257 * cache size. We want the average lookup time to be extremely small.
3258 */
3259 static void
3260 kmem_hash_rescale(kmem_cache_t *cp)
3261 {
3262 kmem_bufctl_t **old_table, **new_table, *bcp;
3263 size_t old_size, new_size, h;
3264
3265 ASSERT(taskq_member(kmem_taskq, curthread));
3266
3267 new_size = MAX(KMEM_HASH_INITIAL,
3268 1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3269 old_size = cp->cache_hash_mask + 1;
3270
3271 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3272 return;
3273
3274 new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3275 VM_NOSLEEP);
3276 if (new_table == NULL)
3277 return;
3278 bzero(new_table, new_size * sizeof (void *));
3279
3280 mutex_enter(&cp->cache_lock);
3281
3282 old_size = cp->cache_hash_mask + 1;
3283 old_table = cp->cache_hash_table;
3284
3285 cp->cache_hash_mask = new_size - 1;
3286 cp->cache_hash_table = new_table;
3287 cp->cache_rescale++;
3288
3289 for (h = 0; h < old_size; h++) {
3290 bcp = old_table[h];
3291 while (bcp != NULL) {
3292 void *addr = bcp->bc_addr;
3293 kmem_bufctl_t *next_bcp = bcp->bc_next;
3294 kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3295 bcp->bc_next = *hash_bucket;
3296 *hash_bucket = bcp;
3297 bcp = next_bcp;
3298 }
3299 }
3300
3301 mutex_exit(&cp->cache_lock);
3302
3303 vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3304 }
3305
3306 /*
3307 * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3308 * update, magazine resizing, and slab consolidation.
3309 */
3310 static void
3311 kmem_cache_update(kmem_cache_t *cp)
3312 {
3313 int need_hash_rescale = 0;
3314 int need_magazine_resize = 0;
3315
3316 ASSERT(MUTEX_HELD(&kmem_cache_lock));
3317
3318 /*
3319 * If the cache has become much larger or smaller than its hash table,
3320 * fire off a request to rescale the hash table.
3321 */
3322 mutex_enter(&cp->cache_lock);
3323
3324 if ((cp->cache_flags & KMF_HASH) &&
3325 (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3326 (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3327 cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3328 need_hash_rescale = 1;
3329
3330 mutex_exit(&cp->cache_lock);
3331
3332 /*
3333 * Update the depot working set statistics.
3334 */
3335 kmem_depot_ws_update(cp);
3336
3337 /*
3338 * If there's a lot of contention in the depot,
3339 * increase the magazine size.
3340 */
3341 mutex_enter(&cp->cache_depot_lock);
3342
3343 if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3344 (int)(cp->cache_depot_contention -
3345 cp->cache_depot_contention_prev) > kmem_depot_contention)
3346 need_magazine_resize = 1;
3347
3348 cp->cache_depot_contention_prev = cp->cache_depot_contention;
3349
3350 mutex_exit(&cp->cache_depot_lock);
3351
3352 if (need_hash_rescale)
3353 (void) taskq_dispatch(kmem_taskq,
3354 (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3355
3356 if (need_magazine_resize)
3357 (void) taskq_dispatch(kmem_taskq,
3358 (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3359
3360 if (cp->cache_defrag != NULL)
3361 (void) taskq_dispatch(kmem_taskq,
3362 (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3363 }
3364
3365 static void kmem_update(void *);
3366
3367 static void
3368 kmem_update_timeout(void *dummy)
3369 {
3370 (void) timeout(kmem_update, dummy, kmem_reap_interval);
3371 }
3372
3373 static void
3374 kmem_update(void *dummy)
3375 {
3376 kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3377
3378 /*
3379 * We use taskq_dispatch() to reschedule the timeout so that
3380 * kmem_update() becomes self-throttling: it won't schedule
3381 * new tasks until all previous tasks have completed.
3382 */
3383 if (taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP)
3384 == TASKQID_INVALID)
3385 kmem_update_timeout(NULL);
3386 }
3387
3388 static int
3389 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3390 {
3391 struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3392 kmem_cache_t *cp = ksp->ks_private;
3393 uint64_t cpu_buf_avail;
3394 uint64_t buf_avail = 0;
3395 int cpu_seqid;
3396 long reap;
3397
3398 ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3399
3400 if (rw == KSTAT_WRITE)
3401 return (EACCES);
3402
3403 mutex_enter(&cp->cache_lock);
3404
3405 kmcp->kmc_alloc_fail.value.ui64 = cp->cache_alloc_fail;
3406 kmcp->kmc_alloc.value.ui64 = cp->cache_slab_alloc;
3407 kmcp->kmc_free.value.ui64 = cp->cache_slab_free;
3408 kmcp->kmc_slab_alloc.value.ui64 = cp->cache_slab_alloc;
3409 kmcp->kmc_slab_free.value.ui64 = cp->cache_slab_free;
3410
3411 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3412 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3413
3414 mutex_enter(&ccp->cc_lock);
3415
3416 cpu_buf_avail = 0;
3417 if (ccp->cc_rounds > 0)
3418 cpu_buf_avail += ccp->cc_rounds;
3419 if (ccp->cc_prounds > 0)
3420 cpu_buf_avail += ccp->cc_prounds;
3421
3422 kmcp->kmc_alloc.value.ui64 += ccp->cc_alloc;
3423 kmcp->kmc_free.value.ui64 += ccp->cc_free;
3424 buf_avail += cpu_buf_avail;
3425
3426 mutex_exit(&ccp->cc_lock);
3427 }
3428
3429 mutex_enter(&cp->cache_depot_lock);
3430
3431 kmcp->kmc_depot_alloc.value.ui64 = cp->cache_full.ml_alloc;
3432 kmcp->kmc_depot_free.value.ui64 = cp->cache_empty.ml_alloc;
3433 kmcp->kmc_depot_contention.value.ui64 = cp->cache_depot_contention;
3434 kmcp->kmc_full_magazines.value.ui64 = cp->cache_full.ml_total;
3435 kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3436 kmcp->kmc_magazine_size.value.ui64 =
3437 (cp->cache_flags & KMF_NOMAGAZINE) ?
3438 0 : cp->cache_magtype->mt_magsize;
3439
3440 kmcp->kmc_alloc.value.ui64 += cp->cache_full.ml_alloc;
3441 kmcp->kmc_free.value.ui64 += cp->cache_empty.ml_alloc;
3442 buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3443
3444 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3445 reap = MIN(reap, cp->cache_full.ml_total);
3446
3447 mutex_exit(&cp->cache_depot_lock);
3448
3449 kmcp->kmc_buf_size.value.ui64 = cp->cache_bufsize;
3450 kmcp->kmc_align.value.ui64 = cp->cache_align;
3451 kmcp->kmc_chunk_size.value.ui64 = cp->cache_chunksize;
3452 kmcp->kmc_slab_size.value.ui64 = cp->cache_slabsize;
3453 kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3454 buf_avail += cp->cache_bufslab;
3455 kmcp->kmc_buf_avail.value.ui64 = buf_avail;
3456 kmcp->kmc_buf_inuse.value.ui64 = cp->cache_buftotal - buf_avail;
3457 kmcp->kmc_buf_total.value.ui64 = cp->cache_buftotal;
3458 kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3459 kmcp->kmc_slab_create.value.ui64 = cp->cache_slab_create;
3460 kmcp->kmc_slab_destroy.value.ui64 = cp->cache_slab_destroy;
3461 kmcp->kmc_hash_size.value.ui64 = (cp->cache_flags & KMF_HASH) ?
3462 cp->cache_hash_mask + 1 : 0;
3463 kmcp->kmc_hash_lookup_depth.value.ui64 = cp->cache_lookup_depth;
3464 kmcp->kmc_hash_rescale.value.ui64 = cp->cache_rescale;
3465 kmcp->kmc_vmem_source.value.ui64 = cp->cache_arena->vm_id;
3466 kmcp->kmc_reap.value.ui64 = cp->cache_reap;
3467
3468 if (cp->cache_defrag == NULL) {
3469 kmcp->kmc_move_callbacks.value.ui64 = 0;
3470 kmcp->kmc_move_yes.value.ui64 = 0;
3471 kmcp->kmc_move_no.value.ui64 = 0;
3472 kmcp->kmc_move_later.value.ui64 = 0;
3473 kmcp->kmc_move_dont_need.value.ui64 = 0;
3474 kmcp->kmc_move_dont_know.value.ui64 = 0;
3475 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3476 kmcp->kmc_move_slabs_freed.value.ui64 = 0;
3477 kmcp->kmc_defrag.value.ui64 = 0;
3478 kmcp->kmc_scan.value.ui64 = 0;
3479 kmcp->kmc_move_reclaimable.value.ui64 = 0;
3480 } else {
3481 int64_t reclaimable;
3482
3483 kmem_defrag_t *kd = cp->cache_defrag;
3484 kmcp->kmc_move_callbacks.value.ui64 = kd->kmd_callbacks;
3485 kmcp->kmc_move_yes.value.ui64 = kd->kmd_yes;
3486 kmcp->kmc_move_no.value.ui64 = kd->kmd_no;
3487 kmcp->kmc_move_later.value.ui64 = kd->kmd_later;
3488 kmcp->kmc_move_dont_need.value.ui64 = kd->kmd_dont_need;
3489 kmcp->kmc_move_dont_know.value.ui64 = kd->kmd_dont_know;
3490 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3491 kmcp->kmc_move_slabs_freed.value.ui64 = kd->kmd_slabs_freed;
3492 kmcp->kmc_defrag.value.ui64 = kd->kmd_defrags;
3493 kmcp->kmc_scan.value.ui64 = kd->kmd_scans;
3494
3495 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3496 reclaimable = MAX(reclaimable, 0);
3497 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3498 kmcp->kmc_move_reclaimable.value.ui64 = reclaimable;
3499 }
3500
3501 mutex_exit(&cp->cache_lock);
3502 return (0);
3503 }
3504
3505 /*
3506 * Return a named statistic about a particular cache.
3507 * This shouldn't be called very often, so it's currently designed for
3508 * simplicity (leverages existing kstat support) rather than efficiency.
3509 */
3510 uint64_t
3511 kmem_cache_stat(kmem_cache_t *cp, char *name)
3512 {
3513 int i;
3514 kstat_t *ksp = cp->cache_kstat;
3515 kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3516 uint64_t value = 0;
3517
3518 if (ksp != NULL) {
3519 mutex_enter(&kmem_cache_kstat_lock);
3520 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3521 for (i = 0; i < ksp->ks_ndata; i++) {
3522 if (strcmp(knp[i].name, name) == 0) {
3523 value = knp[i].value.ui64;
3524 break;
3525 }
3526 }
3527 mutex_exit(&kmem_cache_kstat_lock);
3528 }
3529 return (value);
3530 }
3531
3532 /*
3533 * Return an estimate of currently available kernel heap memory.
3534 * On 32-bit systems, physical memory may exceed virtual memory,
3535 * we just truncate the result at 1GB.
3536 */
3537 size_t
3538 kmem_avail(void)
3539 {
3540 spgcnt_t rmem = availrmem - tune.t_minarmem;
3541 spgcnt_t fmem = freemem - minfree;
3542
3543 return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3544 1 << (30 - PAGESHIFT))));
3545 }
3546
3547 /*
3548 * Return the maximum amount of memory that is (in theory) allocatable
3549 * from the heap. This may be used as an estimate only since there
3550 * is no guarentee this space will still be available when an allocation
3551 * request is made, nor that the space may be allocated in one big request
3552 * due to kernel heap fragmentation.
3553 */
3554 size_t
3555 kmem_maxavail(void)
3556 {
3557 spgcnt_t pmem = availrmem - tune.t_minarmem;
3558 spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3559
3560 return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3561 }
3562
3563 /*
3564 * Indicate whether memory-intensive kmem debugging is enabled.
3565 */
3566 int
3567 kmem_debugging(void)
3568 {
3569 return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3570 }
3571
3572 /* binning function, sorts finely at the two extremes */
3573 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift) \
3574 ((((sp)->slab_refcnt <= (binshift)) || \
3575 (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift))) \
3576 ? -(sp)->slab_refcnt \
3577 : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3578
3579 /*
3580 * Minimizing the number of partial slabs on the freelist minimizes
3581 * fragmentation (the ratio of unused buffers held by the slab layer). There are
3582 * two ways to get a slab off of the freelist: 1) free all the buffers on the
3583 * slab, and 2) allocate all the buffers on the slab. It follows that we want
3584 * the most-used slabs at the front of the list where they have the best chance
3585 * of being completely allocated, and the least-used slabs at a safe distance
3586 * from the front to improve the odds that the few remaining buffers will all be
3587 * freed before another allocation can tie up the slab. For that reason a slab
3588 * with a higher slab_refcnt sorts less than than a slab with a lower
3589 * slab_refcnt.
3590 *
3591 * However, if a slab has at least one buffer that is deemed unfreeable, we
3592 * would rather have that slab at the front of the list regardless of
3593 * slab_refcnt, since even one unfreeable buffer makes the entire slab
3594 * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3595 * callback, the slab is marked unfreeable for as long as it remains on the
3596 * freelist.
3597 */
3598 static int
3599 kmem_partial_slab_cmp(const void *p0, const void *p1)
3600 {
3601 const kmem_cache_t *cp;
3602 const kmem_slab_t *s0 = p0;
3603 const kmem_slab_t *s1 = p1;
3604 int w0, w1;
3605 size_t binshift;
3606
3607 ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3608 ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3609 ASSERT(s0->slab_cache == s1->slab_cache);
3610 cp = s1->slab_cache;
3611 ASSERT(MUTEX_HELD(&cp->cache_lock));
3612 binshift = cp->cache_partial_binshift;
3613
3614 /* weight of first slab */
3615 w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3616 if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3617 w0 -= cp->cache_maxchunks;
3618 }
3619
3620 /* weight of second slab */
3621 w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3622 if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3623 w1 -= cp->cache_maxchunks;
3624 }
3625
3626 if (w0 < w1)
3627 return (-1);
3628 if (w0 > w1)
3629 return (1);
3630
3631 /* compare pointer values */
3632 if ((uintptr_t)s0 < (uintptr_t)s1)
3633 return (-1);
3634 if ((uintptr_t)s0 > (uintptr_t)s1)
3635 return (1);
3636
3637 return (0);
3638 }
3639
3640 /*
3641 * It must be valid to call the destructor (if any) on a newly created object.
3642 * That is, the constructor (if any) must leave the object in a valid state for
3643 * the destructor.
3644 */
3645 kmem_cache_t *
3646 kmem_cache_create(
3647 char *name, /* descriptive name for this cache */
3648 size_t bufsize, /* size of the objects it manages */
3649 size_t align, /* required object alignment */
3650 int (*constructor)(void *, void *, int), /* object constructor */
3651 void (*destructor)(void *, void *), /* object destructor */
3652 void (*reclaim)(void *), /* memory reclaim callback */
3653 void *private, /* pass-thru arg for constr/destr/reclaim */
3654 vmem_t *vmp, /* vmem source for slab allocation */
3655 int cflags) /* cache creation flags */
3656 {
3657 int cpu_seqid;
3658 size_t chunksize;
3659 kmem_cache_t *cp;
3660 kmem_magtype_t *mtp;
3661 size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3662
3663 #ifdef DEBUG
3664 /*
3665 * Cache names should conform to the rules for valid C identifiers
3666 */
3667 if (!strident_valid(name)) {
3668 cmn_err(CE_CONT,
3669 "kmem_cache_create: '%s' is an invalid cache name\n"
3670 "cache names must conform to the rules for "
3671 "C identifiers\n", name);
3672 }
3673 #endif /* DEBUG */
3674
3675 if (vmp == NULL)
3676 vmp = kmem_default_arena;
3677
3678 /*
3679 * If this kmem cache has an identifier vmem arena as its source, mark
3680 * it such to allow kmem_reap_idspace().
3681 */
3682 ASSERT(!(cflags & KMC_IDENTIFIER)); /* consumer should not set this */
3683 if (vmp->vm_cflags & VMC_IDENTIFIER)
3684 cflags |= KMC_IDENTIFIER;
3685
3686 /*
3687 * Get a kmem_cache structure. We arrange that cp->cache_cpu[]
3688 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3689 * false sharing of per-CPU data.
3690 */
3691 cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3692 P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3693 bzero(cp, csize);
3694 list_link_init(&cp->cache_link);
3695
3696 if (align == 0)
3697 align = KMEM_ALIGN;
3698
3699 /*
3700 * If we're not at least KMEM_ALIGN aligned, we can't use free
3701 * memory to hold bufctl information (because we can't safely
3702 * perform word loads and stores on it).
3703 */
3704 if (align < KMEM_ALIGN)
3705 cflags |= KMC_NOTOUCH;
3706
3707 if (!ISP2(align) || align > vmp->vm_quantum)
3708 panic("kmem_cache_create: bad alignment %lu", align);
3709
3710 mutex_enter(&kmem_flags_lock);
3711 if (kmem_flags & KMF_RANDOMIZE)
3712 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3713 KMF_RANDOMIZE;
3714 cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3715 mutex_exit(&kmem_flags_lock);
3716
3717 /*
3718 * Make sure all the various flags are reasonable.
3719 */
3720 ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3721
3722 if (cp->cache_flags & KMF_LITE) {
3723 if (bufsize >= kmem_lite_minsize &&
3724 align <= kmem_lite_maxalign &&
3725 P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3726 cp->cache_flags |= KMF_BUFTAG;
3727 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3728 } else {
3729 cp->cache_flags &= ~KMF_DEBUG;
3730 }
3731 }
3732
3733 if (cp->cache_flags & KMF_DEADBEEF)
3734 cp->cache_flags |= KMF_REDZONE;
3735
3736 if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3737 cp->cache_flags |= KMF_NOMAGAZINE;
3738
3739 if (cflags & KMC_NODEBUG)
3740 cp->cache_flags &= ~KMF_DEBUG;
3741
3742 if (cflags & KMC_NOTOUCH)
3743 cp->cache_flags &= ~KMF_TOUCH;
3744
3745 if (cflags & KMC_PREFILL)
3746 cp->cache_flags |= KMF_PREFILL;
3747
3748 if (cflags & KMC_NOHASH)
3749 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3750
3751 if (cflags & KMC_NOMAGAZINE)
3752 cp->cache_flags |= KMF_NOMAGAZINE;
3753
3754 if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3755 cp->cache_flags |= KMF_REDZONE;
3756
3757 if (!(cp->cache_flags & KMF_AUDIT))
3758 cp->cache_flags &= ~KMF_CONTENTS;
3759
3760 if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3761 !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3762 cp->cache_flags |= KMF_FIREWALL;
3763
3764 if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3765 cp->cache_flags &= ~KMF_FIREWALL;
3766
3767 if (cp->cache_flags & KMF_FIREWALL) {
3768 cp->cache_flags &= ~KMF_BUFTAG;
3769 cp->cache_flags |= KMF_NOMAGAZINE;
3770 ASSERT(vmp == kmem_default_arena);
3771 vmp = kmem_firewall_arena;
3772 }
3773
3774 /*
3775 * Set cache properties.
3776 */
3777 (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3778 strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3779 cp->cache_bufsize = bufsize;
3780 cp->cache_align = align;
3781 cp->cache_constructor = constructor;
3782 cp->cache_destructor = destructor;
3783 cp->cache_reclaim = reclaim;
3784 cp->cache_private = private;
3785 cp->cache_arena = vmp;
3786 cp->cache_cflags = cflags;
3787
3788 /*
3789 * Determine the chunk size.
3790 */
3791 chunksize = bufsize;
3792
3793 if (align >= KMEM_ALIGN) {
3794 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3795 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3796 }
3797
3798 if (cp->cache_flags & KMF_BUFTAG) {
3799 cp->cache_bufctl = chunksize;
3800 cp->cache_buftag = chunksize;
3801 if (cp->cache_flags & KMF_LITE)
3802 chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3803 else
3804 chunksize += sizeof (kmem_buftag_t);
3805 }
3806
3807 if (cp->cache_flags & KMF_DEADBEEF) {
3808 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3809 if (cp->cache_flags & KMF_LITE)
3810 cp->cache_verify = sizeof (uint64_t);
3811 }
3812
3813 cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3814
3815 cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3816
3817 /*
3818 * Now that we know the chunk size, determine the optimal slab size.
3819 */
3820 if (vmp == kmem_firewall_arena) {
3821 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3822 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3823 cp->cache_maxcolor = cp->cache_mincolor;
3824 cp->cache_flags |= KMF_HASH;
3825 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3826 } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3827 !(cp->cache_flags & KMF_AUDIT) &&
3828 chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3829 cp->cache_slabsize = vmp->vm_quantum;
3830 cp->cache_mincolor = 0;
3831 cp->cache_maxcolor =
3832 (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3833 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3834 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3835 } else {
3836 size_t chunks, bestfit, waste, slabsize;
3837 size_t minwaste = LONG_MAX;
3838
3839 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3840 slabsize = P2ROUNDUP(chunksize * chunks,
3841 vmp->vm_quantum);
3842 chunks = slabsize / chunksize;
3843 waste = (slabsize % chunksize) / chunks;
3844 if (waste < minwaste) {
3845 minwaste = waste;
3846 bestfit = slabsize;
3847 }
3848 }
3849 if (cflags & KMC_QCACHE)
3850 bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3851 cp->cache_slabsize = bestfit;
3852 cp->cache_mincolor = 0;
3853 cp->cache_maxcolor = bestfit % chunksize;
3854 cp->cache_flags |= KMF_HASH;
3855 }
3856
3857 cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3858 cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3859
3860 /*
3861 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3862 * there is a constructor avoids some tricky issues with debug setup
3863 * that may be revisited later. We cannot allow prefill in a
3864 * metadata cache because of potential recursion.
3865 */
3866 if (vmp == kmem_msb_arena ||
3867 cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3868 cp->cache_constructor != NULL)
3869 cp->cache_flags &= ~KMF_PREFILL;
3870
3871 if (cp->cache_flags & KMF_HASH) {
3872 ASSERT(!(cflags & KMC_NOHASH));
3873 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3874 kmem_bufctl_audit_cache : kmem_bufctl_cache;
3875 }
3876
3877 if (cp->cache_maxcolor >= vmp->vm_quantum)
3878 cp->cache_maxcolor = vmp->vm_quantum - 1;
3879
3880 cp->cache_color = cp->cache_mincolor;
3881
3882 /*
3883 * Initialize the rest of the slab layer.
3884 */
3885 mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3886
3887 avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3888 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3889 /* LINTED: E_TRUE_LOGICAL_EXPR */
3890 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3891 /* reuse partial slab AVL linkage for complete slab list linkage */
3892 list_create(&cp->cache_complete_slabs,
3893 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3894
3895 if (cp->cache_flags & KMF_HASH) {
3896 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3897 KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3898 bzero(cp->cache_hash_table,
3899 KMEM_HASH_INITIAL * sizeof (void *));
3900 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3901 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3902 }
3903
3904 /*
3905 * Initialize the depot.
3906 */
3907 mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3908
3909 for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3910 continue;
3911
3912 cp->cache_magtype = mtp;
3913
3914 /*
3915 * Initialize the CPU layer.
3916 */
3917 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3918 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3919 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3920 ccp->cc_flags = cp->cache_flags;
3921 ccp->cc_rounds = -1;
3922 ccp->cc_prounds = -1;
3923 }
3924
3925 /*
3926 * Create the cache's kstats.
3927 */
3928 if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
3929 "kmem_cache", KSTAT_TYPE_NAMED,
3930 sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
3931 KSTAT_FLAG_VIRTUAL)) != NULL) {
3932 cp->cache_kstat->ks_data = &kmem_cache_kstat;
3933 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
3934 cp->cache_kstat->ks_private = cp;
3935 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
3936 kstat_install(cp->cache_kstat);
3937 }
3938
3939 /*
3940 * Add the cache to the global list. This makes it visible
3941 * to kmem_update(), so the cache must be ready for business.
3942 */
3943 mutex_enter(&kmem_cache_lock);
3944 list_insert_tail(&kmem_caches, cp);
3945 mutex_exit(&kmem_cache_lock);
3946
3947 if (kmem_ready)
3948 kmem_cache_magazine_enable(cp);
3949
3950 return (cp);
3951 }
3952
3953 static int
3954 kmem_move_cmp(const void *buf, const void *p)
3955 {
3956 const kmem_move_t *kmm = p;
3957 uintptr_t v1 = (uintptr_t)buf;
3958 uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
3959 return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
3960 }
3961
3962 static void
3963 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
3964 {
3965 kmd->kmd_reclaim_numer = 1;
3966 }
3967
3968 /*
3969 * Initially, when choosing candidate slabs for buffers to move, we want to be
3970 * very selective and take only slabs that are less than
3971 * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
3972 * slabs, then we raise the allocation ceiling incrementally. The reclaim
3973 * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
3974 * longer fragmented.
3975 */
3976 static void
3977 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
3978 {
3979 if (direction > 0) {
3980 /* make it easier to find a candidate slab */
3981 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
3982 kmd->kmd_reclaim_numer++;
3983 }
3984 } else {
3985 /* be more selective */
3986 if (kmd->kmd_reclaim_numer > 1) {
3987 kmd->kmd_reclaim_numer--;
3988 }
3989 }
3990 }
3991
3992 void
3993 kmem_cache_set_move(kmem_cache_t *cp,
3994 kmem_cbrc_t (*move)(void *, void *, size_t, void *))
3995 {
3996 kmem_defrag_t *defrag;
3997
3998 ASSERT(move != NULL);
3999 /*
4000 * The consolidator does not support NOTOUCH caches because kmem cannot
4001 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4002 * a low order bit usable by clients to distinguish uninitialized memory
4003 * from known objects (see kmem_slab_create).
4004 */
4005 ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4006 ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4007
4008 /*
4009 * We should not be holding anyone's cache lock when calling
4010 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4011 * lock.
4012 */
4013 defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4014
4015 mutex_enter(&cp->cache_lock);
4016
4017 if (KMEM_IS_MOVABLE(cp)) {
4018 if (cp->cache_move == NULL) {
4019 ASSERT(cp->cache_slab_alloc == 0);
4020
4021 cp->cache_defrag = defrag;
4022 defrag = NULL; /* nothing to free */
4023 bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4024 avl_create(&cp->cache_defrag->kmd_moves_pending,
4025 kmem_move_cmp, sizeof (kmem_move_t),
4026 offsetof(kmem_move_t, kmm_entry));
4027 /* LINTED: E_TRUE_LOGICAL_EXPR */
4028 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4029 /* reuse the slab's AVL linkage for deadlist linkage */
4030 list_create(&cp->cache_defrag->kmd_deadlist,
4031 sizeof (kmem_slab_t),
4032 offsetof(kmem_slab_t, slab_link));
4033 kmem_reset_reclaim_threshold(cp->cache_defrag);
4034 }
4035 cp->cache_move = move;
4036 }
4037
4038 mutex_exit(&cp->cache_lock);
4039
4040 if (defrag != NULL) {
4041 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4042 }
4043 }
4044
4045 void
4046 kmem_cache_destroy(kmem_cache_t *cp)
4047 {
4048 int cpu_seqid;
4049
4050 /*
4051 * Remove the cache from the global cache list so that no one else
4052 * can schedule tasks on its behalf, wait for any pending tasks to
4053 * complete, purge the cache, and then destroy it.
4054 */
4055 mutex_enter(&kmem_cache_lock);
4056 list_remove(&kmem_caches, cp);
4057 mutex_exit(&kmem_cache_lock);
4058
4059 if (kmem_taskq != NULL)
4060 taskq_wait(kmem_taskq);
4061
4062 if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4063 taskq_wait(kmem_move_taskq);
4064
4065 kmem_cache_magazine_purge(cp);
4066
4067 mutex_enter(&cp->cache_lock);
4068 if (cp->cache_buftotal != 0)
4069 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4070 cp->cache_name, (void *)cp);
4071 if (cp->cache_defrag != NULL) {
4072 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4073 list_destroy(&cp->cache_defrag->kmd_deadlist);
4074 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4075 cp->cache_defrag = NULL;
4076 }
4077 /*
4078 * The cache is now dead. There should be no further activity. We
4079 * enforce this by setting land mines in the constructor, destructor,
4080 * reclaim, and move routines that induce a kernel text fault if
4081 * invoked.
4082 */
4083 cp->cache_constructor = (int (*)(void *, void *, int))1;
4084 cp->cache_destructor = (void (*)(void *, void *))2;
4085 cp->cache_reclaim = (void (*)(void *))3;
4086 cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4087 mutex_exit(&cp->cache_lock);
4088
4089 kstat_delete(cp->cache_kstat);
4090
4091 if (cp->cache_hash_table != NULL)
4092 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4093 (cp->cache_hash_mask + 1) * sizeof (void *));
4094
4095 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4096 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4097
4098 mutex_destroy(&cp->cache_depot_lock);
4099 mutex_destroy(&cp->cache_lock);
4100
4101 vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4102 }
4103
4104 /*ARGSUSED*/
4105 static int
4106 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4107 {
4108 ASSERT(MUTEX_HELD(&cpu_lock));
4109 if (what == CPU_UNCONFIG) {
4110 kmem_cache_applyall(kmem_cache_magazine_purge,
4111 kmem_taskq, TQ_SLEEP);
4112 kmem_cache_applyall(kmem_cache_magazine_enable,
4113 kmem_taskq, TQ_SLEEP);
4114 }
4115 return (0);
4116 }
4117
4118 static void
4119 kmem_alloc_caches_create(const int *array, size_t count,
4120 kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4121 {
4122 char name[KMEM_CACHE_NAMELEN + 1];
4123 size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4124 size_t size = table_unit;
4125 int i;
4126
4127 for (i = 0; i < count; i++) {
4128 size_t cache_size = array[i];
4129 size_t align = KMEM_ALIGN;
4130 kmem_cache_t *cp;
4131
4132 /* if the table has an entry for maxbuf, we're done */
4133 if (size > maxbuf)
4134 break;
4135
4136 /* cache size must be a multiple of the table unit */
4137 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4138
4139 /*
4140 * If they allocate a multiple of the coherency granularity,
4141 * they get a coherency-granularity-aligned address.
4142 */
4143 if (IS_P2ALIGNED(cache_size, 64))
4144 align = 64;
4145 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4146 align = PAGESIZE;
4147 (void) snprintf(name, sizeof (name),
4148 "kmem_alloc_%lu", cache_size);
4149 cp = kmem_cache_create(name, cache_size, align,
4150 NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4151
4152 while (size <= cache_size) {
4153 alloc_table[(size - 1) >> shift] = cp;
4154 size += table_unit;
4155 }
4156 }
4157
4158 ASSERT(size > maxbuf); /* i.e. maxbuf <= max(cache_size) */
4159 }
4160
4161 static void
4162 kmem_cache_init(int pass, int use_large_pages)
4163 {
4164 int i;
4165 size_t maxbuf;
4166 kmem_magtype_t *mtp;
4167
4168 for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4169 char name[KMEM_CACHE_NAMELEN + 1];
4170
4171 mtp = &kmem_magtype[i];
4172 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4173 mtp->mt_cache = kmem_cache_create(name,
4174 (mtp->mt_magsize + 1) * sizeof (void *),
4175 mtp->mt_align, NULL, NULL, NULL, NULL,
4176 kmem_msb_arena, KMC_NOHASH);
4177 }
4178
4179 kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4180 sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4181 kmem_msb_arena, KMC_NOHASH);
4182
4183 kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4184 sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4185 kmem_msb_arena, KMC_NOHASH);
4186
4187 kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4188 sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4189 kmem_msb_arena, KMC_NOHASH);
4190
4191 if (pass == 2) {
4192 kmem_va_arena = vmem_create("kmem_va",
4193 NULL, 0, PAGESIZE,
4194 vmem_alloc, vmem_free, heap_arena,
4195 8 * PAGESIZE, VM_SLEEP);
4196
4197 if (use_large_pages) {
4198 kmem_default_arena = vmem_xcreate("kmem_default",
4199 NULL, 0, PAGESIZE,
4200 segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4201 0, VMC_DUMPSAFE | VM_SLEEP);
4202 } else {
4203 kmem_default_arena = vmem_create("kmem_default",
4204 NULL, 0, PAGESIZE,
4205 segkmem_alloc, segkmem_free, kmem_va_arena,
4206 0, VMC_DUMPSAFE | VM_SLEEP);
4207 }
4208
4209 /* Figure out what our maximum cache size is */
4210 maxbuf = kmem_max_cached;
4211 if (maxbuf <= KMEM_MAXBUF) {
4212 maxbuf = 0;
4213 kmem_max_cached = KMEM_MAXBUF;
4214 } else {
4215 size_t size = 0;
4216 size_t max =
4217 sizeof (kmem_big_alloc_sizes) / sizeof (int);
4218 /*
4219 * Round maxbuf up to an existing cache size. If maxbuf
4220 * is larger than the largest cache, we truncate it to
4221 * the largest cache's size.
4222 */
4223 for (i = 0; i < max; i++) {
4224 size = kmem_big_alloc_sizes[i];
4225 if (maxbuf <= size)
4226 break;
4227 }
4228 kmem_max_cached = maxbuf = size;
4229 }
4230
4231 /*
4232 * The big alloc table may not be completely overwritten, so
4233 * we clear out any stale cache pointers from the first pass.
4234 */
4235 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4236 } else {
4237 /*
4238 * During the first pass, the kmem_alloc_* caches
4239 * are treated as metadata.
4240 */
4241 kmem_default_arena = kmem_msb_arena;
4242 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4243 }
4244
4245 /*
4246 * Set up the default caches to back kmem_alloc()
4247 */
4248 kmem_alloc_caches_create(
4249 kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4250 kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4251
4252 kmem_alloc_caches_create(
4253 kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4254 kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4255
4256 kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4257 }
4258
4259 void
4260 kmem_init(void)
4261 {
4262 kmem_cache_t *cp;
4263 int old_kmem_flags = kmem_flags;
4264 int use_large_pages = 0;
4265 size_t maxverify, minfirewall;
4266
4267 kstat_init();
4268
4269 /*
4270 * Don't do firewalled allocations if the heap is less than 1TB
4271 * (i.e. on a 32-bit kernel)
4272 * The resulting VM_NEXTFIT allocations would create too much
4273 * fragmentation in a small heap.
4274 */
4275 #if defined(_LP64)
4276 maxverify = minfirewall = PAGESIZE / 2;
4277 #else
4278 maxverify = minfirewall = ULONG_MAX;
4279 #endif
4280
4281 /* LINTED */
4282 ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4283
4284 list_create(&kmem_caches, sizeof (kmem_cache_t),
4285 offsetof(kmem_cache_t, cache_link));
4286
4287 kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4288 vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4289 VM_SLEEP | VMC_NO_QCACHE);
4290
4291 kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4292 PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4293 VMC_DUMPSAFE | VM_SLEEP);
4294
4295 kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4296 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4297
4298 kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4299 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4300
4301 kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4302 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4303
4304 kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4305 NULL, 0, PAGESIZE,
4306 kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4307 0, VM_SLEEP);
4308
4309 kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4310 segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4311 VMC_DUMPSAFE | VM_SLEEP);
4312
4313 /* temporary oversize arena for mod_read_system_file */
4314 kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4315 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4316
4317 kmem_reap_interval = 15 * hz;
4318
4319 /*
4320 * Read /etc/system. This is a chicken-and-egg problem because
4321 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4322 * needs to use the allocator. The simplest solution is to create
4323 * all the standard kmem caches, read /etc/system, destroy all the
4324 * caches we just created, and then create them all again in light
4325 * of the (possibly) new kmem_flags and other kmem tunables.
4326 */
4327 kmem_cache_init(1, 0);
4328
4329 mod_read_system_file(boothowto & RB_ASKNAME);
4330
4331 while ((cp = list_tail(&kmem_caches)) != NULL)
4332 kmem_cache_destroy(cp);
4333
4334 vmem_destroy(kmem_oversize_arena);
4335
4336 if (old_kmem_flags & KMF_STICKY)
4337 kmem_flags = old_kmem_flags;
4338
4339 if (!(kmem_flags & KMF_AUDIT))
4340 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4341
4342 if (kmem_maxverify == 0)
4343 kmem_maxverify = maxverify;
4344
4345 if (kmem_minfirewall == 0)
4346 kmem_minfirewall = minfirewall;
4347
4348 /*
4349 * give segkmem a chance to figure out if we are using large pages
4350 * for the kernel heap
4351 */
4352 use_large_pages = segkmem_lpsetup();
4353
4354 /*
4355 * To protect against corruption, we keep the actual number of callers
4356 * KMF_LITE records seperate from the tunable. We arbitrarily clamp
4357 * to 16, since the overhead for small buffers quickly gets out of
4358 * hand.
4359 *
4360 * The real limit would depend on the needs of the largest KMC_NOHASH
4361 * cache.
4362 */
4363 kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4364 kmem_lite_pcs = kmem_lite_count;
4365
4366 /*
4367 * Normally, we firewall oversized allocations when possible, but
4368 * if we are using large pages for kernel memory, and we don't have
4369 * any non-LITE debugging flags set, we want to allocate oversized
4370 * buffers from large pages, and so skip the firewalling.
4371 */
4372 if (use_large_pages &&
4373 ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4374 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4375 PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4376 0, VMC_DUMPSAFE | VM_SLEEP);
4377 } else {
4378 kmem_oversize_arena = vmem_create("kmem_oversize",
4379 NULL, 0, PAGESIZE,
4380 segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4381 kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4382 VM_SLEEP);
4383 }
4384
4385 kmem_cache_init(2, use_large_pages);
4386
4387 if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4388 if (kmem_transaction_log_size == 0)
4389 kmem_transaction_log_size = kmem_maxavail() / 50;
4390 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4391 }
4392
4393 if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4394 if (kmem_content_log_size == 0)
4395 kmem_content_log_size = kmem_maxavail() / 50;
4396 kmem_content_log = kmem_log_init(kmem_content_log_size);
4397 }
4398
4399 kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4400
4401 kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4402
4403 /*
4404 * Initialize STREAMS message caches so allocb() is available.
4405 * This allows us to initialize the logging framework (cmn_err(9F),
4406 * strlog(9F), etc) so we can start recording messages.
4407 */
4408 streams_msg_init();
4409
4410 /*
4411 * Initialize the ZSD framework in Zones so modules loaded henceforth
4412 * can register their callbacks.
4413 */
4414 zone_zsd_init();
4415
4416 log_init();
4417 taskq_init();
4418
4419 /*
4420 * Warn about invalid or dangerous values of kmem_flags.
4421 * Always warn about unsupported values.
4422 */
4423 if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4424 KMF_CONTENTS | KMF_LITE)) != 0) ||
4425 ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4426 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4427 "See the Solaris Tunable Parameters Reference Manual.",
4428 kmem_flags);
4429
4430 #ifdef DEBUG
4431 if ((kmem_flags & KMF_DEBUG) == 0)
4432 cmn_err(CE_NOTE, "kmem debugging disabled.");
4433 #else
4434 /*
4435 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4436 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4437 * if KMF_AUDIT is set). We should warn the user about the performance
4438 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4439 * isn't set (since that disables AUDIT).
4440 */
4441 if (!(kmem_flags & KMF_LITE) &&
4442 (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4443 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4444 "enabled (kmem_flags = 0x%x). Performance degradation "
4445 "and large memory overhead possible. See the Solaris "
4446 "Tunable Parameters Reference Manual.", kmem_flags);
4447 #endif /* not DEBUG */
4448
4449 kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4450
4451 kmem_ready = 1;
4452
4453 /*
4454 * Initialize the platform-specific aligned/DMA memory allocator.
4455 */
4456 ka_init();
4457
4458 /*
4459 * Initialize 32-bit ID cache.
4460 */
4461 id32_init();
4462
4463 /*
4464 * Initialize the networking stack so modules loaded can
4465 * register their callbacks.
4466 */
4467 netstack_init();
4468 }
4469
4470 static void
4471 kmem_move_init(void)
4472 {
4473 kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4474 sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4475 kmem_msb_arena, KMC_NOHASH);
4476 kmem_move_cache = kmem_cache_create("kmem_move_cache",
4477 sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4478 kmem_msb_arena, KMC_NOHASH);
4479
4480 /*
4481 * kmem guarantees that move callbacks are sequential and that even
4482 * across multiple caches no two moves ever execute simultaneously.
4483 * Move callbacks are processed on a separate taskq so that client code
4484 * does not interfere with internal maintenance tasks.
4485 */
4486 kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4487 minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4488 }
4489
4490 void
4491 kmem_thread_init(void)
4492 {
4493 kmem_move_init();
4494 kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4495 300, INT_MAX, TASKQ_PREPOPULATE);
4496 }
4497
4498 void
4499 kmem_mp_init(void)
4500 {
4501 mutex_enter(&cpu_lock);
4502 register_cpu_setup_func(kmem_cpu_setup, NULL);
4503 mutex_exit(&cpu_lock);
4504
4505 kmem_update_timeout(NULL);
4506
4507 taskq_mp_init();
4508 }
4509
4510 /*
4511 * Return the slab of the allocated buffer, or NULL if the buffer is not
4512 * allocated. This function may be called with a known slab address to determine
4513 * whether or not the buffer is allocated, or with a NULL slab address to obtain
4514 * an allocated buffer's slab.
4515 */
4516 static kmem_slab_t *
4517 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4518 {
4519 kmem_bufctl_t *bcp, *bufbcp;
4520
4521 ASSERT(MUTEX_HELD(&cp->cache_lock));
4522 ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4523
4524 if (cp->cache_flags & KMF_HASH) {
4525 for (bcp = *KMEM_HASH(cp, buf);
4526 (bcp != NULL) && (bcp->bc_addr != buf);
4527 bcp = bcp->bc_next) {
4528 continue;
4529 }
4530 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4531 return (bcp == NULL ? NULL : bcp->bc_slab);
4532 }
4533
4534 if (sp == NULL) {
4535 sp = KMEM_SLAB(cp, buf);
4536 }
4537 bufbcp = KMEM_BUFCTL(cp, buf);
4538 for (bcp = sp->slab_head;
4539 (bcp != NULL) && (bcp != bufbcp);
4540 bcp = bcp->bc_next) {
4541 continue;
4542 }
4543 return (bcp == NULL ? sp : NULL);
4544 }
4545
4546 static boolean_t
4547 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4548 {
4549 long refcnt = sp->slab_refcnt;
4550
4551 ASSERT(cp->cache_defrag != NULL);
4552
4553 /*
4554 * For code coverage we want to be able to move an object within the
4555 * same slab (the only partial slab) even if allocating the destination
4556 * buffer resulted in a completely allocated slab.
4557 */
4558 if (flags & KMM_DEBUG) {
4559 return ((flags & KMM_DESPERATE) ||
4560 ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4561 }
4562
4563 /* If we're desperate, we don't care if the client said NO. */
4564 if (flags & KMM_DESPERATE) {
4565 return (refcnt < sp->slab_chunks); /* any partial */
4566 }
4567
4568 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4569 return (B_FALSE);
4570 }
4571
4572 if ((refcnt == 1) || kmem_move_any_partial) {
4573 return (refcnt < sp->slab_chunks);
4574 }
4575
4576 /*
4577 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4578 * slabs with a progressively higher percentage of used buffers can be
4579 * reclaimed until the cache as a whole is no longer fragmented.
4580 *
4581 * sp->slab_refcnt kmd_reclaim_numer
4582 * --------------- < ------------------
4583 * sp->slab_chunks KMEM_VOID_FRACTION
4584 */
4585 return ((refcnt * KMEM_VOID_FRACTION) <
4586 (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4587 }
4588
4589 /*
4590 * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4591 * or when the buffer is freed.
4592 */
4593 static void
4594 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4595 {
4596 ASSERT(MUTEX_HELD(&cp->cache_lock));
4597 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4598
4599 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4600 return;
4601 }
4602
4603 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4604 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4605 avl_remove(&cp->cache_partial_slabs, sp);
4606 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4607 sp->slab_stuck_offset = (uint32_t)-1;
4608 avl_add(&cp->cache_partial_slabs, sp);
4609 }
4610 } else {
4611 sp->slab_later_count = 0;
4612 sp->slab_stuck_offset = (uint32_t)-1;
4613 }
4614 }
4615
4616 static void
4617 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4618 {
4619 ASSERT(taskq_member(kmem_move_taskq, curthread));
4620 ASSERT(MUTEX_HELD(&cp->cache_lock));
4621 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4622
4623 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4624 return;
4625 }
4626
4627 avl_remove(&cp->cache_partial_slabs, sp);
4628 sp->slab_later_count = 0;
4629 sp->slab_flags |= KMEM_SLAB_NOMOVE;
4630 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4631 avl_add(&cp->cache_partial_slabs, sp);
4632 }
4633
4634 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4635
4636 /*
4637 * The move callback takes two buffer addresses, the buffer to be moved, and a
4638 * newly allocated and constructed buffer selected by kmem as the destination.
4639 * It also takes the size of the buffer and an optional user argument specified
4640 * at cache creation time. kmem guarantees that the buffer to be moved has not
4641 * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4642 * guarantee the present whereabouts of the buffer to be moved, so it is up to
4643 * the client to safely determine whether or not it is still using the buffer.
4644 * The client must not free either of the buffers passed to the move callback,
4645 * since kmem wants to free them directly to the slab layer. The client response
4646 * tells kmem which of the two buffers to free:
4647 *
4648 * YES kmem frees the old buffer (the move was successful)
4649 * NO kmem frees the new buffer, marks the slab of the old buffer
4650 * non-reclaimable to avoid bothering the client again
4651 * LATER kmem frees the new buffer, increments slab_later_count
4652 * DONT_KNOW kmem frees the new buffer
4653 * DONT_NEED kmem frees both the old buffer and the new buffer
4654 *
4655 * The pending callback argument now being processed contains both of the
4656 * buffers (old and new) passed to the move callback function, the slab of the
4657 * old buffer, and flags related to the move request, such as whether or not the
4658 * system was desperate for memory.
4659 *
4660 * Slabs are not freed while there is a pending callback, but instead are kept
4661 * on a deadlist, which is drained after the last callback completes. This means
4662 * that slabs are safe to access until kmem_move_end(), no matter how many of
4663 * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4664 * zero for as long as the slab remains on the deadlist and until the slab is
4665 * freed.
4666 */
4667 static void
4668 kmem_move_buffer(kmem_move_t *callback)
4669 {
4670 kmem_cbrc_t response;
4671 kmem_slab_t *sp = callback->kmm_from_slab;
4672 kmem_cache_t *cp = sp->slab_cache;
4673 boolean_t free_on_slab;
4674
4675 ASSERT(taskq_member(kmem_move_taskq, curthread));
4676 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4677 ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4678
4679 /*
4680 * The number of allocated buffers on the slab may have changed since we
4681 * last checked the slab's reclaimability (when the pending move was
4682 * enqueued), or the client may have responded NO when asked to move
4683 * another buffer on the same slab.
4684 */
4685 if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4686 kmem_slab_free(cp, callback->kmm_to_buf);
4687 kmem_move_end(cp, callback);
4688 return;
4689 }
4690
4691 /*
4692 * Checking the slab layer is easy, so we might as well do that here
4693 * in case we can avoid bothering the client.
4694 */
4695 mutex_enter(&cp->cache_lock);
4696 free_on_slab = (kmem_slab_allocated(cp, sp,
4697 callback->kmm_from_buf) == NULL);
4698 mutex_exit(&cp->cache_lock);
4699
4700 if (free_on_slab) {
4701 kmem_slab_free(cp, callback->kmm_to_buf);
4702 kmem_move_end(cp, callback);
4703 return;
4704 }
4705
4706 if (cp->cache_flags & KMF_BUFTAG) {
4707 /*
4708 * Make kmem_cache_alloc_debug() apply the constructor for us.
4709 */
4710 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4711 KM_NOSLEEP, 1, caller()) != 0) {
4712 kmem_move_end(cp, callback);
4713 return;
4714 }
4715 } else if (cp->cache_constructor != NULL &&
4716 cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4717 KM_NOSLEEP) != 0) {
4718 atomic_inc_64(&cp->cache_alloc_fail);
4719 kmem_slab_free(cp, callback->kmm_to_buf);
4720 kmem_move_end(cp, callback);
4721 return;
4722 }
4723
4724 cp->cache_defrag->kmd_callbacks++;
4725 cp->cache_defrag->kmd_thread = curthread;
4726 cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4727 cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4728 DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4729 callback);
4730
4731 response = cp->cache_move(callback->kmm_from_buf,
4732 callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4733
4734 DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4735 callback, kmem_cbrc_t, response);
4736 cp->cache_defrag->kmd_thread = NULL;
4737 cp->cache_defrag->kmd_from_buf = NULL;
4738 cp->cache_defrag->kmd_to_buf = NULL;
4739
4740 if (response == KMEM_CBRC_YES) {
4741 cp->cache_defrag->kmd_yes++;
4742 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4743 /* slab safe to access until kmem_move_end() */
4744 if (sp->slab_refcnt == 0)
4745 cp->cache_defrag->kmd_slabs_freed++;
4746 mutex_enter(&cp->cache_lock);
4747 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4748 mutex_exit(&cp->cache_lock);
4749 kmem_move_end(cp, callback);
4750 return;
4751 }
4752
4753 switch (response) {
4754 case KMEM_CBRC_NO:
4755 cp->cache_defrag->kmd_no++;
4756 mutex_enter(&cp->cache_lock);
4757 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4758 mutex_exit(&cp->cache_lock);
4759 break;
4760 case KMEM_CBRC_LATER:
4761 cp->cache_defrag->kmd_later++;
4762 mutex_enter(&cp->cache_lock);
4763 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4764 mutex_exit(&cp->cache_lock);
4765 break;
4766 }
4767
4768 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4769 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4770 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4771 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4772 callback->kmm_from_buf);
4773 }
4774 mutex_exit(&cp->cache_lock);
4775 break;
4776 case KMEM_CBRC_DONT_NEED:
4777 cp->cache_defrag->kmd_dont_need++;
4778 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4779 if (sp->slab_refcnt == 0)
4780 cp->cache_defrag->kmd_slabs_freed++;
4781 mutex_enter(&cp->cache_lock);
4782 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4783 mutex_exit(&cp->cache_lock);
4784 break;
4785 case KMEM_CBRC_DONT_KNOW:
4786 /*
4787 * If we don't know if we can move this buffer or not, we'll
4788 * just assume that we can't: if the buffer is in fact free,
4789 * then it is sitting in one of the per-CPU magazines or in
4790 * a full magazine in the depot layer. Either way, because
4791 * defrag is induced in the same logic that reaps a cache,
4792 * it's likely that full magazines will be returned to the
4793 * system soon (thereby accomplishing what we're trying to
4794 * accomplish here: return those magazines to their slabs).
4795 * Given this, any work that we might do now to locate a buffer
4796 * in a magazine is wasted (and expensive!) work; we bump
4797 * a counter in this case and otherwise assume that we can't
4798 * move it.
4799 */
4800 cp->cache_defrag->kmd_dont_know++;
4801 break;
4802 default:
4803 panic("'%s' (%p) unexpected move callback response %d\n",
4804 cp->cache_name, (void *)cp, response);
4805 }
4806
4807 kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4808 kmem_move_end(cp, callback);
4809 }
4810
4811 /* Return B_FALSE if there is insufficient memory for the move request. */
4812 static boolean_t
4813 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4814 {
4815 void *to_buf;
4816 avl_index_t index;
4817 kmem_move_t *callback, *pending;
4818 ulong_t n;
4819
4820 ASSERT(taskq_member(kmem_taskq, curthread));
4821 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4822 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4823
4824 callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4825
4826 if (callback == NULL)
4827 return (B_FALSE);
4828
4829 callback->kmm_from_slab = sp;
4830 callback->kmm_from_buf = buf;
4831 callback->kmm_flags = flags;
4832
4833 mutex_enter(&cp->cache_lock);
4834
4835 n = avl_numnodes(&cp->cache_partial_slabs);
4836 if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
4837 mutex_exit(&cp->cache_lock);
4838 kmem_cache_free(kmem_move_cache, callback);
4839 return (B_TRUE); /* there is no need for the move request */
4840 }
4841
4842 pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
4843 if (pending != NULL) {
4844 /*
4845 * If the move is already pending and we're desperate now,
4846 * update the move flags.
4847 */
4848 if (flags & KMM_DESPERATE) {
4849 pending->kmm_flags |= KMM_DESPERATE;
4850 }
4851 mutex_exit(&cp->cache_lock);
4852 kmem_cache_free(kmem_move_cache, callback);
4853 return (B_TRUE);
4854 }
4855
4856 to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
4857 B_FALSE);
4858 callback->kmm_to_buf = to_buf;
4859 avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
4860
4861 mutex_exit(&cp->cache_lock);
4862
4863 if (taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
4864 callback, TQ_NOSLEEP) == TASKQID_INVALID) {
4865 mutex_enter(&cp->cache_lock);
4866 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4867 mutex_exit(&cp->cache_lock);
4868 kmem_slab_free(cp, to_buf);
4869 kmem_cache_free(kmem_move_cache, callback);
4870 return (B_FALSE);
4871 }
4872
4873 return (B_TRUE);
4874 }
4875
4876 static void
4877 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
4878 {
4879 avl_index_t index;
4880
4881 ASSERT(cp->cache_defrag != NULL);
4882 ASSERT(taskq_member(kmem_move_taskq, curthread));
4883 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4884
4885 mutex_enter(&cp->cache_lock);
4886 VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
4887 callback->kmm_from_buf, &index) != NULL);
4888 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4889 if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
4890 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
4891 kmem_slab_t *sp;
4892
4893 /*
4894 * The last pending move completed. Release all slabs from the
4895 * front of the dead list except for any slab at the tail that
4896 * needs to be released from the context of kmem_move_buffers().
4897 * kmem deferred unmapping the buffers on these slabs in order
4898 * to guarantee that buffers passed to the move callback have
4899 * been touched only by kmem or by the client itself.
4900 */
4901 while ((sp = list_remove_head(deadlist)) != NULL) {
4902 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
4903 list_insert_tail(deadlist, sp);
4904 break;
4905 }
4906 cp->cache_defrag->kmd_deadcount--;
4907 cp->cache_slab_destroy++;
4908 mutex_exit(&cp->cache_lock);
4909 kmem_slab_destroy(cp, sp);
4910 mutex_enter(&cp->cache_lock);
4911 }
4912 }
4913 mutex_exit(&cp->cache_lock);
4914 kmem_cache_free(kmem_move_cache, callback);
4915 }
4916
4917 /*
4918 * Move buffers from least used slabs first by scanning backwards from the end
4919 * of the partial slab list. Scan at most max_scan candidate slabs and move
4920 * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
4921 * If desperate to reclaim memory, move buffers from any partial slab, otherwise
4922 * skip slabs with a ratio of allocated buffers at or above the current
4923 * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
4924 * scan is aborted) so that the caller can adjust the reclaimability threshold
4925 * depending on how many reclaimable slabs it finds.
4926 *
4927 * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
4928 * move request, since it is not valid for kmem_move_begin() to call
4929 * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
4930 */
4931 static int
4932 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
4933 int flags)
4934 {
4935 kmem_slab_t *sp;
4936 void *buf;
4937 int i, j; /* slab index, buffer index */
4938 int s; /* reclaimable slabs */
4939 int b; /* allocated (movable) buffers on reclaimable slab */
4940 boolean_t success;
4941 int refcnt;
4942 int nomove;
4943
4944 ASSERT(taskq_member(kmem_taskq, curthread));
4945 ASSERT(MUTEX_HELD(&cp->cache_lock));
4946 ASSERT(kmem_move_cache != NULL);
4947 ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
4948 ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
4949 avl_numnodes(&cp->cache_partial_slabs) > 1);
4950
4951 if (kmem_move_blocked) {
4952 return (0);
4953 }
4954
4955 if (kmem_move_fulltilt) {
4956 flags |= KMM_DESPERATE;
4957 }
4958
4959 if (max_scan == 0 || (flags & KMM_DESPERATE)) {
4960 /*
4961 * Scan as many slabs as needed to find the desired number of
4962 * candidate slabs.
4963 */
4964 max_scan = (size_t)-1;
4965 }
4966
4967 if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
4968 /* Find as many candidate slabs as possible. */
4969 max_slabs = (size_t)-1;
4970 }
4971
4972 sp = avl_last(&cp->cache_partial_slabs);
4973 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
4974 for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
4975 ((sp != avl_first(&cp->cache_partial_slabs)) ||
4976 (flags & KMM_DEBUG));
4977 sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
4978
4979 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
4980 continue;
4981 }
4982 s++;
4983
4984 /* Look for allocated buffers to move. */
4985 for (j = 0, b = 0, buf = sp->slab_base;
4986 (j < sp->slab_chunks) && (b < sp->slab_refcnt);
4987 buf = (((char *)buf) + cp->cache_chunksize), j++) {
4988
4989 if (kmem_slab_allocated(cp, sp, buf) == NULL) {
4990 continue;
4991 }
4992
4993 b++;
4994
4995 /*
4996 * Prevent the slab from being destroyed while we drop
4997 * cache_lock and while the pending move is not yet
4998 * registered. Flag the pending move while
4999 * kmd_moves_pending may still be empty, since we can't
5000 * yet rely on a non-zero pending move count to prevent
5001 * the slab from being destroyed.
5002 */
5003 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5004 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5005 /*
5006 * Recheck refcnt and nomove after reacquiring the lock,
5007 * since these control the order of partial slabs, and
5008 * we want to know if we can pick up the scan where we
5009 * left off.
5010 */
5011 refcnt = sp->slab_refcnt;
5012 nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5013 mutex_exit(&cp->cache_lock);
5014
5015 success = kmem_move_begin(cp, sp, buf, flags);
5016
5017 /*
5018 * Now, before the lock is reacquired, kmem could
5019 * process all pending move requests and purge the
5020 * deadlist, so that upon reacquiring the lock, sp has
5021 * been remapped. Or, the client may free all the
5022 * objects on the slab while the pending moves are still
5023 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5024 * flag causes the slab to be put at the end of the
5025 * deadlist and prevents it from being destroyed, since
5026 * we plan to destroy it here after reacquiring the
5027 * lock.
5028 */
5029 mutex_enter(&cp->cache_lock);
5030 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5031 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5032
5033 if (sp->slab_refcnt == 0) {
5034 list_t *deadlist =
5035 &cp->cache_defrag->kmd_deadlist;
5036 list_remove(deadlist, sp);
5037
5038 if (!avl_is_empty(
5039 &cp->cache_defrag->kmd_moves_pending)) {
5040 /*
5041 * A pending move makes it unsafe to
5042 * destroy the slab, because even though
5043 * the move is no longer needed, the
5044 * context where that is determined
5045 * requires the slab to exist.
5046 * Fortunately, a pending move also
5047 * means we don't need to destroy the
5048 * slab here, since it will get
5049 * destroyed along with any other slabs
5050 * on the deadlist after the last
5051 * pending move completes.
5052 */
5053 list_insert_head(deadlist, sp);
5054 return (-1);
5055 }
5056
5057 /*
5058 * Destroy the slab now if it was completely
5059 * freed while we dropped cache_lock and there
5060 * are no pending moves. Since slab_refcnt
5061 * cannot change once it reaches zero, no new
5062 * pending moves from that slab are possible.
5063 */
5064 cp->cache_defrag->kmd_deadcount--;
5065 cp->cache_slab_destroy++;
5066 mutex_exit(&cp->cache_lock);
5067 kmem_slab_destroy(cp, sp);
5068 mutex_enter(&cp->cache_lock);
5069 /*
5070 * Since we can't pick up the scan where we left
5071 * off, abort the scan and say nothing about the
5072 * number of reclaimable slabs.
5073 */
5074 return (-1);
5075 }
5076
5077 if (!success) {
5078 /*
5079 * Abort the scan if there is not enough memory
5080 * for the request and say nothing about the
5081 * number of reclaimable slabs.
5082 */
5083 return (-1);
5084 }
5085
5086 /*
5087 * The slab's position changed while the lock was
5088 * dropped, so we don't know where we are in the
5089 * sequence any more.
5090 */
5091 if (sp->slab_refcnt != refcnt) {
5092 /*
5093 * If this is a KMM_DEBUG move, the slab_refcnt
5094 * may have changed because we allocated a
5095 * destination buffer on the same slab. In that
5096 * case, we're not interested in counting it.
5097 */
5098 return (-1);
5099 }
5100 if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove)
5101 return (-1);
5102
5103 /*
5104 * Generating a move request allocates a destination
5105 * buffer from the slab layer, bumping the first partial
5106 * slab if it is completely allocated. If the current
5107 * slab becomes the first partial slab as a result, we
5108 * can't continue to scan backwards.
5109 *
5110 * If this is a KMM_DEBUG move and we allocated the
5111 * destination buffer from the last partial slab, then
5112 * the buffer we're moving is on the same slab and our
5113 * slab_refcnt has changed, causing us to return before
5114 * reaching here if there are no partial slabs left.
5115 */
5116 ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5117 if (sp == avl_first(&cp->cache_partial_slabs)) {
5118 /*
5119 * We're not interested in a second KMM_DEBUG
5120 * move.
5121 */
5122 goto end_scan;
5123 }
5124 }
5125 }
5126 end_scan:
5127
5128 return (s);
5129 }
5130
5131 typedef struct kmem_move_notify_args {
5132 kmem_cache_t *kmna_cache;
5133 void *kmna_buf;
5134 } kmem_move_notify_args_t;
5135
5136 static void
5137 kmem_cache_move_notify_task(void *arg)
5138 {
5139 kmem_move_notify_args_t *args = arg;
5140 kmem_cache_t *cp = args->kmna_cache;
5141 void *buf = args->kmna_buf;
5142 kmem_slab_t *sp;
5143
5144 ASSERT(taskq_member(kmem_taskq, curthread));
5145 ASSERT(list_link_active(&cp->cache_link));
5146
5147 kmem_free(args, sizeof (kmem_move_notify_args_t));
5148 mutex_enter(&cp->cache_lock);
5149 sp = kmem_slab_allocated(cp, NULL, buf);
5150
5151 /* Ignore the notification if the buffer is no longer allocated. */
5152 if (sp == NULL) {
5153 mutex_exit(&cp->cache_lock);
5154 return;
5155 }
5156
5157 /* Ignore the notification if there's no reason to move the buffer. */
5158 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5159 /*
5160 * So far the notification is not ignored. Ignore the
5161 * notification if the slab is not marked by an earlier refusal
5162 * to move a buffer.
5163 */
5164 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5165 (sp->slab_later_count == 0)) {
5166 mutex_exit(&cp->cache_lock);
5167 return;
5168 }
5169
5170 kmem_slab_move_yes(cp, sp, buf);
5171 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5172 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5173 mutex_exit(&cp->cache_lock);
5174 /* see kmem_move_buffers() about dropping the lock */
5175 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5176 mutex_enter(&cp->cache_lock);
5177 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5178 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5179 if (sp->slab_refcnt == 0) {
5180 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5181 list_remove(deadlist, sp);
5182
5183 if (!avl_is_empty(
5184 &cp->cache_defrag->kmd_moves_pending)) {
5185 list_insert_head(deadlist, sp);
5186 mutex_exit(&cp->cache_lock);
5187 return;
5188 }
5189
5190 cp->cache_defrag->kmd_deadcount--;
5191 cp->cache_slab_destroy++;
5192 mutex_exit(&cp->cache_lock);
5193 kmem_slab_destroy(cp, sp);
5194 return;
5195 }
5196 } else {
5197 kmem_slab_move_yes(cp, sp, buf);
5198 }
5199 mutex_exit(&cp->cache_lock);
5200 }
5201
5202 void
5203 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5204 {
5205 kmem_move_notify_args_t *args;
5206
5207 args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5208 if (args != NULL) {
5209 args->kmna_cache = cp;
5210 args->kmna_buf = buf;
5211 if (taskq_dispatch(kmem_taskq,
5212 (task_func_t *)kmem_cache_move_notify_task, args,
5213 TQ_NOSLEEP) == TASKQID_INVALID)
5214 kmem_free(args, sizeof (kmem_move_notify_args_t));
5215 }
5216 }
5217
5218 static void
5219 kmem_cache_defrag(kmem_cache_t *cp)
5220 {
5221 size_t n;
5222
5223 ASSERT(cp->cache_defrag != NULL);
5224
5225 mutex_enter(&cp->cache_lock);
5226 n = avl_numnodes(&cp->cache_partial_slabs);
5227 if (n > 1) {
5228 /* kmem_move_buffers() drops and reacquires cache_lock */
5229 cp->cache_defrag->kmd_defrags++;
5230 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5231 }
5232 mutex_exit(&cp->cache_lock);
5233 }
5234
5235 /* Is this cache above the fragmentation threshold? */
5236 static boolean_t
5237 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5238 {
5239 /*
5240 * nfree kmem_frag_numer
5241 * ------------------ > ---------------
5242 * cp->cache_buftotal kmem_frag_denom
5243 */
5244 return ((nfree * kmem_frag_denom) >
5245 (cp->cache_buftotal * kmem_frag_numer));
5246 }
5247
5248 static boolean_t
5249 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5250 {
5251 boolean_t fragmented;
5252 uint64_t nfree;
5253
5254 ASSERT(MUTEX_HELD(&cp->cache_lock));
5255 *doreap = B_FALSE;
5256
5257 if (kmem_move_fulltilt) {
5258 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5259 return (B_TRUE);
5260 }
5261 } else {
5262 if ((cp->cache_complete_slab_count + avl_numnodes(
5263 &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5264 return (B_FALSE);
5265 }
5266 }
5267
5268 nfree = cp->cache_bufslab;
5269 fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5270 kmem_cache_frag_threshold(cp, nfree));
5271
5272 /*
5273 * Free buffers in the magazine layer appear allocated from the point of
5274 * view of the slab layer. We want to know if the slab layer would
5275 * appear fragmented if we included free buffers from magazines that
5276 * have fallen out of the working set.
5277 */
5278 if (!fragmented) {
5279 long reap;
5280
5281 mutex_enter(&cp->cache_depot_lock);
5282 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5283 reap = MIN(reap, cp->cache_full.ml_total);
5284 mutex_exit(&cp->cache_depot_lock);
5285
5286 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5287 if (kmem_cache_frag_threshold(cp, nfree)) {
5288 *doreap = B_TRUE;
5289 }
5290 }
5291
5292 return (fragmented);
5293 }
5294
5295 /* Called periodically from kmem_taskq */
5296 static void
5297 kmem_cache_scan(kmem_cache_t *cp)
5298 {
5299 boolean_t reap = B_FALSE;
5300 kmem_defrag_t *kmd;
5301
5302 ASSERT(taskq_member(kmem_taskq, curthread));
5303
5304 mutex_enter(&cp->cache_lock);
5305
5306 kmd = cp->cache_defrag;
5307 if (kmd->kmd_consolidate > 0) {
5308 kmd->kmd_consolidate--;
5309 mutex_exit(&cp->cache_lock);
5310 kmem_cache_reap(cp);
5311 return;
5312 }
5313
5314 if (kmem_cache_is_fragmented(cp, &reap)) {
5315 size_t slabs_found;
5316
5317 /*
5318 * Consolidate reclaimable slabs from the end of the partial
5319 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5320 * reclaimable slabs). Keep track of how many candidate slabs we
5321 * looked for and how many we actually found so we can adjust
5322 * the definition of a candidate slab if we're having trouble
5323 * finding them.
5324 *
5325 * kmem_move_buffers() drops and reacquires cache_lock.
5326 */
5327 kmd->kmd_scans++;
5328 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5329 kmem_reclaim_max_slabs, 0);
5330 if (slabs_found >= 0) {
5331 kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5332 kmd->kmd_slabs_found += slabs_found;
5333 }
5334
5335 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5336 kmd->kmd_tries = 0;
5337
5338 /*
5339 * If we had difficulty finding candidate slabs in
5340 * previous scans, adjust the threshold so that
5341 * candidates are easier to find.
5342 */
5343 if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5344 kmem_adjust_reclaim_threshold(kmd, -1);
5345 } else if ((kmd->kmd_slabs_found * 2) <
5346 kmd->kmd_slabs_sought) {
5347 kmem_adjust_reclaim_threshold(kmd, 1);
5348 }
5349 kmd->kmd_slabs_sought = 0;
5350 kmd->kmd_slabs_found = 0;
5351 }
5352 } else {
5353 kmem_reset_reclaim_threshold(cp->cache_defrag);
5354 #ifdef DEBUG
5355 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5356 /*
5357 * In a debug kernel we want the consolidator to
5358 * run occasionally even when there is plenty of
5359 * memory.
5360 */
5361 uint16_t debug_rand;
5362
5363 (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5364 if (!kmem_move_noreap &&
5365 ((debug_rand % kmem_mtb_reap) == 0)) {
5366 mutex_exit(&cp->cache_lock);
5367 kmem_cache_reap(cp);
5368 return;
5369 } else if ((debug_rand % kmem_mtb_move) == 0) {
5370 kmd->kmd_scans++;
5371 (void) kmem_move_buffers(cp,
5372 kmem_reclaim_scan_range, 1, KMM_DEBUG);
5373 }
5374 }
5375 #endif /* DEBUG */
5376 }
5377
5378 mutex_exit(&cp->cache_lock);
5379
5380 if (reap)
5381 kmem_depot_ws_reap(cp);
5382 }