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) 2015 Joyent, Inc. All rights reserved.
24 * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
25 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
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_zerosized_log_size; /* zero-sized log [4 pages per CPU] */
1015 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1016 size_t kmem_lite_minsize = 0; /* minimum buffer size for KMF_LITE */
1017 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1018 int kmem_lite_pcs = 4; /* number of PCs to store in KMF_LITE mode */
1019 size_t kmem_maxverify; /* maximum bytes to inspect in debug routines */
1020 size_t kmem_minfirewall; /* hardware-enforced redzone threshold */
1021
1022 #ifdef DEBUG
1023 int kmem_warn_zerosized = 1; /* whether to warn on zero-sized KM_SLEEP */
1024 #else
1025 int kmem_warn_zerosized = 0; /* whether to warn on zero-sized KM_SLEEP */
1026 #endif
1027
1028 int kmem_panic_zerosized = 0; /* whether to panic on zero-sized KM_SLEEP */
1029
1030 #ifdef _LP64
1031 size_t kmem_max_cached = KMEM_BIG_MAXBUF; /* maximum kmem_alloc cache */
1032 #else
1033 size_t kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1034 #endif
1035
1036 #ifdef DEBUG
1037 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1038 #else
1039 int kmem_flags = 0;
1040 #endif
1041 int kmem_ready;
1042
1043 static kmem_cache_t *kmem_slab_cache;
1044 static kmem_cache_t *kmem_bufctl_cache;
1045 static kmem_cache_t *kmem_bufctl_audit_cache;
1046
1047 static kmutex_t kmem_cache_lock; /* inter-cache linkage only */
1048 static list_t kmem_caches;
1049
1050 static taskq_t *kmem_taskq;
1051 static kmutex_t kmem_flags_lock;
1052 static vmem_t *kmem_metadata_arena;
1053 static vmem_t *kmem_msb_arena; /* arena for metadata caches */
1054 static vmem_t *kmem_cache_arena;
1055 static vmem_t *kmem_hash_arena;
1056 static vmem_t *kmem_log_arena;
1057 static vmem_t *kmem_oversize_arena;
1058 static vmem_t *kmem_va_arena;
1059 static vmem_t *kmem_default_arena;
1060 static vmem_t *kmem_firewall_va_arena;
1061 static vmem_t *kmem_firewall_arena;
1062
1063 static int kmem_zerosized; /* # of zero-sized allocs */
1064
1065 /*
1066 * kmem slab consolidator thresholds (tunables)
1067 */
1068 size_t kmem_frag_minslabs = 101; /* minimum total slabs */
1069 size_t kmem_frag_numer = 1; /* free buffers (numerator) */
1070 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1071 /*
1072 * Maximum number of slabs from which to move buffers during a single
1073 * maintenance interval while the system is not low on memory.
1074 */
1075 size_t kmem_reclaim_max_slabs = 1;
1076 /*
1077 * Number of slabs to scan backwards from the end of the partial slab list
1078 * when searching for buffers to relocate.
1079 */
1080 size_t kmem_reclaim_scan_range = 12;
1081
1082 /* consolidator knobs */
1083 static boolean_t kmem_move_noreap;
1084 static boolean_t kmem_move_blocked;
1085 static boolean_t kmem_move_fulltilt;
1086 static boolean_t kmem_move_any_partial;
1087
1088 #ifdef DEBUG
1089 /*
1090 * kmem consolidator debug tunables:
1091 * Ensure code coverage by occasionally running the consolidator even when the
1092 * caches are not fragmented (they may never be). These intervals are mean time
1093 * in cache maintenance intervals (kmem_cache_update).
1094 */
1095 uint32_t kmem_mtb_move = 60; /* defrag 1 slab (~15min) */
1096 uint32_t kmem_mtb_reap = 1800; /* defrag all slabs (~7.5hrs) */
1097 #endif /* DEBUG */
1098
1099 static kmem_cache_t *kmem_defrag_cache;
1100 static kmem_cache_t *kmem_move_cache;
1101 static taskq_t *kmem_move_taskq;
1102
1103 static void kmem_cache_scan(kmem_cache_t *);
1104 static void kmem_cache_defrag(kmem_cache_t *);
1105 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1106
1107
1108 kmem_log_header_t *kmem_transaction_log;
1109 kmem_log_header_t *kmem_content_log;
1110 kmem_log_header_t *kmem_failure_log;
1111 kmem_log_header_t *kmem_slab_log;
1112 kmem_log_header_t *kmem_zerosized_log;
1113
1114 static int kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1115
1116 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller) \
1117 if ((count) > 0) { \
1118 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1119 pc_t *_e; \
1120 /* memmove() the old entries down one notch */ \
1121 for (_e = &_s[(count) - 1]; _e > _s; _e--) \
1122 *_e = *(_e - 1); \
1123 *_s = (uintptr_t)(caller); \
1124 }
1125
1126 #define KMERR_MODIFIED 0 /* buffer modified while on freelist */
1127 #define KMERR_REDZONE 1 /* redzone violation (write past end of buf) */
1128 #define KMERR_DUPFREE 2 /* freed a buffer twice */
1129 #define KMERR_BADADDR 3 /* freed a bad (unallocated) address */
1130 #define KMERR_BADBUFTAG 4 /* buftag corrupted */
1131 #define KMERR_BADBUFCTL 5 /* bufctl corrupted */
1132 #define KMERR_BADCACHE 6 /* freed a buffer to the wrong cache */
1133 #define KMERR_BADSIZE 7 /* alloc size != free size */
1134 #define KMERR_BADBASE 8 /* buffer base address wrong */
1135
1136 struct {
1137 hrtime_t kmp_timestamp; /* timestamp of panic */
1138 int kmp_error; /* type of kmem error */
1139 void *kmp_buffer; /* buffer that induced panic */
1140 void *kmp_realbuf; /* real start address for buffer */
1141 kmem_cache_t *kmp_cache; /* buffer's cache according to client */
1142 kmem_cache_t *kmp_realcache; /* actual cache containing buffer */
1143 kmem_slab_t *kmp_slab; /* slab accoring to kmem_findslab() */
1144 kmem_bufctl_t *kmp_bufctl; /* bufctl */
1145 } kmem_panic_info;
1146
1147
1148 static void
1149 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1150 {
1151 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1152 uint64_t *buf = buf_arg;
1153
1154 while (buf < bufend)
1155 *buf++ = pattern;
1156 }
1157
1158 static void *
1159 verify_pattern(uint64_t pattern, 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 != pattern)
1166 return (buf);
1167 return (NULL);
1168 }
1169
1170 static void *
1171 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1172 {
1173 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1174 uint64_t *buf;
1175
1176 for (buf = buf_arg; buf < bufend; buf++) {
1177 if (*buf != old) {
1178 copy_pattern(old, buf_arg,
1179 (char *)buf - (char *)buf_arg);
1180 return (buf);
1181 }
1182 *buf = new;
1183 }
1184
1185 return (NULL);
1186 }
1187
1188 static void
1189 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1190 {
1191 kmem_cache_t *cp;
1192
1193 mutex_enter(&kmem_cache_lock);
1194 for (cp = list_head(&kmem_caches); cp != NULL;
1195 cp = list_next(&kmem_caches, cp))
1196 if (tq != NULL)
1197 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1198 tqflag);
1199 else
1200 func(cp);
1201 mutex_exit(&kmem_cache_lock);
1202 }
1203
1204 static void
1205 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1206 {
1207 kmem_cache_t *cp;
1208
1209 mutex_enter(&kmem_cache_lock);
1210 for (cp = list_head(&kmem_caches); cp != NULL;
1211 cp = list_next(&kmem_caches, cp)) {
1212 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1213 continue;
1214 if (tq != NULL)
1215 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1216 tqflag);
1217 else
1218 func(cp);
1219 }
1220 mutex_exit(&kmem_cache_lock);
1221 }
1222
1223 /*
1224 * Debugging support. Given a buffer address, find its slab.
1225 */
1226 static kmem_slab_t *
1227 kmem_findslab(kmem_cache_t *cp, void *buf)
1228 {
1229 kmem_slab_t *sp;
1230
1231 mutex_enter(&cp->cache_lock);
1232 for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1233 sp = list_next(&cp->cache_complete_slabs, sp)) {
1234 if (KMEM_SLAB_MEMBER(sp, buf)) {
1235 mutex_exit(&cp->cache_lock);
1236 return (sp);
1237 }
1238 }
1239 for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1240 sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1241 if (KMEM_SLAB_MEMBER(sp, buf)) {
1242 mutex_exit(&cp->cache_lock);
1243 return (sp);
1244 }
1245 }
1246 mutex_exit(&cp->cache_lock);
1247
1248 return (NULL);
1249 }
1250
1251 static void
1252 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1253 {
1254 kmem_buftag_t *btp = NULL;
1255 kmem_bufctl_t *bcp = NULL;
1256 kmem_cache_t *cp = cparg;
1257 kmem_slab_t *sp;
1258 uint64_t *off;
1259 void *buf = bufarg;
1260
1261 kmem_logging = 0; /* stop logging when a bad thing happens */
1262
1263 kmem_panic_info.kmp_timestamp = gethrtime();
1264
1265 sp = kmem_findslab(cp, buf);
1266 if (sp == NULL) {
1267 for (cp = list_tail(&kmem_caches); cp != NULL;
1268 cp = list_prev(&kmem_caches, cp)) {
1269 if ((sp = kmem_findslab(cp, buf)) != NULL)
1270 break;
1271 }
1272 }
1273
1274 if (sp == NULL) {
1275 cp = NULL;
1276 error = KMERR_BADADDR;
1277 } else {
1278 if (cp != cparg)
1279 error = KMERR_BADCACHE;
1280 else
1281 buf = (char *)bufarg - ((uintptr_t)bufarg -
1282 (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1283 if (buf != bufarg)
1284 error = KMERR_BADBASE;
1285 if (cp->cache_flags & KMF_BUFTAG)
1286 btp = KMEM_BUFTAG(cp, buf);
1287 if (cp->cache_flags & KMF_HASH) {
1288 mutex_enter(&cp->cache_lock);
1289 for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1290 if (bcp->bc_addr == buf)
1291 break;
1292 mutex_exit(&cp->cache_lock);
1293 if (bcp == NULL && btp != NULL)
1294 bcp = btp->bt_bufctl;
1295 if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1296 NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1297 bcp->bc_addr != buf) {
1298 error = KMERR_BADBUFCTL;
1299 bcp = NULL;
1300 }
1301 }
1302 }
1303
1304 kmem_panic_info.kmp_error = error;
1305 kmem_panic_info.kmp_buffer = bufarg;
1306 kmem_panic_info.kmp_realbuf = buf;
1307 kmem_panic_info.kmp_cache = cparg;
1308 kmem_panic_info.kmp_realcache = cp;
1309 kmem_panic_info.kmp_slab = sp;
1310 kmem_panic_info.kmp_bufctl = bcp;
1311
1312 printf("kernel memory allocator: ");
1313
1314 switch (error) {
1315
1316 case KMERR_MODIFIED:
1317 printf("buffer modified after being freed\n");
1318 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1319 if (off == NULL) /* shouldn't happen */
1320 off = buf;
1321 printf("modification occurred at offset 0x%lx "
1322 "(0x%llx replaced by 0x%llx)\n",
1323 (uintptr_t)off - (uintptr_t)buf,
1324 (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1325 break;
1326
1327 case KMERR_REDZONE:
1328 printf("redzone violation: write past end of buffer\n");
1329 break;
1330
1331 case KMERR_BADADDR:
1332 printf("invalid free: buffer not in cache\n");
1333 break;
1334
1335 case KMERR_DUPFREE:
1336 printf("duplicate free: buffer freed twice\n");
1337 break;
1338
1339 case KMERR_BADBUFTAG:
1340 printf("boundary tag corrupted\n");
1341 printf("bcp ^ bxstat = %lx, should be %lx\n",
1342 (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1343 KMEM_BUFTAG_FREE);
1344 break;
1345
1346 case KMERR_BADBUFCTL:
1347 printf("bufctl corrupted\n");
1348 break;
1349
1350 case KMERR_BADCACHE:
1351 printf("buffer freed to wrong cache\n");
1352 printf("buffer was allocated from %s,\n", cp->cache_name);
1353 printf("caller attempting free to %s.\n", cparg->cache_name);
1354 break;
1355
1356 case KMERR_BADSIZE:
1357 printf("bad free: free size (%u) != alloc size (%u)\n",
1358 KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1359 KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1360 break;
1361
1362 case KMERR_BADBASE:
1363 printf("bad free: free address (%p) != alloc address (%p)\n",
1364 bufarg, buf);
1365 break;
1366 }
1367
1368 printf("buffer=%p bufctl=%p cache: %s\n",
1369 bufarg, (void *)bcp, cparg->cache_name);
1370
1371 if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1372 error != KMERR_BADBUFCTL) {
1373 int d;
1374 timestruc_t ts;
1375 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1376
1377 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1378 printf("previous transaction on buffer %p:\n", buf);
1379 printf("thread=%p time=T-%ld.%09ld slab=%p cache: %s\n",
1380 (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1381 (void *)sp, cp->cache_name);
1382 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1383 ulong_t off;
1384 char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1385 printf("%s+%lx\n", sym ? sym : "?", off);
1386 }
1387 }
1388 if (kmem_panic > 0)
1389 panic("kernel heap corruption detected");
1390 if (kmem_panic == 0)
1391 debug_enter(NULL);
1392 kmem_logging = 1; /* resume logging */
1393 }
1394
1395 static kmem_log_header_t *
1396 kmem_log_init(size_t logsize)
1397 {
1398 kmem_log_header_t *lhp;
1399 int nchunks = 4 * max_ncpus;
1400 size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1401 int i;
1402
1403 /*
1404 * Make sure that lhp->lh_cpu[] is nicely aligned
1405 * to prevent false sharing of cache lines.
1406 */
1407 lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1408 lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1409 NULL, NULL, VM_SLEEP);
1410 bzero(lhp, lhsize);
1411
1412 mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1413 lhp->lh_nchunks = nchunks;
1414 lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1415 lhp->lh_base = vmem_alloc(kmem_log_arena,
1416 lhp->lh_chunksize * nchunks, VM_SLEEP);
1417 lhp->lh_free = vmem_alloc(kmem_log_arena,
1418 nchunks * sizeof (int), VM_SLEEP);
1419 bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1420
1421 for (i = 0; i < max_ncpus; i++) {
1422 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1423 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1424 clhp->clh_chunk = i;
1425 }
1426
1427 for (i = max_ncpus; i < nchunks; i++)
1428 lhp->lh_free[i] = i;
1429
1430 lhp->lh_head = max_ncpus;
1431 lhp->lh_tail = 0;
1432
1433 return (lhp);
1434 }
1435
1436 static void *
1437 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1438 {
1439 void *logspace;
1440 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1441
1442 if (lhp == NULL || kmem_logging == 0 || panicstr)
1443 return (NULL);
1444
1445 mutex_enter(&clhp->clh_lock);
1446 clhp->clh_hits++;
1447 if (size > clhp->clh_avail) {
1448 mutex_enter(&lhp->lh_lock);
1449 lhp->lh_hits++;
1450 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1451 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1452 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1453 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1454 clhp->clh_current = lhp->lh_base +
1455 clhp->clh_chunk * lhp->lh_chunksize;
1456 clhp->clh_avail = lhp->lh_chunksize;
1457 if (size > lhp->lh_chunksize)
1458 size = lhp->lh_chunksize;
1459 mutex_exit(&lhp->lh_lock);
1460 }
1461 logspace = clhp->clh_current;
1462 clhp->clh_current += size;
1463 clhp->clh_avail -= size;
1464 bcopy(data, logspace, size);
1465 mutex_exit(&clhp->clh_lock);
1466 return (logspace);
1467 }
1468
1469 #define KMEM_AUDIT(lp, cp, bcp) \
1470 { \
1471 kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp); \
1472 _bcp->bc_timestamp = gethrtime(); \
1473 _bcp->bc_thread = curthread; \
1474 _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH); \
1475 _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp)); \
1476 }
1477
1478 static void
1479 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1480 kmem_slab_t *sp, void *addr)
1481 {
1482 kmem_bufctl_audit_t bca;
1483
1484 bzero(&bca, sizeof (kmem_bufctl_audit_t));
1485 bca.bc_addr = addr;
1486 bca.bc_slab = sp;
1487 bca.bc_cache = cp;
1488 KMEM_AUDIT(lp, cp, &bca);
1489 }
1490
1491 /*
1492 * Create a new slab for cache cp.
1493 */
1494 static kmem_slab_t *
1495 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1496 {
1497 size_t slabsize = cp->cache_slabsize;
1498 size_t chunksize = cp->cache_chunksize;
1499 int cache_flags = cp->cache_flags;
1500 size_t color, chunks;
1501 char *buf, *slab;
1502 kmem_slab_t *sp;
1503 kmem_bufctl_t *bcp;
1504 vmem_t *vmp = cp->cache_arena;
1505
1506 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1507
1508 color = cp->cache_color + cp->cache_align;
1509 if (color > cp->cache_maxcolor)
1510 color = cp->cache_mincolor;
1511 cp->cache_color = color;
1512
1513 slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1514
1515 if (slab == NULL)
1516 goto vmem_alloc_failure;
1517
1518 ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1519
1520 /*
1521 * Reverify what was already checked in kmem_cache_set_move(), since the
1522 * consolidator depends (for correctness) on slabs being initialized
1523 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1524 * clients to distinguish uninitialized memory from known objects).
1525 */
1526 ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1527 if (!(cp->cache_cflags & KMC_NOTOUCH))
1528 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1529
1530 if (cache_flags & KMF_HASH) {
1531 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1532 goto slab_alloc_failure;
1533 chunks = (slabsize - color) / chunksize;
1534 } else {
1535 sp = KMEM_SLAB(cp, slab);
1536 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1537 }
1538
1539 sp->slab_cache = cp;
1540 sp->slab_head = NULL;
1541 sp->slab_refcnt = 0;
1542 sp->slab_base = buf = slab + color;
1543 sp->slab_chunks = chunks;
1544 sp->slab_stuck_offset = (uint32_t)-1;
1545 sp->slab_later_count = 0;
1546 sp->slab_flags = 0;
1547
1548 ASSERT(chunks > 0);
1549 while (chunks-- != 0) {
1550 if (cache_flags & KMF_HASH) {
1551 bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1552 if (bcp == NULL)
1553 goto bufctl_alloc_failure;
1554 if (cache_flags & KMF_AUDIT) {
1555 kmem_bufctl_audit_t *bcap =
1556 (kmem_bufctl_audit_t *)bcp;
1557 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1558 bcap->bc_cache = cp;
1559 }
1560 bcp->bc_addr = buf;
1561 bcp->bc_slab = sp;
1562 } else {
1563 bcp = KMEM_BUFCTL(cp, buf);
1564 }
1565 if (cache_flags & KMF_BUFTAG) {
1566 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1567 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1568 btp->bt_bufctl = bcp;
1569 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1570 if (cache_flags & KMF_DEADBEEF) {
1571 copy_pattern(KMEM_FREE_PATTERN, buf,
1572 cp->cache_verify);
1573 }
1574 }
1575 bcp->bc_next = sp->slab_head;
1576 sp->slab_head = bcp;
1577 buf += chunksize;
1578 }
1579
1580 kmem_log_event(kmem_slab_log, cp, sp, slab);
1581
1582 return (sp);
1583
1584 bufctl_alloc_failure:
1585
1586 while ((bcp = sp->slab_head) != NULL) {
1587 sp->slab_head = bcp->bc_next;
1588 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1589 }
1590 kmem_cache_free(kmem_slab_cache, sp);
1591
1592 slab_alloc_failure:
1593
1594 vmem_free(vmp, slab, slabsize);
1595
1596 vmem_alloc_failure:
1597
1598 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1599 atomic_inc_64(&cp->cache_alloc_fail);
1600
1601 return (NULL);
1602 }
1603
1604 /*
1605 * Destroy a slab.
1606 */
1607 static void
1608 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1609 {
1610 vmem_t *vmp = cp->cache_arena;
1611 void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1612
1613 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1614 ASSERT(sp->slab_refcnt == 0);
1615
1616 if (cp->cache_flags & KMF_HASH) {
1617 kmem_bufctl_t *bcp;
1618 while ((bcp = sp->slab_head) != NULL) {
1619 sp->slab_head = bcp->bc_next;
1620 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1621 }
1622 kmem_cache_free(kmem_slab_cache, sp);
1623 }
1624 vmem_free(vmp, slab, cp->cache_slabsize);
1625 }
1626
1627 static void *
1628 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1629 {
1630 kmem_bufctl_t *bcp, **hash_bucket;
1631 void *buf;
1632 boolean_t new_slab = (sp->slab_refcnt == 0);
1633
1634 ASSERT(MUTEX_HELD(&cp->cache_lock));
1635 /*
1636 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1637 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1638 * slab is newly created.
1639 */
1640 ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1641 (sp == avl_first(&cp->cache_partial_slabs))));
1642 ASSERT(sp->slab_cache == cp);
1643
1644 cp->cache_slab_alloc++;
1645 cp->cache_bufslab--;
1646 sp->slab_refcnt++;
1647
1648 bcp = sp->slab_head;
1649 sp->slab_head = bcp->bc_next;
1650
1651 if (cp->cache_flags & KMF_HASH) {
1652 /*
1653 * Add buffer to allocated-address hash table.
1654 */
1655 buf = bcp->bc_addr;
1656 hash_bucket = KMEM_HASH(cp, buf);
1657 bcp->bc_next = *hash_bucket;
1658 *hash_bucket = bcp;
1659 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1660 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1661 }
1662 } else {
1663 buf = KMEM_BUF(cp, bcp);
1664 }
1665
1666 ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1667
1668 if (sp->slab_head == NULL) {
1669 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1670 if (new_slab) {
1671 ASSERT(sp->slab_chunks == 1);
1672 } else {
1673 ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1674 avl_remove(&cp->cache_partial_slabs, sp);
1675 sp->slab_later_count = 0; /* clear history */
1676 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1677 sp->slab_stuck_offset = (uint32_t)-1;
1678 }
1679 list_insert_head(&cp->cache_complete_slabs, sp);
1680 cp->cache_complete_slab_count++;
1681 return (buf);
1682 }
1683
1684 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1685 /*
1686 * Peek to see if the magazine layer is enabled before
1687 * we prefill. We're not holding the cpu cache lock,
1688 * so the peek could be wrong, but there's no harm in it.
1689 */
1690 if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1691 (KMEM_CPU_CACHE(cp)->cc_magsize != 0)) {
1692 kmem_slab_prefill(cp, sp);
1693 return (buf);
1694 }
1695
1696 if (new_slab) {
1697 avl_add(&cp->cache_partial_slabs, sp);
1698 return (buf);
1699 }
1700
1701 /*
1702 * The slab is now more allocated than it was, so the
1703 * order remains unchanged.
1704 */
1705 ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1706 return (buf);
1707 }
1708
1709 /*
1710 * Allocate a raw (unconstructed) buffer from cp's slab layer.
1711 */
1712 static void *
1713 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1714 {
1715 kmem_slab_t *sp;
1716 void *buf;
1717 boolean_t test_destructor;
1718
1719 mutex_enter(&cp->cache_lock);
1720 test_destructor = (cp->cache_slab_alloc == 0);
1721 sp = avl_first(&cp->cache_partial_slabs);
1722 if (sp == NULL) {
1723 ASSERT(cp->cache_bufslab == 0);
1724
1725 /*
1726 * The freelist is empty. Create a new slab.
1727 */
1728 mutex_exit(&cp->cache_lock);
1729 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1730 return (NULL);
1731 }
1732 mutex_enter(&cp->cache_lock);
1733 cp->cache_slab_create++;
1734 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1735 cp->cache_bufmax = cp->cache_buftotal;
1736 cp->cache_bufslab += sp->slab_chunks;
1737 }
1738
1739 buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1740 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1741 (cp->cache_complete_slab_count +
1742 avl_numnodes(&cp->cache_partial_slabs) +
1743 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1744 mutex_exit(&cp->cache_lock);
1745
1746 if (test_destructor && cp->cache_destructor != NULL) {
1747 /*
1748 * On the first kmem_slab_alloc(), assert that it is valid to
1749 * call the destructor on a newly constructed object without any
1750 * client involvement.
1751 */
1752 if ((cp->cache_constructor == NULL) ||
1753 cp->cache_constructor(buf, cp->cache_private,
1754 kmflag) == 0) {
1755 cp->cache_destructor(buf, cp->cache_private);
1756 }
1757 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1758 cp->cache_bufsize);
1759 if (cp->cache_flags & KMF_DEADBEEF) {
1760 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1761 }
1762 }
1763
1764 return (buf);
1765 }
1766
1767 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1768
1769 /*
1770 * Free a raw (unconstructed) buffer to cp's slab layer.
1771 */
1772 static void
1773 kmem_slab_free(kmem_cache_t *cp, void *buf)
1774 {
1775 kmem_slab_t *sp;
1776 kmem_bufctl_t *bcp, **prev_bcpp;
1777
1778 ASSERT(buf != NULL);
1779
1780 mutex_enter(&cp->cache_lock);
1781 cp->cache_slab_free++;
1782
1783 if (cp->cache_flags & KMF_HASH) {
1784 /*
1785 * Look up buffer in allocated-address hash table.
1786 */
1787 prev_bcpp = KMEM_HASH(cp, buf);
1788 while ((bcp = *prev_bcpp) != NULL) {
1789 if (bcp->bc_addr == buf) {
1790 *prev_bcpp = bcp->bc_next;
1791 sp = bcp->bc_slab;
1792 break;
1793 }
1794 cp->cache_lookup_depth++;
1795 prev_bcpp = &bcp->bc_next;
1796 }
1797 } else {
1798 bcp = KMEM_BUFCTL(cp, buf);
1799 sp = KMEM_SLAB(cp, buf);
1800 }
1801
1802 if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1803 mutex_exit(&cp->cache_lock);
1804 kmem_error(KMERR_BADADDR, cp, buf);
1805 return;
1806 }
1807
1808 if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1809 /*
1810 * If this is the buffer that prevented the consolidator from
1811 * clearing the slab, we can reset the slab flags now that the
1812 * buffer is freed. (It makes sense to do this in
1813 * kmem_cache_free(), where the client gives up ownership of the
1814 * buffer, but on the hot path the test is too expensive.)
1815 */
1816 kmem_slab_move_yes(cp, sp, buf);
1817 }
1818
1819 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1820 if (cp->cache_flags & KMF_CONTENTS)
1821 ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1822 kmem_log_enter(kmem_content_log, buf,
1823 cp->cache_contents);
1824 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1825 }
1826
1827 bcp->bc_next = sp->slab_head;
1828 sp->slab_head = bcp;
1829
1830 cp->cache_bufslab++;
1831 ASSERT(sp->slab_refcnt >= 1);
1832
1833 if (--sp->slab_refcnt == 0) {
1834 /*
1835 * There are no outstanding allocations from this slab,
1836 * so we can reclaim the memory.
1837 */
1838 if (sp->slab_chunks == 1) {
1839 list_remove(&cp->cache_complete_slabs, sp);
1840 cp->cache_complete_slab_count--;
1841 } else {
1842 avl_remove(&cp->cache_partial_slabs, sp);
1843 }
1844
1845 cp->cache_buftotal -= sp->slab_chunks;
1846 cp->cache_bufslab -= sp->slab_chunks;
1847 /*
1848 * Defer releasing the slab to the virtual memory subsystem
1849 * while there is a pending move callback, since we guarantee
1850 * that buffers passed to the move callback have only been
1851 * touched by kmem or by the client itself. Since the memory
1852 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1853 * set at least one of the two lowest order bits, the client can
1854 * test those bits in the move callback to determine whether or
1855 * not it knows about the buffer (assuming that the client also
1856 * sets one of those low order bits whenever it frees a buffer).
1857 */
1858 if (cp->cache_defrag == NULL ||
1859 (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1860 !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1861 cp->cache_slab_destroy++;
1862 mutex_exit(&cp->cache_lock);
1863 kmem_slab_destroy(cp, sp);
1864 } else {
1865 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1866 /*
1867 * Slabs are inserted at both ends of the deadlist to
1868 * distinguish between slabs freed while move callbacks
1869 * are pending (list head) and a slab freed while the
1870 * lock is dropped in kmem_move_buffers() (list tail) so
1871 * that in both cases slab_destroy() is called from the
1872 * right context.
1873 */
1874 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1875 list_insert_tail(deadlist, sp);
1876 } else {
1877 list_insert_head(deadlist, sp);
1878 }
1879 cp->cache_defrag->kmd_deadcount++;
1880 mutex_exit(&cp->cache_lock);
1881 }
1882 return;
1883 }
1884
1885 if (bcp->bc_next == NULL) {
1886 /* Transition the slab from completely allocated to partial. */
1887 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1888 ASSERT(sp->slab_chunks > 1);
1889 list_remove(&cp->cache_complete_slabs, sp);
1890 cp->cache_complete_slab_count--;
1891 avl_add(&cp->cache_partial_slabs, sp);
1892 } else {
1893 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1894 }
1895
1896 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1897 (cp->cache_complete_slab_count +
1898 avl_numnodes(&cp->cache_partial_slabs) +
1899 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1900 mutex_exit(&cp->cache_lock);
1901 }
1902
1903 /*
1904 * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1905 */
1906 static int
1907 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1908 caddr_t caller)
1909 {
1910 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1911 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1912 uint32_t mtbf;
1913
1914 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1915 kmem_error(KMERR_BADBUFTAG, cp, buf);
1916 return (-1);
1917 }
1918
1919 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1920
1921 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1922 kmem_error(KMERR_BADBUFCTL, cp, buf);
1923 return (-1);
1924 }
1925
1926 if (cp->cache_flags & KMF_DEADBEEF) {
1927 if (!construct && (cp->cache_flags & KMF_LITE)) {
1928 if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1929 kmem_error(KMERR_MODIFIED, cp, buf);
1930 return (-1);
1931 }
1932 if (cp->cache_constructor != NULL)
1933 *(uint64_t *)buf = btp->bt_redzone;
1934 else
1935 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1936 } else {
1937 construct = 1;
1938 if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1939 KMEM_UNINITIALIZED_PATTERN, buf,
1940 cp->cache_verify)) {
1941 kmem_error(KMERR_MODIFIED, cp, buf);
1942 return (-1);
1943 }
1944 }
1945 }
1946 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1947
1948 if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1949 gethrtime() % mtbf == 0 &&
1950 (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1951 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1952 if (!construct && cp->cache_destructor != NULL)
1953 cp->cache_destructor(buf, cp->cache_private);
1954 } else {
1955 mtbf = 0;
1956 }
1957
1958 if (mtbf || (construct && cp->cache_constructor != NULL &&
1959 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1960 atomic_inc_64(&cp->cache_alloc_fail);
1961 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1962 if (cp->cache_flags & KMF_DEADBEEF)
1963 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1964 kmem_slab_free(cp, buf);
1965 return (1);
1966 }
1967
1968 if (cp->cache_flags & KMF_AUDIT) {
1969 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1970 }
1971
1972 if ((cp->cache_flags & KMF_LITE) &&
1973 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1974 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1975 }
1976
1977 return (0);
1978 }
1979
1980 static int
1981 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1982 {
1983 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1984 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1985 kmem_slab_t *sp;
1986
1987 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1988 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1989 kmem_error(KMERR_DUPFREE, cp, buf);
1990 return (-1);
1991 }
1992 sp = kmem_findslab(cp, buf);
1993 if (sp == NULL || sp->slab_cache != cp)
1994 kmem_error(KMERR_BADADDR, cp, buf);
1995 else
1996 kmem_error(KMERR_REDZONE, cp, buf);
1997 return (-1);
1998 }
1999
2000 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2001
2002 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2003 kmem_error(KMERR_BADBUFCTL, cp, buf);
2004 return (-1);
2005 }
2006
2007 if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2008 kmem_error(KMERR_REDZONE, cp, buf);
2009 return (-1);
2010 }
2011
2012 if (cp->cache_flags & KMF_AUDIT) {
2013 if (cp->cache_flags & KMF_CONTENTS)
2014 bcp->bc_contents = kmem_log_enter(kmem_content_log,
2015 buf, cp->cache_contents);
2016 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2017 }
2018
2019 if ((cp->cache_flags & KMF_LITE) &&
2020 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2021 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2022 }
2023
2024 if (cp->cache_flags & KMF_DEADBEEF) {
2025 if (cp->cache_flags & KMF_LITE)
2026 btp->bt_redzone = *(uint64_t *)buf;
2027 else if (cp->cache_destructor != NULL)
2028 cp->cache_destructor(buf, cp->cache_private);
2029
2030 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2031 }
2032
2033 return (0);
2034 }
2035
2036 /*
2037 * Free each object in magazine mp to cp's slab layer, and free mp itself.
2038 */
2039 static void
2040 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2041 {
2042 int round;
2043
2044 ASSERT(!list_link_active(&cp->cache_link) ||
2045 taskq_member(kmem_taskq, curthread));
2046
2047 for (round = 0; round < nrounds; round++) {
2048 void *buf = mp->mag_round[round];
2049
2050 if (cp->cache_flags & KMF_DEADBEEF) {
2051 if (verify_pattern(KMEM_FREE_PATTERN, buf,
2052 cp->cache_verify) != NULL) {
2053 kmem_error(KMERR_MODIFIED, cp, buf);
2054 continue;
2055 }
2056 if ((cp->cache_flags & KMF_LITE) &&
2057 cp->cache_destructor != NULL) {
2058 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2059 *(uint64_t *)buf = btp->bt_redzone;
2060 cp->cache_destructor(buf, cp->cache_private);
2061 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2062 }
2063 } else if (cp->cache_destructor != NULL) {
2064 cp->cache_destructor(buf, cp->cache_private);
2065 }
2066
2067 kmem_slab_free(cp, buf);
2068 }
2069 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2070 kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2071 }
2072
2073 /*
2074 * Allocate a magazine from the depot.
2075 */
2076 static kmem_magazine_t *
2077 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2078 {
2079 kmem_magazine_t *mp;
2080
2081 /*
2082 * If we can't get the depot lock without contention,
2083 * update our contention count. We use the depot
2084 * contention rate to determine whether we need to
2085 * increase the magazine size for better scalability.
2086 */
2087 if (!mutex_tryenter(&cp->cache_depot_lock)) {
2088 mutex_enter(&cp->cache_depot_lock);
2089 cp->cache_depot_contention++;
2090 }
2091
2092 if ((mp = mlp->ml_list) != NULL) {
2093 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2094 mlp->ml_list = mp->mag_next;
2095 if (--mlp->ml_total < mlp->ml_min)
2096 mlp->ml_min = mlp->ml_total;
2097 mlp->ml_alloc++;
2098 }
2099
2100 mutex_exit(&cp->cache_depot_lock);
2101
2102 return (mp);
2103 }
2104
2105 /*
2106 * Free a magazine to the depot.
2107 */
2108 static void
2109 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2110 {
2111 mutex_enter(&cp->cache_depot_lock);
2112 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2113 mp->mag_next = mlp->ml_list;
2114 mlp->ml_list = mp;
2115 mlp->ml_total++;
2116 mutex_exit(&cp->cache_depot_lock);
2117 }
2118
2119 /*
2120 * Update the working set statistics for cp's depot.
2121 */
2122 static void
2123 kmem_depot_ws_update(kmem_cache_t *cp)
2124 {
2125 mutex_enter(&cp->cache_depot_lock);
2126 cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2127 cp->cache_full.ml_min = cp->cache_full.ml_total;
2128 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2129 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2130 mutex_exit(&cp->cache_depot_lock);
2131 }
2132
2133 /*
2134 * Set the working set statistics for cp's depot to zero. (Everything is
2135 * eligible for reaping.)
2136 */
2137 static void
2138 kmem_depot_ws_zero(kmem_cache_t *cp)
2139 {
2140 mutex_enter(&cp->cache_depot_lock);
2141 cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2142 cp->cache_full.ml_min = cp->cache_full.ml_total;
2143 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2144 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2145 mutex_exit(&cp->cache_depot_lock);
2146 }
2147
2148 /*
2149 * The number of bytes to reap before we call kpreempt(). The default (1MB)
2150 * causes us to preempt reaping up to hundreds of times per second. Using a
2151 * larger value (1GB) causes this to have virtually no effect.
2152 */
2153 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2154
2155 /*
2156 * Reap all magazines that have fallen out of the depot's working set.
2157 */
2158 static void
2159 kmem_depot_ws_reap(kmem_cache_t *cp)
2160 {
2161 size_t bytes = 0;
2162 long reap;
2163 kmem_magazine_t *mp;
2164
2165 ASSERT(!list_link_active(&cp->cache_link) ||
2166 taskq_member(kmem_taskq, curthread));
2167
2168 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2169 while (reap-- &&
2170 (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2171 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2172 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2173 if (bytes > kmem_reap_preempt_bytes) {
2174 kpreempt(KPREEMPT_SYNC);
2175 bytes = 0;
2176 }
2177 }
2178
2179 reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2180 while (reap-- &&
2181 (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2182 kmem_magazine_destroy(cp, mp, 0);
2183 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2184 if (bytes > kmem_reap_preempt_bytes) {
2185 kpreempt(KPREEMPT_SYNC);
2186 bytes = 0;
2187 }
2188 }
2189 }
2190
2191 static void
2192 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2193 {
2194 ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2195 (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2196 ASSERT(ccp->cc_magsize > 0);
2197
2198 ccp->cc_ploaded = ccp->cc_loaded;
2199 ccp->cc_prounds = ccp->cc_rounds;
2200 ccp->cc_loaded = mp;
2201 ccp->cc_rounds = rounds;
2202 }
2203
2204 /*
2205 * Intercept kmem alloc/free calls during crash dump in order to avoid
2206 * changing kmem state while memory is being saved to the dump device.
2207 * Otherwise, ::kmem_verify will report "corrupt buffers". Note that
2208 * there are no locks because only one CPU calls kmem during a crash
2209 * dump. To enable this feature, first create the associated vmem
2210 * arena with VMC_DUMPSAFE.
2211 */
2212 static void *kmem_dump_start; /* start of pre-reserved heap */
2213 static void *kmem_dump_end; /* end of heap area */
2214 static void *kmem_dump_curr; /* current free heap pointer */
2215 static size_t kmem_dump_size; /* size of heap area */
2216
2217 /* append to each buf created in the pre-reserved heap */
2218 typedef struct kmem_dumpctl {
2219 void *kdc_next; /* cache dump free list linkage */
2220 } kmem_dumpctl_t;
2221
2222 #define KMEM_DUMPCTL(cp, buf) \
2223 ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2224 sizeof (void *)))
2225
2226 /* Keep some simple stats. */
2227 #define KMEM_DUMP_LOGS (100)
2228
2229 typedef struct kmem_dump_log {
2230 kmem_cache_t *kdl_cache;
2231 uint_t kdl_allocs; /* # of dump allocations */
2232 uint_t kdl_frees; /* # of dump frees */
2233 uint_t kdl_alloc_fails; /* # of allocation failures */
2234 uint_t kdl_free_nondump; /* # of non-dump frees */
2235 uint_t kdl_unsafe; /* cache was used, but unsafe */
2236 } kmem_dump_log_t;
2237
2238 static kmem_dump_log_t *kmem_dump_log;
2239 static int kmem_dump_log_idx;
2240
2241 #define KDI_LOG(cp, stat) { \
2242 kmem_dump_log_t *kdl; \
2243 if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) { \
2244 kdl->stat++; \
2245 } else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) { \
2246 kdl = &kmem_dump_log[kmem_dump_log_idx++]; \
2247 kdl->stat++; \
2248 kdl->kdl_cache = (cp); \
2249 (cp)->cache_dumplog = kdl; \
2250 } \
2251 }
2252
2253 /* set non zero for full report */
2254 uint_t kmem_dump_verbose = 0;
2255
2256 /* stats for overize heap */
2257 uint_t kmem_dump_oversize_allocs = 0;
2258 uint_t kmem_dump_oversize_max = 0;
2259
2260 static void
2261 kmem_dumppr(char **pp, char *e, const char *format, ...)
2262 {
2263 char *p = *pp;
2264
2265 if (p < e) {
2266 int n;
2267 va_list ap;
2268
2269 va_start(ap, format);
2270 n = vsnprintf(p, e - p, format, ap);
2271 va_end(ap);
2272 *pp = p + n;
2273 }
2274 }
2275
2276 /*
2277 * Called when dumpadm(1M) configures dump parameters.
2278 */
2279 void
2280 kmem_dump_init(size_t size)
2281 {
2282 if (kmem_dump_start != NULL)
2283 kmem_free(kmem_dump_start, kmem_dump_size);
2284
2285 if (kmem_dump_log == NULL)
2286 kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2287 sizeof (kmem_dump_log_t), KM_SLEEP);
2288
2289 kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2290
2291 if (kmem_dump_start != NULL) {
2292 kmem_dump_size = size;
2293 kmem_dump_curr = kmem_dump_start;
2294 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2295 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2296 } else {
2297 kmem_dump_size = 0;
2298 kmem_dump_curr = NULL;
2299 kmem_dump_end = NULL;
2300 }
2301 }
2302
2303 /*
2304 * Set flag for each kmem_cache_t if is safe to use alternate dump
2305 * memory. Called just before panic crash dump starts. Set the flag
2306 * for the calling CPU.
2307 */
2308 void
2309 kmem_dump_begin(void)
2310 {
2311 ASSERT(panicstr != NULL);
2312 if (kmem_dump_start != NULL) {
2313 kmem_cache_t *cp;
2314
2315 for (cp = list_head(&kmem_caches); cp != NULL;
2316 cp = list_next(&kmem_caches, cp)) {
2317 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2318
2319 if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2320 cp->cache_flags |= KMF_DUMPDIVERT;
2321 ccp->cc_flags |= KMF_DUMPDIVERT;
2322 ccp->cc_dump_rounds = ccp->cc_rounds;
2323 ccp->cc_dump_prounds = ccp->cc_prounds;
2324 ccp->cc_rounds = ccp->cc_prounds = -1;
2325 } else {
2326 cp->cache_flags |= KMF_DUMPUNSAFE;
2327 ccp->cc_flags |= KMF_DUMPUNSAFE;
2328 }
2329 }
2330 }
2331 }
2332
2333 /*
2334 * finished dump intercept
2335 * print any warnings on the console
2336 * return verbose information to dumpsys() in the given buffer
2337 */
2338 size_t
2339 kmem_dump_finish(char *buf, size_t size)
2340 {
2341 int kdi_idx;
2342 int kdi_end = kmem_dump_log_idx;
2343 int percent = 0;
2344 int header = 0;
2345 int warn = 0;
2346 size_t used;
2347 kmem_cache_t *cp;
2348 kmem_dump_log_t *kdl;
2349 char *e = buf + size;
2350 char *p = buf;
2351
2352 if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2353 return (0);
2354
2355 used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2356 percent = (used * 100) / kmem_dump_size;
2357
2358 kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2359 kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2360 kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2361 kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2362 kmem_dump_oversize_allocs);
2363 kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2364 kmem_dump_oversize_max);
2365
2366 for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2367 kdl = &kmem_dump_log[kdi_idx];
2368 cp = kdl->kdl_cache;
2369 if (cp == NULL)
2370 break;
2371 if (kdl->kdl_alloc_fails)
2372 ++warn;
2373 if (header == 0) {
2374 kmem_dumppr(&p, e,
2375 "Cache Name,Allocs,Frees,Alloc Fails,"
2376 "Nondump Frees,Unsafe Allocs/Frees\n");
2377 header = 1;
2378 }
2379 kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2380 cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2381 kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2382 kdl->kdl_unsafe);
2383 }
2384
2385 /* return buffer size used */
2386 if (p < e)
2387 bzero(p, e - p);
2388 return (p - buf);
2389 }
2390
2391 /*
2392 * Allocate a constructed object from alternate dump memory.
2393 */
2394 void *
2395 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2396 {
2397 void *buf;
2398 void *curr;
2399 char *bufend;
2400
2401 /* return a constructed object */
2402 if ((buf = cp->cache_dumpfreelist) != NULL) {
2403 cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2404 KDI_LOG(cp, kdl_allocs);
2405 return (buf);
2406 }
2407
2408 /* create a new constructed object */
2409 curr = kmem_dump_curr;
2410 buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2411 bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2412
2413 /* hat layer objects cannot cross a page boundary */
2414 if (cp->cache_align < PAGESIZE) {
2415 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2416 if (bufend > page) {
2417 bufend += page - (char *)buf;
2418 buf = (void *)page;
2419 }
2420 }
2421
2422 /* fall back to normal alloc if reserved area is used up */
2423 if (bufend > (char *)kmem_dump_end) {
2424 kmem_dump_curr = kmem_dump_end;
2425 KDI_LOG(cp, kdl_alloc_fails);
2426 return (NULL);
2427 }
2428
2429 /*
2430 * Must advance curr pointer before calling a constructor that
2431 * may also allocate memory.
2432 */
2433 kmem_dump_curr = bufend;
2434
2435 /* run constructor */
2436 if (cp->cache_constructor != NULL &&
2437 cp->cache_constructor(buf, cp->cache_private, kmflag)
2438 != 0) {
2439 #ifdef DEBUG
2440 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2441 cp->cache_name, (void *)cp);
2442 #endif
2443 /* reset curr pointer iff no allocs were done */
2444 if (kmem_dump_curr == bufend)
2445 kmem_dump_curr = curr;
2446
2447 /* fall back to normal alloc if the constructor fails */
2448 KDI_LOG(cp, kdl_alloc_fails);
2449 return (NULL);
2450 }
2451
2452 KDI_LOG(cp, kdl_allocs);
2453 return (buf);
2454 }
2455
2456 /*
2457 * Free a constructed object in alternate dump memory.
2458 */
2459 int
2460 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2461 {
2462 /* save constructed buffers for next time */
2463 if ((char *)buf >= (char *)kmem_dump_start &&
2464 (char *)buf < (char *)kmem_dump_end) {
2465 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2466 cp->cache_dumpfreelist = buf;
2467 KDI_LOG(cp, kdl_frees);
2468 return (0);
2469 }
2470
2471 /* count all non-dump buf frees */
2472 KDI_LOG(cp, kdl_free_nondump);
2473
2474 /* just drop buffers that were allocated before dump started */
2475 if (kmem_dump_curr < kmem_dump_end)
2476 return (0);
2477
2478 /* fall back to normal free if reserved area is used up */
2479 return (1);
2480 }
2481
2482 /*
2483 * Allocate a constructed object from cache cp.
2484 */
2485 void *
2486 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2487 {
2488 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2489 kmem_magazine_t *fmp;
2490 void *buf;
2491
2492 mutex_enter(&ccp->cc_lock);
2493 for (;;) {
2494 /*
2495 * If there's an object available in the current CPU's
2496 * loaded magazine, just take it and return.
2497 */
2498 if (ccp->cc_rounds > 0) {
2499 buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2500 ccp->cc_alloc++;
2501 mutex_exit(&ccp->cc_lock);
2502 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2503 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2504 ASSERT(!(ccp->cc_flags &
2505 KMF_DUMPDIVERT));
2506 KDI_LOG(cp, kdl_unsafe);
2507 }
2508 if ((ccp->cc_flags & KMF_BUFTAG) &&
2509 kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2510 caller()) != 0) {
2511 if (kmflag & KM_NOSLEEP)
2512 return (NULL);
2513 mutex_enter(&ccp->cc_lock);
2514 continue;
2515 }
2516 }
2517 return (buf);
2518 }
2519
2520 /*
2521 * The loaded magazine is empty. If the previously loaded
2522 * magazine was full, exchange them and try again.
2523 */
2524 if (ccp->cc_prounds > 0) {
2525 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2526 continue;
2527 }
2528
2529 /*
2530 * Return an alternate buffer at dump time to preserve
2531 * the heap.
2532 */
2533 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2534 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2535 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2536 /* log it so that we can warn about it */
2537 KDI_LOG(cp, kdl_unsafe);
2538 } else {
2539 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2540 NULL) {
2541 mutex_exit(&ccp->cc_lock);
2542 return (buf);
2543 }
2544 break; /* fall back to slab layer */
2545 }
2546 }
2547
2548 /*
2549 * If the magazine layer is disabled, break out now.
2550 */
2551 if (ccp->cc_magsize == 0)
2552 break;
2553
2554 /*
2555 * Try to get a full magazine from the depot.
2556 */
2557 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2558 if (fmp != NULL) {
2559 if (ccp->cc_ploaded != NULL)
2560 kmem_depot_free(cp, &cp->cache_empty,
2561 ccp->cc_ploaded);
2562 kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2563 continue;
2564 }
2565
2566 /*
2567 * There are no full magazines in the depot,
2568 * so fall through to the slab layer.
2569 */
2570 break;
2571 }
2572 mutex_exit(&ccp->cc_lock);
2573
2574 /*
2575 * We couldn't allocate a constructed object from the magazine layer,
2576 * so get a raw buffer from the slab layer and apply its constructor.
2577 */
2578 buf = kmem_slab_alloc(cp, kmflag);
2579
2580 if (buf == NULL)
2581 return (NULL);
2582
2583 if (cp->cache_flags & KMF_BUFTAG) {
2584 /*
2585 * Make kmem_cache_alloc_debug() apply the constructor for us.
2586 */
2587 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2588 if (rc != 0) {
2589 if (kmflag & KM_NOSLEEP)
2590 return (NULL);
2591 /*
2592 * kmem_cache_alloc_debug() detected corruption
2593 * but didn't panic (kmem_panic <= 0). We should not be
2594 * here because the constructor failed (indicated by a
2595 * return code of 1). Try again.
2596 */
2597 ASSERT(rc == -1);
2598 return (kmem_cache_alloc(cp, kmflag));
2599 }
2600 return (buf);
2601 }
2602
2603 if (cp->cache_constructor != NULL &&
2604 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2605 atomic_inc_64(&cp->cache_alloc_fail);
2606 kmem_slab_free(cp, buf);
2607 return (NULL);
2608 }
2609
2610 return (buf);
2611 }
2612
2613 /*
2614 * The freed argument tells whether or not kmem_cache_free_debug() has already
2615 * been called so that we can avoid the duplicate free error. For example, a
2616 * buffer on a magazine has already been freed by the client but is still
2617 * constructed.
2618 */
2619 static void
2620 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2621 {
2622 if (!freed && (cp->cache_flags & KMF_BUFTAG))
2623 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2624 return;
2625
2626 /*
2627 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2628 * kmem_cache_free_debug() will have already applied the destructor.
2629 */
2630 if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2631 cp->cache_destructor != NULL) {
2632 if (cp->cache_flags & KMF_DEADBEEF) { /* KMF_LITE implied */
2633 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2634 *(uint64_t *)buf = btp->bt_redzone;
2635 cp->cache_destructor(buf, cp->cache_private);
2636 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2637 } else {
2638 cp->cache_destructor(buf, cp->cache_private);
2639 }
2640 }
2641
2642 kmem_slab_free(cp, buf);
2643 }
2644
2645 /*
2646 * Used when there's no room to free a buffer to the per-CPU cache.
2647 * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2648 * caller should try freeing to the per-CPU cache again.
2649 * Note that we don't directly install the magazine in the cpu cache,
2650 * since its state may have changed wildly while the lock was dropped.
2651 */
2652 static int
2653 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2654 {
2655 kmem_magazine_t *emp;
2656 kmem_magtype_t *mtp;
2657
2658 ASSERT(MUTEX_HELD(&ccp->cc_lock));
2659 ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2660 ((uint_t)ccp->cc_rounds == -1)) &&
2661 ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2662 ((uint_t)ccp->cc_prounds == -1)));
2663
2664 emp = kmem_depot_alloc(cp, &cp->cache_empty);
2665 if (emp != NULL) {
2666 if (ccp->cc_ploaded != NULL)
2667 kmem_depot_free(cp, &cp->cache_full,
2668 ccp->cc_ploaded);
2669 kmem_cpu_reload(ccp, emp, 0);
2670 return (1);
2671 }
2672 /*
2673 * There are no empty magazines in the depot,
2674 * so try to allocate a new one. We must drop all locks
2675 * across kmem_cache_alloc() because lower layers may
2676 * attempt to allocate from this cache.
2677 */
2678 mtp = cp->cache_magtype;
2679 mutex_exit(&ccp->cc_lock);
2680 emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2681 mutex_enter(&ccp->cc_lock);
2682
2683 if (emp != NULL) {
2684 /*
2685 * We successfully allocated an empty magazine.
2686 * However, we had to drop ccp->cc_lock to do it,
2687 * so the cache's magazine size may have changed.
2688 * If so, free the magazine and try again.
2689 */
2690 if (ccp->cc_magsize != mtp->mt_magsize) {
2691 mutex_exit(&ccp->cc_lock);
2692 kmem_cache_free(mtp->mt_cache, emp);
2693 mutex_enter(&ccp->cc_lock);
2694 return (1);
2695 }
2696
2697 /*
2698 * We got a magazine of the right size. Add it to
2699 * the depot and try the whole dance again.
2700 */
2701 kmem_depot_free(cp, &cp->cache_empty, emp);
2702 return (1);
2703 }
2704
2705 /*
2706 * We couldn't allocate an empty magazine,
2707 * so fall through to the slab layer.
2708 */
2709 return (0);
2710 }
2711
2712 /*
2713 * Free a constructed object to cache cp.
2714 */
2715 void
2716 kmem_cache_free(kmem_cache_t *cp, void *buf)
2717 {
2718 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2719
2720 /*
2721 * The client must not free either of the buffers passed to the move
2722 * callback function.
2723 */
2724 ASSERT(cp->cache_defrag == NULL ||
2725 cp->cache_defrag->kmd_thread != curthread ||
2726 (buf != cp->cache_defrag->kmd_from_buf &&
2727 buf != cp->cache_defrag->kmd_to_buf));
2728
2729 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2730 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2731 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2732 /* log it so that we can warn about it */
2733 KDI_LOG(cp, kdl_unsafe);
2734 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2735 return;
2736 }
2737 if (ccp->cc_flags & KMF_BUFTAG) {
2738 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2739 return;
2740 }
2741 }
2742
2743 mutex_enter(&ccp->cc_lock);
2744 /*
2745 * Any changes to this logic should be reflected in kmem_slab_prefill()
2746 */
2747 for (;;) {
2748 /*
2749 * If there's a slot available in the current CPU's
2750 * loaded magazine, just put the object there and return.
2751 */
2752 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2753 ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2754 ccp->cc_free++;
2755 mutex_exit(&ccp->cc_lock);
2756 return;
2757 }
2758
2759 /*
2760 * The loaded magazine is full. If the previously loaded
2761 * magazine was empty, exchange them and try again.
2762 */
2763 if (ccp->cc_prounds == 0) {
2764 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2765 continue;
2766 }
2767
2768 /*
2769 * If the magazine layer is disabled, break out now.
2770 */
2771 if (ccp->cc_magsize == 0)
2772 break;
2773
2774 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2775 /*
2776 * We couldn't free our constructed object to the
2777 * magazine layer, so apply its destructor and free it
2778 * to the slab layer.
2779 */
2780 break;
2781 }
2782 }
2783 mutex_exit(&ccp->cc_lock);
2784 kmem_slab_free_constructed(cp, buf, B_TRUE);
2785 }
2786
2787 static void
2788 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2789 {
2790 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2791 int cache_flags = cp->cache_flags;
2792
2793 kmem_bufctl_t *next, *head;
2794 size_t nbufs;
2795
2796 /*
2797 * Completely allocate the newly created slab and put the pre-allocated
2798 * buffers in magazines. Any of the buffers that cannot be put in
2799 * magazines must be returned to the slab.
2800 */
2801 ASSERT(MUTEX_HELD(&cp->cache_lock));
2802 ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2803 ASSERT(cp->cache_constructor == NULL);
2804 ASSERT(sp->slab_cache == cp);
2805 ASSERT(sp->slab_refcnt == 1);
2806 ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2807 ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2808
2809 head = sp->slab_head;
2810 nbufs = (sp->slab_chunks - sp->slab_refcnt);
2811 sp->slab_head = NULL;
2812 sp->slab_refcnt += nbufs;
2813 cp->cache_bufslab -= nbufs;
2814 cp->cache_slab_alloc += nbufs;
2815 list_insert_head(&cp->cache_complete_slabs, sp);
2816 cp->cache_complete_slab_count++;
2817 mutex_exit(&cp->cache_lock);
2818 mutex_enter(&ccp->cc_lock);
2819
2820 while (head != NULL) {
2821 void *buf = KMEM_BUF(cp, head);
2822 /*
2823 * If there's a slot available in the current CPU's
2824 * loaded magazine, just put the object there and
2825 * continue.
2826 */
2827 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2828 ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2829 buf;
2830 ccp->cc_free++;
2831 nbufs--;
2832 head = head->bc_next;
2833 continue;
2834 }
2835
2836 /*
2837 * The loaded magazine is full. If the previously
2838 * loaded magazine was empty, exchange them and try
2839 * again.
2840 */
2841 if (ccp->cc_prounds == 0) {
2842 kmem_cpu_reload(ccp, ccp->cc_ploaded,
2843 ccp->cc_prounds);
2844 continue;
2845 }
2846
2847 /*
2848 * If the magazine layer is disabled, break out now.
2849 */
2850
2851 if (ccp->cc_magsize == 0) {
2852 break;
2853 }
2854
2855 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2856 break;
2857 }
2858 mutex_exit(&ccp->cc_lock);
2859 if (nbufs != 0) {
2860 ASSERT(head != NULL);
2861
2862 /*
2863 * If there was a failure, return remaining objects to
2864 * the slab
2865 */
2866 while (head != NULL) {
2867 ASSERT(nbufs != 0);
2868 next = head->bc_next;
2869 head->bc_next = NULL;
2870 kmem_slab_free(cp, KMEM_BUF(cp, head));
2871 head = next;
2872 nbufs--;
2873 }
2874 }
2875 ASSERT(head == NULL);
2876 ASSERT(nbufs == 0);
2877 mutex_enter(&cp->cache_lock);
2878 }
2879
2880 void *
2881 kmem_zalloc(size_t size, int kmflag)
2882 {
2883 size_t index;
2884 void *buf;
2885
2886 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2887 kmem_cache_t *cp = kmem_alloc_table[index];
2888 buf = kmem_cache_alloc(cp, kmflag);
2889 if (buf != NULL) {
2890 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2891 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2892 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2893 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2894
2895 if (cp->cache_flags & KMF_LITE) {
2896 KMEM_BUFTAG_LITE_ENTER(btp,
2897 kmem_lite_count, caller());
2898 }
2899 }
2900 bzero(buf, size);
2901 }
2902 } else {
2903 buf = kmem_alloc(size, kmflag);
2904 if (buf != NULL)
2905 bzero(buf, size);
2906 }
2907 return (buf);
2908 }
2909
2910 void *
2911 kmem_alloc(size_t size, int kmflag)
2912 {
2913 size_t index;
2914 kmem_cache_t *cp;
2915 void *buf;
2916
2917 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2918 cp = kmem_alloc_table[index];
2919 /* fall through to kmem_cache_alloc() */
2920
2921 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2922 kmem_big_alloc_table_max) {
2923 cp = kmem_big_alloc_table[index];
2924 /* fall through to kmem_cache_alloc() */
2925
2926 } else {
2927 if (size == 0) {
2928 if (kmflag != KM_SLEEP && !(kmflag & KM_PANIC))
2929 return (NULL);
2930
2931 /*
2932 * If this is a sleeping allocation or one that has
2933 * been specified to panic on allocation failure, we
2934 * consider it to be deprecated behavior to allocate
2935 * 0 bytes. If we have been configured to panic under
2936 * this condition, we panic; if to warn, we warn -- and
2937 * regardless, we log to the kmem_zerosized_log that
2938 * that this condition has occurred (which gives us
2939 * enough information to be able to debug it).
2940 */
2941 if (kmem_panic && kmem_panic_zerosized)
2942 panic("attempted to kmem_alloc() size of 0");
2943
2944 if (kmem_warn_zerosized) {
2945 cmn_err(CE_WARN, "kmem_alloc(): sleeping "
2946 "allocation with size of 0; "
2947 "see kmem_zerosized_log for details");
2948 }
2949
2950 kmem_log_event(kmem_zerosized_log, NULL, NULL, NULL);
2951
2952 return (NULL);
2953 }
2954
2955 buf = vmem_alloc(kmem_oversize_arena, size,
2956 kmflag & KM_VMFLAGS);
2957 if (buf == NULL)
2958 kmem_log_event(kmem_failure_log, NULL, NULL,
2959 (void *)size);
2960 else if (KMEM_DUMP(kmem_slab_cache)) {
2961 /* stats for dump intercept */
2962 kmem_dump_oversize_allocs++;
2963 if (size > kmem_dump_oversize_max)
2964 kmem_dump_oversize_max = size;
2965 }
2966 return (buf);
2967 }
2968
2969 buf = kmem_cache_alloc(cp, kmflag);
2970 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2971 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2972 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2973 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2974
2975 if (cp->cache_flags & KMF_LITE) {
2976 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2977 }
2978 }
2979 return (buf);
2980 }
2981
2982 void
2983 kmem_free(void *buf, size_t size)
2984 {
2985 size_t index;
2986 kmem_cache_t *cp;
2987
2988 if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2989 cp = kmem_alloc_table[index];
2990 /* fall through to kmem_cache_free() */
2991
2992 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2993 kmem_big_alloc_table_max) {
2994 cp = kmem_big_alloc_table[index];
2995 /* fall through to kmem_cache_free() */
2996
2997 } else {
2998 EQUIV(buf == NULL, size == 0);
2999 if (buf == NULL && size == 0)
3000 return;
3001 vmem_free(kmem_oversize_arena, buf, size);
3002 return;
3003 }
3004
3005 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
3006 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
3007 uint32_t *ip = (uint32_t *)btp;
3008 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3009 if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3010 kmem_error(KMERR_DUPFREE, cp, buf);
3011 return;
3012 }
3013 if (KMEM_SIZE_VALID(ip[1])) {
3014 ip[0] = KMEM_SIZE_ENCODE(size);
3015 kmem_error(KMERR_BADSIZE, cp, buf);
3016 } else {
3017 kmem_error(KMERR_REDZONE, cp, buf);
3018 }
3019 return;
3020 }
3021 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3022 kmem_error(KMERR_REDZONE, cp, buf);
3023 return;
3024 }
3025 btp->bt_redzone = KMEM_REDZONE_PATTERN;
3026 if (cp->cache_flags & KMF_LITE) {
3027 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3028 caller());
3029 }
3030 }
3031 kmem_cache_free(cp, buf);
3032 }
3033
3034 void *
3035 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3036 {
3037 size_t realsize = size + vmp->vm_quantum;
3038 void *addr;
3039
3040 /*
3041 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3042 * vm_quantum will cause integer wraparound. Check for this, and
3043 * blow off the firewall page in this case. Note that such a
3044 * giant allocation (the entire kernel address space) can never
3045 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3046 * or sleep forever (VM_SLEEP). Thus, there is no need for a
3047 * corresponding check in kmem_firewall_va_free().
3048 */
3049 if (realsize < size)
3050 realsize = size;
3051
3052 /*
3053 * While boot still owns resource management, make sure that this
3054 * redzone virtual address allocation is properly accounted for in
3055 * OBPs "virtual-memory" "available" lists because we're
3056 * effectively claiming them for a red zone. If we don't do this,
3057 * the available lists become too fragmented and too large for the
3058 * current boot/kernel memory list interface.
3059 */
3060 addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3061
3062 if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3063 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3064
3065 return (addr);
3066 }
3067
3068 void
3069 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3070 {
3071 ASSERT((kvseg.s_base == NULL ?
3072 va_to_pfn((char *)addr + size) :
3073 hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3074
3075 vmem_free(vmp, addr, size + vmp->vm_quantum);
3076 }
3077
3078 /*
3079 * Try to allocate at least `size' bytes of memory without sleeping or
3080 * panicking. Return actual allocated size in `asize'. If allocation failed,
3081 * try final allocation with sleep or panic allowed.
3082 */
3083 void *
3084 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3085 {
3086 void *p;
3087
3088 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3089 do {
3090 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3091 if (p != NULL)
3092 return (p);
3093 *asize += KMEM_ALIGN;
3094 } while (*asize <= PAGESIZE);
3095
3096 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3097 return (kmem_alloc(*asize, kmflag));
3098 }
3099
3100 /*
3101 * Reclaim all unused memory from a cache.
3102 */
3103 static void
3104 kmem_cache_reap(kmem_cache_t *cp)
3105 {
3106 ASSERT(taskq_member(kmem_taskq, curthread));
3107 cp->cache_reap++;
3108
3109 /*
3110 * Ask the cache's owner to free some memory if possible.
3111 * The idea is to handle things like the inode cache, which
3112 * typically sits on a bunch of memory that it doesn't truly
3113 * *need*. Reclaim policy is entirely up to the owner; this
3114 * callback is just an advisory plea for help.
3115 */
3116 if (cp->cache_reclaim != NULL) {
3117 long delta;
3118
3119 /*
3120 * Reclaimed memory should be reapable (not included in the
3121 * depot's working set).
3122 */
3123 delta = cp->cache_full.ml_total;
3124 cp->cache_reclaim(cp->cache_private);
3125 delta = cp->cache_full.ml_total - delta;
3126 if (delta > 0) {
3127 mutex_enter(&cp->cache_depot_lock);
3128 cp->cache_full.ml_reaplimit += delta;
3129 cp->cache_full.ml_min += delta;
3130 mutex_exit(&cp->cache_depot_lock);
3131 }
3132 }
3133
3134 kmem_depot_ws_reap(cp);
3135
3136 if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3137 kmem_cache_defrag(cp);
3138 }
3139 }
3140
3141 static void
3142 kmem_reap_timeout(void *flag_arg)
3143 {
3144 uint32_t *flag = (uint32_t *)flag_arg;
3145
3146 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3147 *flag = 0;
3148 }
3149
3150 static void
3151 kmem_reap_done(void *flag)
3152 {
3153 if (!callout_init_done) {
3154 /* can't schedule a timeout at this point */
3155 kmem_reap_timeout(flag);
3156 } else {
3157 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3158 }
3159 }
3160
3161 static void
3162 kmem_reap_start(void *flag)
3163 {
3164 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3165
3166 if (flag == &kmem_reaping) {
3167 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3168 /*
3169 * if we have segkp under heap, reap segkp cache.
3170 */
3171 if (segkp_fromheap)
3172 segkp_cache_free();
3173 }
3174 else
3175 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3176
3177 /*
3178 * We use taskq_dispatch() to schedule a timeout to clear
3179 * the flag so that kmem_reap() becomes self-throttling:
3180 * we won't reap again until the current reap completes *and*
3181 * at least kmem_reap_interval ticks have elapsed.
3182 */
3183 if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3184 kmem_reap_done(flag);
3185 }
3186
3187 static void
3188 kmem_reap_common(void *flag_arg)
3189 {
3190 uint32_t *flag = (uint32_t *)flag_arg;
3191
3192 if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3193 atomic_cas_32(flag, 0, 1) != 0)
3194 return;
3195
3196 /*
3197 * It may not be kosher to do memory allocation when a reap is called
3198 * (for example, if vmem_populate() is in the call chain). So we
3199 * start the reap going with a TQ_NOALLOC dispatch. If the dispatch
3200 * fails, we reset the flag, and the next reap will try again.
3201 */
3202 if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3203 *flag = 0;
3204 }
3205
3206 /*
3207 * Reclaim all unused memory from all caches. Called from the VM system
3208 * when memory gets tight.
3209 */
3210 void
3211 kmem_reap(void)
3212 {
3213 kmem_reap_common(&kmem_reaping);
3214 }
3215
3216 /*
3217 * Reclaim all unused memory from identifier arenas, called when a vmem
3218 * arena not back by memory is exhausted. Since reaping memory-backed caches
3219 * cannot help with identifier exhaustion, we avoid both a large amount of
3220 * work and unwanted side-effects from reclaim callbacks.
3221 */
3222 void
3223 kmem_reap_idspace(void)
3224 {
3225 kmem_reap_common(&kmem_reaping_idspace);
3226 }
3227
3228 /*
3229 * Purge all magazines from a cache and set its magazine limit to zero.
3230 * All calls are serialized by the kmem_taskq lock, except for the final
3231 * call from kmem_cache_destroy().
3232 */
3233 static void
3234 kmem_cache_magazine_purge(kmem_cache_t *cp)
3235 {
3236 kmem_cpu_cache_t *ccp;
3237 kmem_magazine_t *mp, *pmp;
3238 int rounds, prounds, cpu_seqid;
3239
3240 ASSERT(!list_link_active(&cp->cache_link) ||
3241 taskq_member(kmem_taskq, curthread));
3242 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3243
3244 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3245 ccp = &cp->cache_cpu[cpu_seqid];
3246
3247 mutex_enter(&ccp->cc_lock);
3248 mp = ccp->cc_loaded;
3249 pmp = ccp->cc_ploaded;
3250 rounds = ccp->cc_rounds;
3251 prounds = ccp->cc_prounds;
3252 ccp->cc_loaded = NULL;
3253 ccp->cc_ploaded = NULL;
3254 ccp->cc_rounds = -1;
3255 ccp->cc_prounds = -1;
3256 ccp->cc_magsize = 0;
3257 mutex_exit(&ccp->cc_lock);
3258
3259 if (mp)
3260 kmem_magazine_destroy(cp, mp, rounds);
3261 if (pmp)
3262 kmem_magazine_destroy(cp, pmp, prounds);
3263 }
3264
3265 kmem_depot_ws_zero(cp);
3266 kmem_depot_ws_reap(cp);
3267 }
3268
3269 /*
3270 * Enable per-cpu magazines on a cache.
3271 */
3272 static void
3273 kmem_cache_magazine_enable(kmem_cache_t *cp)
3274 {
3275 int cpu_seqid;
3276
3277 if (cp->cache_flags & KMF_NOMAGAZINE)
3278 return;
3279
3280 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3281 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3282 mutex_enter(&ccp->cc_lock);
3283 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3284 mutex_exit(&ccp->cc_lock);
3285 }
3286
3287 }
3288
3289 /*
3290 * Reap (almost) everything right now.
3291 */
3292 void
3293 kmem_cache_reap_now(kmem_cache_t *cp)
3294 {
3295 ASSERT(list_link_active(&cp->cache_link));
3296
3297 kmem_depot_ws_zero(cp);
3298
3299 (void) taskq_dispatch(kmem_taskq,
3300 (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3301 taskq_wait(kmem_taskq);
3302 }
3303
3304 /*
3305 * Recompute a cache's magazine size. The trade-off is that larger magazines
3306 * provide a higher transfer rate with the depot, while smaller magazines
3307 * reduce memory consumption. Magazine resizing is an expensive operation;
3308 * it should not be done frequently.
3309 *
3310 * Changes to the magazine size are serialized by the kmem_taskq lock.
3311 *
3312 * Note: at present this only grows the magazine size. It might be useful
3313 * to allow shrinkage too.
3314 */
3315 static void
3316 kmem_cache_magazine_resize(kmem_cache_t *cp)
3317 {
3318 kmem_magtype_t *mtp = cp->cache_magtype;
3319
3320 ASSERT(taskq_member(kmem_taskq, curthread));
3321
3322 if (cp->cache_chunksize < mtp->mt_maxbuf) {
3323 kmem_cache_magazine_purge(cp);
3324 mutex_enter(&cp->cache_depot_lock);
3325 cp->cache_magtype = ++mtp;
3326 cp->cache_depot_contention_prev =
3327 cp->cache_depot_contention + INT_MAX;
3328 mutex_exit(&cp->cache_depot_lock);
3329 kmem_cache_magazine_enable(cp);
3330 }
3331 }
3332
3333 /*
3334 * Rescale a cache's hash table, so that the table size is roughly the
3335 * cache size. We want the average lookup time to be extremely small.
3336 */
3337 static void
3338 kmem_hash_rescale(kmem_cache_t *cp)
3339 {
3340 kmem_bufctl_t **old_table, **new_table, *bcp;
3341 size_t old_size, new_size, h;
3342
3343 ASSERT(taskq_member(kmem_taskq, curthread));
3344
3345 new_size = MAX(KMEM_HASH_INITIAL,
3346 1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3347 old_size = cp->cache_hash_mask + 1;
3348
3349 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3350 return;
3351
3352 new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3353 VM_NOSLEEP);
3354 if (new_table == NULL)
3355 return;
3356 bzero(new_table, new_size * sizeof (void *));
3357
3358 mutex_enter(&cp->cache_lock);
3359
3360 old_size = cp->cache_hash_mask + 1;
3361 old_table = cp->cache_hash_table;
3362
3363 cp->cache_hash_mask = new_size - 1;
3364 cp->cache_hash_table = new_table;
3365 cp->cache_rescale++;
3366
3367 for (h = 0; h < old_size; h++) {
3368 bcp = old_table[h];
3369 while (bcp != NULL) {
3370 void *addr = bcp->bc_addr;
3371 kmem_bufctl_t *next_bcp = bcp->bc_next;
3372 kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3373 bcp->bc_next = *hash_bucket;
3374 *hash_bucket = bcp;
3375 bcp = next_bcp;
3376 }
3377 }
3378
3379 mutex_exit(&cp->cache_lock);
3380
3381 vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3382 }
3383
3384 /*
3385 * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3386 * update, magazine resizing, and slab consolidation.
3387 */
3388 static void
3389 kmem_cache_update(kmem_cache_t *cp)
3390 {
3391 int need_hash_rescale = 0;
3392 int need_magazine_resize = 0;
3393
3394 ASSERT(MUTEX_HELD(&kmem_cache_lock));
3395
3396 /*
3397 * If the cache has become much larger or smaller than its hash table,
3398 * fire off a request to rescale the hash table.
3399 */
3400 mutex_enter(&cp->cache_lock);
3401
3402 if ((cp->cache_flags & KMF_HASH) &&
3403 (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3404 (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3405 cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3406 need_hash_rescale = 1;
3407
3408 mutex_exit(&cp->cache_lock);
3409
3410 /*
3411 * Update the depot working set statistics.
3412 */
3413 kmem_depot_ws_update(cp);
3414
3415 /*
3416 * If there's a lot of contention in the depot,
3417 * increase the magazine size.
3418 */
3419 mutex_enter(&cp->cache_depot_lock);
3420
3421 if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3422 (int)(cp->cache_depot_contention -
3423 cp->cache_depot_contention_prev) > kmem_depot_contention)
3424 need_magazine_resize = 1;
3425
3426 cp->cache_depot_contention_prev = cp->cache_depot_contention;
3427
3428 mutex_exit(&cp->cache_depot_lock);
3429
3430 if (need_hash_rescale)
3431 (void) taskq_dispatch(kmem_taskq,
3432 (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3433
3434 if (need_magazine_resize)
3435 (void) taskq_dispatch(kmem_taskq,
3436 (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3437
3438 if (cp->cache_defrag != NULL)
3439 (void) taskq_dispatch(kmem_taskq,
3440 (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3441 }
3442
3443 static void kmem_update(void *);
3444
3445 static void
3446 kmem_update_timeout(void *dummy)
3447 {
3448 (void) timeout(kmem_update, dummy, kmem_reap_interval);
3449 }
3450
3451 static void
3452 kmem_update(void *dummy)
3453 {
3454 kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3455
3456 /*
3457 * We use taskq_dispatch() to reschedule the timeout so that
3458 * kmem_update() becomes self-throttling: it won't schedule
3459 * new tasks until all previous tasks have completed.
3460 */
3461 if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3462 kmem_update_timeout(NULL);
3463 }
3464
3465 static int
3466 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3467 {
3468 struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3469 kmem_cache_t *cp = ksp->ks_private;
3470 uint64_t cpu_buf_avail;
3471 uint64_t buf_avail = 0;
3472 int cpu_seqid;
3473 long reap;
3474
3475 ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3476
3477 if (rw == KSTAT_WRITE)
3478 return (EACCES);
3479
3480 mutex_enter(&cp->cache_lock);
3481
3482 kmcp->kmc_alloc_fail.value.ui64 = cp->cache_alloc_fail;
3483 kmcp->kmc_alloc.value.ui64 = cp->cache_slab_alloc;
3484 kmcp->kmc_free.value.ui64 = cp->cache_slab_free;
3485 kmcp->kmc_slab_alloc.value.ui64 = cp->cache_slab_alloc;
3486 kmcp->kmc_slab_free.value.ui64 = cp->cache_slab_free;
3487
3488 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3489 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3490
3491 mutex_enter(&ccp->cc_lock);
3492
3493 cpu_buf_avail = 0;
3494 if (ccp->cc_rounds > 0)
3495 cpu_buf_avail += ccp->cc_rounds;
3496 if (ccp->cc_prounds > 0)
3497 cpu_buf_avail += ccp->cc_prounds;
3498
3499 kmcp->kmc_alloc.value.ui64 += ccp->cc_alloc;
3500 kmcp->kmc_free.value.ui64 += ccp->cc_free;
3501 buf_avail += cpu_buf_avail;
3502
3503 mutex_exit(&ccp->cc_lock);
3504 }
3505
3506 mutex_enter(&cp->cache_depot_lock);
3507
3508 kmcp->kmc_depot_alloc.value.ui64 = cp->cache_full.ml_alloc;
3509 kmcp->kmc_depot_free.value.ui64 = cp->cache_empty.ml_alloc;
3510 kmcp->kmc_depot_contention.value.ui64 = cp->cache_depot_contention;
3511 kmcp->kmc_full_magazines.value.ui64 = cp->cache_full.ml_total;
3512 kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3513 kmcp->kmc_magazine_size.value.ui64 =
3514 (cp->cache_flags & KMF_NOMAGAZINE) ?
3515 0 : cp->cache_magtype->mt_magsize;
3516
3517 kmcp->kmc_alloc.value.ui64 += cp->cache_full.ml_alloc;
3518 kmcp->kmc_free.value.ui64 += cp->cache_empty.ml_alloc;
3519 buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3520
3521 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3522 reap = MIN(reap, cp->cache_full.ml_total);
3523
3524 mutex_exit(&cp->cache_depot_lock);
3525
3526 kmcp->kmc_buf_size.value.ui64 = cp->cache_bufsize;
3527 kmcp->kmc_align.value.ui64 = cp->cache_align;
3528 kmcp->kmc_chunk_size.value.ui64 = cp->cache_chunksize;
3529 kmcp->kmc_slab_size.value.ui64 = cp->cache_slabsize;
3530 kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3531 buf_avail += cp->cache_bufslab;
3532 kmcp->kmc_buf_avail.value.ui64 = buf_avail;
3533 kmcp->kmc_buf_inuse.value.ui64 = cp->cache_buftotal - buf_avail;
3534 kmcp->kmc_buf_total.value.ui64 = cp->cache_buftotal;
3535 kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3536 kmcp->kmc_slab_create.value.ui64 = cp->cache_slab_create;
3537 kmcp->kmc_slab_destroy.value.ui64 = cp->cache_slab_destroy;
3538 kmcp->kmc_hash_size.value.ui64 = (cp->cache_flags & KMF_HASH) ?
3539 cp->cache_hash_mask + 1 : 0;
3540 kmcp->kmc_hash_lookup_depth.value.ui64 = cp->cache_lookup_depth;
3541 kmcp->kmc_hash_rescale.value.ui64 = cp->cache_rescale;
3542 kmcp->kmc_vmem_source.value.ui64 = cp->cache_arena->vm_id;
3543 kmcp->kmc_reap.value.ui64 = cp->cache_reap;
3544
3545 if (cp->cache_defrag == NULL) {
3546 kmcp->kmc_move_callbacks.value.ui64 = 0;
3547 kmcp->kmc_move_yes.value.ui64 = 0;
3548 kmcp->kmc_move_no.value.ui64 = 0;
3549 kmcp->kmc_move_later.value.ui64 = 0;
3550 kmcp->kmc_move_dont_need.value.ui64 = 0;
3551 kmcp->kmc_move_dont_know.value.ui64 = 0;
3552 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3553 kmcp->kmc_move_slabs_freed.value.ui64 = 0;
3554 kmcp->kmc_defrag.value.ui64 = 0;
3555 kmcp->kmc_scan.value.ui64 = 0;
3556 kmcp->kmc_move_reclaimable.value.ui64 = 0;
3557 } else {
3558 int64_t reclaimable;
3559
3560 kmem_defrag_t *kd = cp->cache_defrag;
3561 kmcp->kmc_move_callbacks.value.ui64 = kd->kmd_callbacks;
3562 kmcp->kmc_move_yes.value.ui64 = kd->kmd_yes;
3563 kmcp->kmc_move_no.value.ui64 = kd->kmd_no;
3564 kmcp->kmc_move_later.value.ui64 = kd->kmd_later;
3565 kmcp->kmc_move_dont_need.value.ui64 = kd->kmd_dont_need;
3566 kmcp->kmc_move_dont_know.value.ui64 = kd->kmd_dont_know;
3567 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3568 kmcp->kmc_move_slabs_freed.value.ui64 = kd->kmd_slabs_freed;
3569 kmcp->kmc_defrag.value.ui64 = kd->kmd_defrags;
3570 kmcp->kmc_scan.value.ui64 = kd->kmd_scans;
3571
3572 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3573 reclaimable = MAX(reclaimable, 0);
3574 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3575 kmcp->kmc_move_reclaimable.value.ui64 = reclaimable;
3576 }
3577
3578 mutex_exit(&cp->cache_lock);
3579 return (0);
3580 }
3581
3582 /*
3583 * Return a named statistic about a particular cache.
3584 * This shouldn't be called very often, so it's currently designed for
3585 * simplicity (leverages existing kstat support) rather than efficiency.
3586 */
3587 uint64_t
3588 kmem_cache_stat(kmem_cache_t *cp, char *name)
3589 {
3590 int i;
3591 kstat_t *ksp = cp->cache_kstat;
3592 kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3593 uint64_t value = 0;
3594
3595 if (ksp != NULL) {
3596 mutex_enter(&kmem_cache_kstat_lock);
3597 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3598 for (i = 0; i < ksp->ks_ndata; i++) {
3599 if (strcmp(knp[i].name, name) == 0) {
3600 value = knp[i].value.ui64;
3601 break;
3602 }
3603 }
3604 mutex_exit(&kmem_cache_kstat_lock);
3605 }
3606 return (value);
3607 }
3608
3609 /*
3610 * Return an estimate of currently available kernel heap memory.
3611 * On 32-bit systems, physical memory may exceed virtual memory,
3612 * we just truncate the result at 1GB.
3613 */
3614 size_t
3615 kmem_avail(void)
3616 {
3617 spgcnt_t rmem = availrmem - tune.t_minarmem;
3618 spgcnt_t fmem = freemem - minfree;
3619
3620 return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3621 1 << (30 - PAGESHIFT))));
3622 }
3623
3624 /*
3625 * Return the maximum amount of memory that is (in theory) allocatable
3626 * from the heap. This may be used as an estimate only since there
3627 * is no guarentee this space will still be available when an allocation
3628 * request is made, nor that the space may be allocated in one big request
3629 * due to kernel heap fragmentation.
3630 */
3631 size_t
3632 kmem_maxavail(void)
3633 {
3634 spgcnt_t pmem = availrmem - tune.t_minarmem;
3635 spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3636
3637 return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3638 }
3639
3640 /*
3641 * Indicate whether memory-intensive kmem debugging is enabled.
3642 */
3643 int
3644 kmem_debugging(void)
3645 {
3646 return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3647 }
3648
3649 /* binning function, sorts finely at the two extremes */
3650 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift) \
3651 ((((sp)->slab_refcnt <= (binshift)) || \
3652 (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift))) \
3653 ? -(sp)->slab_refcnt \
3654 : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3655
3656 /*
3657 * Minimizing the number of partial slabs on the freelist minimizes
3658 * fragmentation (the ratio of unused buffers held by the slab layer). There are
3659 * two ways to get a slab off of the freelist: 1) free all the buffers on the
3660 * slab, and 2) allocate all the buffers on the slab. It follows that we want
3661 * the most-used slabs at the front of the list where they have the best chance
3662 * of being completely allocated, and the least-used slabs at a safe distance
3663 * from the front to improve the odds that the few remaining buffers will all be
3664 * freed before another allocation can tie up the slab. For that reason a slab
3665 * with a higher slab_refcnt sorts less than than a slab with a lower
3666 * slab_refcnt.
3667 *
3668 * However, if a slab has at least one buffer that is deemed unfreeable, we
3669 * would rather have that slab at the front of the list regardless of
3670 * slab_refcnt, since even one unfreeable buffer makes the entire slab
3671 * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3672 * callback, the slab is marked unfreeable for as long as it remains on the
3673 * freelist.
3674 */
3675 static int
3676 kmem_partial_slab_cmp(const void *p0, const void *p1)
3677 {
3678 const kmem_cache_t *cp;
3679 const kmem_slab_t *s0 = p0;
3680 const kmem_slab_t *s1 = p1;
3681 int w0, w1;
3682 size_t binshift;
3683
3684 ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3685 ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3686 ASSERT(s0->slab_cache == s1->slab_cache);
3687 cp = s1->slab_cache;
3688 ASSERT(MUTEX_HELD(&cp->cache_lock));
3689 binshift = cp->cache_partial_binshift;
3690
3691 /* weight of first slab */
3692 w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3693 if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3694 w0 -= cp->cache_maxchunks;
3695 }
3696
3697 /* weight of second slab */
3698 w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3699 if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3700 w1 -= cp->cache_maxchunks;
3701 }
3702
3703 if (w0 < w1)
3704 return (-1);
3705 if (w0 > w1)
3706 return (1);
3707
3708 /* compare pointer values */
3709 if ((uintptr_t)s0 < (uintptr_t)s1)
3710 return (-1);
3711 if ((uintptr_t)s0 > (uintptr_t)s1)
3712 return (1);
3713
3714 return (0);
3715 }
3716
3717 /*
3718 * It must be valid to call the destructor (if any) on a newly created object.
3719 * That is, the constructor (if any) must leave the object in a valid state for
3720 * the destructor.
3721 */
3722 kmem_cache_t *
3723 kmem_cache_create(
3724 char *name, /* descriptive name for this cache */
3725 size_t bufsize, /* size of the objects it manages */
3726 size_t align, /* required object alignment */
3727 int (*constructor)(void *, void *, int), /* object constructor */
3728 void (*destructor)(void *, void *), /* object destructor */
3729 void (*reclaim)(void *), /* memory reclaim callback */
3730 void *private, /* pass-thru arg for constr/destr/reclaim */
3731 vmem_t *vmp, /* vmem source for slab allocation */
3732 int cflags) /* cache creation flags */
3733 {
3734 int cpu_seqid;
3735 size_t chunksize;
3736 kmem_cache_t *cp;
3737 kmem_magtype_t *mtp;
3738 size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3739
3740 #ifdef DEBUG
3741 /*
3742 * Cache names should conform to the rules for valid C identifiers
3743 */
3744 if (!strident_valid(name)) {
3745 cmn_err(CE_CONT,
3746 "kmem_cache_create: '%s' is an invalid cache name\n"
3747 "cache names must conform to the rules for "
3748 "C identifiers\n", name);
3749 }
3750 #endif /* DEBUG */
3751
3752 if (vmp == NULL)
3753 vmp = kmem_default_arena;
3754
3755 /*
3756 * If this kmem cache has an identifier vmem arena as its source, mark
3757 * it such to allow kmem_reap_idspace().
3758 */
3759 ASSERT(!(cflags & KMC_IDENTIFIER)); /* consumer should not set this */
3760 if (vmp->vm_cflags & VMC_IDENTIFIER)
3761 cflags |= KMC_IDENTIFIER;
3762
3763 /*
3764 * Get a kmem_cache structure. We arrange that cp->cache_cpu[]
3765 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3766 * false sharing of per-CPU data.
3767 */
3768 cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3769 P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3770 bzero(cp, csize);
3771 list_link_init(&cp->cache_link);
3772
3773 if (align == 0)
3774 align = KMEM_ALIGN;
3775
3776 /*
3777 * If we're not at least KMEM_ALIGN aligned, we can't use free
3778 * memory to hold bufctl information (because we can't safely
3779 * perform word loads and stores on it).
3780 */
3781 if (align < KMEM_ALIGN)
3782 cflags |= KMC_NOTOUCH;
3783
3784 if (!ISP2(align) || align > vmp->vm_quantum)
3785 panic("kmem_cache_create: bad alignment %lu", align);
3786
3787 mutex_enter(&kmem_flags_lock);
3788 if (kmem_flags & KMF_RANDOMIZE)
3789 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3790 KMF_RANDOMIZE;
3791 cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3792 mutex_exit(&kmem_flags_lock);
3793
3794 /*
3795 * Make sure all the various flags are reasonable.
3796 */
3797 ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3798
3799 if (cp->cache_flags & KMF_LITE) {
3800 if (bufsize >= kmem_lite_minsize &&
3801 align <= kmem_lite_maxalign &&
3802 P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3803 cp->cache_flags |= KMF_BUFTAG;
3804 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3805 } else {
3806 cp->cache_flags &= ~KMF_DEBUG;
3807 }
3808 }
3809
3810 if (cp->cache_flags & KMF_DEADBEEF)
3811 cp->cache_flags |= KMF_REDZONE;
3812
3813 if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3814 cp->cache_flags |= KMF_NOMAGAZINE;
3815
3816 if (cflags & KMC_NODEBUG)
3817 cp->cache_flags &= ~KMF_DEBUG;
3818
3819 if (cflags & KMC_NOTOUCH)
3820 cp->cache_flags &= ~KMF_TOUCH;
3821
3822 if (cflags & KMC_PREFILL)
3823 cp->cache_flags |= KMF_PREFILL;
3824
3825 if (cflags & KMC_NOHASH)
3826 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3827
3828 if (cflags & KMC_NOMAGAZINE)
3829 cp->cache_flags |= KMF_NOMAGAZINE;
3830
3831 if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3832 cp->cache_flags |= KMF_REDZONE;
3833
3834 if (!(cp->cache_flags & KMF_AUDIT))
3835 cp->cache_flags &= ~KMF_CONTENTS;
3836
3837 if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3838 !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3839 cp->cache_flags |= KMF_FIREWALL;
3840
3841 if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3842 cp->cache_flags &= ~KMF_FIREWALL;
3843
3844 if (cp->cache_flags & KMF_FIREWALL) {
3845 cp->cache_flags &= ~KMF_BUFTAG;
3846 cp->cache_flags |= KMF_NOMAGAZINE;
3847 ASSERT(vmp == kmem_default_arena);
3848 vmp = kmem_firewall_arena;
3849 }
3850
3851 /*
3852 * Set cache properties.
3853 */
3854 (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3855 strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3856 cp->cache_bufsize = bufsize;
3857 cp->cache_align = align;
3858 cp->cache_constructor = constructor;
3859 cp->cache_destructor = destructor;
3860 cp->cache_reclaim = reclaim;
3861 cp->cache_private = private;
3862 cp->cache_arena = vmp;
3863 cp->cache_cflags = cflags;
3864
3865 /*
3866 * Determine the chunk size.
3867 */
3868 chunksize = bufsize;
3869
3870 if (align >= KMEM_ALIGN) {
3871 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3872 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3873 }
3874
3875 if (cp->cache_flags & KMF_BUFTAG) {
3876 cp->cache_bufctl = chunksize;
3877 cp->cache_buftag = chunksize;
3878 if (cp->cache_flags & KMF_LITE)
3879 chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3880 else
3881 chunksize += sizeof (kmem_buftag_t);
3882 }
3883
3884 if (cp->cache_flags & KMF_DEADBEEF) {
3885 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3886 if (cp->cache_flags & KMF_LITE)
3887 cp->cache_verify = sizeof (uint64_t);
3888 }
3889
3890 cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3891
3892 cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3893
3894 /*
3895 * Now that we know the chunk size, determine the optimal slab size.
3896 */
3897 if (vmp == kmem_firewall_arena) {
3898 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3899 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3900 cp->cache_maxcolor = cp->cache_mincolor;
3901 cp->cache_flags |= KMF_HASH;
3902 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3903 } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3904 !(cp->cache_flags & KMF_AUDIT) &&
3905 chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3906 cp->cache_slabsize = vmp->vm_quantum;
3907 cp->cache_mincolor = 0;
3908 cp->cache_maxcolor =
3909 (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3910 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3911 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3912 } else {
3913 size_t chunks, bestfit, waste, slabsize;
3914 size_t minwaste = LONG_MAX;
3915
3916 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3917 slabsize = P2ROUNDUP(chunksize * chunks,
3918 vmp->vm_quantum);
3919 chunks = slabsize / chunksize;
3920 waste = (slabsize % chunksize) / chunks;
3921 if (waste < minwaste) {
3922 minwaste = waste;
3923 bestfit = slabsize;
3924 }
3925 }
3926 if (cflags & KMC_QCACHE)
3927 bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3928 cp->cache_slabsize = bestfit;
3929 cp->cache_mincolor = 0;
3930 cp->cache_maxcolor = bestfit % chunksize;
3931 cp->cache_flags |= KMF_HASH;
3932 }
3933
3934 cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3935 cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3936
3937 /*
3938 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3939 * there is a constructor avoids some tricky issues with debug setup
3940 * that may be revisited later. We cannot allow prefill in a
3941 * metadata cache because of potential recursion.
3942 */
3943 if (vmp == kmem_msb_arena ||
3944 cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3945 cp->cache_constructor != NULL)
3946 cp->cache_flags &= ~KMF_PREFILL;
3947
3948 if (cp->cache_flags & KMF_HASH) {
3949 ASSERT(!(cflags & KMC_NOHASH));
3950 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3951 kmem_bufctl_audit_cache : kmem_bufctl_cache;
3952 }
3953
3954 if (cp->cache_maxcolor >= vmp->vm_quantum)
3955 cp->cache_maxcolor = vmp->vm_quantum - 1;
3956
3957 cp->cache_color = cp->cache_mincolor;
3958
3959 /*
3960 * Initialize the rest of the slab layer.
3961 */
3962 mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3963
3964 avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3965 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3966 /* LINTED: E_TRUE_LOGICAL_EXPR */
3967 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3968 /* reuse partial slab AVL linkage for complete slab list linkage */
3969 list_create(&cp->cache_complete_slabs,
3970 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3971
3972 if (cp->cache_flags & KMF_HASH) {
3973 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3974 KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3975 bzero(cp->cache_hash_table,
3976 KMEM_HASH_INITIAL * sizeof (void *));
3977 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3978 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3979 }
3980
3981 /*
3982 * Initialize the depot.
3983 */
3984 mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3985
3986 for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3987 continue;
3988
3989 cp->cache_magtype = mtp;
3990
3991 /*
3992 * Initialize the CPU layer.
3993 */
3994 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3995 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3996 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3997 ccp->cc_flags = cp->cache_flags;
3998 ccp->cc_rounds = -1;
3999 ccp->cc_prounds = -1;
4000 }
4001
4002 /*
4003 * Create the cache's kstats.
4004 */
4005 if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
4006 "kmem_cache", KSTAT_TYPE_NAMED,
4007 sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
4008 KSTAT_FLAG_VIRTUAL)) != NULL) {
4009 cp->cache_kstat->ks_data = &kmem_cache_kstat;
4010 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
4011 cp->cache_kstat->ks_private = cp;
4012 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
4013 kstat_install(cp->cache_kstat);
4014 }
4015
4016 /*
4017 * Add the cache to the global list. This makes it visible
4018 * to kmem_update(), so the cache must be ready for business.
4019 */
4020 mutex_enter(&kmem_cache_lock);
4021 list_insert_tail(&kmem_caches, cp);
4022 mutex_exit(&kmem_cache_lock);
4023
4024 if (kmem_ready)
4025 kmem_cache_magazine_enable(cp);
4026
4027 return (cp);
4028 }
4029
4030 static int
4031 kmem_move_cmp(const void *buf, const void *p)
4032 {
4033 const kmem_move_t *kmm = p;
4034 uintptr_t v1 = (uintptr_t)buf;
4035 uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
4036 return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
4037 }
4038
4039 static void
4040 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4041 {
4042 kmd->kmd_reclaim_numer = 1;
4043 }
4044
4045 /*
4046 * Initially, when choosing candidate slabs for buffers to move, we want to be
4047 * very selective and take only slabs that are less than
4048 * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4049 * slabs, then we raise the allocation ceiling incrementally. The reclaim
4050 * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4051 * longer fragmented.
4052 */
4053 static void
4054 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4055 {
4056 if (direction > 0) {
4057 /* make it easier to find a candidate slab */
4058 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4059 kmd->kmd_reclaim_numer++;
4060 }
4061 } else {
4062 /* be more selective */
4063 if (kmd->kmd_reclaim_numer > 1) {
4064 kmd->kmd_reclaim_numer--;
4065 }
4066 }
4067 }
4068
4069 void
4070 kmem_cache_set_move(kmem_cache_t *cp,
4071 kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4072 {
4073 kmem_defrag_t *defrag;
4074
4075 ASSERT(move != NULL);
4076 /*
4077 * The consolidator does not support NOTOUCH caches because kmem cannot
4078 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4079 * a low order bit usable by clients to distinguish uninitialized memory
4080 * from known objects (see kmem_slab_create).
4081 */
4082 ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4083 ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4084
4085 /*
4086 * We should not be holding anyone's cache lock when calling
4087 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4088 * lock.
4089 */
4090 defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4091
4092 mutex_enter(&cp->cache_lock);
4093
4094 if (KMEM_IS_MOVABLE(cp)) {
4095 if (cp->cache_move == NULL) {
4096 ASSERT(cp->cache_slab_alloc == 0);
4097
4098 cp->cache_defrag = defrag;
4099 defrag = NULL; /* nothing to free */
4100 bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4101 avl_create(&cp->cache_defrag->kmd_moves_pending,
4102 kmem_move_cmp, sizeof (kmem_move_t),
4103 offsetof(kmem_move_t, kmm_entry));
4104 /* LINTED: E_TRUE_LOGICAL_EXPR */
4105 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4106 /* reuse the slab's AVL linkage for deadlist linkage */
4107 list_create(&cp->cache_defrag->kmd_deadlist,
4108 sizeof (kmem_slab_t),
4109 offsetof(kmem_slab_t, slab_link));
4110 kmem_reset_reclaim_threshold(cp->cache_defrag);
4111 }
4112 cp->cache_move = move;
4113 }
4114
4115 mutex_exit(&cp->cache_lock);
4116
4117 if (defrag != NULL) {
4118 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4119 }
4120 }
4121
4122 void
4123 kmem_cache_destroy(kmem_cache_t *cp)
4124 {
4125 int cpu_seqid;
4126
4127 /*
4128 * Remove the cache from the global cache list so that no one else
4129 * can schedule tasks on its behalf, wait for any pending tasks to
4130 * complete, purge the cache, and then destroy it.
4131 */
4132 mutex_enter(&kmem_cache_lock);
4133 list_remove(&kmem_caches, cp);
4134 mutex_exit(&kmem_cache_lock);
4135
4136 if (kmem_taskq != NULL)
4137 taskq_wait(kmem_taskq);
4138
4139 if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4140 taskq_wait(kmem_move_taskq);
4141
4142 kmem_cache_magazine_purge(cp);
4143
4144 mutex_enter(&cp->cache_lock);
4145 if (cp->cache_buftotal != 0)
4146 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4147 cp->cache_name, (void *)cp);
4148 if (cp->cache_defrag != NULL) {
4149 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4150 list_destroy(&cp->cache_defrag->kmd_deadlist);
4151 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4152 cp->cache_defrag = NULL;
4153 }
4154 /*
4155 * The cache is now dead. There should be no further activity. We
4156 * enforce this by setting land mines in the constructor, destructor,
4157 * reclaim, and move routines that induce a kernel text fault if
4158 * invoked.
4159 */
4160 cp->cache_constructor = (int (*)(void *, void *, int))1;
4161 cp->cache_destructor = (void (*)(void *, void *))2;
4162 cp->cache_reclaim = (void (*)(void *))3;
4163 cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4164 mutex_exit(&cp->cache_lock);
4165
4166 kstat_delete(cp->cache_kstat);
4167
4168 if (cp->cache_hash_table != NULL)
4169 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4170 (cp->cache_hash_mask + 1) * sizeof (void *));
4171
4172 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4173 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4174
4175 mutex_destroy(&cp->cache_depot_lock);
4176 mutex_destroy(&cp->cache_lock);
4177
4178 vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4179 }
4180
4181 /*ARGSUSED*/
4182 static int
4183 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4184 {
4185 ASSERT(MUTEX_HELD(&cpu_lock));
4186 if (what == CPU_UNCONFIG) {
4187 kmem_cache_applyall(kmem_cache_magazine_purge,
4188 kmem_taskq, TQ_SLEEP);
4189 kmem_cache_applyall(kmem_cache_magazine_enable,
4190 kmem_taskq, TQ_SLEEP);
4191 }
4192 return (0);
4193 }
4194
4195 static void
4196 kmem_alloc_caches_create(const int *array, size_t count,
4197 kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4198 {
4199 char name[KMEM_CACHE_NAMELEN + 1];
4200 size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4201 size_t size = table_unit;
4202 int i;
4203
4204 for (i = 0; i < count; i++) {
4205 size_t cache_size = array[i];
4206 size_t align = KMEM_ALIGN;
4207 kmem_cache_t *cp;
4208
4209 /* if the table has an entry for maxbuf, we're done */
4210 if (size > maxbuf)
4211 break;
4212
4213 /* cache size must be a multiple of the table unit */
4214 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4215
4216 /*
4217 * If they allocate a multiple of the coherency granularity,
4218 * they get a coherency-granularity-aligned address.
4219 */
4220 if (IS_P2ALIGNED(cache_size, 64))
4221 align = 64;
4222 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4223 align = PAGESIZE;
4224 (void) snprintf(name, sizeof (name),
4225 "kmem_alloc_%lu", cache_size);
4226 cp = kmem_cache_create(name, cache_size, align,
4227 NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4228
4229 while (size <= cache_size) {
4230 alloc_table[(size - 1) >> shift] = cp;
4231 size += table_unit;
4232 }
4233 }
4234
4235 ASSERT(size > maxbuf); /* i.e. maxbuf <= max(cache_size) */
4236 }
4237
4238 static void
4239 kmem_cache_init(int pass, int use_large_pages)
4240 {
4241 int i;
4242 size_t maxbuf;
4243 kmem_magtype_t *mtp;
4244
4245 for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4246 char name[KMEM_CACHE_NAMELEN + 1];
4247
4248 mtp = &kmem_magtype[i];
4249 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4250 mtp->mt_cache = kmem_cache_create(name,
4251 (mtp->mt_magsize + 1) * sizeof (void *),
4252 mtp->mt_align, NULL, NULL, NULL, NULL,
4253 kmem_msb_arena, KMC_NOHASH);
4254 }
4255
4256 kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4257 sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4258 kmem_msb_arena, KMC_NOHASH);
4259
4260 kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4261 sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4262 kmem_msb_arena, KMC_NOHASH);
4263
4264 kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4265 sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4266 kmem_msb_arena, KMC_NOHASH);
4267
4268 if (pass == 2) {
4269 kmem_va_arena = vmem_create("kmem_va",
4270 NULL, 0, PAGESIZE,
4271 vmem_alloc, vmem_free, heap_arena,
4272 8 * PAGESIZE, VM_SLEEP);
4273
4274 if (use_large_pages) {
4275 kmem_default_arena = vmem_xcreate("kmem_default",
4276 NULL, 0, PAGESIZE,
4277 segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4278 0, VMC_DUMPSAFE | VM_SLEEP);
4279 } else {
4280 kmem_default_arena = vmem_create("kmem_default",
4281 NULL, 0, PAGESIZE,
4282 segkmem_alloc, segkmem_free, kmem_va_arena,
4283 0, VMC_DUMPSAFE | VM_SLEEP);
4284 }
4285
4286 /* Figure out what our maximum cache size is */
4287 maxbuf = kmem_max_cached;
4288 if (maxbuf <= KMEM_MAXBUF) {
4289 maxbuf = 0;
4290 kmem_max_cached = KMEM_MAXBUF;
4291 } else {
4292 size_t size = 0;
4293 size_t max =
4294 sizeof (kmem_big_alloc_sizes) / sizeof (int);
4295 /*
4296 * Round maxbuf up to an existing cache size. If maxbuf
4297 * is larger than the largest cache, we truncate it to
4298 * the largest cache's size.
4299 */
4300 for (i = 0; i < max; i++) {
4301 size = kmem_big_alloc_sizes[i];
4302 if (maxbuf <= size)
4303 break;
4304 }
4305 kmem_max_cached = maxbuf = size;
4306 }
4307
4308 /*
4309 * The big alloc table may not be completely overwritten, so
4310 * we clear out any stale cache pointers from the first pass.
4311 */
4312 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4313 } else {
4314 /*
4315 * During the first pass, the kmem_alloc_* caches
4316 * are treated as metadata.
4317 */
4318 kmem_default_arena = kmem_msb_arena;
4319 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4320 }
4321
4322 /*
4323 * Set up the default caches to back kmem_alloc()
4324 */
4325 kmem_alloc_caches_create(
4326 kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4327 kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4328
4329 kmem_alloc_caches_create(
4330 kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4331 kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4332
4333 kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4334 }
4335
4336 void
4337 kmem_init(void)
4338 {
4339 kmem_cache_t *cp;
4340 int old_kmem_flags = kmem_flags;
4341 int use_large_pages = 0;
4342 size_t maxverify, minfirewall;
4343
4344 kstat_init();
4345
4346 /*
4347 * Don't do firewalled allocations if the heap is less than 1TB
4348 * (i.e. on a 32-bit kernel)
4349 * The resulting VM_NEXTFIT allocations would create too much
4350 * fragmentation in a small heap.
4351 */
4352 #if defined(_LP64)
4353 maxverify = minfirewall = PAGESIZE / 2;
4354 #else
4355 maxverify = minfirewall = ULONG_MAX;
4356 #endif
4357
4358 /* LINTED */
4359 ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4360
4361 list_create(&kmem_caches, sizeof (kmem_cache_t),
4362 offsetof(kmem_cache_t, cache_link));
4363
4364 kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4365 vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4366 VM_SLEEP | VMC_NO_QCACHE);
4367
4368 kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4369 PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4370 VMC_DUMPSAFE | VM_SLEEP);
4371
4372 kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4373 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4374
4375 kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4376 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4377
4378 kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4379 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4380
4381 kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4382 NULL, 0, PAGESIZE,
4383 kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4384 0, VM_SLEEP);
4385
4386 kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4387 segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4388 VMC_DUMPSAFE | VM_SLEEP);
4389
4390 /* temporary oversize arena for mod_read_system_file */
4391 kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4392 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4393
4394 kmem_reap_interval = 15 * hz;
4395
4396 /*
4397 * Read /etc/system. This is a chicken-and-egg problem because
4398 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4399 * needs to use the allocator. The simplest solution is to create
4400 * all the standard kmem caches, read /etc/system, destroy all the
4401 * caches we just created, and then create them all again in light
4402 * of the (possibly) new kmem_flags and other kmem tunables.
4403 */
4404 kmem_cache_init(1, 0);
4405
4406 mod_read_system_file(boothowto & RB_ASKNAME);
4407
4408 while ((cp = list_tail(&kmem_caches)) != NULL)
4409 kmem_cache_destroy(cp);
4410
4411 vmem_destroy(kmem_oversize_arena);
4412
4413 if (old_kmem_flags & KMF_STICKY)
4414 kmem_flags = old_kmem_flags;
4415
4416 if (!(kmem_flags & KMF_AUDIT))
4417 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4418
4419 if (kmem_maxverify == 0)
4420 kmem_maxverify = maxverify;
4421
4422 if (kmem_minfirewall == 0)
4423 kmem_minfirewall = minfirewall;
4424
4425 /*
4426 * give segkmem a chance to figure out if we are using large pages
4427 * for the kernel heap
4428 */
4429 use_large_pages = segkmem_lpsetup();
4430
4431 /*
4432 * To protect against corruption, we keep the actual number of callers
4433 * KMF_LITE records seperate from the tunable. We arbitrarily clamp
4434 * to 16, since the overhead for small buffers quickly gets out of
4435 * hand.
4436 *
4437 * The real limit would depend on the needs of the largest KMC_NOHASH
4438 * cache.
4439 */
4440 kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4441 kmem_lite_pcs = kmem_lite_count;
4442
4443 /*
4444 * Normally, we firewall oversized allocations when possible, but
4445 * if we are using large pages for kernel memory, and we don't have
4446 * any non-LITE debugging flags set, we want to allocate oversized
4447 * buffers from large pages, and so skip the firewalling.
4448 */
4449 if (use_large_pages &&
4450 ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4451 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4452 PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4453 0, VMC_DUMPSAFE | VM_SLEEP);
4454 } else {
4455 kmem_oversize_arena = vmem_create("kmem_oversize",
4456 NULL, 0, PAGESIZE,
4457 segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4458 kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4459 VM_SLEEP);
4460 }
4461
4462 kmem_cache_init(2, use_large_pages);
4463
4464 if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4465 if (kmem_transaction_log_size == 0)
4466 kmem_transaction_log_size = kmem_maxavail() / 50;
4467 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4468 }
4469
4470 if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4471 if (kmem_content_log_size == 0)
4472 kmem_content_log_size = kmem_maxavail() / 50;
4473 kmem_content_log = kmem_log_init(kmem_content_log_size);
4474 }
4475
4476 kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4477 kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4478 kmem_zerosized_log = kmem_log_init(kmem_zerosized_log_size);
4479
4480 /*
4481 * Initialize STREAMS message caches so allocb() is available.
4482 * This allows us to initialize the logging framework (cmn_err(9F),
4483 * strlog(9F), etc) so we can start recording messages.
4484 */
4485 streams_msg_init();
4486
4487 /*
4488 * Initialize the ZSD framework in Zones so modules loaded henceforth
4489 * can register their callbacks.
4490 */
4491 zone_zsd_init();
4492
4493 log_init();
4494 taskq_init();
4495
4496 /*
4497 * Warn about invalid or dangerous values of kmem_flags.
4498 * Always warn about unsupported values.
4499 */
4500 if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4501 KMF_CONTENTS | KMF_LITE)) != 0) ||
4502 ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4503 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4504 "See the Solaris Tunable Parameters Reference Manual.",
4505 kmem_flags);
4506
4507 #ifdef DEBUG
4508 if ((kmem_flags & KMF_DEBUG) == 0)
4509 cmn_err(CE_NOTE, "kmem debugging disabled.");
4510 #else
4511 /*
4512 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4513 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4514 * if KMF_AUDIT is set). We should warn the user about the performance
4515 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4516 * isn't set (since that disables AUDIT).
4517 */
4518 if (!(kmem_flags & KMF_LITE) &&
4519 (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4520 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4521 "enabled (kmem_flags = 0x%x). Performance degradation "
4522 "and large memory overhead possible. See the Solaris "
4523 "Tunable Parameters Reference Manual.", kmem_flags);
4524 #endif /* not DEBUG */
4525
4526 kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4527
4528 kmem_ready = 1;
4529
4530 /*
4531 * Initialize the platform-specific aligned/DMA memory allocator.
4532 */
4533 ka_init();
4534
4535 /*
4536 * Initialize 32-bit ID cache.
4537 */
4538 id32_init();
4539
4540 /*
4541 * Initialize the networking stack so modules loaded can
4542 * register their callbacks.
4543 */
4544 netstack_init();
4545 }
4546
4547 static void
4548 kmem_move_init(void)
4549 {
4550 kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4551 sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4552 kmem_msb_arena, KMC_NOHASH);
4553 kmem_move_cache = kmem_cache_create("kmem_move_cache",
4554 sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4555 kmem_msb_arena, KMC_NOHASH);
4556
4557 /*
4558 * kmem guarantees that move callbacks are sequential and that even
4559 * across multiple caches no two moves ever execute simultaneously.
4560 * Move callbacks are processed on a separate taskq so that client code
4561 * does not interfere with internal maintenance tasks.
4562 */
4563 kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4564 minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4565 }
4566
4567 void
4568 kmem_thread_init(void)
4569 {
4570 kmem_move_init();
4571 kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4572 300, INT_MAX, TASKQ_PREPOPULATE);
4573 }
4574
4575 void
4576 kmem_mp_init(void)
4577 {
4578 mutex_enter(&cpu_lock);
4579 register_cpu_setup_func(kmem_cpu_setup, NULL);
4580 mutex_exit(&cpu_lock);
4581
4582 kmem_update_timeout(NULL);
4583
4584 taskq_mp_init();
4585 }
4586
4587 /*
4588 * Return the slab of the allocated buffer, or NULL if the buffer is not
4589 * allocated. This function may be called with a known slab address to determine
4590 * whether or not the buffer is allocated, or with a NULL slab address to obtain
4591 * an allocated buffer's slab.
4592 */
4593 static kmem_slab_t *
4594 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4595 {
4596 kmem_bufctl_t *bcp, *bufbcp;
4597
4598 ASSERT(MUTEX_HELD(&cp->cache_lock));
4599 ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4600
4601 if (cp->cache_flags & KMF_HASH) {
4602 for (bcp = *KMEM_HASH(cp, buf);
4603 (bcp != NULL) && (bcp->bc_addr != buf);
4604 bcp = bcp->bc_next) {
4605 continue;
4606 }
4607 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4608 return (bcp == NULL ? NULL : bcp->bc_slab);
4609 }
4610
4611 if (sp == NULL) {
4612 sp = KMEM_SLAB(cp, buf);
4613 }
4614 bufbcp = KMEM_BUFCTL(cp, buf);
4615 for (bcp = sp->slab_head;
4616 (bcp != NULL) && (bcp != bufbcp);
4617 bcp = bcp->bc_next) {
4618 continue;
4619 }
4620 return (bcp == NULL ? sp : NULL);
4621 }
4622
4623 static boolean_t
4624 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4625 {
4626 long refcnt = sp->slab_refcnt;
4627
4628 ASSERT(cp->cache_defrag != NULL);
4629
4630 /*
4631 * For code coverage we want to be able to move an object within the
4632 * same slab (the only partial slab) even if allocating the destination
4633 * buffer resulted in a completely allocated slab.
4634 */
4635 if (flags & KMM_DEBUG) {
4636 return ((flags & KMM_DESPERATE) ||
4637 ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4638 }
4639
4640 /* If we're desperate, we don't care if the client said NO. */
4641 if (flags & KMM_DESPERATE) {
4642 return (refcnt < sp->slab_chunks); /* any partial */
4643 }
4644
4645 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4646 return (B_FALSE);
4647 }
4648
4649 if ((refcnt == 1) || kmem_move_any_partial) {
4650 return (refcnt < sp->slab_chunks);
4651 }
4652
4653 /*
4654 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4655 * slabs with a progressively higher percentage of used buffers can be
4656 * reclaimed until the cache as a whole is no longer fragmented.
4657 *
4658 * sp->slab_refcnt kmd_reclaim_numer
4659 * --------------- < ------------------
4660 * sp->slab_chunks KMEM_VOID_FRACTION
4661 */
4662 return ((refcnt * KMEM_VOID_FRACTION) <
4663 (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4664 }
4665
4666 /*
4667 * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4668 * or when the buffer is freed.
4669 */
4670 static void
4671 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4672 {
4673 ASSERT(MUTEX_HELD(&cp->cache_lock));
4674 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4675
4676 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4677 return;
4678 }
4679
4680 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4681 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4682 avl_remove(&cp->cache_partial_slabs, sp);
4683 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4684 sp->slab_stuck_offset = (uint32_t)-1;
4685 avl_add(&cp->cache_partial_slabs, sp);
4686 }
4687 } else {
4688 sp->slab_later_count = 0;
4689 sp->slab_stuck_offset = (uint32_t)-1;
4690 }
4691 }
4692
4693 static void
4694 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4695 {
4696 ASSERT(taskq_member(kmem_move_taskq, curthread));
4697 ASSERT(MUTEX_HELD(&cp->cache_lock));
4698 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4699
4700 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4701 return;
4702 }
4703
4704 avl_remove(&cp->cache_partial_slabs, sp);
4705 sp->slab_later_count = 0;
4706 sp->slab_flags |= KMEM_SLAB_NOMOVE;
4707 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4708 avl_add(&cp->cache_partial_slabs, sp);
4709 }
4710
4711 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4712
4713 /*
4714 * The move callback takes two buffer addresses, the buffer to be moved, and a
4715 * newly allocated and constructed buffer selected by kmem as the destination.
4716 * It also takes the size of the buffer and an optional user argument specified
4717 * at cache creation time. kmem guarantees that the buffer to be moved has not
4718 * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4719 * guarantee the present whereabouts of the buffer to be moved, so it is up to
4720 * the client to safely determine whether or not it is still using the buffer.
4721 * The client must not free either of the buffers passed to the move callback,
4722 * since kmem wants to free them directly to the slab layer. The client response
4723 * tells kmem which of the two buffers to free:
4724 *
4725 * YES kmem frees the old buffer (the move was successful)
4726 * NO kmem frees the new buffer, marks the slab of the old buffer
4727 * non-reclaimable to avoid bothering the client again
4728 * LATER kmem frees the new buffer, increments slab_later_count
4729 * DONT_KNOW kmem frees the new buffer
4730 * DONT_NEED kmem frees both the old buffer and the new buffer
4731 *
4732 * The pending callback argument now being processed contains both of the
4733 * buffers (old and new) passed to the move callback function, the slab of the
4734 * old buffer, and flags related to the move request, such as whether or not the
4735 * system was desperate for memory.
4736 *
4737 * Slabs are not freed while there is a pending callback, but instead are kept
4738 * on a deadlist, which is drained after the last callback completes. This means
4739 * that slabs are safe to access until kmem_move_end(), no matter how many of
4740 * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4741 * zero for as long as the slab remains on the deadlist and until the slab is
4742 * freed.
4743 */
4744 static void
4745 kmem_move_buffer(kmem_move_t *callback)
4746 {
4747 kmem_cbrc_t response;
4748 kmem_slab_t *sp = callback->kmm_from_slab;
4749 kmem_cache_t *cp = sp->slab_cache;
4750 boolean_t free_on_slab;
4751
4752 ASSERT(taskq_member(kmem_move_taskq, curthread));
4753 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4754 ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4755
4756 /*
4757 * The number of allocated buffers on the slab may have changed since we
4758 * last checked the slab's reclaimability (when the pending move was
4759 * enqueued), or the client may have responded NO when asked to move
4760 * another buffer on the same slab.
4761 */
4762 if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4763 kmem_slab_free(cp, callback->kmm_to_buf);
4764 kmem_move_end(cp, callback);
4765 return;
4766 }
4767
4768 /*
4769 * Checking the slab layer is easy, so we might as well do that here
4770 * in case we can avoid bothering the client.
4771 */
4772 mutex_enter(&cp->cache_lock);
4773 free_on_slab = (kmem_slab_allocated(cp, sp,
4774 callback->kmm_from_buf) == NULL);
4775 mutex_exit(&cp->cache_lock);
4776
4777 if (free_on_slab) {
4778 kmem_slab_free(cp, callback->kmm_to_buf);
4779 kmem_move_end(cp, callback);
4780 return;
4781 }
4782
4783 if (cp->cache_flags & KMF_BUFTAG) {
4784 /*
4785 * Make kmem_cache_alloc_debug() apply the constructor for us.
4786 */
4787 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4788 KM_NOSLEEP, 1, caller()) != 0) {
4789 kmem_move_end(cp, callback);
4790 return;
4791 }
4792 } else if (cp->cache_constructor != NULL &&
4793 cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4794 KM_NOSLEEP) != 0) {
4795 atomic_inc_64(&cp->cache_alloc_fail);
4796 kmem_slab_free(cp, callback->kmm_to_buf);
4797 kmem_move_end(cp, callback);
4798 return;
4799 }
4800
4801 cp->cache_defrag->kmd_callbacks++;
4802 cp->cache_defrag->kmd_thread = curthread;
4803 cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4804 cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4805 DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4806 callback);
4807
4808 response = cp->cache_move(callback->kmm_from_buf,
4809 callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4810
4811 DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4812 callback, kmem_cbrc_t, response);
4813 cp->cache_defrag->kmd_thread = NULL;
4814 cp->cache_defrag->kmd_from_buf = NULL;
4815 cp->cache_defrag->kmd_to_buf = NULL;
4816
4817 if (response == KMEM_CBRC_YES) {
4818 cp->cache_defrag->kmd_yes++;
4819 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4820 /* slab safe to access until kmem_move_end() */
4821 if (sp->slab_refcnt == 0)
4822 cp->cache_defrag->kmd_slabs_freed++;
4823 mutex_enter(&cp->cache_lock);
4824 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4825 mutex_exit(&cp->cache_lock);
4826 kmem_move_end(cp, callback);
4827 return;
4828 }
4829
4830 switch (response) {
4831 case KMEM_CBRC_NO:
4832 cp->cache_defrag->kmd_no++;
4833 mutex_enter(&cp->cache_lock);
4834 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4835 mutex_exit(&cp->cache_lock);
4836 break;
4837 case KMEM_CBRC_LATER:
4838 cp->cache_defrag->kmd_later++;
4839 mutex_enter(&cp->cache_lock);
4840 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4841 mutex_exit(&cp->cache_lock);
4842 break;
4843 }
4844
4845 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4846 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4847 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4848 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4849 callback->kmm_from_buf);
4850 }
4851 mutex_exit(&cp->cache_lock);
4852 break;
4853 case KMEM_CBRC_DONT_NEED:
4854 cp->cache_defrag->kmd_dont_need++;
4855 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4856 if (sp->slab_refcnt == 0)
4857 cp->cache_defrag->kmd_slabs_freed++;
4858 mutex_enter(&cp->cache_lock);
4859 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4860 mutex_exit(&cp->cache_lock);
4861 break;
4862 case KMEM_CBRC_DONT_KNOW:
4863 /*
4864 * If we don't know if we can move this buffer or not, we'll
4865 * just assume that we can't: if the buffer is in fact free,
4866 * then it is sitting in one of the per-CPU magazines or in
4867 * a full magazine in the depot layer. Either way, because
4868 * defrag is induced in the same logic that reaps a cache,
4869 * it's likely that full magazines will be returned to the
4870 * system soon (thereby accomplishing what we're trying to
4871 * accomplish here: return those magazines to their slabs).
4872 * Given this, any work that we might do now to locate a buffer
4873 * in a magazine is wasted (and expensive!) work; we bump
4874 * a counter in this case and otherwise assume that we can't
4875 * move it.
4876 */
4877 cp->cache_defrag->kmd_dont_know++;
4878 break;
4879 default:
4880 panic("'%s' (%p) unexpected move callback response %d\n",
4881 cp->cache_name, (void *)cp, response);
4882 }
4883
4884 kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4885 kmem_move_end(cp, callback);
4886 }
4887
4888 /* Return B_FALSE if there is insufficient memory for the move request. */
4889 static boolean_t
4890 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4891 {
4892 void *to_buf;
4893 avl_index_t index;
4894 kmem_move_t *callback, *pending;
4895 ulong_t n;
4896
4897 ASSERT(taskq_member(kmem_taskq, curthread));
4898 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4899 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4900
4901 callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4902
4903 if (callback == NULL)
4904 return (B_FALSE);
4905
4906 callback->kmm_from_slab = sp;
4907 callback->kmm_from_buf = buf;
4908 callback->kmm_flags = flags;
4909
4910 mutex_enter(&cp->cache_lock);
4911
4912 n = avl_numnodes(&cp->cache_partial_slabs);
4913 if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
4914 mutex_exit(&cp->cache_lock);
4915 kmem_cache_free(kmem_move_cache, callback);
4916 return (B_TRUE); /* there is no need for the move request */
4917 }
4918
4919 pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
4920 if (pending != NULL) {
4921 /*
4922 * If the move is already pending and we're desperate now,
4923 * update the move flags.
4924 */
4925 if (flags & KMM_DESPERATE) {
4926 pending->kmm_flags |= KMM_DESPERATE;
4927 }
4928 mutex_exit(&cp->cache_lock);
4929 kmem_cache_free(kmem_move_cache, callback);
4930 return (B_TRUE);
4931 }
4932
4933 to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
4934 B_FALSE);
4935 callback->kmm_to_buf = to_buf;
4936 avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
4937
4938 mutex_exit(&cp->cache_lock);
4939
4940 if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
4941 callback, TQ_NOSLEEP)) {
4942 mutex_enter(&cp->cache_lock);
4943 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4944 mutex_exit(&cp->cache_lock);
4945 kmem_slab_free(cp, to_buf);
4946 kmem_cache_free(kmem_move_cache, callback);
4947 return (B_FALSE);
4948 }
4949
4950 return (B_TRUE);
4951 }
4952
4953 static void
4954 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
4955 {
4956 avl_index_t index;
4957
4958 ASSERT(cp->cache_defrag != NULL);
4959 ASSERT(taskq_member(kmem_move_taskq, curthread));
4960 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4961
4962 mutex_enter(&cp->cache_lock);
4963 VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
4964 callback->kmm_from_buf, &index) != NULL);
4965 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4966 if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
4967 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
4968 kmem_slab_t *sp;
4969
4970 /*
4971 * The last pending move completed. Release all slabs from the
4972 * front of the dead list except for any slab at the tail that
4973 * needs to be released from the context of kmem_move_buffers().
4974 * kmem deferred unmapping the buffers on these slabs in order
4975 * to guarantee that buffers passed to the move callback have
4976 * been touched only by kmem or by the client itself.
4977 */
4978 while ((sp = list_remove_head(deadlist)) != NULL) {
4979 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
4980 list_insert_tail(deadlist, sp);
4981 break;
4982 }
4983 cp->cache_defrag->kmd_deadcount--;
4984 cp->cache_slab_destroy++;
4985 mutex_exit(&cp->cache_lock);
4986 kmem_slab_destroy(cp, sp);
4987 mutex_enter(&cp->cache_lock);
4988 }
4989 }
4990 mutex_exit(&cp->cache_lock);
4991 kmem_cache_free(kmem_move_cache, callback);
4992 }
4993
4994 /*
4995 * Move buffers from least used slabs first by scanning backwards from the end
4996 * of the partial slab list. Scan at most max_scan candidate slabs and move
4997 * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
4998 * If desperate to reclaim memory, move buffers from any partial slab, otherwise
4999 * skip slabs with a ratio of allocated buffers at or above the current
5000 * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5001 * scan is aborted) so that the caller can adjust the reclaimability threshold
5002 * depending on how many reclaimable slabs it finds.
5003 *
5004 * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5005 * move request, since it is not valid for kmem_move_begin() to call
5006 * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5007 */
5008 static int
5009 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5010 int flags)
5011 {
5012 kmem_slab_t *sp;
5013 void *buf;
5014 int i, j; /* slab index, buffer index */
5015 int s; /* reclaimable slabs */
5016 int b; /* allocated (movable) buffers on reclaimable slab */
5017 boolean_t success;
5018 int refcnt;
5019 int nomove;
5020
5021 ASSERT(taskq_member(kmem_taskq, curthread));
5022 ASSERT(MUTEX_HELD(&cp->cache_lock));
5023 ASSERT(kmem_move_cache != NULL);
5024 ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5025 ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5026 avl_numnodes(&cp->cache_partial_slabs) > 1);
5027
5028 if (kmem_move_blocked) {
5029 return (0);
5030 }
5031
5032 if (kmem_move_fulltilt) {
5033 flags |= KMM_DESPERATE;
5034 }
5035
5036 if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5037 /*
5038 * Scan as many slabs as needed to find the desired number of
5039 * candidate slabs.
5040 */
5041 max_scan = (size_t)-1;
5042 }
5043
5044 if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5045 /* Find as many candidate slabs as possible. */
5046 max_slabs = (size_t)-1;
5047 }
5048
5049 sp = avl_last(&cp->cache_partial_slabs);
5050 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5051 for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5052 ((sp != avl_first(&cp->cache_partial_slabs)) ||
5053 (flags & KMM_DEBUG));
5054 sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5055
5056 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5057 continue;
5058 }
5059 s++;
5060
5061 /* Look for allocated buffers to move. */
5062 for (j = 0, b = 0, buf = sp->slab_base;
5063 (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5064 buf = (((char *)buf) + cp->cache_chunksize), j++) {
5065
5066 if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5067 continue;
5068 }
5069
5070 b++;
5071
5072 /*
5073 * Prevent the slab from being destroyed while we drop
5074 * cache_lock and while the pending move is not yet
5075 * registered. Flag the pending move while
5076 * kmd_moves_pending may still be empty, since we can't
5077 * yet rely on a non-zero pending move count to prevent
5078 * the slab from being destroyed.
5079 */
5080 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5081 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5082 /*
5083 * Recheck refcnt and nomove after reacquiring the lock,
5084 * since these control the order of partial slabs, and
5085 * we want to know if we can pick up the scan where we
5086 * left off.
5087 */
5088 refcnt = sp->slab_refcnt;
5089 nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5090 mutex_exit(&cp->cache_lock);
5091
5092 success = kmem_move_begin(cp, sp, buf, flags);
5093
5094 /*
5095 * Now, before the lock is reacquired, kmem could
5096 * process all pending move requests and purge the
5097 * deadlist, so that upon reacquiring the lock, sp has
5098 * been remapped. Or, the client may free all the
5099 * objects on the slab while the pending moves are still
5100 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5101 * flag causes the slab to be put at the end of the
5102 * deadlist and prevents it from being destroyed, since
5103 * we plan to destroy it here after reacquiring the
5104 * lock.
5105 */
5106 mutex_enter(&cp->cache_lock);
5107 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5108 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5109
5110 if (sp->slab_refcnt == 0) {
5111 list_t *deadlist =
5112 &cp->cache_defrag->kmd_deadlist;
5113 list_remove(deadlist, sp);
5114
5115 if (!avl_is_empty(
5116 &cp->cache_defrag->kmd_moves_pending)) {
5117 /*
5118 * A pending move makes it unsafe to
5119 * destroy the slab, because even though
5120 * the move is no longer needed, the
5121 * context where that is determined
5122 * requires the slab to exist.
5123 * Fortunately, a pending move also
5124 * means we don't need to destroy the
5125 * slab here, since it will get
5126 * destroyed along with any other slabs
5127 * on the deadlist after the last
5128 * pending move completes.
5129 */
5130 list_insert_head(deadlist, sp);
5131 return (-1);
5132 }
5133
5134 /*
5135 * Destroy the slab now if it was completely
5136 * freed while we dropped cache_lock and there
5137 * are no pending moves. Since slab_refcnt
5138 * cannot change once it reaches zero, no new
5139 * pending moves from that slab are possible.
5140 */
5141 cp->cache_defrag->kmd_deadcount--;
5142 cp->cache_slab_destroy++;
5143 mutex_exit(&cp->cache_lock);
5144 kmem_slab_destroy(cp, sp);
5145 mutex_enter(&cp->cache_lock);
5146 /*
5147 * Since we can't pick up the scan where we left
5148 * off, abort the scan and say nothing about the
5149 * number of reclaimable slabs.
5150 */
5151 return (-1);
5152 }
5153
5154 if (!success) {
5155 /*
5156 * Abort the scan if there is not enough memory
5157 * for the request and say nothing about the
5158 * number of reclaimable slabs.
5159 */
5160 return (-1);
5161 }
5162
5163 /*
5164 * The slab's position changed while the lock was
5165 * dropped, so we don't know where we are in the
5166 * sequence any more.
5167 */
5168 if (sp->slab_refcnt != refcnt) {
5169 /*
5170 * If this is a KMM_DEBUG move, the slab_refcnt
5171 * may have changed because we allocated a
5172 * destination buffer on the same slab. In that
5173 * case, we're not interested in counting it.
5174 */
5175 return (-1);
5176 }
5177 if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove)
5178 return (-1);
5179
5180 /*
5181 * Generating a move request allocates a destination
5182 * buffer from the slab layer, bumping the first partial
5183 * slab if it is completely allocated. If the current
5184 * slab becomes the first partial slab as a result, we
5185 * can't continue to scan backwards.
5186 *
5187 * If this is a KMM_DEBUG move and we allocated the
5188 * destination buffer from the last partial slab, then
5189 * the buffer we're moving is on the same slab and our
5190 * slab_refcnt has changed, causing us to return before
5191 * reaching here if there are no partial slabs left.
5192 */
5193 ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5194 if (sp == avl_first(&cp->cache_partial_slabs)) {
5195 /*
5196 * We're not interested in a second KMM_DEBUG
5197 * move.
5198 */
5199 goto end_scan;
5200 }
5201 }
5202 }
5203 end_scan:
5204
5205 return (s);
5206 }
5207
5208 typedef struct kmem_move_notify_args {
5209 kmem_cache_t *kmna_cache;
5210 void *kmna_buf;
5211 } kmem_move_notify_args_t;
5212
5213 static void
5214 kmem_cache_move_notify_task(void *arg)
5215 {
5216 kmem_move_notify_args_t *args = arg;
5217 kmem_cache_t *cp = args->kmna_cache;
5218 void *buf = args->kmna_buf;
5219 kmem_slab_t *sp;
5220
5221 ASSERT(taskq_member(kmem_taskq, curthread));
5222 ASSERT(list_link_active(&cp->cache_link));
5223
5224 kmem_free(args, sizeof (kmem_move_notify_args_t));
5225 mutex_enter(&cp->cache_lock);
5226 sp = kmem_slab_allocated(cp, NULL, buf);
5227
5228 /* Ignore the notification if the buffer is no longer allocated. */
5229 if (sp == NULL) {
5230 mutex_exit(&cp->cache_lock);
5231 return;
5232 }
5233
5234 /* Ignore the notification if there's no reason to move the buffer. */
5235 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5236 /*
5237 * So far the notification is not ignored. Ignore the
5238 * notification if the slab is not marked by an earlier refusal
5239 * to move a buffer.
5240 */
5241 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5242 (sp->slab_later_count == 0)) {
5243 mutex_exit(&cp->cache_lock);
5244 return;
5245 }
5246
5247 kmem_slab_move_yes(cp, sp, buf);
5248 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5249 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5250 mutex_exit(&cp->cache_lock);
5251 /* see kmem_move_buffers() about dropping the lock */
5252 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5253 mutex_enter(&cp->cache_lock);
5254 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5255 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5256 if (sp->slab_refcnt == 0) {
5257 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5258 list_remove(deadlist, sp);
5259
5260 if (!avl_is_empty(
5261 &cp->cache_defrag->kmd_moves_pending)) {
5262 list_insert_head(deadlist, sp);
5263 mutex_exit(&cp->cache_lock);
5264 return;
5265 }
5266
5267 cp->cache_defrag->kmd_deadcount--;
5268 cp->cache_slab_destroy++;
5269 mutex_exit(&cp->cache_lock);
5270 kmem_slab_destroy(cp, sp);
5271 return;
5272 }
5273 } else {
5274 kmem_slab_move_yes(cp, sp, buf);
5275 }
5276 mutex_exit(&cp->cache_lock);
5277 }
5278
5279 void
5280 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5281 {
5282 kmem_move_notify_args_t *args;
5283
5284 args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5285 if (args != NULL) {
5286 args->kmna_cache = cp;
5287 args->kmna_buf = buf;
5288 if (!taskq_dispatch(kmem_taskq,
5289 (task_func_t *)kmem_cache_move_notify_task, args,
5290 TQ_NOSLEEP))
5291 kmem_free(args, sizeof (kmem_move_notify_args_t));
5292 }
5293 }
5294
5295 static void
5296 kmem_cache_defrag(kmem_cache_t *cp)
5297 {
5298 size_t n;
5299
5300 ASSERT(cp->cache_defrag != NULL);
5301
5302 mutex_enter(&cp->cache_lock);
5303 n = avl_numnodes(&cp->cache_partial_slabs);
5304 if (n > 1) {
5305 /* kmem_move_buffers() drops and reacquires cache_lock */
5306 cp->cache_defrag->kmd_defrags++;
5307 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5308 }
5309 mutex_exit(&cp->cache_lock);
5310 }
5311
5312 /* Is this cache above the fragmentation threshold? */
5313 static boolean_t
5314 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5315 {
5316 /*
5317 * nfree kmem_frag_numer
5318 * ------------------ > ---------------
5319 * cp->cache_buftotal kmem_frag_denom
5320 */
5321 return ((nfree * kmem_frag_denom) >
5322 (cp->cache_buftotal * kmem_frag_numer));
5323 }
5324
5325 static boolean_t
5326 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5327 {
5328 boolean_t fragmented;
5329 uint64_t nfree;
5330
5331 ASSERT(MUTEX_HELD(&cp->cache_lock));
5332 *doreap = B_FALSE;
5333
5334 if (kmem_move_fulltilt) {
5335 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5336 return (B_TRUE);
5337 }
5338 } else {
5339 if ((cp->cache_complete_slab_count + avl_numnodes(
5340 &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5341 return (B_FALSE);
5342 }
5343 }
5344
5345 nfree = cp->cache_bufslab;
5346 fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5347 kmem_cache_frag_threshold(cp, nfree));
5348
5349 /*
5350 * Free buffers in the magazine layer appear allocated from the point of
5351 * view of the slab layer. We want to know if the slab layer would
5352 * appear fragmented if we included free buffers from magazines that
5353 * have fallen out of the working set.
5354 */
5355 if (!fragmented) {
5356 long reap;
5357
5358 mutex_enter(&cp->cache_depot_lock);
5359 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5360 reap = MIN(reap, cp->cache_full.ml_total);
5361 mutex_exit(&cp->cache_depot_lock);
5362
5363 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5364 if (kmem_cache_frag_threshold(cp, nfree)) {
5365 *doreap = B_TRUE;
5366 }
5367 }
5368
5369 return (fragmented);
5370 }
5371
5372 /* Called periodically from kmem_taskq */
5373 static void
5374 kmem_cache_scan(kmem_cache_t *cp)
5375 {
5376 boolean_t reap = B_FALSE;
5377 kmem_defrag_t *kmd;
5378
5379 ASSERT(taskq_member(kmem_taskq, curthread));
5380
5381 mutex_enter(&cp->cache_lock);
5382
5383 kmd = cp->cache_defrag;
5384 if (kmd->kmd_consolidate > 0) {
5385 kmd->kmd_consolidate--;
5386 mutex_exit(&cp->cache_lock);
5387 kmem_cache_reap(cp);
5388 return;
5389 }
5390
5391 if (kmem_cache_is_fragmented(cp, &reap)) {
5392 size_t slabs_found;
5393
5394 /*
5395 * Consolidate reclaimable slabs from the end of the partial
5396 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5397 * reclaimable slabs). Keep track of how many candidate slabs we
5398 * looked for and how many we actually found so we can adjust
5399 * the definition of a candidate slab if we're having trouble
5400 * finding them.
5401 *
5402 * kmem_move_buffers() drops and reacquires cache_lock.
5403 */
5404 kmd->kmd_scans++;
5405 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5406 kmem_reclaim_max_slabs, 0);
5407 if (slabs_found >= 0) {
5408 kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5409 kmd->kmd_slabs_found += slabs_found;
5410 }
5411
5412 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5413 kmd->kmd_tries = 0;
5414
5415 /*
5416 * If we had difficulty finding candidate slabs in
5417 * previous scans, adjust the threshold so that
5418 * candidates are easier to find.
5419 */
5420 if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5421 kmem_adjust_reclaim_threshold(kmd, -1);
5422 } else if ((kmd->kmd_slabs_found * 2) <
5423 kmd->kmd_slabs_sought) {
5424 kmem_adjust_reclaim_threshold(kmd, 1);
5425 }
5426 kmd->kmd_slabs_sought = 0;
5427 kmd->kmd_slabs_found = 0;
5428 }
5429 } else {
5430 kmem_reset_reclaim_threshold(cp->cache_defrag);
5431 #ifdef DEBUG
5432 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5433 /*
5434 * In a debug kernel we want the consolidator to
5435 * run occasionally even when there is plenty of
5436 * memory.
5437 */
5438 uint16_t debug_rand;
5439
5440 (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5441 if (!kmem_move_noreap &&
5442 ((debug_rand % kmem_mtb_reap) == 0)) {
5443 mutex_exit(&cp->cache_lock);
5444 kmem_cache_reap(cp);
5445 return;
5446 } else if ((debug_rand % kmem_mtb_move) == 0) {
5447 kmd->kmd_scans++;
5448 (void) kmem_move_buffers(cp,
5449 kmem_reclaim_scan_range, 1, KMM_DEBUG);
5450 }
5451 }
5452 #endif /* DEBUG */
5453 }
5454
5455 mutex_exit(&cp->cache_lock);
5456
5457 if (reap)
5458 kmem_depot_ws_reap(cp);
5459 }