xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision dfec2ecf)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
25  * Copyright 2018, Joyent, Inc.
26  */
27 
28 /*
29  * Kernel memory allocator, as described in the following two papers and a
30  * statement about the consolidator:
31  *
32  * Jeff Bonwick,
33  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
34  * Proceedings of the Summer 1994 Usenix Conference.
35  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
36  *
37  * Jeff Bonwick and Jonathan Adams,
38  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
39  * Arbitrary Resources.
40  * Proceedings of the 2001 Usenix Conference.
41  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
42  *
43  * kmem Slab Consolidator Big Theory Statement:
44  *
45  * 1. Motivation
46  *
47  * As stated in Bonwick94, slabs provide the following advantages over other
48  * allocation structures in terms of memory fragmentation:
49  *
50  *  - Internal fragmentation (per-buffer wasted space) is minimal.
51  *  - Severe external fragmentation (unused buffers on the free list) is
52  *    unlikely.
53  *
54  * Segregating objects by size eliminates one source of external fragmentation,
55  * and according to Bonwick:
56  *
57  *   The other reason that slabs reduce external fragmentation is that all
58  *   objects in a slab are of the same type, so they have the same lifetime
59  *   distribution. The resulting segregation of short-lived and long-lived
60  *   objects at slab granularity reduces the likelihood of an entire page being
61  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
62  *
63  * While unlikely, severe external fragmentation remains possible. Clients that
64  * allocate both short- and long-lived objects from the same cache cannot
65  * anticipate the distribution of long-lived objects within the allocator's slab
66  * implementation. Even a small percentage of long-lived objects distributed
67  * randomly across many slabs can lead to a worst case scenario where the client
68  * frees the majority of its objects and the system gets back almost none of the
69  * slabs. Despite the client doing what it reasonably can to help the system
70  * reclaim memory, the allocator cannot shake free enough slabs because of
71  * lonely allocations stubbornly hanging on. Although the allocator is in a
72  * position to diagnose the fragmentation, there is nothing that the allocator
73  * by itself can do about it. It only takes a single allocated object to prevent
74  * an entire slab from being reclaimed, and any object handed out by
75  * kmem_cache_alloc() is by definition in the client's control. Conversely,
76  * although the client is in a position to move a long-lived object, it has no
77  * way of knowing if the object is causing fragmentation, and if so, where to
78  * move it. A solution necessarily requires further cooperation between the
79  * allocator and the client.
80  *
81  * 2. Move Callback
82  *
83  * The kmem slab consolidator therefore adds a move callback to the
84  * allocator/client interface, improving worst-case external fragmentation in
85  * kmem caches that supply a function to move objects from one memory location
86  * to another. In a situation of low memory kmem attempts to consolidate all of
87  * a cache's slabs at once; otherwise it works slowly to bring external
88  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
89  * thereby helping to avoid a low memory situation in the future.
90  *
91  * The callback has the following signature:
92  *
93  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
94  *
95  * It supplies the kmem client with two addresses: the allocated object that
96  * kmem wants to move and a buffer selected by kmem for the client to use as the
97  * copy destination. The callback is kmem's way of saying "Please get off of
98  * this buffer and use this one instead." kmem knows where it wants to move the
99  * object in order to best reduce fragmentation. All the client needs to know
100  * about the second argument (void *new) is that it is an allocated, constructed
101  * object ready to take the contents of the old object. When the move function
102  * is called, the system is likely to be low on memory, and the new object
103  * spares the client from having to worry about allocating memory for the
104  * requested move. The third argument supplies the size of the object, in case a
105  * single move function handles multiple caches whose objects differ only in
106  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
107  * user argument passed to the constructor, destructor, and reclaim functions is
108  * also passed to the move callback.
109  *
110  * 2.1 Setting the Move Callback
111  *
112  * The client sets the move callback after creating the cache and before
113  * allocating from it:
114  *
115  *	object_cache = kmem_cache_create(...);
116  *      kmem_cache_set_move(object_cache, object_move);
117  *
118  * 2.2 Move Callback Return Values
119  *
120  * Only the client knows about its own data and when is a good time to move it.
121  * The client is cooperating with kmem to return unused memory to the system,
122  * and kmem respectfully accepts this help at the client's convenience. When
123  * asked to move an object, the client can respond with any of the following:
124  *
125  *   typedef enum kmem_cbrc {
126  *           KMEM_CBRC_YES,
127  *           KMEM_CBRC_NO,
128  *           KMEM_CBRC_LATER,
129  *           KMEM_CBRC_DONT_NEED,
130  *           KMEM_CBRC_DONT_KNOW
131  *   } kmem_cbrc_t;
132  *
133  * The client must not explicitly kmem_cache_free() either of the objects passed
134  * to the callback, since kmem wants to free them directly to the slab layer
135  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
136  * objects to free:
137  *
138  *       YES: (Did it) The client moved the object, so kmem frees the old one.
139  *        NO: (Never) The client refused, so kmem frees the new object (the
140  *            unused copy destination). kmem also marks the slab of the old
141  *            object so as not to bother the client with further callbacks for
142  *            that object as long as the slab remains on the partial slab list.
143  *            (The system won't be getting the slab back as long as the
144  *            immovable object holds it hostage, so there's no point in moving
145  *            any of its objects.)
146  *     LATER: The client is using the object and cannot move it now, so kmem
147  *            frees the new object (the unused copy destination). kmem still
148  *            attempts to move other objects off the slab, since it expects to
149  *            succeed in clearing the slab in a later callback. The client
150  *            should use LATER instead of NO if the object is likely to become
151  *            movable very soon.
152  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
153  *            with the new object (the unused copy destination). This response
154  *            is the client's opportunity to be a model citizen and give back as
155  *            much as it can.
156  * DONT_KNOW: The client does not know about the object because
157  *            a) the client has just allocated the object and not yet put it
158  *               wherever it expects to find known objects
159  *            b) the client has removed the object from wherever it expects to
160  *               find known objects and is about to free it, or
161  *            c) the client has freed the object.
162  *            In all these cases (a, b, and c) kmem frees the new object (the
163  *            unused copy destination).  In the first case, the object is in
164  *            use and the correct action is that for LATER; in the latter two
165  *            cases, we know that the object is either freed or about to be
166  *            freed, in which case it is either already in a magazine or about
167  *            to be in one.  In these cases, we know that the object will either
168  *            be reallocated and reused, or it will end up in a full magazine
169  *            that will be reaped (thereby liberating the slab).  Because it
170  *            is prohibitively expensive to differentiate these cases, and
171  *            because the defrag code is executed when we're low on memory
172  *            (thereby biasing the system to reclaim full magazines) we treat
173  *            all DONT_KNOW cases as LATER and rely on cache reaping to
174  *            generally clean up full magazines.  While we take the same action
175  *            for these cases, we maintain their semantic distinction:  if
176  *            defragmentation is not occurring, it is useful to know if this
177  *            is due to objects in use (LATER) or objects in an unknown state
178  *            of transition (DONT_KNOW).
179  *
180  * 2.3 Object States
181  *
182  * Neither kmem nor the client can be assumed to know the object's whereabouts
183  * at the time of the callback. An object belonging to a kmem cache may be in
184  * any of the following states:
185  *
186  * 1. Uninitialized on the slab
187  * 2. Allocated from the slab but not constructed (still uninitialized)
188  * 3. Allocated from the slab, constructed, but not yet ready for business
189  *    (not in a valid state for the move callback)
190  * 4. In use (valid and known to the client)
191  * 5. About to be freed (no longer in a valid state for the move callback)
192  * 6. Freed to a magazine (still constructed)
193  * 7. Allocated from a magazine, not yet ready for business (not in a valid
194  *    state for the move callback), and about to return to state #4
195  * 8. Deconstructed on a magazine that is about to be freed
196  * 9. Freed to the slab
197  *
198  * Since the move callback may be called at any time while the object is in any
199  * of the above states (except state #1), the client needs a safe way to
200  * determine whether or not it knows about the object. Specifically, the client
201  * needs to know whether or not the object is in state #4, the only state in
202  * which a move is valid. If the object is in any other state, the client should
203  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
204  * the object's fields.
205  *
206  * Note that although an object may be in state #4 when kmem initiates the move
207  * request, the object may no longer be in that state by the time kmem actually
208  * calls the move function. Not only does the client free objects
209  * asynchronously, kmem itself puts move requests on a queue where thay are
210  * pending until kmem processes them from another context. Also, objects freed
211  * to a magazine appear allocated from the point of view of the slab layer, so
212  * kmem may even initiate requests for objects in a state other than state #4.
213  *
214  * 2.3.1 Magazine Layer
215  *
216  * An important insight revealed by the states listed above is that the magazine
217  * layer is populated only by kmem_cache_free(). Magazines of constructed
218  * objects are never populated directly from the slab layer (which contains raw,
219  * unconstructed objects). Whenever an allocation request cannot be satisfied
220  * from the magazine layer, the magazines are bypassed and the request is
221  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
222  * the object constructor only when allocating from the slab layer, and only in
223  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
224  * the move callback. kmem does not preconstruct objects in anticipation of
225  * kmem_cache_alloc().
226  *
227  * 2.3.2 Object Constructor and Destructor
228  *
229  * If the client supplies a destructor, it must be valid to call the destructor
230  * on a newly created object (immediately after the constructor).
231  *
232  * 2.4 Recognizing Known Objects
233  *
234  * There is a simple test to determine safely whether or not the client knows
235  * about a given object in the move callback. It relies on the fact that kmem
236  * guarantees that the object of the move callback has only been touched by the
237  * client itself or else by kmem. kmem does this by ensuring that none of the
238  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
239  * callback is pending. When the last object on a slab is freed, if there is a
240  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
241  * slabs on that list until all pending callbacks are completed. That way,
242  * clients can be certain that the object of a move callback is in one of the
243  * states listed above, making it possible to distinguish known objects (in
244  * state #4) using the two low order bits of any pointer member (with the
245  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
246  * platforms).
247  *
248  * The test works as long as the client always transitions objects from state #4
249  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
250  * order bit of the client-designated pointer member. Since kmem only writes
251  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
252  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
253  * guaranteed to set at least one of the two low order bits. Therefore, given an
254  * object with a back pointer to a 'container_t *o_container', the client can
255  * test
256  *
257  *      container_t *container = object->o_container;
258  *      if ((uintptr_t)container & 0x3) {
259  *              return (KMEM_CBRC_DONT_KNOW);
260  *      }
261  *
262  * Typically, an object will have a pointer to some structure with a list or
263  * hash where objects from the cache are kept while in use. Assuming that the
264  * client has some way of knowing that the container structure is valid and will
265  * not go away during the move, and assuming that the structure includes a lock
266  * to protect whatever collection is used, then the client would continue as
267  * follows:
268  *
269  *	// Ensure that the container structure does not go away.
270  *      if (container_hold(container) == 0) {
271  *              return (KMEM_CBRC_DONT_KNOW);
272  *      }
273  *      mutex_enter(&container->c_objects_lock);
274  *      if (container != object->o_container) {
275  *              mutex_exit(&container->c_objects_lock);
276  *              container_rele(container);
277  *              return (KMEM_CBRC_DONT_KNOW);
278  *      }
279  *
280  * At this point the client knows that the object cannot be freed as long as
281  * c_objects_lock is held. Note that after acquiring the lock, the client must
282  * recheck the o_container pointer in case the object was removed just before
283  * acquiring the lock.
284  *
285  * When the client is about to free an object, it must first remove that object
286  * from the list, hash, or other structure where it is kept. At that time, to
287  * mark the object so it can be distinguished from the remaining, known objects,
288  * the client sets the designated low order bit:
289  *
290  *      mutex_enter(&container->c_objects_lock);
291  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
292  *      list_remove(&container->c_objects, object);
293  *      mutex_exit(&container->c_objects_lock);
294  *
295  * In the common case, the object is freed to the magazine layer, where it may
296  * be reused on a subsequent allocation without the overhead of calling the
297  * constructor. While in the magazine it appears allocated from the point of
298  * view of the slab layer, making it a candidate for the move callback. Most
299  * objects unrecognized by the client in the move callback fall into this
300  * category and are cheaply distinguished from known objects by the test
301  * described earlier. Because searching magazines is prohibitively expensive
302  * for kmem, clients that do not mark freed objects (and therefore return
303  * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
304  * efficacy reduced.
305  *
306  * Invalidating the designated pointer member before freeing the object marks
307  * the object to be avoided in the callback, and conversely, assigning a valid
308  * value to the designated pointer member after allocating the object makes the
309  * object fair game for the callback:
310  *
311  *      ... allocate object ...
312  *      ... set any initial state not set by the constructor ...
313  *
314  *      mutex_enter(&container->c_objects_lock);
315  *      list_insert_tail(&container->c_objects, object);
316  *      membar_producer();
317  *      object->o_container = container;
318  *      mutex_exit(&container->c_objects_lock);
319  *
320  * Note that everything else must be valid before setting o_container makes the
321  * object fair game for the move callback. The membar_producer() call ensures
322  * that all the object's state is written to memory before setting the pointer
323  * that transitions the object from state #3 or #7 (allocated, constructed, not
324  * yet in use) to state #4 (in use, valid). That's important because the move
325  * function has to check the validity of the pointer before it can safely
326  * acquire the lock protecting the collection where it expects to find known
327  * objects.
328  *
329  * This method of distinguishing known objects observes the usual symmetry:
330  * invalidating the designated pointer is the first thing the client does before
331  * freeing the object, and setting the designated pointer is the last thing the
332  * client does after allocating the object. Of course, the client is not
333  * required to use this method. Fundamentally, how the client recognizes known
334  * objects is completely up to the client, but this method is recommended as an
335  * efficient and safe way to take advantage of the guarantees made by kmem. If
336  * the entire object is arbitrary data without any markable bits from a suitable
337  * pointer member, then the client must find some other method, such as
338  * searching a hash table of known objects.
339  *
340  * 2.5 Preventing Objects From Moving
341  *
342  * Besides a way to distinguish known objects, the other thing that the client
343  * needs is a strategy to ensure that an object will not move while the client
344  * is actively using it. The details of satisfying this requirement tend to be
345  * highly cache-specific. It might seem that the same rules that let a client
346  * remove an object safely should also decide when an object can be moved
347  * safely. However, any object state that makes a removal attempt invalid is
348  * likely to be long-lasting for objects that the client does not expect to
349  * remove. kmem knows nothing about the object state and is equally likely (from
350  * the client's point of view) to request a move for any object in the cache,
351  * whether prepared for removal or not. Even a low percentage of objects stuck
352  * in place by unremovability will defeat the consolidator if the stuck objects
353  * are the same long-lived allocations likely to hold slabs hostage.
354  * Fundamentally, the consolidator is not aimed at common cases. Severe external
355  * fragmentation is a worst case scenario manifested as sparsely allocated
356  * slabs, by definition a low percentage of the cache's objects. When deciding
357  * what makes an object movable, keep in mind the goal of the consolidator: to
358  * bring worst-case external fragmentation within the limits guaranteed for
359  * internal fragmentation. Removability is a poor criterion if it is likely to
360  * exclude more than an insignificant percentage of objects for long periods of
361  * time.
362  *
363  * A tricky general solution exists, and it has the advantage of letting you
364  * move any object at almost any moment, practically eliminating the likelihood
365  * that an object can hold a slab hostage. However, if there is a cache-specific
366  * way to ensure that an object is not actively in use in the vast majority of
367  * cases, a simpler solution that leverages this cache-specific knowledge is
368  * preferred.
369  *
370  * 2.5.1 Cache-Specific Solution
371  *
372  * As an example of a cache-specific solution, the ZFS znode cache takes
373  * advantage of the fact that the vast majority of znodes are only being
374  * referenced from the DNLC. (A typical case might be a few hundred in active
375  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
376  * client has established that it recognizes the znode and can access its fields
377  * safely (using the method described earlier), it then tests whether the znode
378  * is referenced by anything other than the DNLC. If so, it assumes that the
379  * znode may be in active use and is unsafe to move, so it drops its locks and
380  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
381  * else znodes are used, no change is needed to protect against the possibility
382  * of the znode moving. The disadvantage is that it remains possible for an
383  * application to hold a znode slab hostage with an open file descriptor.
384  * However, this case ought to be rare and the consolidator has a way to deal
385  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
386  * object, kmem eventually stops believing it and treats the slab as if the
387  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
388  * then focus on getting it off of the partial slab list by allocating rather
389  * than freeing all of its objects. (Either way of getting a slab off the
390  * free list reduces fragmentation.)
391  *
392  * 2.5.2 General Solution
393  *
394  * The general solution, on the other hand, requires an explicit hold everywhere
395  * the object is used to prevent it from moving. To keep the client locking
396  * strategy as uncomplicated as possible, kmem guarantees the simplifying
397  * assumption that move callbacks are sequential, even across multiple caches.
398  * Internally, a global queue processed by a single thread supports all caches
399  * implementing the callback function. No matter how many caches supply a move
400  * function, the consolidator never moves more than one object at a time, so the
401  * client does not have to worry about tricky lock ordering involving several
402  * related objects from different kmem caches.
403  *
404  * The general solution implements the explicit hold as a read-write lock, which
405  * allows multiple readers to access an object from the cache simultaneously
406  * while a single writer is excluded from moving it. A single rwlock for the
407  * entire cache would lock out all threads from using any of the cache's objects
408  * even though only a single object is being moved, so to reduce contention,
409  * the client can fan out the single rwlock into an array of rwlocks hashed by
410  * the object address, making it probable that moving one object will not
411  * prevent other threads from using a different object. The rwlock cannot be a
412  * member of the object itself, because the possibility of the object moving
413  * makes it unsafe to access any of the object's fields until the lock is
414  * acquired.
415  *
416  * Assuming a small, fixed number of locks, it's possible that multiple objects
417  * will hash to the same lock. A thread that needs to use multiple objects in
418  * the same function may acquire the same lock multiple times. Since rwlocks are
419  * reentrant for readers, and since there is never more than a single writer at
420  * a time (assuming that the client acquires the lock as a writer only when
421  * moving an object inside the callback), there would seem to be no problem.
422  * However, a client locking multiple objects in the same function must handle
423  * one case of potential deadlock: Assume that thread A needs to prevent both
424  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
425  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
426  * same lock, that thread A will acquire the lock for object 1 as a reader
427  * before thread B sets the lock's write-wanted bit, preventing thread A from
428  * reacquiring the lock for object 2 as a reader. Unable to make forward
429  * progress, thread A will never release the lock for object 1, resulting in
430  * deadlock.
431  *
432  * There are two ways of avoiding the deadlock just described. The first is to
433  * use rw_tryenter() rather than rw_enter() in the callback function when
434  * attempting to acquire the lock as a writer. If tryenter discovers that the
435  * same object (or another object hashed to the same lock) is already in use, it
436  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
437  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
438  * since it allows a thread to acquire the lock as a reader in spite of a
439  * waiting writer. This second approach insists on moving the object now, no
440  * matter how many readers the move function must wait for in order to do so,
441  * and could delay the completion of the callback indefinitely (blocking
442  * callbacks to other clients). In practice, a less insistent callback using
443  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
444  * little reason to use anything else.
445  *
446  * Avoiding deadlock is not the only problem that an implementation using an
447  * explicit hold needs to solve. Locking the object in the first place (to
448  * prevent it from moving) remains a problem, since the object could move
449  * between the time you obtain a pointer to the object and the time you acquire
450  * the rwlock hashed to that pointer value. Therefore the client needs to
451  * recheck the value of the pointer after acquiring the lock, drop the lock if
452  * the value has changed, and try again. This requires a level of indirection:
453  * something that points to the object rather than the object itself, that the
454  * client can access safely while attempting to acquire the lock. (The object
455  * itself cannot be referenced safely because it can move at any time.)
456  * The following lock-acquisition function takes whatever is safe to reference
457  * (arg), follows its pointer to the object (using function f), and tries as
458  * often as necessary to acquire the hashed lock and verify that the object
459  * still has not moved:
460  *
461  *      object_t *
462  *      object_hold(object_f f, void *arg)
463  *      {
464  *              object_t *op;
465  *
466  *              op = f(arg);
467  *              if (op == NULL) {
468  *                      return (NULL);
469  *              }
470  *
471  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
472  *              while (op != f(arg)) {
473  *                      rw_exit(OBJECT_RWLOCK(op));
474  *                      op = f(arg);
475  *                      if (op == NULL) {
476  *                              break;
477  *                      }
478  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
479  *              }
480  *
481  *              return (op);
482  *      }
483  *
484  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
485  * lock reacquisition loop, while necessary, almost never executes. The function
486  * pointer f (used to obtain the object pointer from arg) has the following type
487  * definition:
488  *
489  *      typedef object_t *(*object_f)(void *arg);
490  *
491  * An object_f implementation is likely to be as simple as accessing a structure
492  * member:
493  *
494  *      object_t *
495  *      s_object(void *arg)
496  *      {
497  *              something_t *sp = arg;
498  *              return (sp->s_object);
499  *      }
500  *
501  * The flexibility of a function pointer allows the path to the object to be
502  * arbitrarily complex and also supports the notion that depending on where you
503  * are using the object, you may need to get it from someplace different.
504  *
505  * The function that releases the explicit hold is simpler because it does not
506  * have to worry about the object moving:
507  *
508  *      void
509  *      object_rele(object_t *op)
510  *      {
511  *              rw_exit(OBJECT_RWLOCK(op));
512  *      }
513  *
514  * The caller is spared these details so that obtaining and releasing an
515  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
516  * of object_hold() only needs to know that the returned object pointer is valid
517  * if not NULL and that the object will not move until released.
518  *
519  * Although object_hold() prevents an object from moving, it does not prevent it
520  * from being freed. The caller must take measures before calling object_hold()
521  * (afterwards is too late) to ensure that the held object cannot be freed. The
522  * caller must do so without accessing the unsafe object reference, so any lock
523  * or reference count used to ensure the continued existence of the object must
524  * live outside the object itself.
525  *
526  * Obtaining a new object is a special case where an explicit hold is impossible
527  * for the caller. Any function that returns a newly allocated object (either as
528  * a return value, or as an in-out paramter) must return it already held; after
529  * the caller gets it is too late, since the object cannot be safely accessed
530  * without the level of indirection described earlier. The following
531  * object_alloc() example uses the same code shown earlier to transition a new
532  * object into the state of being recognized (by the client) as a known object.
533  * The function must acquire the hold (rw_enter) before that state transition
534  * makes the object movable:
535  *
536  *      static object_t *
537  *      object_alloc(container_t *container)
538  *      {
539  *              object_t *object = kmem_cache_alloc(object_cache, 0);
540  *              ... set any initial state not set by the constructor ...
541  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
542  *              mutex_enter(&container->c_objects_lock);
543  *              list_insert_tail(&container->c_objects, object);
544  *              membar_producer();
545  *              object->o_container = container;
546  *              mutex_exit(&container->c_objects_lock);
547  *              return (object);
548  *      }
549  *
550  * Functions that implicitly acquire an object hold (any function that calls
551  * object_alloc() to supply an object for the caller) need to be carefully noted
552  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
553  * prevent all objects hashed to the affected rwlocks from ever being moved.
554  *
555  * The pointer to a held object can be hashed to the holding rwlock even after
556  * the object has been freed. Although it is possible to release the hold
557  * after freeing the object, you may decide to release the hold implicitly in
558  * whatever function frees the object, so as to release the hold as soon as
559  * possible, and for the sake of symmetry with the function that implicitly
560  * acquires the hold when it allocates the object. Here, object_free() releases
561  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
562  * matching pair with object_hold():
563  *
564  *      void
565  *      object_free(object_t *object)
566  *      {
567  *              container_t *container;
568  *
569  *              ASSERT(object_held(object));
570  *              container = object->o_container;
571  *              mutex_enter(&container->c_objects_lock);
572  *              object->o_container =
573  *                  (void *)((uintptr_t)object->o_container | 0x1);
574  *              list_remove(&container->c_objects, object);
575  *              mutex_exit(&container->c_objects_lock);
576  *              object_rele(object);
577  *              kmem_cache_free(object_cache, object);
578  *      }
579  *
580  * Note that object_free() cannot safely accept an object pointer as an argument
581  * unless the object is already held. Any function that calls object_free()
582  * needs to be carefully noted since it similarly forms a matching pair with
583  * object_hold().
584  *
585  * To complete the picture, the following callback function implements the
586  * general solution by moving objects only if they are currently unheld:
587  *
588  *      static kmem_cbrc_t
589  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
590  *      {
591  *              object_t *op = buf, *np = newbuf;
592  *              container_t *container;
593  *
594  *              container = op->o_container;
595  *              if ((uintptr_t)container & 0x3) {
596  *                      return (KMEM_CBRC_DONT_KNOW);
597  *              }
598  *
599  *	        // Ensure that the container structure does not go away.
600  *              if (container_hold(container) == 0) {
601  *                      return (KMEM_CBRC_DONT_KNOW);
602  *              }
603  *
604  *              mutex_enter(&container->c_objects_lock);
605  *              if (container != op->o_container) {
606  *                      mutex_exit(&container->c_objects_lock);
607  *                      container_rele(container);
608  *                      return (KMEM_CBRC_DONT_KNOW);
609  *              }
610  *
611  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
612  *                      mutex_exit(&container->c_objects_lock);
613  *                      container_rele(container);
614  *                      return (KMEM_CBRC_LATER);
615  *              }
616  *
617  *              object_move_impl(op, np); // critical section
618  *              rw_exit(OBJECT_RWLOCK(op));
619  *
620  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
621  *              list_link_replace(&op->o_link_node, &np->o_link_node);
622  *              mutex_exit(&container->c_objects_lock);
623  *              container_rele(container);
624  *              return (KMEM_CBRC_YES);
625  *      }
626  *
627  * Note that object_move() must invalidate the designated o_container pointer of
628  * the old object in the same way that object_free() does, since kmem will free
629  * the object in response to the KMEM_CBRC_YES return value.
630  *
631  * The lock order in object_move() differs from object_alloc(), which locks
632  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
633  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
634  * not a problem. Holding the lock on the object list in the example above
635  * through the entire callback not only prevents the object from going away, it
636  * also allows you to lock the list elsewhere and know that none of its elements
637  * will move during iteration.
638  *
639  * Adding an explicit hold everywhere an object from the cache is used is tricky
640  * and involves much more change to client code than a cache-specific solution
641  * that leverages existing state to decide whether or not an object is
642  * movable. However, this approach has the advantage that no object remains
643  * immovable for any significant length of time, making it extremely unlikely
644  * that long-lived allocations can continue holding slabs hostage; and it works
645  * for any cache.
646  *
647  * 3. Consolidator Implementation
648  *
649  * Once the client supplies a move function that a) recognizes known objects and
650  * b) avoids moving objects that are actively in use, the remaining work is up
651  * to the consolidator to decide which objects to move and when to issue
652  * callbacks.
653  *
654  * The consolidator relies on the fact that a cache's slabs are ordered by
655  * usage. Each slab has a fixed number of objects. Depending on the slab's
656  * "color" (the offset of the first object from the beginning of the slab;
657  * offsets are staggered to mitigate false sharing of cache lines) it is either
658  * the maximum number of objects per slab determined at cache creation time or
659  * else the number closest to the maximum that fits within the space remaining
660  * after the initial offset. A completely allocated slab may contribute some
661  * internal fragmentation (per-slab overhead) but no external fragmentation, so
662  * it is of no interest to the consolidator. At the other extreme, slabs whose
663  * objects have all been freed to the slab are released to the virtual memory
664  * (VM) subsystem (objects freed to magazines are still allocated as far as the
665  * slab is concerned). External fragmentation exists when there are slabs
666  * somewhere between these extremes. A partial slab has at least one but not all
667  * of its objects allocated. The more partial slabs, and the fewer allocated
668  * objects on each of them, the higher the fragmentation. Hence the
669  * consolidator's overall strategy is to reduce the number of partial slabs by
670  * moving allocated objects from the least allocated slabs to the most allocated
671  * slabs.
672  *
673  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
674  * slabs are kept separately in an unordered list. Since the majority of slabs
675  * tend to be completely allocated (a typical unfragmented cache may have
676  * thousands of complete slabs and only a single partial slab), separating
677  * complete slabs improves the efficiency of partial slab ordering, since the
678  * complete slabs do not affect the depth or balance of the AVL tree. This
679  * ordered sequence of partial slabs acts as a "free list" supplying objects for
680  * allocation requests.
681  *
682  * Objects are always allocated from the first partial slab in the free list,
683  * where the allocation is most likely to eliminate a partial slab (by
684  * completely allocating it). Conversely, when a single object from a completely
685  * allocated slab is freed to the slab, that slab is added to the front of the
686  * free list. Since most free list activity involves highly allocated slabs
687  * coming and going at the front of the list, slabs tend naturally toward the
688  * ideal order: highly allocated at the front, sparsely allocated at the back.
689  * Slabs with few allocated objects are likely to become completely free if they
690  * keep a safe distance away from the front of the free list. Slab misorders
691  * interfere with the natural tendency of slabs to become completely free or
692  * completely allocated. For example, a slab with a single allocated object
693  * needs only a single free to escape the cache; its natural desire is
694  * frustrated when it finds itself at the front of the list where a second
695  * allocation happens just before the free could have released it. Another slab
696  * with all but one object allocated might have supplied the buffer instead, so
697  * that both (as opposed to neither) of the slabs would have been taken off the
698  * free list.
699  *
700  * Although slabs tend naturally toward the ideal order, misorders allowed by a
701  * simple list implementation defeat the consolidator's strategy of merging
702  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
703  * needs another way to fix misorders to optimize its callback strategy. One
704  * approach is to periodically scan a limited number of slabs, advancing a
705  * marker to hold the current scan position, and to move extreme misorders to
706  * the front or back of the free list and to the front or back of the current
707  * scan range. By making consecutive scan ranges overlap by one slab, the least
708  * allocated slab in the current range can be carried along from the end of one
709  * scan to the start of the next.
710  *
711  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
712  * task, however. Since most of the cache's activity is in the magazine layer,
713  * and allocations from the slab layer represent only a startup cost, the
714  * overhead of maintaining a balanced tree is not a significant concern compared
715  * to the opportunity of reducing complexity by eliminating the partial slab
716  * scanner just described. The overhead of an AVL tree is minimized by
717  * maintaining only partial slabs in the tree and keeping completely allocated
718  * slabs separately in a list. To avoid increasing the size of the slab
719  * structure the AVL linkage pointers are reused for the slab's list linkage,
720  * since the slab will always be either partial or complete, never stored both
721  * ways at the same time. To further minimize the overhead of the AVL tree the
722  * compare function that orders partial slabs by usage divides the range of
723  * allocated object counts into bins such that counts within the same bin are
724  * considered equal. Binning partial slabs makes it less likely that allocating
725  * or freeing a single object will change the slab's order, requiring a tree
726  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
727  * requiring some rebalancing of the tree). Allocation counts closest to
728  * completely free and completely allocated are left unbinned (finely sorted) to
729  * better support the consolidator's strategy of merging slabs at either
730  * extreme.
731  *
732  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
733  *
734  * The consolidator piggybacks on the kmem maintenance thread and is called on
735  * the same interval as kmem_cache_update(), once per cache every fifteen
736  * seconds. kmem maintains a running count of unallocated objects in the slab
737  * layer (cache_bufslab). The consolidator checks whether that number exceeds
738  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
739  * there is a significant number of slabs in the cache (arbitrarily a minimum
740  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
741  * working set are included in the assessment, and magazines in the depot are
742  * reaped if those objects would lift cache_bufslab above the fragmentation
743  * threshold. Once the consolidator decides that a cache is fragmented, it looks
744  * for a candidate slab to reclaim, starting at the end of the partial slab free
745  * list and scanning backwards. At first the consolidator is choosy: only a slab
746  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
747  * single allocated object, regardless of percentage). If there is difficulty
748  * finding a candidate slab, kmem raises the allocation threshold incrementally,
749  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
750  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
751  * even in the worst case of every slab in the cache being almost 7/8 allocated.
752  * The threshold can also be lowered incrementally when candidate slabs are easy
753  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
754  * is no longer fragmented.
755  *
756  * 3.2 Generating Callbacks
757  *
758  * Once an eligible slab is chosen, a callback is generated for every allocated
759  * object on the slab, in the hope that the client will move everything off the
760  * slab and make it reclaimable. Objects selected as move destinations are
761  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
762  * order (most allocated at the front, least allocated at the back) and a
763  * cooperative client, the consolidator will succeed in removing slabs from both
764  * ends of the free list, completely allocating on the one hand and completely
765  * freeing on the other. Objects selected as move destinations are allocated in
766  * the kmem maintenance thread where move requests are enqueued. A separate
767  * callback thread removes pending callbacks from the queue and calls the
768  * client. The separate thread ensures that client code (the move function) does
769  * not interfere with internal kmem maintenance tasks. A map of pending
770  * callbacks keyed by object address (the object to be moved) is checked to
771  * ensure that duplicate callbacks are not generated for the same object.
772  * Allocating the move destination (the object to move to) prevents subsequent
773  * callbacks from selecting the same destination as an earlier pending callback.
774  *
775  * Move requests can also be generated by kmem_cache_reap() when the system is
776  * desperate for memory and by kmem_cache_move_notify(), called by the client to
777  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
778  * The map of pending callbacks is protected by the same lock that protects the
779  * slab layer.
780  *
781  * When the system is desperate for memory, kmem does not bother to determine
782  * whether or not the cache exceeds the fragmentation threshold, but tries to
783  * consolidate as many slabs as possible. Normally, the consolidator chews
784  * slowly, one sparsely allocated slab at a time during each maintenance
785  * interval that the cache is fragmented. When desperate, the consolidator
786  * starts at the last partial slab and enqueues callbacks for every allocated
787  * object on every partial slab, working backwards until it reaches the first
788  * partial slab. The first partial slab, meanwhile, advances in pace with the
789  * consolidator as allocations to supply move destinations for the enqueued
790  * callbacks use up the highly allocated slabs at the front of the free list.
791  * Ideally, the overgrown free list collapses like an accordion, starting at
792  * both ends and ending at the center with a single partial slab.
793  *
794  * 3.3 Client Responses
795  *
796  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
797  * marks the slab that supplied the stuck object non-reclaimable and moves it to
798  * front of the free list. The slab remains marked as long as it remains on the
799  * free list, and it appears more allocated to the partial slab compare function
800  * than any unmarked slab, no matter how many of its objects are allocated.
801  * Since even one immovable object ties up the entire slab, the goal is to
802  * completely allocate any slab that cannot be completely freed. kmem does not
803  * bother generating callbacks to move objects from a marked slab unless the
804  * system is desperate.
805  *
806  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
807  * slab. If the client responds LATER too many times, kmem disbelieves and
808  * treats the response as a NO. The count is cleared when the slab is taken off
809  * the partial slab list or when the client moves one of the slab's objects.
810  *
811  * 4. Observability
812  *
813  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
814  * the ::kmem_slabs dcmd. For a complete description of the command, enter
815  * '::help kmem_slabs' at the mdb prompt.
816  */
817 
818 #include <sys/kmem_impl.h>
819 #include <sys/vmem_impl.h>
820 #include <sys/param.h>
821 #include <sys/sysmacros.h>
822 #include <sys/vm.h>
823 #include <sys/proc.h>
824 #include <sys/tuneable.h>
825 #include <sys/systm.h>
826 #include <sys/cmn_err.h>
827 #include <sys/debug.h>
828 #include <sys/sdt.h>
829 #include <sys/mutex.h>
830 #include <sys/bitmap.h>
831 #include <sys/atomic.h>
832 #include <sys/kobj.h>
833 #include <sys/disp.h>
834 #include <vm/seg_kmem.h>
835 #include <sys/log.h>
836 #include <sys/callb.h>
837 #include <sys/taskq.h>
838 #include <sys/modctl.h>
839 #include <sys/reboot.h>
840 #include <sys/id32.h>
841 #include <sys/zone.h>
842 #include <sys/netstack.h>
843 #ifdef	DEBUG
844 #include <sys/random.h>
845 #endif
846 
847 extern void streams_msg_init(void);
848 extern int segkp_fromheap;
849 extern void segkp_cache_free(void);
850 extern int callout_init_done;
851 
852 struct kmem_cache_kstat {
853 	kstat_named_t	kmc_buf_size;
854 	kstat_named_t	kmc_align;
855 	kstat_named_t	kmc_chunk_size;
856 	kstat_named_t	kmc_slab_size;
857 	kstat_named_t	kmc_alloc;
858 	kstat_named_t	kmc_alloc_fail;
859 	kstat_named_t	kmc_free;
860 	kstat_named_t	kmc_depot_alloc;
861 	kstat_named_t	kmc_depot_free;
862 	kstat_named_t	kmc_depot_contention;
863 	kstat_named_t	kmc_slab_alloc;
864 	kstat_named_t	kmc_slab_free;
865 	kstat_named_t	kmc_buf_constructed;
866 	kstat_named_t	kmc_buf_avail;
867 	kstat_named_t	kmc_buf_inuse;
868 	kstat_named_t	kmc_buf_total;
869 	kstat_named_t	kmc_buf_max;
870 	kstat_named_t	kmc_slab_create;
871 	kstat_named_t	kmc_slab_destroy;
872 	kstat_named_t	kmc_vmem_source;
873 	kstat_named_t	kmc_hash_size;
874 	kstat_named_t	kmc_hash_lookup_depth;
875 	kstat_named_t	kmc_hash_rescale;
876 	kstat_named_t	kmc_full_magazines;
877 	kstat_named_t	kmc_empty_magazines;
878 	kstat_named_t	kmc_magazine_size;
879 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
880 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
881 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
882 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
883 	kstat_named_t	kmc_move_yes;
884 	kstat_named_t	kmc_move_no;
885 	kstat_named_t	kmc_move_later;
886 	kstat_named_t	kmc_move_dont_need;
887 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
888 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
889 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
890 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
891 } kmem_cache_kstat = {
892 	{ "buf_size",		KSTAT_DATA_UINT64 },
893 	{ "align",		KSTAT_DATA_UINT64 },
894 	{ "chunk_size",		KSTAT_DATA_UINT64 },
895 	{ "slab_size",		KSTAT_DATA_UINT64 },
896 	{ "alloc",		KSTAT_DATA_UINT64 },
897 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
898 	{ "free",		KSTAT_DATA_UINT64 },
899 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
900 	{ "depot_free",		KSTAT_DATA_UINT64 },
901 	{ "depot_contention",	KSTAT_DATA_UINT64 },
902 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
903 	{ "slab_free",		KSTAT_DATA_UINT64 },
904 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
905 	{ "buf_avail",		KSTAT_DATA_UINT64 },
906 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
907 	{ "buf_total",		KSTAT_DATA_UINT64 },
908 	{ "buf_max",		KSTAT_DATA_UINT64 },
909 	{ "slab_create",	KSTAT_DATA_UINT64 },
910 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
911 	{ "vmem_source",	KSTAT_DATA_UINT64 },
912 	{ "hash_size",		KSTAT_DATA_UINT64 },
913 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
914 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
915 	{ "full_magazines",	KSTAT_DATA_UINT64 },
916 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
917 	{ "magazine_size",	KSTAT_DATA_UINT64 },
918 	{ "reap",		KSTAT_DATA_UINT64 },
919 	{ "defrag",		KSTAT_DATA_UINT64 },
920 	{ "scan",		KSTAT_DATA_UINT64 },
921 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
922 	{ "move_yes",		KSTAT_DATA_UINT64 },
923 	{ "move_no",		KSTAT_DATA_UINT64 },
924 	{ "move_later",		KSTAT_DATA_UINT64 },
925 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
926 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
927 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
928 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
929 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
930 };
931 
932 static kmutex_t kmem_cache_kstat_lock;
933 
934 /*
935  * The default set of caches to back kmem_alloc().
936  * These sizes should be reevaluated periodically.
937  *
938  * We want allocations that are multiples of the coherency granularity
939  * (64 bytes) to be satisfied from a cache which is a multiple of 64
940  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
941  * the next kmem_cache_size greater than or equal to it must be a
942  * multiple of 64.
943  *
944  * We split the table into two sections:  size <= 4k and size > 4k.  This
945  * saves a lot of space and cache footprint in our cache tables.
946  */
947 static const int kmem_alloc_sizes[] = {
948 	1 * 8,
949 	2 * 8,
950 	3 * 8,
951 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
952 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
953 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
954 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
955 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
956 	P2ALIGN(8192 / 7, 64),
957 	P2ALIGN(8192 / 6, 64),
958 	P2ALIGN(8192 / 5, 64),
959 	P2ALIGN(8192 / 4, 64),
960 	P2ALIGN(8192 / 3, 64),
961 	P2ALIGN(8192 / 2, 64),
962 };
963 
964 static const int kmem_big_alloc_sizes[] = {
965 	2 * 4096,	3 * 4096,
966 	2 * 8192,	3 * 8192,
967 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
968 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
969 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
970 	16 * 8192
971 };
972 
973 #define	KMEM_MAXBUF		4096
974 #define	KMEM_BIG_MAXBUF_32BIT	32768
975 #define	KMEM_BIG_MAXBUF		131072
976 
977 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
978 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
979 
980 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
981 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
982 
983 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
984 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
985 
986 static kmem_magtype_t kmem_magtype[] = {
987 	{ 1,	8,	3200,	65536	},
988 	{ 3,	16,	256,	32768	},
989 	{ 7,	32,	64,	16384	},
990 	{ 15,	64,	0,	8192	},
991 	{ 31,	64,	0,	4096	},
992 	{ 47,	64,	0,	2048	},
993 	{ 63,	64,	0,	1024	},
994 	{ 95,	64,	0,	512	},
995 	{ 143,	64,	0,	0	},
996 };
997 
998 static uint32_t kmem_reaping;
999 static uint32_t kmem_reaping_idspace;
1000 
1001 /*
1002  * kmem tunables
1003  */
1004 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
1005 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
1006 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
1007 int kmem_panic = 1;		/* whether to panic on error */
1008 int kmem_logging = 1;		/* kmem_log_enter() override */
1009 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
1010 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
1011 size_t kmem_content_log_size;	/* content log size [2% of memory] */
1012 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
1013 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
1014 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1015 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1016 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1017 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1018 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1019 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1020 
1021 #ifdef _LP64
1022 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1023 #else
1024 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1025 #endif
1026 
1027 #ifdef DEBUG
1028 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1029 #else
1030 int kmem_flags = 0;
1031 #endif
1032 int kmem_ready;
1033 
1034 static kmem_cache_t	*kmem_slab_cache;
1035 static kmem_cache_t	*kmem_bufctl_cache;
1036 static kmem_cache_t	*kmem_bufctl_audit_cache;
1037 
1038 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1039 static list_t		kmem_caches;
1040 
1041 static taskq_t		*kmem_taskq;
1042 static kmutex_t		kmem_flags_lock;
1043 static vmem_t		*kmem_metadata_arena;
1044 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1045 static vmem_t		*kmem_cache_arena;
1046 static vmem_t		*kmem_hash_arena;
1047 static vmem_t		*kmem_log_arena;
1048 static vmem_t		*kmem_oversize_arena;
1049 static vmem_t		*kmem_va_arena;
1050 static vmem_t		*kmem_default_arena;
1051 static vmem_t		*kmem_firewall_va_arena;
1052 static vmem_t		*kmem_firewall_arena;
1053 
1054 /*
1055  * kmem slab consolidator thresholds (tunables)
1056  */
1057 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1058 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1059 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1060 /*
1061  * Maximum number of slabs from which to move buffers during a single
1062  * maintenance interval while the system is not low on memory.
1063  */
1064 size_t kmem_reclaim_max_slabs = 1;
1065 /*
1066  * Number of slabs to scan backwards from the end of the partial slab list
1067  * when searching for buffers to relocate.
1068  */
1069 size_t kmem_reclaim_scan_range = 12;
1070 
1071 /* consolidator knobs */
1072 boolean_t kmem_move_noreap;
1073 boolean_t kmem_move_blocked;
1074 boolean_t kmem_move_fulltilt;
1075 boolean_t kmem_move_any_partial;
1076 
1077 #ifdef	DEBUG
1078 /*
1079  * kmem consolidator debug tunables:
1080  * Ensure code coverage by occasionally running the consolidator even when the
1081  * caches are not fragmented (they may never be). These intervals are mean time
1082  * in cache maintenance intervals (kmem_cache_update).
1083  */
1084 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1085 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1086 #endif	/* DEBUG */
1087 
1088 static kmem_cache_t	*kmem_defrag_cache;
1089 static kmem_cache_t	*kmem_move_cache;
1090 static taskq_t		*kmem_move_taskq;
1091 
1092 static void kmem_cache_scan(kmem_cache_t *);
1093 static void kmem_cache_defrag(kmem_cache_t *);
1094 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1095 
1096 
1097 kmem_log_header_t	*kmem_transaction_log;
1098 kmem_log_header_t	*kmem_content_log;
1099 kmem_log_header_t	*kmem_failure_log;
1100 kmem_log_header_t	*kmem_slab_log;
1101 
1102 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1103 
1104 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1105 	if ((count) > 0) {						\
1106 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1107 		pc_t *_e;						\
1108 		/* memmove() the old entries down one notch */		\
1109 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1110 			*_e = *(_e - 1);				\
1111 		*_s = (uintptr_t)(caller);				\
1112 	}
1113 
1114 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1115 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1116 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1117 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1118 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1119 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1120 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1121 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1122 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1123 
1124 struct {
1125 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1126 	int		kmp_error;	/* type of kmem error */
1127 	void		*kmp_buffer;	/* buffer that induced panic */
1128 	void		*kmp_realbuf;	/* real start address for buffer */
1129 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1130 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1131 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1132 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1133 } kmem_panic_info;
1134 
1135 
1136 static void
1137 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1138 {
1139 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1140 	uint64_t *buf = buf_arg;
1141 
1142 	while (buf < bufend)
1143 		*buf++ = pattern;
1144 }
1145 
1146 static void *
1147 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1148 {
1149 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1150 	uint64_t *buf;
1151 
1152 	for (buf = buf_arg; buf < bufend; buf++)
1153 		if (*buf != pattern)
1154 			return (buf);
1155 	return (NULL);
1156 }
1157 
1158 static void *
1159 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1160 {
1161 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1162 	uint64_t *buf;
1163 
1164 	for (buf = buf_arg; buf < bufend; buf++) {
1165 		if (*buf != old) {
1166 			copy_pattern(old, buf_arg,
1167 			    (char *)buf - (char *)buf_arg);
1168 			return (buf);
1169 		}
1170 		*buf = new;
1171 	}
1172 
1173 	return (NULL);
1174 }
1175 
1176 static void
1177 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1178 {
1179 	kmem_cache_t *cp;
1180 
1181 	mutex_enter(&kmem_cache_lock);
1182 	for (cp = list_head(&kmem_caches); cp != NULL;
1183 	    cp = list_next(&kmem_caches, cp))
1184 		if (tq != NULL)
1185 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1186 			    tqflag);
1187 		else
1188 			func(cp);
1189 	mutex_exit(&kmem_cache_lock);
1190 }
1191 
1192 static void
1193 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1194 {
1195 	kmem_cache_t *cp;
1196 
1197 	mutex_enter(&kmem_cache_lock);
1198 	for (cp = list_head(&kmem_caches); cp != NULL;
1199 	    cp = list_next(&kmem_caches, cp)) {
1200 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1201 			continue;
1202 		if (tq != NULL)
1203 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1204 			    tqflag);
1205 		else
1206 			func(cp);
1207 	}
1208 	mutex_exit(&kmem_cache_lock);
1209 }
1210 
1211 /*
1212  * Debugging support.  Given a buffer address, find its slab.
1213  */
1214 static kmem_slab_t *
1215 kmem_findslab(kmem_cache_t *cp, void *buf)
1216 {
1217 	kmem_slab_t *sp;
1218 
1219 	mutex_enter(&cp->cache_lock);
1220 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1221 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1222 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1223 			mutex_exit(&cp->cache_lock);
1224 			return (sp);
1225 		}
1226 	}
1227 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1228 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1229 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1230 			mutex_exit(&cp->cache_lock);
1231 			return (sp);
1232 		}
1233 	}
1234 	mutex_exit(&cp->cache_lock);
1235 
1236 	return (NULL);
1237 }
1238 
1239 static void
1240 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1241 {
1242 	kmem_buftag_t *btp = NULL;
1243 	kmem_bufctl_t *bcp = NULL;
1244 	kmem_cache_t *cp = cparg;
1245 	kmem_slab_t *sp;
1246 	uint64_t *off;
1247 	void *buf = bufarg;
1248 
1249 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1250 
1251 	kmem_panic_info.kmp_timestamp = gethrtime();
1252 
1253 	sp = kmem_findslab(cp, buf);
1254 	if (sp == NULL) {
1255 		for (cp = list_tail(&kmem_caches); cp != NULL;
1256 		    cp = list_prev(&kmem_caches, cp)) {
1257 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1258 				break;
1259 		}
1260 	}
1261 
1262 	if (sp == NULL) {
1263 		cp = NULL;
1264 		error = KMERR_BADADDR;
1265 	} else {
1266 		if (cp != cparg)
1267 			error = KMERR_BADCACHE;
1268 		else
1269 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1270 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1271 		if (buf != bufarg)
1272 			error = KMERR_BADBASE;
1273 		if (cp->cache_flags & KMF_BUFTAG)
1274 			btp = KMEM_BUFTAG(cp, buf);
1275 		if (cp->cache_flags & KMF_HASH) {
1276 			mutex_enter(&cp->cache_lock);
1277 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1278 				if (bcp->bc_addr == buf)
1279 					break;
1280 			mutex_exit(&cp->cache_lock);
1281 			if (bcp == NULL && btp != NULL)
1282 				bcp = btp->bt_bufctl;
1283 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1284 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1285 			    bcp->bc_addr != buf) {
1286 				error = KMERR_BADBUFCTL;
1287 				bcp = NULL;
1288 			}
1289 		}
1290 	}
1291 
1292 	kmem_panic_info.kmp_error = error;
1293 	kmem_panic_info.kmp_buffer = bufarg;
1294 	kmem_panic_info.kmp_realbuf = buf;
1295 	kmem_panic_info.kmp_cache = cparg;
1296 	kmem_panic_info.kmp_realcache = cp;
1297 	kmem_panic_info.kmp_slab = sp;
1298 	kmem_panic_info.kmp_bufctl = bcp;
1299 
1300 	printf("kernel memory allocator: ");
1301 
1302 	switch (error) {
1303 
1304 	case KMERR_MODIFIED:
1305 		printf("buffer modified after being freed\n");
1306 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1307 		if (off == NULL)	/* shouldn't happen */
1308 			off = buf;
1309 		printf("modification occurred at offset 0x%lx "
1310 		    "(0x%llx replaced by 0x%llx)\n",
1311 		    (uintptr_t)off - (uintptr_t)buf,
1312 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1313 		break;
1314 
1315 	case KMERR_REDZONE:
1316 		printf("redzone violation: write past end of buffer\n");
1317 		break;
1318 
1319 	case KMERR_BADADDR:
1320 		printf("invalid free: buffer not in cache\n");
1321 		break;
1322 
1323 	case KMERR_DUPFREE:
1324 		printf("duplicate free: buffer freed twice\n");
1325 		break;
1326 
1327 	case KMERR_BADBUFTAG:
1328 		printf("boundary tag corrupted\n");
1329 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1330 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1331 		    KMEM_BUFTAG_FREE);
1332 		break;
1333 
1334 	case KMERR_BADBUFCTL:
1335 		printf("bufctl corrupted\n");
1336 		break;
1337 
1338 	case KMERR_BADCACHE:
1339 		printf("buffer freed to wrong cache\n");
1340 		printf("buffer was allocated from %s,\n", cp->cache_name);
1341 		printf("caller attempting free to %s.\n", cparg->cache_name);
1342 		break;
1343 
1344 	case KMERR_BADSIZE:
1345 		printf("bad free: free size (%u) != alloc size (%u)\n",
1346 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1347 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1348 		break;
1349 
1350 	case KMERR_BADBASE:
1351 		printf("bad free: free address (%p) != alloc address (%p)\n",
1352 		    bufarg, buf);
1353 		break;
1354 	}
1355 
1356 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1357 	    bufarg, (void *)bcp, cparg->cache_name);
1358 
1359 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1360 	    error != KMERR_BADBUFCTL) {
1361 		int d;
1362 		timestruc_t ts;
1363 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1364 
1365 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1366 		printf("previous transaction on buffer %p:\n", buf);
1367 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1368 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1369 		    (void *)sp, cp->cache_name);
1370 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1371 			ulong_t off;
1372 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1373 			printf("%s+%lx\n", sym ? sym : "?", off);
1374 		}
1375 	}
1376 	if (kmem_panic > 0)
1377 		panic("kernel heap corruption detected");
1378 	if (kmem_panic == 0)
1379 		debug_enter(NULL);
1380 	kmem_logging = 1;	/* resume logging */
1381 }
1382 
1383 static kmem_log_header_t *
1384 kmem_log_init(size_t logsize)
1385 {
1386 	kmem_log_header_t *lhp;
1387 	int nchunks = 4 * max_ncpus;
1388 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1389 	int i;
1390 
1391 	/*
1392 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1393 	 * to prevent false sharing of cache lines.
1394 	 */
1395 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1396 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1397 	    NULL, NULL, VM_SLEEP);
1398 	bzero(lhp, lhsize);
1399 
1400 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1401 	lhp->lh_nchunks = nchunks;
1402 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1403 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1404 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1405 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1406 	    nchunks * sizeof (int), VM_SLEEP);
1407 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1408 
1409 	for (i = 0; i < max_ncpus; i++) {
1410 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1411 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1412 		clhp->clh_chunk = i;
1413 	}
1414 
1415 	for (i = max_ncpus; i < nchunks; i++)
1416 		lhp->lh_free[i] = i;
1417 
1418 	lhp->lh_head = max_ncpus;
1419 	lhp->lh_tail = 0;
1420 
1421 	return (lhp);
1422 }
1423 
1424 static void *
1425 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1426 {
1427 	void *logspace;
1428 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1429 
1430 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1431 		return (NULL);
1432 
1433 	mutex_enter(&clhp->clh_lock);
1434 	clhp->clh_hits++;
1435 	if (size > clhp->clh_avail) {
1436 		mutex_enter(&lhp->lh_lock);
1437 		lhp->lh_hits++;
1438 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1439 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1440 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1441 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1442 		clhp->clh_current = lhp->lh_base +
1443 		    clhp->clh_chunk * lhp->lh_chunksize;
1444 		clhp->clh_avail = lhp->lh_chunksize;
1445 		if (size > lhp->lh_chunksize)
1446 			size = lhp->lh_chunksize;
1447 		mutex_exit(&lhp->lh_lock);
1448 	}
1449 	logspace = clhp->clh_current;
1450 	clhp->clh_current += size;
1451 	clhp->clh_avail -= size;
1452 	bcopy(data, logspace, size);
1453 	mutex_exit(&clhp->clh_lock);
1454 	return (logspace);
1455 }
1456 
1457 #define	KMEM_AUDIT(lp, cp, bcp)						\
1458 {									\
1459 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1460 	_bcp->bc_timestamp = gethrtime();				\
1461 	_bcp->bc_thread = curthread;					\
1462 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1463 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1464 }
1465 
1466 static void
1467 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1468     kmem_slab_t *sp, void *addr)
1469 {
1470 	kmem_bufctl_audit_t bca;
1471 
1472 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1473 	bca.bc_addr = addr;
1474 	bca.bc_slab = sp;
1475 	bca.bc_cache = cp;
1476 	KMEM_AUDIT(lp, cp, &bca);
1477 }
1478 
1479 /*
1480  * Create a new slab for cache cp.
1481  */
1482 static kmem_slab_t *
1483 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1484 {
1485 	size_t slabsize = cp->cache_slabsize;
1486 	size_t chunksize = cp->cache_chunksize;
1487 	int cache_flags = cp->cache_flags;
1488 	size_t color, chunks;
1489 	char *buf, *slab;
1490 	kmem_slab_t *sp;
1491 	kmem_bufctl_t *bcp;
1492 	vmem_t *vmp = cp->cache_arena;
1493 
1494 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1495 
1496 	color = cp->cache_color + cp->cache_align;
1497 	if (color > cp->cache_maxcolor)
1498 		color = cp->cache_mincolor;
1499 	cp->cache_color = color;
1500 
1501 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1502 
1503 	if (slab == NULL)
1504 		goto vmem_alloc_failure;
1505 
1506 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1507 
1508 	/*
1509 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1510 	 * consolidator depends (for correctness) on slabs being initialized
1511 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1512 	 * clients to distinguish uninitialized memory from known objects).
1513 	 */
1514 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1515 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1516 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1517 
1518 	if (cache_flags & KMF_HASH) {
1519 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1520 			goto slab_alloc_failure;
1521 		chunks = (slabsize - color) / chunksize;
1522 	} else {
1523 		sp = KMEM_SLAB(cp, slab);
1524 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1525 	}
1526 
1527 	sp->slab_cache	= cp;
1528 	sp->slab_head	= NULL;
1529 	sp->slab_refcnt	= 0;
1530 	sp->slab_base	= buf = slab + color;
1531 	sp->slab_chunks	= chunks;
1532 	sp->slab_stuck_offset = (uint32_t)-1;
1533 	sp->slab_later_count = 0;
1534 	sp->slab_flags = 0;
1535 
1536 	ASSERT(chunks > 0);
1537 	while (chunks-- != 0) {
1538 		if (cache_flags & KMF_HASH) {
1539 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1540 			if (bcp == NULL)
1541 				goto bufctl_alloc_failure;
1542 			if (cache_flags & KMF_AUDIT) {
1543 				kmem_bufctl_audit_t *bcap =
1544 				    (kmem_bufctl_audit_t *)bcp;
1545 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1546 				bcap->bc_cache = cp;
1547 			}
1548 			bcp->bc_addr = buf;
1549 			bcp->bc_slab = sp;
1550 		} else {
1551 			bcp = KMEM_BUFCTL(cp, buf);
1552 		}
1553 		if (cache_flags & KMF_BUFTAG) {
1554 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1555 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1556 			btp->bt_bufctl = bcp;
1557 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1558 			if (cache_flags & KMF_DEADBEEF) {
1559 				copy_pattern(KMEM_FREE_PATTERN, buf,
1560 				    cp->cache_verify);
1561 			}
1562 		}
1563 		bcp->bc_next = sp->slab_head;
1564 		sp->slab_head = bcp;
1565 		buf += chunksize;
1566 	}
1567 
1568 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1569 
1570 	return (sp);
1571 
1572 bufctl_alloc_failure:
1573 
1574 	while ((bcp = sp->slab_head) != NULL) {
1575 		sp->slab_head = bcp->bc_next;
1576 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1577 	}
1578 	kmem_cache_free(kmem_slab_cache, sp);
1579 
1580 slab_alloc_failure:
1581 
1582 	vmem_free(vmp, slab, slabsize);
1583 
1584 vmem_alloc_failure:
1585 
1586 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1587 	atomic_inc_64(&cp->cache_alloc_fail);
1588 
1589 	return (NULL);
1590 }
1591 
1592 /*
1593  * Destroy a slab.
1594  */
1595 static void
1596 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1597 {
1598 	vmem_t *vmp = cp->cache_arena;
1599 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1600 
1601 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1602 	ASSERT(sp->slab_refcnt == 0);
1603 
1604 	if (cp->cache_flags & KMF_HASH) {
1605 		kmem_bufctl_t *bcp;
1606 		while ((bcp = sp->slab_head) != NULL) {
1607 			sp->slab_head = bcp->bc_next;
1608 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1609 		}
1610 		kmem_cache_free(kmem_slab_cache, sp);
1611 	}
1612 	vmem_free(vmp, slab, cp->cache_slabsize);
1613 }
1614 
1615 static void *
1616 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1617 {
1618 	kmem_bufctl_t *bcp, **hash_bucket;
1619 	void *buf;
1620 	boolean_t new_slab = (sp->slab_refcnt == 0);
1621 
1622 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1623 	/*
1624 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1625 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1626 	 * slab is newly created.
1627 	 */
1628 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1629 	    (sp == avl_first(&cp->cache_partial_slabs))));
1630 	ASSERT(sp->slab_cache == cp);
1631 
1632 	cp->cache_slab_alloc++;
1633 	cp->cache_bufslab--;
1634 	sp->slab_refcnt++;
1635 
1636 	bcp = sp->slab_head;
1637 	sp->slab_head = bcp->bc_next;
1638 
1639 	if (cp->cache_flags & KMF_HASH) {
1640 		/*
1641 		 * Add buffer to allocated-address hash table.
1642 		 */
1643 		buf = bcp->bc_addr;
1644 		hash_bucket = KMEM_HASH(cp, buf);
1645 		bcp->bc_next = *hash_bucket;
1646 		*hash_bucket = bcp;
1647 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1648 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1649 		}
1650 	} else {
1651 		buf = KMEM_BUF(cp, bcp);
1652 	}
1653 
1654 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1655 
1656 	if (sp->slab_head == NULL) {
1657 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1658 		if (new_slab) {
1659 			ASSERT(sp->slab_chunks == 1);
1660 		} else {
1661 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1662 			avl_remove(&cp->cache_partial_slabs, sp);
1663 			sp->slab_later_count = 0; /* clear history */
1664 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1665 			sp->slab_stuck_offset = (uint32_t)-1;
1666 		}
1667 		list_insert_head(&cp->cache_complete_slabs, sp);
1668 		cp->cache_complete_slab_count++;
1669 		return (buf);
1670 	}
1671 
1672 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1673 	/*
1674 	 * Peek to see if the magazine layer is enabled before
1675 	 * we prefill.  We're not holding the cpu cache lock,
1676 	 * so the peek could be wrong, but there's no harm in it.
1677 	 */
1678 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1679 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1680 		kmem_slab_prefill(cp, sp);
1681 		return (buf);
1682 	}
1683 
1684 	if (new_slab) {
1685 		avl_add(&cp->cache_partial_slabs, sp);
1686 		return (buf);
1687 	}
1688 
1689 	/*
1690 	 * The slab is now more allocated than it was, so the
1691 	 * order remains unchanged.
1692 	 */
1693 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1694 	return (buf);
1695 }
1696 
1697 /*
1698  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1699  */
1700 static void *
1701 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1702 {
1703 	kmem_slab_t *sp;
1704 	void *buf;
1705 	boolean_t test_destructor;
1706 
1707 	mutex_enter(&cp->cache_lock);
1708 	test_destructor = (cp->cache_slab_alloc == 0);
1709 	sp = avl_first(&cp->cache_partial_slabs);
1710 	if (sp == NULL) {
1711 		ASSERT(cp->cache_bufslab == 0);
1712 
1713 		/*
1714 		 * The freelist is empty.  Create a new slab.
1715 		 */
1716 		mutex_exit(&cp->cache_lock);
1717 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1718 			return (NULL);
1719 		}
1720 		mutex_enter(&cp->cache_lock);
1721 		cp->cache_slab_create++;
1722 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1723 			cp->cache_bufmax = cp->cache_buftotal;
1724 		cp->cache_bufslab += sp->slab_chunks;
1725 	}
1726 
1727 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1728 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1729 	    (cp->cache_complete_slab_count +
1730 	    avl_numnodes(&cp->cache_partial_slabs) +
1731 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1732 	mutex_exit(&cp->cache_lock);
1733 
1734 	if (test_destructor && cp->cache_destructor != NULL) {
1735 		/*
1736 		 * On the first kmem_slab_alloc(), assert that it is valid to
1737 		 * call the destructor on a newly constructed object without any
1738 		 * client involvement.
1739 		 */
1740 		if ((cp->cache_constructor == NULL) ||
1741 		    cp->cache_constructor(buf, cp->cache_private,
1742 		    kmflag) == 0) {
1743 			cp->cache_destructor(buf, cp->cache_private);
1744 		}
1745 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1746 		    cp->cache_bufsize);
1747 		if (cp->cache_flags & KMF_DEADBEEF) {
1748 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1749 		}
1750 	}
1751 
1752 	return (buf);
1753 }
1754 
1755 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1756 
1757 /*
1758  * Free a raw (unconstructed) buffer to cp's slab layer.
1759  */
1760 static void
1761 kmem_slab_free(kmem_cache_t *cp, void *buf)
1762 {
1763 	kmem_slab_t *sp;
1764 	kmem_bufctl_t *bcp, **prev_bcpp;
1765 
1766 	ASSERT(buf != NULL);
1767 
1768 	mutex_enter(&cp->cache_lock);
1769 	cp->cache_slab_free++;
1770 
1771 	if (cp->cache_flags & KMF_HASH) {
1772 		/*
1773 		 * Look up buffer in allocated-address hash table.
1774 		 */
1775 		prev_bcpp = KMEM_HASH(cp, buf);
1776 		while ((bcp = *prev_bcpp) != NULL) {
1777 			if (bcp->bc_addr == buf) {
1778 				*prev_bcpp = bcp->bc_next;
1779 				sp = bcp->bc_slab;
1780 				break;
1781 			}
1782 			cp->cache_lookup_depth++;
1783 			prev_bcpp = &bcp->bc_next;
1784 		}
1785 	} else {
1786 		bcp = KMEM_BUFCTL(cp, buf);
1787 		sp = KMEM_SLAB(cp, buf);
1788 	}
1789 
1790 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1791 		mutex_exit(&cp->cache_lock);
1792 		kmem_error(KMERR_BADADDR, cp, buf);
1793 		return;
1794 	}
1795 
1796 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1797 		/*
1798 		 * If this is the buffer that prevented the consolidator from
1799 		 * clearing the slab, we can reset the slab flags now that the
1800 		 * buffer is freed. (It makes sense to do this in
1801 		 * kmem_cache_free(), where the client gives up ownership of the
1802 		 * buffer, but on the hot path the test is too expensive.)
1803 		 */
1804 		kmem_slab_move_yes(cp, sp, buf);
1805 	}
1806 
1807 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1808 		if (cp->cache_flags & KMF_CONTENTS)
1809 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1810 			    kmem_log_enter(kmem_content_log, buf,
1811 			    cp->cache_contents);
1812 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1813 	}
1814 
1815 	bcp->bc_next = sp->slab_head;
1816 	sp->slab_head = bcp;
1817 
1818 	cp->cache_bufslab++;
1819 	ASSERT(sp->slab_refcnt >= 1);
1820 
1821 	if (--sp->slab_refcnt == 0) {
1822 		/*
1823 		 * There are no outstanding allocations from this slab,
1824 		 * so we can reclaim the memory.
1825 		 */
1826 		if (sp->slab_chunks == 1) {
1827 			list_remove(&cp->cache_complete_slabs, sp);
1828 			cp->cache_complete_slab_count--;
1829 		} else {
1830 			avl_remove(&cp->cache_partial_slabs, sp);
1831 		}
1832 
1833 		cp->cache_buftotal -= sp->slab_chunks;
1834 		cp->cache_bufslab -= sp->slab_chunks;
1835 		/*
1836 		 * Defer releasing the slab to the virtual memory subsystem
1837 		 * while there is a pending move callback, since we guarantee
1838 		 * that buffers passed to the move callback have only been
1839 		 * touched by kmem or by the client itself. Since the memory
1840 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1841 		 * set at least one of the two lowest order bits, the client can
1842 		 * test those bits in the move callback to determine whether or
1843 		 * not it knows about the buffer (assuming that the client also
1844 		 * sets one of those low order bits whenever it frees a buffer).
1845 		 */
1846 		if (cp->cache_defrag == NULL ||
1847 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1848 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1849 			cp->cache_slab_destroy++;
1850 			mutex_exit(&cp->cache_lock);
1851 			kmem_slab_destroy(cp, sp);
1852 		} else {
1853 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1854 			/*
1855 			 * Slabs are inserted at both ends of the deadlist to
1856 			 * distinguish between slabs freed while move callbacks
1857 			 * are pending (list head) and a slab freed while the
1858 			 * lock is dropped in kmem_move_buffers() (list tail) so
1859 			 * that in both cases slab_destroy() is called from the
1860 			 * right context.
1861 			 */
1862 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1863 				list_insert_tail(deadlist, sp);
1864 			} else {
1865 				list_insert_head(deadlist, sp);
1866 			}
1867 			cp->cache_defrag->kmd_deadcount++;
1868 			mutex_exit(&cp->cache_lock);
1869 		}
1870 		return;
1871 	}
1872 
1873 	if (bcp->bc_next == NULL) {
1874 		/* Transition the slab from completely allocated to partial. */
1875 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1876 		ASSERT(sp->slab_chunks > 1);
1877 		list_remove(&cp->cache_complete_slabs, sp);
1878 		cp->cache_complete_slab_count--;
1879 		avl_add(&cp->cache_partial_slabs, sp);
1880 	} else {
1881 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1882 	}
1883 
1884 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1885 	    (cp->cache_complete_slab_count +
1886 	    avl_numnodes(&cp->cache_partial_slabs) +
1887 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1888 	mutex_exit(&cp->cache_lock);
1889 }
1890 
1891 /*
1892  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1893  */
1894 static int
1895 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1896     caddr_t caller)
1897 {
1898 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1899 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1900 	uint32_t mtbf;
1901 
1902 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1903 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1904 		return (-1);
1905 	}
1906 
1907 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1908 
1909 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1910 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1911 		return (-1);
1912 	}
1913 
1914 	if (cp->cache_flags & KMF_DEADBEEF) {
1915 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1916 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1917 				kmem_error(KMERR_MODIFIED, cp, buf);
1918 				return (-1);
1919 			}
1920 			if (cp->cache_constructor != NULL)
1921 				*(uint64_t *)buf = btp->bt_redzone;
1922 			else
1923 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1924 		} else {
1925 			construct = 1;
1926 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1927 			    KMEM_UNINITIALIZED_PATTERN, buf,
1928 			    cp->cache_verify)) {
1929 				kmem_error(KMERR_MODIFIED, cp, buf);
1930 				return (-1);
1931 			}
1932 		}
1933 	}
1934 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1935 
1936 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1937 	    gethrtime() % mtbf == 0 &&
1938 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1939 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1940 		if (!construct && cp->cache_destructor != NULL)
1941 			cp->cache_destructor(buf, cp->cache_private);
1942 	} else {
1943 		mtbf = 0;
1944 	}
1945 
1946 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1947 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1948 		atomic_inc_64(&cp->cache_alloc_fail);
1949 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1950 		if (cp->cache_flags & KMF_DEADBEEF)
1951 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1952 		kmem_slab_free(cp, buf);
1953 		return (1);
1954 	}
1955 
1956 	if (cp->cache_flags & KMF_AUDIT) {
1957 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1958 	}
1959 
1960 	if ((cp->cache_flags & KMF_LITE) &&
1961 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1962 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1963 	}
1964 
1965 	return (0);
1966 }
1967 
1968 static int
1969 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1970 {
1971 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1972 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1973 	kmem_slab_t *sp;
1974 
1975 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1976 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1977 			kmem_error(KMERR_DUPFREE, cp, buf);
1978 			return (-1);
1979 		}
1980 		sp = kmem_findslab(cp, buf);
1981 		if (sp == NULL || sp->slab_cache != cp)
1982 			kmem_error(KMERR_BADADDR, cp, buf);
1983 		else
1984 			kmem_error(KMERR_REDZONE, cp, buf);
1985 		return (-1);
1986 	}
1987 
1988 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1989 
1990 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1991 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1992 		return (-1);
1993 	}
1994 
1995 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
1996 		kmem_error(KMERR_REDZONE, cp, buf);
1997 		return (-1);
1998 	}
1999 
2000 	if (cp->cache_flags & KMF_AUDIT) {
2001 		if (cp->cache_flags & KMF_CONTENTS)
2002 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2003 			    buf, cp->cache_contents);
2004 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2005 	}
2006 
2007 	if ((cp->cache_flags & KMF_LITE) &&
2008 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2009 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2010 	}
2011 
2012 	if (cp->cache_flags & KMF_DEADBEEF) {
2013 		if (cp->cache_flags & KMF_LITE)
2014 			btp->bt_redzone = *(uint64_t *)buf;
2015 		else if (cp->cache_destructor != NULL)
2016 			cp->cache_destructor(buf, cp->cache_private);
2017 
2018 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2019 	}
2020 
2021 	return (0);
2022 }
2023 
2024 /*
2025  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2026  */
2027 static void
2028 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2029 {
2030 	int round;
2031 
2032 	ASSERT(!list_link_active(&cp->cache_link) ||
2033 	    taskq_member(kmem_taskq, curthread));
2034 
2035 	for (round = 0; round < nrounds; round++) {
2036 		void *buf = mp->mag_round[round];
2037 
2038 		if (cp->cache_flags & KMF_DEADBEEF) {
2039 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2040 			    cp->cache_verify) != NULL) {
2041 				kmem_error(KMERR_MODIFIED, cp, buf);
2042 				continue;
2043 			}
2044 			if ((cp->cache_flags & KMF_LITE) &&
2045 			    cp->cache_destructor != NULL) {
2046 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2047 				*(uint64_t *)buf = btp->bt_redzone;
2048 				cp->cache_destructor(buf, cp->cache_private);
2049 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2050 			}
2051 		} else if (cp->cache_destructor != NULL) {
2052 			cp->cache_destructor(buf, cp->cache_private);
2053 		}
2054 
2055 		kmem_slab_free(cp, buf);
2056 	}
2057 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2058 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2059 }
2060 
2061 /*
2062  * Allocate a magazine from the depot.
2063  */
2064 static kmem_magazine_t *
2065 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2066 {
2067 	kmem_magazine_t *mp;
2068 
2069 	/*
2070 	 * If we can't get the depot lock without contention,
2071 	 * update our contention count.  We use the depot
2072 	 * contention rate to determine whether we need to
2073 	 * increase the magazine size for better scalability.
2074 	 */
2075 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2076 		mutex_enter(&cp->cache_depot_lock);
2077 		cp->cache_depot_contention++;
2078 	}
2079 
2080 	if ((mp = mlp->ml_list) != NULL) {
2081 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2082 		mlp->ml_list = mp->mag_next;
2083 		if (--mlp->ml_total < mlp->ml_min)
2084 			mlp->ml_min = mlp->ml_total;
2085 		mlp->ml_alloc++;
2086 	}
2087 
2088 	mutex_exit(&cp->cache_depot_lock);
2089 
2090 	return (mp);
2091 }
2092 
2093 /*
2094  * Free a magazine to the depot.
2095  */
2096 static void
2097 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2098 {
2099 	mutex_enter(&cp->cache_depot_lock);
2100 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2101 	mp->mag_next = mlp->ml_list;
2102 	mlp->ml_list = mp;
2103 	mlp->ml_total++;
2104 	mutex_exit(&cp->cache_depot_lock);
2105 }
2106 
2107 /*
2108  * Update the working set statistics for cp's depot.
2109  */
2110 static void
2111 kmem_depot_ws_update(kmem_cache_t *cp)
2112 {
2113 	mutex_enter(&cp->cache_depot_lock);
2114 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2115 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2116 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2117 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2118 	mutex_exit(&cp->cache_depot_lock);
2119 }
2120 
2121 /*
2122  * Set the working set statistics for cp's depot to zero.  (Everything is
2123  * eligible for reaping.)
2124  */
2125 static void
2126 kmem_depot_ws_zero(kmem_cache_t *cp)
2127 {
2128 	mutex_enter(&cp->cache_depot_lock);
2129 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2130 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2131 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2132 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2133 	mutex_exit(&cp->cache_depot_lock);
2134 }
2135 
2136 /*
2137  * The number of bytes to reap before we call kpreempt(). The default (1MB)
2138  * causes us to preempt reaping up to hundreds of times per second. Using a
2139  * larger value (1GB) causes this to have virtually no effect.
2140  */
2141 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2142 
2143 /*
2144  * Reap all magazines that have fallen out of the depot's working set.
2145  */
2146 static void
2147 kmem_depot_ws_reap(kmem_cache_t *cp)
2148 {
2149 	size_t bytes = 0;
2150 	long reap;
2151 	kmem_magazine_t *mp;
2152 
2153 	ASSERT(!list_link_active(&cp->cache_link) ||
2154 	    taskq_member(kmem_taskq, curthread));
2155 
2156 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2157 	while (reap-- &&
2158 	    (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2159 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2160 		bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2161 		if (bytes > kmem_reap_preempt_bytes) {
2162 			kpreempt(KPREEMPT_SYNC);
2163 			bytes = 0;
2164 		}
2165 	}
2166 
2167 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2168 	while (reap-- &&
2169 	    (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2170 		kmem_magazine_destroy(cp, mp, 0);
2171 		bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2172 		if (bytes > kmem_reap_preempt_bytes) {
2173 			kpreempt(KPREEMPT_SYNC);
2174 			bytes = 0;
2175 		}
2176 	}
2177 }
2178 
2179 static void
2180 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2181 {
2182 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2183 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2184 	ASSERT(ccp->cc_magsize > 0);
2185 
2186 	ccp->cc_ploaded = ccp->cc_loaded;
2187 	ccp->cc_prounds = ccp->cc_rounds;
2188 	ccp->cc_loaded = mp;
2189 	ccp->cc_rounds = rounds;
2190 }
2191 
2192 /*
2193  * Intercept kmem alloc/free calls during crash dump in order to avoid
2194  * changing kmem state while memory is being saved to the dump device.
2195  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2196  * there are no locks because only one CPU calls kmem during a crash
2197  * dump. To enable this feature, first create the associated vmem
2198  * arena with VMC_DUMPSAFE.
2199  */
2200 static void *kmem_dump_start;	/* start of pre-reserved heap */
2201 static void *kmem_dump_end;	/* end of heap area */
2202 static void *kmem_dump_curr;	/* current free heap pointer */
2203 static size_t kmem_dump_size;	/* size of heap area */
2204 
2205 /* append to each buf created in the pre-reserved heap */
2206 typedef struct kmem_dumpctl {
2207 	void	*kdc_next;	/* cache dump free list linkage */
2208 } kmem_dumpctl_t;
2209 
2210 #define	KMEM_DUMPCTL(cp, buf)	\
2211 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2212 	    sizeof (void *)))
2213 
2214 /* set non zero for full report */
2215 uint_t kmem_dump_verbose = 0;
2216 
2217 /* stats for overize heap */
2218 uint_t kmem_dump_oversize_allocs = 0;
2219 uint_t kmem_dump_oversize_max = 0;
2220 
2221 static void
2222 kmem_dumppr(char **pp, char *e, const char *format, ...)
2223 {
2224 	char *p = *pp;
2225 
2226 	if (p < e) {
2227 		int n;
2228 		va_list ap;
2229 
2230 		va_start(ap, format);
2231 		n = vsnprintf(p, e - p, format, ap);
2232 		va_end(ap);
2233 		*pp = p + n;
2234 	}
2235 }
2236 
2237 /*
2238  * Called when dumpadm(1M) configures dump parameters.
2239  */
2240 void
2241 kmem_dump_init(size_t size)
2242 {
2243 	/* Our caller ensures size is always set. */
2244 	ASSERT3U(size, >, 0);
2245 
2246 	if (kmem_dump_start != NULL)
2247 		kmem_free(kmem_dump_start, kmem_dump_size);
2248 
2249 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2250 	kmem_dump_size = size;
2251 	kmem_dump_curr = kmem_dump_start;
2252 	kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2253 	copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2254 }
2255 
2256 /*
2257  * Set flag for each kmem_cache_t if is safe to use alternate dump
2258  * memory. Called just before panic crash dump starts. Set the flag
2259  * for the calling CPU.
2260  */
2261 void
2262 kmem_dump_begin(void)
2263 {
2264 	kmem_cache_t *cp;
2265 
2266 	ASSERT(panicstr != NULL);
2267 
2268 	for (cp = list_head(&kmem_caches); cp != NULL;
2269 	    cp = list_next(&kmem_caches, cp)) {
2270 		kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2271 
2272 		if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2273 			cp->cache_flags |= KMF_DUMPDIVERT;
2274 			ccp->cc_flags |= KMF_DUMPDIVERT;
2275 			ccp->cc_dump_rounds = ccp->cc_rounds;
2276 			ccp->cc_dump_prounds = ccp->cc_prounds;
2277 			ccp->cc_rounds = ccp->cc_prounds = -1;
2278 		} else {
2279 			cp->cache_flags |= KMF_DUMPUNSAFE;
2280 			ccp->cc_flags |= KMF_DUMPUNSAFE;
2281 		}
2282 	}
2283 }
2284 
2285 /*
2286  * finished dump intercept
2287  * print any warnings on the console
2288  * return verbose information to dumpsys() in the given buffer
2289  */
2290 size_t
2291 kmem_dump_finish(char *buf, size_t size)
2292 {
2293 	int percent = 0;
2294 	size_t used;
2295 	char *e = buf + size;
2296 	char *p = buf;
2297 
2298 	if (kmem_dump_curr == kmem_dump_end) {
2299 		cmn_err(CE_WARN, "exceeded kmem_dump space of %lu "
2300 		    "bytes: kmem state in dump may be inconsistent",
2301 		    kmem_dump_size);
2302 	}
2303 
2304 	if (kmem_dump_verbose == 0)
2305 		return (0);
2306 
2307 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2308 	percent = (used * 100) / kmem_dump_size;
2309 
2310 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2311 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2312 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2313 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2314 	    kmem_dump_oversize_allocs);
2315 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2316 	    kmem_dump_oversize_max);
2317 
2318 	/* return buffer size used */
2319 	if (p < e)
2320 		bzero(p, e - p);
2321 	return (p - buf);
2322 }
2323 
2324 /*
2325  * Allocate a constructed object from alternate dump memory.
2326  */
2327 void *
2328 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2329 {
2330 	void *buf;
2331 	void *curr;
2332 	char *bufend;
2333 
2334 	/* return a constructed object */
2335 	if ((buf = cp->cache_dump.kd_freelist) != NULL) {
2336 		cp->cache_dump.kd_freelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2337 		return (buf);
2338 	}
2339 
2340 	/* create a new constructed object */
2341 	curr = kmem_dump_curr;
2342 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2343 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2344 
2345 	/* hat layer objects cannot cross a page boundary */
2346 	if (cp->cache_align < PAGESIZE) {
2347 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2348 		if (bufend > page) {
2349 			bufend += page - (char *)buf;
2350 			buf = (void *)page;
2351 		}
2352 	}
2353 
2354 	/* fall back to normal alloc if reserved area is used up */
2355 	if (bufend > (char *)kmem_dump_end) {
2356 		kmem_dump_curr = kmem_dump_end;
2357 		cp->cache_dump.kd_alloc_fails++;
2358 		return (NULL);
2359 	}
2360 
2361 	/*
2362 	 * Must advance curr pointer before calling a constructor that
2363 	 * may also allocate memory.
2364 	 */
2365 	kmem_dump_curr = bufend;
2366 
2367 	/* run constructor */
2368 	if (cp->cache_constructor != NULL &&
2369 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2370 	    != 0) {
2371 #ifdef DEBUG
2372 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2373 		    cp->cache_name, (void *)cp);
2374 #endif
2375 		/* reset curr pointer iff no allocs were done */
2376 		if (kmem_dump_curr == bufend)
2377 			kmem_dump_curr = curr;
2378 
2379 		cp->cache_dump.kd_alloc_fails++;
2380 		/* fall back to normal alloc if the constructor fails */
2381 		return (NULL);
2382 	}
2383 
2384 	return (buf);
2385 }
2386 
2387 /*
2388  * Free a constructed object in alternate dump memory.
2389  */
2390 int
2391 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2392 {
2393 	/* save constructed buffers for next time */
2394 	if ((char *)buf >= (char *)kmem_dump_start &&
2395 	    (char *)buf < (char *)kmem_dump_end) {
2396 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dump.kd_freelist;
2397 		cp->cache_dump.kd_freelist = buf;
2398 		return (0);
2399 	}
2400 
2401 	/* just drop buffers that were allocated before dump started */
2402 	if (kmem_dump_curr < kmem_dump_end)
2403 		return (0);
2404 
2405 	/* fall back to normal free if reserved area is used up */
2406 	return (1);
2407 }
2408 
2409 /*
2410  * Allocate a constructed object from cache cp.
2411  */
2412 void *
2413 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2414 {
2415 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2416 	kmem_magazine_t *fmp;
2417 	void *buf;
2418 
2419 	mutex_enter(&ccp->cc_lock);
2420 	for (;;) {
2421 		/*
2422 		 * If there's an object available in the current CPU's
2423 		 * loaded magazine, just take it and return.
2424 		 */
2425 		if (ccp->cc_rounds > 0) {
2426 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2427 			ccp->cc_alloc++;
2428 			mutex_exit(&ccp->cc_lock);
2429 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2430 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2431 					ASSERT(!(ccp->cc_flags &
2432 					    KMF_DUMPDIVERT));
2433 					cp->cache_dump.kd_unsafe++;
2434 				}
2435 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2436 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2437 				    caller()) != 0) {
2438 					if (kmflag & KM_NOSLEEP)
2439 						return (NULL);
2440 					mutex_enter(&ccp->cc_lock);
2441 					continue;
2442 				}
2443 			}
2444 			return (buf);
2445 		}
2446 
2447 		/*
2448 		 * The loaded magazine is empty.  If the previously loaded
2449 		 * magazine was full, exchange them and try again.
2450 		 */
2451 		if (ccp->cc_prounds > 0) {
2452 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2453 			continue;
2454 		}
2455 
2456 		/*
2457 		 * Return an alternate buffer at dump time to preserve
2458 		 * the heap.
2459 		 */
2460 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2461 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2462 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2463 				/* log it so that we can warn about it */
2464 				cp->cache_dump.kd_unsafe++;
2465 			} else {
2466 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2467 				    NULL) {
2468 					mutex_exit(&ccp->cc_lock);
2469 					return (buf);
2470 				}
2471 				break;		/* fall back to slab layer */
2472 			}
2473 		}
2474 
2475 		/*
2476 		 * If the magazine layer is disabled, break out now.
2477 		 */
2478 		if (ccp->cc_magsize == 0)
2479 			break;
2480 
2481 		/*
2482 		 * Try to get a full magazine from the depot.
2483 		 */
2484 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2485 		if (fmp != NULL) {
2486 			if (ccp->cc_ploaded != NULL)
2487 				kmem_depot_free(cp, &cp->cache_empty,
2488 				    ccp->cc_ploaded);
2489 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2490 			continue;
2491 		}
2492 
2493 		/*
2494 		 * There are no full magazines in the depot,
2495 		 * so fall through to the slab layer.
2496 		 */
2497 		break;
2498 	}
2499 	mutex_exit(&ccp->cc_lock);
2500 
2501 	/*
2502 	 * We couldn't allocate a constructed object from the magazine layer,
2503 	 * so get a raw buffer from the slab layer and apply its constructor.
2504 	 */
2505 	buf = kmem_slab_alloc(cp, kmflag);
2506 
2507 	if (buf == NULL)
2508 		return (NULL);
2509 
2510 	if (cp->cache_flags & KMF_BUFTAG) {
2511 		/*
2512 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2513 		 */
2514 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2515 		if (rc != 0) {
2516 			if (kmflag & KM_NOSLEEP)
2517 				return (NULL);
2518 			/*
2519 			 * kmem_cache_alloc_debug() detected corruption
2520 			 * but didn't panic (kmem_panic <= 0). We should not be
2521 			 * here because the constructor failed (indicated by a
2522 			 * return code of 1). Try again.
2523 			 */
2524 			ASSERT(rc == -1);
2525 			return (kmem_cache_alloc(cp, kmflag));
2526 		}
2527 		return (buf);
2528 	}
2529 
2530 	if (cp->cache_constructor != NULL &&
2531 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2532 		atomic_inc_64(&cp->cache_alloc_fail);
2533 		kmem_slab_free(cp, buf);
2534 		return (NULL);
2535 	}
2536 
2537 	return (buf);
2538 }
2539 
2540 /*
2541  * The freed argument tells whether or not kmem_cache_free_debug() has already
2542  * been called so that we can avoid the duplicate free error. For example, a
2543  * buffer on a magazine has already been freed by the client but is still
2544  * constructed.
2545  */
2546 static void
2547 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2548 {
2549 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2550 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2551 			return;
2552 
2553 	/*
2554 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2555 	 * kmem_cache_free_debug() will have already applied the destructor.
2556 	 */
2557 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2558 	    cp->cache_destructor != NULL) {
2559 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2560 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2561 			*(uint64_t *)buf = btp->bt_redzone;
2562 			cp->cache_destructor(buf, cp->cache_private);
2563 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2564 		} else {
2565 			cp->cache_destructor(buf, cp->cache_private);
2566 		}
2567 	}
2568 
2569 	kmem_slab_free(cp, buf);
2570 }
2571 
2572 /*
2573  * Used when there's no room to free a buffer to the per-CPU cache.
2574  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2575  * caller should try freeing to the per-CPU cache again.
2576  * Note that we don't directly install the magazine in the cpu cache,
2577  * since its state may have changed wildly while the lock was dropped.
2578  */
2579 static int
2580 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2581 {
2582 	kmem_magazine_t *emp;
2583 	kmem_magtype_t *mtp;
2584 
2585 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2586 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2587 	    ((uint_t)ccp->cc_rounds == -1)) &&
2588 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2589 	    ((uint_t)ccp->cc_prounds == -1)));
2590 
2591 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2592 	if (emp != NULL) {
2593 		if (ccp->cc_ploaded != NULL)
2594 			kmem_depot_free(cp, &cp->cache_full,
2595 			    ccp->cc_ploaded);
2596 		kmem_cpu_reload(ccp, emp, 0);
2597 		return (1);
2598 	}
2599 	/*
2600 	 * There are no empty magazines in the depot,
2601 	 * so try to allocate a new one.  We must drop all locks
2602 	 * across kmem_cache_alloc() because lower layers may
2603 	 * attempt to allocate from this cache.
2604 	 */
2605 	mtp = cp->cache_magtype;
2606 	mutex_exit(&ccp->cc_lock);
2607 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2608 	mutex_enter(&ccp->cc_lock);
2609 
2610 	if (emp != NULL) {
2611 		/*
2612 		 * We successfully allocated an empty magazine.
2613 		 * However, we had to drop ccp->cc_lock to do it,
2614 		 * so the cache's magazine size may have changed.
2615 		 * If so, free the magazine and try again.
2616 		 */
2617 		if (ccp->cc_magsize != mtp->mt_magsize) {
2618 			mutex_exit(&ccp->cc_lock);
2619 			kmem_cache_free(mtp->mt_cache, emp);
2620 			mutex_enter(&ccp->cc_lock);
2621 			return (1);
2622 		}
2623 
2624 		/*
2625 		 * We got a magazine of the right size.  Add it to
2626 		 * the depot and try the whole dance again.
2627 		 */
2628 		kmem_depot_free(cp, &cp->cache_empty, emp);
2629 		return (1);
2630 	}
2631 
2632 	/*
2633 	 * We couldn't allocate an empty magazine,
2634 	 * so fall through to the slab layer.
2635 	 */
2636 	return (0);
2637 }
2638 
2639 /*
2640  * Free a constructed object to cache cp.
2641  */
2642 void
2643 kmem_cache_free(kmem_cache_t *cp, void *buf)
2644 {
2645 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2646 
2647 	/*
2648 	 * The client must not free either of the buffers passed to the move
2649 	 * callback function.
2650 	 */
2651 	ASSERT(cp->cache_defrag == NULL ||
2652 	    cp->cache_defrag->kmd_thread != curthread ||
2653 	    (buf != cp->cache_defrag->kmd_from_buf &&
2654 	    buf != cp->cache_defrag->kmd_to_buf));
2655 
2656 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2657 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2658 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2659 			/* log it so that we can warn about it */
2660 			cp->cache_dump.kd_unsafe++;
2661 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2662 			return;
2663 		}
2664 		if (ccp->cc_flags & KMF_BUFTAG) {
2665 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2666 				return;
2667 		}
2668 	}
2669 
2670 	mutex_enter(&ccp->cc_lock);
2671 	/*
2672 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2673 	 */
2674 	for (;;) {
2675 		/*
2676 		 * If there's a slot available in the current CPU's
2677 		 * loaded magazine, just put the object there and return.
2678 		 */
2679 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2680 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2681 			ccp->cc_free++;
2682 			mutex_exit(&ccp->cc_lock);
2683 			return;
2684 		}
2685 
2686 		/*
2687 		 * The loaded magazine is full.  If the previously loaded
2688 		 * magazine was empty, exchange them and try again.
2689 		 */
2690 		if (ccp->cc_prounds == 0) {
2691 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2692 			continue;
2693 		}
2694 
2695 		/*
2696 		 * If the magazine layer is disabled, break out now.
2697 		 */
2698 		if (ccp->cc_magsize == 0)
2699 			break;
2700 
2701 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2702 			/*
2703 			 * We couldn't free our constructed object to the
2704 			 * magazine layer, so apply its destructor and free it
2705 			 * to the slab layer.
2706 			 */
2707 			break;
2708 		}
2709 	}
2710 	mutex_exit(&ccp->cc_lock);
2711 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2712 }
2713 
2714 static void
2715 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2716 {
2717 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2718 	int cache_flags = cp->cache_flags;
2719 
2720 	kmem_bufctl_t *next, *head;
2721 	size_t nbufs;
2722 
2723 	/*
2724 	 * Completely allocate the newly created slab and put the pre-allocated
2725 	 * buffers in magazines. Any of the buffers that cannot be put in
2726 	 * magazines must be returned to the slab.
2727 	 */
2728 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2729 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2730 	ASSERT(cp->cache_constructor == NULL);
2731 	ASSERT(sp->slab_cache == cp);
2732 	ASSERT(sp->slab_refcnt == 1);
2733 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2734 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2735 
2736 	head = sp->slab_head;
2737 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2738 	sp->slab_head = NULL;
2739 	sp->slab_refcnt += nbufs;
2740 	cp->cache_bufslab -= nbufs;
2741 	cp->cache_slab_alloc += nbufs;
2742 	list_insert_head(&cp->cache_complete_slabs, sp);
2743 	cp->cache_complete_slab_count++;
2744 	mutex_exit(&cp->cache_lock);
2745 	mutex_enter(&ccp->cc_lock);
2746 
2747 	while (head != NULL) {
2748 		void *buf = KMEM_BUF(cp, head);
2749 		/*
2750 		 * If there's a slot available in the current CPU's
2751 		 * loaded magazine, just put the object there and
2752 		 * continue.
2753 		 */
2754 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2755 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2756 			    buf;
2757 			ccp->cc_free++;
2758 			nbufs--;
2759 			head = head->bc_next;
2760 			continue;
2761 		}
2762 
2763 		/*
2764 		 * The loaded magazine is full.  If the previously
2765 		 * loaded magazine was empty, exchange them and try
2766 		 * again.
2767 		 */
2768 		if (ccp->cc_prounds == 0) {
2769 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2770 			    ccp->cc_prounds);
2771 			continue;
2772 		}
2773 
2774 		/*
2775 		 * If the magazine layer is disabled, break out now.
2776 		 */
2777 
2778 		if (ccp->cc_magsize == 0) {
2779 			break;
2780 		}
2781 
2782 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2783 			break;
2784 	}
2785 	mutex_exit(&ccp->cc_lock);
2786 	if (nbufs != 0) {
2787 		ASSERT(head != NULL);
2788 
2789 		/*
2790 		 * If there was a failure, return remaining objects to
2791 		 * the slab
2792 		 */
2793 		while (head != NULL) {
2794 			ASSERT(nbufs != 0);
2795 			next = head->bc_next;
2796 			head->bc_next = NULL;
2797 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2798 			head = next;
2799 			nbufs--;
2800 		}
2801 	}
2802 	ASSERT(head == NULL);
2803 	ASSERT(nbufs == 0);
2804 	mutex_enter(&cp->cache_lock);
2805 }
2806 
2807 void *
2808 kmem_zalloc(size_t size, int kmflag)
2809 {
2810 	size_t index;
2811 	void *buf;
2812 
2813 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2814 		kmem_cache_t *cp = kmem_alloc_table[index];
2815 		buf = kmem_cache_alloc(cp, kmflag);
2816 		if (buf != NULL) {
2817 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2818 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2819 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2820 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2821 
2822 				if (cp->cache_flags & KMF_LITE) {
2823 					KMEM_BUFTAG_LITE_ENTER(btp,
2824 					    kmem_lite_count, caller());
2825 				}
2826 			}
2827 			bzero(buf, size);
2828 		}
2829 	} else {
2830 		buf = kmem_alloc(size, kmflag);
2831 		if (buf != NULL)
2832 			bzero(buf, size);
2833 	}
2834 	return (buf);
2835 }
2836 
2837 void *
2838 kmem_alloc(size_t size, int kmflag)
2839 {
2840 	size_t index;
2841 	kmem_cache_t *cp;
2842 	void *buf;
2843 
2844 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2845 		cp = kmem_alloc_table[index];
2846 		/* fall through to kmem_cache_alloc() */
2847 
2848 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2849 	    kmem_big_alloc_table_max) {
2850 		cp = kmem_big_alloc_table[index];
2851 		/* fall through to kmem_cache_alloc() */
2852 
2853 	} else {
2854 		if (size == 0)
2855 			return (NULL);
2856 
2857 		buf = vmem_alloc(kmem_oversize_arena, size,
2858 		    kmflag & KM_VMFLAGS);
2859 		if (buf == NULL)
2860 			kmem_log_event(kmem_failure_log, NULL, NULL,
2861 			    (void *)size);
2862 		else if (KMEM_DUMP(kmem_slab_cache)) {
2863 			/* stats for dump intercept */
2864 			kmem_dump_oversize_allocs++;
2865 			if (size > kmem_dump_oversize_max)
2866 				kmem_dump_oversize_max = size;
2867 		}
2868 		return (buf);
2869 	}
2870 
2871 	buf = kmem_cache_alloc(cp, kmflag);
2872 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2873 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2874 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2875 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2876 
2877 		if (cp->cache_flags & KMF_LITE) {
2878 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2879 		}
2880 	}
2881 	return (buf);
2882 }
2883 
2884 void
2885 kmem_free(void *buf, size_t size)
2886 {
2887 	size_t index;
2888 	kmem_cache_t *cp;
2889 
2890 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2891 		cp = kmem_alloc_table[index];
2892 		/* fall through to kmem_cache_free() */
2893 
2894 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2895 	    kmem_big_alloc_table_max) {
2896 		cp = kmem_big_alloc_table[index];
2897 		/* fall through to kmem_cache_free() */
2898 
2899 	} else {
2900 		EQUIV(buf == NULL, size == 0);
2901 		if (buf == NULL && size == 0)
2902 			return;
2903 		vmem_free(kmem_oversize_arena, buf, size);
2904 		return;
2905 	}
2906 
2907 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2908 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2909 		uint32_t *ip = (uint32_t *)btp;
2910 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2911 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2912 				kmem_error(KMERR_DUPFREE, cp, buf);
2913 				return;
2914 			}
2915 			if (KMEM_SIZE_VALID(ip[1])) {
2916 				ip[0] = KMEM_SIZE_ENCODE(size);
2917 				kmem_error(KMERR_BADSIZE, cp, buf);
2918 			} else {
2919 				kmem_error(KMERR_REDZONE, cp, buf);
2920 			}
2921 			return;
2922 		}
2923 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2924 			kmem_error(KMERR_REDZONE, cp, buf);
2925 			return;
2926 		}
2927 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
2928 		if (cp->cache_flags & KMF_LITE) {
2929 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2930 			    caller());
2931 		}
2932 	}
2933 	kmem_cache_free(cp, buf);
2934 }
2935 
2936 void *
2937 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2938 {
2939 	size_t realsize = size + vmp->vm_quantum;
2940 	void *addr;
2941 
2942 	/*
2943 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2944 	 * vm_quantum will cause integer wraparound.  Check for this, and
2945 	 * blow off the firewall page in this case.  Note that such a
2946 	 * giant allocation (the entire kernel address space) can never
2947 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
2948 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2949 	 * corresponding check in kmem_firewall_va_free().
2950 	 */
2951 	if (realsize < size)
2952 		realsize = size;
2953 
2954 	/*
2955 	 * While boot still owns resource management, make sure that this
2956 	 * redzone virtual address allocation is properly accounted for in
2957 	 * OBPs "virtual-memory" "available" lists because we're
2958 	 * effectively claiming them for a red zone.  If we don't do this,
2959 	 * the available lists become too fragmented and too large for the
2960 	 * current boot/kernel memory list interface.
2961 	 */
2962 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
2963 
2964 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
2965 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
2966 
2967 	return (addr);
2968 }
2969 
2970 void
2971 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2972 {
2973 	ASSERT((kvseg.s_base == NULL ?
2974 	    va_to_pfn((char *)addr + size) :
2975 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
2976 
2977 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2978 }
2979 
2980 /*
2981  * Try to allocate at least `size' bytes of memory without sleeping or
2982  * panicking. Return actual allocated size in `asize'. If allocation failed,
2983  * try final allocation with sleep or panic allowed.
2984  */
2985 void *
2986 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
2987 {
2988 	void *p;
2989 
2990 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2991 	do {
2992 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
2993 		if (p != NULL)
2994 			return (p);
2995 		*asize += KMEM_ALIGN;
2996 	} while (*asize <= PAGESIZE);
2997 
2998 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2999 	return (kmem_alloc(*asize, kmflag));
3000 }
3001 
3002 /*
3003  * Reclaim all unused memory from a cache.
3004  */
3005 static void
3006 kmem_cache_reap(kmem_cache_t *cp)
3007 {
3008 	ASSERT(taskq_member(kmem_taskq, curthread));
3009 	cp->cache_reap++;
3010 
3011 	/*
3012 	 * Ask the cache's owner to free some memory if possible.
3013 	 * The idea is to handle things like the inode cache, which
3014 	 * typically sits on a bunch of memory that it doesn't truly
3015 	 * *need*.  Reclaim policy is entirely up to the owner; this
3016 	 * callback is just an advisory plea for help.
3017 	 */
3018 	if (cp->cache_reclaim != NULL) {
3019 		long delta;
3020 
3021 		/*
3022 		 * Reclaimed memory should be reapable (not included in the
3023 		 * depot's working set).
3024 		 */
3025 		delta = cp->cache_full.ml_total;
3026 		cp->cache_reclaim(cp->cache_private);
3027 		delta = cp->cache_full.ml_total - delta;
3028 		if (delta > 0) {
3029 			mutex_enter(&cp->cache_depot_lock);
3030 			cp->cache_full.ml_reaplimit += delta;
3031 			cp->cache_full.ml_min += delta;
3032 			mutex_exit(&cp->cache_depot_lock);
3033 		}
3034 	}
3035 
3036 	kmem_depot_ws_reap(cp);
3037 
3038 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3039 		kmem_cache_defrag(cp);
3040 	}
3041 }
3042 
3043 static void
3044 kmem_reap_timeout(void *flag_arg)
3045 {
3046 	uint32_t *flag = (uint32_t *)flag_arg;
3047 
3048 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3049 	*flag = 0;
3050 }
3051 
3052 static void
3053 kmem_reap_done(void *flag)
3054 {
3055 	if (!callout_init_done) {
3056 		/* can't schedule a timeout at this point */
3057 		kmem_reap_timeout(flag);
3058 	} else {
3059 		(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3060 	}
3061 }
3062 
3063 static void
3064 kmem_reap_start(void *flag)
3065 {
3066 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3067 
3068 	if (flag == &kmem_reaping) {
3069 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3070 		/*
3071 		 * if we have segkp under heap, reap segkp cache.
3072 		 */
3073 		if (segkp_fromheap)
3074 			segkp_cache_free();
3075 	}
3076 	else
3077 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3078 
3079 	/*
3080 	 * We use taskq_dispatch() to schedule a timeout to clear
3081 	 * the flag so that kmem_reap() becomes self-throttling:
3082 	 * we won't reap again until the current reap completes *and*
3083 	 * at least kmem_reap_interval ticks have elapsed.
3084 	 */
3085 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3086 		kmem_reap_done(flag);
3087 }
3088 
3089 static void
3090 kmem_reap_common(void *flag_arg)
3091 {
3092 	uint32_t *flag = (uint32_t *)flag_arg;
3093 
3094 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3095 	    atomic_cas_32(flag, 0, 1) != 0)
3096 		return;
3097 
3098 	/*
3099 	 * It may not be kosher to do memory allocation when a reap is called
3100 	 * (for example, if vmem_populate() is in the call chain).  So we
3101 	 * start the reap going with a TQ_NOALLOC dispatch.  If the dispatch
3102 	 * fails, we reset the flag, and the next reap will try again.
3103 	 */
3104 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3105 		*flag = 0;
3106 }
3107 
3108 /*
3109  * Reclaim all unused memory from all caches.  Called from the VM system
3110  * when memory gets tight.
3111  */
3112 void
3113 kmem_reap(void)
3114 {
3115 	kmem_reap_common(&kmem_reaping);
3116 }
3117 
3118 /*
3119  * Reclaim all unused memory from identifier arenas, called when a vmem
3120  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3121  * cannot help with identifier exhaustion, we avoid both a large amount of
3122  * work and unwanted side-effects from reclaim callbacks.
3123  */
3124 void
3125 kmem_reap_idspace(void)
3126 {
3127 	kmem_reap_common(&kmem_reaping_idspace);
3128 }
3129 
3130 /*
3131  * Purge all magazines from a cache and set its magazine limit to zero.
3132  * All calls are serialized by the kmem_taskq lock, except for the final
3133  * call from kmem_cache_destroy().
3134  */
3135 static void
3136 kmem_cache_magazine_purge(kmem_cache_t *cp)
3137 {
3138 	kmem_cpu_cache_t *ccp;
3139 	kmem_magazine_t *mp, *pmp;
3140 	int rounds, prounds, cpu_seqid;
3141 
3142 	ASSERT(!list_link_active(&cp->cache_link) ||
3143 	    taskq_member(kmem_taskq, curthread));
3144 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3145 
3146 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3147 		ccp = &cp->cache_cpu[cpu_seqid];
3148 
3149 		mutex_enter(&ccp->cc_lock);
3150 		mp = ccp->cc_loaded;
3151 		pmp = ccp->cc_ploaded;
3152 		rounds = ccp->cc_rounds;
3153 		prounds = ccp->cc_prounds;
3154 		ccp->cc_loaded = NULL;
3155 		ccp->cc_ploaded = NULL;
3156 		ccp->cc_rounds = -1;
3157 		ccp->cc_prounds = -1;
3158 		ccp->cc_magsize = 0;
3159 		mutex_exit(&ccp->cc_lock);
3160 
3161 		if (mp)
3162 			kmem_magazine_destroy(cp, mp, rounds);
3163 		if (pmp)
3164 			kmem_magazine_destroy(cp, pmp, prounds);
3165 	}
3166 
3167 	kmem_depot_ws_zero(cp);
3168 	kmem_depot_ws_reap(cp);
3169 }
3170 
3171 /*
3172  * Enable per-cpu magazines on a cache.
3173  */
3174 static void
3175 kmem_cache_magazine_enable(kmem_cache_t *cp)
3176 {
3177 	int cpu_seqid;
3178 
3179 	if (cp->cache_flags & KMF_NOMAGAZINE)
3180 		return;
3181 
3182 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3183 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3184 		mutex_enter(&ccp->cc_lock);
3185 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3186 		mutex_exit(&ccp->cc_lock);
3187 	}
3188 
3189 }
3190 
3191 /*
3192  * Allow our caller to determine if there are running reaps.
3193  *
3194  * This call is very conservative and may return B_TRUE even when
3195  * reaping activity isn't active. If it returns B_FALSE, then reaping
3196  * activity is definitely inactive.
3197  */
3198 boolean_t
3199 kmem_cache_reap_active(void)
3200 {
3201 	return (!taskq_empty(kmem_taskq));
3202 }
3203 
3204 /*
3205  * Reap (almost) everything soon.
3206  *
3207  * Note: this does not wait for the reap-tasks to complete. Caller
3208  * should use kmem_cache_reap_active() (above) and/or moderation to
3209  * avoid scheduling too many reap-tasks.
3210  */
3211 void
3212 kmem_cache_reap_soon(kmem_cache_t *cp)
3213 {
3214 	ASSERT(list_link_active(&cp->cache_link));
3215 
3216 	kmem_depot_ws_zero(cp);
3217 
3218 	(void) taskq_dispatch(kmem_taskq,
3219 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3220 }
3221 
3222 /*
3223  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3224  * provide a higher transfer rate with the depot, while smaller magazines
3225  * reduce memory consumption.  Magazine resizing is an expensive operation;
3226  * it should not be done frequently.
3227  *
3228  * Changes to the magazine size are serialized by the kmem_taskq lock.
3229  *
3230  * Note: at present this only grows the magazine size.  It might be useful
3231  * to allow shrinkage too.
3232  */
3233 static void
3234 kmem_cache_magazine_resize(kmem_cache_t *cp)
3235 {
3236 	kmem_magtype_t *mtp = cp->cache_magtype;
3237 
3238 	ASSERT(taskq_member(kmem_taskq, curthread));
3239 
3240 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3241 		kmem_cache_magazine_purge(cp);
3242 		mutex_enter(&cp->cache_depot_lock);
3243 		cp->cache_magtype = ++mtp;
3244 		cp->cache_depot_contention_prev =
3245 		    cp->cache_depot_contention + INT_MAX;
3246 		mutex_exit(&cp->cache_depot_lock);
3247 		kmem_cache_magazine_enable(cp);
3248 	}
3249 }
3250 
3251 /*
3252  * Rescale a cache's hash table, so that the table size is roughly the
3253  * cache size.  We want the average lookup time to be extremely small.
3254  */
3255 static void
3256 kmem_hash_rescale(kmem_cache_t *cp)
3257 {
3258 	kmem_bufctl_t **old_table, **new_table, *bcp;
3259 	size_t old_size, new_size, h;
3260 
3261 	ASSERT(taskq_member(kmem_taskq, curthread));
3262 
3263 	new_size = MAX(KMEM_HASH_INITIAL,
3264 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3265 	old_size = cp->cache_hash_mask + 1;
3266 
3267 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3268 		return;
3269 
3270 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3271 	    VM_NOSLEEP);
3272 	if (new_table == NULL)
3273 		return;
3274 	bzero(new_table, new_size * sizeof (void *));
3275 
3276 	mutex_enter(&cp->cache_lock);
3277 
3278 	old_size = cp->cache_hash_mask + 1;
3279 	old_table = cp->cache_hash_table;
3280 
3281 	cp->cache_hash_mask = new_size - 1;
3282 	cp->cache_hash_table = new_table;
3283 	cp->cache_rescale++;
3284 
3285 	for (h = 0; h < old_size; h++) {
3286 		bcp = old_table[h];
3287 		while (bcp != NULL) {
3288 			void *addr = bcp->bc_addr;
3289 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3290 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3291 			bcp->bc_next = *hash_bucket;
3292 			*hash_bucket = bcp;
3293 			bcp = next_bcp;
3294 		}
3295 	}
3296 
3297 	mutex_exit(&cp->cache_lock);
3298 
3299 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3300 }
3301 
3302 /*
3303  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3304  * update, magazine resizing, and slab consolidation.
3305  */
3306 static void
3307 kmem_cache_update(kmem_cache_t *cp)
3308 {
3309 	int need_hash_rescale = 0;
3310 	int need_magazine_resize = 0;
3311 
3312 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3313 
3314 	/*
3315 	 * If the cache has become much larger or smaller than its hash table,
3316 	 * fire off a request to rescale the hash table.
3317 	 */
3318 	mutex_enter(&cp->cache_lock);
3319 
3320 	if ((cp->cache_flags & KMF_HASH) &&
3321 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3322 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3323 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3324 		need_hash_rescale = 1;
3325 
3326 	mutex_exit(&cp->cache_lock);
3327 
3328 	/*
3329 	 * Update the depot working set statistics.
3330 	 */
3331 	kmem_depot_ws_update(cp);
3332 
3333 	/*
3334 	 * If there's a lot of contention in the depot,
3335 	 * increase the magazine size.
3336 	 */
3337 	mutex_enter(&cp->cache_depot_lock);
3338 
3339 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3340 	    (int)(cp->cache_depot_contention -
3341 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3342 		need_magazine_resize = 1;
3343 
3344 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3345 
3346 	mutex_exit(&cp->cache_depot_lock);
3347 
3348 	if (need_hash_rescale)
3349 		(void) taskq_dispatch(kmem_taskq,
3350 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3351 
3352 	if (need_magazine_resize)
3353 		(void) taskq_dispatch(kmem_taskq,
3354 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3355 
3356 	if (cp->cache_defrag != NULL)
3357 		(void) taskq_dispatch(kmem_taskq,
3358 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3359 }
3360 
3361 static void kmem_update(void *);
3362 
3363 static void
3364 kmem_update_timeout(void *dummy)
3365 {
3366 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3367 }
3368 
3369 static void
3370 kmem_update(void *dummy)
3371 {
3372 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3373 
3374 	/*
3375 	 * We use taskq_dispatch() to reschedule the timeout so that
3376 	 * kmem_update() becomes self-throttling: it won't schedule
3377 	 * new tasks until all previous tasks have completed.
3378 	 */
3379 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3380 		kmem_update_timeout(NULL);
3381 }
3382 
3383 static int
3384 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3385 {
3386 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3387 	kmem_cache_t *cp = ksp->ks_private;
3388 	uint64_t cpu_buf_avail;
3389 	uint64_t buf_avail = 0;
3390 	int cpu_seqid;
3391 	long reap;
3392 
3393 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3394 
3395 	if (rw == KSTAT_WRITE)
3396 		return (EACCES);
3397 
3398 	mutex_enter(&cp->cache_lock);
3399 
3400 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3401 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3402 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3403 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3404 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3405 
3406 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3407 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3408 
3409 		mutex_enter(&ccp->cc_lock);
3410 
3411 		cpu_buf_avail = 0;
3412 		if (ccp->cc_rounds > 0)
3413 			cpu_buf_avail += ccp->cc_rounds;
3414 		if (ccp->cc_prounds > 0)
3415 			cpu_buf_avail += ccp->cc_prounds;
3416 
3417 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3418 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3419 		buf_avail			+= cpu_buf_avail;
3420 
3421 		mutex_exit(&ccp->cc_lock);
3422 	}
3423 
3424 	mutex_enter(&cp->cache_depot_lock);
3425 
3426 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3427 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3428 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3429 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3430 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3431 	kmcp->kmc_magazine_size.value.ui64	=
3432 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3433 	    0 : cp->cache_magtype->mt_magsize;
3434 
3435 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3436 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3437 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3438 
3439 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3440 	reap = MIN(reap, cp->cache_full.ml_total);
3441 
3442 	mutex_exit(&cp->cache_depot_lock);
3443 
3444 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3445 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3446 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3447 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3448 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3449 	buf_avail += cp->cache_bufslab;
3450 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3451 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3452 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3453 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3454 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3455 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3456 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3457 	    cp->cache_hash_mask + 1 : 0;
3458 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3459 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3460 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3461 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3462 
3463 	if (cp->cache_defrag == NULL) {
3464 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3465 		kmcp->kmc_move_yes.value.ui64		= 0;
3466 		kmcp->kmc_move_no.value.ui64		= 0;
3467 		kmcp->kmc_move_later.value.ui64		= 0;
3468 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3469 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3470 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3471 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3472 		kmcp->kmc_defrag.value.ui64		= 0;
3473 		kmcp->kmc_scan.value.ui64		= 0;
3474 		kmcp->kmc_move_reclaimable.value.ui64	= 0;
3475 	} else {
3476 		int64_t reclaimable;
3477 
3478 		kmem_defrag_t *kd = cp->cache_defrag;
3479 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3480 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3481 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3482 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3483 		kmcp->kmc_move_dont_need.value.ui64	= kd->kmd_dont_need;
3484 		kmcp->kmc_move_dont_know.value.ui64	= kd->kmd_dont_know;
3485 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3486 		kmcp->kmc_move_slabs_freed.value.ui64	= kd->kmd_slabs_freed;
3487 		kmcp->kmc_defrag.value.ui64		= kd->kmd_defrags;
3488 		kmcp->kmc_scan.value.ui64		= kd->kmd_scans;
3489 
3490 		reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3491 		reclaimable = MAX(reclaimable, 0);
3492 		reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3493 		kmcp->kmc_move_reclaimable.value.ui64	= reclaimable;
3494 	}
3495 
3496 	mutex_exit(&cp->cache_lock);
3497 	return (0);
3498 }
3499 
3500 /*
3501  * Return a named statistic about a particular cache.
3502  * This shouldn't be called very often, so it's currently designed for
3503  * simplicity (leverages existing kstat support) rather than efficiency.
3504  */
3505 uint64_t
3506 kmem_cache_stat(kmem_cache_t *cp, char *name)
3507 {
3508 	int i;
3509 	kstat_t *ksp = cp->cache_kstat;
3510 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3511 	uint64_t value = 0;
3512 
3513 	if (ksp != NULL) {
3514 		mutex_enter(&kmem_cache_kstat_lock);
3515 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3516 		for (i = 0; i < ksp->ks_ndata; i++) {
3517 			if (strcmp(knp[i].name, name) == 0) {
3518 				value = knp[i].value.ui64;
3519 				break;
3520 			}
3521 		}
3522 		mutex_exit(&kmem_cache_kstat_lock);
3523 	}
3524 	return (value);
3525 }
3526 
3527 /*
3528  * Return an estimate of currently available kernel heap memory.
3529  * On 32-bit systems, physical memory may exceed virtual memory,
3530  * we just truncate the result at 1GB.
3531  */
3532 size_t
3533 kmem_avail(void)
3534 {
3535 	spgcnt_t rmem = availrmem - tune.t_minarmem;
3536 	spgcnt_t fmem = freemem - minfree;
3537 
3538 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3539 	    1 << (30 - PAGESHIFT))));
3540 }
3541 
3542 /*
3543  * Return the maximum amount of memory that is (in theory) allocatable
3544  * from the heap. This may be used as an estimate only since there
3545  * is no guarentee this space will still be available when an allocation
3546  * request is made, nor that the space may be allocated in one big request
3547  * due to kernel heap fragmentation.
3548  */
3549 size_t
3550 kmem_maxavail(void)
3551 {
3552 	spgcnt_t pmem = availrmem - tune.t_minarmem;
3553 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3554 
3555 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3556 }
3