xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision 1c207ae9)
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, 2016 by Delphix. All rights reserved.
24  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
25  */
26 
27 /*
28  * Kernel memory allocator, as described in the following two papers and a
29  * statement about the consolidator:
30  *
31  * Jeff Bonwick,
32  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
33  * Proceedings of the Summer 1994 Usenix Conference.
34  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
35  *
36  * Jeff Bonwick and Jonathan Adams,
37  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
38  * Arbitrary Resources.
39  * Proceedings of the 2001 Usenix Conference.
40  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
41  *
42  * kmem Slab Consolidator Big Theory Statement:
43  *
44  * 1. Motivation
45  *
46  * As stated in Bonwick94, slabs provide the following advantages over other
47  * allocation structures in terms of memory fragmentation:
48  *
49  *  - Internal fragmentation (per-buffer wasted space) is minimal.
50  *  - Severe external fragmentation (unused buffers on the free list) is
51  *    unlikely.
52  *
53  * Segregating objects by size eliminates one source of external fragmentation,
54  * and according to Bonwick:
55  *
56  *   The other reason that slabs reduce external fragmentation is that all
57  *   objects in a slab are of the same type, so they have the same lifetime
58  *   distribution. The resulting segregation of short-lived and long-lived
59  *   objects at slab granularity reduces the likelihood of an entire page being
60  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
61  *
62  * While unlikely, severe external fragmentation remains possible. Clients that
63  * allocate both short- and long-lived objects from the same cache cannot
64  * anticipate the distribution of long-lived objects within the allocator's slab
65  * implementation. Even a small percentage of long-lived objects distributed
66  * randomly across many slabs can lead to a worst case scenario where the client
67  * frees the majority of its objects and the system gets back almost none of the
68  * slabs. Despite the client doing what it reasonably can to help the system
69  * reclaim memory, the allocator cannot shake free enough slabs because of
70  * lonely allocations stubbornly hanging on. Although the allocator is in a
71  * position to diagnose the fragmentation, there is nothing that the allocator
72  * by itself can do about it. It only takes a single allocated object to prevent
73  * an entire slab from being reclaimed, and any object handed out by
74  * kmem_cache_alloc() is by definition in the client's control. Conversely,
75  * although the client is in a position to move a long-lived object, it has no
76  * way of knowing if the object is causing fragmentation, and if so, where to
77  * move it. A solution necessarily requires further cooperation between the
78  * allocator and the client.
79  *
80  * 2. Move Callback
81  *
82  * The kmem slab consolidator therefore adds a move callback to the
83  * allocator/client interface, improving worst-case external fragmentation in
84  * kmem caches that supply a function to move objects from one memory location
85  * to another. In a situation of low memory kmem attempts to consolidate all of
86  * a cache's slabs at once; otherwise it works slowly to bring external
87  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
88  * thereby helping to avoid a low memory situation in the future.
89  *
90  * The callback has the following signature:
91  *
92  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
93  *
94  * It supplies the kmem client with two addresses: the allocated object that
95  * kmem wants to move and a buffer selected by kmem for the client to use as the
96  * copy destination. The callback is kmem's way of saying "Please get off of
97  * this buffer and use this one instead." kmem knows where it wants to move the
98  * object in order to best reduce fragmentation. All the client needs to know
99  * about the second argument (void *new) is that it is an allocated, constructed
100  * object ready to take the contents of the old object. When the move function
101  * is called, the system is likely to be low on memory, and the new object
102  * spares the client from having to worry about allocating memory for the
103  * requested move. The third argument supplies the size of the object, in case a
104  * single move function handles multiple caches whose objects differ only in
105  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
106  * user argument passed to the constructor, destructor, and reclaim functions is
107  * also passed to the move callback.
108  *
109  * 2.1 Setting the Move Callback
110  *
111  * The client sets the move callback after creating the cache and before
112  * allocating from it:
113  *
114  *	object_cache = kmem_cache_create(...);
115  *      kmem_cache_set_move(object_cache, object_move);
116  *
117  * 2.2 Move Callback Return Values
118  *
119  * Only the client knows about its own data and when is a good time to move it.
120  * The client is cooperating with kmem to return unused memory to the system,
121  * and kmem respectfully accepts this help at the client's convenience. When
122  * asked to move an object, the client can respond with any of the following:
123  *
124  *   typedef enum kmem_cbrc {
125  *           KMEM_CBRC_YES,
126  *           KMEM_CBRC_NO,
127  *           KMEM_CBRC_LATER,
128  *           KMEM_CBRC_DONT_NEED,
129  *           KMEM_CBRC_DONT_KNOW
130  *   } kmem_cbrc_t;
131  *
132  * The client must not explicitly kmem_cache_free() either of the objects passed
133  * to the callback, since kmem wants to free them directly to the slab layer
134  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
135  * objects to free:
136  *
137  *       YES: (Did it) The client moved the object, so kmem frees the old one.
138  *        NO: (Never) The client refused, so kmem frees the new object (the
139  *            unused copy destination). kmem also marks the slab of the old
140  *            object so as not to bother the client with further callbacks for
141  *            that object as long as the slab remains on the partial slab list.
142  *            (The system won't be getting the slab back as long as the
143  *            immovable object holds it hostage, so there's no point in moving
144  *            any of its objects.)
145  *     LATER: The client is using the object and cannot move it now, so kmem
146  *            frees the new object (the unused copy destination). kmem still
147  *            attempts to move other objects off the slab, since it expects to
148  *            succeed in clearing the slab in a later callback. The client
149  *            should use LATER instead of NO if the object is likely to become
150  *            movable very soon.
151  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
152  *            with the new object (the unused copy destination). This response
153  *            is the client's opportunity to be a model citizen and give back as
154  *            much as it can.
155  * DONT_KNOW: The client does not know about the object because
156  *            a) the client has just allocated the object and not yet put it
157  *               wherever it expects to find known objects
158  *            b) the client has removed the object from wherever it expects to
159  *               find known objects and is about to free it, or
160  *            c) the client has freed the object.
161  *            In all these cases (a, b, and c) kmem frees the new object (the
162  *            unused copy destination) and searches for the old object in the
163  *            magazine layer. If found, the object is removed from the magazine
164  *            layer and freed to the slab layer so it will no longer hold the
165  *            slab hostage.
166  *
167  * 2.3 Object States
168  *
169  * Neither kmem nor the client can be assumed to know the object's whereabouts
170  * at the time of the callback. An object belonging to a kmem cache may be in
171  * any of the following states:
172  *
173  * 1. Uninitialized on the slab
174  * 2. Allocated from the slab but not constructed (still uninitialized)
175  * 3. Allocated from the slab, constructed, but not yet ready for business
176  *    (not in a valid state for the move callback)
177  * 4. In use (valid and known to the client)
178  * 5. About to be freed (no longer in a valid state for the move callback)
179  * 6. Freed to a magazine (still constructed)
180  * 7. Allocated from a magazine, not yet ready for business (not in a valid
181  *    state for the move callback), and about to return to state #4
182  * 8. Deconstructed on a magazine that is about to be freed
183  * 9. Freed to the slab
184  *
185  * Since the move callback may be called at any time while the object is in any
186  * of the above states (except state #1), the client needs a safe way to
187  * determine whether or not it knows about the object. Specifically, the client
188  * needs to know whether or not the object is in state #4, the only state in
189  * which a move is valid. If the object is in any other state, the client should
190  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
191  * the object's fields.
192  *
193  * Note that although an object may be in state #4 when kmem initiates the move
194  * request, the object may no longer be in that state by the time kmem actually
195  * calls the move function. Not only does the client free objects
196  * asynchronously, kmem itself puts move requests on a queue where thay are
197  * pending until kmem processes them from another context. Also, objects freed
198  * to a magazine appear allocated from the point of view of the slab layer, so
199  * kmem may even initiate requests for objects in a state other than state #4.
200  *
201  * 2.3.1 Magazine Layer
202  *
203  * An important insight revealed by the states listed above is that the magazine
204  * layer is populated only by kmem_cache_free(). Magazines of constructed
205  * objects are never populated directly from the slab layer (which contains raw,
206  * unconstructed objects). Whenever an allocation request cannot be satisfied
207  * from the magazine layer, the magazines are bypassed and the request is
208  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
209  * the object constructor only when allocating from the slab layer, and only in
210  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
211  * the move callback. kmem does not preconstruct objects in anticipation of
212  * kmem_cache_alloc().
213  *
214  * 2.3.2 Object Constructor and Destructor
215  *
216  * If the client supplies a destructor, it must be valid to call the destructor
217  * on a newly created object (immediately after the constructor).
218  *
219  * 2.4 Recognizing Known Objects
220  *
221  * There is a simple test to determine safely whether or not the client knows
222  * about a given object in the move callback. It relies on the fact that kmem
223  * guarantees that the object of the move callback has only been touched by the
224  * client itself or else by kmem. kmem does this by ensuring that none of the
225  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
226  * callback is pending. When the last object on a slab is freed, if there is a
227  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
228  * slabs on that list until all pending callbacks are completed. That way,
229  * clients can be certain that the object of a move callback is in one of the
230  * states listed above, making it possible to distinguish known objects (in
231  * state #4) using the two low order bits of any pointer member (with the
232  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
233  * platforms).
234  *
235  * The test works as long as the client always transitions objects from state #4
236  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
237  * order bit of the client-designated pointer member. Since kmem only writes
238  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
239  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
240  * guaranteed to set at least one of the two low order bits. Therefore, given an
241  * object with a back pointer to a 'container_t *o_container', the client can
242  * test
243  *
244  *      container_t *container = object->o_container;
245  *      if ((uintptr_t)container & 0x3) {
246  *              return (KMEM_CBRC_DONT_KNOW);
247  *      }
248  *
249  * Typically, an object will have a pointer to some structure with a list or
250  * hash where objects from the cache are kept while in use. Assuming that the
251  * client has some way of knowing that the container structure is valid and will
252  * not go away during the move, and assuming that the structure includes a lock
253  * to protect whatever collection is used, then the client would continue as
254  * follows:
255  *
256  *	// Ensure that the container structure does not go away.
257  *      if (container_hold(container) == 0) {
258  *              return (KMEM_CBRC_DONT_KNOW);
259  *      }
260  *      mutex_enter(&container->c_objects_lock);
261  *      if (container != object->o_container) {
262  *              mutex_exit(&container->c_objects_lock);
263  *              container_rele(container);
264  *              return (KMEM_CBRC_DONT_KNOW);
265  *      }
266  *
267  * At this point the client knows that the object cannot be freed as long as
268  * c_objects_lock is held. Note that after acquiring the lock, the client must
269  * recheck the o_container pointer in case the object was removed just before
270  * acquiring the lock.
271  *
272  * When the client is about to free an object, it must first remove that object
273  * from the list, hash, or other structure where it is kept. At that time, to
274  * mark the object so it can be distinguished from the remaining, known objects,
275  * the client sets the designated low order bit:
276  *
277  *      mutex_enter(&container->c_objects_lock);
278  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
279  *      list_remove(&container->c_objects, object);
280  *      mutex_exit(&container->c_objects_lock);
281  *
282  * In the common case, the object is freed to the magazine layer, where it may
283  * be reused on a subsequent allocation without the overhead of calling the
284  * constructor. While in the magazine it appears allocated from the point of
285  * view of the slab layer, making it a candidate for the move callback. Most
286  * objects unrecognized by the client in the move callback fall into this
287  * category and are cheaply distinguished from known objects by the test
288  * described earlier. Since recognition is cheap for the client, and searching
289  * magazines is expensive for kmem, kmem defers searching until the client first
290  * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
291  * elsewhere does what it can to avoid bothering the client unnecessarily.
292  *
293  * Invalidating the designated pointer member before freeing the object marks
294  * the object to be avoided in the callback, and conversely, assigning a valid
295  * value to the designated pointer member after allocating the object makes the
296  * object fair game for the callback:
297  *
298  *      ... allocate object ...
299  *      ... set any initial state not set by the constructor ...
300  *
301  *      mutex_enter(&container->c_objects_lock);
302  *      list_insert_tail(&container->c_objects, object);
303  *      membar_producer();
304  *      object->o_container = container;
305  *      mutex_exit(&container->c_objects_lock);
306  *
307  * Note that everything else must be valid before setting o_container makes the
308  * object fair game for the move callback. The membar_producer() call ensures
309  * that all the object's state is written to memory before setting the pointer
310  * that transitions the object from state #3 or #7 (allocated, constructed, not
311  * yet in use) to state #4 (in use, valid). That's important because the move
312  * function has to check the validity of the pointer before it can safely
313  * acquire the lock protecting the collection where it expects to find known
314  * objects.
315  *
316  * This method of distinguishing known objects observes the usual symmetry:
317  * invalidating the designated pointer is the first thing the client does before
318  * freeing the object, and setting the designated pointer is the last thing the
319  * client does after allocating the object. Of course, the client is not
320  * required to use this method. Fundamentally, how the client recognizes known
321  * objects is completely up to the client, but this method is recommended as an
322  * efficient and safe way to take advantage of the guarantees made by kmem. If
323  * the entire object is arbitrary data without any markable bits from a suitable
324  * pointer member, then the client must find some other method, such as
325  * searching a hash table of known objects.
326  *
327  * 2.5 Preventing Objects From Moving
328  *
329  * Besides a way to distinguish known objects, the other thing that the client
330  * needs is a strategy to ensure that an object will not move while the client
331  * is actively using it. The details of satisfying this requirement tend to be
332  * highly cache-specific. It might seem that the same rules that let a client
333  * remove an object safely should also decide when an object can be moved
334  * safely. However, any object state that makes a removal attempt invalid is
335  * likely to be long-lasting for objects that the client does not expect to
336  * remove. kmem knows nothing about the object state and is equally likely (from
337  * the client's point of view) to request a move for any object in the cache,
338  * whether prepared for removal or not. Even a low percentage of objects stuck
339  * in place by unremovability will defeat the consolidator if the stuck objects
340  * are the same long-lived allocations likely to hold slabs hostage.
341  * Fundamentally, the consolidator is not aimed at common cases. Severe external
342  * fragmentation is a worst case scenario manifested as sparsely allocated
343  * slabs, by definition a low percentage of the cache's objects. When deciding
344  * what makes an object movable, keep in mind the goal of the consolidator: to
345  * bring worst-case external fragmentation within the limits guaranteed for
346  * internal fragmentation. Removability is a poor criterion if it is likely to
347  * exclude more than an insignificant percentage of objects for long periods of
348  * time.
349  *
350  * A tricky general solution exists, and it has the advantage of letting you
351  * move any object at almost any moment, practically eliminating the likelihood
352  * that an object can hold a slab hostage. However, if there is a cache-specific
353  * way to ensure that an object is not actively in use in the vast majority of
354  * cases, a simpler solution that leverages this cache-specific knowledge is
355  * preferred.
356  *
357  * 2.5.1 Cache-Specific Solution
358  *
359  * As an example of a cache-specific solution, the ZFS znode cache takes
360  * advantage of the fact that the vast majority of znodes are only being
361  * referenced from the DNLC. (A typical case might be a few hundred in active
362  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
363  * client has established that it recognizes the znode and can access its fields
364  * safely (using the method described earlier), it then tests whether the znode
365  * is referenced by anything other than the DNLC. If so, it assumes that the
366  * znode may be in active use and is unsafe to move, so it drops its locks and
367  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
368  * else znodes are used, no change is needed to protect against the possibility
369  * of the znode moving. The disadvantage is that it remains possible for an
370  * application to hold a znode slab hostage with an open file descriptor.
371  * However, this case ought to be rare and the consolidator has a way to deal
372  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
373  * object, kmem eventually stops believing it and treats the slab as if the
374  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
375  * then focus on getting it off of the partial slab list by allocating rather
376  * than freeing all of its objects. (Either way of getting a slab off the
377  * free list reduces fragmentation.)
378  *
379  * 2.5.2 General Solution
380  *
381  * The general solution, on the other hand, requires an explicit hold everywhere
382  * the object is used to prevent it from moving. To keep the client locking
383  * strategy as uncomplicated as possible, kmem guarantees the simplifying
384  * assumption that move callbacks are sequential, even across multiple caches.
385  * Internally, a global queue processed by a single thread supports all caches
386  * implementing the callback function. No matter how many caches supply a move
387  * function, the consolidator never moves more than one object at a time, so the
388  * client does not have to worry about tricky lock ordering involving several
389  * related objects from different kmem caches.
390  *
391  * The general solution implements the explicit hold as a read-write lock, which
392  * allows multiple readers to access an object from the cache simultaneously
393  * while a single writer is excluded from moving it. A single rwlock for the
394  * entire cache would lock out all threads from using any of the cache's objects
395  * even though only a single object is being moved, so to reduce contention,
396  * the client can fan out the single rwlock into an array of rwlocks hashed by
397  * the object address, making it probable that moving one object will not
398  * prevent other threads from using a different object. The rwlock cannot be a
399  * member of the object itself, because the possibility of the object moving
400  * makes it unsafe to access any of the object's fields until the lock is
401  * acquired.
402  *
403  * Assuming a small, fixed number of locks, it's possible that multiple objects
404  * will hash to the same lock. A thread that needs to use multiple objects in
405  * the same function may acquire the same lock multiple times. Since rwlocks are
406  * reentrant for readers, and since there is never more than a single writer at
407  * a time (assuming that the client acquires the lock as a writer only when
408  * moving an object inside the callback), there would seem to be no problem.
409  * However, a client locking multiple objects in the same function must handle
410  * one case of potential deadlock: Assume that thread A needs to prevent both
411  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
412  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
413  * same lock, that thread A will acquire the lock for object 1 as a reader
414  * before thread B sets the lock's write-wanted bit, preventing thread A from
415  * reacquiring the lock for object 2 as a reader. Unable to make forward
416  * progress, thread A will never release the lock for object 1, resulting in
417  * deadlock.
418  *
419  * There are two ways of avoiding the deadlock just described. The first is to
420  * use rw_tryenter() rather than rw_enter() in the callback function when
421  * attempting to acquire the lock as a writer. If tryenter discovers that the
422  * same object (or another object hashed to the same lock) is already in use, it
423  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
424  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
425  * since it allows a thread to acquire the lock as a reader in spite of a
426  * waiting writer. This second approach insists on moving the object now, no
427  * matter how many readers the move function must wait for in order to do so,
428  * and could delay the completion of the callback indefinitely (blocking
429  * callbacks to other clients). In practice, a less insistent callback using
430  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
431  * little reason to use anything else.
432  *
433  * Avoiding deadlock is not the only problem that an implementation using an
434  * explicit hold needs to solve. Locking the object in the first place (to
435  * prevent it from moving) remains a problem, since the object could move
436  * between the time you obtain a pointer to the object and the time you acquire
437  * the rwlock hashed to that pointer value. Therefore the client needs to
438  * recheck the value of the pointer after acquiring the lock, drop the lock if
439  * the value has changed, and try again. This requires a level of indirection:
440  * something that points to the object rather than the object itself, that the
441  * client can access safely while attempting to acquire the lock. (The object
442  * itself cannot be referenced safely because it can move at any time.)
443  * The following lock-acquisition function takes whatever is safe to reference
444  * (arg), follows its pointer to the object (using function f), and tries as
445  * often as necessary to acquire the hashed lock and verify that the object
446  * still has not moved:
447  *
448  *      object_t *
449  *      object_hold(object_f f, void *arg)
450  *      {
451  *              object_t *op;
452  *
453  *              op = f(arg);
454  *              if (op == NULL) {
455  *                      return (NULL);
456  *              }
457  *
458  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
459  *              while (op != f(arg)) {
460  *                      rw_exit(OBJECT_RWLOCK(op));
461  *                      op = f(arg);
462  *                      if (op == NULL) {
463  *                              break;
464  *                      }
465  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
466  *              }
467  *
468  *              return (op);
469  *      }
470  *
471  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
472  * lock reacquisition loop, while necessary, almost never executes. The function
473  * pointer f (used to obtain the object pointer from arg) has the following type
474  * definition:
475  *
476  *      typedef object_t *(*object_f)(void *arg);
477  *
478  * An object_f implementation is likely to be as simple as accessing a structure
479  * member:
480  *
481  *      object_t *
482  *      s_object(void *arg)
483  *      {
484  *              something_t *sp = arg;
485  *              return (sp->s_object);
486  *      }
487  *
488  * The flexibility of a function pointer allows the path to the object to be
489  * arbitrarily complex and also supports the notion that depending on where you
490  * are using the object, you may need to get it from someplace different.
491  *
492  * The function that releases the explicit hold is simpler because it does not
493  * have to worry about the object moving:
494  *
495  *      void
496  *      object_rele(object_t *op)
497  *      {
498  *              rw_exit(OBJECT_RWLOCK(op));
499  *      }
500  *
501  * The caller is spared these details so that obtaining and releasing an
502  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
503  * of object_hold() only needs to know that the returned object pointer is valid
504  * if not NULL and that the object will not move until released.
505  *
506  * Although object_hold() prevents an object from moving, it does not prevent it
507  * from being freed. The caller must take measures before calling object_hold()
508  * (afterwards is too late) to ensure that the held object cannot be freed. The
509  * caller must do so without accessing the unsafe object reference, so any lock
510  * or reference count used to ensure the continued existence of the object must
511  * live outside the object itself.
512  *
513  * Obtaining a new object is a special case where an explicit hold is impossible
514  * for the caller. Any function that returns a newly allocated object (either as
515  * a return value, or as an in-out paramter) must return it already held; after
516  * the caller gets it is too late, since the object cannot be safely accessed
517  * without the level of indirection described earlier. The following
518  * object_alloc() example uses the same code shown earlier to transition a new
519  * object into the state of being recognized (by the client) as a known object.
520  * The function must acquire the hold (rw_enter) before that state transition
521  * makes the object movable:
522  *
523  *      static object_t *
524  *      object_alloc(container_t *container)
525  *      {
526  *              object_t *object = kmem_cache_alloc(object_cache, 0);
527  *              ... set any initial state not set by the constructor ...
528  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
529  *              mutex_enter(&container->c_objects_lock);
530  *              list_insert_tail(&container->c_objects, object);
531  *              membar_producer();
532  *              object->o_container = container;
533  *              mutex_exit(&container->c_objects_lock);
534  *              return (object);
535  *      }
536  *
537  * Functions that implicitly acquire an object hold (any function that calls
538  * object_alloc() to supply an object for the caller) need to be carefully noted
539  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
540  * prevent all objects hashed to the affected rwlocks from ever being moved.
541  *
542  * The pointer to a held object can be hashed to the holding rwlock even after
543  * the object has been freed. Although it is possible to release the hold
544  * after freeing the object, you may decide to release the hold implicitly in
545  * whatever function frees the object, so as to release the hold as soon as
546  * possible, and for the sake of symmetry with the function that implicitly
547  * acquires the hold when it allocates the object. Here, object_free() releases
548  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
549  * matching pair with object_hold():
550  *
551  *      void
552  *      object_free(object_t *object)
553  *      {
554  *              container_t *container;
555  *
556  *              ASSERT(object_held(object));
557  *              container = object->o_container;
558  *              mutex_enter(&container->c_objects_lock);
559  *              object->o_container =
560  *                  (void *)((uintptr_t)object->o_container | 0x1);
561  *              list_remove(&container->c_objects, object);
562  *              mutex_exit(&container->c_objects_lock);
563  *              object_rele(object);
564  *              kmem_cache_free(object_cache, object);
565  *      }
566  *
567  * Note that object_free() cannot safely accept an object pointer as an argument
568  * unless the object is already held. Any function that calls object_free()
569  * needs to be carefully noted since it similarly forms a matching pair with
570  * object_hold().
571  *
572  * To complete the picture, the following callback function implements the
573  * general solution by moving objects only if they are currently unheld:
574  *
575  *      static kmem_cbrc_t
576  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
577  *      {
578  *              object_t *op = buf, *np = newbuf;
579  *              container_t *container;
580  *
581  *              container = op->o_container;
582  *              if ((uintptr_t)container & 0x3) {
583  *                      return (KMEM_CBRC_DONT_KNOW);
584  *              }
585  *
586  *	        // Ensure that the container structure does not go away.
587  *              if (container_hold(container) == 0) {
588  *                      return (KMEM_CBRC_DONT_KNOW);
589  *              }
590  *
591  *              mutex_enter(&container->c_objects_lock);
592  *              if (container != op->o_container) {
593  *                      mutex_exit(&container->c_objects_lock);
594  *                      container_rele(container);
595  *                      return (KMEM_CBRC_DONT_KNOW);
596  *              }
597  *
598  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
599  *                      mutex_exit(&container->c_objects_lock);
600  *                      container_rele(container);
601  *                      return (KMEM_CBRC_LATER);
602  *              }
603  *
604  *              object_move_impl(op, np); // critical section
605  *              rw_exit(OBJECT_RWLOCK(op));
606  *
607  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
608  *              list_link_replace(&op->o_link_node, &np->o_link_node);
609  *              mutex_exit(&container->c_objects_lock);
610  *              container_rele(container);
611  *              return (KMEM_CBRC_YES);
612  *      }
613  *
614  * Note that object_move() must invalidate the designated o_container pointer of
615  * the old object in the same way that object_free() does, since kmem will free
616  * the object in response to the KMEM_CBRC_YES return value.
617  *
618  * The lock order in object_move() differs from object_alloc(), which locks
619  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
620  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
621  * not a problem. Holding the lock on the object list in the example above
622  * through the entire callback not only prevents the object from going away, it
623  * also allows you to lock the list elsewhere and know that none of its elements
624  * will move during iteration.
625  *
626  * Adding an explicit hold everywhere an object from the cache is used is tricky
627  * and involves much more change to client code than a cache-specific solution
628  * that leverages existing state to decide whether or not an object is
629  * movable. However, this approach has the advantage that no object remains
630  * immovable for any significant length of time, making it extremely unlikely
631  * that long-lived allocations can continue holding slabs hostage; and it works
632  * for any cache.
633  *
634  * 3. Consolidator Implementation
635  *
636  * Once the client supplies a move function that a) recognizes known objects and
637  * b) avoids moving objects that are actively in use, the remaining work is up
638  * to the consolidator to decide which objects to move and when to issue
639  * callbacks.
640  *
641  * The consolidator relies on the fact that a cache's slabs are ordered by
642  * usage. Each slab has a fixed number of objects. Depending on the slab's
643  * "color" (the offset of the first object from the beginning of the slab;
644  * offsets are staggered to mitigate false sharing of cache lines) it is either
645  * the maximum number of objects per slab determined at cache creation time or
646  * else the number closest to the maximum that fits within the space remaining
647  * after the initial offset. A completely allocated slab may contribute some
648  * internal fragmentation (per-slab overhead) but no external fragmentation, so
649  * it is of no interest to the consolidator. At the other extreme, slabs whose
650  * objects have all been freed to the slab are released to the virtual memory
651  * (VM) subsystem (objects freed to magazines are still allocated as far as the
652  * slab is concerned). External fragmentation exists when there are slabs
653  * somewhere between these extremes. A partial slab has at least one but not all
654  * of its objects allocated. The more partial slabs, and the fewer allocated
655  * objects on each of them, the higher the fragmentation. Hence the
656  * consolidator's overall strategy is to reduce the number of partial slabs by
657  * moving allocated objects from the least allocated slabs to the most allocated
658  * slabs.
659  *
660  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
661  * slabs are kept separately in an unordered list. Since the majority of slabs
662  * tend to be completely allocated (a typical unfragmented cache may have
663  * thousands of complete slabs and only a single partial slab), separating
664  * complete slabs improves the efficiency of partial slab ordering, since the
665  * complete slabs do not affect the depth or balance of the AVL tree. This
666  * ordered sequence of partial slabs acts as a "free list" supplying objects for
667  * allocation requests.
668  *
669  * Objects are always allocated from the first partial slab in the free list,
670  * where the allocation is most likely to eliminate a partial slab (by
671  * completely allocating it). Conversely, when a single object from a completely
672  * allocated slab is freed to the slab, that slab is added to the front of the
673  * free list. Since most free list activity involves highly allocated slabs
674  * coming and going at the front of the list, slabs tend naturally toward the
675  * ideal order: highly allocated at the front, sparsely allocated at the back.
676  * Slabs with few allocated objects are likely to become completely free if they
677  * keep a safe distance away from the front of the free list. Slab misorders
678  * interfere with the natural tendency of slabs to become completely free or
679  * completely allocated. For example, a slab with a single allocated object
680  * needs only a single free to escape the cache; its natural desire is
681  * frustrated when it finds itself at the front of the list where a second
682  * allocation happens just before the free could have released it. Another slab
683  * with all but one object allocated might have supplied the buffer instead, so
684  * that both (as opposed to neither) of the slabs would have been taken off the
685  * free list.
686  *
687  * Although slabs tend naturally toward the ideal order, misorders allowed by a
688  * simple list implementation defeat the consolidator's strategy of merging
689  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
690  * needs another way to fix misorders to optimize its callback strategy. One
691  * approach is to periodically scan a limited number of slabs, advancing a
692  * marker to hold the current scan position, and to move extreme misorders to
693  * the front or back of the free list and to the front or back of the current
694  * scan range. By making consecutive scan ranges overlap by one slab, the least
695  * allocated slab in the current range can be carried along from the end of one
696  * scan to the start of the next.
697  *
698  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
699  * task, however. Since most of the cache's activity is in the magazine layer,
700  * and allocations from the slab layer represent only a startup cost, the
701  * overhead of maintaining a balanced tree is not a significant concern compared
702  * to the opportunity of reducing complexity by eliminating the partial slab
703  * scanner just described. The overhead of an AVL tree is minimized by
704  * maintaining only partial slabs in the tree and keeping completely allocated
705  * slabs separately in a list. To avoid increasing the size of the slab
706  * structure the AVL linkage pointers are reused for the slab's list linkage,
707  * since the slab will always be either partial or complete, never stored both
708  * ways at the same time. To further minimize the overhead of the AVL tree the
709  * compare function that orders partial slabs by usage divides the range of
710  * allocated object counts into bins such that counts within the same bin are
711  * considered equal. Binning partial slabs makes it less likely that allocating
712  * or freeing a single object will change the slab's order, requiring a tree
713  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
714  * requiring some rebalancing of the tree). Allocation counts closest to
715  * completely free and completely allocated are left unbinned (finely sorted) to
716  * better support the consolidator's strategy of merging slabs at either
717  * extreme.
718  *
719  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
720  *
721  * The consolidator piggybacks on the kmem maintenance thread and is called on
722  * the same interval as kmem_cache_update(), once per cache every fifteen
723  * seconds. kmem maintains a running count of unallocated objects in the slab
724  * layer (cache_bufslab). The consolidator checks whether that number exceeds
725  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
726  * there is a significant number of slabs in the cache (arbitrarily a minimum
727  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
728  * working set are included in the assessment, and magazines in the depot are
729  * reaped if those objects would lift cache_bufslab above the fragmentation
730  * threshold. Once the consolidator decides that a cache is fragmented, it looks
731  * for a candidate slab to reclaim, starting at the end of the partial slab free
732  * list and scanning backwards. At first the consolidator is choosy: only a slab
733  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
734  * single allocated object, regardless of percentage). If there is difficulty
735  * finding a candidate slab, kmem raises the allocation threshold incrementally,
736  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
737  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
738  * even in the worst case of every slab in the cache being almost 7/8 allocated.
739  * The threshold can also be lowered incrementally when candidate slabs are easy
740  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
741  * is no longer fragmented.
742  *
743  * 3.2 Generating Callbacks
744  *
745  * Once an eligible slab is chosen, a callback is generated for every allocated
746  * object on the slab, in the hope that the client will move everything off the
747  * slab and make it reclaimable. Objects selected as move destinations are
748  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
749  * order (most allocated at the front, least allocated at the back) and a
750  * cooperative client, the consolidator will succeed in removing slabs from both
751  * ends of the free list, completely allocating on the one hand and completely
752  * freeing on the other. Objects selected as move destinations are allocated in
753  * the kmem maintenance thread where move requests are enqueued. A separate
754  * callback thread removes pending callbacks from the queue and calls the
755  * client. The separate thread ensures that client code (the move function) does
756  * not interfere with internal kmem maintenance tasks. A map of pending
757  * callbacks keyed by object address (the object to be moved) is checked to
758  * ensure that duplicate callbacks are not generated for the same object.
759  * Allocating the move destination (the object to move to) prevents subsequent
760  * callbacks from selecting the same destination as an earlier pending callback.
761  *
762  * Move requests can also be generated by kmem_cache_reap() when the system is
763  * desperate for memory and by kmem_cache_move_notify(), called by the client to
764  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
765  * The map of pending callbacks is protected by the same lock that protects the
766  * slab layer.
767  *
768  * When the system is desperate for memory, kmem does not bother to determine
769  * whether or not the cache exceeds the fragmentation threshold, but tries to
770  * consolidate as many slabs as possible. Normally, the consolidator chews
771  * slowly, one sparsely allocated slab at a time during each maintenance
772  * interval that the cache is fragmented. When desperate, the consolidator
773  * starts at the last partial slab and enqueues callbacks for every allocated
774  * object on every partial slab, working backwards until it reaches the first
775  * partial slab. The first partial slab, meanwhile, advances in pace with the
776  * consolidator as allocations to supply move destinations for the enqueued
777  * callbacks use up the highly allocated slabs at the front of the free list.
778  * Ideally, the overgrown free list collapses like an accordion, starting at
779  * both ends and ending at the center with a single partial slab.
780  *
781  * 3.3 Client Responses
782  *
783  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
784  * marks the slab that supplied the stuck object non-reclaimable and moves it to
785  * front of the free list. The slab remains marked as long as it remains on the
786  * free list, and it appears more allocated to the partial slab compare function
787  * than any unmarked slab, no matter how many of its objects are allocated.
788  * Since even one immovable object ties up the entire slab, the goal is to
789  * completely allocate any slab that cannot be completely freed. kmem does not
790  * bother generating callbacks to move objects from a marked slab unless the
791  * system is desperate.
792  *
793  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
794  * slab. If the client responds LATER too many times, kmem disbelieves and
795  * treats the response as a NO. The count is cleared when the slab is taken off
796  * the partial slab list or when the client moves one of the slab's objects.
797  *
798  * 4. Observability
799  *
800  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
801  * the ::kmem_slabs dcmd. For a complete description of the command, enter
802  * '::help kmem_slabs' at the mdb prompt.
803  */
804 
805 #include <sys/kmem_impl.h>
806 #include <sys/vmem_impl.h>
807 #include <sys/param.h>
808 #include <sys/sysmacros.h>
809 #include <sys/vm.h>
810 #include <sys/proc.h>
811 #include <sys/tuneable.h>
812 #include <sys/systm.h>
813 #include <sys/cmn_err.h>
814 #include <sys/debug.h>
815 #include <sys/sdt.h>
816 #include <sys/mutex.h>
817 #include <sys/bitmap.h>
818 #include <sys/atomic.h>
819 #include <sys/kobj.h>
820 #include <sys/disp.h>
821 #include <vm/seg_kmem.h>
822 #include <sys/log.h>
823 #include <sys/callb.h>
824 #include <sys/taskq.h>
825 #include <sys/modctl.h>
826 #include <sys/reboot.h>
827 #include <sys/id32.h>
828 #include <sys/zone.h>
829 #include <sys/netstack.h>
830 #ifdef	DEBUG
831 #include <sys/random.h>
832 #endif
833 
834 extern void streams_msg_init(void);
835 extern int segkp_fromheap;
836 extern void segkp_cache_free(void);
837 extern int callout_init_done;
838 
839 struct kmem_cache_kstat {
840 	kstat_named_t	kmc_buf_size;
841 	kstat_named_t	kmc_align;
842 	kstat_named_t	kmc_chunk_size;
843 	kstat_named_t	kmc_slab_size;
844 	kstat_named_t	kmc_alloc;
845 	kstat_named_t	kmc_alloc_fail;
846 	kstat_named_t	kmc_free;
847 	kstat_named_t	kmc_depot_alloc;
848 	kstat_named_t	kmc_depot_free;
849 	kstat_named_t	kmc_depot_contention;
850 	kstat_named_t	kmc_slab_alloc;
851 	kstat_named_t	kmc_slab_free;
852 	kstat_named_t	kmc_buf_constructed;
853 	kstat_named_t	kmc_buf_avail;
854 	kstat_named_t	kmc_buf_inuse;
855 	kstat_named_t	kmc_buf_total;
856 	kstat_named_t	kmc_buf_max;
857 	kstat_named_t	kmc_slab_create;
858 	kstat_named_t	kmc_slab_destroy;
859 	kstat_named_t	kmc_vmem_source;
860 	kstat_named_t	kmc_hash_size;
861 	kstat_named_t	kmc_hash_lookup_depth;
862 	kstat_named_t	kmc_hash_rescale;
863 	kstat_named_t	kmc_full_magazines;
864 	kstat_named_t	kmc_empty_magazines;
865 	kstat_named_t	kmc_magazine_size;
866 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
867 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
868 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
869 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
870 	kstat_named_t	kmc_move_yes;
871 	kstat_named_t	kmc_move_no;
872 	kstat_named_t	kmc_move_later;
873 	kstat_named_t	kmc_move_dont_need;
874 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
875 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
876 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
877 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
878 } kmem_cache_kstat = {
879 	{ "buf_size",		KSTAT_DATA_UINT64 },
880 	{ "align",		KSTAT_DATA_UINT64 },
881 	{ "chunk_size",		KSTAT_DATA_UINT64 },
882 	{ "slab_size",		KSTAT_DATA_UINT64 },
883 	{ "alloc",		KSTAT_DATA_UINT64 },
884 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
885 	{ "free",		KSTAT_DATA_UINT64 },
886 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
887 	{ "depot_free",		KSTAT_DATA_UINT64 },
888 	{ "depot_contention",	KSTAT_DATA_UINT64 },
889 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
890 	{ "slab_free",		KSTAT_DATA_UINT64 },
891 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
892 	{ "buf_avail",		KSTAT_DATA_UINT64 },
893 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
894 	{ "buf_total",		KSTAT_DATA_UINT64 },
895 	{ "buf_max",		KSTAT_DATA_UINT64 },
896 	{ "slab_create",	KSTAT_DATA_UINT64 },
897 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
898 	{ "vmem_source",	KSTAT_DATA_UINT64 },
899 	{ "hash_size",		KSTAT_DATA_UINT64 },
900 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
901 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
902 	{ "full_magazines",	KSTAT_DATA_UINT64 },
903 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
904 	{ "magazine_size",	KSTAT_DATA_UINT64 },
905 	{ "reap",		KSTAT_DATA_UINT64 },
906 	{ "defrag",		KSTAT_DATA_UINT64 },
907 	{ "scan",		KSTAT_DATA_UINT64 },
908 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
909 	{ "move_yes",		KSTAT_DATA_UINT64 },
910 	{ "move_no",		KSTAT_DATA_UINT64 },
911 	{ "move_later",		KSTAT_DATA_UINT64 },
912 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
913 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
914 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
915 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
916 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
917 };
918 
919 static kmutex_t kmem_cache_kstat_lock;
920 
921 /*
922  * The default set of caches to back kmem_alloc().
923  * These sizes should be reevaluated periodically.
924  *
925  * We want allocations that are multiples of the coherency granularity
926  * (64 bytes) to be satisfied from a cache which is a multiple of 64
927  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
928  * the next kmem_cache_size greater than or equal to it must be a
929  * multiple of 64.
930  *
931  * We split the table into two sections:  size <= 4k and size > 4k.  This
932  * saves a lot of space and cache footprint in our cache tables.
933  */
934 static const int kmem_alloc_sizes[] = {
935 	1 * 8,
936 	2 * 8,
937 	3 * 8,
938 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
939 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
940 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
941 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
942 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
943 	P2ALIGN(8192 / 7, 64),
944 	P2ALIGN(8192 / 6, 64),
945 	P2ALIGN(8192 / 5, 64),
946 	P2ALIGN(8192 / 4, 64),
947 	P2ALIGN(8192 / 3, 64),
948 	P2ALIGN(8192 / 2, 64),
949 };
950 
951 static const int kmem_big_alloc_sizes[] = {
952 	2 * 4096,	3 * 4096,
953 	2 * 8192,	3 * 8192,
954 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
955 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
956 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
957 	16 * 8192
958 };
959 
960 #define	KMEM_MAXBUF		4096
961 #define	KMEM_BIG_MAXBUF_32BIT	32768
962 #define	KMEM_BIG_MAXBUF		131072
963 
964 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
965 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
966 
967 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
968 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
969 
970 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
971 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
972 
973 static kmem_magtype_t kmem_magtype[] = {
974 	{ 1,	8,	3200,	65536	},
975 	{ 3,	16,	256,	32768	},
976 	{ 7,	32,	64,	16384	},
977 	{ 15,	64,	0,	8192	},
978 	{ 31,	64,	0,	4096	},
979 	{ 47,	64,	0,	2048	},
980 	{ 63,	64,	0,	1024	},
981 	{ 95,	64,	0,	512	},
982 	{ 143,	64,	0,	0	},
983 };
984 
985 static uint32_t kmem_reaping;
986 static uint32_t kmem_reaping_idspace;
987 
988 /*
989  * kmem tunables
990  */
991 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
992 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
993 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
994 int kmem_panic = 1;		/* whether to panic on error */
995 int kmem_logging = 1;		/* kmem_log_enter() override */
996 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
997 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
998 size_t kmem_content_log_size;	/* content log size [2% of memory] */
999 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
1000 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
1001 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1002 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1003 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1004 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1005 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1006 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1007 
1008 #ifdef _LP64
1009 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1010 #else
1011 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1012 #endif
1013 
1014 #ifdef DEBUG
1015 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1016 #else
1017 int kmem_flags = 0;
1018 #endif
1019 int kmem_ready;
1020 
1021 static kmem_cache_t	*kmem_slab_cache;
1022 static kmem_cache_t	*kmem_bufctl_cache;
1023 static kmem_cache_t	*kmem_bufctl_audit_cache;
1024 
1025 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1026 static list_t		kmem_caches;
1027 
1028 static taskq_t		*kmem_taskq;
1029 static kmutex_t		kmem_flags_lock;
1030 static vmem_t		*kmem_metadata_arena;
1031 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1032 static vmem_t		*kmem_cache_arena;
1033 static vmem_t		*kmem_hash_arena;
1034 static vmem_t		*kmem_log_arena;
1035 static vmem_t		*kmem_oversize_arena;
1036 static vmem_t		*kmem_va_arena;
1037 static vmem_t		*kmem_default_arena;
1038 static vmem_t		*kmem_firewall_va_arena;
1039 static vmem_t		*kmem_firewall_arena;
1040 
1041 /*
1042  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1043  * turned on when DEBUG is also defined.
1044  */
1045 #ifdef	DEBUG
1046 #define	KMEM_STATS
1047 #endif	/* DEBUG */
1048 
1049 #ifdef	KMEM_STATS
1050 #define	KMEM_STAT_ADD(stat)			((stat)++)
1051 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1052 #else
1053 #define	KMEM_STAT_ADD(stat)			/* nothing */
1054 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1055 #endif	/* KMEM_STATS */
1056 
1057 /*
1058  * kmem slab consolidator thresholds (tunables)
1059  */
1060 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1061 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1062 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1063 /*
1064  * Maximum number of slabs from which to move buffers during a single
1065  * maintenance interval while the system is not low on memory.
1066  */
1067 size_t kmem_reclaim_max_slabs = 1;
1068 /*
1069  * Number of slabs to scan backwards from the end of the partial slab list
1070  * when searching for buffers to relocate.
1071  */
1072 size_t kmem_reclaim_scan_range = 12;
1073 
1074 #ifdef	KMEM_STATS
1075 static struct {
1076 	uint64_t kms_callbacks;
1077 	uint64_t kms_yes;
1078 	uint64_t kms_no;
1079 	uint64_t kms_later;
1080 	uint64_t kms_dont_need;
1081 	uint64_t kms_dont_know;
1082 	uint64_t kms_hunt_found_mag;
1083 	uint64_t kms_hunt_found_slab;
1084 	uint64_t kms_hunt_alloc_fail;
1085 	uint64_t kms_hunt_lucky;
1086 	uint64_t kms_notify;
1087 	uint64_t kms_notify_callbacks;
1088 	uint64_t kms_disbelief;
1089 	uint64_t kms_already_pending;
1090 	uint64_t kms_callback_alloc_fail;
1091 	uint64_t kms_callback_taskq_fail;
1092 	uint64_t kms_endscan_slab_dead;
1093 	uint64_t kms_endscan_slab_destroyed;
1094 	uint64_t kms_endscan_nomem;
1095 	uint64_t kms_endscan_refcnt_changed;
1096 	uint64_t kms_endscan_nomove_changed;
1097 	uint64_t kms_endscan_freelist;
1098 	uint64_t kms_avl_update;
1099 	uint64_t kms_avl_noupdate;
1100 	uint64_t kms_no_longer_reclaimable;
1101 	uint64_t kms_notify_no_longer_reclaimable;
1102 	uint64_t kms_notify_slab_dead;
1103 	uint64_t kms_notify_slab_destroyed;
1104 	uint64_t kms_alloc_fail;
1105 	uint64_t kms_constructor_fail;
1106 	uint64_t kms_dead_slabs_freed;
1107 	uint64_t kms_defrags;
1108 	uint64_t kms_scans;
1109 	uint64_t kms_scan_depot_ws_reaps;
1110 	uint64_t kms_debug_reaps;
1111 	uint64_t kms_debug_scans;
1112 } kmem_move_stats;
1113 #endif	/* KMEM_STATS */
1114 
1115 /* consolidator knobs */
1116 static boolean_t kmem_move_noreap;
1117 static boolean_t kmem_move_blocked;
1118 static boolean_t kmem_move_fulltilt;
1119 static boolean_t kmem_move_any_partial;
1120 
1121 #ifdef	DEBUG
1122 /*
1123  * kmem consolidator debug tunables:
1124  * Ensure code coverage by occasionally running the consolidator even when the
1125  * caches are not fragmented (they may never be). These intervals are mean time
1126  * in cache maintenance intervals (kmem_cache_update).
1127  */
1128 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1129 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1130 #endif	/* DEBUG */
1131 
1132 static kmem_cache_t	*kmem_defrag_cache;
1133 static kmem_cache_t	*kmem_move_cache;
1134 static taskq_t		*kmem_move_taskq;
1135 
1136 static void kmem_cache_scan(kmem_cache_t *);
1137 static void kmem_cache_defrag(kmem_cache_t *);
1138 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1139 
1140 
1141 kmem_log_header_t	*kmem_transaction_log;
1142 kmem_log_header_t	*kmem_content_log;
1143 kmem_log_header_t	*kmem_failure_log;
1144 kmem_log_header_t	*kmem_slab_log;
1145 
1146 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1147 
1148 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1149 	if ((count) > 0) {						\
1150 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1151 		pc_t *_e;						\
1152 		/* memmove() the old entries down one notch */		\
1153 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1154 			*_e = *(_e - 1);				\
1155 		*_s = (uintptr_t)(caller);				\
1156 	}
1157 
1158 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1159 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1160 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1161 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1162 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1163 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1164 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1165 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1166 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1167 
1168 struct {
1169 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1170 	int		kmp_error;	/* type of kmem error */
1171 	void		*kmp_buffer;	/* buffer that induced panic */
1172 	void		*kmp_realbuf;	/* real start address for buffer */
1173 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1174 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1175 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1176 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1177 } kmem_panic_info;
1178 
1179 
1180 static void
1181 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1182 {
1183 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1184 	uint64_t *buf = buf_arg;
1185 
1186 	while (buf < bufend)
1187 		*buf++ = pattern;
1188 }
1189 
1190 static void *
1191 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1192 {
1193 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1194 	uint64_t *buf;
1195 
1196 	for (buf = buf_arg; buf < bufend; buf++)
1197 		if (*buf != pattern)
1198 			return (buf);
1199 	return (NULL);
1200 }
1201 
1202 static void *
1203 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1204 {
1205 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1206 	uint64_t *buf;
1207 
1208 	for (buf = buf_arg; buf < bufend; buf++) {
1209 		if (*buf != old) {
1210 			copy_pattern(old, buf_arg,
1211 			    (char *)buf - (char *)buf_arg);
1212 			return (buf);
1213 		}
1214 		*buf = new;
1215 	}
1216 
1217 	return (NULL);
1218 }
1219 
1220 static void
1221 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1222 {
1223 	kmem_cache_t *cp;
1224 
1225 	mutex_enter(&kmem_cache_lock);
1226 	for (cp = list_head(&kmem_caches); cp != NULL;
1227 	    cp = list_next(&kmem_caches, cp))
1228 		if (tq != NULL)
1229 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1230 			    tqflag);
1231 		else
1232 			func(cp);
1233 	mutex_exit(&kmem_cache_lock);
1234 }
1235 
1236 static void
1237 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1238 {
1239 	kmem_cache_t *cp;
1240 
1241 	mutex_enter(&kmem_cache_lock);
1242 	for (cp = list_head(&kmem_caches); cp != NULL;
1243 	    cp = list_next(&kmem_caches, cp)) {
1244 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1245 			continue;
1246 		if (tq != NULL)
1247 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1248 			    tqflag);
1249 		else
1250 			func(cp);
1251 	}
1252 	mutex_exit(&kmem_cache_lock);
1253 }
1254 
1255 /*
1256  * Debugging support.  Given a buffer address, find its slab.
1257  */
1258 static kmem_slab_t *
1259 kmem_findslab(kmem_cache_t *cp, void *buf)
1260 {
1261 	kmem_slab_t *sp;
1262 
1263 	mutex_enter(&cp->cache_lock);
1264 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1265 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1266 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1267 			mutex_exit(&cp->cache_lock);
1268 			return (sp);
1269 		}
1270 	}
1271 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1272 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1273 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1274 			mutex_exit(&cp->cache_lock);
1275 			return (sp);
1276 		}
1277 	}
1278 	mutex_exit(&cp->cache_lock);
1279 
1280 	return (NULL);
1281 }
1282 
1283 static void
1284 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1285 {
1286 	kmem_buftag_t *btp = NULL;
1287 	kmem_bufctl_t *bcp = NULL;
1288 	kmem_cache_t *cp = cparg;
1289 	kmem_slab_t *sp;
1290 	uint64_t *off;
1291 	void *buf = bufarg;
1292 
1293 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1294 
1295 	kmem_panic_info.kmp_timestamp = gethrtime();
1296 
1297 	sp = kmem_findslab(cp, buf);
1298 	if (sp == NULL) {
1299 		for (cp = list_tail(&kmem_caches); cp != NULL;
1300 		    cp = list_prev(&kmem_caches, cp)) {
1301 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1302 				break;
1303 		}
1304 	}
1305 
1306 	if (sp == NULL) {
1307 		cp = NULL;
1308 		error = KMERR_BADADDR;
1309 	} else {
1310 		if (cp != cparg)
1311 			error = KMERR_BADCACHE;
1312 		else
1313 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1314 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1315 		if (buf != bufarg)
1316 			error = KMERR_BADBASE;
1317 		if (cp->cache_flags & KMF_BUFTAG)
1318 			btp = KMEM_BUFTAG(cp, buf);
1319 		if (cp->cache_flags & KMF_HASH) {
1320 			mutex_enter(&cp->cache_lock);
1321 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1322 				if (bcp->bc_addr == buf)
1323 					break;
1324 			mutex_exit(&cp->cache_lock);
1325 			if (bcp == NULL && btp != NULL)
1326 				bcp = btp->bt_bufctl;
1327 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1328 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1329 			    bcp->bc_addr != buf) {
1330 				error = KMERR_BADBUFCTL;
1331 				bcp = NULL;
1332 			}
1333 		}
1334 	}
1335 
1336 	kmem_panic_info.kmp_error = error;
1337 	kmem_panic_info.kmp_buffer = bufarg;
1338 	kmem_panic_info.kmp_realbuf = buf;
1339 	kmem_panic_info.kmp_cache = cparg;
1340 	kmem_panic_info.kmp_realcache = cp;
1341 	kmem_panic_info.kmp_slab = sp;
1342 	kmem_panic_info.kmp_bufctl = bcp;
1343 
1344 	printf("kernel memory allocator: ");
1345 
1346 	switch (error) {
1347 
1348 	case KMERR_MODIFIED:
1349 		printf("buffer modified after being freed\n");
1350 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1351 		if (off == NULL)	/* shouldn't happen */
1352 			off = buf;
1353 		printf("modification occurred at offset 0x%lx "
1354 		    "(0x%llx replaced by 0x%llx)\n",
1355 		    (uintptr_t)off - (uintptr_t)buf,
1356 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1357 		break;
1358 
1359 	case KMERR_REDZONE:
1360 		printf("redzone violation: write past end of buffer\n");
1361 		break;
1362 
1363 	case KMERR_BADADDR:
1364 		printf("invalid free: buffer not in cache\n");
1365 		break;
1366 
1367 	case KMERR_DUPFREE:
1368 		printf("duplicate free: buffer freed twice\n");
1369 		break;
1370 
1371 	case KMERR_BADBUFTAG:
1372 		printf("boundary tag corrupted\n");
1373 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1374 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1375 		    KMEM_BUFTAG_FREE);
1376 		break;
1377 
1378 	case KMERR_BADBUFCTL:
1379 		printf("bufctl corrupted\n");
1380 		break;
1381 
1382 	case KMERR_BADCACHE:
1383 		printf("buffer freed to wrong cache\n");
1384 		printf("buffer was allocated from %s,\n", cp->cache_name);
1385 		printf("caller attempting free to %s.\n", cparg->cache_name);
1386 		break;
1387 
1388 	case KMERR_BADSIZE:
1389 		printf("bad free: free size (%u) != alloc size (%u)\n",
1390 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1391 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1392 		break;
1393 
1394 	case KMERR_BADBASE:
1395 		printf("bad free: free address (%p) != alloc address (%p)\n",
1396 		    bufarg, buf);
1397 		break;
1398 	}
1399 
1400 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1401 	    bufarg, (void *)bcp, cparg->cache_name);
1402 
1403 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1404 	    error != KMERR_BADBUFCTL) {
1405 		int d;
1406 		timestruc_t ts;
1407 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1408 
1409 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1410 		printf("previous transaction on buffer %p:\n", buf);
1411 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1412 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1413 		    (void *)sp, cp->cache_name);
1414 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1415 			ulong_t off;
1416 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1417 			printf("%s+%lx\n", sym ? sym : "?", off);
1418 		}
1419 	}
1420 	if (kmem_panic > 0)
1421 		panic("kernel heap corruption detected");
1422 	if (kmem_panic == 0)
1423 		debug_enter(NULL);
1424 	kmem_logging = 1;	/* resume logging */
1425 }
1426 
1427 static kmem_log_header_t *
1428 kmem_log_init(size_t logsize)
1429 {
1430 	kmem_log_header_t *lhp;
1431 	int nchunks = 4 * max_ncpus;
1432 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1433 	int i;
1434 
1435 	/*
1436 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1437 	 * to prevent false sharing of cache lines.
1438 	 */
1439 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1440 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1441 	    NULL, NULL, VM_SLEEP);
1442 	bzero(lhp, lhsize);
1443 
1444 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1445 	lhp->lh_nchunks = nchunks;
1446 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1447 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1448 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1449 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1450 	    nchunks * sizeof (int), VM_SLEEP);
1451 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1452 
1453 	for (i = 0; i < max_ncpus; i++) {
1454 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1455 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1456 		clhp->clh_chunk = i;
1457 	}
1458 
1459 	for (i = max_ncpus; i < nchunks; i++)
1460 		lhp->lh_free[i] = i;
1461 
1462 	lhp->lh_head = max_ncpus;
1463 	lhp->lh_tail = 0;
1464 
1465 	return (lhp);
1466 }
1467 
1468 static void *
1469 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1470 {
1471 	void *logspace;
1472 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1473 
1474 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1475 		return (NULL);
1476 
1477 	mutex_enter(&clhp->clh_lock);
1478 	clhp->clh_hits++;
1479 	if (size > clhp->clh_avail) {
1480 		mutex_enter(&lhp->lh_lock);
1481 		lhp->lh_hits++;
1482 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1483 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1484 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1485 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1486 		clhp->clh_current = lhp->lh_base +
1487 		    clhp->clh_chunk * lhp->lh_chunksize;
1488 		clhp->clh_avail = lhp->lh_chunksize;
1489 		if (size > lhp->lh_chunksize)
1490 			size = lhp->lh_chunksize;
1491 		mutex_exit(&lhp->lh_lock);
1492 	}
1493 	logspace = clhp->clh_current;
1494 	clhp->clh_current += size;
1495 	clhp->clh_avail -= size;
1496 	bcopy(data, logspace, size);
1497 	mutex_exit(&clhp->clh_lock);
1498 	return (logspace);
1499 }
1500 
1501 #define	KMEM_AUDIT(lp, cp, bcp)						\
1502 {									\
1503 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1504 	_bcp->bc_timestamp = gethrtime();				\
1505 	_bcp->bc_thread = curthread;					\
1506 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1507 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1508 }
1509 
1510 static void
1511 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1512     kmem_slab_t *sp, void *addr)
1513 {
1514 	kmem_bufctl_audit_t bca;
1515 
1516 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1517 	bca.bc_addr = addr;
1518 	bca.bc_slab = sp;
1519 	bca.bc_cache = cp;
1520 	KMEM_AUDIT(lp, cp, &bca);
1521 }
1522 
1523 /*
1524  * Create a new slab for cache cp.
1525  */
1526 static kmem_slab_t *
1527 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1528 {
1529 	size_t slabsize = cp->cache_slabsize;
1530 	size_t chunksize = cp->cache_chunksize;
1531 	int cache_flags = cp->cache_flags;
1532 	size_t color, chunks;
1533 	char *buf, *slab;
1534 	kmem_slab_t *sp;
1535 	kmem_bufctl_t *bcp;
1536 	vmem_t *vmp = cp->cache_arena;
1537 
1538 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1539 
1540 	color = cp->cache_color + cp->cache_align;
1541 	if (color > cp->cache_maxcolor)
1542 		color = cp->cache_mincolor;
1543 	cp->cache_color = color;
1544 
1545 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1546 
1547 	if (slab == NULL)
1548 		goto vmem_alloc_failure;
1549 
1550 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1551 
1552 	/*
1553 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1554 	 * consolidator depends (for correctness) on slabs being initialized
1555 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1556 	 * clients to distinguish uninitialized memory from known objects).
1557 	 */
1558 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1559 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1560 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1561 
1562 	if (cache_flags & KMF_HASH) {
1563 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1564 			goto slab_alloc_failure;
1565 		chunks = (slabsize - color) / chunksize;
1566 	} else {
1567 		sp = KMEM_SLAB(cp, slab);
1568 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1569 	}
1570 
1571 	sp->slab_cache	= cp;
1572 	sp->slab_head	= NULL;
1573 	sp->slab_refcnt	= 0;
1574 	sp->slab_base	= buf = slab + color;
1575 	sp->slab_chunks	= chunks;
1576 	sp->slab_stuck_offset = (uint32_t)-1;
1577 	sp->slab_later_count = 0;
1578 	sp->slab_flags = 0;
1579 
1580 	ASSERT(chunks > 0);
1581 	while (chunks-- != 0) {
1582 		if (cache_flags & KMF_HASH) {
1583 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1584 			if (bcp == NULL)
1585 				goto bufctl_alloc_failure;
1586 			if (cache_flags & KMF_AUDIT) {
1587 				kmem_bufctl_audit_t *bcap =
1588 				    (kmem_bufctl_audit_t *)bcp;
1589 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1590 				bcap->bc_cache = cp;
1591 			}
1592 			bcp->bc_addr = buf;
1593 			bcp->bc_slab = sp;
1594 		} else {
1595 			bcp = KMEM_BUFCTL(cp, buf);
1596 		}
1597 		if (cache_flags & KMF_BUFTAG) {
1598 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1599 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1600 			btp->bt_bufctl = bcp;
1601 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1602 			if (cache_flags & KMF_DEADBEEF) {
1603 				copy_pattern(KMEM_FREE_PATTERN, buf,
1604 				    cp->cache_verify);
1605 			}
1606 		}
1607 		bcp->bc_next = sp->slab_head;
1608 		sp->slab_head = bcp;
1609 		buf += chunksize;
1610 	}
1611 
1612 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1613 
1614 	return (sp);
1615 
1616 bufctl_alloc_failure:
1617 
1618 	while ((bcp = sp->slab_head) != NULL) {
1619 		sp->slab_head = bcp->bc_next;
1620 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1621 	}
1622 	kmem_cache_free(kmem_slab_cache, sp);
1623 
1624 slab_alloc_failure:
1625 
1626 	vmem_free(vmp, slab, slabsize);
1627 
1628 vmem_alloc_failure:
1629 
1630 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1631 	atomic_inc_64(&cp->cache_alloc_fail);
1632 
1633 	return (NULL);
1634 }
1635 
1636 /*
1637  * Destroy a slab.
1638  */
1639 static void
1640 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1641 {
1642 	vmem_t *vmp = cp->cache_arena;
1643 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1644 
1645 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1646 	ASSERT(sp->slab_refcnt == 0);
1647 
1648 	if (cp->cache_flags & KMF_HASH) {
1649 		kmem_bufctl_t *bcp;
1650 		while ((bcp = sp->slab_head) != NULL) {
1651 			sp->slab_head = bcp->bc_next;
1652 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1653 		}
1654 		kmem_cache_free(kmem_slab_cache, sp);
1655 	}
1656 	vmem_free(vmp, slab, cp->cache_slabsize);
1657 }
1658 
1659 static void *
1660 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1661 {
1662 	kmem_bufctl_t *bcp, **hash_bucket;
1663 	void *buf;
1664 	boolean_t new_slab = (sp->slab_refcnt == 0);
1665 
1666 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1667 	/*
1668 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1669 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1670 	 * slab is newly created.
1671 	 */
1672 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1673 	    (sp == avl_first(&cp->cache_partial_slabs))));
1674 	ASSERT(sp->slab_cache == cp);
1675 
1676 	cp->cache_slab_alloc++;
1677 	cp->cache_bufslab--;
1678 	sp->slab_refcnt++;
1679 
1680 	bcp = sp->slab_head;
1681 	sp->slab_head = bcp->bc_next;
1682 
1683 	if (cp->cache_flags & KMF_HASH) {
1684 		/*
1685 		 * Add buffer to allocated-address hash table.
1686 		 */
1687 		buf = bcp->bc_addr;
1688 		hash_bucket = KMEM_HASH(cp, buf);
1689 		bcp->bc_next = *hash_bucket;
1690 		*hash_bucket = bcp;
1691 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1692 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1693 		}
1694 	} else {
1695 		buf = KMEM_BUF(cp, bcp);
1696 	}
1697 
1698 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1699 
1700 	if (sp->slab_head == NULL) {
1701 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1702 		if (new_slab) {
1703 			ASSERT(sp->slab_chunks == 1);
1704 		} else {
1705 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1706 			avl_remove(&cp->cache_partial_slabs, sp);
1707 			sp->slab_later_count = 0; /* clear history */
1708 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1709 			sp->slab_stuck_offset = (uint32_t)-1;
1710 		}
1711 		list_insert_head(&cp->cache_complete_slabs, sp);
1712 		cp->cache_complete_slab_count++;
1713 		return (buf);
1714 	}
1715 
1716 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1717 	/*
1718 	 * Peek to see if the magazine layer is enabled before
1719 	 * we prefill.  We're not holding the cpu cache lock,
1720 	 * so the peek could be wrong, but there's no harm in it.
1721 	 */
1722 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1723 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1724 		kmem_slab_prefill(cp, sp);
1725 		return (buf);
1726 	}
1727 
1728 	if (new_slab) {
1729 		avl_add(&cp->cache_partial_slabs, sp);
1730 		return (buf);
1731 	}
1732 
1733 	/*
1734 	 * The slab is now more allocated than it was, so the
1735 	 * order remains unchanged.
1736 	 */
1737 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1738 	return (buf);
1739 }
1740 
1741 /*
1742  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1743  */
1744 static void *
1745 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1746 {
1747 	kmem_slab_t *sp;
1748 	void *buf;
1749 	boolean_t test_destructor;
1750 
1751 	mutex_enter(&cp->cache_lock);
1752 	test_destructor = (cp->cache_slab_alloc == 0);
1753 	sp = avl_first(&cp->cache_partial_slabs);
1754 	if (sp == NULL) {
1755 		ASSERT(cp->cache_bufslab == 0);
1756 
1757 		/*
1758 		 * The freelist is empty.  Create a new slab.
1759 		 */
1760 		mutex_exit(&cp->cache_lock);
1761 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1762 			return (NULL);
1763 		}
1764 		mutex_enter(&cp->cache_lock);
1765 		cp->cache_slab_create++;
1766 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1767 			cp->cache_bufmax = cp->cache_buftotal;
1768 		cp->cache_bufslab += sp->slab_chunks;
1769 	}
1770 
1771 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1772 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1773 	    (cp->cache_complete_slab_count +
1774 	    avl_numnodes(&cp->cache_partial_slabs) +
1775 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1776 	mutex_exit(&cp->cache_lock);
1777 
1778 	if (test_destructor && cp->cache_destructor != NULL) {
1779 		/*
1780 		 * On the first kmem_slab_alloc(), assert that it is valid to
1781 		 * call the destructor on a newly constructed object without any
1782 		 * client involvement.
1783 		 */
1784 		if ((cp->cache_constructor == NULL) ||
1785 		    cp->cache_constructor(buf, cp->cache_private,
1786 		    kmflag) == 0) {
1787 			cp->cache_destructor(buf, cp->cache_private);
1788 		}
1789 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1790 		    cp->cache_bufsize);
1791 		if (cp->cache_flags & KMF_DEADBEEF) {
1792 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1793 		}
1794 	}
1795 
1796 	return (buf);
1797 }
1798 
1799 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1800 
1801 /*
1802  * Free a raw (unconstructed) buffer to cp's slab layer.
1803  */
1804 static void
1805 kmem_slab_free(kmem_cache_t *cp, void *buf)
1806 {
1807 	kmem_slab_t *sp;
1808 	kmem_bufctl_t *bcp, **prev_bcpp;
1809 
1810 	ASSERT(buf != NULL);
1811 
1812 	mutex_enter(&cp->cache_lock);
1813 	cp->cache_slab_free++;
1814 
1815 	if (cp->cache_flags & KMF_HASH) {
1816 		/*
1817 		 * Look up buffer in allocated-address hash table.
1818 		 */
1819 		prev_bcpp = KMEM_HASH(cp, buf);
1820 		while ((bcp = *prev_bcpp) != NULL) {
1821 			if (bcp->bc_addr == buf) {
1822 				*prev_bcpp = bcp->bc_next;
1823 				sp = bcp->bc_slab;
1824 				break;
1825 			}
1826 			cp->cache_lookup_depth++;
1827 			prev_bcpp = &bcp->bc_next;
1828 		}
1829 	} else {
1830 		bcp = KMEM_BUFCTL(cp, buf);
1831 		sp = KMEM_SLAB(cp, buf);
1832 	}
1833 
1834 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1835 		mutex_exit(&cp->cache_lock);
1836 		kmem_error(KMERR_BADADDR, cp, buf);
1837 		return;
1838 	}
1839 
1840 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1841 		/*
1842 		 * If this is the buffer that prevented the consolidator from
1843 		 * clearing the slab, we can reset the slab flags now that the
1844 		 * buffer is freed. (It makes sense to do this in
1845 		 * kmem_cache_free(), where the client gives up ownership of the
1846 		 * buffer, but on the hot path the test is too expensive.)
1847 		 */
1848 		kmem_slab_move_yes(cp, sp, buf);
1849 	}
1850 
1851 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1852 		if (cp->cache_flags & KMF_CONTENTS)
1853 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1854 			    kmem_log_enter(kmem_content_log, buf,
1855 			    cp->cache_contents);
1856 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1857 	}
1858 
1859 	bcp->bc_next = sp->slab_head;
1860 	sp->slab_head = bcp;
1861 
1862 	cp->cache_bufslab++;
1863 	ASSERT(sp->slab_refcnt >= 1);
1864 
1865 	if (--sp->slab_refcnt == 0) {
1866 		/*
1867 		 * There are no outstanding allocations from this slab,
1868 		 * so we can reclaim the memory.
1869 		 */
1870 		if (sp->slab_chunks == 1) {
1871 			list_remove(&cp->cache_complete_slabs, sp);
1872 			cp->cache_complete_slab_count--;
1873 		} else {
1874 			avl_remove(&cp->cache_partial_slabs, sp);
1875 		}
1876 
1877 		cp->cache_buftotal -= sp->slab_chunks;
1878 		cp->cache_bufslab -= sp->slab_chunks;
1879 		/*
1880 		 * Defer releasing the slab to the virtual memory subsystem
1881 		 * while there is a pending move callback, since we guarantee
1882 		 * that buffers passed to the move callback have only been
1883 		 * touched by kmem or by the client itself. Since the memory
1884 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1885 		 * set at least one of the two lowest order bits, the client can
1886 		 * test those bits in the move callback to determine whether or
1887 		 * not it knows about the buffer (assuming that the client also
1888 		 * sets one of those low order bits whenever it frees a buffer).
1889 		 */
1890 		if (cp->cache_defrag == NULL ||
1891 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1892 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1893 			cp->cache_slab_destroy++;
1894 			mutex_exit(&cp->cache_lock);
1895 			kmem_slab_destroy(cp, sp);
1896 		} else {
1897 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1898 			/*
1899 			 * Slabs are inserted at both ends of the deadlist to
1900 			 * distinguish between slabs freed while move callbacks
1901 			 * are pending (list head) and a slab freed while the
1902 			 * lock is dropped in kmem_move_buffers() (list tail) so
1903 			 * that in both cases slab_destroy() is called from the
1904 			 * right context.
1905 			 */
1906 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1907 				list_insert_tail(deadlist, sp);
1908 			} else {
1909 				list_insert_head(deadlist, sp);
1910 			}
1911 			cp->cache_defrag->kmd_deadcount++;
1912 			mutex_exit(&cp->cache_lock);
1913 		}
1914 		return;
1915 	}
1916 
1917 	if (bcp->bc_next == NULL) {
1918 		/* Transition the slab from completely allocated to partial. */
1919 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1920 		ASSERT(sp->slab_chunks > 1);
1921 		list_remove(&cp->cache_complete_slabs, sp);
1922 		cp->cache_complete_slab_count--;
1923 		avl_add(&cp->cache_partial_slabs, sp);
1924 	} else {
1925 #ifdef	DEBUG
1926 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1927 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1928 		} else {
1929 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1930 		}
1931 #else
1932 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1933 #endif
1934 	}
1935 
1936 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1937 	    (cp->cache_complete_slab_count +
1938 	    avl_numnodes(&cp->cache_partial_slabs) +
1939 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1940 	mutex_exit(&cp->cache_lock);
1941 }
1942 
1943 /*
1944  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1945  */
1946 static int
1947 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1948     caddr_t caller)
1949 {
1950 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1951 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1952 	uint32_t mtbf;
1953 
1954 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1955 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1956 		return (-1);
1957 	}
1958 
1959 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1960 
1961 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1962 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1963 		return (-1);
1964 	}
1965 
1966 	if (cp->cache_flags & KMF_DEADBEEF) {
1967 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1968 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1969 				kmem_error(KMERR_MODIFIED, cp, buf);
1970 				return (-1);
1971 			}
1972 			if (cp->cache_constructor != NULL)
1973 				*(uint64_t *)buf = btp->bt_redzone;
1974 			else
1975 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1976 		} else {
1977 			construct = 1;
1978 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1979 			    KMEM_UNINITIALIZED_PATTERN, buf,
1980 			    cp->cache_verify)) {
1981 				kmem_error(KMERR_MODIFIED, cp, buf);
1982 				return (-1);
1983 			}
1984 		}
1985 	}
1986 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1987 
1988 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1989 	    gethrtime() % mtbf == 0 &&
1990 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1991 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1992 		if (!construct && cp->cache_destructor != NULL)
1993 			cp->cache_destructor(buf, cp->cache_private);
1994 	} else {
1995 		mtbf = 0;
1996 	}
1997 
1998 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1999 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
2000 		atomic_inc_64(&cp->cache_alloc_fail);
2001 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2002 		if (cp->cache_flags & KMF_DEADBEEF)
2003 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2004 		kmem_slab_free(cp, buf);
2005 		return (1);
2006 	}
2007 
2008 	if (cp->cache_flags & KMF_AUDIT) {
2009 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2010 	}
2011 
2012 	if ((cp->cache_flags & KMF_LITE) &&
2013 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2014 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2015 	}
2016 
2017 	return (0);
2018 }
2019 
2020 static int
2021 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2022 {
2023 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2024 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2025 	kmem_slab_t *sp;
2026 
2027 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2028 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2029 			kmem_error(KMERR_DUPFREE, cp, buf);
2030 			return (-1);
2031 		}
2032 		sp = kmem_findslab(cp, buf);
2033 		if (sp == NULL || sp->slab_cache != cp)
2034 			kmem_error(KMERR_BADADDR, cp, buf);
2035 		else
2036 			kmem_error(KMERR_REDZONE, cp, buf);
2037 		return (-1);
2038 	}
2039 
2040 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2041 
2042 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2043 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2044 		return (-1);
2045 	}
2046 
2047 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2048 		kmem_error(KMERR_REDZONE, cp, buf);
2049 		return (-1);
2050 	}
2051 
2052 	if (cp->cache_flags & KMF_AUDIT) {
2053 		if (cp->cache_flags & KMF_CONTENTS)
2054 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2055 			    buf, cp->cache_contents);
2056 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2057 	}
2058 
2059 	if ((cp->cache_flags & KMF_LITE) &&
2060 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2061 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2062 	}
2063 
2064 	if (cp->cache_flags & KMF_DEADBEEF) {
2065 		if (cp->cache_flags & KMF_LITE)
2066 			btp->bt_redzone = *(uint64_t *)buf;
2067 		else if (cp->cache_destructor != NULL)
2068 			cp->cache_destructor(buf, cp->cache_private);
2069 
2070 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2071 	}
2072 
2073 	return (0);
2074 }
2075 
2076 /*
2077  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2078  */
2079 static void
2080 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2081 {
2082 	int round;
2083 
2084 	ASSERT(!list_link_active(&cp->cache_link) ||
2085 	    taskq_member(kmem_taskq, curthread));
2086 
2087 	for (round = 0; round < nrounds; round++) {
2088 		void *buf = mp->mag_round[round];
2089 
2090 		if (cp->cache_flags & KMF_DEADBEEF) {
2091 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2092 			    cp->cache_verify) != NULL) {
2093 				kmem_error(KMERR_MODIFIED, cp, buf);
2094 				continue;
2095 			}
2096 			if ((cp->cache_flags & KMF_LITE) &&
2097 			    cp->cache_destructor != NULL) {
2098 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2099 				*(uint64_t *)buf = btp->bt_redzone;
2100 				cp->cache_destructor(buf, cp->cache_private);
2101 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2102 			}
2103 		} else if (cp->cache_destructor != NULL) {
2104 			cp->cache_destructor(buf, cp->cache_private);
2105 		}
2106 
2107 		kmem_slab_free(cp, buf);
2108 	}
2109 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2110 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2111 }
2112 
2113 /*
2114  * Allocate a magazine from the depot.
2115  */
2116 static kmem_magazine_t *
2117 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2118 {
2119 	kmem_magazine_t *mp;
2120 
2121 	/*
2122 	 * If we can't get the depot lock without contention,
2123 	 * update our contention count.  We use the depot
2124 	 * contention rate to determine whether we need to
2125 	 * increase the magazine size for better scalability.
2126 	 */
2127 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2128 		mutex_enter(&cp->cache_depot_lock);
2129 		cp->cache_depot_contention++;
2130 	}
2131 
2132 	if ((mp = mlp->ml_list) != NULL) {
2133 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2134 		mlp->ml_list = mp->mag_next;
2135 		if (--mlp->ml_total < mlp->ml_min)
2136 			mlp->ml_min = mlp->ml_total;
2137 		mlp->ml_alloc++;
2138 	}
2139 
2140 	mutex_exit(&cp->cache_depot_lock);
2141 
2142 	return (mp);
2143 }
2144 
2145 /*
2146  * Free a magazine to the depot.
2147  */
2148 static void
2149 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2150 {
2151 	mutex_enter(&cp->cache_depot_lock);
2152 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2153 	mp->mag_next = mlp->ml_list;
2154 	mlp->ml_list = mp;
2155 	mlp->ml_total++;
2156 	mutex_exit(&cp->cache_depot_lock);
2157 }
2158 
2159 /*
2160  * Update the working set statistics for cp's depot.
2161  */
2162 static void
2163 kmem_depot_ws_update(kmem_cache_t *cp)
2164 {
2165 	mutex_enter(&cp->cache_depot_lock);
2166 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2167 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2168 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2169 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2170 	mutex_exit(&cp->cache_depot_lock);
2171 }
2172 
2173 /*
2174  * Set the working set statistics for cp's depot to zero.  (Everything is
2175  * eligible for reaping.)
2176  */
2177 static void
2178 kmem_depot_ws_zero(kmem_cache_t *cp)
2179 {
2180 	mutex_enter(&cp->cache_depot_lock);
2181 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2182 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2183 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2184 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2185 	mutex_exit(&cp->cache_depot_lock);
2186 }
2187 
2188 /*
2189  * The number of bytes to reap before we call kpreempt(). The default (1MB)
2190  * causes us to preempt reaping up to hundreds of times per second. Using a
2191  * larger value (1GB) causes this to have virtually no effect.
2192  */
2193 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2194 
2195 /*
2196  * Reap all magazines that have fallen out of the depot's working set.
2197  */
2198 static void
2199 kmem_depot_ws_reap(kmem_cache_t *cp)
2200 {
2201 	size_t bytes = 0;
2202 	long reap;
2203 	kmem_magazine_t *mp;
2204 
2205 	ASSERT(!list_link_active(&cp->cache_link) ||
2206 	    taskq_member(kmem_taskq, curthread));
2207 
2208 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2209 	while (reap-- &&
2210 	    (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2211 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2212 		bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2213 		if (bytes > kmem_reap_preempt_bytes) {
2214 			kpreempt(KPREEMPT_SYNC);
2215 			bytes = 0;
2216 		}
2217 	}
2218 
2219 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2220 	while (reap-- &&
2221 	    (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2222 		kmem_magazine_destroy(cp, mp, 0);
2223 		bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2224 		if (bytes > kmem_reap_preempt_bytes) {
2225 			kpreempt(KPREEMPT_SYNC);
2226 			bytes = 0;
2227 		}
2228 	}
2229 }
2230 
2231 static void
2232 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2233 {
2234 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2235 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2236 	ASSERT(ccp->cc_magsize > 0);
2237 
2238 	ccp->cc_ploaded = ccp->cc_loaded;
2239 	ccp->cc_prounds = ccp->cc_rounds;
2240 	ccp->cc_loaded = mp;
2241 	ccp->cc_rounds = rounds;
2242 }
2243 
2244 /*
2245  * Intercept kmem alloc/free calls during crash dump in order to avoid
2246  * changing kmem state while memory is being saved to the dump device.
2247  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2248  * there are no locks because only one CPU calls kmem during a crash
2249  * dump. To enable this feature, first create the associated vmem
2250  * arena with VMC_DUMPSAFE.
2251  */
2252 static void *kmem_dump_start;	/* start of pre-reserved heap */
2253 static void *kmem_dump_end;	/* end of heap area */
2254 static void *kmem_dump_curr;	/* current free heap pointer */
2255 static size_t kmem_dump_size;	/* size of heap area */
2256 
2257 /* append to each buf created in the pre-reserved heap */
2258 typedef struct kmem_dumpctl {
2259 	void	*kdc_next;	/* cache dump free list linkage */
2260 } kmem_dumpctl_t;
2261 
2262 #define	KMEM_DUMPCTL(cp, buf)	\
2263 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2264 	    sizeof (void *)))
2265 
2266 /* Keep some simple stats. */
2267 #define	KMEM_DUMP_LOGS	(100)
2268 
2269 typedef struct kmem_dump_log {
2270 	kmem_cache_t	*kdl_cache;
2271 	uint_t		kdl_allocs;		/* # of dump allocations */
2272 	uint_t		kdl_frees;		/* # of dump frees */
2273 	uint_t		kdl_alloc_fails;	/* # of allocation failures */
2274 	uint_t		kdl_free_nondump;	/* # of non-dump frees */
2275 	uint_t		kdl_unsafe;		/* cache was used, but unsafe */
2276 } kmem_dump_log_t;
2277 
2278 static kmem_dump_log_t *kmem_dump_log;
2279 static int kmem_dump_log_idx;
2280 
2281 #define	KDI_LOG(cp, stat) {						\
2282 	kmem_dump_log_t *kdl;						\
2283 	if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) {	\
2284 		kdl->stat++;						\
2285 	} else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) {		\
2286 		kdl = &kmem_dump_log[kmem_dump_log_idx++];		\
2287 		kdl->stat++;						\
2288 		kdl->kdl_cache = (cp);					\
2289 		(cp)->cache_dumplog = kdl;				\
2290 	}								\
2291 }
2292 
2293 /* set non zero for full report */
2294 uint_t kmem_dump_verbose = 0;
2295 
2296 /* stats for overize heap */
2297 uint_t kmem_dump_oversize_allocs = 0;
2298 uint_t kmem_dump_oversize_max = 0;
2299 
2300 static void
2301 kmem_dumppr(char **pp, char *e, const char *format, ...)
2302 {
2303 	char *p = *pp;
2304 
2305 	if (p < e) {
2306 		int n;
2307 		va_list ap;
2308 
2309 		va_start(ap, format);
2310 		n = vsnprintf(p, e - p, format, ap);
2311 		va_end(ap);
2312 		*pp = p + n;
2313 	}
2314 }
2315 
2316 /*
2317  * Called when dumpadm(1M) configures dump parameters.
2318  */
2319 void
2320 kmem_dump_init(size_t size)
2321 {
2322 	if (kmem_dump_start != NULL)
2323 		kmem_free(kmem_dump_start, kmem_dump_size);
2324 
2325 	if (kmem_dump_log == NULL)
2326 		kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2327 		    sizeof (kmem_dump_log_t), KM_SLEEP);
2328 
2329 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2330 
2331 	if (kmem_dump_start != NULL) {
2332 		kmem_dump_size = size;
2333 		kmem_dump_curr = kmem_dump_start;
2334 		kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2335 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2336 	} else {
2337 		kmem_dump_size = 0;
2338 		kmem_dump_curr = NULL;
2339 		kmem_dump_end = NULL;
2340 	}
2341 }
2342 
2343 /*
2344  * Set flag for each kmem_cache_t if is safe to use alternate dump
2345  * memory. Called just before panic crash dump starts. Set the flag
2346  * for the calling CPU.
2347  */
2348 void
2349 kmem_dump_begin(void)
2350 {
2351 	ASSERT(panicstr != NULL);
2352 	if (kmem_dump_start != NULL) {
2353 		kmem_cache_t *cp;
2354 
2355 		for (cp = list_head(&kmem_caches); cp != NULL;
2356 		    cp = list_next(&kmem_caches, cp)) {
2357 			kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2358 
2359 			if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2360 				cp->cache_flags |= KMF_DUMPDIVERT;
2361 				ccp->cc_flags |= KMF_DUMPDIVERT;
2362 				ccp->cc_dump_rounds = ccp->cc_rounds;
2363 				ccp->cc_dump_prounds = ccp->cc_prounds;
2364 				ccp->cc_rounds = ccp->cc_prounds = -1;
2365 			} else {
2366 				cp->cache_flags |= KMF_DUMPUNSAFE;
2367 				ccp->cc_flags |= KMF_DUMPUNSAFE;
2368 			}
2369 		}
2370 	}
2371 }
2372 
2373 /*
2374  * finished dump intercept
2375  * print any warnings on the console
2376  * return verbose information to dumpsys() in the given buffer
2377  */
2378 size_t
2379 kmem_dump_finish(char *buf, size_t size)
2380 {
2381 	int kdi_idx;
2382 	int kdi_end = kmem_dump_log_idx;
2383 	int percent = 0;
2384 	int header = 0;
2385 	int warn = 0;
2386 	size_t used;
2387 	kmem_cache_t *cp;
2388 	kmem_dump_log_t *kdl;
2389 	char *e = buf + size;
2390 	char *p = buf;
2391 
2392 	if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2393 		return (0);
2394 
2395 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2396 	percent = (used * 100) / kmem_dump_size;
2397 
2398 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2399 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2400 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2401 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2402 	    kmem_dump_oversize_allocs);
2403 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2404 	    kmem_dump_oversize_max);
2405 
2406 	for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2407 		kdl = &kmem_dump_log[kdi_idx];
2408 		cp = kdl->kdl_cache;
2409 		if (cp == NULL)
2410 			break;
2411 		if (kdl->kdl_alloc_fails)
2412 			++warn;
2413 		if (header == 0) {
2414 			kmem_dumppr(&p, e,
2415 			    "Cache Name,Allocs,Frees,Alloc Fails,"
2416 			    "Nondump Frees,Unsafe Allocs/Frees\n");
2417 			header = 1;
2418 		}
2419 		kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2420 		    cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2421 		    kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2422 		    kdl->kdl_unsafe);
2423 	}
2424 
2425 	/* return buffer size used */
2426 	if (p < e)
2427 		bzero(p, e - p);
2428 	return (p - buf);
2429 }
2430 
2431 /*
2432  * Allocate a constructed object from alternate dump memory.
2433  */
2434 void *
2435 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2436 {
2437 	void *buf;
2438 	void *curr;
2439 	char *bufend;
2440 
2441 	/* return a constructed object */
2442 	if ((buf = cp->cache_dumpfreelist) != NULL) {
2443 		cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2444 		KDI_LOG(cp, kdl_allocs);
2445 		return (buf);
2446 	}
2447 
2448 	/* create a new constructed object */
2449 	curr = kmem_dump_curr;
2450 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2451 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2452 
2453 	/* hat layer objects cannot cross a page boundary */
2454 	if (cp->cache_align < PAGESIZE) {
2455 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2456 		if (bufend > page) {
2457 			bufend += page - (char *)buf;
2458 			buf = (void *)page;
2459 		}
2460 	}
2461 
2462 	/* fall back to normal alloc if reserved area is used up */
2463 	if (bufend > (char *)kmem_dump_end) {
2464 		kmem_dump_curr = kmem_dump_end;
2465 		KDI_LOG(cp, kdl_alloc_fails);
2466 		return (NULL);
2467 	}
2468 
2469 	/*
2470 	 * Must advance curr pointer before calling a constructor that
2471 	 * may also allocate memory.
2472 	 */
2473 	kmem_dump_curr = bufend;
2474 
2475 	/* run constructor */
2476 	if (cp->cache_constructor != NULL &&
2477 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2478 	    != 0) {
2479 #ifdef DEBUG
2480 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2481 		    cp->cache_name, (void *)cp);
2482 #endif
2483 		/* reset curr pointer iff no allocs were done */
2484 		if (kmem_dump_curr == bufend)
2485 			kmem_dump_curr = curr;
2486 
2487 		/* fall back to normal alloc if the constructor fails */
2488 		KDI_LOG(cp, kdl_alloc_fails);
2489 		return (NULL);
2490 	}
2491 
2492 	KDI_LOG(cp, kdl_allocs);
2493 	return (buf);
2494 }
2495 
2496 /*
2497  * Free a constructed object in alternate dump memory.
2498  */
2499 int
2500 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2501 {
2502 	/* save constructed buffers for next time */
2503 	if ((char *)buf >= (char *)kmem_dump_start &&
2504 	    (char *)buf < (char *)kmem_dump_end) {
2505 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2506 		cp->cache_dumpfreelist = buf;
2507 		KDI_LOG(cp, kdl_frees);
2508 		return (0);
2509 	}
2510 
2511 	/* count all non-dump buf frees */
2512 	KDI_LOG(cp, kdl_free_nondump);
2513 
2514 	/* just drop buffers that were allocated before dump started */
2515 	if (kmem_dump_curr < kmem_dump_end)
2516 		return (0);
2517 
2518 	/* fall back to normal free if reserved area is used up */
2519 	return (1);
2520 }
2521 
2522 /*
2523  * Allocate a constructed object from cache cp.
2524  */
2525 void *
2526 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2527 {
2528 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2529 	kmem_magazine_t *fmp;
2530 	void *buf;
2531 
2532 	mutex_enter(&ccp->cc_lock);
2533 	for (;;) {
2534 		/*
2535 		 * If there's an object available in the current CPU's
2536 		 * loaded magazine, just take it and return.
2537 		 */
2538 		if (ccp->cc_rounds > 0) {
2539 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2540 			ccp->cc_alloc++;
2541 			mutex_exit(&ccp->cc_lock);
2542 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2543 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2544 					ASSERT(!(ccp->cc_flags &
2545 					    KMF_DUMPDIVERT));
2546 					KDI_LOG(cp, kdl_unsafe);
2547 				}
2548 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2549 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2550 				    caller()) != 0) {
2551 					if (kmflag & KM_NOSLEEP)
2552 						return (NULL);
2553 					mutex_enter(&ccp->cc_lock);
2554 					continue;
2555 				}
2556 			}
2557 			return (buf);
2558 		}
2559 
2560 		/*
2561 		 * The loaded magazine is empty.  If the previously loaded
2562 		 * magazine was full, exchange them and try again.
2563 		 */
2564 		if (ccp->cc_prounds > 0) {
2565 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2566 			continue;
2567 		}
2568 
2569 		/*
2570 		 * Return an alternate buffer at dump time to preserve
2571 		 * the heap.
2572 		 */
2573 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2574 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2575 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2576 				/* log it so that we can warn about it */
2577 				KDI_LOG(cp, kdl_unsafe);
2578 			} else {
2579 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2580 				    NULL) {
2581 					mutex_exit(&ccp->cc_lock);
2582 					return (buf);
2583 				}
2584 				break;		/* fall back to slab layer */
2585 			}
2586 		}
2587 
2588 		/*
2589 		 * If the magazine layer is disabled, break out now.
2590 		 */
2591 		if (ccp->cc_magsize == 0)
2592 			break;
2593 
2594 		/*
2595 		 * Try to get a full magazine from the depot.
2596 		 */
2597 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2598 		if (fmp != NULL) {
2599 			if (ccp->cc_ploaded != NULL)
2600 				kmem_depot_free(cp, &cp->cache_empty,
2601 				    ccp->cc_ploaded);
2602 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2603 			continue;
2604 		}
2605 
2606 		/*
2607 		 * There are no full magazines in the depot,
2608 		 * so fall through to the slab layer.
2609 		 */
2610 		break;
2611 	}
2612 	mutex_exit(&ccp->cc_lock);
2613 
2614 	/*
2615 	 * We couldn't allocate a constructed object from the magazine layer,
2616 	 * so get a raw buffer from the slab layer and apply its constructor.
2617 	 */
2618 	buf = kmem_slab_alloc(cp, kmflag);
2619 
2620 	if (buf == NULL)
2621 		return (NULL);
2622 
2623 	if (cp->cache_flags & KMF_BUFTAG) {
2624 		/*
2625 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2626 		 */
2627 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2628 		if (rc != 0) {
2629 			if (kmflag & KM_NOSLEEP)
2630 				return (NULL);
2631 			/*
2632 			 * kmem_cache_alloc_debug() detected corruption
2633 			 * but didn't panic (kmem_panic <= 0). We should not be
2634 			 * here because the constructor failed (indicated by a
2635 			 * return code of 1). Try again.
2636 			 */
2637 			ASSERT(rc == -1);
2638 			return (kmem_cache_alloc(cp, kmflag));
2639 		}
2640 		return (buf);
2641 	}
2642 
2643 	if (cp->cache_constructor != NULL &&
2644 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2645 		atomic_inc_64(&cp->cache_alloc_fail);
2646 		kmem_slab_free(cp, buf);
2647 		return (NULL);
2648 	}
2649 
2650 	return (buf);
2651 }
2652 
2653 /*
2654  * The freed argument tells whether or not kmem_cache_free_debug() has already
2655  * been called so that we can avoid the duplicate free error. For example, a
2656  * buffer on a magazine has already been freed by the client but is still
2657  * constructed.
2658  */
2659 static void
2660 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2661 {
2662 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2663 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2664 			return;
2665 
2666 	/*
2667 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2668 	 * kmem_cache_free_debug() will have already applied the destructor.
2669 	 */
2670 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2671 	    cp->cache_destructor != NULL) {
2672 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2673 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2674 			*(uint64_t *)buf = btp->bt_redzone;
2675 			cp->cache_destructor(buf, cp->cache_private);
2676 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2677 		} else {
2678 			cp->cache_destructor(buf, cp->cache_private);
2679 		}
2680 	}
2681 
2682 	kmem_slab_free(cp, buf);
2683 }
2684 
2685 /*
2686  * Used when there's no room to free a buffer to the per-CPU cache.
2687  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2688  * caller should try freeing to the per-CPU cache again.
2689  * Note that we don't directly install the magazine in the cpu cache,
2690  * since its state may have changed wildly while the lock was dropped.
2691  */
2692 static int
2693 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2694 {
2695 	kmem_magazine_t *emp;
2696 	kmem_magtype_t *mtp;
2697 
2698 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2699 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2700 	    ((uint_t)ccp->cc_rounds == -1)) &&
2701 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2702 	    ((uint_t)ccp->cc_prounds == -1)));
2703 
2704 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2705 	if (emp != NULL) {
2706 		if (ccp->cc_ploaded != NULL)
2707 			kmem_depot_free(cp, &cp->cache_full,
2708 			    ccp->cc_ploaded);
2709 		kmem_cpu_reload(ccp, emp, 0);
2710 		return (1);
2711 	}
2712 	/*
2713 	 * There are no empty magazines in the depot,
2714 	 * so try to allocate a new one.  We must drop all locks
2715 	 * across kmem_cache_alloc() because lower layers may
2716 	 * attempt to allocate from this cache.
2717 	 */
2718 	mtp = cp->cache_magtype;
2719 	mutex_exit(&ccp->cc_lock);
2720 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2721 	mutex_enter(&ccp->cc_lock);
2722 
2723 	if (emp != NULL) {
2724 		/*
2725 		 * We successfully allocated an empty magazine.
2726 		 * However, we had to drop ccp->cc_lock to do it,
2727 		 * so the cache's magazine size may have changed.
2728 		 * If so, free the magazine and try again.
2729 		 */
2730 		if (ccp->cc_magsize != mtp->mt_magsize) {
2731 			mutex_exit(&ccp->cc_lock);
2732 			kmem_cache_free(mtp->mt_cache, emp);
2733 			mutex_enter(&ccp->cc_lock);
2734 			return (1);
2735 		}
2736 
2737 		/*
2738 		 * We got a magazine of the right size.  Add it to
2739 		 * the depot and try the whole dance again.
2740 		 */
2741 		kmem_depot_free(cp, &cp->cache_empty, emp);
2742 		return (1);
2743 	}
2744 
2745 	/*
2746 	 * We couldn't allocate an empty magazine,
2747 	 * so fall through to the slab layer.
2748 	 */
2749 	return (0);
2750 }
2751 
2752 /*
2753  * Free a constructed object to cache cp.
2754  */
2755 void
2756 kmem_cache_free(kmem_cache_t *cp, void *buf)
2757 {
2758 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2759 
2760 	/*
2761 	 * The client must not free either of the buffers passed to the move
2762 	 * callback function.
2763 	 */
2764 	ASSERT(cp->cache_defrag == NULL ||
2765 	    cp->cache_defrag->kmd_thread != curthread ||
2766 	    (buf != cp->cache_defrag->kmd_from_buf &&
2767 	    buf != cp->cache_defrag->kmd_to_buf));
2768 
2769 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2770 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2771 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2772 			/* log it so that we can warn about it */
2773 			KDI_LOG(cp, kdl_unsafe);
2774 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2775 			return;
2776 		}
2777 		if (ccp->cc_flags & KMF_BUFTAG) {
2778 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2779 				return;
2780 		}
2781 	}
2782 
2783 	mutex_enter(&ccp->cc_lock);
2784 	/*
2785 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2786 	 */
2787 	for (;;) {
2788 		/*
2789 		 * If there's a slot available in the current CPU's
2790 		 * loaded magazine, just put the object there and return.
2791 		 */
2792 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2793 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2794 			ccp->cc_free++;
2795 			mutex_exit(&ccp->cc_lock);
2796 			return;
2797 		}
2798 
2799 		/*
2800 		 * The loaded magazine is full.  If the previously loaded
2801 		 * magazine was empty, exchange them and try again.
2802 		 */
2803 		if (ccp->cc_prounds == 0) {
2804 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2805 			continue;
2806 		}
2807 
2808 		/*
2809 		 * If the magazine layer is disabled, break out now.
2810 		 */
2811 		if (ccp->cc_magsize == 0)
2812 			break;
2813 
2814 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2815 			/*
2816 			 * We couldn't free our constructed object to the
2817 			 * magazine layer, so apply its destructor and free it
2818 			 * to the slab layer.
2819 			 */
2820 			break;
2821 		}
2822 	}
2823 	mutex_exit(&ccp->cc_lock);
2824 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2825 }
2826 
2827 static void
2828 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2829 {
2830 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2831 	int cache_flags = cp->cache_flags;
2832 
2833 	kmem_bufctl_t *next, *head;
2834 	size_t nbufs;
2835 
2836 	/*
2837 	 * Completely allocate the newly created slab and put the pre-allocated
2838 	 * buffers in magazines. Any of the buffers that cannot be put in
2839 	 * magazines must be returned to the slab.
2840 	 */
2841 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2842 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2843 	ASSERT(cp->cache_constructor == NULL);
2844 	ASSERT(sp->slab_cache == cp);
2845 	ASSERT(sp->slab_refcnt == 1);
2846 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2847 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2848 
2849 	head = sp->slab_head;
2850 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2851 	sp->slab_head = NULL;
2852 	sp->slab_refcnt += nbufs;
2853 	cp->cache_bufslab -= nbufs;
2854 	cp->cache_slab_alloc += nbufs;
2855 	list_insert_head(&cp->cache_complete_slabs, sp);
2856 	cp->cache_complete_slab_count++;
2857 	mutex_exit(&cp->cache_lock);
2858 	mutex_enter(&ccp->cc_lock);
2859 
2860 	while (head != NULL) {
2861 		void *buf = KMEM_BUF(cp, head);
2862 		/*
2863 		 * If there's a slot available in the current CPU's
2864 		 * loaded magazine, just put the object there and
2865 		 * continue.
2866 		 */
2867 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2868 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2869 			    buf;
2870 			ccp->cc_free++;
2871 			nbufs--;
2872 			head = head->bc_next;
2873 			continue;
2874 		}
2875 
2876 		/*
2877 		 * The loaded magazine is full.  If the previously
2878 		 * loaded magazine was empty, exchange them and try
2879 		 * again.
2880 		 */
2881 		if (ccp->cc_prounds == 0) {
2882 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2883 			    ccp->cc_prounds);
2884 			continue;
2885 		}
2886 
2887 		/*
2888 		 * If the magazine layer is disabled, break out now.
2889 		 */
2890 
2891 		if (ccp->cc_magsize == 0) {
2892 			break;
2893 		}
2894 
2895 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2896 			break;
2897 	}
2898 	mutex_exit(&ccp->cc_lock);
2899 	if (nbufs != 0) {
2900 		ASSERT(head != NULL);
2901 
2902 		/*
2903 		 * If there was a failure, return remaining objects to
2904 		 * the slab
2905 		 */
2906 		while (head != NULL) {
2907 			ASSERT(nbufs != 0);
2908 			next = head->bc_next;
2909 			head->bc_next = NULL;
2910 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2911 			head = next;
2912 			nbufs--;
2913 		}
2914 	}
2915 	ASSERT(head == NULL);
2916 	ASSERT(nbufs == 0);
2917 	mutex_enter(&cp->cache_lock);
2918 }
2919 
2920 void *
2921 kmem_zalloc(size_t size, int kmflag)
2922 {
2923 	size_t index;
2924 	void *buf;
2925 
2926 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2927 		kmem_cache_t *cp = kmem_alloc_table[index];
2928 		buf = kmem_cache_alloc(cp, kmflag);
2929 		if (buf != NULL) {
2930 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2931 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2932 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2933 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2934 
2935 				if (cp->cache_flags & KMF_LITE) {
2936 					KMEM_BUFTAG_LITE_ENTER(btp,
2937 					    kmem_lite_count, caller());
2938 				}
2939 			}
2940 			bzero(buf, size);
2941 		}
2942 	} else {
2943 		buf = kmem_alloc(size, kmflag);
2944 		if (buf != NULL)
2945 			bzero(buf, size);
2946 	}
2947 	return (buf);
2948 }
2949 
2950 void *
2951 kmem_alloc(size_t size, int kmflag)
2952 {
2953 	size_t index;
2954 	kmem_cache_t *cp;
2955 	void *buf;
2956 
2957 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2958 		cp = kmem_alloc_table[index];
2959 		/* fall through to kmem_cache_alloc() */
2960 
2961 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2962 	    kmem_big_alloc_table_max) {
2963 		cp = kmem_big_alloc_table[index];
2964 		/* fall through to kmem_cache_alloc() */
2965 
2966 	} else {
2967 		if (size == 0)
2968 			return (NULL);
2969 
2970 		buf = vmem_alloc(kmem_oversize_arena, size,
2971 		    kmflag & KM_VMFLAGS);
2972 		if (buf == NULL)
2973 			kmem_log_event(kmem_failure_log, NULL, NULL,
2974 			    (void *)size);
2975 		else if (KMEM_DUMP(kmem_slab_cache)) {
2976 			/* stats for dump intercept */
2977 			kmem_dump_oversize_allocs++;
2978 			if (size > kmem_dump_oversize_max)
2979 				kmem_dump_oversize_max = size;
2980 		}
2981 		return (buf);
2982 	}
2983 
2984 	buf = kmem_cache_alloc(cp, kmflag);
2985 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2986 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2987 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2988 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2989 
2990 		if (cp->cache_flags & KMF_LITE) {
2991 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2992 		}
2993 	}
2994 	return (buf);
2995 }
2996 
2997 void
2998 kmem_free(void *buf, size_t size)
2999 {
3000 	size_t index;
3001 	kmem_cache_t *cp;
3002 
3003 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
3004 		cp = kmem_alloc_table[index];
3005 		/* fall through to kmem_cache_free() */
3006 
3007 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
3008 	    kmem_big_alloc_table_max) {
3009 		cp = kmem_big_alloc_table[index];
3010 		/* fall through to kmem_cache_free() */
3011 
3012 	} else {
3013 		EQUIV(buf == NULL, size == 0);
3014 		if (buf == NULL && size == 0)
3015 			return;
3016 		vmem_free(kmem_oversize_arena, buf, size);
3017 		return;
3018 	}
3019 
3020 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
3021 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
3022 		uint32_t *ip = (uint32_t *)btp;
3023 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3024 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3025 				kmem_error(KMERR_DUPFREE, cp, buf);
3026 				return;
3027 			}
3028 			if (KMEM_SIZE_VALID(ip[1])) {
3029 				ip[0] = KMEM_SIZE_ENCODE(size);
3030 				kmem_error(KMERR_BADSIZE, cp, buf);
3031 			} else {
3032 				kmem_error(KMERR_REDZONE, cp, buf);
3033 			}
3034 			return;
3035 		}
3036 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3037 			kmem_error(KMERR_REDZONE, cp, buf);
3038 			return;
3039 		}
3040 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
3041 		if (cp->cache_flags & KMF_LITE) {
3042 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3043 			    caller());
3044 		}
3045 	}
3046 	kmem_cache_free(cp, buf);
3047 }
3048 
3049 void *
3050 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3051 {
3052 	size_t realsize = size + vmp->vm_quantum;
3053 	void *addr;
3054 
3055 	/*
3056 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3057 	 * vm_quantum will cause integer wraparound.  Check for this, and
3058 	 * blow off the firewall page in this case.  Note that such a
3059 	 * giant allocation (the entire kernel address space) can never
3060 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3061 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
3062 	 * corresponding check in kmem_firewall_va_free().
3063 	 */
3064 	if (realsize < size)
3065 		realsize = size;
3066 
3067 	/*
3068 	 * While boot still owns resource management, make sure that this
3069 	 * redzone virtual address allocation is properly accounted for in
3070 	 * OBPs "virtual-memory" "available" lists because we're
3071 	 * effectively claiming them for a red zone.  If we don't do this,
3072 	 * the available lists become too fragmented and too large for the
3073 	 * current boot/kernel memory list interface.
3074 	 */
3075 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3076 
3077 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3078 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3079 
3080 	return (addr);
3081 }
3082 
3083 void
3084 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3085 {
3086 	ASSERT((kvseg.s_base == NULL ?
3087 	    va_to_pfn((char *)addr + size) :
3088 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3089 
3090 	vmem_free(vmp, addr, size + vmp->vm_quantum);
3091 }
3092 
3093 /*
3094  * Try to allocate at least `size' bytes of memory without sleeping or
3095  * panicking. Return actual allocated size in `asize'. If allocation failed,
3096  * try final allocation with sleep or panic allowed.
3097  */
3098 void *
3099 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3100 {
3101 	void *p;
3102 
3103 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3104 	do {
3105 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3106 		if (p != NULL)
3107 			return (p);
3108 		*asize += KMEM_ALIGN;
3109 	} while (*asize <= PAGESIZE);
3110 
3111 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3112 	return (kmem_alloc(*asize, kmflag));
3113 }
3114 
3115 /*
3116  * Reclaim all unused memory from a cache.
3117  */
3118 static void
3119 kmem_cache_reap(kmem_cache_t *cp)
3120 {
3121 	ASSERT(taskq_member(kmem_taskq, curthread));
3122 	cp->cache_reap++;
3123 
3124 	/*
3125 	 * Ask the cache's owner to free some memory if possible.
3126 	 * The idea is to handle things like the inode cache, which
3127 	 * typically sits on a bunch of memory that it doesn't truly
3128 	 * *need*.  Reclaim policy is entirely up to the owner; this
3129 	 * callback is just an advisory plea for help.
3130 	 */
3131 	if (cp->cache_reclaim != NULL) {
3132 		long delta;
3133 
3134 		/*
3135 		 * Reclaimed memory should be reapable (not included in the
3136 		 * depot's working set).
3137 		 */
3138 		delta = cp->cache_full.ml_total;
3139 		cp->cache_reclaim(cp->cache_private);
3140 		delta = cp->cache_full.ml_total - delta;
3141 		if (delta > 0) {
3142 			mutex_enter(&cp->cache_depot_lock);
3143 			cp->cache_full.ml_reaplimit += delta;
3144 			cp->cache_full.ml_min += delta;
3145 			mutex_exit(&cp->cache_depot_lock);
3146 		}
3147 	}
3148 
3149 	kmem_depot_ws_reap(cp);
3150 
3151 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3152 		kmem_cache_defrag(cp);
3153 	}
3154 }
3155 
3156 static void
3157 kmem_reap_timeout(void *flag_arg)
3158 {
3159 	uint32_t *flag = (uint32_t *)flag_arg;
3160 
3161 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3162 	*flag = 0;
3163 }
3164 
3165 static void
3166 kmem_reap_done(void *flag)
3167 {
3168 	if (!callout_init_done) {
3169 		/* can't schedule a timeout at this point */
3170 		kmem_reap_timeout(flag);
3171 	} else {
3172 		(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3173 	}
3174 }
3175 
3176 static void
3177 kmem_reap_start(void *flag)
3178 {
3179 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3180 
3181 	if (flag == &kmem_reaping) {
3182 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3183 		/*
3184 		 * if we have segkp under heap, reap segkp cache.
3185 		 */
3186 		if (segkp_fromheap)
3187 			segkp_cache_free();
3188 	}
3189 	else
3190 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3191 
3192 	/*
3193 	 * We use taskq_dispatch() to schedule a timeout to clear
3194 	 * the flag so that kmem_reap() becomes self-throttling:
3195 	 * we won't reap again until the current reap completes *and*
3196 	 * at least kmem_reap_interval ticks have elapsed.
3197 	 */
3198 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3199 		kmem_reap_done(flag);
3200 }
3201 
3202 static void
3203 kmem_reap_common(void *flag_arg)
3204 {
3205 	uint32_t *flag = (uint32_t *)flag_arg;
3206 
3207 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3208 	    atomic_cas_32(flag, 0, 1) != 0)
3209 		return;
3210 
3211 	/*
3212 	 * It may not be kosher to do memory allocation when a reap is called
3213 	 * (for example, if vmem_populate() is in the call chain).  So we
3214 	 * start the reap going with a TQ_NOALLOC dispatch.  If the dispatch
3215 	 * fails, we reset the flag, and the next reap will try again.
3216 	 */
3217 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3218 		*flag = 0;
3219 }
3220 
3221 /*
3222  * Reclaim all unused memory from all caches.  Called from the VM system
3223  * when memory gets tight.
3224  */
3225 void
3226 kmem_reap(void)
3227 {
3228 	kmem_reap_common(&kmem_reaping);
3229 }
3230 
3231 /*
3232  * Reclaim all unused memory from identifier arenas, called when a vmem
3233  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3234  * cannot help with identifier exhaustion, we avoid both a large amount of
3235  * work and unwanted side-effects from reclaim callbacks.
3236  */
3237 void
3238 kmem_reap_idspace(void)
3239 {
3240 	kmem_reap_common(&kmem_reaping_idspace);
3241 }
3242 
3243 /*
3244  * Purge all magazines from a cache and set its magazine limit to zero.
3245  * All calls are serialized by the kmem_taskq lock, except for the final
3246  * call from kmem_cache_destroy().
3247  */
3248 static void
3249 kmem_cache_magazine_purge(kmem_cache_t *cp)
3250 {
3251 	kmem_cpu_cache_t *ccp;
3252 	kmem_magazine_t *mp, *pmp;
3253 	int rounds, prounds, cpu_seqid;
3254 
3255 	ASSERT(!list_link_active(&cp->cache_link) ||
3256 	    taskq_member(kmem_taskq, curthread));
3257 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3258 
3259 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3260 		ccp = &cp->cache_cpu[cpu_seqid];
3261 
3262 		mutex_enter(&ccp->cc_lock);
3263 		mp = ccp->cc_loaded;
3264 		pmp = ccp->cc_ploaded;
3265 		rounds = ccp->cc_rounds;
3266 		prounds = ccp->cc_prounds;
3267 		ccp->cc_loaded = NULL;
3268 		ccp->cc_ploaded = NULL;
3269 		ccp->cc_rounds = -1;
3270 		ccp->cc_prounds = -1;
3271 		ccp->cc_magsize = 0;
3272 		mutex_exit(&ccp->cc_lock);
3273 
3274 		if (mp)
3275 			kmem_magazine_destroy(cp, mp, rounds);
3276 		if (pmp)
3277 			kmem_magazine_destroy(cp, pmp, prounds);
3278 	}
3279 
3280 	kmem_depot_ws_zero(cp);
3281 	kmem_depot_ws_reap(cp);
3282 }
3283 
3284 /*
3285  * Enable per-cpu magazines on a cache.
3286  */
3287 static void
3288 kmem_cache_magazine_enable(kmem_cache_t *cp)
3289 {
3290 	int cpu_seqid;
3291 
3292 	if (cp->cache_flags & KMF_NOMAGAZINE)
3293 		return;
3294 
3295 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3296 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3297 		mutex_enter(&ccp->cc_lock);
3298 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3299 		mutex_exit(&ccp->cc_lock);
3300 	}
3301 
3302 }
3303 
3304 /*
3305  * Reap (almost) everything right now.
3306  */
3307 void
3308 kmem_cache_reap_now(kmem_cache_t *cp)
3309 {
3310 	ASSERT(list_link_active(&cp->cache_link));
3311 
3312 	kmem_depot_ws_zero(cp);
3313 
3314 	(void) taskq_dispatch(kmem_taskq,
3315 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3316 	taskq_wait(kmem_taskq);
3317 }
3318 
3319 /*
3320  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3321  * provide a higher transfer rate with the depot, while smaller magazines
3322  * reduce memory consumption.  Magazine resizing is an expensive operation;
3323  * it should not be done frequently.
3324  *
3325  * Changes to the magazine size are serialized by the kmem_taskq lock.
3326  *
3327  * Note: at present this only grows the magazine size.  It might be useful
3328  * to allow shrinkage too.
3329  */
3330 static void
3331 kmem_cache_magazine_resize(kmem_cache_t *cp)
3332 {
3333 	kmem_magtype_t *mtp = cp->cache_magtype;
3334 
3335 	ASSERT(taskq_member(kmem_taskq, curthread));
3336 
3337 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3338 		kmem_cache_magazine_purge(cp);
3339 		mutex_enter(&cp->cache_depot_lock);
3340 		cp->cache_magtype = ++mtp;
3341 		cp->cache_depot_contention_prev =
3342 		    cp->cache_depot_contention + INT_MAX;
3343 		mutex_exit(&cp->cache_depot_lock);
3344 		kmem_cache_magazine_enable(cp);
3345 	}
3346 }
3347 
3348 /*
3349  * Rescale a cache's hash table, so that the table size is roughly the
3350  * cache size.  We want the average lookup time to be extremely small.
3351  */
3352 static void
3353 kmem_hash_rescale(kmem_cache_t *cp)
3354 {
3355 	kmem_bufctl_t **old_table, **new_table, *bcp;
3356 	size_t old_size, new_size, h;
3357 
3358 	ASSERT(taskq_member(kmem_taskq, curthread));
3359 
3360 	new_size = MAX(KMEM_HASH_INITIAL,
3361 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3362 	old_size = cp->cache_hash_mask + 1;
3363 
3364 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3365 		return;
3366 
3367 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3368 	    VM_NOSLEEP);
3369 	if (new_table == NULL)
3370 		return;
3371 	bzero(new_table, new_size * sizeof (void *));
3372 
3373 	mutex_enter(&cp->cache_lock);
3374 
3375 	old_size = cp->cache_hash_mask + 1;
3376 	old_table = cp->cache_hash_table;
3377 
3378 	cp->cache_hash_mask = new_size - 1;
3379 	cp->cache_hash_table = new_table;
3380 	cp->cache_rescale++;
3381 
3382 	for (h = 0; h < old_size; h++) {
3383 		bcp = old_table[h];
3384 		while (bcp != NULL) {
3385 			void *addr = bcp->bc_addr;
3386 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3387 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3388 			bcp->bc_next = *hash_bucket;
3389 			*hash_bucket = bcp;
3390 			bcp = next_bcp;
3391 		}
3392 	}
3393 
3394 	mutex_exit(&cp->cache_lock);
3395 
3396 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3397 }
3398 
3399 /*
3400  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3401  * update, magazine resizing, and slab consolidation.
3402  */
3403 static void
3404 kmem_cache_update(kmem_cache_t *cp)
3405 {
3406 	int need_hash_rescale = 0;
3407 	int need_magazine_resize = 0;
3408 
3409 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3410 
3411 	/*
3412 	 * If the cache has become much larger or smaller than its hash table,
3413 	 * fire off a request to rescale the hash table.
3414 	 */
3415 	mutex_enter(&cp->cache_lock);
3416 
3417 	if ((cp->cache_flags & KMF_HASH) &&
3418 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3419 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3420 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3421 		need_hash_rescale = 1;
3422 
3423 	mutex_exit(&cp->cache_lock);
3424 
3425 	/*
3426 	 * Update the depot working set statistics.
3427 	 */
3428 	kmem_depot_ws_update(cp);
3429 
3430 	/*
3431 	 * If there's a lot of contention in the depot,
3432 	 * increase the magazine size.
3433 	 */
3434 	mutex_enter(&cp->cache_depot_lock);
3435 
3436 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3437 	    (int)(cp->cache_depot_contention -
3438 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3439 		need_magazine_resize = 1;
3440 
3441 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3442 
3443 	mutex_exit(&cp->cache_depot_lock);
3444 
3445 	if (need_hash_rescale)
3446 		(void) taskq_dispatch(kmem_taskq,
3447 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3448 
3449 	if (need_magazine_resize)
3450 		(void) taskq_dispatch(kmem_taskq,
3451 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3452 
3453 	if (cp->cache_defrag != NULL)
3454 		(void) taskq_dispatch(kmem_taskq,
3455 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3456 }
3457 
3458 static void kmem_update(void *);
3459 
3460 static void
3461 kmem_update_timeout(void *dummy)
3462 {
3463 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3464 }
3465 
3466 static void
3467 kmem_update(void *dummy)
3468 {
3469 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3470 
3471 	/*
3472 	 * We use taskq_dispatch() to reschedule the timeout so that
3473 	 * kmem_update() becomes self-throttling: it won't schedule
3474 	 * new tasks until all previous tasks have completed.
3475 	 */
3476 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3477 		kmem_update_timeout(NULL);
3478 }
3479 
3480 static int
3481 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3482 {
3483 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3484 	kmem_cache_t *cp = ksp->ks_private;
3485 	uint64_t cpu_buf_avail;
3486 	uint64_t buf_avail = 0;
3487 	int cpu_seqid;
3488 	long reap;
3489 
3490 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3491 
3492 	if (rw == KSTAT_WRITE)
3493 		return (EACCES);
3494 
3495 	mutex_enter(&cp->cache_lock);
3496 
3497 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3498 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3499 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3500 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3501 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3502 
3503 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3504 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3505 
3506 		mutex_enter(&ccp->cc_lock);
3507 
3508 		cpu_buf_avail = 0;
3509 		if (ccp->cc_rounds > 0)
3510 			cpu_buf_avail += ccp->cc_rounds;
3511 		if (ccp->cc_prounds > 0)
3512 			cpu_buf_avail += ccp->cc_prounds;
3513 
3514 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3515 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3516 		buf_avail			+= cpu_buf_avail;
3517 
3518 		mutex_exit(&ccp->cc_lock);
3519 	}
3520 
3521 	mutex_enter(&cp->cache_depot_lock);
3522 
3523 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3524 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3525 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3526 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3527 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3528 	kmcp->kmc_magazine_size.value.ui64	=
3529 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3530 	    0 : cp->cache_magtype->mt_magsize;
3531 
3532 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3533 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3534 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3535 
3536 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3537 	reap = MIN(reap, cp->cache_full.ml_total);
3538 
3539 	mutex_exit(&cp->cache_depot_lock);
3540 
3541 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3542 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3543 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3544 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3545 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3546 	buf_avail += cp->cache_bufslab;
3547 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3548 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3549 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3550 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3551 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3552 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3553 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3554 	    cp->cache_hash_mask + 1 : 0;
3555 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3556 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3557 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3558 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3559 
3560 	if (cp->cache_defrag == NULL) {