xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision 0c833d64)
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 2015 Nexenta Systems, Inc.  All rights reserved.
24  */
25 
26 /*
27  * Kernel memory allocator, as described in the following two papers and a
28  * statement about the consolidator:
29  *
30  * Jeff Bonwick,
31  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
32  * Proceedings of the Summer 1994 Usenix Conference.
33  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
34  *
35  * Jeff Bonwick and Jonathan Adams,
36  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
37  * Arbitrary Resources.
38  * Proceedings of the 2001 Usenix Conference.
39  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
40  *
41  * kmem Slab Consolidator Big Theory Statement:
42  *
43  * 1. Motivation
44  *
45  * As stated in Bonwick94, slabs provide the following advantages over other
46  * allocation structures in terms of memory fragmentation:
47  *
48  *  - Internal fragmentation (per-buffer wasted space) is minimal.
49  *  - Severe external fragmentation (unused buffers on the free list) is
50  *    unlikely.
51  *
52  * Segregating objects by size eliminates one source of external fragmentation,
53  * and according to Bonwick:
54  *
55  *   The other reason that slabs reduce external fragmentation is that all
56  *   objects in a slab are of the same type, so they have the same lifetime
57  *   distribution. The resulting segregation of short-lived and long-lived
58  *   objects at slab granularity reduces the likelihood of an entire page being
59  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
60  *
61  * While unlikely, severe external fragmentation remains possible. Clients that
62  * allocate both short- and long-lived objects from the same cache cannot
63  * anticipate the distribution of long-lived objects within the allocator's slab
64  * implementation. Even a small percentage of long-lived objects distributed
65  * randomly across many slabs can lead to a worst case scenario where the client
66  * frees the majority of its objects and the system gets back almost none of the
67  * slabs. Despite the client doing what it reasonably can to help the system
68  * reclaim memory, the allocator cannot shake free enough slabs because of
69  * lonely allocations stubbornly hanging on. Although the allocator is in a
70  * position to diagnose the fragmentation, there is nothing that the allocator
71  * by itself can do about it. It only takes a single allocated object to prevent
72  * an entire slab from being reclaimed, and any object handed out by
73  * kmem_cache_alloc() is by definition in the client's control. Conversely,
74  * although the client is in a position to move a long-lived object, it has no
75  * way of knowing if the object is causing fragmentation, and if so, where to
76  * move it. A solution necessarily requires further cooperation between the
77  * allocator and the client.
78  *
79  * 2. Move Callback
80  *
81  * The kmem slab consolidator therefore adds a move callback to the
82  * allocator/client interface, improving worst-case external fragmentation in
83  * kmem caches that supply a function to move objects from one memory location
84  * to another. In a situation of low memory kmem attempts to consolidate all of
85  * a cache's slabs at once; otherwise it works slowly to bring external
86  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
87  * thereby helping to avoid a low memory situation in the future.
88  *
89  * The callback has the following signature:
90  *
91  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
92  *
93  * It supplies the kmem client with two addresses: the allocated object that
94  * kmem wants to move and a buffer selected by kmem for the client to use as the
95  * copy destination. The callback is kmem's way of saying "Please get off of
96  * this buffer and use this one instead." kmem knows where it wants to move the
97  * object in order to best reduce fragmentation. All the client needs to know
98  * about the second argument (void *new) is that it is an allocated, constructed
99  * object ready to take the contents of the old object. When the move function
100  * is called, the system is likely to be low on memory, and the new object
101  * spares the client from having to worry about allocating memory for the
102  * requested move. The third argument supplies the size of the object, in case a
103  * single move function handles multiple caches whose objects differ only in
104  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
105  * user argument passed to the constructor, destructor, and reclaim functions is
106  * also passed to the move callback.
107  *
108  * 2.1 Setting the Move Callback
109  *
110  * The client sets the move callback after creating the cache and before
111  * allocating from it:
112  *
113  *	object_cache = kmem_cache_create(...);
114  *      kmem_cache_set_move(object_cache, object_move);
115  *
116  * 2.2 Move Callback Return Values
117  *
118  * Only the client knows about its own data and when is a good time to move it.
119  * The client is cooperating with kmem to return unused memory to the system,
120  * and kmem respectfully accepts this help at the client's convenience. When
121  * asked to move an object, the client can respond with any of the following:
122  *
123  *   typedef enum kmem_cbrc {
124  *           KMEM_CBRC_YES,
125  *           KMEM_CBRC_NO,
126  *           KMEM_CBRC_LATER,
127  *           KMEM_CBRC_DONT_NEED,
128  *           KMEM_CBRC_DONT_KNOW
129  *   } kmem_cbrc_t;
130  *
131  * The client must not explicitly kmem_cache_free() either of the objects passed
132  * to the callback, since kmem wants to free them directly to the slab layer
133  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
134  * objects to free:
135  *
136  *       YES: (Did it) The client moved the object, so kmem frees the old one.
137  *        NO: (Never) The client refused, so kmem frees the new object (the
138  *            unused copy destination). kmem also marks the slab of the old
139  *            object so as not to bother the client with further callbacks for
140  *            that object as long as the slab remains on the partial slab list.
141  *            (The system won't be getting the slab back as long as the
142  *            immovable object holds it hostage, so there's no point in moving
143  *            any of its objects.)
144  *     LATER: The client is using the object and cannot move it now, so kmem
145  *            frees the new object (the unused copy destination). kmem still
146  *            attempts to move other objects off the slab, since it expects to
147  *            succeed in clearing the slab in a later callback. The client
148  *            should use LATER instead of NO if the object is likely to become
149  *            movable very soon.
150  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
151  *            with the new object (the unused copy destination). This response
152  *            is the client's opportunity to be a model citizen and give back as
153  *            much as it can.
154  * DONT_KNOW: The client does not know about the object because
155  *            a) the client has just allocated the object and not yet put it
156  *               wherever it expects to find known objects
157  *            b) the client has removed the object from wherever it expects to
158  *               find known objects and is about to free it, or
159  *            c) the client has freed the object.
160  *            In all these cases (a, b, and c) kmem frees the new object (the
161  *            unused copy destination) and searches for the old object in the
162  *            magazine layer. If found, the object is removed from the magazine
163  *            layer and freed to the slab layer so it will no longer hold the
164  *            slab hostage.
165  *
166  * 2.3 Object States
167  *
168  * Neither kmem nor the client can be assumed to know the object's whereabouts
169  * at the time of the callback. An object belonging to a kmem cache may be in
170  * any of the following states:
171  *
172  * 1. Uninitialized on the slab
173  * 2. Allocated from the slab but not constructed (still uninitialized)
174  * 3. Allocated from the slab, constructed, but not yet ready for business
175  *    (not in a valid state for the move callback)
176  * 4. In use (valid and known to the client)
177  * 5. About to be freed (no longer in a valid state for the move callback)
178  * 6. Freed to a magazine (still constructed)
179  * 7. Allocated from a magazine, not yet ready for business (not in a valid
180  *    state for the move callback), and about to return to state #4
181  * 8. Deconstructed on a magazine that is about to be freed
182  * 9. Freed to the slab
183  *
184  * Since the move callback may be called at any time while the object is in any
185  * of the above states (except state #1), the client needs a safe way to
186  * determine whether or not it knows about the object. Specifically, the client
187  * needs to know whether or not the object is in state #4, the only state in
188  * which a move is valid. If the object is in any other state, the client should
189  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
190  * the object's fields.
191  *
192  * Note that although an object may be in state #4 when kmem initiates the move
193  * request, the object may no longer be in that state by the time kmem actually
194  * calls the move function. Not only does the client free objects
195  * asynchronously, kmem itself puts move requests on a queue where thay are
196  * pending until kmem processes them from another context. Also, objects freed
197  * to a magazine appear allocated from the point of view of the slab layer, so
198  * kmem may even initiate requests for objects in a state other than state #4.
199  *
200  * 2.3.1 Magazine Layer
201  *
202  * An important insight revealed by the states listed above is that the magazine
203  * layer is populated only by kmem_cache_free(). Magazines of constructed
204  * objects are never populated directly from the slab layer (which contains raw,
205  * unconstructed objects). Whenever an allocation request cannot be satisfied
206  * from the magazine layer, the magazines are bypassed and the request is
207  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
208  * the object constructor only when allocating from the slab layer, and only in
209  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
210  * the move callback. kmem does not preconstruct objects in anticipation of
211  * kmem_cache_alloc().
212  *
213  * 2.3.2 Object Constructor and Destructor
214  *
215  * If the client supplies a destructor, it must be valid to call the destructor
216  * on a newly created object (immediately after the constructor).
217  *
218  * 2.4 Recognizing Known Objects
219  *
220  * There is a simple test to determine safely whether or not the client knows
221  * about a given object in the move callback. It relies on the fact that kmem
222  * guarantees that the object of the move callback has only been touched by the
223  * client itself or else by kmem. kmem does this by ensuring that none of the
224  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
225  * callback is pending. When the last object on a slab is freed, if there is a
226  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
227  * slabs on that list until all pending callbacks are completed. That way,
228  * clients can be certain that the object of a move callback is in one of the
229  * states listed above, making it possible to distinguish known objects (in
230  * state #4) using the two low order bits of any pointer member (with the
231  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
232  * platforms).
233  *
234  * The test works as long as the client always transitions objects from state #4
235  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
236  * order bit of the client-designated pointer member. Since kmem only writes
237  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
238  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
239  * guaranteed to set at least one of the two low order bits. Therefore, given an
240  * object with a back pointer to a 'container_t *o_container', the client can
241  * test
242  *
243  *      container_t *container = object->o_container;
244  *      if ((uintptr_t)container & 0x3) {
245  *              return (KMEM_CBRC_DONT_KNOW);
246  *      }
247  *
248  * Typically, an object will have a pointer to some structure with a list or
249  * hash where objects from the cache are kept while in use. Assuming that the
250  * client has some way of knowing that the container structure is valid and will
251  * not go away during the move, and assuming that the structure includes a lock
252  * to protect whatever collection is used, then the client would continue as
253  * follows:
254  *
255  *	// Ensure that the container structure does not go away.
256  *      if (container_hold(container) == 0) {
257  *              return (KMEM_CBRC_DONT_KNOW);
258  *      }
259  *      mutex_enter(&container->c_objects_lock);
260  *      if (container != object->o_container) {
261  *              mutex_exit(&container->c_objects_lock);
262  *              container_rele(container);
263  *              return (KMEM_CBRC_DONT_KNOW);
264  *      }
265  *
266  * At this point the client knows that the object cannot be freed as long as
267  * c_objects_lock is held. Note that after acquiring the lock, the client must
268  * recheck the o_container pointer in case the object was removed just before
269  * acquiring the lock.
270  *
271  * When the client is about to free an object, it must first remove that object
272  * from the list, hash, or other structure where it is kept. At that time, to
273  * mark the object so it can be distinguished from the remaining, known objects,
274  * the client sets the designated low order bit:
275  *
276  *      mutex_enter(&container->c_objects_lock);
277  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
278  *      list_remove(&container->c_objects, object);
279  *      mutex_exit(&container->c_objects_lock);
280  *
281  * In the common case, the object is freed to the magazine layer, where it may
282  * be reused on a subsequent allocation without the overhead of calling the
283  * constructor. While in the magazine it appears allocated from the point of
284  * view of the slab layer, making it a candidate for the move callback. Most
285  * objects unrecognized by the client in the move callback fall into this
286  * category and are cheaply distinguished from known objects by the test
287  * described earlier. Since recognition is cheap for the client, and searching
288  * magazines is expensive for kmem, kmem defers searching until the client first
289  * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
290  * elsewhere does what it can to avoid bothering the client unnecessarily.
291  *
292  * Invalidating the designated pointer member before freeing the object marks
293  * the object to be avoided in the callback, and conversely, assigning a valid
294  * value to the designated pointer member after allocating the object makes the
295  * object fair game for the callback:
296  *
297  *      ... allocate object ...
298  *      ... set any initial state not set by the constructor ...
299  *
300  *      mutex_enter(&container->c_objects_lock);
301  *      list_insert_tail(&container->c_objects, object);
302  *      membar_producer();
303  *      object->o_container = container;
304  *      mutex_exit(&container->c_objects_lock);
305  *
306  * Note that everything else must be valid before setting o_container makes the
307  * object fair game for the move callback. The membar_producer() call ensures
308  * that all the object's state is written to memory before setting the pointer
309  * that transitions the object from state #3 or #7 (allocated, constructed, not
310  * yet in use) to state #4 (in use, valid). That's important because the move
311  * function has to check the validity of the pointer before it can safely
312  * acquire the lock protecting the collection where it expects to find known
313  * objects.
314  *
315  * This method of distinguishing known objects observes the usual symmetry:
316  * invalidating the designated pointer is the first thing the client does before
317  * freeing the object, and setting the designated pointer is the last thing the
318  * client does after allocating the object. Of course, the client is not
319  * required to use this method. Fundamentally, how the client recognizes known
320  * objects is completely up to the client, but this method is recommended as an
321  * efficient and safe way to take advantage of the guarantees made by kmem. If
322  * the entire object is arbitrary data without any markable bits from a suitable
323  * pointer member, then the client must find some other method, such as
324  * searching a hash table of known objects.
325  *
326  * 2.5 Preventing Objects From Moving
327  *
328  * Besides a way to distinguish known objects, the other thing that the client
329  * needs is a strategy to ensure that an object will not move while the client
330  * is actively using it. The details of satisfying this requirement tend to be
331  * highly cache-specific. It might seem that the same rules that let a client
332  * remove an object safely should also decide when an object can be moved
333  * safely. However, any object state that makes a removal attempt invalid is
334  * likely to be long-lasting for objects that the client does not expect to
335  * remove. kmem knows nothing about the object state and is equally likely (from
336  * the client's point of view) to request a move for any object in the cache,
337  * whether prepared for removal or not. Even a low percentage of objects stuck
338  * in place by unremovability will defeat the consolidator if the stuck objects
339  * are the same long-lived allocations likely to hold slabs hostage.
340  * Fundamentally, the consolidator is not aimed at common cases. Severe external
341  * fragmentation is a worst case scenario manifested as sparsely allocated
342  * slabs, by definition a low percentage of the cache's objects. When deciding
343  * what makes an object movable, keep in mind the goal of the consolidator: to
344  * bring worst-case external fragmentation within the limits guaranteed for
345  * internal fragmentation. Removability is a poor criterion if it is likely to
346  * exclude more than an insignificant percentage of objects for long periods of
347  * time.
348  *
349  * A tricky general solution exists, and it has the advantage of letting you
350  * move any object at almost any moment, practically eliminating the likelihood
351  * that an object can hold a slab hostage. However, if there is a cache-specific
352  * way to ensure that an object is not actively in use in the vast majority of
353  * cases, a simpler solution that leverages this cache-specific knowledge is
354  * preferred.
355  *
356  * 2.5.1 Cache-Specific Solution
357  *
358  * As an example of a cache-specific solution, the ZFS znode cache takes
359  * advantage of the fact that the vast majority of znodes are only being
360  * referenced from the DNLC. (A typical case might be a few hundred in active
361  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
362  * client has established that it recognizes the znode and can access its fields
363  * safely (using the method described earlier), it then tests whether the znode
364  * is referenced by anything other than the DNLC. If so, it assumes that the
365  * znode may be in active use and is unsafe to move, so it drops its locks and
366  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
367  * else znodes are used, no change is needed to protect against the possibility
368  * of the znode moving. The disadvantage is that it remains possible for an
369  * application to hold a znode slab hostage with an open file descriptor.
370  * However, this case ought to be rare and the consolidator has a way to deal
371  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
372  * object, kmem eventually stops believing it and treats the slab as if the
373  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
374  * then focus on getting it off of the partial slab list by allocating rather
375  * than freeing all of its objects. (Either way of getting a slab off the
376  * free list reduces fragmentation.)
377  *
378  * 2.5.2 General Solution
379  *
380  * The general solution, on the other hand, requires an explicit hold everywhere
381  * the object is used to prevent it from moving. To keep the client locking
382  * strategy as uncomplicated as possible, kmem guarantees the simplifying
383  * assumption that move callbacks are sequential, even across multiple caches.
384  * Internally, a global queue processed by a single thread supports all caches
385  * implementing the callback function. No matter how many caches supply a move
386  * function, the consolidator never moves more than one object at a time, so the
387  * client does not have to worry about tricky lock ordering involving several
388  * related objects from different kmem caches.
389  *
390  * The general solution implements the explicit hold as a read-write lock, which
391  * allows multiple readers to access an object from the cache simultaneously
392  * while a single writer is excluded from moving it. A single rwlock for the
393  * entire cache would lock out all threads from using any of the cache's objects
394  * even though only a single object is being moved, so to reduce contention,
395  * the client can fan out the single rwlock into an array of rwlocks hashed by
396  * the object address, making it probable that moving one object will not
397  * prevent other threads from using a different object. The rwlock cannot be a
398  * member of the object itself, because the possibility of the object moving
399  * makes it unsafe to access any of the object's fields until the lock is
400  * acquired.
401  *
402  * Assuming a small, fixed number of locks, it's possible that multiple objects
403  * will hash to the same lock. A thread that needs to use multiple objects in
404  * the same function may acquire the same lock multiple times. Since rwlocks are
405  * reentrant for readers, and since there is never more than a single writer at
406  * a time (assuming that the client acquires the lock as a writer only when
407  * moving an object inside the callback), there would seem to be no problem.
408  * However, a client locking multiple objects in the same function must handle
409  * one case of potential deadlock: Assume that thread A needs to prevent both
410  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
411  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
412  * same lock, that thread A will acquire the lock for object 1 as a reader
413  * before thread B sets the lock's write-wanted bit, preventing thread A from
414  * reacquiring the lock for object 2 as a reader. Unable to make forward
415  * progress, thread A will never release the lock for object 1, resulting in
416  * deadlock.
417  *
418  * There are two ways of avoiding the deadlock just described. The first is to
419  * use rw_tryenter() rather than rw_enter() in the callback function when
420  * attempting to acquire the lock as a writer. If tryenter discovers that the
421  * same object (or another object hashed to the same lock) is already in use, it
422  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
423  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
424  * since it allows a thread to acquire the lock as a reader in spite of a
425  * waiting writer. This second approach insists on moving the object now, no
426  * matter how many readers the move function must wait for in order to do so,
427  * and could delay the completion of the callback indefinitely (blocking
428  * callbacks to other clients). In practice, a less insistent callback using
429  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
430  * little reason to use anything else.
431  *
432  * Avoiding deadlock is not the only problem that an implementation using an
433  * explicit hold needs to solve. Locking the object in the first place (to
434  * prevent it from moving) remains a problem, since the object could move
435  * between the time you obtain a pointer to the object and the time you acquire
436  * the rwlock hashed to that pointer value. Therefore the client needs to
437  * recheck the value of the pointer after acquiring the lock, drop the lock if
438  * the value has changed, and try again. This requires a level of indirection:
439  * something that points to the object rather than the object itself, that the
440  * client can access safely while attempting to acquire the lock. (The object
441  * itself cannot be referenced safely because it can move at any time.)
442  * The following lock-acquisition function takes whatever is safe to reference
443  * (arg), follows its pointer to the object (using function f), and tries as
444  * often as necessary to acquire the hashed lock and verify that the object
445  * still has not moved:
446  *
447  *      object_t *
448  *      object_hold(object_f f, void *arg)
449  *      {
450  *              object_t *op;
451  *
452  *              op = f(arg);
453  *              if (op == NULL) {
454  *                      return (NULL);
455  *              }
456  *
457  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
458  *              while (op != f(arg)) {
459  *                      rw_exit(OBJECT_RWLOCK(op));
460  *                      op = f(arg);
461  *                      if (op == NULL) {
462  *                              break;
463  *                      }
464  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
465  *              }
466  *
467  *              return (op);
468  *      }
469  *
470  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
471  * lock reacquisition loop, while necessary, almost never executes. The function
472  * pointer f (used to obtain the object pointer from arg) has the following type
473  * definition:
474  *
475  *      typedef object_t *(*object_f)(void *arg);
476  *
477  * An object_f implementation is likely to be as simple as accessing a structure
478  * member:
479  *
480  *      object_t *
481  *      s_object(void *arg)
482  *      {
483  *              something_t *sp = arg;
484  *              return (sp->s_object);
485  *      }
486  *
487  * The flexibility of a function pointer allows the path to the object to be
488  * arbitrarily complex and also supports the notion that depending on where you
489  * are using the object, you may need to get it from someplace different.
490  *
491  * The function that releases the explicit hold is simpler because it does not
492  * have to worry about the object moving:
493  *
494  *      void
495  *      object_rele(object_t *op)
496  *      {
497  *              rw_exit(OBJECT_RWLOCK(op));
498  *      }
499  *
500  * The caller is spared these details so that obtaining and releasing an
501  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
502  * of object_hold() only needs to know that the returned object pointer is valid
503  * if not NULL and that the object will not move until released.
504  *
505  * Although object_hold() prevents an object from moving, it does not prevent it
506  * from being freed. The caller must take measures before calling object_hold()
507  * (afterwards is too late) to ensure that the held object cannot be freed. The
508  * caller must do so without accessing the unsafe object reference, so any lock
509  * or reference count used to ensure the continued existence of the object must
510  * live outside the object itself.
511  *
512  * Obtaining a new object is a special case where an explicit hold is impossible
513  * for the caller. Any function that returns a newly allocated object (either as
514  * a return value, or as an in-out paramter) must return it already held; after
515  * the caller gets it is too late, since the object cannot be safely accessed
516  * without the level of indirection described earlier. The following
517  * object_alloc() example uses the same code shown earlier to transition a new
518  * object into the state of being recognized (by the client) as a known object.
519  * The function must acquire the hold (rw_enter) before that state transition
520  * makes the object movable:
521  *
522  *      static object_t *
523  *      object_alloc(container_t *container)
524  *      {
525  *              object_t *object = kmem_cache_alloc(object_cache, 0);
526  *              ... set any initial state not set by the constructor ...
527  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
528  *              mutex_enter(&container->c_objects_lock);
529  *              list_insert_tail(&container->c_objects, object);
530  *              membar_producer();
531  *              object->o_container = container;
532  *              mutex_exit(&container->c_objects_lock);
533  *              return (object);
534  *      }
535  *
536  * Functions that implicitly acquire an object hold (any function that calls
537  * object_alloc() to supply an object for the caller) need to be carefully noted
538  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
539  * prevent all objects hashed to the affected rwlocks from ever being moved.
540  *
541  * The pointer to a held object can be hashed to the holding rwlock even after
542  * the object has been freed. Although it is possible to release the hold
543  * after freeing the object, you may decide to release the hold implicitly in
544  * whatever function frees the object, so as to release the hold as soon as
545  * possible, and for the sake of symmetry with the function that implicitly
546  * acquires the hold when it allocates the object. Here, object_free() releases
547  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
548  * matching pair with object_hold():
549  *
550  *      void
551  *      object_free(object_t *object)
552  *      {
553  *              container_t *container;
554  *
555  *              ASSERT(object_held(object));
556  *              container = object->o_container;
557  *              mutex_enter(&container->c_objects_lock);
558  *              object->o_container =
559  *                  (void *)((uintptr_t)object->o_container | 0x1);
560  *              list_remove(&container->c_objects, object);
561  *              mutex_exit(&container->c_objects_lock);
562  *              object_rele(object);
563  *              kmem_cache_free(object_cache, object);
564  *      }
565  *
566  * Note that object_free() cannot safely accept an object pointer as an argument
567  * unless the object is already held. Any function that calls object_free()
568  * needs to be carefully noted since it similarly forms a matching pair with
569  * object_hold().
570  *
571  * To complete the picture, the following callback function implements the
572  * general solution by moving objects only if they are currently unheld:
573  *
574  *      static kmem_cbrc_t
575  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
576  *      {
577  *              object_t *op = buf, *np = newbuf;
578  *              container_t *container;
579  *
580  *              container = op->o_container;
581  *              if ((uintptr_t)container & 0x3) {
582  *                      return (KMEM_CBRC_DONT_KNOW);
583  *              }
584  *
585  *	        // Ensure that the container structure does not go away.
586  *              if (container_hold(container) == 0) {
587  *                      return (KMEM_CBRC_DONT_KNOW);
588  *              }
589  *
590  *              mutex_enter(&container->c_objects_lock);
591  *              if (container != op->o_container) {
592  *                      mutex_exit(&container->c_objects_lock);
593  *                      container_rele(container);
594  *                      return (KMEM_CBRC_DONT_KNOW);
595  *              }
596  *
597  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
598  *                      mutex_exit(&container->c_objects_lock);
599  *                      container_rele(container);
600  *                      return (KMEM_CBRC_LATER);
601  *              }
602  *
603  *              object_move_impl(op, np); // critical section
604  *              rw_exit(OBJECT_RWLOCK(op));
605  *
606  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
607  *              list_link_replace(&op->o_link_node, &np->o_link_node);
608  *              mutex_exit(&container->c_objects_lock);
609  *              container_rele(container);
610  *              return (KMEM_CBRC_YES);
611  *      }
612  *
613  * Note that object_move() must invalidate the designated o_container pointer of
614  * the old object in the same way that object_free() does, since kmem will free
615  * the object in response to the KMEM_CBRC_YES return value.
616  *
617  * The lock order in object_move() differs from object_alloc(), which locks
618  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
619  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
620  * not a problem. Holding the lock on the object list in the example above
621  * through the entire callback not only prevents the object from going away, it
622  * also allows you to lock the list elsewhere and know that none of its elements
623  * will move during iteration.
624  *
625  * Adding an explicit hold everywhere an object from the cache is used is tricky
626  * and involves much more change to client code than a cache-specific solution
627  * that leverages existing state to decide whether or not an object is
628  * movable. However, this approach has the advantage that no object remains
629  * immovable for any significant length of time, making it extremely unlikely
630  * that long-lived allocations can continue holding slabs hostage; and it works
631  * for any cache.
632  *
633  * 3. Consolidator Implementation
634  *
635  * Once the client supplies a move function that a) recognizes known objects and
636  * b) avoids moving objects that are actively in use, the remaining work is up
637  * to the consolidator to decide which objects to move and when to issue
638  * callbacks.
639  *
640  * The consolidator relies on the fact that a cache's slabs are ordered by
641  * usage. Each slab has a fixed number of objects. Depending on the slab's
642  * "color" (the offset of the first object from the beginning of the slab;
643  * offsets are staggered to mitigate false sharing of cache lines) it is either
644  * the maximum number of objects per slab determined at cache creation time or
645  * else the number closest to the maximum that fits within the space remaining
646  * after the initial offset. A completely allocated slab may contribute some
647  * internal fragmentation (per-slab overhead) but no external fragmentation, so
648  * it is of no interest to the consolidator. At the other extreme, slabs whose
649  * objects have all been freed to the slab are released to the virtual memory
650  * (VM) subsystem (objects freed to magazines are still allocated as far as the
651  * slab is concerned). External fragmentation exists when there are slabs
652  * somewhere between these extremes. A partial slab has at least one but not all
653  * of its objects allocated. The more partial slabs, and the fewer allocated
654  * objects on each of them, the higher the fragmentation. Hence the
655  * consolidator's overall strategy is to reduce the number of partial slabs by
656  * moving allocated objects from the least allocated slabs to the most allocated
657  * slabs.
658  *
659  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
660  * slabs are kept separately in an unordered list. Since the majority of slabs
661  * tend to be completely allocated (a typical unfragmented cache may have
662  * thousands of complete slabs and only a single partial slab), separating
663  * complete slabs improves the efficiency of partial slab ordering, since the
664  * complete slabs do not affect the depth or balance of the AVL tree. This
665  * ordered sequence of partial slabs acts as a "free list" supplying objects for
666  * allocation requests.
667  *
668  * Objects are always allocated from the first partial slab in the free list,
669  * where the allocation is most likely to eliminate a partial slab (by
670  * completely allocating it). Conversely, when a single object from a completely
671  * allocated slab is freed to the slab, that slab is added to the front of the
672  * free list. Since most free list activity involves highly allocated slabs
673  * coming and going at the front of the list, slabs tend naturally toward the
674  * ideal order: highly allocated at the front, sparsely allocated at the back.
675  * Slabs with few allocated objects are likely to become completely free if they
676  * keep a safe distance away from the front of the free list. Slab misorders
677  * interfere with the natural tendency of slabs to become completely free or
678  * completely allocated. For example, a slab with a single allocated object
679  * needs only a single free to escape the cache; its natural desire is
680  * frustrated when it finds itself at the front of the list where a second
681  * allocation happens just before the free could have released it. Another slab
682  * with all but one object allocated might have supplied the buffer instead, so
683  * that both (as opposed to neither) of the slabs would have been taken off the
684  * free list.
685  *
686  * Although slabs tend naturally toward the ideal order, misorders allowed by a
687  * simple list implementation defeat the consolidator's strategy of merging
688  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
689  * needs another way to fix misorders to optimize its callback strategy. One
690  * approach is to periodically scan a limited number of slabs, advancing a
691  * marker to hold the current scan position, and to move extreme misorders to
692  * the front or back of the free list and to the front or back of the current
693  * scan range. By making consecutive scan ranges overlap by one slab, the least
694  * allocated slab in the current range can be carried along from the end of one
695  * scan to the start of the next.
696  *
697  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
698  * task, however. Since most of the cache's activity is in the magazine layer,
699  * and allocations from the slab layer represent only a startup cost, the
700  * overhead of maintaining a balanced tree is not a significant concern compared
701  * to the opportunity of reducing complexity by eliminating the partial slab
702  * scanner just described. The overhead of an AVL tree is minimized by
703  * maintaining only partial slabs in the tree and keeping completely allocated
704  * slabs separately in a list. To avoid increasing the size of the slab
705  * structure the AVL linkage pointers are reused for the slab's list linkage,
706  * since the slab will always be either partial or complete, never stored both
707  * ways at the same time. To further minimize the overhead of the AVL tree the
708  * compare function that orders partial slabs by usage divides the range of
709  * allocated object counts into bins such that counts within the same bin are
710  * considered equal. Binning partial slabs makes it less likely that allocating
711  * or freeing a single object will change the slab's order, requiring a tree
712  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
713  * requiring some rebalancing of the tree). Allocation counts closest to
714  * completely free and completely allocated are left unbinned (finely sorted) to
715  * better support the consolidator's strategy of merging slabs at either
716  * extreme.
717  *
718  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
719  *
720  * The consolidator piggybacks on the kmem maintenance thread and is called on
721  * the same interval as kmem_cache_update(), once per cache every fifteen
722  * seconds. kmem maintains a running count of unallocated objects in the slab
723  * layer (cache_bufslab). The consolidator checks whether that number exceeds
724  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
725  * there is a significant number of slabs in the cache (arbitrarily a minimum
726  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
727  * working set are included in the assessment, and magazines in the depot are
728  * reaped if those objects would lift cache_bufslab above the fragmentation
729  * threshold. Once the consolidator decides that a cache is fragmented, it looks
730  * for a candidate slab to reclaim, starting at the end of the partial slab free
731  * list and scanning backwards. At first the consolidator is choosy: only a slab
732  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
733  * single allocated object, regardless of percentage). If there is difficulty
734  * finding a candidate slab, kmem raises the allocation threshold incrementally,
735  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
736  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
737  * even in the worst case of every slab in the cache being almost 7/8 allocated.
738  * The threshold can also be lowered incrementally when candidate slabs are easy
739  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
740  * is no longer fragmented.
741  *
742  * 3.2 Generating Callbacks
743  *
744  * Once an eligible slab is chosen, a callback is generated for every allocated
745  * object on the slab, in the hope that the client will move everything off the
746  * slab and make it reclaimable. Objects selected as move destinations are
747  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
748  * order (most allocated at the front, least allocated at the back) and a
749  * cooperative client, the consolidator will succeed in removing slabs from both
750  * ends of the free list, completely allocating on the one hand and completely
751  * freeing on the other. Objects selected as move destinations are allocated in
752  * the kmem maintenance thread where move requests are enqueued. A separate
753  * callback thread removes pending callbacks from the queue and calls the
754  * client. The separate thread ensures that client code (the move function) does
755  * not interfere with internal kmem maintenance tasks. A map of pending
756  * callbacks keyed by object address (the object to be moved) is checked to
757  * ensure that duplicate callbacks are not generated for the same object.
758  * Allocating the move destination (the object to move to) prevents subsequent
759  * callbacks from selecting the same destination as an earlier pending callback.
760  *
761  * Move requests can also be generated by kmem_cache_reap() when the system is
762  * desperate for memory and by kmem_cache_move_notify(), called by the client to
763  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
764  * The map of pending callbacks is protected by the same lock that protects the
765  * slab layer.
766  *
767  * When the system is desperate for memory, kmem does not bother to determine
768  * whether or not the cache exceeds the fragmentation threshold, but tries to
769  * consolidate as many slabs as possible. Normally, the consolidator chews
770  * slowly, one sparsely allocated slab at a time during each maintenance
771  * interval that the cache is fragmented. When desperate, the consolidator
772  * starts at the last partial slab and enqueues callbacks for every allocated
773  * object on every partial slab, working backwards until it reaches the first
774  * partial slab. The first partial slab, meanwhile, advances in pace with the
775  * consolidator as allocations to supply move destinations for the enqueued
776  * callbacks use up the highly allocated slabs at the front of the free list.
777  * Ideally, the overgrown free list collapses like an accordion, starting at
778  * both ends and ending at the center with a single partial slab.
779  *
780  * 3.3 Client Responses
781  *
782  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
783  * marks the slab that supplied the stuck object non-reclaimable and moves it to
784  * front of the free list. The slab remains marked as long as it remains on the
785  * free list, and it appears more allocated to the partial slab compare function
786  * than any unmarked slab, no matter how many of its objects are allocated.
787  * Since even one immovable object ties up the entire slab, the goal is to
788  * completely allocate any slab that cannot be completely freed. kmem does not
789  * bother generating callbacks to move objects from a marked slab unless the
790  * system is desperate.
791  *
792  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
793  * slab. If the client responds LATER too many times, kmem disbelieves and
794  * treats the response as a NO. The count is cleared when the slab is taken off
795  * the partial slab list or when the client moves one of the slab's objects.
796  *
797  * 4. Observability
798  *
799  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
800  * the ::kmem_slabs dcmd. For a complete description of the command, enter
801  * '::help kmem_slabs' at the mdb prompt.
802  */
803 
804 #include <sys/kmem_impl.h>
805 #include <sys/vmem_impl.h>
806 #include <sys/param.h>
807 #include <sys/sysmacros.h>
808 #include <sys/vm.h>
809 #include <sys/proc.h>
810 #include <sys/tuneable.h>
811 #include <sys/systm.h>
812 #include <sys/cmn_err.h>
813 #include <sys/debug.h>
814 #include <sys/sdt.h>
815 #include <sys/mutex.h>
816 #include <sys/bitmap.h>
817 #include <sys/atomic.h>
818 #include <sys/kobj.h>
819 #include <sys/disp.h>
820 #include <vm/seg_kmem.h>
821 #include <sys/log.h>
822 #include <sys/callb.h>
823 #include <sys/taskq.h>
824 #include <sys/modctl.h>
825 #include <sys/reboot.h>
826 #include <sys/id32.h>
827 #include <sys/zone.h>
828 #include <sys/netstack.h>
829 #ifdef	DEBUG
830 #include <sys/random.h>
831 #endif
832 
833 extern void streams_msg_init(void);
834 extern int segkp_fromheap;
835 extern void segkp_cache_free(void);
836 extern int callout_init_done;
837 
838 struct kmem_cache_kstat {
839 	kstat_named_t	kmc_buf_size;
840 	kstat_named_t	kmc_align;
841 	kstat_named_t	kmc_chunk_size;
842 	kstat_named_t	kmc_slab_size;
843 	kstat_named_t	kmc_alloc;
844 	kstat_named_t	kmc_alloc_fail;
845 	kstat_named_t	kmc_free;
846 	kstat_named_t	kmc_depot_alloc;
847 	kstat_named_t	kmc_depot_free;
848 	kstat_named_t	kmc_depot_contention;
849 	kstat_named_t	kmc_slab_alloc;
850 	kstat_named_t	kmc_slab_free;
851 	kstat_named_t	kmc_buf_constructed;
852 	kstat_named_t	kmc_buf_avail;
853 	kstat_named_t	kmc_buf_inuse;
854 	kstat_named_t	kmc_buf_total;
855 	kstat_named_t	kmc_buf_max;
856 	kstat_named_t	kmc_slab_create;
857 	kstat_named_t	kmc_slab_destroy;
858 	kstat_named_t	kmc_vmem_source;
859 	kstat_named_t	kmc_hash_size;
860 	kstat_named_t	kmc_hash_lookup_depth;
861 	kstat_named_t	kmc_hash_rescale;
862 	kstat_named_t	kmc_full_magazines;
863 	kstat_named_t	kmc_empty_magazines;
864 	kstat_named_t	kmc_magazine_size;
865 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
866 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
867 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
868 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
869 	kstat_named_t	kmc_move_yes;
870 	kstat_named_t	kmc_move_no;
871 	kstat_named_t	kmc_move_later;
872 	kstat_named_t	kmc_move_dont_need;
873 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
874 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
875 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
876 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
877 } kmem_cache_kstat = {
878 	{ "buf_size",		KSTAT_DATA_UINT64 },
879 	{ "align",		KSTAT_DATA_UINT64 },
880 	{ "chunk_size",		KSTAT_DATA_UINT64 },
881 	{ "slab_size",		KSTAT_DATA_UINT64 },
882 	{ "alloc",		KSTAT_DATA_UINT64 },
883 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
884 	{ "free",		KSTAT_DATA_UINT64 },
885 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
886 	{ "depot_free",		KSTAT_DATA_UINT64 },
887 	{ "depot_contention",	KSTAT_DATA_UINT64 },
888 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
889 	{ "slab_free",		KSTAT_DATA_UINT64 },
890 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
891 	{ "buf_avail",		KSTAT_DATA_UINT64 },
892 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
893 	{ "buf_total",		KSTAT_DATA_UINT64 },
894 	{ "buf_max",		KSTAT_DATA_UINT64 },
895 	{ "slab_create",	KSTAT_DATA_UINT64 },
896 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
897 	{ "vmem_source",	KSTAT_DATA_UINT64 },
898 	{ "hash_size",		KSTAT_DATA_UINT64 },
899 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
900 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
901 	{ "full_magazines",	KSTAT_DATA_UINT64 },
902 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
903 	{ "magazine_size",	KSTAT_DATA_UINT64 },
904 	{ "reap",		KSTAT_DATA_UINT64 },
905 	{ "defrag",		KSTAT_DATA_UINT64 },
906 	{ "scan",		KSTAT_DATA_UINT64 },
907 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
908 	{ "move_yes",		KSTAT_DATA_UINT64 },
909 	{ "move_no",		KSTAT_DATA_UINT64 },
910 	{ "move_later",		KSTAT_DATA_UINT64 },
911 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
912 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
913 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
914 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
915 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
916 };
917 
918 static kmutex_t kmem_cache_kstat_lock;
919 
920 /*
921  * The default set of caches to back kmem_alloc().
922  * These sizes should be reevaluated periodically.
923  *
924  * We want allocations that are multiples of the coherency granularity
925  * (64 bytes) to be satisfied from a cache which is a multiple of 64
926  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
927  * the next kmem_cache_size greater than or equal to it must be a
928  * multiple of 64.
929  *
930  * We split the table into two sections:  size <= 4k and size > 4k.  This
931  * saves a lot of space and cache footprint in our cache tables.
932  */
933 static const int kmem_alloc_sizes[] = {
934 	1 * 8,
935 	2 * 8,
936 	3 * 8,
937 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
938 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
939 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
940 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
941 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
942 	P2ALIGN(8192 / 7, 64),
943 	P2ALIGN(8192 / 6, 64),
944 	P2ALIGN(8192 / 5, 64),
945 	P2ALIGN(8192 / 4, 64),
946 	P2ALIGN(8192 / 3, 64),
947 	P2ALIGN(8192 / 2, 64),
948 };
949 
950 static const int kmem_big_alloc_sizes[] = {
951 	2 * 4096,	3 * 4096,
952 	2 * 8192,	3 * 8192,
953 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
954 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
955 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
956 	16 * 8192
957 };
958 
959 #define	KMEM_MAXBUF		4096
960 #define	KMEM_BIG_MAXBUF_32BIT	32768
961 #define	KMEM_BIG_MAXBUF		131072
962 
963 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
964 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
965 
966 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
967 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
968 
969 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
970 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
971 
972 static kmem_magtype_t kmem_magtype[] = {
973 	{ 1,	8,	3200,	65536	},
974 	{ 3,	16,	256,	32768	},
975 	{ 7,	32,	64,	16384	},
976 	{ 15,	64,	0,	8192	},
977 	{ 31,	64,	0,	4096	},
978 	{ 47,	64,	0,	2048	},
979 	{ 63,	64,	0,	1024	},
980 	{ 95,	64,	0,	512	},
981 	{ 143,	64,	0,	0	},
982 };
983 
984 static uint32_t kmem_reaping;
985 static uint32_t kmem_reaping_idspace;
986 
987 /*
988  * kmem tunables
989  */
990 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
991 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
992 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
993 int kmem_panic = 1;		/* whether to panic on error */
994 int kmem_logging = 1;		/* kmem_log_enter() override */
995 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
996 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
997 size_t kmem_content_log_size;	/* content log size [2% of memory] */
998 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
999 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
1000 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1001 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1002 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1003 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1004 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1005 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1006 
1007 #ifdef _LP64
1008 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1009 #else
1010 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1011 #endif
1012 
1013 #ifdef DEBUG
1014 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1015 #else
1016 int kmem_flags = 0;
1017 #endif
1018 int kmem_ready;
1019 
1020 static kmem_cache_t	*kmem_slab_cache;
1021 static kmem_cache_t	*kmem_bufctl_cache;
1022 static kmem_cache_t	*kmem_bufctl_audit_cache;
1023 
1024 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1025 static list_t		kmem_caches;
1026 
1027 static taskq_t		*kmem_taskq;
1028 static kmutex_t		kmem_flags_lock;
1029 static vmem_t		*kmem_metadata_arena;
1030 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1031 static vmem_t		*kmem_cache_arena;
1032 static vmem_t		*kmem_hash_arena;
1033 static vmem_t		*kmem_log_arena;
1034 static vmem_t		*kmem_oversize_arena;
1035 static vmem_t		*kmem_va_arena;
1036 static vmem_t		*kmem_default_arena;
1037 static vmem_t		*kmem_firewall_va_arena;
1038 static vmem_t		*kmem_firewall_arena;
1039 
1040 /*
1041  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1042  * turned on when DEBUG is also defined.
1043  */
1044 #ifdef	DEBUG
1045 #define	KMEM_STATS
1046 #endif	/* DEBUG */
1047 
1048 #ifdef	KMEM_STATS
1049 #define	KMEM_STAT_ADD(stat)			((stat)++)
1050 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1051 #else
1052 #define	KMEM_STAT_ADD(stat)			/* nothing */
1053 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1054 #endif	/* KMEM_STATS */
1055 
1056 /*
1057  * kmem slab consolidator thresholds (tunables)
1058  */
1059 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1060 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1061 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1062 /*
1063  * Maximum number of slabs from which to move buffers during a single
1064  * maintenance interval while the system is not low on memory.
1065  */
1066 size_t kmem_reclaim_max_slabs = 1;
1067 /*
1068  * Number of slabs to scan backwards from the end of the partial slab list
1069  * when searching for buffers to relocate.
1070  */
1071 size_t kmem_reclaim_scan_range = 12;
1072 
1073 #ifdef	KMEM_STATS
1074 static struct {
1075 	uint64_t kms_callbacks;
1076 	uint64_t kms_yes;
1077 	uint64_t kms_no;
1078 	uint64_t kms_later;
1079 	uint64_t kms_dont_need;
1080 	uint64_t kms_dont_know;
1081 	uint64_t kms_hunt_found_mag;
1082 	uint64_t kms_hunt_found_slab;
1083 	uint64_t kms_hunt_alloc_fail;
1084 	uint64_t kms_hunt_lucky;
1085 	uint64_t kms_notify;
1086 	uint64_t kms_notify_callbacks;
1087 	uint64_t kms_disbelief;
1088 	uint64_t kms_already_pending;
1089 	uint64_t kms_callback_alloc_fail;
1090 	uint64_t kms_callback_taskq_fail;
1091 	uint64_t kms_endscan_slab_dead;
1092 	uint64_t kms_endscan_slab_destroyed;
1093 	uint64_t kms_endscan_nomem;
1094 	uint64_t kms_endscan_refcnt_changed;
1095 	uint64_t kms_endscan_nomove_changed;
1096 	uint64_t kms_endscan_freelist;
1097 	uint64_t kms_avl_update;
1098 	uint64_t kms_avl_noupdate;
1099 	uint64_t kms_no_longer_reclaimable;
1100 	uint64_t kms_notify_no_longer_reclaimable;
1101 	uint64_t kms_notify_slab_dead;
1102 	uint64_t kms_notify_slab_destroyed;
1103 	uint64_t kms_alloc_fail;
1104 	uint64_t kms_constructor_fail;
1105 	uint64_t kms_dead_slabs_freed;
1106 	uint64_t kms_defrags;
1107 	uint64_t kms_scans;
1108 	uint64_t kms_scan_depot_ws_reaps;
1109 	uint64_t kms_debug_reaps;
1110 	uint64_t kms_debug_scans;
1111 } kmem_move_stats;
1112 #endif	/* KMEM_STATS */
1113 
1114 /* consolidator knobs */
1115 static boolean_t kmem_move_noreap;
1116 static boolean_t kmem_move_blocked;
1117 static boolean_t kmem_move_fulltilt;
1118 static boolean_t kmem_move_any_partial;
1119 
1120 #ifdef	DEBUG
1121 /*
1122  * kmem consolidator debug tunables:
1123  * Ensure code coverage by occasionally running the consolidator even when the
1124  * caches are not fragmented (they may never be). These intervals are mean time
1125  * in cache maintenance intervals (kmem_cache_update).
1126  */
1127 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1128 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1129 #endif	/* DEBUG */
1130 
1131 static kmem_cache_t	*kmem_defrag_cache;
1132 static kmem_cache_t	*kmem_move_cache;
1133 static taskq_t		*kmem_move_taskq;
1134 
1135 static void kmem_cache_scan(kmem_cache_t *);
1136 static void kmem_cache_defrag(kmem_cache_t *);
1137 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1138 
1139 
1140 kmem_log_header_t	*kmem_transaction_log;
1141 kmem_log_header_t	*kmem_content_log;
1142 kmem_log_header_t	*kmem_failure_log;
1143 kmem_log_header_t	*kmem_slab_log;
1144 
1145 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1146 
1147 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1148 	if ((count) > 0) {						\
1149 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1150 		pc_t *_e;						\
1151 		/* memmove() the old entries down one notch */		\
1152 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1153 			*_e = *(_e - 1);				\
1154 		*_s = (uintptr_t)(caller);				\
1155 	}
1156 
1157 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1158 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1159 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1160 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1161 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1162 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1163 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1164 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1165 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1166 
1167 struct {
1168 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1169 	int		kmp_error;	/* type of kmem error */
1170 	void		*kmp_buffer;	/* buffer that induced panic */
1171 	void		*kmp_realbuf;	/* real start address for buffer */
1172 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1173 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1174 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1175 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1176 } kmem_panic_info;
1177 
1178 
1179 static void
1180 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1181 {
1182 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1183 	uint64_t *buf = buf_arg;
1184 
1185 	while (buf < bufend)
1186 		*buf++ = pattern;
1187 }
1188 
1189 static void *
1190 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1191 {
1192 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1193 	uint64_t *buf;
1194 
1195 	for (buf = buf_arg; buf < bufend; buf++)
1196 		if (*buf != pattern)
1197 			return (buf);
1198 	return (NULL);
1199 }
1200 
1201 static void *
1202 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1203 {
1204 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1205 	uint64_t *buf;
1206 
1207 	for (buf = buf_arg; buf < bufend; buf++) {
1208 		if (*buf != old) {
1209 			copy_pattern(old, buf_arg,
1210 			    (char *)buf - (char *)buf_arg);
1211 			return (buf);
1212 		}
1213 		*buf = new;
1214 	}
1215 
1216 	return (NULL);
1217 }
1218 
1219 static void
1220 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1221 {
1222 	kmem_cache_t *cp;
1223 
1224 	mutex_enter(&kmem_cache_lock);
1225 	for (cp = list_head(&kmem_caches); cp != NULL;
1226 	    cp = list_next(&kmem_caches, cp))
1227 		if (tq != NULL)
1228 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1229 			    tqflag);
1230 		else
1231 			func(cp);
1232 	mutex_exit(&kmem_cache_lock);
1233 }
1234 
1235 static void
1236 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1237 {
1238 	kmem_cache_t *cp;
1239 
1240 	mutex_enter(&kmem_cache_lock);
1241 	for (cp = list_head(&kmem_caches); cp != NULL;
1242 	    cp = list_next(&kmem_caches, cp)) {
1243 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1244 			continue;
1245 		if (tq != NULL)
1246 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1247 			    tqflag);
1248 		else
1249 			func(cp);
1250 	}
1251 	mutex_exit(&kmem_cache_lock);
1252 }
1253 
1254 /*
1255  * Debugging support.  Given a buffer address, find its slab.
1256  */
1257 static kmem_slab_t *
1258 kmem_findslab(kmem_cache_t *cp, void *buf)
1259 {
1260 	kmem_slab_t *sp;
1261 
1262 	mutex_enter(&cp->cache_lock);
1263 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1264 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1265 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1266 			mutex_exit(&cp->cache_lock);
1267 			return (sp);
1268 		}
1269 	}
1270 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1271 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1272 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1273 			mutex_exit(&cp->cache_lock);
1274 			return (sp);
1275 		}
1276 	}
1277 	mutex_exit(&cp->cache_lock);
1278 
1279 	return (NULL);
1280 }
1281 
1282 static void
1283 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1284 {
1285 	kmem_buftag_t *btp = NULL;
1286 	kmem_bufctl_t *bcp = NULL;
1287 	kmem_cache_t *cp = cparg;
1288 	kmem_slab_t *sp;
1289 	uint64_t *off;
1290 	void *buf = bufarg;
1291 
1292 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1293 
1294 	kmem_panic_info.kmp_timestamp = gethrtime();
1295 
1296 	sp = kmem_findslab(cp, buf);
1297 	if (sp == NULL) {
1298 		for (cp = list_tail(&kmem_caches); cp != NULL;
1299 		    cp = list_prev(&kmem_caches, cp)) {
1300 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1301 				break;
1302 		}
1303 	}
1304 
1305 	if (sp == NULL) {
1306 		cp = NULL;
1307 		error = KMERR_BADADDR;
1308 	} else {
1309 		if (cp != cparg)
1310 			error = KMERR_BADCACHE;
1311 		else
1312 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1313 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1314 		if (buf != bufarg)
1315 			error = KMERR_BADBASE;
1316 		if (cp->cache_flags & KMF_BUFTAG)
1317 			btp = KMEM_BUFTAG(cp, buf);
1318 		if (cp->cache_flags & KMF_HASH) {
1319 			mutex_enter(&cp->cache_lock);
1320 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1321 				if (bcp->bc_addr == buf)
1322 					break;
1323 			mutex_exit(&cp->cache_lock);
1324 			if (bcp == NULL && btp != NULL)
1325 				bcp = btp->bt_bufctl;
1326 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1327 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1328 			    bcp->bc_addr != buf) {
1329 				error = KMERR_BADBUFCTL;
1330 				bcp = NULL;
1331 			}
1332 		}
1333 	}
1334 
1335 	kmem_panic_info.kmp_error = error;
1336 	kmem_panic_info.kmp_buffer = bufarg;
1337 	kmem_panic_info.kmp_realbuf = buf;
1338 	kmem_panic_info.kmp_cache = cparg;
1339 	kmem_panic_info.kmp_realcache = cp;
1340 	kmem_panic_info.kmp_slab = sp;
1341 	kmem_panic_info.kmp_bufctl = bcp;
1342 
1343 	printf("kernel memory allocator: ");
1344 
1345 	switch (error) {
1346 
1347 	case KMERR_MODIFIED:
1348 		printf("buffer modified after being freed\n");
1349 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1350 		if (off == NULL)	/* shouldn't happen */
1351 			off = buf;
1352 		printf("modification occurred at offset 0x%lx "
1353 		    "(0x%llx replaced by 0x%llx)\n",
1354 		    (uintptr_t)off - (uintptr_t)buf,
1355 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1356 		break;
1357 
1358 	case KMERR_REDZONE:
1359 		printf("redzone violation: write past end of buffer\n");
1360 		break;
1361 
1362 	case KMERR_BADADDR:
1363 		printf("invalid free: buffer not in cache\n");
1364 		break;
1365 
1366 	case KMERR_DUPFREE:
1367 		printf("duplicate free: buffer freed twice\n");
1368 		break;
1369 
1370 	case KMERR_BADBUFTAG:
1371 		printf("boundary tag corrupted\n");
1372 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1373 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1374 		    KMEM_BUFTAG_FREE);
1375 		break;
1376 
1377 	case KMERR_BADBUFCTL:
1378 		printf("bufctl corrupted\n");
1379 		break;
1380 
1381 	case KMERR_BADCACHE:
1382 		printf("buffer freed to wrong cache\n");
1383 		printf("buffer was allocated from %s,\n", cp->cache_name);
1384 		printf("caller attempting free to %s.\n", cparg->cache_name);
1385 		break;
1386 
1387 	case KMERR_BADSIZE:
1388 		printf("bad free: free size (%u) != alloc size (%u)\n",
1389 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1390 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1391 		break;
1392 
1393 	case KMERR_BADBASE:
1394 		printf("bad free: free address (%p) != alloc address (%p)\n",
1395 		    bufarg, buf);
1396 		break;
1397 	}
1398 
1399 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1400 	    bufarg, (void *)bcp, cparg->cache_name);
1401 
1402 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1403 	    error != KMERR_BADBUFCTL) {
1404 		int d;
1405 		timestruc_t ts;
1406 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1407 
1408 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1409 		printf("previous transaction on buffer %p:\n", buf);
1410 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1411 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1412 		    (void *)sp, cp->cache_name);
1413 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1414 			ulong_t off;
1415 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1416 			printf("%s+%lx\n", sym ? sym : "?", off);
1417 		}
1418 	}
1419 	if (kmem_panic > 0)
1420 		panic("kernel heap corruption detected");
1421 	if (kmem_panic == 0)
1422 		debug_enter(NULL);
1423 	kmem_logging = 1;	/* resume logging */
1424 }
1425 
1426 static kmem_log_header_t *
1427 kmem_log_init(size_t logsize)
1428 {
1429 	kmem_log_header_t *lhp;
1430 	int nchunks = 4 * max_ncpus;
1431 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1432 	int i;
1433 
1434 	/*
1435 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1436 	 * to prevent false sharing of cache lines.
1437 	 */
1438 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1439 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1440 	    NULL, NULL, VM_SLEEP);
1441 	bzero(lhp, lhsize);
1442 
1443 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1444 	lhp->lh_nchunks = nchunks;
1445 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1446 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1447 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1448 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1449 	    nchunks * sizeof (int), VM_SLEEP);
1450 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1451 
1452 	for (i = 0; i < max_ncpus; i++) {
1453 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1454 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1455 		clhp->clh_chunk = i;
1456 	}
1457 
1458 	for (i = max_ncpus; i < nchunks; i++)
1459 		lhp->lh_free[i] = i;
1460 
1461 	lhp->lh_head = max_ncpus;
1462 	lhp->lh_tail = 0;
1463 
1464 	return (lhp);
1465 }
1466 
1467 static void *
1468 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1469 {
1470 	void *logspace;
1471 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1472 
1473 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1474 		return (NULL);
1475 
1476 	mutex_enter(&clhp->clh_lock);
1477 	clhp->clh_hits++;
1478 	if (size > clhp->clh_avail) {
1479 		mutex_enter(&lhp->lh_lock);
1480 		lhp->lh_hits++;
1481 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1482 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1483 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1484 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1485 		clhp->clh_current = lhp->lh_base +
1486 		    clhp->clh_chunk * lhp->lh_chunksize;
1487 		clhp->clh_avail = lhp->lh_chunksize;
1488 		if (size > lhp->lh_chunksize)
1489 			size = lhp->lh_chunksize;
1490 		mutex_exit(&lhp->lh_lock);
1491 	}
1492 	logspace = clhp->clh_current;
1493 	clhp->clh_current += size;
1494 	clhp->clh_avail -= size;
1495 	bcopy(data, logspace, size);
1496 	mutex_exit(&clhp->clh_lock);
1497 	return (logspace);
1498 }
1499 
1500 #define	KMEM_AUDIT(lp, cp, bcp)						\
1501 {									\
1502 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1503 	_bcp->bc_timestamp = gethrtime();				\
1504 	_bcp->bc_thread = curthread;					\
1505 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1506 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1507 }
1508 
1509 static void
1510 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1511 	kmem_slab_t *sp, void *addr)
1512 {
1513 	kmem_bufctl_audit_t bca;
1514 
1515 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1516 	bca.bc_addr = addr;
1517 	bca.bc_slab = sp;
1518 	bca.bc_cache = cp;
1519 	KMEM_AUDIT(lp, cp, &bca);
1520 }
1521 
1522 /*
1523  * Create a new slab for cache cp.
1524  */
1525 static kmem_slab_t *
1526 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1527 {
1528 	size_t slabsize = cp->cache_slabsize;
1529 	size_t chunksize = cp->cache_chunksize;
1530 	int cache_flags = cp->cache_flags;
1531 	size_t color, chunks;
1532 	char *buf, *slab;
1533 	kmem_slab_t *sp;
1534 	kmem_bufctl_t *bcp;
1535 	vmem_t *vmp = cp->cache_arena;
1536 
1537 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1538 
1539 	color = cp->cache_color + cp->cache_align;
1540 	if (color > cp->cache_maxcolor)
1541 		color = cp->cache_mincolor;
1542 	cp->cache_color = color;
1543 
1544 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1545 
1546 	if (slab == NULL)
1547 		goto vmem_alloc_failure;
1548 
1549 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1550 
1551 	/*
1552 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1553 	 * consolidator depends (for correctness) on slabs being initialized
1554 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1555 	 * clients to distinguish uninitialized memory from known objects).
1556 	 */
1557 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1558 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1559 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1560 
1561 	if (cache_flags & KMF_HASH) {
1562 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1563 			goto slab_alloc_failure;
1564 		chunks = (slabsize - color) / chunksize;
1565 	} else {
1566 		sp = KMEM_SLAB(cp, slab);
1567 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1568 	}
1569 
1570 	sp->slab_cache	= cp;
1571 	sp->slab_head	= NULL;
1572 	sp->slab_refcnt	= 0;
1573 	sp->slab_base	= buf = slab + color;
1574 	sp->slab_chunks	= chunks;
1575 	sp->slab_stuck_offset = (uint32_t)-1;
1576 	sp->slab_later_count = 0;
1577 	sp->slab_flags = 0;
1578 
1579 	ASSERT(chunks > 0);
1580 	while (chunks-- != 0) {
1581 		if (cache_flags & KMF_HASH) {
1582 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1583 			if (bcp == NULL)
1584 				goto bufctl_alloc_failure;
1585 			if (cache_flags & KMF_AUDIT) {
1586 				kmem_bufctl_audit_t *bcap =
1587 				    (kmem_bufctl_audit_t *)bcp;
1588 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1589 				bcap->bc_cache = cp;
1590 			}
1591 			bcp->bc_addr = buf;
1592 			bcp->bc_slab = sp;
1593 		} else {
1594 			bcp = KMEM_BUFCTL(cp, buf);
1595 		}
1596 		if (cache_flags & KMF_BUFTAG) {
1597 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1598 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1599 			btp->bt_bufctl = bcp;
1600 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1601 			if (cache_flags & KMF_DEADBEEF) {
1602 				copy_pattern(KMEM_FREE_PATTERN, buf,
1603 				    cp->cache_verify);
1604 			}
1605 		}
1606 		bcp->bc_next = sp->slab_head;
1607 		sp->slab_head = bcp;
1608 		buf += chunksize;
1609 	}
1610 
1611 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1612 
1613 	return (sp);
1614 
1615 bufctl_alloc_failure:
1616 
1617 	while ((bcp = sp->slab_head) != NULL) {
1618 		sp->slab_head = bcp->bc_next;
1619 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1620 	}
1621 	kmem_cache_free(kmem_slab_cache, sp);
1622 
1623 slab_alloc_failure:
1624 
1625 	vmem_free(vmp, slab, slabsize);
1626 
1627 vmem_alloc_failure:
1628 
1629 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1630 	atomic_inc_64(&cp->cache_alloc_fail);
1631 
1632 	return (NULL);
1633 }
1634 
1635 /*
1636  * Destroy a slab.
1637  */
1638 static void
1639 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1640 {
1641 	vmem_t *vmp = cp->cache_arena;
1642 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1643 
1644 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1645 	ASSERT(sp->slab_refcnt == 0);
1646 
1647 	if (cp->cache_flags & KMF_HASH) {
1648 		kmem_bufctl_t *bcp;
1649 		while ((bcp = sp->slab_head) != NULL) {
1650 			sp->slab_head = bcp->bc_next;
1651 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1652 		}
1653 		kmem_cache_free(kmem_slab_cache, sp);
1654 	}
1655 	vmem_free(vmp, slab, cp->cache_slabsize);
1656 }
1657 
1658 static void *
1659 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1660 {
1661 	kmem_bufctl_t *bcp, **hash_bucket;
1662 	void *buf;
1663 	boolean_t new_slab = (sp->slab_refcnt == 0);
1664 
1665 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1666 	/*
1667 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1668 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1669 	 * slab is newly created.
1670 	 */
1671 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1672 	    (sp == avl_first(&cp->cache_partial_slabs))));
1673 	ASSERT(sp->slab_cache == cp);
1674 
1675 	cp->cache_slab_alloc++;
1676 	cp->cache_bufslab--;
1677 	sp->slab_refcnt++;
1678 
1679 	bcp = sp->slab_head;
1680 	sp->slab_head = bcp->bc_next;
1681 
1682 	if (cp->cache_flags & KMF_HASH) {
1683 		/*
1684 		 * Add buffer to allocated-address hash table.
1685 		 */
1686 		buf = bcp->bc_addr;
1687 		hash_bucket = KMEM_HASH(cp, buf);
1688 		bcp->bc_next = *hash_bucket;
1689 		*hash_bucket = bcp;
1690 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1691 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1692 		}
1693 	} else {
1694 		buf = KMEM_BUF(cp, bcp);
1695 	}
1696 
1697 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1698 
1699 	if (sp->slab_head == NULL) {
1700 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1701 		if (new_slab) {
1702 			ASSERT(sp->slab_chunks == 1);
1703 		} else {
1704 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1705 			avl_remove(&cp->cache_partial_slabs, sp);
1706 			sp->slab_later_count = 0; /* clear history */
1707 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1708 			sp->slab_stuck_offset = (uint32_t)-1;
1709 		}
1710 		list_insert_head(&cp->cache_complete_slabs, sp);
1711 		cp->cache_complete_slab_count++;
1712 		return (buf);
1713 	}
1714 
1715 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1716 	/*
1717 	 * Peek to see if the magazine layer is enabled before
1718 	 * we prefill.  We're not holding the cpu cache lock,
1719 	 * so the peek could be wrong, but there's no harm in it.
1720 	 */
1721 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1722 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1723 		kmem_slab_prefill(cp, sp);
1724 		return (buf);
1725 	}
1726 
1727 	if (new_slab) {
1728 		avl_add(&cp->cache_partial_slabs, sp);
1729 		return (buf);
1730 	}
1731 
1732 	/*
1733 	 * The slab is now more allocated than it was, so the
1734 	 * order remains unchanged.
1735 	 */
1736 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1737 	return (buf);
1738 }
1739 
1740 /*
1741  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1742  */
1743 static void *
1744 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1745 {
1746 	kmem_slab_t *sp;
1747 	void *buf;
1748 	boolean_t test_destructor;
1749 
1750 	mutex_enter(&cp->cache_lock);
1751 	test_destructor = (cp->cache_slab_alloc == 0);
1752 	sp = avl_first(&cp->cache_partial_slabs);
1753 	if (sp == NULL) {
1754 		ASSERT(cp->cache_bufslab == 0);
1755 
1756 		/*
1757 		 * The freelist is empty.  Create a new slab.
1758 		 */
1759 		mutex_exit(&cp->cache_lock);
1760 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1761 			return (NULL);
1762 		}
1763 		mutex_enter(&cp->cache_lock);
1764 		cp->cache_slab_create++;
1765 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1766 			cp->cache_bufmax = cp->cache_buftotal;
1767 		cp->cache_bufslab += sp->slab_chunks;
1768 	}
1769 
1770 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1771 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1772 	    (cp->cache_complete_slab_count +
1773 	    avl_numnodes(&cp->cache_partial_slabs) +
1774 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1775 	mutex_exit(&cp->cache_lock);
1776 
1777 	if (test_destructor && cp->cache_destructor != NULL) {
1778 		/*
1779 		 * On the first kmem_slab_alloc(), assert that it is valid to
1780 		 * call the destructor on a newly constructed object without any
1781 		 * client involvement.
1782 		 */
1783 		if ((cp->cache_constructor == NULL) ||
1784 		    cp->cache_constructor(buf, cp->cache_private,
1785 		    kmflag) == 0) {
1786 			cp->cache_destructor(buf, cp->cache_private);
1787 		}
1788 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1789 		    cp->cache_bufsize);
1790 		if (cp->cache_flags & KMF_DEADBEEF) {
1791 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1792 		}
1793 	}
1794 
1795 	return (buf);
1796 }
1797 
1798 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1799 
1800 /*
1801  * Free a raw (unconstructed) buffer to cp's slab layer.
1802  */
1803 static void
1804 kmem_slab_free(kmem_cache_t *cp, void *buf)
1805 {
1806 	kmem_slab_t *sp;
1807 	kmem_bufctl_t *bcp, **prev_bcpp;
1808 
1809 	ASSERT(buf != NULL);
1810 
1811 	mutex_enter(&cp->cache_lock);
1812 	cp->cache_slab_free++;
1813 
1814 	if (cp->cache_flags & KMF_HASH) {
1815 		/*
1816 		 * Look up buffer in allocated-address hash table.
1817 		 */
1818 		prev_bcpp = KMEM_HASH(cp, buf);
1819 		while ((bcp = *prev_bcpp) != NULL) {
1820 			if (bcp->bc_addr == buf) {
1821 				*prev_bcpp = bcp->bc_next;
1822 				sp = bcp->bc_slab;
1823 				break;
1824 			}
1825 			cp->cache_lookup_depth++;
1826 			prev_bcpp = &bcp->bc_next;
1827 		}
1828 	} else {
1829 		bcp = KMEM_BUFCTL(cp, buf);
1830 		sp = KMEM_SLAB(cp, buf);
1831 	}
1832 
1833 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1834 		mutex_exit(&cp->cache_lock);
1835 		kmem_error(KMERR_BADADDR, cp, buf);
1836 		return;
1837 	}
1838 
1839 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1840 		/*
1841 		 * If this is the buffer that prevented the consolidator from
1842 		 * clearing the slab, we can reset the slab flags now that the
1843 		 * buffer is freed. (It makes sense to do this in
1844 		 * kmem_cache_free(), where the client gives up ownership of the
1845 		 * buffer, but on the hot path the test is too expensive.)
1846 		 */
1847 		kmem_slab_move_yes(cp, sp, buf);
1848 	}
1849 
1850 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1851 		if (cp->cache_flags & KMF_CONTENTS)
1852 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1853 			    kmem_log_enter(kmem_content_log, buf,
1854 			    cp->cache_contents);
1855 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1856 	}
1857 
1858 	bcp->bc_next = sp->slab_head;
1859 	sp->slab_head = bcp;
1860 
1861 	cp->cache_bufslab++;
1862 	ASSERT(sp->slab_refcnt >= 1);
1863 
1864 	if (--sp->slab_refcnt == 0) {
1865 		/*
1866 		 * There are no outstanding allocations from this slab,
1867 		 * so we can reclaim the memory.
1868 		 */
1869 		if (sp->slab_chunks == 1) {
1870 			list_remove(&cp->cache_complete_slabs, sp);
1871 			cp->cache_complete_slab_count--;
1872 		} else {
1873 			avl_remove(&cp->cache_partial_slabs, sp);
1874 		}
1875 
1876 		cp->cache_buftotal -= sp->slab_chunks;
1877 		cp->cache_bufslab -= sp->slab_chunks;
1878 		/*
1879 		 * Defer releasing the slab to the virtual memory subsystem
1880 		 * while there is a pending move callback, since we guarantee
1881 		 * that buffers passed to the move callback have only been
1882 		 * touched by kmem or by the client itself. Since the memory
1883 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1884 		 * set at least one of the two lowest order bits, the client can
1885 		 * test those bits in the move callback to determine whether or
1886 		 * not it knows about the buffer (assuming that the client also
1887 		 * sets one of those low order bits whenever it frees a buffer).
1888 		 */
1889 		if (cp->cache_defrag == NULL ||
1890 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1891 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1892 			cp->cache_slab_destroy++;
1893 			mutex_exit(&cp->cache_lock);
1894 			kmem_slab_destroy(cp, sp);
1895 		} else {
1896 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1897 			/*
1898 			 * Slabs are inserted at both ends of the deadlist to
1899 			 * distinguish between slabs freed while move callbacks
1900 			 * are pending (list head) and a slab freed while the
1901 			 * lock is dropped in kmem_move_buffers() (list tail) so
1902 			 * that in both cases slab_destroy() is called from the
1903 			 * right context.
1904 			 */
1905 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1906 				list_insert_tail(deadlist, sp);
1907 			} else {
1908 				list_insert_head(deadlist, sp);
1909 			}
1910 			cp->cache_defrag->kmd_deadcount++;
1911 			mutex_exit(&cp->cache_lock);
1912 		}
1913 		return;
1914 	}
1915 
1916 	if (bcp->bc_next == NULL) {
1917 		/* Transition the slab from completely allocated to partial. */
1918 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1919 		ASSERT(sp->slab_chunks > 1);
1920 		list_remove(&cp->cache_complete_slabs, sp);
1921 		cp->cache_complete_slab_count--;
1922 		avl_add(&cp->cache_partial_slabs, sp);
1923 	} else {
1924 #ifdef	DEBUG
1925 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1926 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1927 		} else {
1928 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1929 		}
1930 #else
1931 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1932 #endif
1933 	}
1934 
1935 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1936 	    (cp->cache_complete_slab_count +
1937 	    avl_numnodes(&cp->cache_partial_slabs) +
1938 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1939 	mutex_exit(&cp->cache_lock);
1940 }
1941 
1942 /*
1943  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1944  */
1945 static int
1946 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1947     caddr_t caller)
1948 {
1949 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1950 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1951 	uint32_t mtbf;
1952 
1953 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1954 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1955 		return (-1);
1956 	}
1957 
1958 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1959 
1960 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1961 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1962 		return (-1);
1963 	}
1964 
1965 	if (cp->cache_flags & KMF_DEADBEEF) {
1966 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1967 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1968 				kmem_error(KMERR_MODIFIED, cp, buf);
1969 				return (-1);
1970 			}
1971 			if (cp->cache_constructor != NULL)
1972 				*(uint64_t *)buf = btp->bt_redzone;
1973 			else
1974 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1975 		} else {
1976 			construct = 1;
1977 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1978 			    KMEM_UNINITIALIZED_PATTERN, buf,
1979 			    cp->cache_verify)) {
1980 				kmem_error(KMERR_MODIFIED, cp, buf);
1981 				return (-1);
1982 			}
1983 		}
1984 	}
1985 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1986 
1987 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1988 	    gethrtime() % mtbf == 0 &&
1989 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1990 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1991 		if (!construct && cp->cache_destructor != NULL)
1992 			cp->cache_destructor(buf, cp->cache_private);
1993 	} else {
1994 		mtbf = 0;
1995 	}
1996 
1997 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1998 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1999 		atomic_inc_64(&cp->cache_alloc_fail);
2000 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2001 		if (cp->cache_flags & KMF_DEADBEEF)
2002 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2003 		kmem_slab_free(cp, buf);
2004 		return (1);
2005 	}
2006 
2007 	if (cp->cache_flags & KMF_AUDIT) {
2008 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2009 	}
2010 
2011 	if ((cp->cache_flags & KMF_LITE) &&
2012 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2013 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2014 	}
2015 
2016 	return (0);
2017 }
2018 
2019 static int
2020 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2021 {
2022 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2023 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2024 	kmem_slab_t *sp;
2025 
2026 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2027 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2028 			kmem_error(KMERR_DUPFREE, cp, buf);
2029 			return (-1);
2030 		}
2031 		sp = kmem_findslab(cp, buf);
2032 		if (sp == NULL || sp->slab_cache != cp)
2033 			kmem_error(KMERR_BADADDR, cp, buf);
2034 		else
2035 			kmem_error(KMERR_REDZONE, cp, buf);
2036 		return (-1);
2037 	}
2038 
2039 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2040 
2041 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2042 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2043 		return (-1);
2044 	}
2045 
2046 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2047 		kmem_error(KMERR_REDZONE, cp, buf);
2048 		return (-1);
2049 	}
2050 
2051 	if (cp->cache_flags & KMF_AUDIT) {
2052 		if (cp->cache_flags & KMF_CONTENTS)
2053 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2054 			    buf, cp->cache_contents);
2055 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2056 	}
2057 
2058 	if ((cp->cache_flags & KMF_LITE) &&
2059 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2060 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2061 	}
2062 
2063 	if (cp->cache_flags & KMF_DEADBEEF) {
2064 		if (cp->cache_flags & KMF_LITE)
2065 			btp->bt_redzone = *(uint64_t *)buf;
2066 		else if (cp->cache_destructor != NULL)
2067 			cp->cache_destructor(buf, cp->cache_private);
2068 
2069 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2070 	}
2071 
2072 	return (0);
2073 }
2074 
2075 /*
2076  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2077  */
2078 static void
2079 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2080 {
2081 	int round;
2082 
2083 	ASSERT(!list_link_active(&cp->cache_link) ||
2084 	    taskq_member(kmem_taskq, curthread));
2085 
2086 	for (round = 0; round < nrounds; round++) {
2087 		void *buf = mp->mag_round[round];
2088 
2089 		if (cp->cache_flags & KMF_DEADBEEF) {
2090 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2091 			    cp->cache_verify) != NULL) {
2092 				kmem_error(KMERR_MODIFIED, cp, buf);
2093 				continue;
2094 			}
2095 			if ((cp->cache_flags & KMF_LITE) &&
2096 			    cp->cache_destructor != NULL) {
2097 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2098 				*(uint64_t *)buf = btp->bt_redzone;
2099 				cp->cache_destructor(buf, cp->cache_private);
2100 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2101 			}
2102 		} else if (cp->cache_destructor != NULL) {
2103 			cp->cache_destructor(buf, cp->cache_private);
2104 		}
2105 
2106 		kmem_slab_free(cp, buf);
2107 	}
2108 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2109 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2110 }
2111 
2112 /*
2113  * Allocate a magazine from the depot.
2114  */
2115 static kmem_magazine_t *
2116 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2117 {
2118 	kmem_magazine_t *mp;
2119 
2120 	/*
2121 	 * If we can't get the depot lock without contention,
2122 	 * update our contention count.  We use the depot
2123 	 * contention rate to determine whether we need to
2124 	 * increase the magazine size for better scalability.
2125 	 */
2126 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2127 		mutex_enter(&cp->cache_depot_lock);
2128 		cp->cache_depot_contention++;
2129 	}
2130 
2131 	if ((mp = mlp->ml_list) != NULL) {
2132 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2133 		mlp->ml_list = mp->mag_next;
2134 		if (--mlp->ml_total < mlp->ml_min)
2135 			mlp->ml_min = mlp->ml_total;
2136 		mlp->ml_alloc++;
2137 	}
2138 
2139 	mutex_exit(&cp->cache_depot_lock);
2140 
2141 	return (mp);
2142 }
2143 
2144 /*
2145  * Free a magazine to the depot.
2146  */
2147 static void
2148 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2149 {
2150 	mutex_enter(&cp->cache_depot_lock);
2151 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2152 	mp->mag_next = mlp->ml_list;
2153 	mlp->ml_list = mp;
2154 	mlp->ml_total++;
2155 	mutex_exit(&cp->cache_depot_lock);
2156 }
2157 
2158 /*
2159  * Update the working set statistics for cp's depot.
2160  */
2161 static void
2162 kmem_depot_ws_update(kmem_cache_t *cp)
2163 {
2164 	mutex_enter(&cp->cache_depot_lock);
2165 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2166 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2167 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2168 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2169 	mutex_exit(&cp->cache_depot_lock);
2170 }
2171 
2172 /*
2173  * Set the working set statistics for cp's depot to zero.  (Everything is
2174  * eligible for reaping.)
2175  */
2176 static void
2177 kmem_depot_ws_zero(kmem_cache_t *cp)
2178 {
2179 	mutex_enter(&cp->cache_depot_lock);
2180 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2181 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2182 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2183 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2184 	mutex_exit(&cp->cache_depot_lock);
2185 }
2186 
2187 /*
2188  * Reap all magazines that have fallen out of the depot's working set.
2189  */
2190 static void
2191 kmem_depot_ws_reap(kmem_cache_t *cp)
2192 {
2193 	long reap;
2194 	kmem_magazine_t *mp;
2195 
2196 	ASSERT(!list_link_active(&cp->cache_link) ||
2197 	    taskq_member(kmem_taskq, curthread));
2198 
2199 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2200 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2201 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2202 
2203 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2204 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2205 		kmem_magazine_destroy(cp, mp, 0);
2206 }
2207 
2208 static void
2209 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2210 {
2211 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2212 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2213 	ASSERT(ccp->cc_magsize > 0);
2214 
2215 	ccp->cc_ploaded = ccp->cc_loaded;
2216 	ccp->cc_prounds = ccp->cc_rounds;
2217 	ccp->cc_loaded = mp;
2218 	ccp->cc_rounds = rounds;
2219 }
2220 
2221 /*
2222  * Intercept kmem alloc/free calls during crash dump in order to avoid
2223  * changing kmem state while memory is being saved to the dump device.
2224  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2225  * there are no locks because only one CPU calls kmem during a crash
2226  * dump. To enable this feature, first create the associated vmem
2227  * arena with VMC_DUMPSAFE.
2228  */
2229 static void *kmem_dump_start;	/* start of pre-reserved heap */
2230 static void *kmem_dump_end;	/* end of heap area */
2231 static void *kmem_dump_curr;	/* current free heap pointer */
2232 static size_t kmem_dump_size;	/* size of heap area */
2233 
2234 /* append to each buf created in the pre-reserved heap */
2235 typedef struct kmem_dumpctl {
2236 	void	*kdc_next;	/* cache dump free list linkage */
2237 } kmem_dumpctl_t;
2238 
2239 #define	KMEM_DUMPCTL(cp, buf)	\
2240 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2241 	    sizeof (void *)))
2242 
2243 /* Keep some simple stats. */
2244 #define	KMEM_DUMP_LOGS	(100)
2245 
2246 typedef struct kmem_dump_log {
2247 	kmem_cache_t	*kdl_cache;
2248 	uint_t		kdl_allocs;		/* # of dump allocations */
2249 	uint_t		kdl_frees;		/* # of dump frees */
2250 	uint_t		kdl_alloc_fails;	/* # of allocation failures */
2251 	uint_t		kdl_free_nondump;	/* # of non-dump frees */
2252 	uint_t		kdl_unsafe;		/* cache was used, but unsafe */
2253 } kmem_dump_log_t;
2254 
2255 static kmem_dump_log_t *kmem_dump_log;
2256 static int kmem_dump_log_idx;
2257 
2258 #define	KDI_LOG(cp, stat) {						\
2259 	kmem_dump_log_t *kdl;						\
2260 	if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) {	\
2261 		kdl->stat++;						\
2262 	} else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) {		\
2263 		kdl = &kmem_dump_log[kmem_dump_log_idx++];		\
2264 		kdl->stat++;						\
2265 		kdl->kdl_cache = (cp);					\
2266 		(cp)->cache_dumplog = kdl;				\
2267 	}								\
2268 }
2269 
2270 /* set non zero for full report */
2271 uint_t kmem_dump_verbose = 0;
2272 
2273 /* stats for overize heap */
2274 uint_t kmem_dump_oversize_allocs = 0;
2275 uint_t kmem_dump_oversize_max = 0;
2276 
2277 static void
2278 kmem_dumppr(char **pp, char *e, const char *format, ...)
2279 {
2280 	char *p = *pp;
2281 
2282 	if (p < e) {
2283 		int n;
2284 		va_list ap;
2285 
2286 		va_start(ap, format);
2287 		n = vsnprintf(p, e - p, format, ap);
2288 		va_end(ap);
2289 		*pp = p + n;
2290 	}
2291 }
2292 
2293 /*
2294  * Called when dumpadm(1M) configures dump parameters.
2295  */
2296 void
2297 kmem_dump_init(size_t size)
2298 {
2299 	if (kmem_dump_start != NULL)
2300 		kmem_free(kmem_dump_start, kmem_dump_size);
2301 
2302 	if (kmem_dump_log == NULL)
2303 		kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2304 		    sizeof (kmem_dump_log_t), KM_SLEEP);
2305 
2306 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2307 
2308 	if (kmem_dump_start != NULL) {
2309 		kmem_dump_size = size;
2310 		kmem_dump_curr = kmem_dump_start;
2311 		kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2312 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2313 	} else {
2314 		kmem_dump_size = 0;
2315 		kmem_dump_curr = NULL;
2316 		kmem_dump_end = NULL;
2317 	}
2318 }
2319 
2320 /*
2321  * Set flag for each kmem_cache_t if is safe to use alternate dump
2322  * memory. Called just before panic crash dump starts. Set the flag
2323  * for the calling CPU.
2324  */
2325 void
2326 kmem_dump_begin(void)
2327 {
2328 	ASSERT(panicstr != NULL);
2329 	if (kmem_dump_start != NULL) {
2330 		kmem_cache_t *cp;
2331 
2332 		for (cp = list_head(&kmem_caches); cp != NULL;
2333 		    cp = list_next(&kmem_caches, cp)) {
2334 			kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2335 
2336 			if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2337 				cp->cache_flags |= KMF_DUMPDIVERT;
2338 				ccp->cc_flags |= KMF_DUMPDIVERT;
2339 				ccp->cc_dump_rounds = ccp->cc_rounds;
2340 				ccp->cc_dump_prounds = ccp->cc_prounds;
2341 				ccp->cc_rounds = ccp->cc_prounds = -1;
2342 			} else {
2343 				cp->cache_flags |= KMF_DUMPUNSAFE;
2344 				ccp->cc_flags |= KMF_DUMPUNSAFE;
2345 			}
2346 		}
2347 	}
2348 }
2349 
2350 /*
2351  * finished dump intercept
2352  * print any warnings on the console
2353  * return verbose information to dumpsys() in the given buffer
2354  */
2355 size_t
2356 kmem_dump_finish(char *buf, size_t size)
2357 {
2358 	int kdi_idx;
2359 	int kdi_end = kmem_dump_log_idx;
2360 	int percent = 0;
2361 	int header = 0;
2362 	int warn = 0;
2363 	size_t used;
2364 	kmem_cache_t *cp;
2365 	kmem_dump_log_t *kdl;
2366 	char *e = buf + size;
2367 	char *p = buf;
2368 
2369 	if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2370 		return (0);
2371 
2372 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2373 	percent = (used * 100) / kmem_dump_size;
2374 
2375 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2376 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2377 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2378 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2379 	    kmem_dump_oversize_allocs);
2380 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2381 	    kmem_dump_oversize_max);
2382 
2383 	for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2384 		kdl = &kmem_dump_log[kdi_idx];
2385 		cp = kdl->kdl_cache;
2386 		if (cp == NULL)
2387 			break;
2388 		if (kdl->kdl_alloc_fails)
2389 			++warn;
2390 		if (header == 0) {
2391 			kmem_dumppr(&p, e,
2392 			    "Cache Name,Allocs,Frees,Alloc Fails,"
2393 			    "Nondump Frees,Unsafe Allocs/Frees\n");
2394 			header = 1;
2395 		}
2396 		kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2397 		    cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2398 		    kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2399 		    kdl->kdl_unsafe);
2400 	}
2401 
2402 	/* return buffer size used */
2403 	if (p < e)
2404 		bzero(p, e - p);
2405 	return (p - buf);
2406 }
2407 
2408 /*
2409  * Allocate a constructed object from alternate dump memory.
2410  */
2411 void *
2412 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2413 {
2414 	void *buf;
2415 	void *curr;
2416 	char *bufend;
2417 
2418 	/* return a constructed object */
2419 	if ((buf = cp->cache_dumpfreelist) != NULL) {
2420 		cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2421 		KDI_LOG(cp, kdl_allocs);
2422 		return (buf);
2423 	}
2424 
2425 	/* create a new constructed object */
2426 	curr = kmem_dump_curr;
2427 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2428 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2429 
2430 	/* hat layer objects cannot cross a page boundary */
2431 	if (cp->cache_align < PAGESIZE) {
2432 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2433 		if (bufend > page) {
2434 			bufend += page - (char *)buf;
2435 			buf = (void *)page;
2436 		}
2437 	}
2438 
2439 	/* fall back to normal alloc if reserved area is used up */
2440 	if (bufend > (char *)kmem_dump_end) {
2441 		kmem_dump_curr = kmem_dump_end;
2442 		KDI_LOG(cp, kdl_alloc_fails);
2443 		return (NULL);
2444 	}
2445 
2446 	/*
2447 	 * Must advance curr pointer before calling a constructor that
2448 	 * may also allocate memory.
2449 	 */
2450 	kmem_dump_curr = bufend;
2451 
2452 	/* run constructor */
2453 	if (cp->cache_constructor != NULL &&
2454 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2455 	    != 0) {
2456 #ifdef DEBUG
2457 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2458 		    cp->cache_name, (void *)cp);
2459 #endif
2460 		/* reset curr pointer iff no allocs were done */
2461 		if (kmem_dump_curr == bufend)
2462 			kmem_dump_curr = curr;
2463 
2464 		/* fall back to normal alloc if the constructor fails */
2465 		KDI_LOG(cp, kdl_alloc_fails);
2466 		return (NULL);
2467 	}
2468 
2469 	KDI_LOG(cp, kdl_allocs);
2470 	return (buf);
2471 }
2472 
2473 /*
2474  * Free a constructed object in alternate dump memory.
2475  */
2476 int
2477 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2478 {
2479 	/* save constructed buffers for next time */
2480 	if ((char *)buf >= (char *)kmem_dump_start &&
2481 	    (char *)buf < (char *)kmem_dump_end) {
2482 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2483 		cp->cache_dumpfreelist = buf;
2484 		KDI_LOG(cp, kdl_frees);
2485 		return (0);
2486 	}
2487 
2488 	/* count all non-dump buf frees */
2489 	KDI_LOG(cp, kdl_free_nondump);
2490 
2491 	/* just drop buffers that were allocated before dump started */
2492 	if (kmem_dump_curr < kmem_dump_end)
2493 		return (0);
2494 
2495 	/* fall back to normal free if reserved area is used up */
2496 	return (1);
2497 }
2498 
2499 /*
2500  * Allocate a constructed object from cache cp.
2501  */
2502 void *
2503 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2504 {
2505 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2506 	kmem_magazine_t *fmp;
2507 	void *buf;
2508 
2509 	mutex_enter(&ccp->cc_lock);
2510 	for (;;) {
2511 		/*
2512 		 * If there's an object available in the current CPU's
2513 		 * loaded magazine, just take it and return.
2514 		 */
2515 		if (ccp->cc_rounds > 0) {
2516 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2517 			ccp->cc_alloc++;
2518 			mutex_exit(&ccp->cc_lock);
2519 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2520 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2521 					ASSERT(!(ccp->cc_flags &
2522 					    KMF_DUMPDIVERT));
2523 					KDI_LOG(cp, kdl_unsafe);
2524 				}
2525 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2526 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2527 				    caller()) != 0) {
2528 					if (kmflag & KM_NOSLEEP)
2529 						return (NULL);
2530 					mutex_enter(&ccp->cc_lock);
2531 					continue;
2532 				}
2533 			}
2534 			return (buf);
2535 		}
2536 
2537 		/*
2538 		 * The loaded magazine is empty.  If the previously loaded
2539 		 * magazine was full, exchange them and try again.
2540 		 */
2541 		if (ccp->cc_prounds > 0) {
2542 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2543 			continue;
2544 		}
2545 
2546 		/*
2547 		 * Return an alternate buffer at dump time to preserve
2548 		 * the heap.
2549 		 */
2550 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2551 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2552 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2553 				/* log it so that we can warn about it */
2554 				KDI_LOG(cp, kdl_unsafe);
2555 			} else {
2556 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2557 				    NULL) {
2558 					mutex_exit(&ccp->cc_lock);
2559 					return (buf);
2560 				}
2561 				break;		/* fall back to slab layer */
2562 			}
2563 		}
2564 
2565 		/*
2566 		 * If the magazine layer is disabled, break out now.
2567 		 */
2568 		if (ccp->cc_magsize == 0)
2569 			break;
2570 
2571 		/*
2572 		 * Try to get a full magazine from the depot.
2573 		 */
2574 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2575 		if (fmp != NULL) {
2576 			if (ccp->cc_ploaded != NULL)
2577 				kmem_depot_free(cp, &cp->cache_empty,
2578 				    ccp->cc_ploaded);
2579 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2580 			continue;
2581 		}
2582 
2583 		/*
2584 		 * There are no full magazines in the depot,
2585 		 * so fall through to the slab layer.
2586 		 */
2587 		break;
2588 	}
2589 	mutex_exit(&ccp->cc_lock);
2590 
2591 	/*
2592 	 * We couldn't allocate a constructed object from the magazine layer,
2593 	 * so get a raw buffer from the slab layer and apply its constructor.
2594 	 */
2595 	buf = kmem_slab_alloc(cp, kmflag);
2596 
2597 	if (buf == NULL)
2598 		return (NULL);
2599 
2600 	if (cp->cache_flags & KMF_BUFTAG) {
2601 		/*
2602 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2603 		 */
2604 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2605 		if (rc != 0) {
2606 			if (kmflag & KM_NOSLEEP)
2607 				return (NULL);
2608 			/*
2609 			 * kmem_cache_alloc_debug() detected corruption
2610 			 * but didn't panic (kmem_panic <= 0). We should not be
2611 			 * here because the constructor failed (indicated by a
2612 			 * return code of 1). Try again.
2613 			 */
2614 			ASSERT(rc == -1);
2615 			return (kmem_cache_alloc(cp, kmflag));
2616 		}
2617 		return (buf);
2618 	}
2619 
2620 	if (cp->cache_constructor != NULL &&
2621 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2622 		atomic_inc_64(&cp->cache_alloc_fail);
2623 		kmem_slab_free(cp, buf);
2624 		return (NULL);
2625 	}
2626 
2627 	return (buf);
2628 }
2629 
2630 /*
2631  * The freed argument tells whether or not kmem_cache_free_debug() has already
2632  * been called so that we can avoid the duplicate free error. For example, a
2633  * buffer on a magazine has already been freed by the client but is still
2634  * constructed.
2635  */
2636 static void
2637 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2638 {
2639 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2640 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2641 			return;
2642 
2643 	/*
2644 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2645 	 * kmem_cache_free_debug() will have already applied the destructor.
2646 	 */
2647 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2648 	    cp->cache_destructor != NULL) {
2649 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2650 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2651 			*(uint64_t *)buf = btp->bt_redzone;
2652 			cp->cache_destructor(buf, cp->cache_private);
2653 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2654 		} else {
2655 			cp->cache_destructor(buf, cp->cache_private);
2656 		}
2657 	}
2658 
2659 	kmem_slab_free(cp, buf);
2660 }
2661 
2662 /*
2663  * Used when there's no room to free a buffer to the per-CPU cache.
2664  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2665  * caller should try freeing to the per-CPU cache again.
2666  * Note that we don't directly install the magazine in the cpu cache,
2667  * since its state may have changed wildly while the lock was dropped.
2668  */
2669 static int
2670 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2671 {
2672 	kmem_magazine_t *emp;
2673 	kmem_magtype_t *mtp;
2674 
2675 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2676 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2677 	    ((uint_t)ccp->cc_rounds == -1)) &&
2678 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2679 	    ((uint_t)ccp->cc_prounds == -1)));
2680 
2681 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2682 	if (emp != NULL) {
2683 		if (ccp->cc_ploaded != NULL)
2684 			kmem_depot_free(cp, &cp->cache_full,
2685 			    ccp->cc_ploaded);
2686 		kmem_cpu_reload(ccp, emp, 0);
2687 		return (1);
2688 	}
2689 	/*
2690 	 * There are no empty magazines in the depot,
2691 	 * so try to allocate a new one.  We must drop all locks
2692 	 * across kmem_cache_alloc() because lower layers may
2693 	 * attempt to allocate from this cache.
2694 	 */
2695 	mtp = cp->cache_magtype;
2696 	mutex_exit(&ccp->cc_lock);
2697 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2698 	mutex_enter(&ccp->cc_lock);
2699 
2700 	if (emp != NULL) {
2701 		/*
2702 		 * We successfully allocated an empty magazine.
2703 		 * However, we had to drop ccp->cc_lock to do it,
2704 		 * so the cache's magazine size may have changed.
2705 		 * If so, free the magazine and try again.
2706 		 */
2707 		if (ccp->cc_magsize != mtp->mt_magsize) {
2708 			mutex_exit(&ccp->cc_lock);
2709 			kmem_cache_free(mtp->mt_cache, emp);
2710 			mutex_enter(&ccp->cc_lock);
2711 			return (1);
2712 		}
2713 
2714 		/*
2715 		 * We got a magazine of the right size.  Add it to
2716 		 * the depot and try the whole dance again.
2717 		 */
2718 		kmem_depot_free(cp, &cp->cache_empty, emp);
2719 		return (1);
2720 	}
2721 
2722 	/*
2723 	 * We couldn't allocate an empty magazine,
2724 	 * so fall through to the slab layer.
2725 	 */
2726 	return (0);
2727 }
2728 
2729 /*
2730  * Free a constructed object to cache cp.
2731  */
2732 void
2733 kmem_cache_free(kmem_cache_t *cp, void *buf)
2734 {
2735 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2736 
2737 	/*
2738 	 * The client must not free either of the buffers passed to the move
2739 	 * callback function.
2740 	 */
2741 	ASSERT(cp->cache_defrag == NULL ||
2742 	    cp->cache_defrag->kmd_thread != curthread ||
2743 	    (buf != cp->cache_defrag->kmd_from_buf &&
2744 	    buf != cp->cache_defrag->kmd_to_buf));
2745 
2746 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2747 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2748 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2749 			/* log it so that we can warn about it */
2750 			KDI_LOG(cp, kdl_unsafe);
2751 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2752 			return;
2753 		}
2754 		if (ccp->cc_flags & KMF_BUFTAG) {
2755 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2756 				return;
2757 		}
2758 	}
2759 
2760 	mutex_enter(&ccp->cc_lock);
2761 	/*
2762 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2763 	 */
2764 	for (;;) {
2765 		/*
2766 		 * If there's a slot available in the current CPU's
2767 		 * loaded magazine, just put the object there and return.
2768 		 */
2769 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2770 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2771 			ccp->cc_free++;
2772 			mutex_exit(&ccp->cc_lock);
2773 			return;
2774 		}
2775 
2776 		/*
2777 		 * The loaded magazine is full.  If the previously loaded
2778 		 * magazine was empty, exchange them and try again.
2779 		 */
2780 		if (ccp->cc_prounds == 0) {
2781 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2782 			continue;
2783 		}
2784 
2785 		/*
2786 		 * If the magazine layer is disabled, break out now.
2787 		 */
2788 		if (ccp->cc_magsize == 0)
2789 			break;
2790 
2791 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2792 			/*
2793 			 * We couldn't free our constructed object to the
2794 			 * magazine layer, so apply its destructor and free it
2795 			 * to the slab layer.
2796 			 */
2797 			break;
2798 		}
2799 	}
2800 	mutex_exit(&ccp->cc_lock);
2801 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2802 }
2803 
2804 static void
2805 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2806 {
2807 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2808 	int cache_flags = cp->cache_flags;
2809 
2810 	kmem_bufctl_t *next, *head;
2811 	size_t nbufs;
2812 
2813 	/*
2814 	 * Completely allocate the newly created slab and put the pre-allocated
2815 	 * buffers in magazines. Any of the buffers that cannot be put in
2816 	 * magazines must be returned to the slab.
2817 	 */
2818 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2819 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2820 	ASSERT(cp->cache_constructor == NULL);
2821 	ASSERT(sp->slab_cache == cp);
2822 	ASSERT(sp->slab_refcnt == 1);
2823 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2824 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2825 
2826 	head = sp->slab_head;
2827 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2828 	sp->slab_head = NULL;
2829 	sp->slab_refcnt += nbufs;
2830 	cp->cache_bufslab -= nbufs;
2831 	cp->cache_slab_alloc += nbufs;
2832 	list_insert_head(&cp->cache_complete_slabs, sp);
2833 	cp->cache_complete_slab_count++;
2834 	mutex_exit(&cp->cache_lock);
2835 	mutex_enter(&ccp->cc_lock);
2836 
2837 	while (head != NULL) {
2838 		void *buf = KMEM_BUF(cp, head);
2839 		/*
2840 		 * If there's a slot available in the current CPU's
2841 		 * loaded magazine, just put the object there and
2842 		 * continue.
2843 		 */
2844 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2845 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2846 			    buf;
2847 			ccp->cc_free++;
2848 			nbufs--;
2849 			head = head->bc_next;
2850 			continue;
2851 		}
2852 
2853 		/*
2854 		 * The loaded magazine is full.  If the previously
2855 		 * loaded magazine was empty, exchange them and try
2856 		 * again.
2857 		 */
2858 		if (ccp->cc_prounds == 0) {
2859 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2860 			    ccp->cc_prounds);
2861 			continue;
2862 		}
2863 
2864 		/*
2865 		 * If the magazine layer is disabled, break out now.
2866 		 */
2867 
2868 		if (ccp->cc_magsize == 0) {
2869 			break;
2870 		}
2871 
2872 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2873 			break;
2874 	}
2875 	mutex_exit(&ccp->cc_lock);
2876 	if (nbufs != 0) {
2877 		ASSERT(head != NULL);
2878 
2879 		/*
2880 		 * If there was a failure, return remaining objects to
2881 		 * the slab
2882 		 */
2883 		while (head != NULL) {
2884 			ASSERT(nbufs != 0);
2885 			next = head->bc_next;
2886 			head->bc_next = NULL;
2887 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2888 			head = next;
2889 			nbufs--;
2890 		}
2891 	}
2892 	ASSERT(head == NULL);
2893 	ASSERT(nbufs == 0);
2894 	mutex_enter(&cp->cache_lock);
2895 }
2896 
2897 void *
2898 kmem_zalloc(size_t size, int kmflag)
2899 {
2900 	size_t index;
2901 	void *buf;
2902 
2903 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2904 		kmem_cache_t *cp = kmem_alloc_table[index];
2905 		buf = kmem_cache_alloc(cp, kmflag);
2906 		if (buf != NULL) {
2907 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2908 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2909 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2910 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2911 
2912 				if (cp->cache_flags & KMF_LITE) {
2913 					KMEM_BUFTAG_LITE_ENTER(btp,
2914 					    kmem_lite_count, caller());
2915 				}
2916 			}
2917 			bzero(buf, size);
2918 		}
2919 	} else {
2920 		buf = kmem_alloc(size, kmflag);
2921 		if (buf != NULL)
2922 			bzero(buf, size);
2923 	}
2924 	return (buf);
2925 }
2926 
2927 void *
2928 kmem_alloc(size_t size, int kmflag)
2929 {
2930 	size_t index;
2931 	kmem_cache_t *cp;
2932 	void *buf;
2933 
2934 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2935 		cp = kmem_alloc_table[index];
2936 		/* fall through to kmem_cache_alloc() */
2937 
2938 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2939 	    kmem_big_alloc_table_max) {
2940 		cp = kmem_big_alloc_table[index];
2941 		/* fall through to kmem_cache_alloc() */
2942 
2943 	} else {
2944 		if (size == 0)
2945 			return (NULL);
2946 
2947 		buf = vmem_alloc(kmem_oversize_arena, size,
2948 		    kmflag & KM_VMFLAGS);
2949 		if (buf == NULL)
2950 			kmem_log_event(kmem_failure_log, NULL, NULL,
2951 			    (void *)size);
2952 		else if (KMEM_DUMP(kmem_slab_cache)) {
2953 			/* stats for dump intercept */
2954 			kmem_dump_oversize_allocs++;
2955 			if (size > kmem_dump_oversize_max)
2956 				kmem_dump_oversize_max = size;
2957 		}
2958 		return (buf);
2959 	}
2960 
2961 	buf = kmem_cache_alloc(cp, kmflag);
2962 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2963 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2964 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2965 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2966 
2967 		if (cp->cache_flags & KMF_LITE) {
2968 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2969 		}
2970 	}
2971 	return (buf);
2972 }
2973 
2974 void
2975 kmem_free(void *buf, size_t size)
2976 {
2977 	size_t index;
2978 	kmem_cache_t *cp;
2979 
2980 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2981 		cp = kmem_alloc_table[index];
2982 		/* fall through to kmem_cache_free() */
2983 
2984 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2985 	    kmem_big_alloc_table_max) {
2986 		cp = kmem_big_alloc_table[index];
2987 		/* fall through to kmem_cache_free() */
2988 
2989 	} else {
2990 		EQUIV(buf == NULL, size == 0);
2991 		if (buf == NULL && size == 0)
2992 			return;
2993 		vmem_free(kmem_oversize_arena, buf, size);
2994 		return;
2995 	}
2996 
2997 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2998 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2999 		uint32_t *ip = (uint32_t *)btp;
3000 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3001 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3002 				kmem_error(KMERR_DUPFREE, cp, buf);
3003 				return;
3004 			}
3005 			if (KMEM_SIZE_VALID(ip[1])) {
3006 				ip[0] = KMEM_SIZE_ENCODE(size);
3007 				kmem_error(KMERR_BADSIZE, cp, buf);
3008 			} else {
3009 				kmem_error(KMERR_REDZONE, cp, buf);
3010 			}
3011 			return;
3012 		}
3013 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3014 			kmem_error(KMERR_REDZONE, cp, buf);
3015 			return;
3016 		}
3017 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
3018 		if (cp->cache_flags & KMF_LITE) {
3019 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3020 			    caller());
3021 		}
3022 	}
3023 	kmem_cache_free(cp, buf);
3024 }
3025 
3026 void *
3027 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3028 {
3029 	size_t realsize = size + vmp->vm_quantum;
3030 	void *addr;
3031 
3032 	/*
3033 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3034 	 * vm_quantum will cause integer wraparound.  Check for this, and
3035 	 * blow off the firewall page in this case.  Note that such a
3036 	 * giant allocation (the entire kernel address space) can never
3037 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3038 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
3039 	 * corresponding check in kmem_firewall_va_free().
3040 	 */
3041 	if (realsize < size)
3042 		realsize = size;
3043 
3044 	/*
3045 	 * While boot still owns resource management, make sure that this
3046 	 * redzone virtual address allocation is properly accounted for in
3047 	 * OBPs "virtual-memory" "available" lists because we're
3048 	 * effectively claiming them for a red zone.  If we don't do this,
3049 	 * the available lists become too fragmented and too large for the
3050 	 * current boot/kernel memory list interface.
3051 	 */
3052 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3053 
3054 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3055 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3056 
3057 	return (addr);
3058 }
3059 
3060 void
3061 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3062 {
3063 	ASSERT((kvseg.s_base == NULL ?
3064 	    va_to_pfn((char *)addr + size) :
3065 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3066 
3067 	vmem_free(vmp, addr, size + vmp->vm_quantum);
3068 }
3069 
3070 /*
3071  * Try to allocate at least `size' bytes of memory without sleeping or
3072  * panicking. Return actual allocated size in `asize'. If allocation failed,
3073  * try final allocation with sleep or panic allowed.
3074  */
3075 void *
3076 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3077 {
3078 	void *p;
3079 
3080 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3081 	do {
3082 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3083 		if (p != NULL)
3084 			return (p);
3085 		*asize += KMEM_ALIGN;
3086 	} while (*asize <= PAGESIZE);
3087 
3088 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3089 	return (kmem_alloc(*asize, kmflag));
3090 }
3091 
3092 /*
3093  * Reclaim all unused memory from a cache.
3094  */
3095 static void
3096 kmem_cache_reap(kmem_cache_t *cp)
3097 {
3098 	ASSERT(taskq_member(kmem_taskq, curthread));
3099 	cp->cache_reap++;
3100 
3101 	/*
3102 	 * Ask the cache's owner to free some memory if possible.
3103 	 * The idea is to handle things like the inode cache, which
3104 	 * typically sits on a bunch of memory that it doesn't truly
3105 	 * *need*.  Reclaim policy is entirely up to the owner; this
3106 	 * callback is just an advisory plea for help.
3107 	 */
3108 	if (cp->cache_reclaim != NULL) {
3109 		long delta;
3110 
3111 		/*
3112 		 * Reclaimed memory should be reapable (not included in the
3113 		 * depot's working set).
3114 		 */
3115 		delta = cp->cache_full.ml_total;
3116 		cp->cache_reclaim(cp->cache_private);
3117 		delta = cp->cache_full.ml_total - delta;
3118 		if (delta > 0) {
3119 			mutex_enter(&cp->cache_depot_lock);
3120 			cp->cache_full.ml_reaplimit += delta;
3121 			cp->cache_full.ml_min += delta;
3122 			mutex_exit(&cp->cache_depot_lock);
3123 		}
3124 	}
3125 
3126 	kmem_depot_ws_reap(cp);
3127 
3128 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3129 		kmem_cache_defrag(cp);
3130 	}
3131 }
3132 
3133 static void
3134 kmem_reap_timeout(void *flag_arg)
3135 {
3136 	uint32_t *flag = (uint32_t *)flag_arg;
3137 
3138 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3139 	*flag = 0;
3140 }
3141 
3142 static void
3143 kmem_reap_done(void *flag)
3144 {
3145 	if (!callout_init_done) {
3146 		/* can't schedule a timeout at this point */
3147 		kmem_reap_timeout(flag);
3148 	} else {
3149 		(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3150 	}
3151 }
3152 
3153 static void
3154 kmem_reap_start(void *flag)
3155 {
3156 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3157 
3158 	if (flag == &kmem_reaping) {
3159 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3160 		/*
3161 		 * if we have segkp under heap, reap segkp cache.
3162 		 */
3163 		if (segkp_fromheap)
3164 			segkp_cache_free();
3165 	}
3166 	else
3167 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3168 
3169 	/*
3170 	 * We use taskq_dispatch() to schedule a timeout to clear
3171 	 * the flag so that kmem_reap() becomes self-throttling:
3172 	 * we won't reap again until the current reap completes *and*
3173 	 * at least kmem_reap_interval ticks have elapsed.
3174 	 */
3175 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3176 		kmem_reap_done(flag);
3177 }
3178 
3179 static void
3180 kmem_reap_common(void *flag_arg)
3181 {
3182 	uint32_t *flag = (uint32_t *)flag_arg;
3183 
3184 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3185 	    atomic_cas_32(flag, 0, 1) != 0)
3186 		return;
3187 
3188 	/*
3189 	 * It may not be kosher to do memory allocation when a reap is called
3190 	 * (for example, if vmem_populate() is in the call chain).  So we
3191 	 * start the reap going with a TQ_NOALLOC dispatch.  If the dispatch
3192 	 * fails, we reset the flag, and the next reap will try again.
3193 	 */
3194 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3195 		*flag = 0;
3196 }
3197 
3198 /*
3199  * Reclaim all unused memory from all caches.  Called from the VM system
3200  * when memory gets tight.
3201  */
3202 void
3203 kmem_reap(void)
3204 {
3205 	kmem_reap_common(&kmem_reaping);
3206 }
3207 
3208 /*
3209  * Reclaim all unused memory from identifier arenas, called when a vmem
3210  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3211  * cannot help with identifier exhaustion, we avoid both a large amount of
3212  * work and unwanted side-effects from reclaim callbacks.
3213  */
3214 void
3215 kmem_reap_idspace(void)
3216 {
3217 	kmem_reap_common(&kmem_reaping_idspace);
3218 }
3219 
3220 /*
3221  * Purge all magazines from a cache and set its magazine limit to zero.
3222  * All calls are serialized by the kmem_taskq lock, except for the final
3223  * call from kmem_cache_destroy().
3224  */
3225 static void
3226 kmem_cache_magazine_purge(kmem_cache_t *cp)
3227 {
3228 	kmem_cpu_cache_t *ccp;
3229 	kmem_magazine_t *mp, *pmp;
3230 	int rounds, prounds, cpu_seqid;
3231 
3232 	ASSERT(!list_link_active(&cp->cache_link) ||
3233 	    taskq_member(kmem_taskq, curthread));
3234 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3235 
3236 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3237 		ccp = &cp->cache_cpu[cpu_seqid];
3238 
3239 		mutex_enter(&ccp->cc_lock);
3240 		mp = ccp->cc_loaded;
3241 		pmp = ccp->cc_ploaded;
3242 		rounds = ccp->cc_rounds;
3243 		prounds = ccp->cc_prounds;
3244 		ccp->cc_loaded = NULL;
3245 		ccp->cc_ploaded = NULL;
3246 		ccp->cc_rounds = -1;
3247 		ccp->cc_prounds = -1;
3248 		ccp->cc_magsize = 0;
3249 		mutex_exit(&ccp->cc_lock);
3250 
3251 		if (mp)
3252 			kmem_magazine_destroy(cp, mp, rounds);
3253 		if (pmp)
3254 			kmem_magazine_destroy(cp, pmp, prounds);
3255 	}
3256 
3257 	kmem_depot_ws_zero(cp);
3258 	kmem_depot_ws_reap(cp);
3259 }
3260 
3261 /*
3262  * Enable per-cpu magazines on a cache.
3263  */
3264 static void
3265 kmem_cache_magazine_enable(kmem_cache_t *cp)
3266 {
3267 	int cpu_seqid;
3268 
3269 	if (cp->cache_flags & KMF_NOMAGAZINE)
3270 		return;
3271 
3272 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3273 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3274 		mutex_enter(&ccp->cc_lock);
3275 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3276 		mutex_exit(&ccp->cc_lock);
3277 	}
3278 
3279 }
3280 
3281 /*
3282  * Reap (almost) everything right now.
3283  */
3284 void
3285 kmem_cache_reap_now(kmem_cache_t *cp)
3286 {
3287 	ASSERT(list_link_active(&cp->cache_link));
3288 
3289 	kmem_depot_ws_zero(cp);
3290 
3291 	(void) taskq_dispatch(kmem_taskq,
3292 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3293 	taskq_wait(kmem_taskq);
3294 }
3295 
3296 /*
3297  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3298  * provide a higher transfer rate with the depot, while smaller magazines
3299  * reduce memory consumption.  Magazine resizing is an expensive operation;
3300  * it should not be done frequently.
3301  *
3302  * Changes to the magazine size are serialized by the kmem_taskq lock.
3303  *
3304  * Note: at present this only grows the magazine size.  It might be useful
3305  * to allow shrinkage too.
3306  */
3307 static void
3308 kmem_cache_magazine_resize(kmem_cache_t *cp)
3309 {
3310 	kmem_magtype_t *mtp = cp->cache_magtype;
3311 
3312 	ASSERT(taskq_member(kmem_taskq, curthread));
3313 
3314 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3315 		kmem_cache_magazine_purge(cp);
3316 		mutex_enter(&cp->cache_depot_lock);
3317 		cp->cache_magtype = ++mtp;
3318 		cp->cache_depot_contention_prev =
3319 		    cp->cache_depot_contention + INT_MAX;
3320 		mutex_exit(&cp->cache_depot_lock);
3321 		kmem_cache_magazine_enable(cp);
3322 	}
3323 }
3324 
3325 /*
3326  * Rescale a cache's hash table, so that the table size is roughly the
3327  * cache size.  We want the average lookup time to be extremely small.
3328  */
3329 static void
3330 kmem_hash_rescale(kmem_cache_t *cp)
3331 {
3332 	kmem_bufctl_t **old_table, **new_table, *bcp;
3333 	size_t old_size, new_size, h;
3334 
3335 	ASSERT(taskq_member(kmem_taskq, curthread));
3336 
3337 	new_size = MAX(KMEM_HASH_INITIAL,
3338 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3339 	old_size = cp->cache_hash_mask + 1;
3340 
3341 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3342 		return;
3343 
3344 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3345 	    VM_NOSLEEP);
3346 	if (new_table == NULL)
3347 		return;
3348 	bzero(new_table, new_size * sizeof (void *));
3349 
3350 	mutex_enter(&cp->cache_lock);
3351 
3352 	old_size = cp->cache_hash_mask + 1;
3353 	old_table = cp->cache_hash_table;
3354 
3355 	cp->cache_hash_mask = new_size - 1;
3356 	cp->cache_hash_table = new_table;
3357 	cp->cache_rescale++;
3358 
3359 	for (h = 0; h < old_size; h++) {
3360 		bcp = old_table[h];
3361 		while (bcp != NULL) {
3362 			void *addr = bcp->bc_addr;
3363 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3364 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3365 			bcp->bc_next = *hash_bucket;
3366 			*hash_bucket = bcp;
3367 			bcp = next_bcp;
3368 		}
3369 	}
3370 
3371 	mutex_exit(&cp->cache_lock);
3372 
3373 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3374 }
3375 
3376 /*
3377  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3378  * update, magazine resizing, and slab consolidation.
3379  */
3380 static void
3381 kmem_cache_update(kmem_cache_t *cp)
3382 {
3383 	int need_hash_rescale = 0;
3384 	int need_magazine_resize = 0;
3385 
3386 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3387 
3388 	/*
3389 	 * If the cache has become much larger or smaller than its hash table,
3390 	 * fire off a request to rescale the hash table.
3391 	 */
3392 	mutex_enter(&cp->cache_lock);
3393 
3394 	if ((cp->cache_flags & KMF_HASH) &&
3395 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3396 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3397 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3398 		need_hash_rescale = 1;
3399 
3400 	mutex_exit(&cp->cache_lock);
3401 
3402 	/*
3403 	 * Update the depot working set statistics.
3404 	 */
3405 	kmem_depot_ws_update(cp);
3406 
3407 	/*
3408 	 * If there's a lot of contention in the depot,
3409 	 * increase the magazine size.
3410 	 */
3411 	mutex_enter(&cp->cache_depot_lock);
3412 
3413 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3414 	    (int)(cp->cache_depot_contention -
3415 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3416 		need_magazine_resize = 1;
3417 
3418 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3419 
3420 	mutex_exit(&cp->cache_depot_lock);
3421 
3422 	if (need_hash_rescale)
3423 		(void) taskq_dispatch(kmem_taskq,
3424 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3425 
3426 	if (need_magazine_resize)
3427 		(void) taskq_dispatch(kmem_taskq,
3428 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3429 
3430 	if (cp->cache_defrag != NULL)
3431 		(void) taskq_dispatch(kmem_taskq,
3432 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3433 }
3434 
3435 static void kmem_update(void *);
3436 
3437 static void
3438 kmem_update_timeout(void *dummy)
3439 {
3440 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3441 }
3442 
3443 static void
3444 kmem_update(void *dummy)
3445 {
3446 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3447 
3448 	/*
3449 	 * We use taskq_dispatch() to reschedule the timeout so that
3450 	 * kmem_update() becomes self-throttling: it won't schedule
3451 	 * new tasks until all previous tasks have completed.
3452 	 */
3453 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3454 		kmem_update_timeout(NULL);
3455 }
3456 
3457 static int
3458 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3459 {
3460 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3461 	kmem_cache_t *cp = ksp->ks_private;
3462 	uint64_t cpu_buf_avail;
3463 	uint64_t buf_avail = 0;
3464 	int cpu_seqid;
3465 	long reap;
3466 
3467 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3468 
3469 	if (rw == KSTAT_WRITE)
3470 		return (EACCES);
3471 
3472 	mutex_enter(&cp->cache_lock);
3473 
3474 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3475 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3476 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3477 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3478 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3479 
3480 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3481 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3482 
3483 		mutex_enter(&ccp->cc_lock);
3484 
3485 		cpu_buf_avail = 0;
3486 		if (ccp->cc_rounds > 0)
3487 			cpu_buf_avail += ccp->cc_rounds;
3488 		if (ccp->cc_prounds > 0)
3489 			cpu_buf_avail += ccp->cc_prounds;
3490 
3491 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3492 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3493 		buf_avail			+= cpu_buf_avail;
3494 
3495 		mutex_exit(&ccp->cc_lock);
3496 	}
3497 
3498 	mutex_enter(&cp->cache_depot_lock);
3499 
3500 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3501 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3502 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3503 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3504 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3505 	kmcp->kmc_magazine_size.value.ui64	=
3506 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3507 	    0 : cp->cache_magtype->mt_magsize;
3508 
3509 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3510 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3511 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3512 
3513 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3514 	reap = MIN(reap, cp->cache_full.ml_total);
3515 
3516 	mutex_exit(&cp->cache_depot_lock);
3517 
3518 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3519 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3520 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3521 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3522 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3523 	buf_avail += cp->cache_bufslab;
3524 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3525 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3526 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3527 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3528 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3529 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3530 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3531 	    cp->cache_hash_mask + 1 : 0;
3532 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3533 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3534 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3535 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3536 
3537 	if (cp->cache_defrag == NULL) {
3538 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3539 		kmcp->kmc_move_yes.value.ui64		= 0;
3540 		kmcp->kmc_move_no.value.ui64		= 0;
3541 		kmcp->kmc_move_later.value.ui64		= 0;
3542 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3543 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3544 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3545 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3546 		kmcp->kmc_defrag.value.ui64		= 0;
3547 		kmcp->