xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision b942e89b)
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  */
24 
25 /*
26  * Kernel memory allocator, as described in the following two papers and a
27  * statement about the consolidator:
28  *
29  * Jeff Bonwick,
30  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
31  * Proceedings of the Summer 1994 Usenix Conference.
32  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
33  *
34  * Jeff Bonwick and Jonathan Adams,
35  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
36  * Arbitrary Resources.
37  * Proceedings of the 2001 Usenix Conference.
38  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
39  *
40  * kmem Slab Consolidator Big Theory Statement:
41  *
42  * 1. Motivation
43  *
44  * As stated in Bonwick94, slabs provide the following advantages over other
45  * allocation structures in terms of memory fragmentation:
46  *
47  *  - Internal fragmentation (per-buffer wasted space) is minimal.
48  *  - Severe external fragmentation (unused buffers on the free list) is
49  *    unlikely.
50  *
51  * Segregating objects by size eliminates one source of external fragmentation,
52  * and according to Bonwick:
53  *
54  *   The other reason that slabs reduce external fragmentation is that all
55  *   objects in a slab are of the same type, so they have the same lifetime
56  *   distribution. The resulting segregation of short-lived and long-lived
57  *   objects at slab granularity reduces the likelihood of an entire page being
58  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
59  *
60  * While unlikely, severe external fragmentation remains possible. Clients that
61  * allocate both short- and long-lived objects from the same cache cannot
62  * anticipate the distribution of long-lived objects within the allocator's slab
63  * implementation. Even a small percentage of long-lived objects distributed
64  * randomly across many slabs can lead to a worst case scenario where the client
65  * frees the majority of its objects and the system gets back almost none of the
66  * slabs. Despite the client doing what it reasonably can to help the system
67  * reclaim memory, the allocator cannot shake free enough slabs because of
68  * lonely allocations stubbornly hanging on. Although the allocator is in a
69  * position to diagnose the fragmentation, there is nothing that the allocator
70  * by itself can do about it. It only takes a single allocated object to prevent
71  * an entire slab from being reclaimed, and any object handed out by
72  * kmem_cache_alloc() is by definition in the client's control. Conversely,
73  * although the client is in a position to move a long-lived object, it has no
74  * way of knowing if the object is causing fragmentation, and if so, where to
75  * move it. A solution necessarily requires further cooperation between the
76  * allocator and the client.
77  *
78  * 2. Move Callback
79  *
80  * The kmem slab consolidator therefore adds a move callback to the
81  * allocator/client interface, improving worst-case external fragmentation in
82  * kmem caches that supply a function to move objects from one memory location
83  * to another. In a situation of low memory kmem attempts to consolidate all of
84  * a cache's slabs at once; otherwise it works slowly to bring external
85  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
86  * thereby helping to avoid a low memory situation in the future.
87  *
88  * The callback has the following signature:
89  *
90  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
91  *
92  * It supplies the kmem client with two addresses: the allocated object that
93  * kmem wants to move and a buffer selected by kmem for the client to use as the
94  * copy destination. The callback is kmem's way of saying "Please get off of
95  * this buffer and use this one instead." kmem knows where it wants to move the
96  * object in order to best reduce fragmentation. All the client needs to know
97  * about the second argument (void *new) is that it is an allocated, constructed
98  * object ready to take the contents of the old object. When the move function
99  * is called, the system is likely to be low on memory, and the new object
100  * spares the client from having to worry about allocating memory for the
101  * requested move. The third argument supplies the size of the object, in case a
102  * single move function handles multiple caches whose objects differ only in
103  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
104  * user argument passed to the constructor, destructor, and reclaim functions is
105  * also passed to the move callback.
106  *
107  * 2.1 Setting the Move Callback
108  *
109  * The client sets the move callback after creating the cache and before
110  * allocating from it:
111  *
112  *	object_cache = kmem_cache_create(...);
113  *      kmem_cache_set_move(object_cache, object_move);
114  *
115  * 2.2 Move Callback Return Values
116  *
117  * Only the client knows about its own data and when is a good time to move it.
118  * The client is cooperating with kmem to return unused memory to the system,
119  * and kmem respectfully accepts this help at the client's convenience. When
120  * asked to move an object, the client can respond with any of the following:
121  *
122  *   typedef enum kmem_cbrc {
123  *           KMEM_CBRC_YES,
124  *           KMEM_CBRC_NO,
125  *           KMEM_CBRC_LATER,
126  *           KMEM_CBRC_DONT_NEED,
127  *           KMEM_CBRC_DONT_KNOW
128  *   } kmem_cbrc_t;
129  *
130  * The client must not explicitly kmem_cache_free() either of the objects passed
131  * to the callback, since kmem wants to free them directly to the slab layer
132  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
133  * objects to free:
134  *
135  *       YES: (Did it) The client moved the object, so kmem frees the old one.
136  *        NO: (Never) The client refused, so kmem frees the new object (the
137  *            unused copy destination). kmem also marks the slab of the old
138  *            object so as not to bother the client with further callbacks for
139  *            that object as long as the slab remains on the partial slab list.
140  *            (The system won't be getting the slab back as long as the
141  *            immovable object holds it hostage, so there's no point in moving
142  *            any of its objects.)
143  *     LATER: The client is using the object and cannot move it now, so kmem
144  *            frees the new object (the unused copy destination). kmem still
145  *            attempts to move other objects off the slab, since it expects to
146  *            succeed in clearing the slab in a later callback. The client
147  *            should use LATER instead of NO if the object is likely to become
148  *            movable very soon.
149  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
150  *            with the new object (the unused copy destination). This response
151  *            is the client's opportunity to be a model citizen and give back as
152  *            much as it can.
153  * DONT_KNOW: The client does not know about the object because
154  *            a) the client has just allocated the object and not yet put it
155  *               wherever it expects to find known objects
156  *            b) the client has removed the object from wherever it expects to
157  *               find known objects and is about to free it, or
158  *            c) the client has freed the object.
159  *            In all these cases (a, b, and c) kmem frees the new object (the
160  *            unused copy destination) and searches for the old object in the
161  *            magazine layer. If found, the object is removed from the magazine
162  *            layer and freed to the slab layer so it will no longer hold the
163  *            slab hostage.
164  *
165  * 2.3 Object States
166  *
167  * Neither kmem nor the client can be assumed to know the object's whereabouts
168  * at the time of the callback. An object belonging to a kmem cache may be in
169  * any of the following states:
170  *
171  * 1. Uninitialized on the slab
172  * 2. Allocated from the slab but not constructed (still uninitialized)
173  * 3. Allocated from the slab, constructed, but not yet ready for business
174  *    (not in a valid state for the move callback)
175  * 4. In use (valid and known to the client)
176  * 5. About to be freed (no longer in a valid state for the move callback)
177  * 6. Freed to a magazine (still constructed)
178  * 7. Allocated from a magazine, not yet ready for business (not in a valid
179  *    state for the move callback), and about to return to state #4
180  * 8. Deconstructed on a magazine that is about to be freed
181  * 9. Freed to the slab
182  *
183  * Since the move callback may be called at any time while the object is in any
184  * of the above states (except state #1), the client needs a safe way to
185  * determine whether or not it knows about the object. Specifically, the client
186  * needs to know whether or not the object is in state #4, the only state in
187  * which a move is valid. If the object is in any other state, the client should
188  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
189  * the object's fields.
190  *
191  * Note that although an object may be in state #4 when kmem initiates the move
192  * request, the object may no longer be in that state by the time kmem actually
193  * calls the move function. Not only does the client free objects
194  * asynchronously, kmem itself puts move requests on a queue where thay are
195  * pending until kmem processes them from another context. Also, objects freed
196  * to a magazine appear allocated from the point of view of the slab layer, so
197  * kmem may even initiate requests for objects in a state other than state #4.
198  *
199  * 2.3.1 Magazine Layer
200  *
201  * An important insight revealed by the states listed above is that the magazine
202  * layer is populated only by kmem_cache_free(). Magazines of constructed
203  * objects are never populated directly from the slab layer (which contains raw,
204  * unconstructed objects). Whenever an allocation request cannot be satisfied
205  * from the magazine layer, the magazines are bypassed and the request is
206  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
207  * the object constructor only when allocating from the slab layer, and only in
208  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
209  * the move callback. kmem does not preconstruct objects in anticipation of
210  * kmem_cache_alloc().
211  *
212  * 2.3.2 Object Constructor and Destructor
213  *
214  * If the client supplies a destructor, it must be valid to call the destructor
215  * on a newly created object (immediately after the constructor).
216  *
217  * 2.4 Recognizing Known Objects
218  *
219  * There is a simple test to determine safely whether or not the client knows
220  * about a given object in the move callback. It relies on the fact that kmem
221  * guarantees that the object of the move callback has only been touched by the
222  * client itself or else by kmem. kmem does this by ensuring that none of the
223  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
224  * callback is pending. When the last object on a slab is freed, if there is a
225  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
226  * slabs on that list until all pending callbacks are completed. That way,
227  * clients can be certain that the object of a move callback is in one of the
228  * states listed above, making it possible to distinguish known objects (in
229  * state #4) using the two low order bits of any pointer member (with the
230  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
231  * platforms).
232  *
233  * The test works as long as the client always transitions objects from state #4
234  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
235  * order bit of the client-designated pointer member. Since kmem only writes
236  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
237  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
238  * guaranteed to set at least one of the two low order bits. Therefore, given an
239  * object with a back pointer to a 'container_t *o_container', the client can
240  * test
241  *
242  *      container_t *container = object->o_container;
243  *      if ((uintptr_t)container & 0x3) {
244  *              return (KMEM_CBRC_DONT_KNOW);
245  *      }
246  *
247  * Typically, an object will have a pointer to some structure with a list or
248  * hash where objects from the cache are kept while in use. Assuming that the
249  * client has some way of knowing that the container structure is valid and will
250  * not go away during the move, and assuming that the structure includes a lock
251  * to protect whatever collection is used, then the client would continue as
252  * follows:
253  *
254  *	// Ensure that the container structure does not go away.
255  *      if (container_hold(container) == 0) {
256  *              return (KMEM_CBRC_DONT_KNOW);
257  *      }
258  *      mutex_enter(&container->c_objects_lock);
259  *      if (container != object->o_container) {
260  *              mutex_exit(&container->c_objects_lock);
261  *              container_rele(container);
262  *              return (KMEM_CBRC_DONT_KNOW);
263  *      }
264  *
265  * At this point the client knows that the object cannot be freed as long as
266  * c_objects_lock is held. Note that after acquiring the lock, the client must
267  * recheck the o_container pointer in case the object was removed just before
268  * acquiring the lock.
269  *
270  * When the client is about to free an object, it must first remove that object
271  * from the list, hash, or other structure where it is kept. At that time, to
272  * mark the object so it can be distinguished from the remaining, known objects,
273  * the client sets the designated low order bit:
274  *
275  *      mutex_enter(&container->c_objects_lock);
276  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
277  *      list_remove(&container->c_objects, object);
278  *      mutex_exit(&container->c_objects_lock);
279  *
280  * In the common case, the object is freed to the magazine layer, where it may
281  * be reused on a subsequent allocation without the overhead of calling the
282  * constructor. While in the magazine it appears allocated from the point of
283  * view of the slab layer, making it a candidate for the move callback. Most
284  * objects unrecognized by the client in the move callback fall into this
285  * category and are cheaply distinguished from known objects by the test
286  * described earlier. Since recognition is cheap for the client, and searching
287  * magazines is expensive for kmem, kmem defers searching until the client first
288  * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
289  * elsewhere does what it can to avoid bothering the client unnecessarily.
290  *
291  * Invalidating the designated pointer member before freeing the object marks
292  * the object to be avoided in the callback, and conversely, assigning a valid
293  * value to the designated pointer member after allocating the object makes the
294  * object fair game for the callback:
295  *
296  *      ... allocate object ...
297  *      ... set any initial state not set by the constructor ...
298  *
299  *      mutex_enter(&container->c_objects_lock);
300  *      list_insert_tail(&container->c_objects, object);
301  *      membar_producer();
302  *      object->o_container = container;
303  *      mutex_exit(&container->c_objects_lock);
304  *
305  * Note that everything else must be valid before setting o_container makes the
306  * object fair game for the move callback. The membar_producer() call ensures
307  * that all the object's state is written to memory before setting the pointer
308  * that transitions the object from state #3 or #7 (allocated, constructed, not
309  * yet in use) to state #4 (in use, valid). That's important because the move
310  * function has to check the validity of the pointer before it can safely
311  * acquire the lock protecting the collection where it expects to find known
312  * objects.
313  *
314  * This method of distinguishing known objects observes the usual symmetry:
315  * invalidating the designated pointer is the first thing the client does before
316  * freeing the object, and setting the designated pointer is the last thing the
317  * client does after allocating the object. Of course, the client is not
318  * required to use this method. Fundamentally, how the client recognizes known
319  * objects is completely up to the client, but this method is recommended as an
320  * efficient and safe way to take advantage of the guarantees made by kmem. If
321  * the entire object is arbitrary data without any markable bits from a suitable
322  * pointer member, then the client must find some other method, such as
323  * searching a hash table of known objects.
324  *
325  * 2.5 Preventing Objects From Moving
326  *
327  * Besides a way to distinguish known objects, the other thing that the client
328  * needs is a strategy to ensure that an object will not move while the client
329  * is actively using it. The details of satisfying this requirement tend to be
330  * highly cache-specific. It might seem that the same rules that let a client
331  * remove an object safely should also decide when an object can be moved
332  * safely. However, any object state that makes a removal attempt invalid is
333  * likely to be long-lasting for objects that the client does not expect to
334  * remove. kmem knows nothing about the object state and is equally likely (from
335  * the client's point of view) to request a move for any object in the cache,
336  * whether prepared for removal or not. Even a low percentage of objects stuck
337  * in place by unremovability will defeat the consolidator if the stuck objects
338  * are the same long-lived allocations likely to hold slabs hostage.
339  * Fundamentally, the consolidator is not aimed at common cases. Severe external
340  * fragmentation is a worst case scenario manifested as sparsely allocated
341  * slabs, by definition a low percentage of the cache's objects. When deciding
342  * what makes an object movable, keep in mind the goal of the consolidator: to
343  * bring worst-case external fragmentation within the limits guaranteed for
344  * internal fragmentation. Removability is a poor criterion if it is likely to
345  * exclude more than an insignificant percentage of objects for long periods of
346  * time.
347  *
348  * A tricky general solution exists, and it has the advantage of letting you
349  * move any object at almost any moment, practically eliminating the likelihood
350  * that an object can hold a slab hostage. However, if there is a cache-specific
351  * way to ensure that an object is not actively in use in the vast majority of
352  * cases, a simpler solution that leverages this cache-specific knowledge is
353  * preferred.
354  *
355  * 2.5.1 Cache-Specific Solution
356  *
357  * As an example of a cache-specific solution, the ZFS znode cache takes
358  * advantage of the fact that the vast majority of znodes are only being
359  * referenced from the DNLC. (A typical case might be a few hundred in active
360  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
361  * client has established that it recognizes the znode and can access its fields
362  * safely (using the method described earlier), it then tests whether the znode
363  * is referenced by anything other than the DNLC. If so, it assumes that the
364  * znode may be in active use and is unsafe to move, so it drops its locks and
365  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
366  * else znodes are used, no change is needed to protect against the possibility
367  * of the znode moving. The disadvantage is that it remains possible for an
368  * application to hold a znode slab hostage with an open file descriptor.
369  * However, this case ought to be rare and the consolidator has a way to deal
370  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
371  * object, kmem eventually stops believing it and treats the slab as if the
372  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
373  * then focus on getting it off of the partial slab list by allocating rather
374  * than freeing all of its objects. (Either way of getting a slab off the
375  * free list reduces fragmentation.)
376  *
377  * 2.5.2 General Solution
378  *
379  * The general solution, on the other hand, requires an explicit hold everywhere
380  * the object is used to prevent it from moving. To keep the client locking
381  * strategy as uncomplicated as possible, kmem guarantees the simplifying
382  * assumption that move callbacks are sequential, even across multiple caches.
383  * Internally, a global queue processed by a single thread supports all caches
384  * implementing the callback function. No matter how many caches supply a move
385  * function, the consolidator never moves more than one object at a time, so the
386  * client does not have to worry about tricky lock ordering involving several
387  * related objects from different kmem caches.
388  *
389  * The general solution implements the explicit hold as a read-write lock, which
390  * allows multiple readers to access an object from the cache simultaneously
391  * while a single writer is excluded from moving it. A single rwlock for the
392  * entire cache would lock out all threads from using any of the cache's objects
393  * even though only a single object is being moved, so to reduce contention,
394  * the client can fan out the single rwlock into an array of rwlocks hashed by
395  * the object address, making it probable that moving one object will not
396  * prevent other threads from using a different object. The rwlock cannot be a
397  * member of the object itself, because the possibility of the object moving
398  * makes it unsafe to access any of the object's fields until the lock is
399  * acquired.
400  *
401  * Assuming a small, fixed number of locks, it's possible that multiple objects
402  * will hash to the same lock. A thread that needs to use multiple objects in
403  * the same function may acquire the same lock multiple times. Since rwlocks are
404  * reentrant for readers, and since there is never more than a single writer at
405  * a time (assuming that the client acquires the lock as a writer only when
406  * moving an object inside the callback), there would seem to be no problem.
407  * However, a client locking multiple objects in the same function must handle
408  * one case of potential deadlock: Assume that thread A needs to prevent both
409  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
410  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
411  * same lock, that thread A will acquire the lock for object 1 as a reader
412  * before thread B sets the lock's write-wanted bit, preventing thread A from
413  * reacquiring the lock for object 2 as a reader. Unable to make forward
414  * progress, thread A will never release the lock for object 1, resulting in
415  * deadlock.
416  *
417  * There are two ways of avoiding the deadlock just described. The first is to
418  * use rw_tryenter() rather than rw_enter() in the callback function when
419  * attempting to acquire the lock as a writer. If tryenter discovers that the
420  * same object (or another object hashed to the same lock) is already in use, it
421  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
422  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
423  * since it allows a thread to acquire the lock as a reader in spite of a
424  * waiting writer. This second approach insists on moving the object now, no
425  * matter how many readers the move function must wait for in order to do so,
426  * and could delay the completion of the callback indefinitely (blocking
427  * callbacks to other clients). In practice, a less insistent callback using
428  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
429  * little reason to use anything else.
430  *
431  * Avoiding deadlock is not the only problem that an implementation using an
432  * explicit hold needs to solve. Locking the object in the first place (to
433  * prevent it from moving) remains a problem, since the object could move
434  * between the time you obtain a pointer to the object and the time you acquire
435  * the rwlock hashed to that pointer value. Therefore the client needs to
436  * recheck the value of the pointer after acquiring the lock, drop the lock if
437  * the value has changed, and try again. This requires a level of indirection:
438  * something that points to the object rather than the object itself, that the
439  * client can access safely while attempting to acquire the lock. (The object
440  * itself cannot be referenced safely because it can move at any time.)
441  * The following lock-acquisition function takes whatever is safe to reference
442  * (arg), follows its pointer to the object (using function f), and tries as
443  * often as necessary to acquire the hashed lock and verify that the object
444  * still has not moved:
445  *
446  *      object_t *
447  *      object_hold(object_f f, void *arg)
448  *      {
449  *              object_t *op;
450  *
451  *              op = f(arg);
452  *              if (op == NULL) {
453  *                      return (NULL);
454  *              }
455  *
456  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
457  *              while (op != f(arg)) {
458  *                      rw_exit(OBJECT_RWLOCK(op));
459  *                      op = f(arg);
460  *                      if (op == NULL) {
461  *                              break;
462  *                      }
463  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
464  *              }
465  *
466  *              return (op);
467  *      }
468  *
469  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
470  * lock reacquisition loop, while necessary, almost never executes. The function
471  * pointer f (used to obtain the object pointer from arg) has the following type
472  * definition:
473  *
474  *      typedef object_t *(*object_f)(void *arg);
475  *
476  * An object_f implementation is likely to be as simple as accessing a structure
477  * member:
478  *
479  *      object_t *
480  *      s_object(void *arg)
481  *      {
482  *              something_t *sp = arg;
483  *              return (sp->s_object);
484  *      }
485  *
486  * The flexibility of a function pointer allows the path to the object to be
487  * arbitrarily complex and also supports the notion that depending on where you
488  * are using the object, you may need to get it from someplace different.
489  *
490  * The function that releases the explicit hold is simpler because it does not
491  * have to worry about the object moving:
492  *
493  *      void
494  *      object_rele(object_t *op)
495  *      {
496  *              rw_exit(OBJECT_RWLOCK(op));
497  *      }
498  *
499  * The caller is spared these details so that obtaining and releasing an
500  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
501  * of object_hold() only needs to know that the returned object pointer is valid
502  * if not NULL and that the object will not move until released.
503  *
504  * Although object_hold() prevents an object from moving, it does not prevent it
505  * from being freed. The caller must take measures before calling object_hold()
506  * (afterwards is too late) to ensure that the held object cannot be freed. The
507  * caller must do so without accessing the unsafe object reference, so any lock
508  * or reference count used to ensure the continued existence of the object must
509  * live outside the object itself.
510  *
511  * Obtaining a new object is a special case where an explicit hold is impossible
512  * for the caller. Any function that returns a newly allocated object (either as
513  * a return value, or as an in-out paramter) must return it already held; after
514  * the caller gets it is too late, since the object cannot be safely accessed
515  * without the level of indirection described earlier. The following
516  * object_alloc() example uses the same code shown earlier to transition a new
517  * object into the state of being recognized (by the client) as a known object.
518  * The function must acquire the hold (rw_enter) before that state transition
519  * makes the object movable:
520  *
521  *      static object_t *
522  *      object_alloc(container_t *container)
523  *      {
524  *              object_t *object = kmem_cache_alloc(object_cache, 0);
525  *              ... set any initial state not set by the constructor ...
526  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
527  *              mutex_enter(&container->c_objects_lock);
528  *              list_insert_tail(&container->c_objects, object);
529  *              membar_producer();
530  *              object->o_container = container;
531  *              mutex_exit(&container->c_objects_lock);
532  *              return (object);
533  *      }
534  *
535  * Functions that implicitly acquire an object hold (any function that calls
536  * object_alloc() to supply an object for the caller) need to be carefully noted
537  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
538  * prevent all objects hashed to the affected rwlocks from ever being moved.
539  *
540  * The pointer to a held object can be hashed to the holding rwlock even after
541  * the object has been freed. Although it is possible to release the hold
542  * after freeing the object, you may decide to release the hold implicitly in
543  * whatever function frees the object, so as to release the hold as soon as
544  * possible, and for the sake of symmetry with the function that implicitly
545  * acquires the hold when it allocates the object. Here, object_free() releases
546  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
547  * matching pair with object_hold():
548  *
549  *      void
550  *      object_free(object_t *object)
551  *      {
552  *              container_t *container;
553  *
554  *              ASSERT(object_held(object));
555  *              container = object->o_container;
556  *              mutex_enter(&container->c_objects_lock);
557  *              object->o_container =
558  *                  (void *)((uintptr_t)object->o_container | 0x1);
559  *              list_remove(&container->c_objects, object);
560  *              mutex_exit(&container->c_objects_lock);
561  *              object_rele(object);
562  *              kmem_cache_free(object_cache, object);
563  *      }
564  *
565  * Note that object_free() cannot safely accept an object pointer as an argument
566  * unless the object is already held. Any function that calls object_free()
567  * needs to be carefully noted since it similarly forms a matching pair with
568  * object_hold().
569  *
570  * To complete the picture, the following callback function implements the
571  * general solution by moving objects only if they are currently unheld:
572  *
573  *      static kmem_cbrc_t
574  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
575  *      {
576  *              object_t *op = buf, *np = newbuf;
577  *              container_t *container;
578  *
579  *              container = op->o_container;
580  *              if ((uintptr_t)container & 0x3) {
581  *                      return (KMEM_CBRC_DONT_KNOW);
582  *              }
583  *
584  *	        // Ensure that the container structure does not go away.
585  *              if (container_hold(container) == 0) {
586  *                      return (KMEM_CBRC_DONT_KNOW);
587  *              }
588  *
589  *              mutex_enter(&container->c_objects_lock);
590  *              if (container != op->o_container) {
591  *                      mutex_exit(&container->c_objects_lock);
592  *                      container_rele(container);
593  *                      return (KMEM_CBRC_DONT_KNOW);
594  *              }
595  *
596  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
597  *                      mutex_exit(&container->c_objects_lock);
598  *                      container_rele(container);
599  *                      return (KMEM_CBRC_LATER);
600  *              }
601  *
602  *              object_move_impl(op, np); // critical section
603  *              rw_exit(OBJECT_RWLOCK(op));
604  *
605  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
606  *              list_link_replace(&op->o_link_node, &np->o_link_node);
607  *              mutex_exit(&container->c_objects_lock);
608  *              container_rele(container);
609  *              return (KMEM_CBRC_YES);
610  *      }
611  *
612  * Note that object_move() must invalidate the designated o_container pointer of
613  * the old object in the same way that object_free() does, since kmem will free
614  * the object in response to the KMEM_CBRC_YES return value.
615  *
616  * The lock order in object_move() differs from object_alloc(), which locks
617  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
618  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
619  * not a problem. Holding the lock on the object list in the example above
620  * through the entire callback not only prevents the object from going away, it
621  * also allows you to lock the list elsewhere and know that none of its elements
622  * will move during iteration.
623  *
624  * Adding an explicit hold everywhere an object from the cache is used is tricky
625  * and involves much more change to client code than a cache-specific solution
626  * that leverages existing state to decide whether or not an object is
627  * movable. However, this approach has the advantage that no object remains
628  * immovable for any significant length of time, making it extremely unlikely
629  * that long-lived allocations can continue holding slabs hostage; and it works
630  * for any cache.
631  *
632  * 3. Consolidator Implementation
633  *
634  * Once the client supplies a move function that a) recognizes known objects and
635  * b) avoids moving objects that are actively in use, the remaining work is up
636  * to the consolidator to decide which objects to move and when to issue
637  * callbacks.
638  *
639  * The consolidator relies on the fact that a cache's slabs are ordered by
640  * usage. Each slab has a fixed number of objects. Depending on the slab's
641  * "color" (the offset of the first object from the beginning of the slab;
642  * offsets are staggered to mitigate false sharing of cache lines) it is either
643  * the maximum number of objects per slab determined at cache creation time or
644  * else the number closest to the maximum that fits within the space remaining
645  * after the initial offset. A completely allocated slab may contribute some
646  * internal fragmentation (per-slab overhead) but no external fragmentation, so
647  * it is of no interest to the consolidator. At the other extreme, slabs whose
648  * objects have all been freed to the slab are released to the virtual memory
649  * (VM) subsystem (objects freed to magazines are still allocated as far as the
650  * slab is concerned). External fragmentation exists when there are slabs
651  * somewhere between these extremes. A partial slab has at least one but not all
652  * of its objects allocated. The more partial slabs, and the fewer allocated
653  * objects on each of them, the higher the fragmentation. Hence the
654  * consolidator's overall strategy is to reduce the number of partial slabs by
655  * moving allocated objects from the least allocated slabs to the most allocated
656  * slabs.
657  *
658  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
659  * slabs are kept separately in an unordered list. Since the majority of slabs
660  * tend to be completely allocated (a typical unfragmented cache may have
661  * thousands of complete slabs and only a single partial slab), separating
662  * complete slabs improves the efficiency of partial slab ordering, since the
663  * complete slabs do not affect the depth or balance of the AVL tree. This
664  * ordered sequence of partial slabs acts as a "free list" supplying objects for
665  * allocation requests.
666  *
667  * Objects are always allocated from the first partial slab in the free list,
668  * where the allocation is most likely to eliminate a partial slab (by
669  * completely allocating it). Conversely, when a single object from a completely
670  * allocated slab is freed to the slab, that slab is added to the front of the
671  * free list. Since most free list activity involves highly allocated slabs
672  * coming and going at the front of the list, slabs tend naturally toward the
673  * ideal order: highly allocated at the front, sparsely allocated at the back.
674  * Slabs with few allocated objects are likely to become completely free if they
675  * keep a safe distance away from the front of the free list. Slab misorders
676  * interfere with the natural tendency of slabs to become completely free or
677  * completely allocated. For example, a slab with a single allocated object
678  * needs only a single free to escape the cache; its natural desire is
679  * frustrated when it finds itself at the front of the list where a second
680  * allocation happens just before the free could have released it. Another slab
681  * with all but one object allocated might have supplied the buffer instead, so
682  * that both (as opposed to neither) of the slabs would have been taken off the
683  * free list.
684  *
685  * Although slabs tend naturally toward the ideal order, misorders allowed by a
686  * simple list implementation defeat the consolidator's strategy of merging
687  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
688  * needs another way to fix misorders to optimize its callback strategy. One
689  * approach is to periodically scan a limited number of slabs, advancing a
690  * marker to hold the current scan position, and to move extreme misorders to
691  * the front or back of the free list and to the front or back of the current
692  * scan range. By making consecutive scan ranges overlap by one slab, the least
693  * allocated slab in the current range can be carried along from the end of one
694  * scan to the start of the next.
695  *
696  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
697  * task, however. Since most of the cache's activity is in the magazine layer,
698  * and allocations from the slab layer represent only a startup cost, the
699  * overhead of maintaining a balanced tree is not a significant concern compared
700  * to the opportunity of reducing complexity by eliminating the partial slab
701  * scanner just described. The overhead of an AVL tree is minimized by
702  * maintaining only partial slabs in the tree and keeping completely allocated
703  * slabs separately in a list. To avoid increasing the size of the slab
704  * structure the AVL linkage pointers are reused for the slab's list linkage,
705  * since the slab will always be either partial or complete, never stored both
706  * ways at the same time. To further minimize the overhead of the AVL tree the
707  * compare function that orders partial slabs by usage divides the range of
708  * allocated object counts into bins such that counts within the same bin are
709  * considered equal. Binning partial slabs makes it less likely that allocating
710  * or freeing a single object will change the slab's order, requiring a tree
711  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
712  * requiring some rebalancing of the tree). Allocation counts closest to
713  * completely free and completely allocated are left unbinned (finely sorted) to
714  * better support the consolidator's strategy of merging slabs at either
715  * extreme.
716  *
717  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
718  *
719  * The consolidator piggybacks on the kmem maintenance thread and is called on
720  * the same interval as kmem_cache_update(), once per cache every fifteen
721  * seconds. kmem maintains a running count of unallocated objects in the slab
722  * layer (cache_bufslab). The consolidator checks whether that number exceeds
723  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
724  * there is a significant number of slabs in the cache (arbitrarily a minimum
725  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
726  * working set are included in the assessment, and magazines in the depot are
727  * reaped if those objects would lift cache_bufslab above the fragmentation
728  * threshold. Once the consolidator decides that a cache is fragmented, it looks
729  * for a candidate slab to reclaim, starting at the end of the partial slab free
730  * list and scanning backwards. At first the consolidator is choosy: only a slab
731  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
732  * single allocated object, regardless of percentage). If there is difficulty
733  * finding a candidate slab, kmem raises the allocation threshold incrementally,
734  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
735  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
736  * even in the worst case of every slab in the cache being almost 7/8 allocated.
737  * The threshold can also be lowered incrementally when candidate slabs are easy
738  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
739  * is no longer fragmented.
740  *
741  * 3.2 Generating Callbacks
742  *
743  * Once an eligible slab is chosen, a callback is generated for every allocated
744  * object on the slab, in the hope that the client will move everything off the
745  * slab and make it reclaimable. Objects selected as move destinations are
746  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
747  * order (most allocated at the front, least allocated at the back) and a
748  * cooperative client, the consolidator will succeed in removing slabs from both
749  * ends of the free list, completely allocating on the one hand and completely
750  * freeing on the other. Objects selected as move destinations are allocated in
751  * the kmem maintenance thread where move requests are enqueued. A separate
752  * callback thread removes pending callbacks from the queue and calls the
753  * client. The separate thread ensures that client code (the move function) does
754  * not interfere with internal kmem maintenance tasks. A map of pending
755  * callbacks keyed by object address (the object to be moved) is checked to
756  * ensure that duplicate callbacks are not generated for the same object.
757  * Allocating the move destination (the object to move to) prevents subsequent
758  * callbacks from selecting the same destination as an earlier pending callback.
759  *
760  * Move requests can also be generated by kmem_cache_reap() when the system is
761  * desperate for memory and by kmem_cache_move_notify(), called by the client to
762  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
763  * The map of pending callbacks is protected by the same lock that protects the
764  * slab layer.
765  *
766  * When the system is desperate for memory, kmem does not bother to determine
767  * whether or not the cache exceeds the fragmentation threshold, but tries to
768  * consolidate as many slabs as possible. Normally, the consolidator chews
769  * slowly, one sparsely allocated slab at a time during each maintenance
770  * interval that the cache is fragmented. When desperate, the consolidator
771  * starts at the last partial slab and enqueues callbacks for every allocated
772  * object on every partial slab, working backwards until it reaches the first
773  * partial slab. The first partial slab, meanwhile, advances in pace with the
774  * consolidator as allocations to supply move destinations for the enqueued
775  * callbacks use up the highly allocated slabs at the front of the free list.
776  * Ideally, the overgrown free list collapses like an accordion, starting at
777  * both ends and ending at the center with a single partial slab.
778  *
779  * 3.3 Client Responses
780  *
781  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
782  * marks the slab that supplied the stuck object non-reclaimable and moves it to
783  * front of the free list. The slab remains marked as long as it remains on the
784  * free list, and it appears more allocated to the partial slab compare function
785  * than any unmarked slab, no matter how many of its objects are allocated.
786  * Since even one immovable object ties up the entire slab, the goal is to
787  * completely allocate any slab that cannot be completely freed. kmem does not
788  * bother generating callbacks to move objects from a marked slab unless the
789  * system is desperate.
790  *
791  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
792  * slab. If the client responds LATER too many times, kmem disbelieves and
793  * treats the response as a NO. The count is cleared when the slab is taken off
794  * the partial slab list or when the client moves one of the slab's objects.
795  *
796  * 4. Observability
797  *
798  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
799  * the ::kmem_slabs dcmd. For a complete description of the command, enter
800  * '::help kmem_slabs' at the mdb prompt.
801  */
802 
803 #include <sys/kmem_impl.h>
804 #include <sys/vmem_impl.h>
805 #include <sys/param.h>
806 #include <sys/sysmacros.h>
807 #include <sys/vm.h>
808 #include <sys/proc.h>
809 #include <sys/tuneable.h>
810 #include <sys/systm.h>
811 #include <sys/cmn_err.h>
812 #include <sys/debug.h>
813 #include <sys/sdt.h>
814 #include <sys/mutex.h>
815 #include <sys/bitmap.h>
816 #include <sys/atomic.h>
817 #include <sys/kobj.h>
818 #include <sys/disp.h>
819 #include <vm/seg_kmem.h>
820 #include <sys/log.h>
821 #include <sys/callb.h>
822 #include <sys/taskq.h>
823 #include <sys/modctl.h>
824 #include <sys/reboot.h>
825 #include <sys/id32.h>
826 #include <sys/zone.h>
827 #include <sys/netstack.h>
828 #ifdef	DEBUG
829 #include <sys/random.h>
830 #endif
831 
832 extern void streams_msg_init(void);
833 extern int segkp_fromheap;
834 extern void segkp_cache_free(void);
835 
836 struct kmem_cache_kstat {
837 	kstat_named_t	kmc_buf_size;
838 	kstat_named_t	kmc_align;
839 	kstat_named_t	kmc_chunk_size;
840 	kstat_named_t	kmc_slab_size;
841 	kstat_named_t	kmc_alloc;
842 	kstat_named_t	kmc_alloc_fail;
843 	kstat_named_t	kmc_free;
844 	kstat_named_t	kmc_depot_alloc;
845 	kstat_named_t	kmc_depot_free;
846 	kstat_named_t	kmc_depot_contention;
847 	kstat_named_t	kmc_slab_alloc;
848 	kstat_named_t	kmc_slab_free;
849 	kstat_named_t	kmc_buf_constructed;
850 	kstat_named_t	kmc_buf_avail;
851 	kstat_named_t	kmc_buf_inuse;
852 	kstat_named_t	kmc_buf_total;
853 	kstat_named_t	kmc_buf_max;
854 	kstat_named_t	kmc_slab_create;
855 	kstat_named_t	kmc_slab_destroy;
856 	kstat_named_t	kmc_vmem_source;
857 	kstat_named_t	kmc_hash_size;
858 	kstat_named_t	kmc_hash_lookup_depth;
859 	kstat_named_t	kmc_hash_rescale;
860 	kstat_named_t	kmc_full_magazines;
861 	kstat_named_t	kmc_empty_magazines;
862 	kstat_named_t	kmc_magazine_size;
863 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
864 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
865 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
866 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
867 	kstat_named_t	kmc_move_yes;
868 	kstat_named_t	kmc_move_no;
869 	kstat_named_t	kmc_move_later;
870 	kstat_named_t	kmc_move_dont_need;
871 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
872 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
873 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
874 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
875 } kmem_cache_kstat = {
876 	{ "buf_size",		KSTAT_DATA_UINT64 },
877 	{ "align",		KSTAT_DATA_UINT64 },
878 	{ "chunk_size",		KSTAT_DATA_UINT64 },
879 	{ "slab_size",		KSTAT_DATA_UINT64 },
880 	{ "alloc",		KSTAT_DATA_UINT64 },
881 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
882 	{ "free",		KSTAT_DATA_UINT64 },
883 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
884 	{ "depot_free",		KSTAT_DATA_UINT64 },
885 	{ "depot_contention",	KSTAT_DATA_UINT64 },
886 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
887 	{ "slab_free",		KSTAT_DATA_UINT64 },
888 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
889 	{ "buf_avail",		KSTAT_DATA_UINT64 },
890 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
891 	{ "buf_total",		KSTAT_DATA_UINT64 },
892 	{ "buf_max",		KSTAT_DATA_UINT64 },
893 	{ "slab_create",	KSTAT_DATA_UINT64 },
894 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
895 	{ "vmem_source",	KSTAT_DATA_UINT64 },
896 	{ "hash_size",		KSTAT_DATA_UINT64 },
897 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
898 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
899 	{ "full_magazines",	KSTAT_DATA_UINT64 },
900 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
901 	{ "magazine_size",	KSTAT_DATA_UINT64 },
902 	{ "reap",		KSTAT_DATA_UINT64 },
903 	{ "defrag",		KSTAT_DATA_UINT64 },
904 	{ "scan",		KSTAT_DATA_UINT64 },
905 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
906 	{ "move_yes",		KSTAT_DATA_UINT64 },
907 	{ "move_no",		KSTAT_DATA_UINT64 },
908 	{ "move_later",		KSTAT_DATA_UINT64 },
909 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
910 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
911 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
912 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
913 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
914 };
915 
916 static kmutex_t kmem_cache_kstat_lock;
917 
918 /*
919  * The default set of caches to back kmem_alloc().
920  * These sizes should be reevaluated periodically.
921  *
922  * We want allocations that are multiples of the coherency granularity
923  * (64 bytes) to be satisfied from a cache which is a multiple of 64
924  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
925  * the next kmem_cache_size greater than or equal to it must be a
926  * multiple of 64.
927  *
928  * We split the table into two sections:  size <= 4k and size > 4k.  This
929  * saves a lot of space and cache footprint in our cache tables.
930  */
931 static const int kmem_alloc_sizes[] = {
932 	1 * 8,
933 	2 * 8,
934 	3 * 8,
935 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
936 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
937 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
938 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
939 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
940 	P2ALIGN(8192 / 7, 64),
941 	P2ALIGN(8192 / 6, 64),
942 	P2ALIGN(8192 / 5, 64),
943 	P2ALIGN(8192 / 4, 64),
944 	P2ALIGN(8192 / 3, 64),
945 	P2ALIGN(8192 / 2, 64),
946 };
947 
948 static const int kmem_big_alloc_sizes[] = {
949 	2 * 4096,	3 * 4096,
950 	2 * 8192,	3 * 8192,
951 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
952 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
953 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
954 	16 * 8192
955 };
956 
957 #define	KMEM_MAXBUF		4096
958 #define	KMEM_BIG_MAXBUF_32BIT	32768
959 #define	KMEM_BIG_MAXBUF		131072
960 
961 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
962 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
963 
964 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
965 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
966 
967 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
968 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
969 
970 static kmem_magtype_t kmem_magtype[] = {
971 	{ 1,	8,	3200,	65536	},
972 	{ 3,	16,	256,	32768	},
973 	{ 7,	32,	64,	16384	},
974 	{ 15,	64,	0,	8192	},
975 	{ 31,	64,	0,	4096	},
976 	{ 47,	64,	0,	2048	},
977 	{ 63,	64,	0,	1024	},
978 	{ 95,	64,	0,	512	},
979 	{ 143,	64,	0,	0	},
980 };
981 
982 static uint32_t kmem_reaping;
983 static uint32_t kmem_reaping_idspace;
984 
985 /*
986  * kmem tunables
987  */
988 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
989 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
990 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
991 int kmem_panic = 1;		/* whether to panic on error */
992 int kmem_logging = 1;		/* kmem_log_enter() override */
993 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
994 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
995 size_t kmem_content_log_size;	/* content log size [2% of memory] */
996 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
997 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
998 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
999 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1000 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1001 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1002 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1003 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1004 
1005 #ifdef _LP64
1006 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1007 #else
1008 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1009 #endif
1010 
1011 #ifdef DEBUG
1012 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1013 #else
1014 int kmem_flags = 0;
1015 #endif
1016 int kmem_ready;
1017 
1018 static kmem_cache_t	*kmem_slab_cache;
1019 static kmem_cache_t	*kmem_bufctl_cache;
1020 static kmem_cache_t	*kmem_bufctl_audit_cache;
1021 
1022 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1023 static list_t		kmem_caches;
1024 
1025 static taskq_t		*kmem_taskq;
1026 static kmutex_t		kmem_flags_lock;
1027 static vmem_t		*kmem_metadata_arena;
1028 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1029 static vmem_t		*kmem_cache_arena;
1030 static vmem_t		*kmem_hash_arena;
1031 static vmem_t		*kmem_log_arena;
1032 static vmem_t		*kmem_oversize_arena;
1033 static vmem_t		*kmem_va_arena;
1034 static vmem_t		*kmem_default_arena;
1035 static vmem_t		*kmem_firewall_va_arena;
1036 static vmem_t		*kmem_firewall_arena;
1037 
1038 /*
1039  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1040  * turned on when DEBUG is also defined.
1041  */
1042 #ifdef	DEBUG
1043 #define	KMEM_STATS
1044 #endif	/* DEBUG */
1045 
1046 #ifdef	KMEM_STATS
1047 #define	KMEM_STAT_ADD(stat)			((stat)++)
1048 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1049 #else
1050 #define	KMEM_STAT_ADD(stat)			/* nothing */
1051 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1052 #endif	/* KMEM_STATS */
1053 
1054 /*
1055  * kmem slab consolidator thresholds (tunables)
1056  */
1057 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1058 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1059 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1060 /*
1061  * Maximum number of slabs from which to move buffers during a single
1062  * maintenance interval while the system is not low on memory.
1063  */
1064 size_t kmem_reclaim_max_slabs = 1;
1065 /*
1066  * Number of slabs to scan backwards from the end of the partial slab list
1067  * when searching for buffers to relocate.
1068  */
1069 size_t kmem_reclaim_scan_range = 12;
1070 
1071 #ifdef	KMEM_STATS
1072 static struct {
1073 	uint64_t kms_callbacks;
1074 	uint64_t kms_yes;
1075 	uint64_t kms_no;
1076 	uint64_t kms_later;
1077 	uint64_t kms_dont_need;
1078 	uint64_t kms_dont_know;
1079 	uint64_t kms_hunt_found_mag;
1080 	uint64_t kms_hunt_found_slab;
1081 	uint64_t kms_hunt_alloc_fail;
1082 	uint64_t kms_hunt_lucky;
1083 	uint64_t kms_notify;
1084 	uint64_t kms_notify_callbacks;
1085 	uint64_t kms_disbelief;
1086 	uint64_t kms_already_pending;
1087 	uint64_t kms_callback_alloc_fail;
1088 	uint64_t kms_callback_taskq_fail;
1089 	uint64_t kms_endscan_slab_dead;
1090 	uint64_t kms_endscan_slab_destroyed;
1091 	uint64_t kms_endscan_nomem;
1092 	uint64_t kms_endscan_refcnt_changed;
1093 	uint64_t kms_endscan_nomove_changed;
1094 	uint64_t kms_endscan_freelist;
1095 	uint64_t kms_avl_update;
1096 	uint64_t kms_avl_noupdate;
1097 	uint64_t kms_no_longer_reclaimable;
1098 	uint64_t kms_notify_no_longer_reclaimable;
1099 	uint64_t kms_notify_slab_dead;
1100 	uint64_t kms_notify_slab_destroyed;
1101 	uint64_t kms_alloc_fail;
1102 	uint64_t kms_constructor_fail;
1103 	uint64_t kms_dead_slabs_freed;
1104 	uint64_t kms_defrags;
1105 	uint64_t kms_scans;
1106 	uint64_t kms_scan_depot_ws_reaps;
1107 	uint64_t kms_debug_reaps;
1108 	uint64_t kms_debug_scans;
1109 } kmem_move_stats;
1110 #endif	/* KMEM_STATS */
1111 
1112 /* consolidator knobs */
1113 static boolean_t kmem_move_noreap;
1114 static boolean_t kmem_move_blocked;
1115 static boolean_t kmem_move_fulltilt;
1116 static boolean_t kmem_move_any_partial;
1117 
1118 #ifdef	DEBUG
1119 /*
1120  * kmem consolidator debug tunables:
1121  * Ensure code coverage by occasionally running the consolidator even when the
1122  * caches are not fragmented (they may never be). These intervals are mean time
1123  * in cache maintenance intervals (kmem_cache_update).
1124  */
1125 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1126 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1127 #endif	/* DEBUG */
1128 
1129 static kmem_cache_t	*kmem_defrag_cache;
1130 static kmem_cache_t	*kmem_move_cache;
1131 static taskq_t		*kmem_move_taskq;
1132 
1133 static void kmem_cache_scan(kmem_cache_t *);
1134 static void kmem_cache_defrag(kmem_cache_t *);
1135 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1136 
1137 
1138 kmem_log_header_t	*kmem_transaction_log;
1139 kmem_log_header_t	*kmem_content_log;
1140 kmem_log_header_t	*kmem_failure_log;
1141 kmem_log_header_t	*kmem_slab_log;
1142 
1143 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1144 
1145 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1146 	if ((count) > 0) {						\
1147 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1148 		pc_t *_e;						\
1149 		/* memmove() the old entries down one notch */		\
1150 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1151 			*_e = *(_e - 1);				\
1152 		*_s = (uintptr_t)(caller);				\
1153 	}
1154 
1155 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1156 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1157 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1158 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1159 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1160 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1161 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1162 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1163 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1164 
1165 struct {
1166 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1167 	int		kmp_error;	/* type of kmem error */
1168 	void		*kmp_buffer;	/* buffer that induced panic */
1169 	void		*kmp_realbuf;	/* real start address for buffer */
1170 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1171 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1172 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1173 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1174 } kmem_panic_info;
1175 
1176 
1177 static void
1178 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1179 {
1180 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1181 	uint64_t *buf = buf_arg;
1182 
1183 	while (buf < bufend)
1184 		*buf++ = pattern;
1185 }
1186 
1187 static void *
1188 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1189 {
1190 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1191 	uint64_t *buf;
1192 
1193 	for (buf = buf_arg; buf < bufend; buf++)
1194 		if (*buf != pattern)
1195 			return (buf);
1196 	return (NULL);
1197 }
1198 
1199 static void *
1200 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1201 {
1202 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1203 	uint64_t *buf;
1204 
1205 	for (buf = buf_arg; buf < bufend; buf++) {
1206 		if (*buf != old) {
1207 			copy_pattern(old, buf_arg,
1208 			    (char *)buf - (char *)buf_arg);
1209 			return (buf);
1210 		}
1211 		*buf = new;
1212 	}
1213 
1214 	return (NULL);
1215 }
1216 
1217 static void
1218 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1219 {
1220 	kmem_cache_t *cp;
1221 
1222 	mutex_enter(&kmem_cache_lock);
1223 	for (cp = list_head(&kmem_caches); cp != NULL;
1224 	    cp = list_next(&kmem_caches, cp))
1225 		if (tq != NULL)
1226 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1227 			    tqflag);
1228 		else
1229 			func(cp);
1230 	mutex_exit(&kmem_cache_lock);
1231 }
1232 
1233 static void
1234 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1235 {
1236 	kmem_cache_t *cp;
1237 
1238 	mutex_enter(&kmem_cache_lock);
1239 	for (cp = list_head(&kmem_caches); cp != NULL;
1240 	    cp = list_next(&kmem_caches, cp)) {
1241 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1242 			continue;
1243 		if (tq != NULL)
1244 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1245 			    tqflag);
1246 		else
1247 			func(cp);
1248 	}
1249 	mutex_exit(&kmem_cache_lock);
1250 }
1251 
1252 /*
1253  * Debugging support.  Given a buffer address, find its slab.
1254  */
1255 static kmem_slab_t *
1256 kmem_findslab(kmem_cache_t *cp, void *buf)
1257 {
1258 	kmem_slab_t *sp;
1259 
1260 	mutex_enter(&cp->cache_lock);
1261 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1262 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1263 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1264 			mutex_exit(&cp->cache_lock);
1265 			return (sp);
1266 		}
1267 	}
1268 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1269 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1270 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1271 			mutex_exit(&cp->cache_lock);
1272 			return (sp);
1273 		}
1274 	}
1275 	mutex_exit(&cp->cache_lock);
1276 
1277 	return (NULL);
1278 }
1279 
1280 static void
1281 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1282 {
1283 	kmem_buftag_t *btp = NULL;
1284 	kmem_bufctl_t *bcp = NULL;
1285 	kmem_cache_t *cp = cparg;
1286 	kmem_slab_t *sp;
1287 	uint64_t *off;
1288 	void *buf = bufarg;
1289 
1290 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1291 
1292 	kmem_panic_info.kmp_timestamp = gethrtime();
1293 
1294 	sp = kmem_findslab(cp, buf);
1295 	if (sp == NULL) {
1296 		for (cp = list_tail(&kmem_caches); cp != NULL;
1297 		    cp = list_prev(&kmem_caches, cp)) {
1298 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1299 				break;
1300 		}
1301 	}
1302 
1303 	if (sp == NULL) {
1304 		cp = NULL;
1305 		error = KMERR_BADADDR;
1306 	} else {
1307 		if (cp != cparg)
1308 			error = KMERR_BADCACHE;
1309 		else
1310 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1311 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1312 		if (buf != bufarg)
1313 			error = KMERR_BADBASE;
1314 		if (cp->cache_flags & KMF_BUFTAG)
1315 			btp = KMEM_BUFTAG(cp, buf);
1316 		if (cp->cache_flags & KMF_HASH) {
1317 			mutex_enter(&cp->cache_lock);
1318 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1319 				if (bcp->bc_addr == buf)
1320 					break;
1321 			mutex_exit(&cp->cache_lock);
1322 			if (bcp == NULL && btp != NULL)
1323 				bcp = btp->bt_bufctl;
1324 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1325 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1326 			    bcp->bc_addr != buf) {
1327 				error = KMERR_BADBUFCTL;
1328 				bcp = NULL;
1329 			}
1330 		}
1331 	}
1332 
1333 	kmem_panic_info.kmp_error = error;
1334 	kmem_panic_info.kmp_buffer = bufarg;
1335 	kmem_panic_info.kmp_realbuf = buf;
1336 	kmem_panic_info.kmp_cache = cparg;
1337 	kmem_panic_info.kmp_realcache = cp;
1338 	kmem_panic_info.kmp_slab = sp;
1339 	kmem_panic_info.kmp_bufctl = bcp;
1340 
1341 	printf("kernel memory allocator: ");
1342 
1343 	switch (error) {
1344 
1345 	case KMERR_MODIFIED:
1346 		printf("buffer modified after being freed\n");
1347 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1348 		if (off == NULL)	/* shouldn't happen */
1349 			off = buf;
1350 		printf("modification occurred at offset 0x%lx "
1351 		    "(0x%llx replaced by 0x%llx)\n",
1352 		    (uintptr_t)off - (uintptr_t)buf,
1353 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1354 		break;
1355 
1356 	case KMERR_REDZONE:
1357 		printf("redzone violation: write past end of buffer\n");
1358 		break;
1359 
1360 	case KMERR_BADADDR:
1361 		printf("invalid free: buffer not in cache\n");
1362 		break;
1363 
1364 	case KMERR_DUPFREE:
1365 		printf("duplicate free: buffer freed twice\n");
1366 		break;
1367 
1368 	case KMERR_BADBUFTAG:
1369 		printf("boundary tag corrupted\n");
1370 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1371 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1372 		    KMEM_BUFTAG_FREE);
1373 		break;
1374 
1375 	case KMERR_BADBUFCTL:
1376 		printf("bufctl corrupted\n");
1377 		break;
1378 
1379 	case KMERR_BADCACHE:
1380 		printf("buffer freed to wrong cache\n");
1381 		printf("buffer was allocated from %s,\n", cp->cache_name);
1382 		printf("caller attempting free to %s.\n", cparg->cache_name);
1383 		break;
1384 
1385 	case KMERR_BADSIZE:
1386 		printf("bad free: free size (%u) != alloc size (%u)\n",
1387 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1388 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1389 		break;
1390 
1391 	case KMERR_BADBASE:
1392 		printf("bad free: free address (%p) != alloc address (%p)\n",
1393 		    bufarg, buf);
1394 		break;
1395 	}
1396 
1397 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1398 	    bufarg, (void *)bcp, cparg->cache_name);
1399 
1400 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1401 	    error != KMERR_BADBUFCTL) {
1402 		int d;
1403 		timestruc_t ts;
1404 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1405 
1406 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1407 		printf("previous transaction on buffer %p:\n", buf);
1408 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1409 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1410 		    (void *)sp, cp->cache_name);
1411 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1412 			ulong_t off;
1413 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1414 			printf("%s+%lx\n", sym ? sym : "?", off);
1415 		}
1416 	}
1417 	if (kmem_panic > 0)
1418 		panic("kernel heap corruption detected");
1419 	if (kmem_panic == 0)
1420 		debug_enter(NULL);
1421 	kmem_logging = 1;	/* resume logging */
1422 }
1423 
1424 static kmem_log_header_t *
1425 kmem_log_init(size_t logsize)
1426 {
1427 	kmem_log_header_t *lhp;
1428 	int nchunks = 4 * max_ncpus;
1429 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1430 	int i;
1431 
1432 	/*
1433 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1434 	 * to prevent false sharing of cache lines.
1435 	 */
1436 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1437 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1438 	    NULL, NULL, VM_SLEEP);
1439 	bzero(lhp, lhsize);
1440 
1441 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1442 	lhp->lh_nchunks = nchunks;
1443 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1444 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1445 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1446 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1447 	    nchunks * sizeof (int), VM_SLEEP);
1448 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1449 
1450 	for (i = 0; i < max_ncpus; i++) {
1451 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1452 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1453 		clhp->clh_chunk = i;
1454 	}
1455 
1456 	for (i = max_ncpus; i < nchunks; i++)
1457 		lhp->lh_free[i] = i;
1458 
1459 	lhp->lh_head = max_ncpus;
1460 	lhp->lh_tail = 0;
1461 
1462 	return (lhp);
1463 }
1464 
1465 static void *
1466 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1467 {
1468 	void *logspace;
1469 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1470 
1471 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1472 		return (NULL);
1473 
1474 	mutex_enter(&clhp->clh_lock);
1475 	clhp->clh_hits++;
1476 	if (size > clhp->clh_avail) {
1477 		mutex_enter(&lhp->lh_lock);
1478 		lhp->lh_hits++;
1479 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1480 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1481 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1482 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1483 		clhp->clh_current = lhp->lh_base +
1484 		    clhp->clh_chunk * lhp->lh_chunksize;
1485 		clhp->clh_avail = lhp->lh_chunksize;
1486 		if (size > lhp->lh_chunksize)
1487 			size = lhp->lh_chunksize;
1488 		mutex_exit(&lhp->lh_lock);
1489 	}
1490 	logspace = clhp->clh_current;
1491 	clhp->clh_current += size;
1492 	clhp->clh_avail -= size;
1493 	bcopy(data, logspace, size);
1494 	mutex_exit(&clhp->clh_lock);
1495 	return (logspace);
1496 }
1497 
1498 #define	KMEM_AUDIT(lp, cp, bcp)						\
1499 {									\
1500 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1501 	_bcp->bc_timestamp = gethrtime();				\
1502 	_bcp->bc_thread = curthread;					\
1503 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1504 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1505 }
1506 
1507 static void
1508 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1509 	kmem_slab_t *sp, void *addr)
1510 {
1511 	kmem_bufctl_audit_t bca;
1512 
1513 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1514 	bca.bc_addr = addr;
1515 	bca.bc_slab = sp;
1516 	bca.bc_cache = cp;
1517 	KMEM_AUDIT(lp, cp, &bca);
1518 }
1519 
1520 /*
1521  * Create a new slab for cache cp.
1522  */
1523 static kmem_slab_t *
1524 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1525 {
1526 	size_t slabsize = cp->cache_slabsize;
1527 	size_t chunksize = cp->cache_chunksize;
1528 	int cache_flags = cp->cache_flags;
1529 	size_t color, chunks;
1530 	char *buf, *slab;
1531 	kmem_slab_t *sp;
1532 	kmem_bufctl_t *bcp;
1533 	vmem_t *vmp = cp->cache_arena;
1534 
1535 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1536 
1537 	color = cp->cache_color + cp->cache_align;
1538 	if (color > cp->cache_maxcolor)
1539 		color = cp->cache_mincolor;
1540 	cp->cache_color = color;
1541 
1542 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1543 
1544 	if (slab == NULL)
1545 		goto vmem_alloc_failure;
1546 
1547 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1548 
1549 	/*
1550 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1551 	 * consolidator depends (for correctness) on slabs being initialized
1552 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1553 	 * clients to distinguish uninitialized memory from known objects).
1554 	 */
1555 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1556 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1557 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1558 
1559 	if (cache_flags & KMF_HASH) {
1560 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1561 			goto slab_alloc_failure;
1562 		chunks = (slabsize - color) / chunksize;
1563 	} else {
1564 		sp = KMEM_SLAB(cp, slab);
1565 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1566 	}
1567 
1568 	sp->slab_cache	= cp;
1569 	sp->slab_head	= NULL;
1570 	sp->slab_refcnt	= 0;
1571 	sp->slab_base	= buf = slab + color;
1572 	sp->slab_chunks	= chunks;
1573 	sp->slab_stuck_offset = (uint32_t)-1;
1574 	sp->slab_later_count = 0;
1575 	sp->slab_flags = 0;
1576 
1577 	ASSERT(chunks > 0);
1578 	while (chunks-- != 0) {
1579 		if (cache_flags & KMF_HASH) {
1580 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1581 			if (bcp == NULL)
1582 				goto bufctl_alloc_failure;
1583 			if (cache_flags & KMF_AUDIT) {
1584 				kmem_bufctl_audit_t *bcap =
1585 				    (kmem_bufctl_audit_t *)bcp;
1586 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1587 				bcap->bc_cache = cp;
1588 			}
1589 			bcp->bc_addr = buf;
1590 			bcp->bc_slab = sp;
1591 		} else {
1592 			bcp = KMEM_BUFCTL(cp, buf);
1593 		}
1594 		if (cache_flags & KMF_BUFTAG) {
1595 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1596 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1597 			btp->bt_bufctl = bcp;
1598 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1599 			if (cache_flags & KMF_DEADBEEF) {
1600 				copy_pattern(KMEM_FREE_PATTERN, buf,
1601 				    cp->cache_verify);
1602 			}
1603 		}
1604 		bcp->bc_next = sp->slab_head;
1605 		sp->slab_head = bcp;
1606 		buf += chunksize;
1607 	}
1608 
1609 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1610 
1611 	return (sp);
1612 
1613 bufctl_alloc_failure:
1614 
1615 	while ((bcp = sp->slab_head) != NULL) {
1616 		sp->slab_head = bcp->bc_next;
1617 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1618 	}
1619 	kmem_cache_free(kmem_slab_cache, sp);
1620 
1621 slab_alloc_failure:
1622 
1623 	vmem_free(vmp, slab, slabsize);
1624 
1625 vmem_alloc_failure:
1626 
1627 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1628 	atomic_add_64(&cp->cache_alloc_fail, 1);
1629 
1630 	return (NULL);
1631 }
1632 
1633 /*
1634  * Destroy a slab.
1635  */
1636 static void
1637 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1638 {
1639 	vmem_t *vmp = cp->cache_arena;
1640 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1641 
1642 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1643 	ASSERT(sp->slab_refcnt == 0);
1644 
1645 	if (cp->cache_flags & KMF_HASH) {
1646 		kmem_bufctl_t *bcp;
1647 		while ((bcp = sp->slab_head) != NULL) {
1648 			sp->slab_head = bcp->bc_next;
1649 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1650 		}
1651 		kmem_cache_free(kmem_slab_cache, sp);
1652 	}
1653 	vmem_free(vmp, slab, cp->cache_slabsize);
1654 }
1655 
1656 static void *
1657 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1658 {
1659 	kmem_bufctl_t *bcp, **hash_bucket;
1660 	void *buf;
1661 	boolean_t new_slab = (sp->slab_refcnt == 0);
1662 
1663 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1664 	/*
1665 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1666 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1667 	 * slab is newly created.
1668 	 */
1669 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1670 	    (sp == avl_first(&cp->cache_partial_slabs))));
1671 	ASSERT(sp->slab_cache == cp);
1672 
1673 	cp->cache_slab_alloc++;
1674 	cp->cache_bufslab--;
1675 	sp->slab_refcnt++;
1676 
1677 	bcp = sp->slab_head;
1678 	sp->slab_head = bcp->bc_next;
1679 
1680 	if (cp->cache_flags & KMF_HASH) {
1681 		/*
1682 		 * Add buffer to allocated-address hash table.
1683 		 */
1684 		buf = bcp->bc_addr;
1685 		hash_bucket = KMEM_HASH(cp, buf);
1686 		bcp->bc_next = *hash_bucket;
1687 		*hash_bucket = bcp;
1688 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1689 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1690 		}
1691 	} else {
1692 		buf = KMEM_BUF(cp, bcp);
1693 	}
1694 
1695 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1696 
1697 	if (sp->slab_head == NULL) {
1698 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1699 		if (new_slab) {
1700 			ASSERT(sp->slab_chunks == 1);
1701 		} else {
1702 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1703 			avl_remove(&cp->cache_partial_slabs, sp);
1704 			sp->slab_later_count = 0; /* clear history */
1705 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1706 			sp->slab_stuck_offset = (uint32_t)-1;
1707 		}
1708 		list_insert_head(&cp->cache_complete_slabs, sp);
1709 		cp->cache_complete_slab_count++;
1710 		return (buf);
1711 	}
1712 
1713 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1714 	/*
1715 	 * Peek to see if the magazine layer is enabled before
1716 	 * we prefill.  We're not holding the cpu cache lock,
1717 	 * so the peek could be wrong, but there's no harm in it.
1718 	 */
1719 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1720 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1721 		kmem_slab_prefill(cp, sp);
1722 		return (buf);
1723 	}
1724 
1725 	if (new_slab) {
1726 		avl_add(&cp->cache_partial_slabs, sp);
1727 		return (buf);
1728 	}
1729 
1730 	/*
1731 	 * The slab is now more allocated than it was, so the
1732 	 * order remains unchanged.
1733 	 */
1734 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1735 	return (buf);
1736 }
1737 
1738 /*
1739  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1740  */
1741 static void *
1742 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1743 {
1744 	kmem_slab_t *sp;
1745 	void *buf;
1746 	boolean_t test_destructor;
1747 
1748 	mutex_enter(&cp->cache_lock);
1749 	test_destructor = (cp->cache_slab_alloc == 0);
1750 	sp = avl_first(&cp->cache_partial_slabs);
1751 	if (sp == NULL) {
1752 		ASSERT(cp->cache_bufslab == 0);
1753 
1754 		/*
1755 		 * The freelist is empty.  Create a new slab.
1756 		 */
1757 		mutex_exit(&cp->cache_lock);
1758 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1759 			return (NULL);
1760 		}
1761 		mutex_enter(&cp->cache_lock);
1762 		cp->cache_slab_create++;
1763 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1764 			cp->cache_bufmax = cp->cache_buftotal;
1765 		cp->cache_bufslab += sp->slab_chunks;
1766 	}
1767 
1768 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1769 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1770 	    (cp->cache_complete_slab_count +
1771 	    avl_numnodes(&cp->cache_partial_slabs) +
1772 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1773 	mutex_exit(&cp->cache_lock);
1774 
1775 	if (test_destructor && cp->cache_destructor != NULL) {
1776 		/*
1777 		 * On the first kmem_slab_alloc(), assert that it is valid to
1778 		 * call the destructor on a newly constructed object without any
1779 		 * client involvement.
1780 		 */
1781 		if ((cp->cache_constructor == NULL) ||
1782 		    cp->cache_constructor(buf, cp->cache_private,
1783 		    kmflag) == 0) {
1784 			cp->cache_destructor(buf, cp->cache_private);
1785 		}
1786 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1787 		    cp->cache_bufsize);
1788 		if (cp->cache_flags & KMF_DEADBEEF) {
1789 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1790 		}
1791 	}
1792 
1793 	return (buf);
1794 }
1795 
1796 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1797 
1798 /*
1799  * Free a raw (unconstructed) buffer to cp's slab layer.
1800  */
1801 static void
1802 kmem_slab_free(kmem_cache_t *cp, void *buf)
1803 {
1804 	kmem_slab_t *sp;
1805 	kmem_bufctl_t *bcp, **prev_bcpp;
1806 
1807 	ASSERT(buf != NULL);
1808 
1809 	mutex_enter(&cp->cache_lock);
1810 	cp->cache_slab_free++;
1811 
1812 	if (cp->cache_flags & KMF_HASH) {
1813 		/*
1814 		 * Look up buffer in allocated-address hash table.
1815 		 */
1816 		prev_bcpp = KMEM_HASH(cp, buf);
1817 		while ((bcp = *prev_bcpp) != NULL) {
1818 			if (bcp->bc_addr == buf) {
1819 				*prev_bcpp = bcp->bc_next;
1820 				sp = bcp->bc_slab;
1821 				break;
1822 			}
1823 			cp->cache_lookup_depth++;
1824 			prev_bcpp = &bcp->bc_next;
1825 		}
1826 	} else {
1827 		bcp = KMEM_BUFCTL(cp, buf);
1828 		sp = KMEM_SLAB(cp, buf);
1829 	}
1830 
1831 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1832 		mutex_exit(&cp->cache_lock);
1833 		kmem_error(KMERR_BADADDR, cp, buf);
1834 		return;
1835 	}
1836 
1837 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1838 		/*
1839 		 * If this is the buffer that prevented the consolidator from
1840 		 * clearing the slab, we can reset the slab flags now that the
1841 		 * buffer is freed. (It makes sense to do this in
1842 		 * kmem_cache_free(), where the client gives up ownership of the
1843 		 * buffer, but on the hot path the test is too expensive.)
1844 		 */
1845 		kmem_slab_move_yes(cp, sp, buf);
1846 	}
1847 
1848 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1849 		if (cp->cache_flags & KMF_CONTENTS)
1850 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1851 			    kmem_log_enter(kmem_content_log, buf,
1852 			    cp->cache_contents);
1853 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1854 	}
1855 
1856 	bcp->bc_next = sp->slab_head;
1857 	sp->slab_head = bcp;
1858 
1859 	cp->cache_bufslab++;
1860 	ASSERT(sp->slab_refcnt >= 1);
1861 
1862 	if (--sp->slab_refcnt == 0) {
1863 		/*
1864 		 * There are no outstanding allocations from this slab,
1865 		 * so we can reclaim the memory.
1866 		 */
1867 		if (sp->slab_chunks == 1) {
1868 			list_remove(&cp->cache_complete_slabs, sp);
1869 			cp->cache_complete_slab_count--;
1870 		} else {
1871 			avl_remove(&cp->cache_partial_slabs, sp);
1872 		}
1873 
1874 		cp->cache_buftotal -= sp->slab_chunks;
1875 		cp->cache_bufslab -= sp->slab_chunks;
1876 		/*
1877 		 * Defer releasing the slab to the virtual memory subsystem
1878 		 * while there is a pending move callback, since we guarantee
1879 		 * that buffers passed to the move callback have only been
1880 		 * touched by kmem or by the client itself. Since the memory
1881 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1882 		 * set at least one of the two lowest order bits, the client can
1883 		 * test those bits in the move callback to determine whether or
1884 		 * not it knows about the buffer (assuming that the client also
1885 		 * sets one of those low order bits whenever it frees a buffer).
1886 		 */
1887 		if (cp->cache_defrag == NULL ||
1888 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1889 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1890 			cp->cache_slab_destroy++;
1891 			mutex_exit(&cp->cache_lock);
1892 			kmem_slab_destroy(cp, sp);
1893 		} else {
1894 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1895 			/*
1896 			 * Slabs are inserted at both ends of the deadlist to
1897 			 * distinguish between slabs freed while move callbacks
1898 			 * are pending (list head) and a slab freed while the
1899 			 * lock is dropped in kmem_move_buffers() (list tail) so
1900 			 * that in both cases slab_destroy() is called from the
1901 			 * right context.
1902 			 */
1903 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1904 				list_insert_tail(deadlist, sp);
1905 			} else {
1906 				list_insert_head(deadlist, sp);
1907 			}
1908 			cp->cache_defrag->kmd_deadcount++;
1909 			mutex_exit(&cp->cache_lock);
1910 		}
1911 		return;
1912 	}
1913 
1914 	if (bcp->bc_next == NULL) {
1915 		/* Transition the slab from completely allocated to partial. */
1916 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1917 		ASSERT(sp->slab_chunks > 1);
1918 		list_remove(&cp->cache_complete_slabs, sp);
1919 		cp->cache_complete_slab_count--;
1920 		avl_add(&cp->cache_partial_slabs, sp);
1921 	} else {
1922 #ifdef	DEBUG
1923 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1924 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1925 		} else {
1926 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1927 		}
1928 #else
1929 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1930 #endif
1931 	}
1932 
1933 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1934 	    (cp->cache_complete_slab_count +
1935 	    avl_numnodes(&cp->cache_partial_slabs) +
1936 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1937 	mutex_exit(&cp->cache_lock);
1938 }
1939 
1940 /*
1941  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1942  */
1943 static int
1944 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1945     caddr_t caller)
1946 {
1947 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1948 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1949 	uint32_t mtbf;
1950 
1951 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1952 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1953 		return (-1);
1954 	}
1955 
1956 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1957 
1958 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1959 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1960 		return (-1);
1961 	}
1962 
1963 	if (cp->cache_flags & KMF_DEADBEEF) {
1964 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1965 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1966 				kmem_error(KMERR_MODIFIED, cp, buf);
1967 				return (-1);
1968 			}
1969 			if (cp->cache_constructor != NULL)
1970 				*(uint64_t *)buf = btp->bt_redzone;
1971 			else
1972 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1973 		} else {
1974 			construct = 1;
1975 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1976 			    KMEM_UNINITIALIZED_PATTERN, buf,
1977 			    cp->cache_verify)) {
1978 				kmem_error(KMERR_MODIFIED, cp, buf);
1979 				return (-1);
1980 			}
1981 		}
1982 	}
1983 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1984 
1985 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1986 	    gethrtime() % mtbf == 0 &&
1987 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1988 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1989 		if (!construct && cp->cache_destructor != NULL)
1990 			cp->cache_destructor(buf, cp->cache_private);
1991 	} else {
1992 		mtbf = 0;
1993 	}
1994 
1995 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1996 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1997 		atomic_add_64(&cp->cache_alloc_fail, 1);
1998 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1999 		if (cp->cache_flags & KMF_DEADBEEF)
2000 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2001 		kmem_slab_free(cp, buf);
2002 		return (1);
2003 	}
2004 
2005 	if (cp->cache_flags & KMF_AUDIT) {
2006 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2007 	}
2008 
2009 	if ((cp->cache_flags & KMF_LITE) &&
2010 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2011 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2012 	}
2013 
2014 	return (0);
2015 }
2016 
2017 static int
2018 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2019 {
2020 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2021 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2022 	kmem_slab_t *sp;
2023 
2024 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2025 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2026 			kmem_error(KMERR_DUPFREE, cp, buf);
2027 			return (-1);
2028 		}
2029 		sp = kmem_findslab(cp, buf);
2030 		if (sp == NULL || sp->slab_cache != cp)
2031 			kmem_error(KMERR_BADADDR, cp, buf);
2032 		else
2033 			kmem_error(KMERR_REDZONE, cp, buf);
2034 		return (-1);
2035 	}
2036 
2037 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2038 
2039 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2040 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2041 		return (-1);
2042 	}
2043 
2044 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2045 		kmem_error(KMERR_REDZONE, cp, buf);
2046 		return (-1);
2047 	}
2048 
2049 	if (cp->cache_flags & KMF_AUDIT) {
2050 		if (cp->cache_flags & KMF_CONTENTS)
2051 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2052 			    buf, cp->cache_contents);
2053 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2054 	}
2055 
2056 	if ((cp->cache_flags & KMF_LITE) &&
2057 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2058 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2059 	}
2060 
2061 	if (cp->cache_flags & KMF_DEADBEEF) {
2062 		if (cp->cache_flags & KMF_LITE)
2063 			btp->bt_redzone = *(uint64_t *)buf;
2064 		else if (cp->cache_destructor != NULL)
2065 			cp->cache_destructor(buf, cp->cache_private);
2066 
2067 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2068 	}
2069 
2070 	return (0);
2071 }
2072 
2073 /*
2074  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2075  */
2076 static void
2077 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2078 {
2079 	int round;
2080 
2081 	ASSERT(!list_link_active(&cp->cache_link) ||
2082 	    taskq_member(kmem_taskq, curthread));
2083 
2084 	for (round = 0; round < nrounds; round++) {
2085 		void *buf = mp->mag_round[round];
2086 
2087 		if (cp->cache_flags & KMF_DEADBEEF) {
2088 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2089 			    cp->cache_verify) != NULL) {
2090 				kmem_error(KMERR_MODIFIED, cp, buf);
2091 				continue;
2092 			}
2093 			if ((cp->cache_flags & KMF_LITE) &&
2094 			    cp->cache_destructor != NULL) {
2095 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2096 				*(uint64_t *)buf = btp->bt_redzone;
2097 				cp->cache_destructor(buf, cp->cache_private);
2098 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2099 			}
2100 		} else if (cp->cache_destructor != NULL) {
2101 			cp->cache_destructor(buf, cp->cache_private);
2102 		}
2103 
2104 		kmem_slab_free(cp, buf);
2105 	}
2106 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2107 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2108 }
2109 
2110 /*
2111  * Allocate a magazine from the depot.
2112  */
2113 static kmem_magazine_t *
2114 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2115 {
2116 	kmem_magazine_t *mp;
2117 
2118 	/*
2119 	 * If we can't get the depot lock without contention,
2120 	 * update our contention count.  We use the depot
2121 	 * contention rate to determine whether we need to
2122 	 * increase the magazine size for better scalability.
2123 	 */
2124 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2125 		mutex_enter(&cp->cache_depot_lock);
2126 		cp->cache_depot_contention++;
2127 	}
2128 
2129 	if ((mp = mlp->ml_list) != NULL) {
2130 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2131 		mlp->ml_list = mp->mag_next;
2132 		if (--mlp->ml_total < mlp->ml_min)
2133 			mlp->ml_min = mlp->ml_total;
2134 		mlp->ml_alloc++;
2135 	}
2136 
2137 	mutex_exit(&cp->cache_depot_lock);
2138 
2139 	return (mp);
2140 }
2141 
2142 /*
2143  * Free a magazine to the depot.
2144  */
2145 static void
2146 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2147 {
2148 	mutex_enter(&cp->cache_depot_lock);
2149 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2150 	mp->mag_next = mlp->ml_list;
2151 	mlp->ml_list = mp;
2152 	mlp->ml_total++;
2153 	mutex_exit(&cp->cache_depot_lock);
2154 }
2155 
2156 /*
2157  * Update the working set statistics for cp's depot.
2158  */
2159 static void
2160 kmem_depot_ws_update(kmem_cache_t *cp)
2161 {
2162 	mutex_enter(&cp->cache_depot_lock);
2163 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2164 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2165 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2166 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2167 	mutex_exit(&cp->cache_depot_lock);
2168 }
2169 
2170 /*
2171  * Reap all magazines that have fallen out of the depot's working set.
2172  */
2173 static void
2174 kmem_depot_ws_reap(kmem_cache_t *cp)
2175 {
2176 	long reap;
2177 	kmem_magazine_t *mp;
2178 
2179 	ASSERT(!list_link_active(&cp->cache_link) ||
2180 	    taskq_member(kmem_taskq, curthread));
2181 
2182 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2183 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2184 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2185 
2186 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2187 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2188 		kmem_magazine_destroy(cp, mp, 0);
2189 }
2190 
2191 static void
2192 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2193 {
2194 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2195 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2196 	ASSERT(ccp->cc_magsize > 0);
2197 
2198 	ccp->cc_ploaded = ccp->cc_loaded;
2199 	ccp->cc_prounds = ccp->cc_rounds;
2200 	ccp->cc_loaded = mp;
2201 	ccp->cc_rounds = rounds;
2202 }
2203 
2204 /*
2205  * Intercept kmem alloc/free calls during crash dump in order to avoid
2206  * changing kmem state while memory is being saved to the dump device.
2207  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2208  * there are no locks because only one CPU calls kmem during a crash
2209  * dump. To enable this feature, first create the associated vmem
2210  * arena with VMC_DUMPSAFE.
2211  */
2212 static void *kmem_dump_start;	/* start of pre-reserved heap */
2213 static void *kmem_dump_end;	/* end of heap area */
2214 static void *kmem_dump_curr;	/* current free heap pointer */
2215 static size_t kmem_dump_size;	/* size of heap area */
2216 
2217 /* append to each buf created in the pre-reserved heap */
2218 typedef struct kmem_dumpctl {
2219 	void	*kdc_next;	/* cache dump free list linkage */
2220 } kmem_dumpctl_t;
2221 
2222 #define	KMEM_DUMPCTL(cp, buf)	\
2223 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2224 	    sizeof (void *)))
2225 
2226 /* Keep some simple stats. */
2227 #define	KMEM_DUMP_LOGS	(100)
2228 
2229 typedef struct kmem_dump_log {
2230 	kmem_cache_t	*kdl_cache;
2231 	uint_t		kdl_allocs;		/* # of dump allocations */
2232 	uint_t		kdl_frees;		/* # of dump frees */
2233 	uint_t		kdl_alloc_fails;	/* # of allocation failures */
2234 	uint_t		kdl_free_nondump;	/* # of non-dump frees */
2235 	uint_t		kdl_unsafe;		/* cache was used, but unsafe */
2236 } kmem_dump_log_t;
2237 
2238 static kmem_dump_log_t *kmem_dump_log;
2239 static int kmem_dump_log_idx;
2240 
2241 #define	KDI_LOG(cp, stat) {						\
2242 	kmem_dump_log_t *kdl;						\
2243 	if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) {	\
2244 		kdl->stat++;						\
2245 	} else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) {		\
2246 		kdl = &kmem_dump_log[kmem_dump_log_idx++];		\
2247 		kdl->stat++;						\
2248 		kdl->kdl_cache = (cp);					\
2249 		(cp)->cache_dumplog = kdl;				\
2250 	}								\
2251 }
2252 
2253 /* set non zero for full report */
2254 uint_t kmem_dump_verbose = 0;
2255 
2256 /* stats for overize heap */
2257 uint_t kmem_dump_oversize_allocs = 0;
2258 uint_t kmem_dump_oversize_max = 0;
2259 
2260 static void
2261 kmem_dumppr(char **pp, char *e, const char *format, ...)
2262 {
2263 	char *p = *pp;
2264 
2265 	if (p < e) {
2266 		int n;
2267 		va_list ap;
2268 
2269 		va_start(ap, format);
2270 		n = vsnprintf(p, e - p, format, ap);
2271 		va_end(ap);
2272 		*pp = p + n;
2273 	}
2274 }
2275 
2276 /*
2277  * Called when dumpadm(1M) configures dump parameters.
2278  */
2279 void
2280 kmem_dump_init(size_t size)
2281 {
2282 	if (kmem_dump_start != NULL)
2283 		kmem_free(kmem_dump_start, kmem_dump_size);
2284 
2285 	if (kmem_dump_log == NULL)
2286 		kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2287 		    sizeof (kmem_dump_log_t), KM_SLEEP);
2288 
2289 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2290 
2291 	if (kmem_dump_start != NULL) {
2292 		kmem_dump_size = size;
2293 		kmem_dump_curr = kmem_dump_start;
2294 		kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2295 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2296 	} else {
2297 		kmem_dump_size = 0;
2298 		kmem_dump_curr = NULL;
2299 		kmem_dump_end = NULL;
2300 	}
2301 }
2302 
2303 /*
2304  * Set flag for each kmem_cache_t if is safe to use alternate dump
2305  * memory. Called just before panic crash dump starts. Set the flag
2306  * for the calling CPU.
2307  */
2308 void
2309 kmem_dump_begin(void)
2310 {
2311 	ASSERT(panicstr != NULL);
2312 	if (kmem_dump_start != NULL) {
2313 		kmem_cache_t *cp;
2314 
2315 		for (cp = list_head(&kmem_caches); cp != NULL;
2316 		    cp = list_next(&kmem_caches, cp)) {
2317 			kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2318 
2319 			if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2320 				cp->cache_flags |= KMF_DUMPDIVERT;
2321 				ccp->cc_flags |= KMF_DUMPDIVERT;
2322 				ccp->cc_dump_rounds = ccp->cc_rounds;
2323 				ccp->cc_dump_prounds = ccp->cc_prounds;
2324 				ccp->cc_rounds = ccp->cc_prounds = -1;
2325 			} else {
2326 				cp->cache_flags |= KMF_DUMPUNSAFE;
2327 				ccp->cc_flags |= KMF_DUMPUNSAFE;
2328 			}
2329 		}
2330 	}
2331 }
2332 
2333 /*
2334  * finished dump intercept
2335  * print any warnings on the console
2336  * return verbose information to dumpsys() in the given buffer
2337  */
2338 size_t
2339 kmem_dump_finish(char *buf, size_t size)
2340 {
2341 	int kdi_idx;
2342 	int kdi_end = kmem_dump_log_idx;
2343 	int percent = 0;
2344 	int header = 0;
2345 	int warn = 0;
2346 	size_t used;
2347 	kmem_cache_t *cp;
2348 	kmem_dump_log_t *kdl;
2349 	char *e = buf + size;
2350 	char *p = buf;
2351 
2352 	if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2353 		return (0);
2354 
2355 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2356 	percent = (used * 100) / kmem_dump_size;
2357 
2358 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2359 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2360 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2361 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2362 	    kmem_dump_oversize_allocs);
2363 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2364 	    kmem_dump_oversize_max);
2365 
2366 	for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2367 		kdl = &kmem_dump_log[kdi_idx];
2368 		cp = kdl->kdl_cache;
2369 		if (cp == NULL)
2370 			break;
2371 		if (kdl->kdl_alloc_fails)
2372 			++warn;
2373 		if (header == 0) {
2374 			kmem_dumppr(&p, e,
2375 			    "Cache Name,Allocs,Frees,Alloc Fails,"
2376 			    "Nondump Frees,Unsafe Allocs/Frees\n");
2377 			header = 1;
2378 		}
2379 		kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2380 		    cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2381 		    kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2382 		    kdl->kdl_unsafe);
2383 	}
2384 
2385 	/* return buffer size used */
2386 	if (p < e)
2387 		bzero(p, e - p);
2388 	return (p - buf);
2389 }
2390 
2391 /*
2392  * Allocate a constructed object from alternate dump memory.
2393  */
2394 void *
2395 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2396 {
2397 	void *buf;
2398 	void *curr;
2399 	char *bufend;
2400 
2401 	/* return a constructed object */
2402 	if ((buf = cp->cache_dumpfreelist) != NULL) {
2403 		cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2404 		KDI_LOG(cp, kdl_allocs);
2405 		return (buf);
2406 	}
2407 
2408 	/* create a new constructed object */
2409 	curr = kmem_dump_curr;
2410 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2411 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2412 
2413 	/* hat layer objects cannot cross a page boundary */
2414 	if (cp->cache_align < PAGESIZE) {
2415 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2416 		if (bufend > page) {
2417 			bufend += page - (char *)buf;
2418 			buf = (void *)page;
2419 		}
2420 	}
2421 
2422 	/* fall back to normal alloc if reserved area is used up */
2423 	if (bufend > (char *)kmem_dump_end) {
2424 		kmem_dump_curr = kmem_dump_end;
2425 		KDI_LOG(cp, kdl_alloc_fails);
2426 		return (NULL);
2427 	}
2428 
2429 	/*
2430 	 * Must advance curr pointer before calling a constructor that
2431 	 * may also allocate memory.
2432 	 */
2433 	kmem_dump_curr = bufend;
2434 
2435 	/* run constructor */
2436 	if (cp->cache_constructor != NULL &&
2437 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2438 	    != 0) {
2439 #ifdef DEBUG
2440 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2441 		    cp->cache_name, (void *)cp);
2442 #endif
2443 		/* reset curr pointer iff no allocs were done */
2444 		if (kmem_dump_curr == bufend)
2445 			kmem_dump_curr = curr;
2446 
2447 		/* fall back to normal alloc if the constructor fails */
2448 		KDI_LOG(cp, kdl_alloc_fails);
2449 		return (NULL);
2450 	}
2451 
2452 	KDI_LOG(cp, kdl_allocs);
2453 	return (buf);
2454 }
2455 
2456 /*
2457  * Free a constructed object in alternate dump memory.
2458  */
2459 int
2460 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2461 {
2462 	/* save constructed buffers for next time */
2463 	if ((char *)buf >= (char *)kmem_dump_start &&
2464 	    (char *)buf < (char *)kmem_dump_end) {
2465 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2466 		cp->cache_dumpfreelist = buf;
2467 		KDI_LOG(cp, kdl_frees);
2468 		return (0);
2469 	}
2470 
2471 	/* count all non-dump buf frees */
2472 	KDI_LOG(cp, kdl_free_nondump);
2473 
2474 	/* just drop buffers that were allocated before dump started */
2475 	if (kmem_dump_curr < kmem_dump_end)
2476 		return (0);
2477 
2478 	/* fall back to normal free if reserved area is used up */
2479 	return (1);
2480 }
2481 
2482 /*
2483  * Allocate a constructed object from cache cp.
2484  */
2485 void *
2486 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2487 {
2488 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2489 	kmem_magazine_t *fmp;
2490 	void *buf;
2491 
2492 	mutex_enter(&ccp->cc_lock);
2493 	for (;;) {
2494 		/*
2495 		 * If there's an object available in the current CPU's
2496 		 * loaded magazine, just take it and return.
2497 		 */
2498 		if (ccp->cc_rounds > 0) {
2499 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2500 			ccp->cc_alloc++;
2501 			mutex_exit(&ccp->cc_lock);
2502 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2503 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2504 					ASSERT(!(ccp->cc_flags &
2505 					    KMF_DUMPDIVERT));
2506 					KDI_LOG(cp, kdl_unsafe);
2507 				}
2508 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2509 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2510 				    caller()) != 0) {
2511 					if (kmflag & KM_NOSLEEP)
2512 						return (NULL);
2513 					mutex_enter(&ccp->cc_lock);
2514 					continue;
2515 				}
2516 			}
2517 			return (buf);
2518 		}
2519 
2520 		/*
2521 		 * The loaded magazine is empty.  If the previously loaded
2522 		 * magazine was full, exchange them and try again.
2523 		 */
2524 		if (ccp->cc_prounds > 0) {
2525 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2526 			continue;
2527 		}
2528 
2529 		/*
2530 		 * Return an alternate buffer at dump time to preserve
2531 		 * the heap.
2532 		 */
2533 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2534 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2535 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2536 				/* log it so that we can warn about it */
2537 				KDI_LOG(cp, kdl_unsafe);
2538 			} else {
2539 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2540 				    NULL) {
2541 					mutex_exit(&ccp->cc_lock);
2542 					return (buf);
2543 				}
2544 				break;		/* fall back to slab layer */
2545 			}
2546 		}
2547 
2548 		/*
2549 		 * If the magazine layer is disabled, break out now.
2550 		 */
2551 		if (ccp->cc_magsize == 0)
2552 			break;
2553 
2554 		/*
2555 		 * Try to get a full magazine from the depot.
2556 		 */
2557 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2558 		if (fmp != NULL) {
2559 			if (ccp->cc_ploaded != NULL)
2560 				kmem_depot_free(cp, &cp->cache_empty,
2561 				    ccp->cc_ploaded);
2562 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2563 			continue;
2564 		}
2565 
2566 		/*
2567 		 * There are no full magazines in the depot,
2568 		 * so fall through to the slab layer.
2569 		 */
2570 		break;
2571 	}
2572 	mutex_exit(&ccp->cc_lock);
2573 
2574 	/*
2575 	 * We couldn't allocate a constructed object from the magazine layer,
2576 	 * so get a raw buffer from the slab layer and apply its constructor.
2577 	 */
2578 	buf = kmem_slab_alloc(cp, kmflag);
2579 
2580 	if (buf == NULL)
2581 		return (NULL);
2582 
2583 	if (cp->cache_flags & KMF_BUFTAG) {
2584 		/*
2585 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2586 		 */
2587 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2588 		if (rc != 0) {
2589 			if (kmflag & KM_NOSLEEP)
2590 				return (NULL);
2591 			/*
2592 			 * kmem_cache_alloc_debug() detected corruption
2593 			 * but didn't panic (kmem_panic <= 0). We should not be
2594 			 * here because the constructor failed (indicated by a
2595 			 * return code of 1). Try again.
2596 			 */
2597 			ASSERT(rc == -1);
2598 			return (kmem_cache_alloc(cp, kmflag));
2599 		}
2600 		return (buf);
2601 	}
2602 
2603 	if (cp->cache_constructor != NULL &&
2604 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2605 		atomic_add_64(&cp->cache_alloc_fail, 1);
2606 		kmem_slab_free(cp, buf);
2607 		return (NULL);
2608 	}
2609 
2610 	return (buf);
2611 }
2612 
2613 /*
2614  * The freed argument tells whether or not kmem_cache_free_debug() has already
2615  * been called so that we can avoid the duplicate free error. For example, a
2616  * buffer on a magazine has already been freed by the client but is still
2617  * constructed.
2618  */
2619 static void
2620 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2621 {
2622 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2623 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2624 			return;
2625 
2626 	/*
2627 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2628 	 * kmem_cache_free_debug() will have already applied the destructor.
2629 	 */
2630 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2631 	    cp->cache_destructor != NULL) {
2632 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2633 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2634 			*(uint64_t *)buf = btp->bt_redzone;
2635 			cp->cache_destructor(buf, cp->cache_private);
2636 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2637 		} else {
2638 			cp->cache_destructor(buf, cp->cache_private);
2639 		}
2640 	}
2641 
2642 	kmem_slab_free(cp, buf);
2643 }
2644 
2645 /*
2646  * Used when there's no room to free a buffer to the per-CPU cache.
2647  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2648  * caller should try freeing to the per-CPU cache again.
2649  * Note that we don't directly install the magazine in the cpu cache,
2650  * since its state may have changed wildly while the lock was dropped.
2651  */
2652 static int
2653 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2654 {
2655 	kmem_magazine_t *emp;
2656 	kmem_magtype_t *mtp;
2657 
2658 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2659 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2660 	    ((uint_t)ccp->cc_rounds == -1)) &&
2661 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2662 	    ((uint_t)ccp->cc_prounds == -1)));
2663 
2664 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2665 	if (emp != NULL) {
2666 		if (ccp->cc_ploaded != NULL)
2667 			kmem_depot_free(cp, &cp->cache_full,
2668 			    ccp->cc_ploaded);
2669 		kmem_cpu_reload(ccp, emp, 0);
2670 		return (1);
2671 	}
2672 	/*
2673 	 * There are no empty magazines in the depot,
2674 	 * so try to allocate a new one.  We must drop all locks
2675 	 * across kmem_cache_alloc() because lower layers may
2676 	 * attempt to allocate from this cache.
2677 	 */
2678 	mtp = cp->cache_magtype;
2679 	mutex_exit(&ccp->cc_lock);
2680 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2681 	mutex_enter(&ccp->cc_lock);
2682 
2683 	if (emp != NULL) {
2684 		/*
2685 		 * We successfully allocated an empty magazine.
2686 		 * However, we had to drop ccp->cc_lock to do it,
2687 		 * so the cache's magazine size may have changed.
2688 		 * If so, free the magazine and try again.
2689 		 */
2690 		if (ccp->cc_magsize != mtp->mt_magsize) {
2691 			mutex_exit(&ccp->cc_lock);
2692 			kmem_cache_free(mtp->mt_cache, emp);
2693 			mutex_enter(&ccp->cc_lock);
2694 			return (1);
2695 		}
2696 
2697 		/*
2698 		 * We got a magazine of the right size.  Add it to
2699 		 * the depot and try the whole dance again.
2700 		 */
2701 		kmem_depot_free(cp, &cp->cache_empty, emp);
2702 		return (1);
2703 	}
2704 
2705 	/*
2706 	 * We couldn't allocate an empty magazine,
2707 	 * so fall through to the slab layer.
2708 	 */
2709 	return (0);
2710 }
2711 
2712 /*
2713  * Free a constructed object to cache cp.
2714  */
2715 void
2716 kmem_cache_free(kmem_cache_t *cp, void *buf)
2717 {
2718 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2719 
2720 	/*
2721 	 * The client must not free either of the buffers passed to the move
2722 	 * callback function.
2723 	 */
2724 	ASSERT(cp->cache_defrag == NULL ||
2725 	    cp->cache_defrag->kmd_thread != curthread ||
2726 	    (buf != cp->cache_defrag->kmd_from_buf &&
2727 	    buf != cp->cache_defrag->kmd_to_buf));
2728 
2729 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2730 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2731 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2732 			/* log it so that we can warn about it */
2733 			KDI_LOG(cp, kdl_unsafe);
2734 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2735 			return;
2736 		}
2737 		if (ccp->cc_flags & KMF_BUFTAG) {
2738 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2739 				return;
2740 		}
2741 	}
2742 
2743 	mutex_enter(&ccp->cc_lock);
2744 	/*
2745 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2746 	 */
2747 	for (;;) {
2748 		/*
2749 		 * If there's a slot available in the current CPU's
2750 		 * loaded magazine, just put the object there and return.
2751 		 */
2752 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2753 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2754 			ccp->cc_free++;
2755 			mutex_exit(&ccp->cc_lock);
2756 			return;
2757 		}
2758 
2759 		/*
2760 		 * The loaded magazine is full.  If the previously loaded
2761 		 * magazine was empty, exchange them and try again.
2762 		 */
2763 		if (ccp->cc_prounds == 0) {
2764 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2765 			continue;
2766 		}
2767 
2768 		/*
2769 		 * If the magazine layer is disabled, break out now.
2770 		 */
2771 		if (ccp->cc_magsize == 0)
2772 			break;
2773 
2774 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2775 			/*
2776 			 * We couldn't free our constructed object to the
2777 			 * magazine layer, so apply its destructor and free it
2778 			 * to the slab layer.
2779 			 */
2780 			break;
2781 		}
2782 	}
2783 	mutex_exit(&ccp->cc_lock);
2784 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2785 }
2786 
2787 static void
2788 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2789 {
2790 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2791 	int cache_flags = cp->cache_flags;
2792 
2793 	kmem_bufctl_t *next, *head;
2794 	size_t nbufs;
2795 
2796 	/*
2797 	 * Completely allocate the newly created slab and put the pre-allocated
2798 	 * buffers in magazines. Any of the buffers that cannot be put in
2799 	 * magazines must be returned to the slab.
2800 	 */
2801 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2802 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2803 	ASSERT(cp->cache_constructor == NULL);
2804 	ASSERT(sp->slab_cache == cp);
2805 	ASSERT(sp->slab_refcnt == 1);
2806 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2807 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2808 
2809 	head = sp->slab_head;
2810 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2811 	sp->slab_head = NULL;
2812 	sp->slab_refcnt += nbufs;
2813 	cp->cache_bufslab -= nbufs;
2814 	cp->cache_slab_alloc += nbufs;
2815 	list_insert_head(&cp->cache_complete_slabs, sp);
2816 	cp->cache_complete_slab_count++;
2817 	mutex_exit(&cp->cache_lock);
2818 	mutex_enter(&ccp->cc_lock);
2819 
2820 	while (head != NULL) {
2821 		void *buf = KMEM_BUF(cp, head);
2822 		/*
2823 		 * If there's a slot available in the current CPU's
2824 		 * loaded magazine, just put the object there and
2825 		 * continue.
2826 		 */
2827 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2828 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2829 			    buf;
2830 			ccp->cc_free++;
2831 			nbufs--;
2832 			head = head->bc_next;
2833 			continue;
2834 		}
2835 
2836 		/*
2837 		 * The loaded magazine is full.  If the previously
2838 		 * loaded magazine was empty, exchange them and try
2839 		 * again.
2840 		 */
2841 		if (ccp->cc_prounds == 0) {
2842 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2843 			    ccp->cc_prounds);
2844 			continue;
2845 		}
2846 
2847 		/*
2848 		 * If the magazine layer is disabled, break out now.
2849 		 */
2850 
2851 		if (ccp->cc_magsize == 0) {
2852 			break;
2853 		}
2854 
2855 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2856 			break;
2857 	}
2858 	mutex_exit(&ccp->cc_lock);
2859 	if (nbufs != 0) {
2860 		ASSERT(head != NULL);
2861 
2862 		/*
2863 		 * If there was a failure, return remaining objects to
2864 		 * the slab
2865 		 */
2866 		while (head != NULL) {
2867 			ASSERT(nbufs != 0);
2868 			next = head->bc_next;
2869 			head->bc_next = NULL;
2870 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2871 			head = next;
2872 			nbufs--;
2873 		}
2874 	}
2875 	ASSERT(head == NULL);
2876 	ASSERT(nbufs == 0);
2877 	mutex_enter(&cp->cache_lock);
2878 }
2879 
2880 void *
2881 kmem_zalloc(size_t size, int kmflag)
2882 {
2883 	size_t index;
2884 	void *buf;
2885 
2886 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2887 		kmem_cache_t *cp = kmem_alloc_table[index];
2888 		buf = kmem_cache_alloc(cp, kmflag);
2889 		if (buf != NULL) {
2890 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2891 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2892 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2893 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2894 
2895 				if (cp->cache_flags & KMF_LITE) {
2896 					KMEM_BUFTAG_LITE_ENTER(btp,
2897 					    kmem_lite_count, caller());
2898 				}
2899 			}
2900 			bzero(buf, size);
2901 		}
2902 	} else {
2903 		buf = kmem_alloc(size, kmflag);
2904 		if (buf != NULL)
2905 			bzero(buf, size);
2906 	}
2907 	return (buf);
2908 }
2909 
2910 void *
2911 kmem_alloc(size_t size, int kmflag)
2912 {
2913 	size_t index;
2914 	kmem_cache_t *cp;
2915 	void *buf;
2916 
2917 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2918 		cp = kmem_alloc_table[index];
2919 		/* fall through to kmem_cache_alloc() */
2920 
2921 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2922 	    kmem_big_alloc_table_max) {
2923 		cp = kmem_big_alloc_table[index];
2924 		/* fall through to kmem_cache_alloc() */
2925 
2926 	} else {
2927 		if (size == 0)
2928 			return (NULL);
2929 
2930 		buf = vmem_alloc(kmem_oversize_arena, size,
2931 		    kmflag & KM_VMFLAGS);
2932 		if (buf == NULL)
2933 			kmem_log_event(kmem_failure_log, NULL, NULL,
2934 			    (void *)size);
2935 		else if (KMEM_DUMP(kmem_slab_cache)) {
2936 			/* stats for dump intercept */
2937 			kmem_dump_oversize_allocs++;
2938 			if (size > kmem_dump_oversize_max)
2939 				kmem_dump_oversize_max = size;
2940 		}
2941 		return (buf);
2942 	}
2943 
2944 	buf = kmem_cache_alloc(cp, kmflag);
2945 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2946 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2947 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2948 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2949 
2950 		if (cp->cache_flags & KMF_LITE) {
2951 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2952 		}
2953 	}
2954 	return (buf);
2955 }
2956 
2957 void
2958 kmem_free(void *buf, size_t size)
2959 {
2960 	size_t index;
2961 	kmem_cache_t *cp;
2962 
2963 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2964 		cp = kmem_alloc_table[index];
2965 		/* fall through to kmem_cache_free() */
2966 
2967 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2968 	    kmem_big_alloc_table_max) {
2969 		cp = kmem_big_alloc_table[index];
2970 		/* fall through to kmem_cache_free() */
2971 
2972 	} else {
2973 		if (buf == NULL && size == 0)
2974 			return;
2975 		vmem_free(kmem_oversize_arena, buf, size);
2976 		return;
2977 	}
2978 
2979 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2980 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2981 		uint32_t *ip = (uint32_t *)btp;
2982 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2983 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2984 				kmem_error(KMERR_DUPFREE, cp, buf);
2985 				return;
2986 			}
2987 			if (KMEM_SIZE_VALID(ip[1])) {
2988 				ip[0] = KMEM_SIZE_ENCODE(size);
2989 				kmem_error(KMERR_BADSIZE, cp, buf);
2990 			} else {
2991 				kmem_error(KMERR_REDZONE, cp, buf);
2992 			}
2993 			return;
2994 		}
2995 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2996 			kmem_error(KMERR_REDZONE, cp, buf);
2997 			return;
2998 		}
2999 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
3000 		if (cp->cache_flags & KMF_LITE) {
3001 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3002 			    caller());
3003 		}
3004 	}
3005 	kmem_cache_free(cp, buf);
3006 }
3007 
3008 void *
3009 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3010 {
3011 	size_t realsize = size + vmp->vm_quantum;
3012 	void *addr;
3013 
3014 	/*
3015 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3016 	 * vm_quantum will cause integer wraparound.  Check for this, and
3017 	 * blow off the firewall page in this case.  Note that such a
3018 	 * giant allocation (the entire kernel address space) can never
3019 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3020 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
3021 	 * corresponding check in kmem_firewall_va_free().
3022 	 */
3023 	if (realsize < size)
3024 		realsize = size;
3025 
3026 	/*
3027 	 * While boot still owns resource management, make sure that this
3028 	 * redzone virtual address allocation is properly accounted for in
3029 	 * OBPs "virtual-memory" "available" lists because we're
3030 	 * effectively claiming them for a red zone.  If we don't do this,
3031 	 * the available lists become too fragmented and too large for the
3032 	 * current boot/kernel memory list interface.
3033 	 */
3034 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3035 
3036 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3037 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3038 
3039 	return (addr);
3040 }
3041 
3042 void
3043 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3044 {
3045 	ASSERT((kvseg.s_base == NULL ?
3046 	    va_to_pfn((char *)addr + size) :
3047 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3048 
3049 	vmem_free(vmp, addr, size + vmp->vm_quantum);
3050 }
3051 
3052 /*
3053  * Try to allocate at least `size' bytes of memory without sleeping or
3054  * panicking. Return actual allocated size in `asize'. If allocation failed,
3055  * try final allocation with sleep or panic allowed.
3056  */
3057 void *
3058 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3059 {
3060 	void *p;
3061 
3062 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3063 	do {
3064 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3065 		if (p != NULL)
3066 			return (p);
3067 		*asize += KMEM_ALIGN;
3068 	} while (*asize <= PAGESIZE);
3069 
3070 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3071 	return (kmem_alloc(*asize, kmflag));
3072 }
3073 
3074 /*
3075  * Reclaim all unused memory from a cache.
3076  */
3077 static void
3078 kmem_cache_reap(kmem_cache_t *cp)
3079 {
3080 	ASSERT(taskq_member(kmem_taskq, curthread));
3081 	cp->cache_reap++;
3082 
3083 	/*
3084 	 * Ask the cache's owner to free some memory if possible.
3085 	 * The idea is to handle things like the inode cache, which
3086 	 * typically sits on a bunch of memory that it doesn't truly
3087 	 * *need*.  Reclaim policy is entirely up to the owner; this
3088 	 * callback is just an advisory plea for help.
3089 	 */
3090 	if (cp->cache_reclaim != NULL) {
3091 		long delta;
3092 
3093 		/*
3094 		 * Reclaimed memory should be reapable (not included in the
3095 		 * depot's working set).
3096 		 */
3097 		delta = cp->cache_full.ml_total;
3098 		cp->cache_reclaim(cp->cache_private);
3099 		delta = cp->cache_full.ml_total - delta;
3100 		if (delta > 0) {
3101 			mutex_enter(&cp->cache_depot_lock);
3102 			cp->cache_full.ml_reaplimit += delta;
3103 			cp->cache_full.ml_min += delta;
3104 			mutex_exit(&cp->cache_depot_lock);
3105 		}
3106 	}
3107 
3108 	kmem_depot_ws_reap(cp);
3109 
3110 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3111 		kmem_cache_defrag(cp);
3112 	}
3113 }
3114 
3115 static void
3116 kmem_reap_timeout(void *flag_arg)
3117 {
3118 	uint32_t *flag = (uint32_t *)flag_arg;
3119 
3120 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3121 	*flag = 0;
3122 }
3123 
3124 static void
3125 kmem_reap_done(void *flag)
3126 {
3127 	(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3128 }
3129 
3130 static void
3131 kmem_reap_start(void *flag)
3132 {
3133 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3134 
3135 	if (flag == &kmem_reaping) {
3136 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3137 		/*
3138 		 * if we have segkp under heap, reap segkp cache.
3139 		 */
3140 		if (segkp_fromheap)
3141 			segkp_cache_free();
3142 	}
3143 	else
3144 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3145 
3146 	/*
3147 	 * We use taskq_dispatch() to schedule a timeout to clear
3148 	 * the flag so that kmem_reap() becomes self-throttling:
3149 	 * we won't reap again until the current reap completes *and*
3150 	 * at least kmem_reap_interval ticks have elapsed.
3151 	 */
3152 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3153 		kmem_reap_done(flag);
3154 }
3155 
3156 static void
3157 kmem_reap_common(void *flag_arg)
3158 {
3159 	uint32_t *flag = (uint32_t *)flag_arg;
3160 
3161 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3162 	    cas32(flag, 0, 1) != 0)
3163 		return;
3164 
3165 	/*
3166 	 * It may not be kosher to do memory allocation when a reap is called
3167 	 * is called (for example, if vmem_populate() is in the call chain).
3168 	 * So we start the reap going with a TQ_NOALLOC dispatch.  If the
3169 	 * dispatch fails, we reset the flag, and the next reap will try again.
3170 	 */
3171 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3172 		*flag = 0;
3173 }
3174 
3175 /*
3176  * Reclaim all unused memory from all caches.  Called from the VM system
3177  * when memory gets tight.
3178  */
3179 void
3180 kmem_reap(void)
3181 {
3182 	kmem_reap_common(&kmem_reaping);
3183 }
3184 
3185 /*
3186  * Reclaim all unused memory from identifier arenas, called when a vmem
3187  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3188  * cannot help with identifier exhaustion, we avoid both a large amount of
3189  * work and unwanted side-effects from reclaim callbacks.
3190  */
3191 void
3192 kmem_reap_idspace(void)
3193 {
3194 	kmem_reap_common(&kmem_reaping_idspace);
3195 }
3196 
3197 /*
3198  * Purge all magazines from a cache and set its magazine limit to zero.
3199  * All calls are serialized by the kmem_taskq lock, except for the final
3200  * call from kmem_cache_destroy().
3201  */
3202 static void
3203 kmem_cache_magazine_purge(kmem_cache_t *cp)
3204 {
3205 	kmem_cpu_cache_t *ccp;
3206 	kmem_magazine_t *mp, *pmp;
3207 	int rounds, prounds, cpu_seqid;
3208 
3209 	ASSERT(!list_link_active(&cp->cache_link) ||
3210 	    taskq_member(kmem_taskq, curthread));
3211 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3212 
3213 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3214 		ccp = &cp->cache_cpu[cpu_seqid];
3215 
3216 		mutex_enter(&ccp->cc_lock);
3217 		mp = ccp->cc_loaded;
3218 		pmp = ccp->cc_ploaded;
3219 		rounds = ccp->cc_rounds;
3220 		prounds = ccp->cc_prounds;
3221 		ccp->cc_loaded = NULL;
3222 		ccp->cc_ploaded = NULL;
3223 		ccp->cc_rounds = -1;
3224 		ccp->cc_prounds = -1;
3225 		ccp->cc_magsize = 0;
3226 		mutex_exit(&ccp->cc_lock);
3227 
3228 		if (mp)
3229 			kmem_magazine_destroy(cp, mp, rounds);
3230 		if (pmp)
3231 			kmem_magazine_destroy(cp, pmp, prounds);
3232 	}
3233 
3234 	/*
3235 	 * Updating the working set statistics twice in a row has the
3236 	 * effect of setting the working set size to zero, so everything
3237 	 * is eligible for reaping.
3238 	 */
3239 	kmem_depot_ws_update(cp);
3240 	kmem_depot_ws_update(cp);
3241 
3242 	kmem_depot_ws_reap(cp);
3243 }
3244 
3245 /*
3246  * Enable per-cpu magazines on a cache.
3247  */
3248 static void
3249 kmem_cache_magazine_enable(kmem_cache_t *cp)
3250 {
3251 	int cpu_seqid;
3252 
3253 	if (cp->cache_flags & KMF_NOMAGAZINE)
3254 		return;
3255 
3256 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3257 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3258 		mutex_enter(&ccp->cc_lock);
3259 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3260 		mutex_exit(&ccp->cc_lock);
3261 	}
3262 
3263 }
3264 
3265 /*
3266  * Reap (almost) everything right now.  See kmem_cache_magazine_purge()
3267  * for explanation of the back-to-back kmem_depot_ws_update() calls.
3268  */
3269 void
3270 kmem_cache_reap_now(kmem_cache_t *cp)
3271 {
3272 	ASSERT(list_link_active(&cp->cache_link));
3273 
3274 	kmem_depot_ws_update(cp);
3275 	kmem_depot_ws_update(cp);
3276 
3277 	(void) taskq_dispatch(kmem_taskq,
3278 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3279 	taskq_wait(kmem_taskq);
3280 }
3281 
3282 /*
3283  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3284  * provide a higher transfer rate with the depot, while smaller magazines
3285  * reduce memory consumption.  Magazine resizing is an expensive operation;
3286  * it should not be done frequently.
3287  *
3288  * Changes to the magazine size are serialized by the kmem_taskq lock.
3289  *
3290  * Note: at present this only grows the magazine size.  It might be useful
3291  * to allow shrinkage too.
3292  */
3293 static void
3294 kmem_cache_magazine_resize(kmem_cache_t *cp)
3295 {
3296 	kmem_magtype_t *mtp = cp->cache_magtype;
3297 
3298 	ASSERT(taskq_member(kmem_taskq, curthread));
3299 
3300 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3301 		kmem_cache_magazine_purge(cp);
3302 		mutex_enter(&cp->cache_depot_lock);
3303 		cp->cache_magtype = ++mtp;
3304 		cp->cache_depot_contention_prev =
3305 		    cp->cache_depot_contention + INT_MAX;
3306 		mutex_exit(&cp->cache_depot_lock);
3307 		kmem_cache_magazine_enable(cp);
3308 	}
3309 }
3310 
3311 /*
3312  * Rescale a cache's hash table, so that the table size is roughly the
3313  * cache size.  We want the average lookup time to be extremely small.
3314  */
3315 static void
3316 kmem_hash_rescale(kmem_cache_t *cp)
3317 {
3318 	kmem_bufctl_t **old_table, **new_table, *bcp;
3319 	size_t old_size, new_size, h;
3320 
3321 	ASSERT(taskq_member(kmem_taskq, curthread));
3322 
3323 	new_size = MAX(KMEM_HASH_INITIAL,
3324 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3325 	old_size = cp->cache_hash_mask + 1;
3326 
3327 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3328 		return;
3329 
3330 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3331 	    VM_NOSLEEP);
3332 	if (new_table == NULL)
3333 		return;
3334 	bzero(new_table, new_size * sizeof (void *));
3335 
3336 	mutex_enter(&cp->cache_lock);
3337 
3338 	old_size = cp->cache_hash_mask + 1;
3339 	old_table = cp->cache_hash_table;
3340 
3341 	cp->cache_hash_mask = new_size - 1;
3342 	cp->cache_hash_table = new_table;
3343 	cp->cache_rescale++;
3344 
3345 	for (h = 0; h < old_size; h++) {
3346 		bcp = old_table[h];
3347 		while (bcp != NULL) {
3348 			void *addr = bcp->bc_addr;
3349 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3350 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3351 			bcp->bc_next = *hash_bucket;
3352 			*hash_bucket = bcp;
3353 			bcp = next_bcp;
3354 		}
3355 	}
3356 
3357 	mutex_exit(&cp->cache_lock);
3358 
3359 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3360 }
3361 
3362 /*
3363  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3364  * update, magazine resizing, and slab consolidation.
3365  */
3366 static void
3367 kmem_cache_update(kmem_cache_t *cp)
3368 {
3369 	int need_hash_rescale = 0;
3370 	int need_magazine_resize = 0;
3371 
3372 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3373 
3374 	/*
3375 	 * If the cache has become much larger or smaller than its hash table,
3376 	 * fire off a request to rescale the hash table.
3377 	 */
3378 	mutex_enter(&cp->cache_lock);
3379 
3380 	if ((cp->cache_flags & KMF_HASH) &&
3381 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3382 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3383 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3384 		need_hash_rescale = 1;
3385 
3386 	mutex_exit(&cp->cache_lock);
3387 
3388 	/*
3389 	 * Update the depot working set statistics.
3390 	 */
3391 	kmem_depot_ws_update(cp);
3392 
3393 	/*
3394 	 * If there's a lot of contention in the depot,
3395 	 * increase the magazine size.
3396 	 */
3397 	mutex_enter(&cp->cache_depot_lock);
3398 
3399 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3400 	    (int)(cp->cache_depot_contention -
3401 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3402 		need_magazine_resize = 1;
3403 
3404 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3405 
3406 	mutex_exit(&cp->cache_depot_lock);
3407 
3408 	if (need_hash_rescale)
3409 		(void) taskq_dispatch(kmem_taskq,
3410 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3411 
3412 	if (need_magazine_resize)
3413 		(void) taskq_dispatch(kmem_taskq,
3414 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3415 
3416 	if (cp->cache_defrag != NULL)
3417 		(void) taskq_dispatch(kmem_taskq,
3418 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3419 }
3420 
3421 static void kmem_update(void *);
3422 
3423 static void
3424 kmem_update_timeout(void *dummy)
3425 {
3426 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3427 }
3428 
3429 static void
3430 kmem_update(void *dummy)
3431 {
3432 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3433 
3434 	/*
3435 	 * We use taskq_dispatch() to reschedule the timeout so that
3436 	 * kmem_update() becomes self-throttling: it won't schedule
3437 	 * new tasks until all previous tasks have completed.
3438 	 */
3439 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3440 		kmem_update_timeout(NULL);
3441 }
3442 
3443 static int
3444 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3445 {
3446 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3447 	kmem_cache_t *cp = ksp->ks_private;
3448 	uint64_t cpu_buf_avail;
3449 	uint64_t buf_avail = 0;
3450 	int cpu_seqid;
3451 	long reap;
3452 
3453 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3454 
3455 	if (rw == KSTAT_WRITE)
3456 		return (EACCES);
3457 
3458 	mutex_enter(&cp->cache_lock);
3459 
3460 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3461 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3462 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3463 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3464 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3465 
3466 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3467 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3468 
3469 		mutex_enter(&ccp->cc_lock);
3470 
3471 		cpu_buf_avail = 0;
3472 		if (ccp->cc_rounds > 0)
3473 			cpu_buf_avail += ccp->cc_rounds;
3474 		if (ccp->cc_prounds > 0)
3475 			cpu_buf_avail += ccp->cc_prounds;
3476 
3477 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3478 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3479 		buf_avail			+= cpu_buf_avail;
3480 
3481 		mutex_exit(&ccp->cc_lock);
3482 	}
3483 
3484 	mutex_enter(&cp->cache_depot_lock);
3485 
3486 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3487 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3488 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3489 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3490 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3491 	kmcp->kmc_magazine_size.value.ui64	=
3492 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3493 	    0 : cp->cache_magtype->mt_magsize;
3494 
3495 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3496 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3497 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3498 
3499 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3500 	reap = MIN(reap, cp->cache_full.ml_total);
3501 
3502 	mutex_exit(&cp->cache_depot_lock);
3503 
3504 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3505 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3506 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3507 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3508 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3509 	buf_avail += cp->cache_bufslab;
3510 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3511 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3512 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3513 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3514 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3515 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3516 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3517 	    cp->cache_hash_mask + 1 : 0;
3518 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3519 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3520 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3521 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3522 
3523 	if (cp->cache_defrag == NULL) {
3524 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3525 		kmcp->kmc_move_yes.value.ui64		= 0;
3526 		kmcp->kmc_move_no.value.ui64		= 0;
3527 		kmcp->kmc_move_later.value.ui64		= 0;
3528 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3529 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3530 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3531 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3532 		kmcp->kmc_defrag.value.ui64		= 0;
3533 		kmcp->kmc_scan.value.ui64		= 0;
3534 		kmcp->kmc_move_reclaimable.value.ui64	= 0;
3535 	} else {
3536 		int64_t reclaimable;
3537 
3538 		kmem_defrag_t *kd = cp->cache_defrag;
3539 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3540 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3541 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3542 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3543 		kmcp->kmc_move_dont_need.value.