xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision 75d94465)
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 extern int callout_init_done;
836 
837 struct kmem_cache_kstat {
838 	kstat_named_t	kmc_buf_size;
839 	kstat_named_t	kmc_align;
840 	kstat_named_t	kmc_chunk_size;
841 	kstat_named_t	kmc_slab_size;
842 	kstat_named_t	kmc_alloc;
843 	kstat_named_t	kmc_alloc_fail;
844 	kstat_named_t	kmc_free;
845 	kstat_named_t	kmc_depot_alloc;
846 	kstat_named_t	kmc_depot_free;
847 	kstat_named_t	kmc_depot_contention;
848 	kstat_named_t	kmc_slab_alloc;
849 	kstat_named_t	kmc_slab_free;
850 	kstat_named_t	kmc_buf_constructed;
851 	kstat_named_t	kmc_buf_avail;
852 	kstat_named_t	kmc_buf_inuse;
853 	kstat_named_t	kmc_buf_total;
854 	kstat_named_t	kmc_buf_max;
855 	kstat_named_t	kmc_slab_create;
856 	kstat_named_t	kmc_slab_destroy;
857 	kstat_named_t	kmc_vmem_source;
858 	kstat_named_t	kmc_hash_size;
859 	kstat_named_t	kmc_hash_lookup_depth;
860 	kstat_named_t	kmc_hash_rescale;
861 	kstat_named_t	kmc_full_magazines;
862 	kstat_named_t	kmc_empty_magazines;
863 	kstat_named_t	kmc_magazine_size;
864 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
865 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
866 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
867 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
868 	kstat_named_t	kmc_move_yes;
869 	kstat_named_t	kmc_move_no;
870 	kstat_named_t	kmc_move_later;
871 	kstat_named_t	kmc_move_dont_need;
872 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
873 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
874 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
875 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
876 } kmem_cache_kstat = {
877 	{ "buf_size",		KSTAT_DATA_UINT64 },
878 	{ "align",		KSTAT_DATA_UINT64 },
879 	{ "chunk_size",		KSTAT_DATA_UINT64 },
880 	{ "slab_size",		KSTAT_DATA_UINT64 },
881 	{ "alloc",		KSTAT_DATA_UINT64 },
882 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
883 	{ "free",		KSTAT_DATA_UINT64 },
884 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
885 	{ "depot_free",		KSTAT_DATA_UINT64 },
886 	{ "depot_contention",	KSTAT_DATA_UINT64 },
887 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
888 	{ "slab_free",		KSTAT_DATA_UINT64 },
889 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
890 	{ "buf_avail",		KSTAT_DATA_UINT64 },
891 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
892 	{ "buf_total",		KSTAT_DATA_UINT64 },
893 	{ "buf_max",		KSTAT_DATA_UINT64 },
894 	{ "slab_create",	KSTAT_DATA_UINT64 },
895 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
896 	{ "vmem_source",	KSTAT_DATA_UINT64 },
897 	{ "hash_size",		KSTAT_DATA_UINT64 },
898 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
899 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
900 	{ "full_magazines",	KSTAT_DATA_UINT64 },
901 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
902 	{ "magazine_size",	KSTAT_DATA_UINT64 },
903 	{ "reap",		KSTAT_DATA_UINT64 },
904 	{ "defrag",		KSTAT_DATA_UINT64 },
905 	{ "scan",		KSTAT_DATA_UINT64 },
906 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
907 	{ "move_yes",		KSTAT_DATA_UINT64 },
908 	{ "move_no",		KSTAT_DATA_UINT64 },
909 	{ "move_later",		KSTAT_DATA_UINT64 },
910 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
911 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
912 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
913 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
914 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
915 };
916 
917 static kmutex_t kmem_cache_kstat_lock;
918 
919 /*
920  * The default set of caches to back kmem_alloc().
921  * These sizes should be reevaluated periodically.
922  *
923  * We want allocations that are multiples of the coherency granularity
924  * (64 bytes) to be satisfied from a cache which is a multiple of 64
925  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
926  * the next kmem_cache_size greater than or equal to it must be a
927  * multiple of 64.
928  *
929  * We split the table into two sections:  size <= 4k and size > 4k.  This
930  * saves a lot of space and cache footprint in our cache tables.
931  */
932 static const int kmem_alloc_sizes[] = {
933 	1 * 8,
934 	2 * 8,
935 	3 * 8,
936 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
937 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
938 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
939 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
940 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
941 	P2ALIGN(8192 / 7, 64),
942 	P2ALIGN(8192 / 6, 64),
943 	P2ALIGN(8192 / 5, 64),
944 	P2ALIGN(8192 / 4, 64),
945 	P2ALIGN(8192 / 3, 64),
946 	P2ALIGN(8192 / 2, 64),
947 };
948 
949 static const int kmem_big_alloc_sizes[] = {
950 	2 * 4096,	3 * 4096,
951 	2 * 8192,	3 * 8192,
952 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
953 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
954 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
955 	16 * 8192
956 };
957 
958 #define	KMEM_MAXBUF		4096
959 #define	KMEM_BIG_MAXBUF_32BIT	32768
960 #define	KMEM_BIG_MAXBUF		131072
961 
962 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
963 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
964 
965 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
966 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
967 
968 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
969 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
970 
971 static kmem_magtype_t kmem_magtype[] = {
972 	{ 1,	8,	3200,	65536	},
973 	{ 3,	16,	256,	32768	},
974 	{ 7,	32,	64,	16384	},
975 	{ 15,	64,	0,	8192	},
976 	{ 31,	64,	0,	4096	},
977 	{ 47,	64,	0,	2048	},
978 	{ 63,	64,	0,	1024	},
979 	{ 95,	64,	0,	512	},
980 	{ 143,	64,	0,	0	},
981 };
982 
983 static uint32_t kmem_reaping;
984 static uint32_t kmem_reaping_idspace;
985 
986 /*
987  * kmem tunables
988  */
989 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
990 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
991 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
992 int kmem_panic = 1;		/* whether to panic on error */
993 int kmem_logging = 1;		/* kmem_log_enter() override */
994 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
995 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
996 size_t kmem_content_log_size;	/* content log size [2% of memory] */
997 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
998 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
999 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1000 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1001 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1002 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1003 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1004 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1005 
1006 #ifdef _LP64
1007 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1008 #else
1009 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1010 #endif
1011 
1012 #ifdef DEBUG
1013 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1014 #else
1015 int kmem_flags = 0;
1016 #endif
1017 int kmem_ready;
1018 
1019 static kmem_cache_t	*kmem_slab_cache;
1020 static kmem_cache_t	*kmem_bufctl_cache;
1021 static kmem_cache_t	*kmem_bufctl_audit_cache;
1022 
1023 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1024 static list_t		kmem_caches;
1025 
1026 static taskq_t		*kmem_taskq;
1027 static kmutex_t		kmem_flags_lock;
1028 static vmem_t		*kmem_metadata_arena;
1029 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1030 static vmem_t		*kmem_cache_arena;
1031 static vmem_t		*kmem_hash_arena;
1032 static vmem_t		*kmem_log_arena;
1033 static vmem_t		*kmem_oversize_arena;
1034 static vmem_t		*kmem_va_arena;
1035 static vmem_t		*kmem_default_arena;
1036 static vmem_t		*kmem_firewall_va_arena;
1037 static vmem_t		*kmem_firewall_arena;
1038 
1039 /*
1040  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1041  * turned on when DEBUG is also defined.
1042  */
1043 #ifdef	DEBUG
1044 #define	KMEM_STATS
1045 #endif	/* DEBUG */
1046 
1047 #ifdef	KMEM_STATS
1048 #define	KMEM_STAT_ADD(stat)			((stat)++)
1049 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1050 #else
1051 #define	KMEM_STAT_ADD(stat)			/* nothing */
1052 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1053 #endif	/* KMEM_STATS */
1054 
1055 /*
1056  * kmem slab consolidator thresholds (tunables)
1057  */
1058 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1059 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1060 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1061 /*
1062  * Maximum number of slabs from which to move buffers during a single
1063  * maintenance interval while the system is not low on memory.
1064  */
1065 size_t kmem_reclaim_max_slabs = 1;
1066 /*
1067  * Number of slabs to scan backwards from the end of the partial slab list
1068  * when searching for buffers to relocate.
1069  */
1070 size_t kmem_reclaim_scan_range = 12;
1071 
1072 #ifdef	KMEM_STATS
1073 static struct {
1074 	uint64_t kms_callbacks;
1075 	uint64_t kms_yes;
1076 	uint64_t kms_no;
1077 	uint64_t kms_later;
1078 	uint64_t kms_dont_need;
1079 	uint64_t kms_dont_know;
1080 	uint64_t kms_hunt_found_mag;
1081 	uint64_t kms_hunt_found_slab;
1082 	uint64_t kms_hunt_alloc_fail;
1083 	uint64_t kms_hunt_lucky;
1084 	uint64_t kms_notify;
1085 	uint64_t kms_notify_callbacks;
1086 	uint64_t kms_disbelief;
1087 	uint64_t kms_already_pending;
1088 	uint64_t kms_callback_alloc_fail;
1089 	uint64_t kms_callback_taskq_fail;
1090 	uint64_t kms_endscan_slab_dead;
1091 	uint64_t kms_endscan_slab_destroyed;
1092 	uint64_t kms_endscan_nomem;
1093 	uint64_t kms_endscan_refcnt_changed;
1094 	uint64_t kms_endscan_nomove_changed;
1095 	uint64_t kms_endscan_freelist;
1096 	uint64_t kms_avl_update;
1097 	uint64_t kms_avl_noupdate;
1098 	uint64_t kms_no_longer_reclaimable;
1099 	uint64_t kms_notify_no_longer_reclaimable;
1100 	uint64_t kms_notify_slab_dead;
1101 	uint64_t kms_notify_slab_destroyed;
1102 	uint64_t kms_alloc_fail;
1103 	uint64_t kms_constructor_fail;
1104 	uint64_t kms_dead_slabs_freed;
1105 	uint64_t kms_defrags;
1106 	uint64_t kms_scans;
1107 	uint64_t kms_scan_depot_ws_reaps;
1108 	uint64_t kms_debug_reaps;
1109 	uint64_t kms_debug_scans;
1110 } kmem_move_stats;
1111 #endif	/* KMEM_STATS */
1112 
1113 /* consolidator knobs */
1114 static boolean_t kmem_move_noreap;
1115 static boolean_t kmem_move_blocked;
1116 static boolean_t kmem_move_fulltilt;
1117 static boolean_t kmem_move_any_partial;
1118 
1119 #ifdef	DEBUG
1120 /*
1121  * kmem consolidator debug tunables:
1122  * Ensure code coverage by occasionally running the consolidator even when the
1123  * caches are not fragmented (they may never be). These intervals are mean time
1124  * in cache maintenance intervals (kmem_cache_update).
1125  */
1126 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1127 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1128 #endif	/* DEBUG */
1129 
1130 static kmem_cache_t	*kmem_defrag_cache;
1131 static kmem_cache_t	*kmem_move_cache;
1132 static taskq_t		*kmem_move_taskq;
1133 
1134 static void kmem_cache_scan(kmem_cache_t *);
1135 static void kmem_cache_defrag(kmem_cache_t *);
1136 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1137 
1138 
1139 kmem_log_header_t	*kmem_transaction_log;
1140 kmem_log_header_t	*kmem_content_log;
1141 kmem_log_header_t	*kmem_failure_log;
1142 kmem_log_header_t	*kmem_slab_log;
1143 
1144 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1145 
1146 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1147 	if ((count) > 0) {						\
1148 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1149 		pc_t *_e;						\
1150 		/* memmove() the old entries down one notch */		\
1151 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1152 			*_e = *(_e - 1);				\
1153 		*_s = (uintptr_t)(caller);				\
1154 	}
1155 
1156 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1157 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1158 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1159 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1160 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1161 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1162 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1163 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1164 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1165 
1166 struct {
1167 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1168 	int		kmp_error;	/* type of kmem error */
1169 	void		*kmp_buffer;	/* buffer that induced panic */
1170 	void		*kmp_realbuf;	/* real start address for buffer */
1171 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1172 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1173 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1174 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1175 } kmem_panic_info;
1176 
1177 
1178 static void
1179 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1180 {
1181 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1182 	uint64_t *buf = buf_arg;
1183 
1184 	while (buf < bufend)
1185 		*buf++ = pattern;
1186 }
1187 
1188 static void *
1189 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1190 {
1191 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1192 	uint64_t *buf;
1193 
1194 	for (buf = buf_arg; buf < bufend; buf++)
1195 		if (*buf != pattern)
1196 			return (buf);
1197 	return (NULL);
1198 }
1199 
1200 static void *
1201 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1202 {
1203 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1204 	uint64_t *buf;
1205 
1206 	for (buf = buf_arg; buf < bufend; buf++) {
1207 		if (*buf != old) {
1208 			copy_pattern(old, buf_arg,
1209 			    (char *)buf - (char *)buf_arg);
1210 			return (buf);
1211 		}
1212 		*buf = new;
1213 	}
1214 
1215 	return (NULL);
1216 }
1217 
1218 static void
1219 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1220 {
1221 	kmem_cache_t *cp;
1222 
1223 	mutex_enter(&kmem_cache_lock);
1224 	for (cp = list_head(&kmem_caches); cp != NULL;
1225 	    cp = list_next(&kmem_caches, cp))
1226 		if (tq != NULL)
1227 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1228 			    tqflag);
1229 		else
1230 			func(cp);
1231 	mutex_exit(&kmem_cache_lock);
1232 }
1233 
1234 static void
1235 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1236 {
1237 	kmem_cache_t *cp;
1238 
1239 	mutex_enter(&kmem_cache_lock);
1240 	for (cp = list_head(&kmem_caches); cp != NULL;
1241 	    cp = list_next(&kmem_caches, cp)) {
1242 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1243 			continue;
1244 		if (tq != NULL)
1245 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1246 			    tqflag);
1247 		else
1248 			func(cp);
1249 	}
1250 	mutex_exit(&kmem_cache_lock);
1251 }
1252 
1253 /*
1254  * Debugging support.  Given a buffer address, find its slab.
1255  */
1256 static kmem_slab_t *
1257 kmem_findslab(kmem_cache_t *cp, void *buf)
1258 {
1259 	kmem_slab_t *sp;
1260 
1261 	mutex_enter(&cp->cache_lock);
1262 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1263 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1264 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1265 			mutex_exit(&cp->cache_lock);
1266 			return (sp);
1267 		}
1268 	}
1269 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1270 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1271 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1272 			mutex_exit(&cp->cache_lock);
1273 			return (sp);
1274 		}
1275 	}
1276 	mutex_exit(&cp->cache_lock);
1277 
1278 	return (NULL);
1279 }
1280 
1281 static void
1282 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1283 {
1284 	kmem_buftag_t *btp = NULL;
1285 	kmem_bufctl_t *bcp = NULL;
1286 	kmem_cache_t *cp = cparg;
1287 	kmem_slab_t *sp;
1288 	uint64_t *off;
1289 	void *buf = bufarg;
1290 
1291 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1292 
1293 	kmem_panic_info.kmp_timestamp = gethrtime();
1294 
1295 	sp = kmem_findslab(cp, buf);
1296 	if (sp == NULL) {
1297 		for (cp = list_tail(&kmem_caches); cp != NULL;
1298 		    cp = list_prev(&kmem_caches, cp)) {
1299 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1300 				break;
1301 		}
1302 	}
1303 
1304 	if (sp == NULL) {
1305 		cp = NULL;
1306 		error = KMERR_BADADDR;
1307 	} else {
1308 		if (cp != cparg)
1309 			error = KMERR_BADCACHE;
1310 		else
1311 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1312 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1313 		if (buf != bufarg)
1314 			error = KMERR_BADBASE;
1315 		if (cp->cache_flags & KMF_BUFTAG)
1316 			btp = KMEM_BUFTAG(cp, buf);
1317 		if (cp->cache_flags & KMF_HASH) {
1318 			mutex_enter(&cp->cache_lock);
1319 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1320 				if (bcp->bc_addr == buf)
1321 					break;
1322 			mutex_exit(&cp->cache_lock);
1323 			if (bcp == NULL && btp != NULL)
1324 				bcp = btp->bt_bufctl;
1325 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1326 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1327 			    bcp->bc_addr != buf) {
1328 				error = KMERR_BADBUFCTL;
1329 				bcp = NULL;
1330 			}
1331 		}
1332 	}
1333 
1334 	kmem_panic_info.kmp_error = error;
1335 	kmem_panic_info.kmp_buffer = bufarg;
1336 	kmem_panic_info.kmp_realbuf = buf;
1337 	kmem_panic_info.kmp_cache = cparg;
1338 	kmem_panic_info.kmp_realcache = cp;
1339 	kmem_panic_info.kmp_slab = sp;
1340 	kmem_panic_info.kmp_bufctl = bcp;
1341 
1342 	printf("kernel memory allocator: ");
1343 
1344 	switch (error) {
1345 
1346 	case KMERR_MODIFIED:
1347 		printf("buffer modified after being freed\n");
1348 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1349 		if (off == NULL)	/* shouldn't happen */
1350 			off = buf;
1351 		printf("modification occurred at offset 0x%lx "
1352 		    "(0x%llx replaced by 0x%llx)\n",
1353 		    (uintptr_t)off - (uintptr_t)buf,
1354 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1355 		break;
1356 
1357 	case KMERR_REDZONE:
1358 		printf("redzone violation: write past end of buffer\n");
1359 		break;
1360 
1361 	case KMERR_BADADDR:
1362 		printf("invalid free: buffer not in cache\n");
1363 		break;
1364 
1365 	case KMERR_DUPFREE:
1366 		printf("duplicate free: buffer freed twice\n");
1367 		break;
1368 
1369 	case KMERR_BADBUFTAG:
1370 		printf("boundary tag corrupted\n");
1371 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1372 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1373 		    KMEM_BUFTAG_FREE);
1374 		break;
1375 
1376 	case KMERR_BADBUFCTL:
1377 		printf("bufctl corrupted\n");
1378 		break;
1379 
1380 	case KMERR_BADCACHE:
1381 		printf("buffer freed to wrong cache\n");
1382 		printf("buffer was allocated from %s,\n", cp->cache_name);
1383 		printf("caller attempting free to %s.\n", cparg->cache_name);
1384 		break;
1385 
1386 	case KMERR_BADSIZE:
1387 		printf("bad free: free size (%u) != alloc size (%u)\n",
1388 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1389 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1390 		break;
1391 
1392 	case KMERR_BADBASE:
1393 		printf("bad free: free address (%p) != alloc address (%p)\n",
1394 		    bufarg, buf);
1395 		break;
1396 	}
1397 
1398 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1399 	    bufarg, (void *)bcp, cparg->cache_name);
1400 
1401 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1402 	    error != KMERR_BADBUFCTL) {
1403 		int d;
1404 		timestruc_t ts;
1405 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1406 
1407 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1408 		printf("previous transaction on buffer %p:\n", buf);
1409 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1410 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1411 		    (void *)sp, cp->cache_name);
1412 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1413 			ulong_t off;
1414 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1415 			printf("%s+%lx\n", sym ? sym : "?", off);
1416 		}
1417 	}
1418 	if (kmem_panic > 0)
1419 		panic("kernel heap corruption detected");
1420 	if (kmem_panic == 0)
1421 		debug_enter(NULL);
1422 	kmem_logging = 1;	/* resume logging */
1423 }
1424 
1425 static kmem_log_header_t *
1426 kmem_log_init(size_t logsize)
1427 {
1428 	kmem_log_header_t *lhp;
1429 	int nchunks = 4 * max_ncpus;
1430 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1431 	int i;
1432 
1433 	/*
1434 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1435 	 * to prevent false sharing of cache lines.
1436 	 */
1437 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1438 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1439 	    NULL, NULL, VM_SLEEP);
1440 	bzero(lhp, lhsize);
1441 
1442 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1443 	lhp->lh_nchunks = nchunks;
1444 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1445 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1446 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1447 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1448 	    nchunks * sizeof (int), VM_SLEEP);
1449 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1450 
1451 	for (i = 0; i < max_ncpus; i++) {
1452 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1453 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1454 		clhp->clh_chunk = i;
1455 	}
1456 
1457 	for (i = max_ncpus; i < nchunks; i++)
1458 		lhp->lh_free[i] = i;
1459 
1460 	lhp->lh_head = max_ncpus;
1461 	lhp->lh_tail = 0;
1462 
1463 	return (lhp);
1464 }
1465 
1466 static void *
1467 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1468 {
1469 	void *logspace;
1470 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1471 
1472 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1473 		return (NULL);
1474 
1475 	mutex_enter(&clhp->clh_lock);
1476 	clhp->clh_hits++;
1477 	if (size > clhp->clh_avail) {
1478 		mutex_enter(&lhp->lh_lock);
1479 		lhp->lh_hits++;
1480 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1481 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1482 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1483 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1484 		clhp->clh_current = lhp->lh_base +
1485 		    clhp->clh_chunk * lhp->lh_chunksize;
1486 		clhp->clh_avail = lhp->lh_chunksize;
1487 		if (size > lhp->lh_chunksize)
1488 			size = lhp->lh_chunksize;
1489 		mutex_exit(&lhp->lh_lock);
1490 	}
1491 	logspace = clhp->clh_current;
1492 	clhp->clh_current += size;
1493 	clhp->clh_avail -= size;
1494 	bcopy(data, logspace, size);
1495 	mutex_exit(&clhp->clh_lock);
1496 	return (logspace);
1497 }
1498 
1499 #define	KMEM_AUDIT(lp, cp, bcp)						\
1500 {									\
1501 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1502 	_bcp->bc_timestamp = gethrtime();				\
1503 	_bcp->bc_thread = curthread;					\
1504 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1505 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1506 }
1507 
1508 static void
1509 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1510 	kmem_slab_t *sp, void *addr)
1511 {
1512 	kmem_bufctl_audit_t bca;
1513 
1514 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1515 	bca.bc_addr = addr;
1516 	bca.bc_slab = sp;
1517 	bca.bc_cache = cp;
1518 	KMEM_AUDIT(lp, cp, &bca);
1519 }
1520 
1521 /*
1522  * Create a new slab for cache cp.
1523  */
1524 static kmem_slab_t *
1525 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1526 {
1527 	size_t slabsize = cp->cache_slabsize;
1528 	size_t chunksize = cp->cache_chunksize;
1529 	int cache_flags = cp->cache_flags;
1530 	size_t color, chunks;
1531 	char *buf, *slab;
1532 	kmem_slab_t *sp;
1533 	kmem_bufctl_t *bcp;
1534 	vmem_t *vmp = cp->cache_arena;
1535 
1536 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1537 
1538 	color = cp->cache_color + cp->cache_align;
1539 	if (color > cp->cache_maxcolor)
1540 		color = cp->cache_mincolor;
1541 	cp->cache_color = color;
1542 
1543 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1544 
1545 	if (slab == NULL)
1546 		goto vmem_alloc_failure;
1547 
1548 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1549 
1550 	/*
1551 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1552 	 * consolidator depends (for correctness) on slabs being initialized
1553 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1554 	 * clients to distinguish uninitialized memory from known objects).
1555 	 */
1556 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1557 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1558 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1559 
1560 	if (cache_flags & KMF_HASH) {
1561 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1562 			goto slab_alloc_failure;
1563 		chunks = (slabsize - color) / chunksize;
1564 	} else {
1565 		sp = KMEM_SLAB(cp, slab);
1566 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1567 	}
1568 
1569 	sp->slab_cache	= cp;
1570 	sp->slab_head	= NULL;
1571 	sp->slab_refcnt	= 0;
1572 	sp->slab_base	= buf = slab + color;
1573 	sp->slab_chunks	= chunks;
1574 	sp->slab_stuck_offset = (uint32_t)-1;
1575 	sp->slab_later_count = 0;
1576 	sp->slab_flags = 0;
1577 
1578 	ASSERT(chunks > 0);
1579 	while (chunks-- != 0) {
1580 		if (cache_flags & KMF_HASH) {
1581 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1582 			if (bcp == NULL)
1583 				goto bufctl_alloc_failure;
1584 			if (cache_flags & KMF_AUDIT) {
1585 				kmem_bufctl_audit_t *bcap =
1586 				    (kmem_bufctl_audit_t *)bcp;
1587 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1588 				bcap->bc_cache = cp;
1589 			}
1590 			bcp->bc_addr = buf;
1591 			bcp->bc_slab = sp;
1592 		} else {
1593 			bcp = KMEM_BUFCTL(cp, buf);
1594 		}
1595 		if (cache_flags & KMF_BUFTAG) {
1596 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1597 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1598 			btp->bt_bufctl = bcp;
1599 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1600 			if (cache_flags & KMF_DEADBEEF) {
1601 				copy_pattern(KMEM_FREE_PATTERN, buf,
1602 				    cp->cache_verify);
1603 			}
1604 		}
1605 		bcp->bc_next = sp->slab_head;
1606 		sp->slab_head = bcp;
1607 		buf += chunksize;
1608 	}
1609 
1610 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1611 
1612 	return (sp);
1613 
1614 bufctl_alloc_failure:
1615 
1616 	while ((bcp = sp->slab_head) != NULL) {
1617 		sp->slab_head = bcp->bc_next;
1618 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1619 	}
1620 	kmem_cache_free(kmem_slab_cache, sp);
1621 
1622 slab_alloc_failure:
1623 
1624 	vmem_free(vmp, slab, slabsize);
1625 
1626 vmem_alloc_failure:
1627 
1628 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1629 	atomic_add_64(&cp->cache_alloc_fail, 1);
1630 
1631 	return (NULL);
1632 }
1633 
1634 /*
1635  * Destroy a slab.
1636  */
1637 static void
1638 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1639 {
1640 	vmem_t *vmp = cp->cache_arena;
1641 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1642 
1643 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1644 	ASSERT(sp->slab_refcnt == 0);
1645 
1646 	if (cp->cache_flags & KMF_HASH) {
1647 		kmem_bufctl_t *bcp;
1648 		while ((bcp = sp->slab_head) != NULL) {
1649 			sp->slab_head = bcp->bc_next;
1650 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1651 		}
1652 		kmem_cache_free(kmem_slab_cache, sp);
1653 	}
1654 	vmem_free(vmp, slab, cp->cache_slabsize);
1655 }
1656 
1657 static void *
1658 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1659 {
1660 	kmem_bufctl_t *bcp, **hash_bucket;
1661 	void *buf;
1662 	boolean_t new_slab = (sp->slab_refcnt == 0);
1663 
1664 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1665 	/*
1666 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1667 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1668 	 * slab is newly created.
1669 	 */
1670 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1671 	    (sp == avl_first(&cp->cache_partial_slabs))));
1672 	ASSERT(sp->slab_cache == cp);
1673 
1674 	cp->cache_slab_alloc++;
1675 	cp->cache_bufslab--;
1676 	sp->slab_refcnt++;
1677 
1678 	bcp = sp->slab_head;
1679 	sp->slab_head = bcp->bc_next;
1680 
1681 	if (cp->cache_flags & KMF_HASH) {
1682 		/*
1683 		 * Add buffer to allocated-address hash table.
1684 		 */
1685 		buf = bcp->bc_addr;
1686 		hash_bucket = KMEM_HASH(cp, buf);
1687 		bcp->bc_next = *hash_bucket;
1688 		*hash_bucket = bcp;
1689 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1690 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1691 		}
1692 	} else {
1693 		buf = KMEM_BUF(cp, bcp);
1694 	}
1695 
1696 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1697 
1698 	if (sp->slab_head == NULL) {
1699 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1700 		if (new_slab) {
1701 			ASSERT(sp->slab_chunks == 1);
1702 		} else {
1703 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1704 			avl_remove(&cp->cache_partial_slabs, sp);
1705 			sp->slab_later_count = 0; /* clear history */
1706 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1707 			sp->slab_stuck_offset = (uint32_t)-1;
1708 		}
1709 		list_insert_head(&cp->cache_complete_slabs, sp);
1710 		cp->cache_complete_slab_count++;
1711 		return (buf);
1712 	}
1713 
1714 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1715 	/*
1716 	 * Peek to see if the magazine layer is enabled before
1717 	 * we prefill.  We're not holding the cpu cache lock,
1718 	 * so the peek could be wrong, but there's no harm in it.
1719 	 */
1720 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1721 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1722 		kmem_slab_prefill(cp, sp);
1723 		return (buf);
1724 	}
1725 
1726 	if (new_slab) {
1727 		avl_add(&cp->cache_partial_slabs, sp);
1728 		return (buf);
1729 	}
1730 
1731 	/*
1732 	 * The slab is now more allocated than it was, so the
1733 	 * order remains unchanged.
1734 	 */
1735 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1736 	return (buf);
1737 }
1738 
1739 /*
1740  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1741  */
1742 static void *
1743 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1744 {
1745 	kmem_slab_t *sp;
1746 	void *buf;
1747 	boolean_t test_destructor;
1748 
1749 	mutex_enter(&cp->cache_lock);
1750 	test_destructor = (cp->cache_slab_alloc == 0);
1751 	sp = avl_first(&cp->cache_partial_slabs);
1752 	if (sp == NULL) {
1753 		ASSERT(cp->cache_bufslab == 0);
1754 
1755 		/*
1756 		 * The freelist is empty.  Create a new slab.
1757 		 */
1758 		mutex_exit(&cp->cache_lock);
1759 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1760 			return (NULL);
1761 		}
1762 		mutex_enter(&cp->cache_lock);
1763 		cp->cache_slab_create++;
1764 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1765 			cp->cache_bufmax = cp->cache_buftotal;
1766 		cp->cache_bufslab += sp->slab_chunks;
1767 	}
1768 
1769 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1770 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1771 	    (cp->cache_complete_slab_count +
1772 	    avl_numnodes(&cp->cache_partial_slabs) +
1773 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1774 	mutex_exit(&cp->cache_lock);
1775 
1776 	if (test_destructor && cp->cache_destructor != NULL) {
1777 		/*
1778 		 * On the first kmem_slab_alloc(), assert that it is valid to
1779 		 * call the destructor on a newly constructed object without any
1780 		 * client involvement.
1781 		 */
1782 		if ((cp->cache_constructor == NULL) ||
1783 		    cp->cache_constructor(buf, cp->cache_private,
1784 		    kmflag) == 0) {
1785 			cp->cache_destructor(buf, cp->cache_private);
1786 		}
1787 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1788 		    cp->cache_bufsize);
1789 		if (cp->cache_flags & KMF_DEADBEEF) {
1790 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1791 		}
1792 	}
1793 
1794 	return (buf);
1795 }
1796 
1797 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1798 
1799 /*
1800  * Free a raw (unconstructed) buffer to cp's slab layer.
1801  */
1802 static void
1803 kmem_slab_free(kmem_cache_t *cp, void *buf)
1804 {
1805 	kmem_slab_t *sp;
1806 	kmem_bufctl_t *bcp, **prev_bcpp;
1807 
1808 	ASSERT(buf != NULL);
1809 
1810 	mutex_enter(&cp->cache_lock);
1811 	cp->cache_slab_free++;
1812 
1813 	if (cp->cache_flags & KMF_HASH) {
1814 		/*
1815 		 * Look up buffer in allocated-address hash table.
1816 		 */
1817 		prev_bcpp = KMEM_HASH(cp, buf);
1818 		while ((bcp = *prev_bcpp) != NULL) {
1819 			if (bcp->bc_addr == buf) {
1820 				*prev_bcpp = bcp->bc_next;
1821 				sp = bcp->bc_slab;
1822 				break;
1823 			}
1824 			cp->cache_lookup_depth++;
1825 			prev_bcpp = &bcp->bc_next;
1826 		}
1827 	} else {
1828 		bcp = KMEM_BUFCTL(cp, buf);
1829 		sp = KMEM_SLAB(cp, buf);
1830 	}
1831 
1832 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1833 		mutex_exit(&cp->cache_lock);
1834 		kmem_error(KMERR_BADADDR, cp, buf);
1835 		return;
1836 	}
1837 
1838 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1839 		/*
1840 		 * If this is the buffer that prevented the consolidator from
1841 		 * clearing the slab, we can reset the slab flags now that the
1842 		 * buffer is freed. (It makes sense to do this in
1843 		 * kmem_cache_free(), where the client gives up ownership of the
1844 		 * buffer, but on the hot path the test is too expensive.)
1845 		 */
1846 		kmem_slab_move_yes(cp, sp, buf);
1847 	}
1848 
1849 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1850 		if (cp->cache_flags & KMF_CONTENTS)
1851 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1852 			    kmem_log_enter(kmem_content_log, buf,
1853 			    cp->cache_contents);
1854 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1855 	}
1856 
1857 	bcp->bc_next = sp->slab_head;
1858 	sp->slab_head = bcp;
1859 
1860 	cp->cache_bufslab++;
1861 	ASSERT(sp->slab_refcnt >= 1);
1862 
1863 	if (--sp->slab_refcnt == 0) {
1864 		/*
1865 		 * There are no outstanding allocations from this slab,
1866 		 * so we can reclaim the memory.
1867 		 */
1868 		if (sp->slab_chunks == 1) {
1869 			list_remove(&cp->cache_complete_slabs, sp);
1870 			cp->cache_complete_slab_count--;
1871 		} else {
1872 			avl_remove(&cp->cache_partial_slabs, sp);
1873 		}
1874 
1875 		cp->cache_buftotal -= sp->slab_chunks;
1876 		cp->cache_bufslab -= sp->slab_chunks;
1877 		/*
1878 		 * Defer releasing the slab to the virtual memory subsystem
1879 		 * while there is a pending move callback, since we guarantee
1880 		 * that buffers passed to the move callback have only been
1881 		 * touched by kmem or by the client itself. Since the memory
1882 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1883 		 * set at least one of the two lowest order bits, the client can
1884 		 * test those bits in the move callback to determine whether or
1885 		 * not it knows about the buffer (assuming that the client also
1886 		 * sets one of those low order bits whenever it frees a buffer).
1887 		 */
1888 		if (cp->cache_defrag == NULL ||
1889 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1890 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1891 			cp->cache_slab_destroy++;
1892 			mutex_exit(&cp->cache_lock);
1893 			kmem_slab_destroy(cp, sp);
1894 		} else {
1895 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1896 			/*
1897 			 * Slabs are inserted at both ends of the deadlist to
1898 			 * distinguish between slabs freed while move callbacks
1899 			 * are pending (list head) and a slab freed while the
1900 			 * lock is dropped in kmem_move_buffers() (list tail) so
1901 			 * that in both cases slab_destroy() is called from the
1902 			 * right context.
1903 			 */
1904 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1905 				list_insert_tail(deadlist, sp);
1906 			} else {
1907 				list_insert_head(deadlist, sp);
1908 			}
1909 			cp->cache_defrag->kmd_deadcount++;
1910 			mutex_exit(&cp->cache_lock);
1911 		}
1912 		return;
1913 	}
1914 
1915 	if (bcp->bc_next == NULL) {
1916 		/* Transition the slab from completely allocated to partial. */
1917 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1918 		ASSERT(sp->slab_chunks > 1);
1919 		list_remove(&cp->cache_complete_slabs, sp);
1920 		cp->cache_complete_slab_count--;
1921 		avl_add(&cp->cache_partial_slabs, sp);
1922 	} else {
1923 #ifdef	DEBUG
1924 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1925 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1926 		} else {
1927 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1928 		}
1929 #else
1930 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1931 #endif
1932 	}
1933 
1934 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1935 	    (cp->cache_complete_slab_count +
1936 	    avl_numnodes(&cp->cache_partial_slabs) +
1937 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1938 	mutex_exit(&cp->cache_lock);
1939 }
1940 
1941 /*
1942  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1943  */
1944 static int
1945 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1946     caddr_t caller)
1947 {
1948 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1949 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1950 	uint32_t mtbf;
1951 
1952 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1953 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1954 		return (-1);
1955 	}
1956 
1957 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1958 
1959 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1960 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1961 		return (-1);
1962 	}
1963 
1964 	if (cp->cache_flags & KMF_DEADBEEF) {
1965 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1966 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1967 				kmem_error(KMERR_MODIFIED, cp, buf);
1968 				return (-1);
1969 			}
1970 			if (cp->cache_constructor != NULL)
1971 				*(uint64_t *)buf = btp->bt_redzone;
1972 			else
1973 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1974 		} else {
1975 			construct = 1;
1976 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1977 			    KMEM_UNINITIALIZED_PATTERN, buf,
1978 			    cp->cache_verify)) {
1979 				kmem_error(KMERR_MODIFIED, cp, buf);
1980 				return (-1);
1981 			}
1982 		}
1983 	}
1984 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1985 
1986 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1987 	    gethrtime() % mtbf == 0 &&
1988 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1989 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1990 		if (!construct && cp->cache_destructor != NULL)
1991 			cp->cache_destructor(buf, cp->cache_private);
1992 	} else {
1993 		mtbf = 0;
1994 	}
1995 
1996 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1997 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1998 		atomic_add_64(&cp->cache_alloc_fail, 1);
1999 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2000 		if (cp->cache_flags & KMF_DEADBEEF)
2001 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2002 		kmem_slab_free(cp, buf);
2003 		return (1);
2004 	}
2005 
2006 	if (cp->cache_flags & KMF_AUDIT) {
2007 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2008 	}
2009 
2010 	if ((cp->cache_flags & KMF_LITE) &&
2011 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2012 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2013 	}
2014 
2015 	return (0);
2016 }
2017 
2018 static int
2019 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2020 {
2021 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2022 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2023 	kmem_slab_t *sp;
2024 
2025 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2026 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2027 			kmem_error(KMERR_DUPFREE, cp, buf);
2028 			return (-1);
2029 		}
2030 		sp = kmem_findslab(cp, buf);
2031 		if (sp == NULL || sp->slab_cache != cp)
2032 			kmem_error(KMERR_BADADDR, cp, buf);
2033 		else
2034 			kmem_error(KMERR_REDZONE, cp, buf);
2035 		return (-1);
2036 	}
2037 
2038 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2039 
2040 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2041 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2042 		return (-1);
2043 	}
2044 
2045 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2046 		kmem_error(KMERR_REDZONE, cp, buf);
2047 		return (-1);
2048 	}
2049 
2050 	if (cp->cache_flags & KMF_AUDIT) {
2051 		if (cp->cache_flags & KMF_CONTENTS)
2052 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2053 			    buf, cp->cache_contents);
2054 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2055 	}
2056 
2057 	if ((cp->cache_flags & KMF_LITE) &&
2058 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2059 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2060 	}
2061 
2062 	if (cp->cache_flags & KMF_DEADBEEF) {
2063 		if (cp->cache_flags & KMF_LITE)
2064 			btp->bt_redzone = *(uint64_t *)buf;
2065 		else if (cp->cache_destructor != NULL)
2066 			cp->cache_destructor(buf, cp->cache_private);
2067 
2068 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2069 	}
2070 
2071 	return (0);
2072 }
2073 
2074 /*
2075  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2076  */
2077 static void
2078 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2079 {
2080 	int round;
2081 
2082 	ASSERT(!list_link_active(&cp->cache_link) ||
2083 	    taskq_member(kmem_taskq, curthread));
2084 
2085 	for (round = 0; round < nrounds; round++) {
2086 		void *buf = mp->mag_round[round];
2087 
2088 		if (cp->cache_flags & KMF_DEADBEEF) {
2089 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2090 			    cp->cache_verify) != NULL) {
2091 				kmem_error(KMERR_MODIFIED, cp, buf);
2092 				continue;
2093 			}
2094 			if ((cp->cache_flags & KMF_LITE) &&
2095 			    cp->cache_destructor != NULL) {
2096 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2097 				*(uint64_t *)buf = btp->bt_redzone;
2098 				cp->cache_destructor(buf, cp->cache_private);
2099 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2100 			}
2101 		} else if (cp->cache_destructor != NULL) {
2102 			cp->cache_destructor(buf, cp->cache_private);
2103 		}
2104 
2105 		kmem_slab_free(cp, buf);
2106 	}
2107 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2108 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2109 }
2110 
2111 /*
2112  * Allocate a magazine from the depot.
2113  */
2114 static kmem_magazine_t *
2115 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2116 {
2117 	kmem_magazine_t *mp;
2118 
2119 	/*
2120 	 * If we can't get the depot lock without contention,
2121 	 * update our contention count.  We use the depot
2122 	 * contention rate to determine whether we need to
2123 	 * increase the magazine size for better scalability.
2124 	 */
2125 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2126 		mutex_enter(&cp->cache_depot_lock);
2127 		cp->cache_depot_contention++;
2128 	}
2129 
2130 	if ((mp = mlp->ml_list) != NULL) {
2131 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2132 		mlp->ml_list = mp->mag_next;
2133 		if (--mlp->ml_total < mlp->ml_min)
2134 			mlp->ml_min = mlp->ml_total;
2135 		mlp->ml_alloc++;
2136 	}
2137 
2138 	mutex_exit(&cp->cache_depot_lock);
2139 
2140 	return (mp);
2141 }
2142 
2143 /*
2144  * Free a magazine to the depot.
2145  */
2146 static void
2147 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2148 {
2149 	mutex_enter(&cp->cache_depot_lock);
2150 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2151 	mp->mag_next = mlp->ml_list;
2152 	mlp->ml_list = mp;
2153 	mlp->ml_total++;
2154 	mutex_exit(&cp->cache_depot_lock);
2155 }
2156 
2157 /*
2158  * Update the working set statistics for cp's depot.
2159  */
2160 static void
2161 kmem_depot_ws_update(kmem_cache_t *cp)
2162 {
2163 	mutex_enter(&cp->cache_depot_lock);
2164 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2165 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2166 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2167 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2168 	mutex_exit(&cp->cache_depot_lock);
2169 }
2170 
2171 /*
2172  * Reap all magazines that have fallen out of the depot's working set.
2173  */
2174 static void
2175 kmem_depot_ws_reap(kmem_cache_t *cp)
2176 {
2177 	long reap;
2178 	kmem_magazine_t *mp;
2179 
2180 	ASSERT(!list_link_active(&cp->cache_link) ||
2181 	    taskq_member(kmem_taskq, curthread));
2182 
2183 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2184 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2185 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2186 
2187 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2188 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2189 		kmem_magazine_destroy(cp, mp, 0);
2190 }
2191 
2192 static void
2193 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2194 {
2195 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2196 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2197 	ASSERT(ccp->cc_magsize > 0);
2198 
2199 	ccp->cc_ploaded = ccp->cc_loaded;
2200 	ccp->cc_prounds = ccp->cc_rounds;
2201 	ccp->cc_loaded = mp;
2202 	ccp->cc_rounds = rounds;
2203 }
2204 
2205 /*
2206  * Intercept kmem alloc/free calls during crash dump in order to avoid
2207  * changing kmem state while memory is being saved to the dump device.
2208  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2209  * there are no locks because only one CPU calls kmem during a crash
2210  * dump. To enable this feature, first create the associated vmem
2211  * arena with VMC_DUMPSAFE.
2212  */
2213 static void *kmem_dump_start;	/* start of pre-reserved heap */
2214 static void *kmem_dump_end;	/* end of heap area */
2215 static void *kmem_dump_curr;	/* current free heap pointer */
2216 static size_t kmem_dump_size;	/* size of heap area */
2217 
2218 /* append to each buf created in the pre-reserved heap */
2219 typedef struct kmem_dumpctl {
2220 	void	*kdc_next;	/* cache dump free list linkage */
2221 } kmem_dumpctl_t;
2222 
2223 #define	KMEM_DUMPCTL(cp, buf)	\
2224 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2225 	    sizeof (void *)))
2226 
2227 /* Keep some simple stats. */
2228 #define	KMEM_DUMP_LOGS	(100)
2229 
2230 typedef struct kmem_dump_log {
2231 	kmem_cache_t	*kdl_cache;
2232 	uint_t		kdl_allocs;		/* # of dump allocations */
2233 	uint_t		kdl_frees;		/* # of dump frees */
2234 	uint_t		kdl_alloc_fails;	/* # of allocation failures */
2235 	uint_t		kdl_free_nondump;	/* # of non-dump frees */
2236 	uint_t		kdl_unsafe;		/* cache was used, but unsafe */
2237 } kmem_dump_log_t;
2238 
2239 static kmem_dump_log_t *kmem_dump_log;
2240 static int kmem_dump_log_idx;
2241 
2242 #define	KDI_LOG(cp, stat) {						\
2243 	kmem_dump_log_t *kdl;						\
2244 	if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) {	\
2245 		kdl->stat++;						\
2246 	} else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) {		\
2247 		kdl = &kmem_dump_log[kmem_dump_log_idx++];		\
2248 		kdl->stat++;						\
2249 		kdl->kdl_cache = (cp);					\
2250 		(cp)->cache_dumplog = kdl;				\
2251 	}								\
2252 }
2253 
2254 /* set non zero for full report */
2255 uint_t kmem_dump_verbose = 0;
2256 
2257 /* stats for overize heap */
2258 uint_t kmem_dump_oversize_allocs = 0;
2259 uint_t kmem_dump_oversize_max = 0;
2260 
2261 static void
2262 kmem_dumppr(char **pp, char *e, const char *format, ...)
2263 {
2264 	char *p = *pp;
2265 
2266 	if (p < e) {
2267 		int n;
2268 		va_list ap;
2269 
2270 		va_start(ap, format);
2271 		n = vsnprintf(p, e - p, format, ap);
2272 		va_end(ap);
2273 		*pp = p + n;
2274 	}
2275 }
2276 
2277 /*
2278  * Called when dumpadm(1M) configures dump parameters.
2279  */
2280 void
2281 kmem_dump_init(size_t size)
2282 {
2283 	if (kmem_dump_start != NULL)
2284 		kmem_free(kmem_dump_start, kmem_dump_size);
2285 
2286 	if (kmem_dump_log == NULL)
2287 		kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2288 		    sizeof (kmem_dump_log_t), KM_SLEEP);
2289 
2290 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2291 
2292 	if (kmem_dump_start != NULL) {
2293 		kmem_dump_size = size;
2294 		kmem_dump_curr = kmem_dump_start;
2295 		kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2296 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2297 	} else {
2298 		kmem_dump_size = 0;
2299 		kmem_dump_curr = NULL;
2300 		kmem_dump_end = NULL;
2301 	}
2302 }
2303 
2304 /*
2305  * Set flag for each kmem_cache_t if is safe to use alternate dump
2306  * memory. Called just before panic crash dump starts. Set the flag
2307  * for the calling CPU.
2308  */
2309 void
2310 kmem_dump_begin(void)
2311 {
2312 	ASSERT(panicstr != NULL);
2313 	if (kmem_dump_start != NULL) {
2314 		kmem_cache_t *cp;
2315 
2316 		for (cp = list_head(&kmem_caches); cp != NULL;
2317 		    cp = list_next(&kmem_caches, cp)) {
2318 			kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2319 
2320 			if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2321 				cp->cache_flags |= KMF_DUMPDIVERT;
2322 				ccp->cc_flags |= KMF_DUMPDIVERT;
2323 				ccp->cc_dump_rounds = ccp->cc_rounds;
2324 				ccp->cc_dump_prounds = ccp->cc_prounds;
2325 				ccp->cc_rounds = ccp->cc_prounds = -1;
2326 			} else {
2327 				cp->cache_flags |= KMF_DUMPUNSAFE;
2328 				ccp->cc_flags |= KMF_DUMPUNSAFE;
2329 			}
2330 		}
2331 	}
2332 }
2333 
2334 /*
2335  * finished dump intercept
2336  * print any warnings on the console
2337  * return verbose information to dumpsys() in the given buffer
2338  */
2339 size_t
2340 kmem_dump_finish(char *buf, size_t size)
2341 {
2342 	int kdi_idx;
2343 	int kdi_end = kmem_dump_log_idx;
2344 	int percent = 0;
2345 	int header = 0;
2346 	int warn = 0;
2347 	size_t used;
2348 	kmem_cache_t *cp;
2349 	kmem_dump_log_t *kdl;
2350 	char *e = buf + size;
2351 	char *p = buf;
2352 
2353 	if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2354 		return (0);
2355 
2356 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2357 	percent = (used * 100) / kmem_dump_size;
2358 
2359 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2360 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2361 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2362 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2363 	    kmem_dump_oversize_allocs);
2364 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2365 	    kmem_dump_oversize_max);
2366 
2367 	for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2368 		kdl = &kmem_dump_log[kdi_idx];
2369 		cp = kdl->kdl_cache;
2370 		if (cp == NULL)
2371 			break;
2372 		if (kdl->kdl_alloc_fails)
2373 			++warn;
2374 		if (header == 0) {
2375 			kmem_dumppr(&p, e,
2376 			    "Cache Name,Allocs,Frees,Alloc Fails,"
2377 			    "Nondump Frees,Unsafe Allocs/Frees\n");
2378 			header = 1;
2379 		}
2380 		kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2381 		    cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2382 		    kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2383 		    kdl->kdl_unsafe);
2384 	}
2385 
2386 	/* return buffer size used */
2387 	if (p < e)
2388 		bzero(p, e - p);
2389 	return (p - buf);
2390 }
2391 
2392 /*
2393  * Allocate a constructed object from alternate dump memory.
2394  */
2395 void *
2396 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2397 {
2398 	void *buf;
2399 	void *curr;
2400 	char *bufend;
2401 
2402 	/* return a constructed object */
2403 	if ((buf = cp->cache_dumpfreelist) != NULL) {
2404 		cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2405 		KDI_LOG(cp, kdl_allocs);
2406 		return (buf);
2407 	}
2408 
2409 	/* create a new constructed object */
2410 	curr = kmem_dump_curr;
2411 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2412 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2413 
2414 	/* hat layer objects cannot cross a page boundary */
2415 	if (cp->cache_align < PAGESIZE) {
2416 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2417 		if (bufend > page) {
2418 			bufend += page - (char *)buf;
2419 			buf = (void *)page;
2420 		}
2421 	}
2422 
2423 	/* fall back to normal alloc if reserved area is used up */
2424 	if (bufend > (char *)kmem_dump_end) {
2425 		kmem_dump_curr = kmem_dump_end;
2426 		KDI_LOG(cp, kdl_alloc_fails);
2427 		return (NULL);
2428 	}
2429 
2430 	/*
2431 	 * Must advance curr pointer before calling a constructor that
2432 	 * may also allocate memory.
2433 	 */
2434 	kmem_dump_curr = bufend;
2435 
2436 	/* run constructor */
2437 	if (cp->cache_constructor != NULL &&
2438 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2439 	    != 0) {
2440 #ifdef DEBUG
2441 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2442 		    cp->cache_name, (void *)cp);
2443 #endif
2444 		/* reset curr pointer iff no allocs were done */
2445 		if (kmem_dump_curr == bufend)
2446 			kmem_dump_curr = curr;
2447 
2448 		/* fall back to normal alloc if the constructor fails */
2449 		KDI_LOG(cp, kdl_alloc_fails);
2450 		return (NULL);
2451 	}
2452 
2453 	KDI_LOG(cp, kdl_allocs);
2454 	return (buf);
2455 }
2456 
2457 /*
2458  * Free a constructed object in alternate dump memory.
2459  */
2460 int
2461 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2462 {
2463 	/* save constructed buffers for next time */
2464 	if ((char *)buf >= (char *)kmem_dump_start &&
2465 	    (char *)buf < (char *)kmem_dump_end) {
2466 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2467 		cp->cache_dumpfreelist = buf;
2468 		KDI_LOG(cp, kdl_frees);
2469 		return (0);
2470 	}
2471 
2472 	/* count all non-dump buf frees */
2473 	KDI_LOG(cp, kdl_free_nondump);
2474 
2475 	/* just drop buffers that were allocated before dump started */
2476 	if (kmem_dump_curr < kmem_dump_end)
2477 		return (0);
2478 
2479 	/* fall back to normal free if reserved area is used up */
2480 	return (1);
2481 }
2482 
2483 /*
2484  * Allocate a constructed object from cache cp.
2485  */
2486 void *
2487 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2488 {
2489 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2490 	kmem_magazine_t *fmp;
2491 	void *buf;
2492 
2493 	mutex_enter(&ccp->cc_lock);
2494 	for (;;) {
2495 		/*
2496 		 * If there's an object available in the current CPU's
2497 		 * loaded magazine, just take it and return.
2498 		 */
2499 		if (ccp->cc_rounds > 0) {
2500 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2501 			ccp->cc_alloc++;
2502 			mutex_exit(&ccp->cc_lock);
2503 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2504 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2505 					ASSERT(!(ccp->cc_flags &
2506 					    KMF_DUMPDIVERT));
2507 					KDI_LOG(cp, kdl_unsafe);
2508 				}
2509 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2510 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2511 				    caller()) != 0) {
2512 					if (kmflag & KM_NOSLEEP)
2513 						return (NULL);
2514 					mutex_enter(&ccp->cc_lock);
2515 					continue;
2516 				}
2517 			}
2518 			return (buf);
2519 		}
2520 
2521 		/*
2522 		 * The loaded magazine is empty.  If the previously loaded
2523 		 * magazine was full, exchange them and try again.
2524 		 */
2525 		if (ccp->cc_prounds > 0) {
2526 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2527 			continue;
2528 		}
2529 
2530 		/*
2531 		 * Return an alternate buffer at dump time to preserve
2532 		 * the heap.
2533 		 */
2534 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2535 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2536 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2537 				/* log it so that we can warn about it */
2538 				KDI_LOG(cp, kdl_unsafe);
2539 			} else {
2540 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2541 				    NULL) {
2542 					mutex_exit(&ccp->cc_lock);
2543 					return (buf);
2544 				}
2545 				break;		/* fall back to slab layer */
2546 			}
2547 		}
2548 
2549 		/*
2550 		 * If the magazine layer is disabled, break out now.
2551 		 */
2552 		if (ccp->cc_magsize == 0)
2553 			break;
2554 
2555 		/*
2556 		 * Try to get a full magazine from the depot.
2557 		 */
2558 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2559 		if (fmp != NULL) {
2560 			if (ccp->cc_ploaded != NULL)
2561 				kmem_depot_free(cp, &cp->cache_empty,
2562 				    ccp->cc_ploaded);
2563 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2564 			continue;
2565 		}
2566 
2567 		/*
2568 		 * There are no full magazines in the depot,
2569 		 * so fall through to the slab layer.
2570 		 */
2571 		break;
2572 	}
2573 	mutex_exit(&ccp->cc_lock);
2574 
2575 	/*
2576 	 * We couldn't allocate a constructed object from the magazine layer,
2577 	 * so get a raw buffer from the slab layer and apply its constructor.
2578 	 */
2579 	buf = kmem_slab_alloc(cp, kmflag);
2580 
2581 	if (buf == NULL)
2582 		return (NULL);
2583 
2584 	if (cp->cache_flags & KMF_BUFTAG) {
2585 		/*
2586 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2587 		 */
2588 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2589 		if (rc != 0) {
2590 			if (kmflag & KM_NOSLEEP)
2591 				return (NULL);
2592 			/*
2593 			 * kmem_cache_alloc_debug() detected corruption
2594 			 * but didn't panic (kmem_panic <= 0). We should not be
2595 			 * here because the constructor failed (indicated by a
2596 			 * return code of 1). Try again.
2597 			 */
2598 			ASSERT(rc == -1);
2599 			return (kmem_cache_alloc(cp, kmflag));
2600 		}
2601 		return (buf);
2602 	}
2603 
2604 	if (cp->cache_constructor != NULL &&
2605 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2606 		atomic_add_64(&cp->cache_alloc_fail, 1);
2607 		kmem_slab_free(cp, buf);
2608 		return (NULL);
2609 	}
2610 
2611 	return (buf);
2612 }
2613 
2614 /*
2615  * The freed argument tells whether or not kmem_cache_free_debug() has already
2616  * been called so that we can avoid the duplicate free error. For example, a
2617  * buffer on a magazine has already been freed by the client but is still
2618  * constructed.
2619  */
2620 static void
2621 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2622 {
2623 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2624 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2625 			return;
2626 
2627 	/*
2628 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2629 	 * kmem_cache_free_debug() will have already applied the destructor.
2630 	 */
2631 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2632 	    cp->cache_destructor != NULL) {
2633 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2634 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2635 			*(uint64_t *)buf = btp->bt_redzone;
2636 			cp->cache_destructor(buf, cp->cache_private);
2637 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2638 		} else {
2639 			cp->cache_destructor(buf, cp->cache_private);
2640 		}
2641 	}
2642 
2643 	kmem_slab_free(cp, buf);
2644 }
2645 
2646 /*
2647  * Used when there's no room to free a buffer to the per-CPU cache.
2648  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2649  * caller should try freeing to the per-CPU cache again.
2650  * Note that we don't directly install the magazine in the cpu cache,
2651  * since its state may have changed wildly while the lock was dropped.
2652  */
2653 static int
2654 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2655 {
2656 	kmem_magazine_t *emp;
2657 	kmem_magtype_t *mtp;
2658 
2659 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2660 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2661 	    ((uint_t)ccp->cc_rounds == -1)) &&
2662 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2663 	    ((uint_t)ccp->cc_prounds == -1)));
2664 
2665 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2666 	if (emp != NULL) {
2667 		if (ccp->cc_ploaded != NULL)
2668 			kmem_depot_free(cp, &cp->cache_full,
2669 			    ccp->cc_ploaded);
2670 		kmem_cpu_reload(ccp, emp, 0);
2671 		return (1);
2672 	}
2673 	/*
2674 	 * There are no empty magazines in the depot,
2675 	 * so try to allocate a new one.  We must drop all locks
2676 	 * across kmem_cache_alloc() because lower layers may
2677 	 * attempt to allocate from this cache.
2678 	 */
2679 	mtp = cp->cache_magtype;
2680 	mutex_exit(&ccp->cc_lock);
2681 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2682 	mutex_enter(&ccp->cc_lock);
2683 
2684 	if (emp != NULL) {
2685 		/*
2686 		 * We successfully allocated an empty magazine.
2687 		 * However, we had to drop ccp->cc_lock to do it,
2688 		 * so the cache's magazine size may have changed.
2689 		 * If so, free the magazine and try again.
2690 		 */
2691 		if (ccp->cc_magsize != mtp->mt_magsize) {
2692 			mutex_exit(&ccp->cc_lock);
2693 			kmem_cache_free(mtp->mt_cache, emp);
2694 			mutex_enter(&ccp->cc_lock);
2695 			return (1);
2696 		}
2697 
2698 		/*
2699 		 * We got a magazine of the right size.  Add it to
2700 		 * the depot and try the whole dance again.
2701 		 */
2702 		kmem_depot_free(cp, &cp->cache_empty, emp);
2703 		return (1);
2704 	}
2705 
2706 	/*
2707 	 * We couldn't allocate an empty magazine,
2708 	 * so fall through to the slab layer.
2709 	 */
2710 	return (0);
2711 }
2712 
2713 /*
2714  * Free a constructed object to cache cp.
2715  */
2716 void
2717 kmem_cache_free(kmem_cache_t *cp, void *buf)
2718 {
2719 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2720 
2721 	/*
2722 	 * The client must not free either of the buffers passed to the move
2723 	 * callback function.
2724 	 */
2725 	ASSERT(cp->cache_defrag == NULL ||
2726 	    cp->cache_defrag->kmd_thread != curthread ||
2727 	    (buf != cp->cache_defrag->kmd_from_buf &&
2728 	    buf != cp->cache_defrag->kmd_to_buf));
2729 
2730 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2731 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2732 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2733 			/* log it so that we can warn about it */
2734 			KDI_LOG(cp, kdl_unsafe);
2735 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2736 			return;
2737 		}
2738 		if (ccp->cc_flags & KMF_BUFTAG) {
2739 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2740 				return;
2741 		}
2742 	}
2743 
2744 	mutex_enter(&ccp->cc_lock);
2745 	/*
2746 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2747 	 */
2748 	for (;;) {
2749 		/*
2750 		 * If there's a slot available in the current CPU's
2751 		 * loaded magazine, just put the object there and return.
2752 		 */
2753 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2754 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2755 			ccp->cc_free++;
2756 			mutex_exit(&ccp->cc_lock);
2757 			return;
2758 		}
2759 
2760 		/*
2761 		 * The loaded magazine is full.  If the previously loaded
2762 		 * magazine was empty, exchange them and try again.
2763 		 */
2764 		if (ccp->cc_prounds == 0) {
2765 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2766 			continue;
2767 		}
2768 
2769 		/*
2770 		 * If the magazine layer is disabled, break out now.
2771 		 */
2772 		if (ccp->cc_magsize == 0)
2773 			break;
2774 
2775 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2776 			/*
2777 			 * We couldn't free our constructed object to the
2778 			 * magazine layer, so apply its destructor and free it
2779 			 * to the slab layer.
2780 			 */
2781 			break;
2782 		}
2783 	}
2784 	mutex_exit(&ccp->cc_lock);
2785 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2786 }
2787 
2788 static void
2789 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2790 {
2791 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2792 	int cache_flags = cp->cache_flags;
2793 
2794 	kmem_bufctl_t *next, *head;
2795 	size_t nbufs;
2796 
2797 	/*
2798 	 * Completely allocate the newly created slab and put the pre-allocated
2799 	 * buffers in magazines. Any of the buffers that cannot be put in
2800 	 * magazines must be returned to the slab.
2801 	 */
2802 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2803 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2804 	ASSERT(cp->cache_constructor == NULL);
2805 	ASSERT(sp->slab_cache == cp);
2806 	ASSERT(sp->slab_refcnt == 1);
2807 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2808 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2809 
2810 	head = sp->slab_head;
2811 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2812 	sp->slab_head = NULL;
2813 	sp->slab_refcnt += nbufs;
2814 	cp->cache_bufslab -= nbufs;
2815 	cp->cache_slab_alloc += nbufs;
2816 	list_insert_head(&cp->cache_complete_slabs, sp);
2817 	cp->cache_complete_slab_count++;
2818 	mutex_exit(&cp->cache_lock);
2819 	mutex_enter(&ccp->cc_lock);
2820 
2821 	while (head != NULL) {
2822 		void *buf = KMEM_BUF(cp, head);
2823 		/*
2824 		 * If there's a slot available in the current CPU's
2825 		 * loaded magazine, just put the object there and
2826 		 * continue.
2827 		 */
2828 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2829 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2830 			    buf;
2831 			ccp->cc_free++;
2832 			nbufs--;
2833 			head = head->bc_next;
2834 			continue;
2835 		}
2836 
2837 		/*
2838 		 * The loaded magazine is full.  If the previously
2839 		 * loaded magazine was empty, exchange them and try
2840 		 * again.
2841 		 */
2842 		if (ccp->cc_prounds == 0) {
2843 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2844 			    ccp->cc_prounds);
2845 			continue;
2846 		}
2847 
2848 		/*
2849 		 * If the magazine layer is disabled, break out now.
2850 		 */
2851 
2852 		if (ccp->cc_magsize == 0) {
2853 			break;
2854 		}
2855 
2856 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2857 			break;
2858 	}
2859 	mutex_exit(&ccp->cc_lock);
2860 	if (nbufs != 0) {
2861 		ASSERT(head != NULL);
2862 
2863 		/*
2864 		 * If there was a failure, return remaining objects to
2865 		 * the slab
2866 		 */
2867 		while (head != NULL) {
2868 			ASSERT(nbufs != 0);
2869 			next = head->bc_next;
2870 			head->bc_next = NULL;
2871 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2872 			head = next;
2873 			nbufs--;
2874 		}
2875 	}
2876 	ASSERT(head == NULL);
2877 	ASSERT(nbufs == 0);
2878 	mutex_enter(&cp->cache_lock);
2879 }
2880 
2881 void *
2882 kmem_zalloc(size_t size, int kmflag)
2883 {
2884 	size_t index;
2885 	void *buf;
2886 
2887 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2888 		kmem_cache_t *cp = kmem_alloc_table[index];
2889 		buf = kmem_cache_alloc(cp, kmflag);
2890 		if (buf != NULL) {
2891 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2892 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2893 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2894 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2895 
2896 				if (cp->cache_flags & KMF_LITE) {
2897 					KMEM_BUFTAG_LITE_ENTER(btp,
2898 					    kmem_lite_count, caller());
2899 				}
2900 			}
2901 			bzero(buf, size);
2902 		}
2903 	} else {
2904 		buf = kmem_alloc(size, kmflag);
2905 		if (buf != NULL)
2906 			bzero(buf, size);
2907 	}
2908 	return (buf);
2909 }
2910 
2911 void *
2912 kmem_alloc(size_t size, int kmflag)
2913 {
2914 	size_t index;
2915 	kmem_cache_t *cp;
2916 	void *buf;
2917 
2918 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2919 		cp = kmem_alloc_table[index];
2920 		/* fall through to kmem_cache_alloc() */
2921 
2922 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2923 	    kmem_big_alloc_table_max) {
2924 		cp = kmem_big_alloc_table[index];
2925 		/* fall through to kmem_cache_alloc() */
2926 
2927 	} else {
2928 		if (size == 0)
2929 			return (NULL);
2930 
2931 		buf = vmem_alloc(kmem_oversize_arena, size,
2932 		    kmflag & KM_VMFLAGS);
2933 		if (buf == NULL)
2934 			kmem_log_event(kmem_failure_log, NULL, NULL,
2935 			    (void *)size);
2936 		else if (KMEM_DUMP(kmem_slab_cache)) {
2937 			/* stats for dump intercept */
2938 			kmem_dump_oversize_allocs++;
2939 			if (size > kmem_dump_oversize_max)
2940 				kmem_dump_oversize_max = size;
2941 		}
2942 		return (buf);
2943 	}
2944 
2945 	buf = kmem_cache_alloc(cp, kmflag);
2946 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2947 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2948 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2949 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2950 
2951 		if (cp->cache_flags & KMF_LITE) {
2952 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2953 		}
2954 	}
2955 	return (buf);
2956 }
2957 
2958 void
2959 kmem_free(void *buf, size_t size)
2960 {
2961 	size_t index;
2962 	kmem_cache_t *cp;
2963 
2964 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2965 		cp = kmem_alloc_table[index];
2966 		/* fall through to kmem_cache_free() */
2967 
2968 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2969 	    kmem_big_alloc_table_max) {
2970 		cp = kmem_big_alloc_table[index];
2971 		/* fall through to kmem_cache_free() */
2972 
2973 	} else {
2974 		EQUIV(buf == NULL, size == 0);
2975 		if (buf == NULL && size == 0)
2976 			return;
2977 		vmem_free(kmem_oversize_arena, buf, size);
2978 		return;
2979 	}
2980 
2981 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2982 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2983 		uint32_t *ip = (uint32_t *)btp;
2984 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2985 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2986 				kmem_error(KMERR_DUPFREE, cp, buf);
2987 				return;
2988 			}
2989 			if (KMEM_SIZE_VALID(ip[1])) {
2990 				ip[0] = KMEM_SIZE_ENCODE(size);
2991 				kmem_error(KMERR_BADSIZE, cp, buf);
2992 			} else {
2993 				kmem_error(KMERR_REDZONE, cp, buf);
2994 			}
2995 			return;
2996 		}
2997 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2998 			kmem_error(KMERR_REDZONE, cp, buf);
2999 			return;
3000 		}
3001 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
3002 		if (cp->cache_flags & KMF_LITE) {
3003 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3004 			    caller());
3005 		}
3006 	}
3007 	kmem_cache_free(cp, buf);
3008 }
3009 
3010 void *
3011 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3012 {
3013 	size_t realsize = size + vmp->vm_quantum;
3014 	void *addr;
3015 
3016 	/*
3017 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3018 	 * vm_quantum will cause integer wraparound.  Check for this, and
3019 	 * blow off the firewall page in this case.  Note that such a
3020 	 * giant allocation (the entire kernel address space) can never
3021 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3022 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
3023 	 * corresponding check in kmem_firewall_va_free().
3024 	 */
3025 	if (realsize < size)
3026 		realsize = size;
3027 
3028 	/*
3029 	 * While boot still owns resource management, make sure that this
3030 	 * redzone virtual address allocation is properly accounted for in
3031 	 * OBPs "virtual-memory" "available" lists because we're
3032 	 * effectively claiming them for a red zone.  If we don't do this,
3033 	 * the available lists become too fragmented and too large for the
3034 	 * current boot/kernel memory list interface.
3035 	 */
3036 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3037 
3038 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3039 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3040 
3041 	return (addr);
3042 }
3043 
3044 void
3045 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3046 {
3047 	ASSERT((kvseg.s_base == NULL ?
3048 	    va_to_pfn((char *)addr + size) :
3049 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3050 
3051 	vmem_free(vmp, addr, size + vmp->vm_quantum);
3052 }
3053 
3054 /*
3055  * Try to allocate at least `size' bytes of memory without sleeping or
3056  * panicking. Return actual allocated size in `asize'. If allocation failed,
3057  * try final allocation with sleep or panic allowed.
3058  */
3059 void *
3060 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3061 {
3062 	void *p;
3063 
3064 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3065 	do {
3066 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3067 		if (p != NULL)
3068 			return (p);
3069 		*asize += KMEM_ALIGN;
3070 	} while (*asize <= PAGESIZE);
3071 
3072 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3073 	return (kmem_alloc(*asize, kmflag));
3074 }
3075 
3076 /*
3077  * Reclaim all unused memory from a cache.
3078  */
3079 static void
3080 kmem_cache_reap(kmem_cache_t *cp)
3081 {
3082 	ASSERT(taskq_member(kmem_taskq, curthread));
3083 	cp->cache_reap++;
3084 
3085 	/*
3086 	 * Ask the cache's owner to free some memory if possible.
3087 	 * The idea is to handle things like the inode cache, which
3088 	 * typically sits on a bunch of memory that it doesn't truly
3089 	 * *need*.  Reclaim policy is entirely up to the owner; this
3090 	 * callback is just an advisory plea for help.
3091 	 */
3092 	if (cp->cache_reclaim != NULL) {
3093 		long delta;
3094 
3095 		/*
3096 		 * Reclaimed memory should be reapable (not included in the
3097 		 * depot's working set).
3098 		 */
3099 		delta = cp->cache_full.ml_total;
3100 		cp->cache_reclaim(cp->cache_private);
3101 		delta = cp->cache_full.ml_total - delta;
3102 		if (delta > 0) {
3103 			mutex_enter(&cp->cache_depot_lock);
3104 			cp->cache_full.ml_reaplimit += delta;
3105 			cp->cache_full.ml_min += delta;
3106 			mutex_exit(&cp->cache_depot_lock);
3107 		}
3108 	}
3109 
3110 	kmem_depot_ws_reap(cp);
3111 
3112 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3113 		kmem_cache_defrag(cp);
3114 	}
3115 }
3116 
3117 static void
3118 kmem_reap_timeout(void *flag_arg)
3119 {
3120 	uint32_t *flag = (uint32_t *)flag_arg;
3121 
3122 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3123 	*flag = 0;
3124 }
3125 
3126 static void
3127 kmem_reap_done(void *flag)
3128 {
3129 	if (!callout_init_done) {
3130 		/* can't schedule a timeout at this point */
3131 		kmem_reap_timeout(flag);
3132 	} else {
3133 		(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3134 	}
3135 }
3136 
3137 static void
3138 kmem_reap_start(void *flag)
3139 {
3140 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3141 
3142 	if (flag == &kmem_reaping) {
3143 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3144 		/*
3145 		 * if we have segkp under heap, reap segkp cache.
3146 		 */
3147 		if (segkp_fromheap)
3148 			segkp_cache_free();
3149 	}
3150 	else
3151 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3152 
3153 	/*
3154 	 * We use taskq_dispatch() to schedule a timeout to clear
3155 	 * the flag so that kmem_reap() becomes self-throttling:
3156 	 * we won't reap again until the current reap completes *and*
3157 	 * at least kmem_reap_interval ticks have elapsed.
3158 	 */
3159 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3160 		kmem_reap_done(flag);
3161 }
3162 
3163 static void
3164 kmem_reap_common(void *flag_arg)
3165 {
3166 	uint32_t *flag = (uint32_t *)flag_arg;
3167 
3168 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3169 	    atomic_cas_32(flag, 0, 1) != 0)
3170 		return;
3171 
3172 	/*
3173 	 * It may not be kosher to do memory allocation when a reap is called
3174 	 * is called (for example, if vmem_populate() is in the call chain).
3175 	 * So we start the reap going with a TQ_NOALLOC dispatch.  If the
3176 	 * dispatch fails, we reset the flag, and the next reap will try again.
3177 	 */
3178 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3179 		*flag = 0;
3180 }
3181 
3182 /*
3183  * Reclaim all unused memory from all caches.  Called from the VM system
3184  * when memory gets tight.
3185  */
3186 void
3187 kmem_reap(void)
3188 {
3189 	kmem_reap_common(&kmem_reaping);
3190 }
3191 
3192 /*
3193  * Reclaim all unused memory from identifier arenas, called when a vmem
3194  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3195  * cannot help with identifier exhaustion, we avoid both a large amount of
3196  * work and unwanted side-effects from reclaim callbacks.
3197  */
3198 void
3199 kmem_reap_idspace(void)
3200 {
3201 	kmem_reap_common(&kmem_reaping_idspace);
3202 }
3203 
3204 /*
3205  * Purge all magazines from a cache and set its magazine limit to zero.
3206  * All calls are serialized by the kmem_taskq lock, except for the final
3207  * call from kmem_cache_destroy().
3208  */
3209 static void
3210 kmem_cache_magazine_purge(kmem_cache_t *cp)
3211 {
3212 	kmem_cpu_cache_t *ccp;
3213 	kmem_magazine_t *mp, *pmp;
3214 	int rounds, prounds, cpu_seqid;
3215 
3216 	ASSERT(!list_link_active(&cp->cache_link) ||
3217 	    taskq_member(kmem_taskq, curthread));
3218 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3219 
3220 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3221 		ccp = &cp->cache_cpu[cpu_seqid];
3222 
3223 		mutex_enter(&ccp->cc_lock);
3224 		mp = ccp->cc_loaded;
3225 		pmp = ccp->cc_ploaded;
3226 		rounds = ccp->cc_rounds;
3227 		prounds = ccp->cc_prounds;
3228 		ccp->cc_loaded = NULL;
3229 		ccp->cc_ploaded = NULL;
3230 		ccp->cc_rounds = -1;
3231 		ccp->cc_prounds = -1;
3232 		ccp->cc_magsize = 0;
3233 		mutex_exit(&ccp->cc_lock);
3234 
3235 		if (mp)
3236 			kmem_magazine_destroy(cp, mp, rounds);
3237 		if (pmp)
3238 			kmem_magazine_destroy(cp, pmp, prounds);
3239 	}
3240 
3241 	/*
3242 	 * Updating the working set statistics twice in a row has the
3243 	 * effect of setting the working set size to zero, so everything
3244 	 * is eligible for reaping.
3245 	 */
3246 	kmem_depot_ws_update(cp);
3247 	kmem_depot_ws_update(cp);
3248 
3249 	kmem_depot_ws_reap(cp);
3250 }
3251 
3252 /*
3253  * Enable per-cpu magazines on a cache.
3254  */
3255 static void
3256 kmem_cache_magazine_enable(kmem_cache_t *cp)
3257 {
3258 	int cpu_seqid;
3259 
3260 	if (cp->cache_flags & KMF_NOMAGAZINE)
3261 		return;
3262 
3263 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3264 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3265 		mutex_enter(&ccp->cc_lock);
3266 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3267 		mutex_exit(&ccp->cc_lock);
3268 	}
3269 
3270 }
3271 
3272 /*
3273  * Reap (almost) everything right now.  See kmem_cache_magazine_purge()
3274  * for explanation of the back-to-back kmem_depot_ws_update() calls.
3275  */
3276 void
3277 kmem_cache_reap_now(kmem_cache_t *cp)
3278 {
3279 	ASSERT(list_link_active(&cp->cache_link));
3280 
3281 	kmem_depot_ws_update(cp);
3282 	kmem_depot_ws_update(cp);
3283 
3284 	(void) taskq_dispatch(kmem_taskq,
3285 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3286 	taskq_wait(kmem_taskq);
3287 }
3288 
3289 /*
3290  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3291  * provide a higher transfer rate with the depot, while smaller magazines
3292  * reduce memory consumption.  Magazine resizing is an expensive operation;
3293  * it should not be done frequently.
3294  *
3295  * Changes to the magazine size are serialized by the kmem_taskq lock.
3296  *
3297  * Note: at present this only grows the magazine size.  It might be useful
3298  * to allow shrinkage too.
3299  */
3300 static void
3301 kmem_cache_magazine_resize(kmem_cache_t *cp)
3302 {
3303 	kmem_magtype_t *mtp = cp->cache_magtype;
3304 
3305 	ASSERT(taskq_member(kmem_taskq, curthread));
3306 
3307 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3308 		kmem_cache_magazine_purge(cp);
3309 		mutex_enter(&cp->cache_depot_lock);
3310 		cp->cache_magtype = ++mtp;
3311 		cp->cache_depot_contention_prev =
3312 		    cp->cache_depot_contention + INT_MAX;
3313 		mutex_exit(&cp->cache_depot_lock);
3314 		kmem_cache_magazine_enable(cp);
3315 	}
3316 }
3317 
3318 /*
3319  * Rescale a cache's hash table, so that the table size is roughly the
3320  * cache size.  We want the average lookup time to be extremely small.
3321  */
3322 static void
3323 kmem_hash_rescale(kmem_cache_t *cp)
3324 {
3325 	kmem_bufctl_t **old_table, **new_table, *bcp;
3326 	size_t old_size, new_size, h;
3327 
3328 	ASSERT(taskq_member(kmem_taskq, curthread));
3329 
3330 	new_size = MAX(KMEM_HASH_INITIAL,
3331 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3332 	old_size = cp->cache_hash_mask + 1;
3333 
3334 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3335 		return;
3336 
3337 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3338 	    VM_NOSLEEP);
3339 	if (new_table == NULL)
3340 		return;
3341 	bzero(new_table, new_size * sizeof (void *));
3342 
3343 	mutex_enter(&cp->cache_lock);
3344 
3345 	old_size = cp->cache_hash_mask + 1;
3346 	old_table = cp->cache_hash_table;
3347 
3348 	cp->cache_hash_mask = new_size - 1;
3349 	cp->cache_hash_table = new_table;
3350 	cp->cache_rescale++;
3351 
3352 	for (h = 0; h < old_size; h++) {
3353 		bcp = old_table[h];
3354 		while (bcp != NULL) {
3355 			void *addr = bcp->bc_addr;
3356 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3357 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3358 			bcp->bc_next = *hash_bucket;
3359 			*hash_bucket = bcp;
3360 			bcp = next_bcp;
3361 		}
3362 	}
3363 
3364 	mutex_exit(&cp->cache_lock);
3365 
3366 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3367 }
3368 
3369 /*
3370  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3371  * update, magazine resizing, and slab consolidation.
3372  */
3373 static void
3374 kmem_cache_update(kmem_cache_t *cp)
3375 {
3376 	int need_hash_rescale = 0;
3377 	int need_magazine_resize = 0;
3378 
3379 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3380 
3381 	/*
3382 	 * If the cache has become much larger or smaller than its hash table,
3383 	 * fire off a request to rescale the hash table.
3384 	 */
3385 	mutex_enter(&cp->cache_lock);
3386 
3387 	if ((cp->cache_flags & KMF_HASH) &&
3388 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3389 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3390 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3391 		need_hash_rescale = 1;
3392 
3393 	mutex_exit(&cp->cache_lock);
3394 
3395 	/*
3396 	 * Update the depot working set statistics.
3397 	 */
3398 	kmem_depot_ws_update(cp);
3399 
3400 	/*
3401 	 * If there's a lot of contention in the depot,
3402 	 * increase the magazine size.
3403 	 */
3404 	mutex_enter(&cp->cache_depot_lock);
3405 
3406 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3407 	    (int)(cp->cache_depot_contention -
3408 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3409 		need_magazine_resize = 1;
3410 
3411 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3412 
3413 	mutex_exit(&cp->cache_depot_lock);
3414 
3415 	if (need_hash_rescale)
3416 		(void) taskq_dispatch(kmem_taskq,
3417 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3418 
3419 	if (need_magazine_resize)
3420 		(void) taskq_dispatch(kmem_taskq,
3421 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3422 
3423 	if (cp->cache_defrag != NULL)
3424 		(void) taskq_dispatch(kmem_taskq,
3425 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3426 }
3427 
3428 static void kmem_update(void *);
3429 
3430 static void
3431 kmem_update_timeout(void *dummy)
3432 {
3433 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3434 }
3435 
3436 static void
3437 kmem_update(void *dummy)
3438 {
3439 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3440 
3441 	/*
3442 	 * We use taskq_dispatch() to reschedule the timeout so that
3443 	 * kmem_update() becomes self-throttling: it won't schedule
3444 	 * new tasks until all previous tasks have completed.
3445 	 */
3446 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3447 		kmem_update_timeout(NULL);
3448 }
3449 
3450 static int
3451 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3452 {
3453 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3454 	kmem_cache_t *cp = ksp->ks_private;
3455 	uint64_t cpu_buf_avail;
3456 	uint64_t buf_avail = 0;
3457 	int cpu_seqid;
3458 	long reap;
3459 
3460 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3461 
3462 	if (rw == KSTAT_WRITE)
3463 		return (EACCES);
3464 
3465 	mutex_enter(&cp->cache_lock);
3466 
3467 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3468 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3469 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3470 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3471 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3472 
3473 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3474 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3475 
3476 		mutex_enter(&ccp->cc_lock);
3477 
3478 		cpu_buf_avail = 0;
3479 		if (ccp->cc_rounds > 0)
3480 			cpu_buf_avail += ccp->cc_rounds;
3481 		if (ccp->cc_prounds > 0)
3482 			cpu_buf_avail += ccp->cc_prounds;
3483 
3484 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3485 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3486 		buf_avail			+= cpu_buf_avail;
3487 
3488 		mutex_exit(&ccp->cc_lock);
3489 	}
3490 
3491 	mutex_enter(&cp->cache_depot_lock);
3492 
3493 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3494 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3495 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3496 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3497 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3498 	kmcp->kmc_magazine_size.value.ui64	=
3499 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3500 	    0 : cp->cache_magtype->mt_magsize;
3501 
3502 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3503 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3504 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3505 
3506 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3507 	reap = MIN(reap, cp->cache_full.ml_total);
3508 
3509 	mutex_exit(&cp->cache_depot_lock);
3510 
3511 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3512 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3513 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3514 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3515 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3516 	buf_avail += cp->cache_bufslab;
3517 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3518 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3519 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3520 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3521 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3522 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3523 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3524 	    cp->cache_hash_mask + 1 : 0;
3525 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3526 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3527 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3528 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3529 
3530 	if (cp->cache_defrag == NULL) {
3531 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3532 		kmcp->kmc_move_yes.value.ui64		= 0;
3533 		kmcp->kmc_move_no.value.ui64		= 0;
3534 		kmcp->kmc_move_later.value.ui64		= 0;
3535 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3536 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3537 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3538 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3539 		kmcp->kmc_defrag.value.ui64		= 0;
3540 		kmcp->kmc_scan.value.ui64		= 0;
3541 		kmcp->kmc_move_reclaimable.value.ui64	= 0;
3542 	} else {
3543 		int64_t reclaimable;
3544 
3545 		kmem_defrag_t *kd = cp->cache_defrag;
3546 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3547 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3548 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3549 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3550 		kmcp->kmc_move_dont_need.value.ui64	= kd->kmd_dont_need;
3551 		kmcp->kmc_move_dont_know.value.ui64	= kd->kmd_dont_know;
3552 		kmcp->kmc_move_hunt_found.value.ui64	= kd->kmd_hunt_found;
3553 		kmcp->kmc_move_slabs_freed.value.ui64	= kd->kmd_slabs_freed;
3554 		kmcp->kmc_defrag.value.ui64		= kd->kmd_defrags;
3555 		kmcp->kmc_scan.value.ui64		= kd->kmd_scans;
3556 
3557 		reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3558 		reclaimable = MAX(reclaimable, 0);
3559 		reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3560 		kmcp->kmc_move_reclaimable.value.ui64	= reclaimable;
3561 	}
3562 
3563 	mutex_exit(&cp->cache_lock);
3564 	return (0);
3565 }
3566 
3567 /*
3568  * Return a named statistic about a particular cache.
3569  * This shouldn't be called very often, so it's currently designed for
3570  * simplicity (leverages existing kstat support) rather than efficiency.
3571  */
3572 uint64_t
3573 kmem_cache_stat(kmem_cache_t *cp, char *name)
3574 {
3575 	int i;
3576 	kstat_t *ksp = cp->cache_kstat;
3577 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3578 	uint64_t value = 0;
3579 
3580 	if (ksp != NULL) {
3581 		mutex_enter(&kmem_cache_kstat_lock);
3582 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3583 		for (i = 0; i < ksp->ks_ndata; i++) {
3584 			if (strcmp(knp[i].name, name) == 0) {
3585 				value = knp[i].value.ui64;
3586 				break;
3587 			}
3588 		}
3589 		mutex_exit(&kmem_cache_kstat_lock);
3590 	}
3591 	return (value);
3592 }
3593 
3594 /*
3595  * Return an estimate of currently available kernel heap memory.
3596  * On 32-bit systems, physical memory may exceed virtual memory,
3597  * we just truncate the result at 1GB.
3598  */
3599 size_t
3600 kmem_avail(void)
3601 {
3602 	spgcnt_t rmem = availrmem - tune.t_minarmem;
3603 	spgcnt_t fmem = freemem - minfree;
3604 
3605 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3606 	    1 << (30 - PAGESHIFT))));
3607 }
3608 
3609 /*
3610  * Return the maximum amount of memory that is (in theory) allocatable
3611  * from the heap. This may be used as an estimate only since there
3612  * is no guarentee this space will still be available when an allocation
3613  * request is made, nor that the space may be allocated in one big request
3614  * due to kernel heap fragmentation.
3615  */
3616 size_t
3617 kmem_maxavail(void)
3618 {
3619 	spgcnt_t pmem = availrmem - tune.t_minarmem;
3620 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3621 
3622 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3623 }
3624 
3625 /*
3626  * Indicate whether memory-intensive kmem debugging is enabled.
3627  */
3628 int
3629 kmem_debugging(void)
3630 {
3631 	return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3632 }
3633 
3634 /* binning function, sorts finely at the two extremes */
3635 #define	KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift)				\
3636 	((((sp)->slab_refcnt <= (binshift)) ||				\
3637 	    (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift)))	\
3638 	    ? -(sp)->slab_refcnt					\
3639 	    : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3640 
3641 /*
3642  * Minimizing the number of partial slabs on the freelist minimizes
3643  * fragmentation (the ratio of unused buffers held by the slab layer). There are
3644  * two ways to get a slab off of the freelist: 1) free all the buffers on the
3645  * slab, and 2) allocate all the buffers on the slab. It follows that we want
3646  * the most-used slabs at the front of the list where they have the best chance
3647  * of being completely allocated, and the least-used slabs at a safe distance
3648  * from the front to improve the odds that the few remaining buffers will all be
3649  * freed before another allocation can tie up the slab. For that reason a slab
3650  * with a higher slab_refcnt sorts less than than a slab with a lower
3651  * slab_refcnt.
3652  *
3653  * However, if a slab has at least one buffer that is deemed unfreeable, we
3654  * would rather have that slab at the front of the list regardless of
3655  * slab_refcnt, since even one unfreeable buffer makes the entire slab
3656  * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3657  * callback, the slab is marked unfreeable for as long as it remains on the
3658  * freelist.
3659  */
3660 static int
3661 kmem_partial_slab_cmp(const void *p0, const void *p1)
3662 {
3663 	const kmem_cache_t *cp;
3664 	const kmem_slab_t *s0 = p0;
3665 	const kmem_slab_t *s1 = p1;
3666 	int w0, w1;
3667 	size_t binshift;
3668 
3669 	ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3670 	ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3671 	ASSERT(s0->slab_cache == s1->slab_cache);
3672 	cp = s1->slab_cache;
3673 	ASSERT(MUTEX_HELD(&cp->cache_lock));
3674 	binshift = cp->cache_partial_binshift;
3675 
3676 	/* weight of first slab */
3677 	w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3678 	if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3679 		w0 -= cp->cache_maxchunks;
3680 	}
3681 
3682 	/* weight of second slab */
3683 	w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3684 	if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3685 		w1 -= cp->cache_maxchunks;
3686 	}
3687 
3688 	if (w0 < w1)
3689 		return (-1);
3690 	if (w0 > w1)
3691 		return (1);
3692 
3693 	/* compare pointer values */
3694 	if ((uintptr_t)s0 < (uintptr_t)s1)
3695 		return (-1);
3696 	if ((uintptr_t)s0 > (uintptr_t)s1)
3697 		return (1);
3698 
3699 	return (0);
3700 }
3701 
3702 /*
3703  * It must be valid to call the destructor (if any) on a newly created object.
3704  * That is, the constructor (if any) must leave the object in a valid state for
3705  * the destructor.
3706  */
3707 kmem_cache_t *
3708 kmem_cache_create(
3709 	char *name,		/* descriptive name for this cache */
3710 	size_t bufsize,		/* size of the objects it manages */
3711 	size_t align,		/* required object alignment */
3712 	int (*constructor)(void *, void *, int), /* object constructor */
3713 	void (*destructor)(void *, void *),	/* object destructor */
3714 	void (*reclaim)(void *), /* memory reclaim callback */
3715 	void *private,		/* pass-thru arg for constr/destr/reclaim */
3716 	vmem_t *vmp,		/* vmem source for slab allocation */
3717 	int cflags)		/* cache creation flags */
3718 {
3719 	int cpu_seqid;
3720 	size_t chunksize;
3721 	kmem_cache_t *cp;
3722 	kmem_magtype_t *mtp;
3723 	size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3724 
3725 #ifdef	DEBUG
3726 	/*
3727 	 * Cache names should conform to the rules for valid C identifiers
3728 	 */
3729 	if (!strident_valid(name)) {
3730 		cmn_err(CE_CONT,
3731 		    "kmem_cache_create: '%s' is an invalid cache name\n"
3732 		    "cache names must conform to the rules for "
3733 		    "C identifiers\n", name);
3734 	}
3735 #endif	/* DEBUG */
3736 
3737 	if (vmp == NULL)
3738 		vmp = kmem_default_arena;
3739 
3740 	/*
3741 	 * If this kmem cache has an identifier vmem arena as its source, mark
3742 	 * it such to allow kmem_reap_idspace().
3743 	 */
3744 	ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
3745 	if (vmp->vm_cflags & VMC_IDENTIFIER)
3746 		cflags |= KMC_IDENTIFIER;
3747 
3748 	/*
3749 	 * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
3750 	 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3751 	 * false sharing of per-CPU data.
3752 	 */
3753 	cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3754 	    P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3755 	bzero(cp, csize);
3756 	list_link_init(&cp->cache_link);
3757 
3758 	if (align == 0)
3759 		align = KMEM_ALIGN;
3760 
3761 	/*
3762 	 * If we're not at least KMEM_ALIGN aligned, we can't use free
3763 	 * memory to hold bufctl information (because we can't safely
3764 	 * perform word loads and stores on it).
3765 	 */
3766 	if (align < KMEM_ALIGN)
3767 		cflags |= KMC_NOTOUCH;
3768 
3769 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum)
3770 		panic("kmem_cache_create: bad alignment %lu", align);
3771 
3772 	mutex_enter(&kmem_flags_lock);
3773 	if (kmem_flags & KMF_RANDOMIZE)
3774 		kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3775 		    KMF_RANDOMIZE;
3776 	cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3777 	mutex_exit(&kmem_flags_lock);
3778 
3779 	/*
3780 	 * Make sure all the various flags are reasonable.
3781 	 */
3782 	ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3783 
3784 	if (cp->cache_flags & KMF_LITE) {
3785 		if (bufsize >= kmem_lite_minsize &&
3786 		    align <= kmem_lite_maxalign &&
3787 		    P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3788 			cp->cache_flags |= KMF_BUFTAG;
3789 			cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3790 		} else {
3791 			cp->cache_flags &= ~KMF_DEBUG;
3792 		}
3793 	}
3794 
3795 	if (cp->cache_flags & KMF_DEADBEEF)
3796 		cp->cache_flags |= KMF_REDZONE;
3797 
3798 	if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3799 		cp->cache_flags |= KMF_NOMAGAZINE;
3800 
3801 	if (cflags & KMC_NODEBUG)
3802 		cp->cache_flags &= ~KMF_DEBUG;
3803 
3804 	if (cflags & KMC_NOTOUCH)
3805 		cp->cache_flags &= ~KMF_TOUCH;
3806 
3807 	if (cflags & KMC_PREFILL)
3808 		cp->cache_flags |= KMF_PREFILL;
3809 
3810 	if (cflags & KMC_NOHASH)
3811 		cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3812 
3813 	if (cflags & KMC_NOMAGAZINE)
3814 		cp->cache_flags |= KMF_NOMAGAZINE;
3815 
3816 	if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3817 		cp->cache_flags |= KMF_REDZONE;
3818 
3819 	if (!(cp->cache_flags & KMF_AUDIT))
3820 		cp->cache_flags &= ~KMF_CONTENTS;
3821 
3822 	if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3823 	    !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3824 		cp->cache_flags |= KMF_FIREWALL;
3825 
3826 	if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3827 		cp->cache_flags &= ~KMF_FIREWALL;
3828 
3829 	if (cp->cache_flags & KMF_FIREWALL) {
3830 		cp->cache_flags &= ~KMF_BUFTAG;
3831 		cp->cache_flags |= KMF_NOMAGAZINE;
3832 		ASSERT(vmp == kmem_default_arena);
3833 		vmp = kmem_firewall_arena;
3834 	}
3835 
3836 	/*
3837 	 * Set cache properties.
3838 	 */
3839 	(void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3840 	strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3841 	cp->cache_bufsize = bufsize;
3842 	cp->cache_align = align;
3843 	cp->cache_constructor = constructor;
3844 	cp->cache_destructor = destructor;
3845 	cp->cache_reclaim = reclaim;
3846 	cp->cache_private = private;
3847 	cp->cache_arena = vmp;
3848 	cp->cache_cflags = cflags;
3849 
3850 	/*
3851 	 * Determine the chunk size.
3852 	 */
3853 	chunksize = bufsize;
3854 
3855 	if (align >= KMEM_ALIGN) {
3856 		chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3857 		cp->cache_bufctl = chunksize - KMEM_ALIGN;
3858 	}
3859 
3860 	if (cp->cache_flags & KMF_BUFTAG) {
3861 		cp->cache_bufctl = chunksize;
3862 		cp->cache_buftag = chunksize;
3863 		if (cp->cache_flags & KMF_LITE)
3864 			chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3865 		else
3866 			chunksize += sizeof (kmem_buftag_t);
3867 	}
3868 
3869 	if (cp->cache_flags & KMF_DEADBEEF) {
3870 		cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3871 		if (cp->cache_flags & KMF_LITE)
3872 			cp->cache_verify = sizeof (uint64_t);
3873 	}
3874 
3875 	cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3876 
3877 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3878 
3879 	/*
3880 	 * Now that we know the chunk size, determine the optimal slab size.
3881 	 */
3882 	if (vmp == kmem_firewall_arena) {
3883 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3884 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
3885 		cp->cache_maxcolor = cp->cache_mincolor;
3886 		cp->cache_flags |= KMF_HASH;
3887 		ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3888 	} else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3889 	    !(cp->cache_flags & KMF_AUDIT) &&
3890 	    chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3891 		cp->cache_slabsize = vmp->vm_quantum;
3892 		cp->cache_mincolor = 0;
3893 		cp->cache_maxcolor =
3894 		    (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3895 		ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3896 		ASSERT(!(cp->cache_flags & KMF_AUDIT));
3897 	} else {
3898 		size_t chunks, bestfit, waste, slabsize;
3899 		size_t minwaste = LONG_MAX;
3900 
3901 		for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3902 			slabsize = P2ROUNDUP(chunksize * chunks,
3903 			    vmp->vm_quantum);
3904 			chunks = slabsize / chunksize;
3905 			waste = (slabsize % chunksize) / chunks;
3906 			if (waste < minwaste) {
3907 				minwaste = waste;
3908 				bestfit = slabsize;
3909 			}
3910 		}
3911 		if (cflags & KMC_QCACHE)
3912 			bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3913 		cp->cache_slabsize = bestfit;
3914 		cp->cache_mincolor = 0;
3915 		cp->cache_maxcolor = bestfit % chunksize;
3916 		cp->cache_flags |= KMF_HASH;
3917 	}
3918 
3919 	cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3920 	cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3921 
3922 	/*
3923 	 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3924 	 * there is a constructor avoids some tricky issues with debug setup
3925 	 * that may be revisited later. We cannot allow prefill in a
3926 	 * metadata cache because of potential recursion.
3927 	 */
3928 	if (vmp == kmem_msb_arena ||
3929 	    cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3930 	    cp->cache_constructor != NULL)
3931 		cp->cache_flags &= ~KMF_PREFILL;
3932 
3933 	if (cp->cache_flags & KMF_HASH) {
3934 		ASSERT(!(cflags & KMC_NOHASH));
3935 		cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3936 		    kmem_bufctl_audit_cache : kmem_bufctl_cache;
3937 	}
3938 
3939 	if (cp->cache_maxcolor >= vmp->vm_quantum)
3940 		cp->cache_maxcolor = vmp->vm_quantum - 1;
3941 
3942 	cp->cache_color = cp->cache_mincolor;
3943 
3944 	/*
3945 	 * Initialize the rest of the slab layer.
3946 	 */
3947 	mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3948 
3949 	avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3950 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3951 	/* LINTED: E_TRUE_LOGICAL_EXPR */
3952 	ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3953 	/* reuse partial slab AVL linkage for complete slab list linkage */
3954 	list_create(&cp->cache_complete_slabs,
3955 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3956 
3957 	if (cp->cache_flags & KMF_HASH) {
3958 		cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3959 		    KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3960 		bzero(cp->cache_hash_table,
3961 		    KMEM_HASH_INITIAL * sizeof (void *));
3962 		cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3963 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3964 	}
3965 
3966 	/*
3967 	 * Initialize the depot.
3968 	 */
3969 	mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3970 
3971 	for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3972 		continue;
3973 
3974 	cp->cache_magtype = mtp;
3975 
3976 	/*
3977 	 * Initialize the CPU layer.
3978 	 */
3979 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3980 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3981 		mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3982 		ccp->cc_flags = cp->cache_flags;
3983 		ccp->cc_rounds = -1;
3984 		ccp->cc_prounds = -1;
3985 	}
3986 
3987 	/*
3988 	 * Create the cache's kstats.
3989 	 */
3990 	if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
3991 	    "kmem_cache", KSTAT_TYPE_NAMED,
3992 	    sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
3993 	    KSTAT_FLAG_VIRTUAL)) != NULL) {
3994 		cp->cache_kstat->ks_data = &kmem_cache_kstat;
3995 		cp->cache_kstat->ks_update = kmem_cache_kstat_update;
3996 		cp->cache_kstat->ks_private = cp;
3997 		cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
3998 		kstat_install(cp->cache_kstat);
3999 	}
4000 
4001 	/*
4002 	 * Add the cache to the global list.  This makes it visible
4003 	 * to kmem_update(), so the cache must be ready for business.
4004 	 */
4005 	mutex_enter(&kmem_cache_lock);
4006 	list_insert_tail(&kmem_caches, cp);
4007 	mutex_exit(&kmem_cache_lock);
4008 
4009 	if (kmem_ready)
4010 		kmem_cache_magazine_enable(cp);
4011 
4012 	return (cp);
4013 }
4014 
4015 static int
4016 kmem_move_cmp(const void *buf, const void *p)
4017 {
4018 	const kmem_move_t *kmm = p;
4019 	uintptr_t v1 = (uintptr_t)buf;
4020 	uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
4021 	return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
4022 }
4023 
4024 static void
4025 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4026 {
4027 	kmd->kmd_reclaim_numer = 1;
4028 }
4029 
4030 /*
4031  * Initially, when choosing candidate slabs for buffers to move, we want to be
4032  * very selective and take only slabs that are less than
4033  * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4034  * slabs, then we raise the allocation ceiling incrementally. The reclaim
4035  * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4036  * longer fragmented.
4037  */
4038 static void
4039 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4040 {
4041 	if (direction > 0) {
4042 		/* make it easier to find a candidate slab */
4043 		if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4044 			kmd->kmd_reclaim_numer++;
4045 		}
4046 	} else {
4047 		/* be more selective */
4048 		if (kmd->kmd_reclaim_numer > 1) {
4049 			kmd->kmd_reclaim_numer--;
4050 		}
4051 	}
4052 }
4053 
4054 void
4055 kmem_cache_set_move(kmem_cache_t *cp,
4056     kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4057 {
4058 	kmem_defrag_t *defrag;
4059 
4060 	ASSERT(move != NULL);
4061 	/*
4062 	 * The consolidator does not support NOTOUCH caches because kmem cannot
4063 	 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4064 	 * a low order bit usable by clients to distinguish uninitialized memory
4065 	 * from known objects (see kmem_slab_create).
4066 	 */
4067 	ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4068 	ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4069 
4070 	/*
4071 	 * We should not be holding anyone's cache lock when calling
4072 	 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4073 	 * lock.
4074 	 */
4075 	defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4076 
4077 	mutex_enter(&cp->cache_lock);
4078 
4079 	if (KMEM_IS_MOVABLE(cp)) {
4080 		if (cp->cache_move == NULL) {
4081 			ASSERT(cp->cache_slab_alloc == 0);
4082 
4083 			cp->cache_defrag = defrag;
4084 			defrag = NULL; /* nothing to free */
4085 			bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4086 			avl_create(&cp->cache_defrag->kmd_moves_pending,
4087 			    kmem_move_cmp, sizeof (kmem_move_t),
4088 			    offsetof(kmem_move_t, kmm_entry));
4089 			/* LINTED: E_TRUE_LOGICAL_EXPR */
4090 			ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4091 			/* reuse the slab's AVL linkage for deadlist linkage */
4092 			list_create(&cp->cache_defrag->kmd_deadlist,
4093 			    sizeof (kmem_slab_t),
4094 			    offsetof(kmem_slab_t, slab_link));
4095 			kmem_reset_reclaim_threshold(cp->cache_defrag);
4096 		}
4097 		cp->cache_move = move;
4098 	}
4099 
4100 	mutex_exit(&cp->cache_lock);
4101 
4102 	if (defrag != NULL) {
4103 		kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4104 	}
4105 }
4106 
4107 void
4108 kmem_cache_destroy(kmem_cache_t *cp)
4109 {
4110 	int cpu_seqid;
4111 
4112 	/*
4113 	 * Remove the cache from the global cache list so that no one else
4114 	 * can schedule tasks on its behalf, wait for any pending tasks to
4115 	 * complete, purge the cache, and then destroy it.
4116 	 */
4117 	mutex_enter(&kmem_cache_lock);
4118 	list_remove(&kmem_caches, cp);
4119 	mutex_exit(&kmem_cache_lock);
4120 
4121 	if (kmem_taskq != NULL)
4122 		taskq_wait(kmem_taskq);
4123 	if (kmem_move_taskq != NULL)
4124 		taskq_wait(kmem_move_taskq);
4125 
4126 	kmem_cache_magazine_purge(cp);
4127 
4128 	mutex_enter(&cp->cache_lock);
4129 	if (cp->cache_buftotal != 0)
4130 		cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4131 		    cp->cache_name, (void *)cp);
4132 	if (cp->cache_defrag != NULL) {
4133 		avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4134 		list_destroy(&cp->cache_defrag->kmd_deadlist);
4135 		kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4136 		cp->cache_defrag = NULL;
4137 	}
4138 	/*
4139 	 * The cache is now dead.  There should be no further activity.  We
4140 	 * enforce this by setting land mines in the constructor, destructor,
4141 	 * reclaim, and move routines that induce a kernel text fault if
4142 	 * invoked.
4143 	 */
4144 	cp->cache_constructor = (int (*)(void *, void *, int))1;
4145 	cp->cache_destructor = (void (*)(void *, void *))2;
4146 	cp->cache_reclaim = (void (*)(void *))3;
4147 	cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4148 	mutex_exit(&cp->cache_lock);
4149 
4150 	kstat_delete(cp->cache_kstat);
4151 
4152 	if (cp->cache_hash_table != NULL)
4153 		vmem_free(kmem_hash_arena, cp->cache_hash_table,
4154 		    (cp->cache_hash_mask + 1) * sizeof (void *));
4155 
4156 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4157 		mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4158 
4159 	mutex_destroy(&cp->cache_depot_lock);
4160 	mutex_destroy(&cp->cache_lock);
4161 
4162 	vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4163 }
4164 
4165 /*ARGSUSED*/
4166 static int
4167 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4168 {
4169 	ASSERT(MUTEX_HELD(&cpu_lock));
4170 	if (what == CPU_UNCONFIG) {
4171 		kmem_cache_applyall(kmem_cache_magazine_purge,
4172 		    kmem_taskq, TQ_SLEEP);
4173 		kmem_cache_applyall(kmem_cache_magazine_enable,
4174 		    kmem_taskq, TQ_SLEEP);
4175 	}
4176 	return (0);
4177 }
4178 
4179 static void
4180 kmem_alloc_caches_create(const int *array, size_t count,
4181     kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4182 {
4183 	char name[KMEM_CACHE_NAMELEN + 1];
4184 	size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4185 	size_t size = table_unit;
4186 	int i;
4187 
4188 	for (i = 0; i < count; i++) {
4189 		size_t cache_size = array[i];
4190 		size_t align = KMEM_ALIGN;
4191 		kmem_cache_t *cp;
4192 
4193 		/* if the table has an entry for maxbuf, we're done */
4194 		if (size > maxbuf)
4195 			break;
4196 
4197 		/* cache size must be a multiple of the table unit */
4198 		ASSERT(P2PHASE(cache_size, table_unit) == 0);
4199 
4200 		/*
4201 		 * If they allocate a multiple of the coherency granularity,
4202 		 * they get a coherency-granularity-aligned address.
4203 		 */
4204 		if (IS_P2ALIGNED(cache_size, 64))
4205 			align = 64;
4206 		if (IS_P2ALIGNED(cache_size, PAGESIZE))
4207 			align = PAGESIZE;
4208 		(void) snprintf(name, sizeof (name),
4209 		    "kmem_alloc_%lu", cache_size);
4210 		cp = kmem_cache_create(name, cache_size, align,
4211 		    NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4212 
4213 		while (size <= cache_size) {
4214 			alloc_table[(size - 1) >> shift] = cp;
4215 			size += table_unit;
4216 		}
4217 	}
4218 
4219 	ASSERT(size > maxbuf);		/* i.e. maxbuf <= max(cache_size) */
4220 }
4221 
4222 static void
4223 kmem_cache_init(int pass, int use_large_pages)
4224 {
4225 	int i;
4226 	size_t maxbuf;
4227 	kmem_magtype_t *mtp;
4228 
4229 	for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4230 		char name[KMEM_CACHE_NAMELEN + 1];
4231 
4232 		mtp = &kmem_magtype[i];
4233 		(void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4234 		mtp->mt_cache = kmem_cache_create(name,
4235 		    (mtp->mt_magsize + 1) * sizeof (void *),
4236 		    mtp->mt_align, NULL, NULL, NULL, NULL,
4237 		    kmem_msb_arena, KMC_NOHASH);
4238 	}
4239 
4240 	kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4241 	    sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4242 	    kmem_msb_arena, KMC_NOHASH);
4243 
4244 	kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4245 	    sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4246 	    kmem_msb_arena, KMC_NOHASH);
4247 
4248 	kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4249 	    sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4250 	    kmem_msb_arena, KMC_NOHASH);
4251 
4252 	if (pass == 2) {
4253 		kmem_va_arena = vmem_create("kmem_va",
4254 		    NULL, 0, PAGESIZE,
4255 		    vmem_alloc, vmem_free, heap_arena,
4256 		    8 * PAGESIZE, VM_SLEEP);
4257 
4258 		if (use_large_pages) {
4259 			kmem_default_arena = vmem_xcreate("kmem_default",
4260 			    NULL, 0, PAGESIZE,
4261 			    segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4262 			    0, VMC_DUMPSAFE | VM_SLEEP);
4263 		} else {
4264 			kmem_default_arena = vmem_create("kmem_default",
4265 			    NULL, 0, PAGESIZE,
4266 			    segkmem_alloc, segkmem_free, kmem_va_arena,
4267 			    0, VMC_DUMPSAFE | VM_SLEEP);
4268 		}
4269 
4270 		/* Figure out what our maximum cache size is */
4271 		maxbuf = kmem_max_cached;
4272 		if (maxbuf <= KMEM_MAXBUF) {
4273 			maxbuf = 0;
4274 			kmem_max_cached = KMEM_MAXBUF;
4275 		} else {
4276 			size_t size = 0;
4277 			size_t max =
4278 			    sizeof (kmem_big_alloc_sizes) / sizeof (int);
4279 			/*
4280 			 * Round maxbuf up to an existing cache size.  If maxbuf
4281 			 * is larger than the largest cache, we truncate it to
4282 			 * the largest cache's size.
4283 			 */
4284 			for (i = 0; i < max; i++) {
4285 				size = kmem_big_alloc_sizes[i];
4286 				if (maxbuf <= size)
4287 					break;
4288 			}
4289 			kmem_max_cached = maxbuf = size;
4290 		}
4291 
4292 		/*
4293 		 * The big alloc table may not be completely overwritten, so
4294 		 * we clear out any stale cache pointers from the first pass.
4295 		 */
4296 		bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4297 	} else {
4298 		/*
4299 		 * During the first pass, the kmem_alloc_* caches
4300 		 * are treated as metadata.
4301 		 */
4302 		kmem_default_arena = kmem_msb_arena;
4303 		maxbuf = KMEM_BIG_MAXBUF_32BIT;
4304 	}
4305 
4306 	/*
4307 	 * Set up the default caches to back kmem_alloc()
4308 	 */
4309 	kmem_alloc_caches_create(
4310 	    kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4311 	    kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4312 
4313 	kmem_alloc_caches_create(
4314 	    kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4315 	    kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4316 
4317 	kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4318 }
4319 
4320 void
4321 kmem_init(void)
4322 {
4323 	kmem_cache_t *cp;
4324 	int old_kmem_flags = kmem_flags;
4325 	int use_large_pages = 0;
4326 	size_t maxverify, minfirewall;
4327 
4328 	kstat_init();
4329 
4330 	/*
4331 	 * Small-memory systems (< 24 MB) can't handle kmem_flags overhead.
4332 	 */
4333 	if (physmem < btop(24 << 20) && !(old_kmem_flags & KMF_STICKY))
4334 		kmem_flags = 0;
4335 
4336 	/*
4337 	 * Don't do firewalled allocations if the heap is less than 1TB
4338 	 * (i.e. on a 32-bit kernel)
4339 	 * The resulting VM_NEXTFIT allocations would create too much
4340 	 * fragmentation in a small heap.
4341 	 */
4342 #if defined(_LP64)
4343 	maxverify = minfirewall = PAGESIZE / 2;
4344 #else
4345 	maxverify = minfirewall = ULONG_MAX;
4346 #endif
4347 
4348 	/* LINTED */
4349 	ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4350 
4351 	list_create(&kmem_caches, sizeof (kmem_cache_t),
4352 	    offsetof(kmem_cache_t, cache_link));
4353 
4354 	kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4355 	    vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4356 	    VM_SLEEP | VMC_NO_QCACHE);
4357 
4358 	kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4359 	    PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4360 	    VMC_DUMPSAFE | VM_SLEEP);
4361 
4362 	kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4363 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4364 
4365 	kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4366 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4367 
4368 	kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4369 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4370 
4371 	kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4372 	    NULL, 0, PAGESIZE,
4373 	    kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4374 	    0, VM_SLEEP);
4375 
4376 	kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4377 	    segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4378 	    VMC_DUMPSAFE | VM_SLEEP);
4379 
4380 	/* temporary oversize arena for mod_read_system_file */
4381 	kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4382 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4383 
4384 	kmem_reap_interval = 15 * hz;
4385 
4386 	/*
4387 	 * Read /etc/system.  This is a chicken-and-egg problem because
4388 	 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4389 	 * needs to use the allocator.  The simplest solution is to create
4390 	 * all the standard kmem caches, read /etc/system, destroy all the
4391 	 * caches we just created, and then create them all again in light
4392 	 * of the (possibly) new kmem_flags and other kmem tunables.
4393 	 */
4394 	kmem_cache_init(1, 0);
4395 
4396 	mod_read_system_file(boothowto & RB_ASKNAME);
4397 
4398 	while ((cp = list_tail(&kmem_caches)) != NULL)
4399 		kmem_cache_destroy(cp);
4400 
4401 	vmem_destroy(kmem_oversize_arena);
4402 
4403 	if (old_kmem_flags & KMF_STICKY)
4404 		kmem_flags = old_kmem_flags;
4405 
4406 	if (!(kmem_flags & KMF_AUDIT))
4407 		vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4408 
4409 	if (kmem_maxverify == 0)
4410 		kmem_maxverify = maxverify;
4411 
4412 	if (kmem_minfirewall == 0)
4413 		kmem_minfirewall = minfirewall;
4414 
4415 	/*
4416 	 * give segkmem a chance to figure out if we are using large pages
4417 	 * for the kernel heap
4418 	 */
4419 	use_large_pages = segkmem_lpsetup();
4420 
4421 	/*
4422 	 * To protect against corruption, we keep the actual number of callers
4423 	 * KMF_LITE records seperate from the tunable.  We arbitrarily clamp
4424 	 * to 16, since the overhead for small buffers quickly gets out of
4425 	 * hand.
4426 	 *
4427 	 * The real limit would depend on the needs of the largest KMC_NOHASH
4428 	 * cache.
4429 	 */
4430 	kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4431 	kmem_lite_pcs = kmem_lite_count;
4432 
4433 	/*
4434 	 * Normally, we firewall oversized allocations when possible, but
4435 	 * if we are using large pages for kernel memory, and we don't have
4436 	 * any non-LITE debugging flags set, we want to allocate oversized
4437 	 * buffers from large pages, and so skip the firewalling.
4438 	 */
4439 	if (use_large_pages &&
4440 	    ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4441 		kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4442 		    PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4443 		    0, VMC_DUMPSAFE | VM_SLEEP);
4444 	} else {
4445 		kmem_oversize_arena = vmem_create("kmem_oversize",
4446 		    NULL, 0, PAGESIZE,
4447 		    segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4448 		    kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4449 		    VM_SLEEP);
4450 	}
4451 
4452 	kmem_cache_init(2, use_large_pages);
4453 
4454 	if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4455 		if (kmem_transaction_log_size == 0)
4456 			kmem_transaction_log_size = kmem_maxavail() / 50;
4457 		kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4458 	}
4459 
4460 	if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4461 		if (kmem_content_log_size == 0)
4462 			kmem_content_log_size = kmem_maxavail() / 50;
4463 		kmem_content_log = kmem_log_init(kmem_content_log_size);
4464 	}
4465 
4466 	kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4467 
4468 	kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4469 
4470 	/*
4471 	 * Initialize STREAMS message caches so allocb() is available.
4472 	 * This allows us to initialize the logging framework (cmn_err(9F),
4473 	 * strlog(9F), etc) so we can start recording messages.
4474 	 */
4475 	streams_msg_init();
4476 
4477 	/*
4478 	 * Initialize the ZSD framework in Zones so modules loaded henceforth
4479 	 * can register their callbacks.
4480 	 */
4481 	zone_zsd_init();
4482 
4483 	log_init();
4484 	taskq_init();
4485 
4486 	/*
4487 	 * Warn about invalid or dangerous values of kmem_flags.
4488 	 * Always warn about unsupported values.
4489 	 */
4490 	if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4491 	    KMF_CONTENTS | KMF_LITE)) != 0) ||
4492 	    ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4493 		cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4494 		    "See the Solaris Tunable Parameters Reference Manual.",
4495 		    kmem_flags);
4496 
4497 #ifdef DEBUG
4498 	if ((kmem_flags & KMF_DEBUG) == 0)
4499 		cmn_err(CE_NOTE, "kmem debugging disabled.");
4500 #else
4501 	/*
4502 	 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4503 	 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4504 	 * if KMF_AUDIT is set). We should warn the user about the performance
4505 	 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4506 	 * isn't set (since that disables AUDIT).
4507 	 */
4508 	if (!(kmem_flags & KMF_LITE) &&
4509 	    (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4510 		cmn_err(CE_WARN, "High-overhead kmem debugging features "
4511 		    "enabled (kmem_flags = 0x%x).  Performance degradation "
4512 		    "and large memory overhead possible. See the Solaris "
4513 		    "Tunable Parameters Reference Manual.", kmem_flags);
4514 #endif /* not DEBUG */
4515 
4516 	kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4517 
4518 	kmem_ready = 1;
4519 
4520 	/*
4521 	 * Initialize the platform-specific aligned/DMA memory allocator.
4522 	 */
4523 	ka_init();
4524 
4525 	/*
4526 	 * Initialize 32-bit ID cache.
4527 	 */
4528 	id32_init();
4529 
4530 	/*
4531 	 * Initialize the networking stack so modules loaded can
4532 	 * register their callbacks.
4533 	 */
4534 	netstack_init();
4535 }
4536 
4537 static void
4538 kmem_move_init(void)
4539 {
4540 	kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4541 	    sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4542 	    kmem_msb_arena, KMC_NOHASH);
4543 	kmem_move_cache = kmem_cache_create("kmem_move_cache",
4544 	    sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4545 	    kmem_msb_arena, KMC_NOHASH);
4546 
4547 	/*
4548 	 * kmem guarantees that move callbacks are sequential and that even
4549 	 * across multiple caches no two moves ever execute simultaneously.
4550 	 * Move callbacks are processed on a separate taskq so that client code
4551 	 * does not interfere with internal maintenance tasks.
4552 	 */
4553 	kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4554 	    minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4555 }
4556 
4557 void
4558 kmem_thread_init(void)
4559 {
4560 	kmem_move_init();
4561 	kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4562 	    300, INT_MAX, TASKQ_PREPOPULATE);
4563 }
4564 
4565 void
4566 kmem_mp_init(void)
4567 {
4568 	mutex_enter(&cpu_lock);
4569 	register_cpu_setup_func(kmem_cpu_setup, NULL);
4570 	mutex_exit(&cpu_lock);
4571 
4572 	kmem_update_timeout(NULL);
4573 
4574 	taskq_mp_init();
4575 }
4576 
4577 /*
4578  * Return the slab of the allocated buffer, or NULL if the buffer is not
4579  * allocated. This function may be called with a known slab address to determine
4580  * whether or not the buffer is allocated, or with a NULL slab address to obtain
4581  * an allocated buffer's slab.
4582  */
4583 static kmem_slab_t *
4584 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4585 {
4586 	kmem_bufctl_t *bcp, *bufbcp;
4587 
4588 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4589 	ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4590 
4591 	if (cp->cache_flags & KMF_HASH) {
4592 		for (bcp = *KMEM_HASH(cp, buf);
4593 		    (bcp != NULL) && (bcp->bc_addr != buf);
4594 		    bcp = bcp->bc_next) {
4595 			continue;
4596 		}
4597 		ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4598 		return (bcp == NULL ? NULL : bcp->bc_slab);
4599 	}
4600 
4601 	if (sp == NULL) {
4602 		sp = KMEM_SLAB(cp, buf);
4603 	}
4604 	bufbcp = KMEM_BUFCTL(cp, buf);
4605 	for (bcp = sp->slab_head;
4606 	    (bcp != NULL) && (bcp != bufbcp);
4607 	    bcp = bcp->bc_next) {
4608 		continue;
4609 	}
4610 	return (bcp == NULL ? sp : NULL);
4611 }
4612 
4613 static boolean_t
4614 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4615 {
4616 	long refcnt = sp->slab_refcnt;
4617 
4618 	ASSERT(cp->cache_defrag != NULL);
4619 
4620 	/*
4621 	 * For code coverage we want to be able to move an object within the
4622 	 * same slab (the only partial slab) even if allocating the destination
4623 	 * buffer resulted in a completely allocated slab.
4624 	 */
4625 	if (flags & KMM_DEBUG) {
4626 		return ((flags & KMM_DESPERATE) ||
4627 		    ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4628 	}
4629 
4630 	/* If we're desperate, we don't care if the client said NO. */
4631 	if (flags & KMM_DESPERATE) {
4632 		return (refcnt < sp->slab_chunks); /* any partial */
4633 	}
4634 
4635 	if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4636 		return (B_FALSE);
4637 	}
4638 
4639 	if ((refcnt == 1) || kmem_move_any_partial) {
4640 		return (refcnt < sp->slab_chunks);
4641 	}
4642 
4643 	/*
4644 	 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4645 	 * slabs with a progressively higher percentage of used buffers can be
4646 	 * reclaimed until the cache as a whole is no longer fragmented.
4647 	 *
4648 	 *	sp->slab_refcnt   kmd_reclaim_numer
4649 	 *	--------------- < ------------------
4650 	 *	sp->slab_chunks   KMEM_VOID_FRACTION
4651 	 */
4652 	return ((refcnt * KMEM_VOID_FRACTION) <
4653 	    (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4654 }
4655 
4656 static void *
4657 kmem_hunt_mag(kmem_cache_t *cp, kmem_magazine_t *m, int n, void *buf,
4658     void *tbuf)
4659 {
4660 	int i;		/* magazine round index */
4661 
4662 	for (i = 0; i < n; i++) {
4663 		if (buf == m->mag_round[i]) {
4664 			if (cp->cache_flags & KMF_BUFTAG) {
4665 				(void) kmem_cache_free_debug(cp, tbuf,
4666 				    caller());
4667 			}
4668 			m->mag_round[i] = tbuf;
4669 			return (buf);
4670 		}
4671 	}
4672 
4673 	return (NULL);
4674 }
4675 
4676 /*
4677  * Hunt the magazine layer for the given buffer. If found, the buffer is
4678  * removed from the magazine layer and returned, otherwise NULL is returned.
4679  * The state of the returned buffer is freed and constructed.
4680  */
4681 static void *
4682 kmem_hunt_mags(kmem_cache_t *cp, void *buf)
4683 {
4684 	kmem_cpu_cache_t *ccp;
4685 	kmem_magazine_t	*m;
4686 	int cpu_seqid;
4687 	int n;		/* magazine rounds */
4688 	void *tbuf;	/* temporary swap buffer */
4689 
4690 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4691 
4692 	/*
4693 	 * Allocated a buffer to swap with the one we hope to pull out of a
4694 	 * magazine when found.
4695 	 */
4696 	tbuf = kmem_cache_alloc(cp, KM_NOSLEEP);
4697 	if (tbuf == NULL) {
4698 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_alloc_fail);
4699 		return (NULL);
4700 	}
4701 	if (tbuf == buf) {
4702 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_lucky);
4703 		if (cp->cache_flags & KMF_BUFTAG) {
4704 			(void) kmem_cache_free_debug(cp, buf, caller());
4705 		}
4706 		return (buf);
4707 	}
4708 
4709 	/* Hunt the depot. */
4710 	mutex_enter(&cp->cache_depot_lock);
4711 	n = cp->cache_magtype->mt_magsize;
4712 	for (m = cp->cache_full.ml_list; m != NULL; m = m->mag_next) {
4713 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4714 			mutex_exit(&cp->cache_depot_lock);
4715 			return (buf);
4716 		}
4717 	}
4718 	mutex_exit(&cp->cache_depot_lock);
4719 
4720 	/* Hunt the per-CPU magazines. */
4721 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4722 		ccp = &cp->cache_cpu[cpu_seqid];
4723 
4724 		mutex_enter(&ccp->cc_lock);
4725 		m = ccp->cc_loaded;
4726 		n = ccp->cc_rounds;
4727 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4728 			mutex_exit(&ccp->cc_lock);
4729 			return (buf);
4730 		}
4731 		m = ccp->cc_ploaded;
4732 		n = ccp->cc_prounds;
4733 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4734 			mutex_exit(&ccp->cc_lock);
4735 			return (buf);
4736 		}
4737 		mutex_exit(&ccp->cc_lock);
4738 	}
4739 
4740 	kmem_cache_free(cp, tbuf);
4741 	return (NULL);
4742 }
4743 
4744 /*
4745  * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4746  * or when the buffer is freed.
4747  */
4748 static void
4749 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4750 {
4751 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4752 	ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4753 
4754 	if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4755 		return;
4756 	}
4757 
4758 	if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4759 		if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4760 			avl_remove(&cp->cache_partial_slabs, sp);
4761 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4762 			sp->slab_stuck_offset = (uint32_t)-1;
4763 			avl_add(&cp->cache_partial_slabs, sp);
4764 		}
4765 	} else {
4766 		sp->slab_later_count = 0;
4767 		sp->slab_stuck_offset = (uint32_t)-1;
4768 	}
4769 }
4770 
4771 static void
4772 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4773 {
4774 	ASSERT(taskq_member(kmem_move_taskq, curthread));
4775 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4776 	ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4777 
4778 	if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4779 		return;
4780 	}
4781 
4782 	avl_remove(&cp->cache_partial_slabs, sp);
4783 	sp->slab_later_count = 0;
4784 	sp->slab_flags |= KMEM_SLAB_NOMOVE;
4785 	sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4786 	avl_add(&cp->cache_partial_slabs, sp);
4787 }
4788 
4789 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4790 
4791 /*
4792  * The move callback takes two buffer addresses, the buffer to be moved, and a
4793  * newly allocated and constructed buffer selected by kmem as the destination.
4794  * It also takes the size of the buffer and an optional user argument specified
4795  * at cache creation time. kmem guarantees that the buffer to be moved has not
4796  * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4797  * guarantee the present whereabouts of the buffer to be moved, so it is up to
4798  * the client to safely determine whether or not it is still using the buffer.
4799  * The client must not free either of the buffers passed to the move callback,
4800  * since kmem wants to free them directly to the slab layer. The client response
4801  * tells kmem which of the two buffers to free:
4802  *
4803  * YES		kmem frees the old buffer (the move was successful)
4804  * NO		kmem frees the new buffer, marks the slab of the old buffer
4805  *              non-reclaimable to avoid bothering the client again
4806  * LATER	kmem frees the new buffer, increments slab_later_count
4807  * DONT_KNOW	kmem frees the new buffer, searches mags for the old buffer
4808  * DONT_NEED	kmem frees both the old buffer and the new buffer
4809  *
4810  * The pending callback argument now being processed contains both of the
4811  * buffers (old and new) passed to the move callback function, the slab of the
4812  * old buffer, and flags related to the move request, such as whether or not the
4813  * system was desperate for memory.
4814  *
4815  * Slabs are not freed while there is a pending callback, but instead are kept
4816  * on a deadlist, which is drained after the last callback completes. This means
4817  * that slabs are safe to access until kmem_move_end(), no matter how many of
4818  * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4819  * zero for as long as the slab remains on the deadlist and until the slab is
4820  * freed.
4821  */
4822 static void
4823 kmem_move_buffer(kmem_move_t *callback)
4824 {
4825 	kmem_cbrc_t response;
4826 	kmem_slab_t *sp = callback->kmm_from_slab;
4827 	kmem_cache_t *cp = sp->slab_cache;
4828 	boolean_t free_on_slab;
4829 
4830 	ASSERT(taskq_member(kmem_move_taskq, curthread));
4831 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4832 	ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4833 
4834 	/*
4835 	 * The number of allocated buffers on the slab may have changed since we
4836 	 * last checked the slab's reclaimability (when the pending move was
4837 	 * enqueued), or the client may have responded NO when asked to move
4838 	 * another buffer on the same slab.
4839 	 */
4840 	if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4841 		KMEM_STAT_ADD(kmem_move_stats.kms_no_longer_reclaimable);
4842 		KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4843 		    kmem_move_stats.kms_notify_no_longer_reclaimable);
4844 		kmem_slab_free(cp, callback->kmm_to_buf);
4845 		kmem_move_end(cp, callback);
4846 		return;
4847 	}
4848 
4849 	/*
4850 	 * Hunting magazines is expensive, so we'll wait to do that until the
4851 	 * client responds KMEM_CBRC_DONT_KNOW. However, checking the slab layer
4852 	 * is cheap, so we might as well do that here in case we can avoid
4853 	 * bothering the client.
4854 	 */
4855 	mutex_enter(&cp->cache_lock);
4856 	free_on_slab = (kmem_slab_allocated(cp, sp,
4857 	    callback->kmm_from_buf) == NULL);
4858 	mutex_exit(&cp->cache_lock);
4859 
4860 	if (free_on_slab) {
4861 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_slab);
4862 		kmem_slab_free(cp, callback->kmm_to_buf);
4863 		kmem_move_end(cp, callback);
4864 		return;
4865 	}
4866 
4867 	if (cp->cache_flags & KMF_BUFTAG) {
4868 		/*
4869 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
4870 		 */
4871 		if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4872 		    KM_NOSLEEP, 1, caller()) != 0) {
4873 			KMEM_STAT_ADD(kmem_move_stats.kms_alloc_fail);
4874 			kmem_move_end(cp, callback);
4875 			return;
4876 		}
4877 	} else if (cp->cache_constructor != NULL &&
4878 	    cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4879 	    KM_NOSLEEP) != 0) {
4880 		atomic_add_64(&cp->cache_alloc_fail, 1);
4881 		KMEM_STAT_ADD(kmem_move_stats.kms_constructor_fail);
4882 		kmem_slab_free(cp, callback->kmm_to_buf);
4883 		kmem_move_end(cp, callback);
4884 		return;
4885 	}
4886 
4887 	KMEM_STAT_ADD(kmem_move_stats.kms_callbacks);
4888 	KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4889 	    kmem_move_stats.kms_notify_callbacks);
4890 	cp->cache_defrag->kmd_callbacks++;
4891 	cp->cache_defrag->kmd_thread = curthread;
4892 	cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4893 	cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4894 	DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4895 	    callback);
4896 
4897 	response = cp->cache_move(callback->kmm_from_buf,
4898 	    callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4899 
4900 	DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4901 	    callback, kmem_cbrc_t, response);
4902 	cp->cache_defrag->kmd_thread = NULL;
4903 	cp->cache_defrag->kmd_from_buf = NULL;
4904 	cp->cache_defrag->kmd_to_buf = NULL;
4905 
4906 	if (response == KMEM_CBRC_YES) {
4907 		KMEM_STAT_ADD(kmem_move_stats.kms_yes);
4908 		cp->cache_defrag->kmd_yes++;
4909 		kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4910 		/* slab safe to access until kmem_move_end() */
4911 		if (sp->slab_refcnt == 0)
4912 			cp->cache_defrag->kmd_slabs_freed++;
4913 		mutex_enter(&cp->cache_lock);
4914 		kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4915 		mutex_exit(&cp->cache_lock);
4916 		kmem_move_end(cp, callback);
4917 		return;
4918 	}
4919 
4920 	switch (response) {
4921 	case KMEM_CBRC_NO:
4922 		KMEM_STAT_ADD(kmem_move_stats.kms_no);
4923 		cp->cache_defrag->kmd_no++;
4924 		mutex_enter(&cp->cache_lock);
4925 		kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4926 		mutex_exit(&cp->cache_lock);
4927 		break;
4928 	case KMEM_CBRC_LATER:
4929 		KMEM_STAT_ADD(kmem_move_stats.kms_later);
4930 		cp->cache_defrag->kmd_later++;
4931 		mutex_enter(&cp->cache_lock);
4932 		if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4933 			mutex_exit(&cp->cache_lock);
4934 			break;
4935 		}
4936 
4937 		if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4938 			KMEM_STAT_ADD(kmem_move_stats.kms_disbelief);
4939 			kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4940 		} else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4941 			sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4942 			    callback->kmm_from_buf);
4943 		}
4944 		mutex_exit(&cp->cache_lock);
4945 		break;
4946 	case KMEM_CBRC_DONT_NEED:
4947 		KMEM_STAT_ADD(kmem_move_stats.kms_dont_need);
4948 		cp->cache_defrag->kmd_dont_need++;
4949 		kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4950 		if (sp->slab_refcnt == 0)
4951 			cp->cache_defrag->kmd_slabs_freed++;
4952 		mutex_enter(&cp->cache_lock);
4953 		kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4954 		mutex_exit(&cp->cache_lock);
4955 		break;
4956 	case KMEM_CBRC_DONT_KNOW:
4957 		KMEM_STAT_ADD(kmem_move_stats.kms_dont_know);
4958 		cp->cache_defrag->kmd_dont_know++;
4959 		if (kmem_hunt_mags(cp, callback->kmm_from_buf) != NULL) {
4960 			KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_mag);
4961 			cp->cache_defrag->kmd_hunt_found++;
4962 			kmem_slab_free_constructed(cp, callback->kmm_from_buf,
4963 			    B_TRUE);
4964 			if (sp->slab_refcnt == 0)
4965 				cp->cache_defrag->kmd_slabs_freed++;
4966 			mutex_enter(&cp->cache_lock);
4967 			kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4968 			mutex_exit(&cp->cache_lock);
4969 		}
4970 		break;
4971 	default:
4972 		panic("'%s' (%p) unexpected move callback response %d\n",
4973 		    cp->cache_name, (void *)cp, response);
4974 	}
4975 
4976 	kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4977 	kmem_move_end(cp, callback);
4978 }
4979 
4980 /* Return B_FALSE if there is insufficient memory for the move request. */
4981 static boolean_t
4982 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4983 {
4984 	void *to_buf;
4985 	avl_index_t index;
4986 	kmem_move_t *callback, *pending;
4987 	ulong_t n;
4988 
4989 	ASSERT(taskq_member(kmem_taskq, curthread));
4990 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4991 	ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4992 
4993 	callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4994 	if (callback == NULL) {
4995 		KMEM_STAT_ADD(kmem_move_stats.kms_callback_alloc_fail);
4996 		return (B_FALSE);
4997 	}
4998 
4999 	callback->kmm_from_slab = sp;
5000 	callback->kmm_from_buf = buf;
5001 	callback->kmm_flags = flags;
5002 
5003 	mutex_enter(&cp->cache_lock);
5004 
5005 	n = avl_numnodes(&cp->cache_partial_slabs);
5006 	if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
5007 		mutex_exit(&cp->cache_lock);
5008 		kmem_cache_free(kmem_move_cache, callback);
5009 		return (B_TRUE); /* there is no need for the move request */
5010 	}
5011 
5012 	pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
5013 	if (pending != NULL) {
5014 		/*
5015 		 * If the move is already pending and we're desperate now,
5016 		 * update the move flags.
5017 		 */
5018 		if (flags & KMM_DESPERATE) {
5019 			pending->kmm_flags |= KMM_DESPERATE;
5020 		}
5021 		mutex_exit(&cp->cache_lock);
5022 		KMEM_STAT_ADD(kmem_move_stats.kms_already_pending);
5023 		kmem_cache_free(kmem_move_cache, callback);
5024 		return (B_TRUE);
5025 	}
5026 
5027 	to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
5028 	    B_FALSE);
5029 	callback->kmm_to_buf = to_buf;
5030 	avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
5031 
5032 	mutex_exit(&cp->cache_lock);
5033 
5034 	if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
5035 	    callback, TQ_NOSLEEP)) {
5036 		KMEM_STAT_ADD(kmem_move_stats.kms_callback_taskq_fail);
5037 		mutex_enter(&cp->cache_lock);
5038 		avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5039 		mutex_exit(&cp->cache_lock);
5040 		kmem_slab_free(cp, to_buf);
5041 		kmem_cache_free(kmem_move_cache, callback);
5042 		return (B_FALSE);
5043 	}
5044 
5045 	return (B_TRUE);
5046 }
5047 
5048 static void
5049 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
5050 {
5051 	avl_index_t index;
5052 
5053 	ASSERT(cp->cache_defrag != NULL);
5054 	ASSERT(taskq_member(kmem_move_taskq, curthread));
5055 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
5056 
5057 	mutex_enter(&cp->cache_lock);
5058 	VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
5059 	    callback->kmm_from_buf, &index) != NULL);
5060 	avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5061 	if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
5062 		list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5063 		kmem_slab_t *sp;
5064 
5065 		/*
5066 		 * The last pending move completed. Release all slabs from the
5067 		 * front of the dead list except for any slab at the tail that
5068 		 * needs to be released from the context of kmem_move_buffers().
5069 		 * kmem deferred unmapping the buffers on these slabs in order
5070 		 * to guarantee that buffers passed to the move callback have
5071 		 * been touched only by kmem or by the client itself.
5072 		 */
5073 		while ((sp = list_remove_head(deadlist)) != NULL) {
5074 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
5075 				list_insert_tail(deadlist, sp);
5076 				break;
5077 			}
5078 			cp->cache_defrag->kmd_deadcount--;
5079 			cp->cache_slab_destroy++;
5080 			mutex_exit(&cp->cache_lock);
5081 			kmem_slab_destroy(cp, sp);
5082 			KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5083 			mutex_enter(&cp->cache_lock);
5084 		}
5085 	}
5086 	mutex_exit(&cp->cache_lock);
5087 	kmem_cache_free(kmem_move_cache, callback);
5088 }
5089 
5090 /*
5091  * Move buffers from least used slabs first by scanning backwards from the end
5092  * of the partial slab list. Scan at most max_scan candidate slabs and move
5093  * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
5094  * If desperate to reclaim memory, move buffers from any partial slab, otherwise
5095  * skip slabs with a ratio of allocated buffers at or above the current
5096  * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5097  * scan is aborted) so that the caller can adjust the reclaimability threshold
5098  * depending on how many reclaimable slabs it finds.
5099  *
5100  * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5101  * move request, since it is not valid for kmem_move_begin() to call
5102  * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5103  */
5104 static int
5105 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5106     int flags)
5107 {
5108 	kmem_slab_t *sp;
5109 	void *buf;
5110 	int i, j; /* slab index, buffer index */
5111 	int s; /* reclaimable slabs */
5112 	int b; /* allocated (movable) buffers on reclaimable slab */
5113 	boolean_t success;
5114 	int refcnt;
5115 	int nomove;
5116 
5117 	ASSERT(taskq_member(kmem_taskq, curthread));
5118 	ASSERT(MUTEX_HELD(&cp->cache_lock));
5119 	ASSERT(kmem_move_cache != NULL);
5120 	ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5121 	ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5122 	    avl_numnodes(&cp->cache_partial_slabs) > 1);
5123 
5124 	if (kmem_move_blocked) {
5125 		return (0);
5126 	}
5127 
5128 	if (kmem_move_fulltilt) {
5129 		flags |= KMM_DESPERATE;
5130 	}
5131 
5132 	if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5133 		/*
5134 		 * Scan as many slabs as needed to find the desired number of
5135 		 * candidate slabs.
5136 		 */
5137 		max_scan = (size_t)-1;
5138 	}
5139 
5140 	if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5141 		/* Find as many candidate slabs as possible. */
5142 		max_slabs = (size_t)-1;
5143 	}
5144 
5145 	sp = avl_last(&cp->cache_partial_slabs);
5146 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5147 	for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5148 	    ((sp != avl_first(&cp->cache_partial_slabs)) ||
5149 	    (flags & KMM_DEBUG));
5150 	    sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5151 
5152 		if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5153 			continue;
5154 		}
5155 		s++;
5156 
5157 		/* Look for allocated buffers to move. */
5158 		for (j = 0, b = 0, buf = sp->slab_base;
5159 		    (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5160 		    buf = (((char *)buf) + cp->cache_chunksize), j++) {
5161 
5162 			if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5163 				continue;
5164 			}
5165 
5166 			b++;
5167 
5168 			/*
5169 			 * Prevent the slab from being destroyed while we drop
5170 			 * cache_lock and while the pending move is not yet
5171 			 * registered. Flag the pending move while
5172 			 * kmd_moves_pending may still be empty, since we can't
5173 			 * yet rely on a non-zero pending move count to prevent
5174 			 * the slab from being destroyed.
5175 			 */
5176 			ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5177 			sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5178 			/*
5179 			 * Recheck refcnt and nomove after reacquiring the lock,
5180 			 * since these control the order of partial slabs, and
5181 			 * we want to know if we can pick up the scan where we
5182 			 * left off.
5183 			 */
5184 			refcnt = sp->slab_refcnt;
5185 			nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5186 			mutex_exit(&cp->cache_lock);
5187 
5188 			success = kmem_move_begin(cp, sp, buf, flags);
5189 
5190 			/*
5191 			 * Now, before the lock is reacquired, kmem could
5192 			 * process all pending move requests and purge the
5193 			 * deadlist, so that upon reacquiring the lock, sp has
5194 			 * been remapped. Or, the client may free all the
5195 			 * objects on the slab while the pending moves are still
5196 			 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5197 			 * flag causes the slab to be put at the end of the
5198 			 * deadlist and prevents it from being destroyed, since
5199 			 * we plan to destroy it here after reacquiring the
5200 			 * lock.
5201 			 */
5202 			mutex_enter(&cp->cache_lock);
5203 			ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5204 			sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5205 
5206 			if (sp->slab_refcnt == 0) {
5207 				list_t *deadlist =
5208 				    &cp->cache_defrag->kmd_deadlist;
5209 				list_remove(deadlist, sp);
5210 
5211 				if (!avl_is_empty(
5212 				    &cp->cache_defrag->kmd_moves_pending)) {
5213 					/*
5214 					 * A pending move makes it unsafe to
5215 					 * destroy the slab, because even though
5216 					 * the move is no longer needed, the
5217 					 * context where that is determined
5218 					 * requires the slab to exist.
5219 					 * Fortunately, a pending move also
5220 					 * means we don't need to destroy the
5221 					 * slab here, since it will get
5222 					 * destroyed along with any other slabs
5223 					 * on the deadlist after the last
5224 					 * pending move completes.
5225 					 */
5226 					list_insert_head(deadlist, sp);
5227 					KMEM_STAT_ADD(kmem_move_stats.
5228 					    kms_endscan_slab_dead);
5229 					return (-1);
5230 				}
5231 
5232 				/*
5233 				 * Destroy the slab now if it was completely
5234 				 * freed while we dropped cache_lock and there
5235 				 * are no pending moves. Since slab_refcnt
5236 				 * cannot change once it reaches zero, no new
5237 				 * pending moves from that slab are possible.
5238 				 */
5239 				cp->cache_defrag->kmd_deadcount--;
5240 				cp->cache_slab_destroy++;
5241 				mutex_exit(&cp->cache_lock);
5242 				kmem_slab_destroy(cp, sp);
5243 				KMEM_STAT_ADD(kmem_move_stats.
5244 				    kms_dead_slabs_freed);
5245 				KMEM_STAT_ADD(kmem_move_stats.
5246 				    kms_endscan_slab_destroyed);
5247 				mutex_enter(&cp->cache_lock);
5248 				/*
5249 				 * Since we can't pick up the scan where we left
5250 				 * off, abort the scan and say nothing about the
5251 				 * number of reclaimable slabs.
5252 				 */
5253 				return (-1);
5254 			}
5255 
5256 			if (!success) {
5257 				/*
5258 				 * Abort the scan if there is not enough memory
5259 				 * for the request and say nothing about the
5260 				 * number of reclaimable slabs.
5261 				 */
5262 				KMEM_STAT_COND_ADD(s < max_slabs,
5263 				    kmem_move_stats.kms_endscan_nomem);
5264 				return (-1);
5265 			}
5266 
5267 			/*
5268 			 * The slab's position changed while the lock was
5269 			 * dropped, so we don't know where we are in the
5270 			 * sequence any more.
5271 			 */
5272 			if (sp->slab_refcnt != refcnt) {
5273 				/*
5274 				 * If this is a KMM_DEBUG move, the slab_refcnt
5275 				 * may have changed because we allocated a
5276 				 * destination buffer on the same slab. In that
5277 				 * case, we're not interested in counting it.
5278 				 */
5279 				KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5280 				    (s < max_slabs),
5281 				    kmem_move_stats.kms_endscan_refcnt_changed);
5282 				return (-1);
5283 			}
5284 			if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove) {
5285 				KMEM_STAT_COND_ADD(s < max_slabs,
5286 				    kmem_move_stats.kms_endscan_nomove_changed);
5287 				return (-1);
5288 			}
5289 
5290 			/*
5291 			 * Generating a move request allocates a destination
5292 			 * buffer from the slab layer, bumping the first partial
5293 			 * slab if it is completely allocated. If the current
5294 			 * slab becomes the first partial slab as a result, we
5295 			 * can't continue to scan backwards.
5296 			 *
5297 			 * If this is a KMM_DEBUG move and we allocated the
5298 			 * destination buffer from the last partial slab, then
5299 			 * the buffer we're moving is on the same slab and our
5300 			 * slab_refcnt has changed, causing us to return before
5301 			 * reaching here if there are no partial slabs left.
5302 			 */
5303 			ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5304 			if (sp == avl_first(&cp->cache_partial_slabs)) {
5305 				/*
5306 				 * We're not interested in a second KMM_DEBUG
5307 				 * move.
5308 				 */
5309 				goto end_scan;
5310 			}
5311 		}
5312 	}
5313 end_scan:
5314 
5315 	KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5316 	    (s < max_slabs) &&
5317 	    (sp == avl_first(&cp->cache_partial_slabs)),
5318 	    kmem_move_stats.kms_endscan_freelist);
5319 
5320 	return (s);
5321 }
5322 
5323 typedef struct kmem_move_notify_args {
5324 	kmem_cache_t *kmna_cache;
5325 	void *kmna_buf;
5326 } kmem_move_notify_args_t;
5327 
5328 static void
5329 kmem_cache_move_notify_task(void *arg)
5330 {
5331 	kmem_move_notify_args_t *args = arg;
5332 	kmem_cache_t *cp = args->kmna_cache;
5333 	void *buf = args->kmna_buf;
5334 	kmem_slab_t *sp;
5335 
5336 	ASSERT(taskq_member(kmem_taskq, curthread));
5337 	ASSERT(list_link_active(&cp->cache_link));
5338 
5339 	kmem_free(args, sizeof (kmem_move_notify_args_t));
5340 	mutex_enter(&cp->cache_lock);
5341 	sp = kmem_slab_allocated(cp, NULL, buf);
5342 
5343 	/* Ignore the notification if the buffer is no longer allocated. */
5344 	if (sp == NULL) {
5345 		mutex_exit(&cp->cache_lock);
5346 		return;
5347 	}
5348 
5349 	/* Ignore the notification if there's no reason to move the buffer. */
5350 	if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5351 		/*
5352 		 * So far the notification is not ignored. Ignore the
5353 		 * notification if the slab is not marked by an earlier refusal
5354 		 * to move a buffer.
5355 		 */
5356 		if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5357 		    (sp->slab_later_count == 0)) {
5358 			mutex_exit(&cp->cache_lock);
5359 			return;
5360 		}
5361 
5362 		kmem_slab_move_yes(cp, sp, buf);
5363 		ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5364 		sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5365 		mutex_exit(&cp->cache_lock);
5366 		/* see kmem_move_buffers() about dropping the lock */
5367 		(void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5368 		mutex_enter(&cp->cache_lock);
5369 		ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5370 		sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5371 		if (sp->slab_refcnt == 0) {
5372 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5373 			list_remove(deadlist, sp);
5374 
5375 			if (!avl_is_empty(
5376 			    &cp->cache_defrag->kmd_moves_pending)) {
5377 				list_insert_head(deadlist, sp);
5378 				mutex_exit(&cp->cache_lock);
5379 				KMEM_STAT_ADD(kmem_move_stats.
5380 				    kms_notify_slab_dead);
5381 				return;
5382 			}
5383 
5384 			cp->cache_defrag->kmd_deadcount--;
5385 			cp->cache_slab_destroy++;
5386 			mutex_exit(&cp->cache_lock);
5387 			kmem_slab_destroy(cp, sp);
5388 			KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5389 			KMEM_STAT_ADD(kmem_move_stats.
5390 			    kms_notify_slab_destroyed);
5391 			return;
5392 		}
5393 	} else {
5394 		kmem_slab_move_yes(cp, sp, buf);
5395 	}
5396 	mutex_exit(&cp->cache_lock);
5397 }
5398 
5399 void
5400 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5401 {
5402 	kmem_move_notify_args_t *args;
5403 
5404 	KMEM_STAT_ADD(kmem_move_stats.kms_notify);
5405 	args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5406 	if (args != NULL) {
5407 		args->kmna_cache = cp;
5408 		args->kmna_buf = buf;
5409 		if (!taskq_dispatch(kmem_taskq,
5410 		    (task_func_t *)kmem_cache_move_notify_task, args,
5411 		    TQ_NOSLEEP))
5412 			kmem_free(args, sizeof (kmem_move_notify_args_t));
5413 	}
5414 }
5415 
5416 static void
5417 kmem_cache_defrag(kmem_cache_t *cp)
5418 {
5419 	size_t n;
5420 
5421 	ASSERT(cp->cache_defrag != NULL);
5422 
5423 	mutex_enter(&cp->cache_lock);
5424 	n = avl_numnodes(&cp->cache_partial_slabs);
5425 	if (n > 1) {
5426 		/* kmem_move_buffers() drops and reacquires cache_lock */
5427 		KMEM_STAT_ADD(kmem_move_stats.kms_defrags);
5428 		cp->cache_defrag->kmd_defrags++;
5429 		(void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5430 	}
5431 	mutex_exit(&cp->cache_lock);
5432 }
5433 
5434 /* Is this cache above the fragmentation threshold? */
5435 static boolean_t
5436 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5437 {
5438 	/*
5439 	 *	nfree		kmem_frag_numer
5440 	 * ------------------ > ---------------
5441 	 * cp->cache_buftotal	kmem_frag_denom
5442 	 */
5443 	return ((nfree * kmem_frag_denom) >
5444 	    (cp->cache_buftotal * kmem_frag_numer));
5445 }
5446 
5447 static boolean_t
5448 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5449 {
5450 	boolean_t fragmented;
5451 	uint64_t nfree;
5452 
5453 	ASSERT(MUTEX_HELD(&cp->cache_lock));
5454 	*doreap = B_FALSE;
5455 
5456 	if (kmem_move_fulltilt) {
5457 		if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5458 			return (B_TRUE);
5459 		}
5460 	} else {
5461 		if ((cp->cache_complete_slab_count + avl_numnodes(
5462 		    &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5463 			return (B_FALSE);
5464 		}
5465 	}
5466 
5467 	nfree = cp->cache_bufslab;
5468 	fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5469 	    kmem_cache_frag_threshold(cp, nfree));
5470 
5471 	/*
5472 	 * Free buffers in the magazine layer appear allocated from the point of
5473 	 * view of the slab layer. We want to know if the slab layer would
5474 	 * appear fragmented if we included free buffers from magazines that
5475 	 * have fallen out of the working set.
5476 	 */
5477 	if (!fragmented) {
5478 		long reap;
5479 
5480 		mutex_enter(&cp->cache_depot_lock);
5481 		reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5482 		reap = MIN(reap, cp->cache_full.ml_total);
5483 		mutex_exit(&cp->cache_depot_lock);
5484 
5485 		nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5486 		if (kmem_cache_frag_threshold(cp, nfree)) {
5487 			*doreap = B_TRUE;
5488 		}
5489 	}
5490 
5491 	return (fragmented);
5492 }
5493 
5494 /* Called periodically from kmem_taskq */
5495 static void
5496 kmem_cache_scan(kmem_cache_t *cp)
5497 {
5498 	boolean_t reap = B_FALSE;
5499 	kmem_defrag_t *kmd;
5500 
5501 	ASSERT(taskq_member(kmem_taskq, curthread));
5502 
5503 	mutex_enter(&cp->cache_lock);
5504 
5505 	kmd = cp->cache_defrag;
5506 	if (kmd->kmd_consolidate > 0) {
5507 		kmd->kmd_consolidate--;
5508 		mutex_exit(&cp->cache_lock);
5509 		kmem_cache_reap(cp);
5510 		return;
5511 	}
5512 
5513 	if (kmem_cache_is_fragmented(cp, &reap)) {
5514 		size_t slabs_found;
5515 
5516 		/*
5517 		 * Consolidate reclaimable slabs from the end of the partial
5518 		 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5519 		 * reclaimable slabs). Keep track of how many candidate slabs we
5520 		 * looked for and how many we actually found so we can adjust
5521 		 * the definition of a candidate slab if we're having trouble
5522 		 * finding them.
5523 		 *
5524 		 * kmem_move_buffers() drops and reacquires cache_lock.
5525 		 */
5526 		KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5527 		kmd->kmd_scans++;
5528 		slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5529 		    kmem_reclaim_max_slabs, 0);
5530 		if (slabs_found >= 0) {
5531 			kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5532 			kmd->kmd_slabs_found += slabs_found;
5533 		}
5534 
5535 		if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5536 			kmd->kmd_tries = 0;
5537 
5538 			/*
5539 			 * If we had difficulty finding candidate slabs in
5540 			 * previous scans, adjust the threshold so that
5541 			 * candidates are easier to find.
5542 			 */
5543 			if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5544 				kmem_adjust_reclaim_threshold(kmd, -1);
5545 			} else if ((kmd->kmd_slabs_found * 2) <
5546 			    kmd->kmd_slabs_sought) {
5547 				kmem_adjust_reclaim_threshold(kmd, 1);
5548 			}
5549 			kmd->kmd_slabs_sought = 0;
5550 			kmd->kmd_slabs_found = 0;
5551 		}
5552 	} else {
5553 		kmem_reset_reclaim_threshold(cp->cache_defrag);
5554 #ifdef	DEBUG
5555 		if (!avl_is_empty(&cp->cache_partial_slabs)) {
5556 			/*
5557 			 * In a debug kernel we want the consolidator to
5558 			 * run occasionally even when there is plenty of
5559 			 * memory.
5560 			 */
5561 			uint16_t debug_rand;
5562 
5563 			(void) random_get_bytes((uint8_t *)&debug_rand, 2);
5564 			if (!kmem_move_noreap &&
5565 			    ((debug_rand % kmem_mtb_reap) == 0)) {
5566 				mutex_exit(&cp->cache_lock);
5567 				KMEM_STAT_ADD(kmem_move_stats.kms_debug_reaps);
5568 				kmem_cache_reap(cp);
5569 				return;
5570 			} else if ((debug_rand % kmem_mtb_move) == 0) {
5571 				KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5572 				KMEM_STAT_ADD(kmem_move_stats.kms_debug_scans);
5573 				kmd->kmd_scans++;
5574 				(void) kmem_move_buffers(cp,
5575 				    kmem_reclaim_scan_range, 1, KMM_DEBUG);
5576 			}
5577 		}
5578 #endif	/* DEBUG */
5579 	}
5580 
5581 	mutex_exit(&cp->cache_lock);
5582 
5583 	if (reap) {
5584 		KMEM_STAT_ADD(kmem_move_stats.kms_scan_depot_ws_reaps);
5585 		kmem_depot_ws_reap(cp);
5586 	}
5587 }
5588