xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision 8883f1c2)
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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*
27  * Kernel memory allocator, as described in the following two papers and a
28  * statement about the consolidator:
29  *
30  * Jeff Bonwick,
31  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
32  * Proceedings of the Summer 1994 Usenix Conference.
33  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
34  *
35  * Jeff Bonwick and Jonathan Adams,
36  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
37  * Arbitrary Resources.
38  * Proceedings of the 2001 Usenix Conference.
39  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
40  *
41  * kmem Slab Consolidator Big Theory Statement:
42  *
43  * 1. Motivation
44  *
45  * As stated in Bonwick94, slabs provide the following advantages over other
46  * allocation structures in terms of memory fragmentation:
47  *
48  *  - Internal fragmentation (per-buffer wasted space) is minimal.
49  *  - Severe external fragmentation (unused buffers on the free list) is
50  *    unlikely.
51  *
52  * Segregating objects by size eliminates one source of external fragmentation,
53  * and according to Bonwick:
54  *
55  *   The other reason that slabs reduce external fragmentation is that all
56  *   objects in a slab are of the same type, so they have the same lifetime
57  *   distribution. The resulting segregation of short-lived and long-lived
58  *   objects at slab granularity reduces the likelihood of an entire page being
59  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
60  *
61  * While unlikely, severe external fragmentation remains possible. Clients that
62  * allocate both short- and long-lived objects from the same cache cannot
63  * anticipate the distribution of long-lived objects within the allocator's slab
64  * implementation. Even a small percentage of long-lived objects distributed
65  * randomly across many slabs can lead to a worst case scenario where the client
66  * frees the majority of its objects and the system gets back almost none of the
67  * slabs. Despite the client doing what it reasonably can to help the system
68  * reclaim memory, the allocator cannot shake free enough slabs because of
69  * lonely allocations stubbornly hanging on. Although the allocator is in a
70  * position to diagnose the fragmentation, there is nothing that the allocator
71  * by itself can do about it. It only takes a single allocated object to prevent
72  * an entire slab from being reclaimed, and any object handed out by
73  * kmem_cache_alloc() is by definition in the client's control. Conversely,
74  * although the client is in a position to move a long-lived object, it has no
75  * way of knowing if the object is causing fragmentation, and if so, where to
76  * move it. A solution necessarily requires further cooperation between the
77  * allocator and the client.
78  *
79  * 2. Move Callback
80  *
81  * The kmem slab consolidator therefore adds a move callback to the
82  * allocator/client interface, improving worst-case external fragmentation in
83  * kmem caches that supply a function to move objects from one memory location
84  * to another. In a situation of low memory kmem attempts to consolidate all of
85  * a cache's slabs at once; otherwise it works slowly to bring external
86  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
87  * thereby helping to avoid a low memory situation in the future.
88  *
89  * The callback has the following signature:
90  *
91  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
92  *
93  * It supplies the kmem client with two addresses: the allocated object that
94  * kmem wants to move and a buffer selected by kmem for the client to use as the
95  * copy destination. The callback is kmem's way of saying "Please get off of
96  * this buffer and use this one instead." kmem knows where it wants to move the
97  * object in order to best reduce fragmentation. All the client needs to know
98  * about the second argument (void *new) is that it is an allocated, constructed
99  * object ready to take the contents of the old object. When the move function
100  * is called, the system is likely to be low on memory, and the new object
101  * spares the client from having to worry about allocating memory for the
102  * requested move. The third argument supplies the size of the object, in case a
103  * single move function handles multiple caches whose objects differ only in
104  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
105  * user argument passed to the constructor, destructor, and reclaim functions is
106  * also passed to the move callback.
107  *
108  * 2.1 Setting the Move Callback
109  *
110  * The client sets the move callback after creating the cache and before
111  * allocating from it:
112  *
113  *	object_cache = kmem_cache_create(...);
114  *      kmem_cache_set_move(object_cache, object_move);
115  *
116  * 2.2 Move Callback Return Values
117  *
118  * Only the client knows about its own data and when is a good time to move it.
119  * The client is cooperating with kmem to return unused memory to the system,
120  * and kmem respectfully accepts this help at the client's convenience. When
121  * asked to move an object, the client can respond with any of the following:
122  *
123  *   typedef enum kmem_cbrc {
124  *           KMEM_CBRC_YES,
125  *           KMEM_CBRC_NO,
126  *           KMEM_CBRC_LATER,
127  *           KMEM_CBRC_DONT_NEED,
128  *           KMEM_CBRC_DONT_KNOW
129  *   } kmem_cbrc_t;
130  *
131  * The client must not explicitly kmem_cache_free() either of the objects passed
132  * to the callback, since kmem wants to free them directly to the slab layer
133  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
134  * objects to free:
135  *
136  *       YES: (Did it) The client moved the object, so kmem frees the old one.
137  *        NO: (Never) The client refused, so kmem frees the new object (the
138  *            unused copy destination). kmem also marks the slab of the old
139  *            object so as not to bother the client with further callbacks for
140  *            that object as long as the slab remains on the partial slab list.
141  *            (The system won't be getting the slab back as long as the
142  *            immovable object holds it hostage, so there's no point in moving
143  *            any of its objects.)
144  *     LATER: The client is using the object and cannot move it now, so kmem
145  *            frees the new object (the unused copy destination). kmem still
146  *            attempts to move other objects off the slab, since it expects to
147  *            succeed in clearing the slab in a later callback. The client
148  *            should use LATER instead of NO if the object is likely to become
149  *            movable very soon.
150  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
151  *            with the new object (the unused copy destination). This response
152  *            is the client's opportunity to be a model citizen and give back as
153  *            much as it can.
154  * DONT_KNOW: The client does not know about the object because
155  *            a) the client has just allocated the object and not yet put it
156  *               wherever it expects to find known objects
157  *            b) the client has removed the object from wherever it expects to
158  *               find known objects and is about to free it, or
159  *            c) the client has freed the object.
160  *            In all these cases (a, b, and c) kmem frees the new object (the
161  *            unused copy destination) and searches for the old object in the
162  *            magazine layer. If found, the object is removed from the magazine
163  *            layer and freed to the slab layer so it will no longer hold the
164  *            slab hostage.
165  *
166  * 2.3 Object States
167  *
168  * Neither kmem nor the client can be assumed to know the object's whereabouts
169  * at the time of the callback. An object belonging to a kmem cache may be in
170  * any of the following states:
171  *
172  * 1. Uninitialized on the slab
173  * 2. Allocated from the slab but not constructed (still uninitialized)
174  * 3. Allocated from the slab, constructed, but not yet ready for business
175  *    (not in a valid state for the move callback)
176  * 4. In use (valid and known to the client)
177  * 5. About to be freed (no longer in a valid state for the move callback)
178  * 6. Freed to a magazine (still constructed)
179  * 7. Allocated from a magazine, not yet ready for business (not in a valid
180  *    state for the move callback), and about to return to state #4
181  * 8. Deconstructed on a magazine that is about to be freed
182  * 9. Freed to the slab
183  *
184  * Since the move callback may be called at any time while the object is in any
185  * of the above states (except state #1), the client needs a safe way to
186  * determine whether or not it knows about the object. Specifically, the client
187  * needs to know whether or not the object is in state #4, the only state in
188  * which a move is valid. If the object is in any other state, the client should
189  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
190  * the object's fields.
191  *
192  * Note that although an object may be in state #4 when kmem initiates the move
193  * request, the object may no longer be in that state by the time kmem actually
194  * calls the move function. Not only does the client free objects
195  * asynchronously, kmem itself puts move requests on a queue where thay are
196  * pending until kmem processes them from another context. Also, objects freed
197  * to a magazine appear allocated from the point of view of the slab layer, so
198  * kmem may even initiate requests for objects in a state other than state #4.
199  *
200  * 2.3.1 Magazine Layer
201  *
202  * An important insight revealed by the states listed above is that the magazine
203  * layer is populated only by kmem_cache_free(). Magazines of constructed
204  * objects are never populated directly from the slab layer (which contains raw,
205  * unconstructed objects). Whenever an allocation request cannot be satisfied
206  * from the magazine layer, the magazines are bypassed and the request is
207  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
208  * the object constructor only when allocating from the slab layer, and only in
209  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
210  * the move callback. kmem does not preconstruct objects in anticipation of
211  * kmem_cache_alloc().
212  *
213  * 2.3.2 Object Constructor and Destructor
214  *
215  * If the client supplies a destructor, it must be valid to call the destructor
216  * on a newly created object (immediately after the constructor).
217  *
218  * 2.4 Recognizing Known Objects
219  *
220  * There is a simple test to determine safely whether or not the client knows
221  * about a given object in the move callback. It relies on the fact that kmem
222  * guarantees that the object of the move callback has only been touched by the
223  * client itself or else by kmem. kmem does this by ensuring that none of the
224  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
225  * callback is pending. When the last object on a slab is freed, if there is a
226  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
227  * slabs on that list until all pending callbacks are completed. That way,
228  * clients can be certain that the object of a move callback is in one of the
229  * states listed above, making it possible to distinguish known objects (in
230  * state #4) using the two low order bits of any pointer member (with the
231  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
232  * platforms).
233  *
234  * The test works as long as the client always transitions objects from state #4
235  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
236  * order bit of the client-designated pointer member. Since kmem only writes
237  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
238  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
239  * guaranteed to set at least one of the two low order bits. Therefore, given an
240  * object with a back pointer to a 'container_t *o_container', the client can
241  * test
242  *
243  *      container_t *container = object->o_container;
244  *      if ((uintptr_t)container & 0x3) {
245  *              return (KMEM_CBRC_DONT_KNOW);
246  *      }
247  *
248  * Typically, an object will have a pointer to some structure with a list or
249  * hash where objects from the cache are kept while in use. Assuming that the
250  * client has some way of knowing that the container structure is valid and will
251  * not go away during the move, and assuming that the structure includes a lock
252  * to protect whatever collection is used, then the client would continue as
253  * follows:
254  *
255  *	// Ensure that the container structure does not go away.
256  *      if (container_hold(container) == 0) {
257  *              return (KMEM_CBRC_DONT_KNOW);
258  *      }
259  *      mutex_enter(&container->c_objects_lock);
260  *      if (container != object->o_container) {
261  *              mutex_exit(&container->c_objects_lock);
262  *              container_rele(container);
263  *              return (KMEM_CBRC_DONT_KNOW);
264  *      }
265  *
266  * At this point the client knows that the object cannot be freed as long as
267  * c_objects_lock is held. Note that after acquiring the lock, the client must
268  * recheck the o_container pointer in case the object was removed just before
269  * acquiring the lock.
270  *
271  * When the client is about to free an object, it must first remove that object
272  * from the list, hash, or other structure where it is kept. At that time, to
273  * mark the object so it can be distinguished from the remaining, known objects,
274  * the client sets the designated low order bit:
275  *
276  *      mutex_enter(&container->c_objects_lock);
277  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
278  *      list_remove(&container->c_objects, object);
279  *      mutex_exit(&container->c_objects_lock);
280  *
281  * In the common case, the object is freed to the magazine layer, where it may
282  * be reused on a subsequent allocation without the overhead of calling the
283  * constructor. While in the magazine it appears allocated from the point of
284  * view of the slab layer, making it a candidate for the move callback. Most
285  * objects unrecognized by the client in the move callback fall into this
286  * category and are cheaply distinguished from known objects by the test
287  * described earlier. Since recognition is cheap for the client, and searching
288  * magazines is expensive for kmem, kmem defers searching until the client first
289  * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
290  * elsewhere does what it can to avoid bothering the client unnecessarily.
291  *
292  * Invalidating the designated pointer member before freeing the object marks
293  * the object to be avoided in the callback, and conversely, assigning a valid
294  * value to the designated pointer member after allocating the object makes the
295  * object fair game for the callback:
296  *
297  *      ... allocate object ...
298  *      ... set any initial state not set by the constructor ...
299  *
300  *      mutex_enter(&container->c_objects_lock);
301  *      list_insert_tail(&container->c_objects, object);
302  *      membar_producer();
303  *      object->o_container = container;
304  *      mutex_exit(&container->c_objects_lock);
305  *
306  * Note that everything else must be valid before setting o_container makes the
307  * object fair game for the move callback. The membar_producer() call ensures
308  * that all the object's state is written to memory before setting the pointer
309  * that transitions the object from state #3 or #7 (allocated, constructed, not
310  * yet in use) to state #4 (in use, valid). That's important because the move
311  * function has to check the validity of the pointer before it can safely
312  * acquire the lock protecting the collection where it expects to find known
313  * objects.
314  *
315  * This method of distinguishing known objects observes the usual symmetry:
316  * invalidating the designated pointer is the first thing the client does before
317  * freeing the object, and setting the designated pointer is the last thing the
318  * client does after allocating the object. Of course, the client is not
319  * required to use this method. Fundamentally, how the client recognizes known
320  * objects is completely up to the client, but this method is recommended as an
321  * efficient and safe way to take advantage of the guarantees made by kmem. If
322  * the entire object is arbitrary data without any markable bits from a suitable
323  * pointer member, then the client must find some other method, such as
324  * searching a hash table of known objects.
325  *
326  * 2.5 Preventing Objects From Moving
327  *
328  * Besides a way to distinguish known objects, the other thing that the client
329  * needs is a strategy to ensure that an object will not move while the client
330  * is actively using it. The details of satisfying this requirement tend to be
331  * highly cache-specific. It might seem that the same rules that let a client
332  * remove an object safely should also decide when an object can be moved
333  * safely. However, any object state that makes a removal attempt invalid is
334  * likely to be long-lasting for objects that the client does not expect to
335  * remove. kmem knows nothing about the object state and is equally likely (from
336  * the client's point of view) to request a move for any object in the cache,
337  * whether prepared for removal or not. Even a low percentage of objects stuck
338  * in place by unremovability will defeat the consolidator if the stuck objects
339  * are the same long-lived allocations likely to hold slabs hostage.
340  * Fundamentally, the consolidator is not aimed at common cases. Severe external
341  * fragmentation is a worst case scenario manifested as sparsely allocated
342  * slabs, by definition a low percentage of the cache's objects. When deciding
343  * what makes an object movable, keep in mind the goal of the consolidator: to
344  * bring worst-case external fragmentation within the limits guaranteed for
345  * internal fragmentation. Removability is a poor criterion if it is likely to
346  * exclude more than an insignificant percentage of objects for long periods of
347  * time.
348  *
349  * A tricky general solution exists, and it has the advantage of letting you
350  * move any object at almost any moment, practically eliminating the likelihood
351  * that an object can hold a slab hostage. However, if there is a cache-specific
352  * way to ensure that an object is not actively in use in the vast majority of
353  * cases, a simpler solution that leverages this cache-specific knowledge is
354  * preferred.
355  *
356  * 2.5.1 Cache-Specific Solution
357  *
358  * As an example of a cache-specific solution, the ZFS znode cache takes
359  * advantage of the fact that the vast majority of znodes are only being
360  * referenced from the DNLC. (A typical case might be a few hundred in active
361  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
362  * client has established that it recognizes the znode and can access its fields
363  * safely (using the method described earlier), it then tests whether the znode
364  * is referenced by anything other than the DNLC. If so, it assumes that the
365  * znode may be in active use and is unsafe to move, so it drops its locks and
366  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
367  * else znodes are used, no change is needed to protect against the possibility
368  * of the znode moving. The disadvantage is that it remains possible for an
369  * application to hold a znode slab hostage with an open file descriptor.
370  * However, this case ought to be rare and the consolidator has a way to deal
371  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
372  * object, kmem eventually stops believing it and treats the slab as if the
373  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
374  * then focus on getting it off of the partial slab list by allocating rather
375  * than freeing all of its objects. (Either way of getting a slab off the
376  * free list reduces fragmentation.)
377  *
378  * 2.5.2 General Solution
379  *
380  * The general solution, on the other hand, requires an explicit hold everywhere
381  * the object is used to prevent it from moving. To keep the client locking
382  * strategy as uncomplicated as possible, kmem guarantees the simplifying
383  * assumption that move callbacks are sequential, even across multiple caches.
384  * Internally, a global queue processed by a single thread supports all caches
385  * implementing the callback function. No matter how many caches supply a move
386  * function, the consolidator never moves more than one object at a time, so the
387  * client does not have to worry about tricky lock ordering involving several
388  * related objects from different kmem caches.
389  *
390  * The general solution implements the explicit hold as a read-write lock, which
391  * allows multiple readers to access an object from the cache simultaneously
392  * while a single writer is excluded from moving it. A single rwlock for the
393  * entire cache would lock out all threads from using any of the cache's objects
394  * even though only a single object is being moved, so to reduce contention,
395  * the client can fan out the single rwlock into an array of rwlocks hashed by
396  * the object address, making it probable that moving one object will not
397  * prevent other threads from using a different object. The rwlock cannot be a
398  * member of the object itself, because the possibility of the object moving
399  * makes it unsafe to access any of the object's fields until the lock is
400  * acquired.
401  *
402  * Assuming a small, fixed number of locks, it's possible that multiple objects
403  * will hash to the same lock. A thread that needs to use multiple objects in
404  * the same function may acquire the same lock multiple times. Since rwlocks are
405  * reentrant for readers, and since there is never more than a single writer at
406  * a time (assuming that the client acquires the lock as a writer only when
407  * moving an object inside the callback), there would seem to be no problem.
408  * However, a client locking multiple objects in the same function must handle
409  * one case of potential deadlock: Assume that thread A needs to prevent both
410  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
411  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
412  * same lock, that thread A will acquire the lock for object 1 as a reader
413  * before thread B sets the lock's write-wanted bit, preventing thread A from
414  * reacquiring the lock for object 2 as a reader. Unable to make forward
415  * progress, thread A will never release the lock for object 1, resulting in
416  * deadlock.
417  *
418  * There are two ways of avoiding the deadlock just described. The first is to
419  * use rw_tryenter() rather than rw_enter() in the callback function when
420  * attempting to acquire the lock as a writer. If tryenter discovers that the
421  * same object (or another object hashed to the same lock) is already in use, it
422  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
423  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
424  * since it allows a thread to acquire the lock as a reader in spite of a
425  * waiting writer. This second approach insists on moving the object now, no
426  * matter how many readers the move function must wait for in order to do so,
427  * and could delay the completion of the callback indefinitely (blocking
428  * callbacks to other clients). In practice, a less insistent callback using
429  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
430  * little reason to use anything else.
431  *
432  * Avoiding deadlock is not the only problem that an implementation using an
433  * explicit hold needs to solve. Locking the object in the first place (to
434  * prevent it from moving) remains a problem, since the object could move
435  * between the time you obtain a pointer to the object and the time you acquire
436  * the rwlock hashed to that pointer value. Therefore the client needs to
437  * recheck the value of the pointer after acquiring the lock, drop the lock if
438  * the value has changed, and try again. This requires a level of indirection:
439  * something that points to the object rather than the object itself, that the
440  * client can access safely while attempting to acquire the lock. (The object
441  * itself cannot be referenced safely because it can move at any time.)
442  * The following lock-acquisition function takes whatever is safe to reference
443  * (arg), follows its pointer to the object (using function f), and tries as
444  * often as necessary to acquire the hashed lock and verify that the object
445  * still has not moved:
446  *
447  *      object_t *
448  *      object_hold(object_f f, void *arg)
449  *      {
450  *              object_t *op;
451  *
452  *              op = f(arg);
453  *              if (op == NULL) {
454  *                      return (NULL);
455  *              }
456  *
457  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
458  *              while (op != f(arg)) {
459  *                      rw_exit(OBJECT_RWLOCK(op));
460  *                      op = f(arg);
461  *                      if (op == NULL) {
462  *                              break;
463  *                      }
464  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
465  *              }
466  *
467  *              return (op);
468  *      }
469  *
470  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
471  * lock reacquisition loop, while necessary, almost never executes. The function
472  * pointer f (used to obtain the object pointer from arg) has the following type
473  * definition:
474  *
475  *      typedef object_t *(*object_f)(void *arg);
476  *
477  * An object_f implementation is likely to be as simple as accessing a structure
478  * member:
479  *
480  *      object_t *
481  *      s_object(void *arg)
482  *      {
483  *              something_t *sp = arg;
484  *              return (sp->s_object);
485  *      }
486  *
487  * The flexibility of a function pointer allows the path to the object to be
488  * arbitrarily complex and also supports the notion that depending on where you
489  * are using the object, you may need to get it from someplace different.
490  *
491  * The function that releases the explicit hold is simpler because it does not
492  * have to worry about the object moving:
493  *
494  *      void
495  *      object_rele(object_t *op)
496  *      {
497  *              rw_exit(OBJECT_RWLOCK(op));
498  *      }
499  *
500  * The caller is spared these details so that obtaining and releasing an
501  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
502  * of object_hold() only needs to know that the returned object pointer is valid
503  * if not NULL and that the object will not move until released.
504  *
505  * Although object_hold() prevents an object from moving, it does not prevent it
506  * from being freed. The caller must take measures before calling object_hold()
507  * (afterwards is too late) to ensure that the held object cannot be freed. The
508  * caller must do so without accessing the unsafe object reference, so any lock
509  * or reference count used to ensure the continued existence of the object must
510  * live outside the object itself.
511  *
512  * Obtaining a new object is a special case where an explicit hold is impossible
513  * for the caller. Any function that returns a newly allocated object (either as
514  * a return value, or as an in-out paramter) must return it already held; after
515  * the caller gets it is too late, since the object cannot be safely accessed
516  * without the level of indirection described earlier. The following
517  * object_alloc() example uses the same code shown earlier to transition a new
518  * object into the state of being recognized (by the client) as a known object.
519  * The function must acquire the hold (rw_enter) before that state transition
520  * makes the object movable:
521  *
522  *      static object_t *
523  *      object_alloc(container_t *container)
524  *      {
525  *              object_t *object = kmem_cache_alloc(object_cache, 0);
526  *              ... set any initial state not set by the constructor ...
527  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
528  *              mutex_enter(&container->c_objects_lock);
529  *              list_insert_tail(&container->c_objects, object);
530  *              membar_producer();
531  *              object->o_container = container;
532  *              mutex_exit(&container->c_objects_lock);
533  *              return (object);
534  *      }
535  *
536  * Functions that implicitly acquire an object hold (any function that calls
537  * object_alloc() to supply an object for the caller) need to be carefully noted
538  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
539  * prevent all objects hashed to the affected rwlocks from ever being moved.
540  *
541  * The pointer to a held object can be hashed to the holding rwlock even after
542  * the object has been freed. Although it is possible to release the hold
543  * after freeing the object, you may decide to release the hold implicitly in
544  * whatever function frees the object, so as to release the hold as soon as
545  * possible, and for the sake of symmetry with the function that implicitly
546  * acquires the hold when it allocates the object. Here, object_free() releases
547  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
548  * matching pair with object_hold():
549  *
550  *      void
551  *      object_free(object_t *object)
552  *      {
553  *              container_t *container;
554  *
555  *              ASSERT(object_held(object));
556  *              container = object->o_container;
557  *              mutex_enter(&container->c_objects_lock);
558  *              object->o_container =
559  *                  (void *)((uintptr_t)object->o_container | 0x1);
560  *              list_remove(&container->c_objects, object);
561  *              mutex_exit(&container->c_objects_lock);
562  *              object_rele(object);
563  *              kmem_cache_free(object_cache, object);
564  *      }
565  *
566  * Note that object_free() cannot safely accept an object pointer as an argument
567  * unless the object is already held. Any function that calls object_free()
568  * needs to be carefully noted since it similarly forms a matching pair with
569  * object_hold().
570  *
571  * To complete the picture, the following callback function implements the
572  * general solution by moving objects only if they are currently unheld:
573  *
574  *      static kmem_cbrc_t
575  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
576  *      {
577  *              object_t *op = buf, *np = newbuf;
578  *              container_t *container;
579  *
580  *              container = op->o_container;
581  *              if ((uintptr_t)container & 0x3) {
582  *                      return (KMEM_CBRC_DONT_KNOW);
583  *              }
584  *
585  *	        // Ensure that the container structure does not go away.
586  *              if (container_hold(container) == 0) {
587  *                      return (KMEM_CBRC_DONT_KNOW);
588  *              }
589  *
590  *              mutex_enter(&container->c_objects_lock);
591  *              if (container != op->o_container) {
592  *                      mutex_exit(&container->c_objects_lock);
593  *                      container_rele(container);
594  *                      return (KMEM_CBRC_DONT_KNOW);
595  *              }
596  *
597  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
598  *                      mutex_exit(&container->c_objects_lock);
599  *                      container_rele(container);
600  *                      return (KMEM_CBRC_LATER);
601  *              }
602  *
603  *              object_move_impl(op, np); // critical section
604  *              rw_exit(OBJECT_RWLOCK(op));
605  *
606  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
607  *              list_link_replace(&op->o_link_node, &np->o_link_node);
608  *              mutex_exit(&container->c_objects_lock);
609  *              container_rele(container);
610  *              return (KMEM_CBRC_YES);
611  *      }
612  *
613  * Note that object_move() must invalidate the designated o_container pointer of
614  * the old object in the same way that object_free() does, since kmem will free
615  * the object in response to the KMEM_CBRC_YES return value.
616  *
617  * The lock order in object_move() differs from object_alloc(), which locks
618  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
619  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
620  * not a problem. Holding the lock on the object list in the example above
621  * through the entire callback not only prevents the object from going away, it
622  * also allows you to lock the list elsewhere and know that none of its elements
623  * will move during iteration.
624  *
625  * Adding an explicit hold everywhere an object from the cache is used is tricky
626  * and involves much more change to client code than a cache-specific solution
627  * that leverages existing state to decide whether or not an object is
628  * movable. However, this approach has the advantage that no object remains
629  * immovable for any significant length of time, making it extremely unlikely
630  * that long-lived allocations can continue holding slabs hostage; and it works
631  * for any cache.
632  *
633  * 3. Consolidator Implementation
634  *
635  * Once the client supplies a move function that a) recognizes known objects and
636  * b) avoids moving objects that are actively in use, the remaining work is up
637  * to the consolidator to decide which objects to move and when to issue
638  * callbacks.
639  *
640  * The consolidator relies on the fact that a cache's slabs are ordered by
641  * usage. Each slab has a fixed number of objects. Depending on the slab's
642  * "color" (the offset of the first object from the beginning of the slab;
643  * offsets are staggered to mitigate false sharing of cache lines) it is either
644  * the maximum number of objects per slab determined at cache creation time or
645  * else the number closest to the maximum that fits within the space remaining
646  * after the initial offset. A completely allocated slab may contribute some
647  * internal fragmentation (per-slab overhead) but no external fragmentation, so
648  * it is of no interest to the consolidator. At the other extreme, slabs whose
649  * objects have all been freed to the slab are released to the virtual memory
650  * (VM) subsystem (objects freed to magazines are still allocated as far as the
651  * slab is concerned). External fragmentation exists when there are slabs
652  * somewhere between these extremes. A partial slab has at least one but not all
653  * of its objects allocated. The more partial slabs, and the fewer allocated
654  * objects on each of them, the higher the fragmentation. Hence the
655  * consolidator's overall strategy is to reduce the number of partial slabs by
656  * moving allocated objects from the least allocated slabs to the most allocated
657  * slabs.
658  *
659  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
660  * slabs are kept separately in an unordered list. Since the majority of slabs
661  * tend to be completely allocated (a typical unfragmented cache may have
662  * thousands of complete slabs and only a single partial slab), separating
663  * complete slabs improves the efficiency of partial slab ordering, since the
664  * complete slabs do not affect the depth or balance of the AVL tree. This
665  * ordered sequence of partial slabs acts as a "free list" supplying objects for
666  * allocation requests.
667  *
668  * Objects are always allocated from the first partial slab in the free list,
669  * where the allocation is most likely to eliminate a partial slab (by
670  * completely allocating it). Conversely, when a single object from a completely
671  * allocated slab is freed to the slab, that slab is added to the front of the
672  * free list. Since most free list activity involves highly allocated slabs
673  * coming and going at the front of the list, slabs tend naturally toward the
674  * ideal order: highly allocated at the front, sparsely allocated at the back.
675  * Slabs with few allocated objects are likely to become completely free if they
676  * keep a safe distance away from the front of the free list. Slab misorders
677  * interfere with the natural tendency of slabs to become completely free or
678  * completely allocated. For example, a slab with a single allocated object
679  * needs only a single free to escape the cache; its natural desire is
680  * frustrated when it finds itself at the front of the list where a second
681  * allocation happens just before the free could have released it. Another slab
682  * with all but one object allocated might have supplied the buffer instead, so
683  * that both (as opposed to neither) of the slabs would have been taken off the
684  * free list.
685  *
686  * Although slabs tend naturally toward the ideal order, misorders allowed by a
687  * simple list implementation defeat the consolidator's strategy of merging
688  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
689  * needs another way to fix misorders to optimize its callback strategy. One
690  * approach is to periodically scan a limited number of slabs, advancing a
691  * marker to hold the current scan position, and to move extreme misorders to
692  * the front or back of the free list and to the front or back of the current
693  * scan range. By making consecutive scan ranges overlap by one slab, the least
694  * allocated slab in the current range can be carried along from the end of one
695  * scan to the start of the next.
696  *
697  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
698  * task, however. Since most of the cache's activity is in the magazine layer,
699  * and allocations from the slab layer represent only a startup cost, the
700  * overhead of maintaining a balanced tree is not a significant concern compared
701  * to the opportunity of reducing complexity by eliminating the partial slab
702  * scanner just described. The overhead of an AVL tree is minimized by
703  * maintaining only partial slabs in the tree and keeping completely allocated
704  * slabs separately in a list. To avoid increasing the size of the slab
705  * structure the AVL linkage pointers are reused for the slab's list linkage,
706  * since the slab will always be either partial or complete, never stored both
707  * ways at the same time. To further minimize the overhead of the AVL tree the
708  * compare function that orders partial slabs by usage divides the range of
709  * allocated object counts into bins such that counts within the same bin are
710  * considered equal. Binning partial slabs makes it less likely that allocating
711  * or freeing a single object will change the slab's order, requiring a tree
712  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
713  * requiring some rebalancing of the tree). Allocation counts closest to
714  * completely free and completely allocated are left unbinned (finely sorted) to
715  * better support the consolidator's strategy of merging slabs at either
716  * extreme.
717  *
718  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
719  *
720  * The consolidator piggybacks on the kmem maintenance thread and is called on
721  * the same interval as kmem_cache_update(), once per cache every fifteen
722  * seconds. kmem maintains a running count of unallocated objects in the slab
723  * layer (cache_bufslab). The consolidator checks whether that number exceeds
724  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
725  * there is a significant number of slabs in the cache (arbitrarily a minimum
726  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
727  * working set are included in the assessment, and magazines in the depot are
728  * reaped if those objects would lift cache_bufslab above the fragmentation
729  * threshold. Once the consolidator decides that a cache is fragmented, it looks
730  * for a candidate slab to reclaim, starting at the end of the partial slab free
731  * list and scanning backwards. At first the consolidator is choosy: only a slab
732  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
733  * single allocated object, regardless of percentage). If there is difficulty
734  * finding a candidate slab, kmem raises the allocation threshold incrementally,
735  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
736  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
737  * even in the worst case of every slab in the cache being almost 7/8 allocated.
738  * The threshold can also be lowered incrementally when candidate slabs are easy
739  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
740  * is no longer fragmented.
741  *
742  * 3.2 Generating Callbacks
743  *
744  * Once an eligible slab is chosen, a callback is generated for every allocated
745  * object on the slab, in the hope that the client will move everything off the
746  * slab and make it reclaimable. Objects selected as move destinations are
747  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
748  * order (most allocated at the front, least allocated at the back) and a
749  * cooperative client, the consolidator will succeed in removing slabs from both
750  * ends of the free list, completely allocating on the one hand and completely
751  * freeing on the other. Objects selected as move destinations are allocated in
752  * the kmem maintenance thread where move requests are enqueued. A separate
753  * callback thread removes pending callbacks from the queue and calls the
754  * client. The separate thread ensures that client code (the move function) does
755  * not interfere with internal kmem maintenance tasks. A map of pending
756  * callbacks keyed by object address (the object to be moved) is checked to
757  * ensure that duplicate callbacks are not generated for the same object.
758  * Allocating the move destination (the object to move to) prevents subsequent
759  * callbacks from selecting the same destination as an earlier pending callback.
760  *
761  * Move requests can also be generated by kmem_cache_reap() when the system is
762  * desperate for memory and by kmem_cache_move_notify(), called by the client to
763  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
764  * The map of pending callbacks is protected by the same lock that protects the
765  * slab layer.
766  *
767  * When the system is desperate for memory, kmem does not bother to determine
768  * whether or not the cache exceeds the fragmentation threshold, but tries to
769  * consolidate as many slabs as possible. Normally, the consolidator chews
770  * slowly, one sparsely allocated slab at a time during each maintenance
771  * interval that the cache is fragmented. When desperate, the consolidator
772  * starts at the last partial slab and enqueues callbacks for every allocated
773  * object on every partial slab, working backwards until it reaches the first
774  * partial slab. The first partial slab, meanwhile, advances in pace with the
775  * consolidator as allocations to supply move destinations for the enqueued
776  * callbacks use up the highly allocated slabs at the front of the free list.
777  * Ideally, the overgrown free list collapses like an accordion, starting at
778  * both ends and ending at the center with a single partial slab.
779  *
780  * 3.3 Client Responses
781  *
782  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
783  * marks the slab that supplied the stuck object non-reclaimable and moves it to
784  * front of the free list. The slab remains marked as long as it remains on the
785  * free list, and it appears more allocated to the partial slab compare function
786  * than any unmarked slab, no matter how many of its objects are allocated.
787  * Since even one immovable object ties up the entire slab, the goal is to
788  * completely allocate any slab that cannot be completely freed. kmem does not
789  * bother generating callbacks to move objects from a marked slab unless the
790  * system is desperate.
791  *
792  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
793  * slab. If the client responds LATER too many times, kmem disbelieves and
794  * treats the response as a NO. The count is cleared when the slab is taken off
795  * the partial slab list or when the client moves one of the slab's objects.
796  *
797  * 4. Observability
798  *
799  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
800  * the ::kmem_slabs dcmd. For a complete description of the command, enter
801  * '::help kmem_slabs' at the mdb prompt.
802  */
803 
804 #include <sys/kmem_impl.h>
805 #include <sys/vmem_impl.h>
806 #include <sys/param.h>
807 #include <sys/sysmacros.h>
808 #include <sys/vm.h>
809 #include <sys/proc.h>
810 #include <sys/tuneable.h>
811 #include <sys/systm.h>
812 #include <sys/cmn_err.h>
813 #include <sys/debug.h>
814 #include <sys/sdt.h>
815 #include <sys/mutex.h>
816 #include <sys/bitmap.h>
817 #include <sys/atomic.h>
818 #include <sys/kobj.h>
819 #include <sys/disp.h>
820 #include <vm/seg_kmem.h>
821 #include <sys/log.h>
822 #include <sys/callb.h>
823 #include <sys/taskq.h>
824 #include <sys/modctl.h>
825 #include <sys/reboot.h>
826 #include <sys/id32.h>
827 #include <sys/zone.h>
828 #include <sys/netstack.h>
829 #ifdef	DEBUG
830 #include <sys/random.h>
831 #endif
832 
833 extern void streams_msg_init(void);
834 extern int segkp_fromheap;
835 extern void segkp_cache_free(void);
836 
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_move_callbacks;
865 	kstat_named_t	kmc_move_yes;
866 	kstat_named_t	kmc_move_no;
867 	kstat_named_t	kmc_move_later;
868 	kstat_named_t	kmc_move_dont_need;
869 	kstat_named_t	kmc_move_dont_know;
870 	kstat_named_t	kmc_move_hunt_found;
871 } kmem_cache_kstat = {
872 	{ "buf_size",		KSTAT_DATA_UINT64 },
873 	{ "align",		KSTAT_DATA_UINT64 },
874 	{ "chunk_size",		KSTAT_DATA_UINT64 },
875 	{ "slab_size",		KSTAT_DATA_UINT64 },
876 	{ "alloc",		KSTAT_DATA_UINT64 },
877 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
878 	{ "free",		KSTAT_DATA_UINT64 },
879 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
880 	{ "depot_free",		KSTAT_DATA_UINT64 },
881 	{ "depot_contention",	KSTAT_DATA_UINT64 },
882 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
883 	{ "slab_free",		KSTAT_DATA_UINT64 },
884 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
885 	{ "buf_avail",		KSTAT_DATA_UINT64 },
886 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
887 	{ "buf_total",		KSTAT_DATA_UINT64 },
888 	{ "buf_max",		KSTAT_DATA_UINT64 },
889 	{ "slab_create",	KSTAT_DATA_UINT64 },
890 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
891 	{ "vmem_source",	KSTAT_DATA_UINT64 },
892 	{ "hash_size",		KSTAT_DATA_UINT64 },
893 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
894 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
895 	{ "full_magazines",	KSTAT_DATA_UINT64 },
896 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
897 	{ "magazine_size",	KSTAT_DATA_UINT64 },
898 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
899 	{ "move_yes",		KSTAT_DATA_UINT64 },
900 	{ "move_no",		KSTAT_DATA_UINT64 },
901 	{ "move_later",		KSTAT_DATA_UINT64 },
902 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
903 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
904 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
905 };
906 
907 static kmutex_t kmem_cache_kstat_lock;
908 
909 /*
910  * The default set of caches to back kmem_alloc().
911  * These sizes should be reevaluated periodically.
912  *
913  * We want allocations that are multiples of the coherency granularity
914  * (64 bytes) to be satisfied from a cache which is a multiple of 64
915  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
916  * the next kmem_cache_size greater than or equal to it must be a
917  * multiple of 64.
918  *
919  * We split the table into two sections:  size <= 4k and size > 4k.  This
920  * saves a lot of space and cache footprint in our cache tables.
921  */
922 static const int kmem_alloc_sizes[] = {
923 	1 * 8,
924 	2 * 8,
925 	3 * 8,
926 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
927 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
928 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
929 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
930 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
931 	P2ALIGN(8192 / 7, 64),
932 	P2ALIGN(8192 / 6, 64),
933 	P2ALIGN(8192 / 5, 64),
934 	P2ALIGN(8192 / 4, 64),
935 	P2ALIGN(8192 / 3, 64),
936 	P2ALIGN(8192 / 2, 64),
937 };
938 
939 static const int kmem_big_alloc_sizes[] = {
940 	2 * 4096,	3 * 4096,
941 	2 * 8192,	3 * 8192,
942 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
943 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
944 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
945 	16 * 8192
946 };
947 
948 #define	KMEM_MAXBUF		4096
949 #define	KMEM_BIG_MAXBUF_32BIT	32768
950 #define	KMEM_BIG_MAXBUF		131072
951 
952 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
953 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
954 
955 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
956 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
957 
958 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
959 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
960 
961 static kmem_magtype_t kmem_magtype[] = {
962 	{ 1,	8,	3200,	65536	},
963 	{ 3,	16,	256,	32768	},
964 	{ 7,	32,	64,	16384	},
965 	{ 15,	64,	0,	8192	},
966 	{ 31,	64,	0,	4096	},
967 	{ 47,	64,	0,	2048	},
968 	{ 63,	64,	0,	1024	},
969 	{ 95,	64,	0,	512	},
970 	{ 143,	64,	0,	0	},
971 };
972 
973 static uint32_t kmem_reaping;
974 static uint32_t kmem_reaping_idspace;
975 
976 /*
977  * kmem tunables
978  */
979 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
980 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
981 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
982 int kmem_panic = 1;		/* whether to panic on error */
983 int kmem_logging = 1;		/* kmem_log_enter() override */
984 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
985 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
986 size_t kmem_content_log_size;	/* content log size [2% of memory] */
987 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
988 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
989 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
990 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
991 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
992 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
993 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
994 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
995 
996 #ifdef _LP64
997 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
998 #else
999 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1000 #endif
1001 
1002 #ifdef DEBUG
1003 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1004 #else
1005 int kmem_flags = 0;
1006 #endif
1007 int kmem_ready;
1008 
1009 static kmem_cache_t	*kmem_slab_cache;
1010 static kmem_cache_t	*kmem_bufctl_cache;
1011 static kmem_cache_t	*kmem_bufctl_audit_cache;
1012 
1013 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1014 static list_t		kmem_caches;
1015 
1016 static taskq_t		*kmem_taskq;
1017 static kmutex_t		kmem_flags_lock;
1018 static vmem_t		*kmem_metadata_arena;
1019 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1020 static vmem_t		*kmem_cache_arena;
1021 static vmem_t		*kmem_hash_arena;
1022 static vmem_t		*kmem_log_arena;
1023 static vmem_t		*kmem_oversize_arena;
1024 static vmem_t		*kmem_va_arena;
1025 static vmem_t		*kmem_default_arena;
1026 static vmem_t		*kmem_firewall_va_arena;
1027 static vmem_t		*kmem_firewall_arena;
1028 
1029 /*
1030  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1031  * turned on when DEBUG is also defined.
1032  */
1033 #ifdef	DEBUG
1034 #define	KMEM_STATS
1035 #endif	/* DEBUG */
1036 
1037 #ifdef	KMEM_STATS
1038 #define	KMEM_STAT_ADD(stat)			((stat)++)
1039 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1040 #else
1041 #define	KMEM_STAT_ADD(stat)			/* nothing */
1042 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1043 #endif	/* KMEM_STATS */
1044 
1045 /*
1046  * kmem slab consolidator thresholds (tunables)
1047  */
1048 static size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1049 static size_t kmem_frag_numer = 1;	/* free buffers (numerator) */
1050 static size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1051 /*
1052  * Maximum number of slabs from which to move buffers during a single
1053  * maintenance interval while the system is not low on memory.
1054  */
1055 static size_t kmem_reclaim_max_slabs = 1;
1056 /*
1057  * Number of slabs to scan backwards from the end of the partial slab list
1058  * when searching for buffers to relocate.
1059  */
1060 static size_t kmem_reclaim_scan_range = 12;
1061 
1062 #ifdef	KMEM_STATS
1063 static struct {
1064 	uint64_t kms_callbacks;
1065 	uint64_t kms_yes;
1066 	uint64_t kms_no;
1067 	uint64_t kms_later;
1068 	uint64_t kms_dont_need;
1069 	uint64_t kms_dont_know;
1070 	uint64_t kms_hunt_found_slab;
1071 	uint64_t kms_hunt_found_mag;
1072 	uint64_t kms_hunt_alloc_fail;
1073 	uint64_t kms_hunt_lucky;
1074 	uint64_t kms_notify;
1075 	uint64_t kms_notify_callbacks;
1076 	uint64_t kms_disbelief;
1077 	uint64_t kms_already_pending;
1078 	uint64_t kms_callback_alloc_fail;
1079 	uint64_t kms_callback_taskq_fail;
1080 	uint64_t kms_endscan_slab_destroyed;
1081 	uint64_t kms_endscan_nomem;
1082 	uint64_t kms_endscan_slab_all_used;
1083 	uint64_t kms_endscan_refcnt_changed;
1084 	uint64_t kms_endscan_nomove_changed;
1085 	uint64_t kms_endscan_freelist;
1086 	uint64_t kms_avl_update;
1087 	uint64_t kms_avl_noupdate;
1088 	uint64_t kms_no_longer_reclaimable;
1089 	uint64_t kms_notify_no_longer_reclaimable;
1090 	uint64_t kms_alloc_fail;
1091 	uint64_t kms_constructor_fail;
1092 	uint64_t kms_dead_slabs_freed;
1093 	uint64_t kms_defrags;
1094 	uint64_t kms_scan_depot_ws_reaps;
1095 	uint64_t kms_debug_reaps;
1096 	uint64_t kms_debug_move_scans;
1097 } kmem_move_stats;
1098 #endif	/* KMEM_STATS */
1099 
1100 /* consolidator knobs */
1101 static boolean_t kmem_move_noreap;
1102 static boolean_t kmem_move_blocked;
1103 static boolean_t kmem_move_fulltilt;
1104 static boolean_t kmem_move_any_partial;
1105 
1106 #ifdef	DEBUG
1107 /*
1108  * Ensure code coverage by occasionally running the consolidator even when the
1109  * caches are not fragmented (they may never be). These intervals are mean time
1110  * in cache maintenance intervals (kmem_cache_update).
1111  */
1112 static int kmem_mtb_move = 60;		/* defrag 1 slab (~15min) */
1113 static int kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1114 #endif	/* DEBUG */
1115 
1116 static kmem_cache_t	*kmem_defrag_cache;
1117 static kmem_cache_t	*kmem_move_cache;
1118 static taskq_t		*kmem_move_taskq;
1119 
1120 static void kmem_cache_scan(kmem_cache_t *);
1121 static void kmem_cache_defrag(kmem_cache_t *);
1122 
1123 
1124 kmem_log_header_t	*kmem_transaction_log;
1125 kmem_log_header_t	*kmem_content_log;
1126 kmem_log_header_t	*kmem_failure_log;
1127 kmem_log_header_t	*kmem_slab_log;
1128 
1129 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1130 
1131 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1132 	if ((count) > 0) {						\
1133 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1134 		pc_t *_e;						\
1135 		/* memmove() the old entries down one notch */		\
1136 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1137 			*_e = *(_e - 1);				\
1138 		*_s = (uintptr_t)(caller);				\
1139 	}
1140 
1141 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1142 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1143 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1144 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1145 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1146 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1147 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1148 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1149 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1150 
1151 struct {
1152 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1153 	int		kmp_error;	/* type of kmem error */
1154 	void		*kmp_buffer;	/* buffer that induced panic */
1155 	void		*kmp_realbuf;	/* real start address for buffer */
1156 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1157 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1158 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1159 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1160 } kmem_panic_info;
1161 
1162 
1163 static void
1164 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1165 {
1166 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1167 	uint64_t *buf = buf_arg;
1168 
1169 	while (buf < bufend)
1170 		*buf++ = pattern;
1171 }
1172 
1173 static void *
1174 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1175 {
1176 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1177 	uint64_t *buf;
1178 
1179 	for (buf = buf_arg; buf < bufend; buf++)
1180 		if (*buf != pattern)
1181 			return (buf);
1182 	return (NULL);
1183 }
1184 
1185 static void *
1186 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1187 {
1188 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1189 	uint64_t *buf;
1190 
1191 	for (buf = buf_arg; buf < bufend; buf++) {
1192 		if (*buf != old) {
1193 			copy_pattern(old, buf_arg,
1194 			    (char *)buf - (char *)buf_arg);
1195 			return (buf);
1196 		}
1197 		*buf = new;
1198 	}
1199 
1200 	return (NULL);
1201 }
1202 
1203 static void
1204 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1205 {
1206 	kmem_cache_t *cp;
1207 
1208 	mutex_enter(&kmem_cache_lock);
1209 	for (cp = list_head(&kmem_caches); cp != NULL;
1210 	    cp = list_next(&kmem_caches, cp))
1211 		if (tq != NULL)
1212 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1213 			    tqflag);
1214 		else
1215 			func(cp);
1216 	mutex_exit(&kmem_cache_lock);
1217 }
1218 
1219 static void
1220 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1221 {
1222 	kmem_cache_t *cp;
1223 
1224 	mutex_enter(&kmem_cache_lock);
1225 	for (cp = list_head(&kmem_caches); cp != NULL;
1226 	    cp = list_next(&kmem_caches, cp)) {
1227 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1228 			continue;
1229 		if (tq != NULL)
1230 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1231 			    tqflag);
1232 		else
1233 			func(cp);
1234 	}
1235 	mutex_exit(&kmem_cache_lock);
1236 }
1237 
1238 /*
1239  * Debugging support.  Given a buffer address, find its slab.
1240  */
1241 static kmem_slab_t *
1242 kmem_findslab(kmem_cache_t *cp, void *buf)
1243 {
1244 	kmem_slab_t *sp;
1245 
1246 	mutex_enter(&cp->cache_lock);
1247 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1248 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1249 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1250 			mutex_exit(&cp->cache_lock);
1251 			return (sp);
1252 		}
1253 	}
1254 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1255 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1256 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1257 			mutex_exit(&cp->cache_lock);
1258 			return (sp);
1259 		}
1260 	}
1261 	mutex_exit(&cp->cache_lock);
1262 
1263 	return (NULL);
1264 }
1265 
1266 static void
1267 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1268 {
1269 	kmem_buftag_t *btp = NULL;
1270 	kmem_bufctl_t *bcp = NULL;
1271 	kmem_cache_t *cp = cparg;
1272 	kmem_slab_t *sp;
1273 	uint64_t *off;
1274 	void *buf = bufarg;
1275 
1276 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1277 
1278 	kmem_panic_info.kmp_timestamp = gethrtime();
1279 
1280 	sp = kmem_findslab(cp, buf);
1281 	if (sp == NULL) {
1282 		for (cp = list_tail(&kmem_caches); cp != NULL;
1283 		    cp = list_prev(&kmem_caches, cp)) {
1284 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1285 				break;
1286 		}
1287 	}
1288 
1289 	if (sp == NULL) {
1290 		cp = NULL;
1291 		error = KMERR_BADADDR;
1292 	} else {
1293 		if (cp != cparg)
1294 			error = KMERR_BADCACHE;
1295 		else
1296 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1297 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1298 		if (buf != bufarg)
1299 			error = KMERR_BADBASE;
1300 		if (cp->cache_flags & KMF_BUFTAG)
1301 			btp = KMEM_BUFTAG(cp, buf);
1302 		if (cp->cache_flags & KMF_HASH) {
1303 			mutex_enter(&cp->cache_lock);
1304 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1305 				if (bcp->bc_addr == buf)
1306 					break;
1307 			mutex_exit(&cp->cache_lock);
1308 			if (bcp == NULL && btp != NULL)
1309 				bcp = btp->bt_bufctl;
1310 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1311 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1312 			    bcp->bc_addr != buf) {
1313 				error = KMERR_BADBUFCTL;
1314 				bcp = NULL;
1315 			}
1316 		}
1317 	}
1318 
1319 	kmem_panic_info.kmp_error = error;
1320 	kmem_panic_info.kmp_buffer = bufarg;
1321 	kmem_panic_info.kmp_realbuf = buf;
1322 	kmem_panic_info.kmp_cache = cparg;
1323 	kmem_panic_info.kmp_realcache = cp;
1324 	kmem_panic_info.kmp_slab = sp;
1325 	kmem_panic_info.kmp_bufctl = bcp;
1326 
1327 	printf("kernel memory allocator: ");
1328 
1329 	switch (error) {
1330 
1331 	case KMERR_MODIFIED:
1332 		printf("buffer modified after being freed\n");
1333 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1334 		if (off == NULL)	/* shouldn't happen */
1335 			off = buf;
1336 		printf("modification occurred at offset 0x%lx "
1337 		    "(0x%llx replaced by 0x%llx)\n",
1338 		    (uintptr_t)off - (uintptr_t)buf,
1339 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1340 		break;
1341 
1342 	case KMERR_REDZONE:
1343 		printf("redzone violation: write past end of buffer\n");
1344 		break;
1345 
1346 	case KMERR_BADADDR:
1347 		printf("invalid free: buffer not in cache\n");
1348 		break;
1349 
1350 	case KMERR_DUPFREE:
1351 		printf("duplicate free: buffer freed twice\n");
1352 		break;
1353 
1354 	case KMERR_BADBUFTAG:
1355 		printf("boundary tag corrupted\n");
1356 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1357 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1358 		    KMEM_BUFTAG_FREE);
1359 		break;
1360 
1361 	case KMERR_BADBUFCTL:
1362 		printf("bufctl corrupted\n");
1363 		break;
1364 
1365 	case KMERR_BADCACHE:
1366 		printf("buffer freed to wrong cache\n");
1367 		printf("buffer was allocated from %s,\n", cp->cache_name);
1368 		printf("caller attempting free to %s.\n", cparg->cache_name);
1369 		break;
1370 
1371 	case KMERR_BADSIZE:
1372 		printf("bad free: free size (%u) != alloc size (%u)\n",
1373 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1374 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1375 		break;
1376 
1377 	case KMERR_BADBASE:
1378 		printf("bad free: free address (%p) != alloc address (%p)\n",
1379 		    bufarg, buf);
1380 		break;
1381 	}
1382 
1383 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1384 	    bufarg, (void *)bcp, cparg->cache_name);
1385 
1386 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1387 	    error != KMERR_BADBUFCTL) {
1388 		int d;
1389 		timestruc_t ts;
1390 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1391 
1392 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1393 		printf("previous transaction on buffer %p:\n", buf);
1394 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1395 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1396 		    (void *)sp, cp->cache_name);
1397 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1398 			ulong_t off;
1399 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1400 			printf("%s+%lx\n", sym ? sym : "?", off);
1401 		}
1402 	}
1403 	if (kmem_panic > 0)
1404 		panic("kernel heap corruption detected");
1405 	if (kmem_panic == 0)
1406 		debug_enter(NULL);
1407 	kmem_logging = 1;	/* resume logging */
1408 }
1409 
1410 static kmem_log_header_t *
1411 kmem_log_init(size_t logsize)
1412 {
1413 	kmem_log_header_t *lhp;
1414 	int nchunks = 4 * max_ncpus;
1415 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1416 	int i;
1417 
1418 	/*
1419 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1420 	 * to prevent false sharing of cache lines.
1421 	 */
1422 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1423 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1424 	    NULL, NULL, VM_SLEEP);
1425 	bzero(lhp, lhsize);
1426 
1427 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1428 	lhp->lh_nchunks = nchunks;
1429 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1430 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1431 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1432 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1433 	    nchunks * sizeof (int), VM_SLEEP);
1434 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1435 
1436 	for (i = 0; i < max_ncpus; i++) {
1437 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1438 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1439 		clhp->clh_chunk = i;
1440 	}
1441 
1442 	for (i = max_ncpus; i < nchunks; i++)
1443 		lhp->lh_free[i] = i;
1444 
1445 	lhp->lh_head = max_ncpus;
1446 	lhp->lh_tail = 0;
1447 
1448 	return (lhp);
1449 }
1450 
1451 static void *
1452 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1453 {
1454 	void *logspace;
1455 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1456 
1457 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1458 		return (NULL);
1459 
1460 	mutex_enter(&clhp->clh_lock);
1461 	clhp->clh_hits++;
1462 	if (size > clhp->clh_avail) {
1463 		mutex_enter(&lhp->lh_lock);
1464 		lhp->lh_hits++;
1465 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1466 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1467 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1468 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1469 		clhp->clh_current = lhp->lh_base +
1470 		    clhp->clh_chunk * lhp->lh_chunksize;
1471 		clhp->clh_avail = lhp->lh_chunksize;
1472 		if (size > lhp->lh_chunksize)
1473 			size = lhp->lh_chunksize;
1474 		mutex_exit(&lhp->lh_lock);
1475 	}
1476 	logspace = clhp->clh_current;
1477 	clhp->clh_current += size;
1478 	clhp->clh_avail -= size;
1479 	bcopy(data, logspace, size);
1480 	mutex_exit(&clhp->clh_lock);
1481 	return (logspace);
1482 }
1483 
1484 #define	KMEM_AUDIT(lp, cp, bcp)						\
1485 {									\
1486 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1487 	_bcp->bc_timestamp = gethrtime();				\
1488 	_bcp->bc_thread = curthread;					\
1489 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1490 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1491 }
1492 
1493 static void
1494 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1495 	kmem_slab_t *sp, void *addr)
1496 {
1497 	kmem_bufctl_audit_t bca;
1498 
1499 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1500 	bca.bc_addr = addr;
1501 	bca.bc_slab = sp;
1502 	bca.bc_cache = cp;
1503 	KMEM_AUDIT(lp, cp, &bca);
1504 }
1505 
1506 /*
1507  * Create a new slab for cache cp.
1508  */
1509 static kmem_slab_t *
1510 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1511 {
1512 	size_t slabsize = cp->cache_slabsize;
1513 	size_t chunksize = cp->cache_chunksize;
1514 	int cache_flags = cp->cache_flags;
1515 	size_t color, chunks;
1516 	char *buf, *slab;
1517 	kmem_slab_t *sp;
1518 	kmem_bufctl_t *bcp;
1519 	vmem_t *vmp = cp->cache_arena;
1520 
1521 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1522 
1523 	color = cp->cache_color + cp->cache_align;
1524 	if (color > cp->cache_maxcolor)
1525 		color = cp->cache_mincolor;
1526 	cp->cache_color = color;
1527 
1528 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1529 
1530 	if (slab == NULL)
1531 		goto vmem_alloc_failure;
1532 
1533 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1534 
1535 	/*
1536 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1537 	 * consolidator depends (for correctness) on slabs being initialized
1538 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1539 	 * clients to distinguish uninitialized memory from known objects).
1540 	 */
1541 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1542 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1543 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1544 
1545 	if (cache_flags & KMF_HASH) {
1546 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1547 			goto slab_alloc_failure;
1548 		chunks = (slabsize - color) / chunksize;
1549 	} else {
1550 		sp = KMEM_SLAB(cp, slab);
1551 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1552 	}
1553 
1554 	sp->slab_cache	= cp;
1555 	sp->slab_head	= NULL;
1556 	sp->slab_refcnt	= 0;
1557 	sp->slab_base	= buf = slab + color;
1558 	sp->slab_chunks	= chunks;
1559 	sp->slab_stuck_offset = (uint32_t)-1;
1560 	sp->slab_later_count = 0;
1561 	sp->slab_flags = 0;
1562 
1563 	ASSERT(chunks > 0);
1564 	while (chunks-- != 0) {
1565 		if (cache_flags & KMF_HASH) {
1566 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1567 			if (bcp == NULL)
1568 				goto bufctl_alloc_failure;
1569 			if (cache_flags & KMF_AUDIT) {
1570 				kmem_bufctl_audit_t *bcap =
1571 				    (kmem_bufctl_audit_t *)bcp;
1572 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1573 				bcap->bc_cache = cp;
1574 			}
1575 			bcp->bc_addr = buf;
1576 			bcp->bc_slab = sp;
1577 		} else {
1578 			bcp = KMEM_BUFCTL(cp, buf);
1579 		}
1580 		if (cache_flags & KMF_BUFTAG) {
1581 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1582 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1583 			btp->bt_bufctl = bcp;
1584 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1585 			if (cache_flags & KMF_DEADBEEF) {
1586 				copy_pattern(KMEM_FREE_PATTERN, buf,
1587 				    cp->cache_verify);
1588 			}
1589 		}
1590 		bcp->bc_next = sp->slab_head;
1591 		sp->slab_head = bcp;
1592 		buf += chunksize;
1593 	}
1594 
1595 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1596 
1597 	return (sp);
1598 
1599 bufctl_alloc_failure:
1600 
1601 	while ((bcp = sp->slab_head) != NULL) {
1602 		sp->slab_head = bcp->bc_next;
1603 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1604 	}
1605 	kmem_cache_free(kmem_slab_cache, sp);
1606 
1607 slab_alloc_failure:
1608 
1609 	vmem_free(vmp, slab, slabsize);
1610 
1611 vmem_alloc_failure:
1612 
1613 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1614 	atomic_add_64(&cp->cache_alloc_fail, 1);
1615 
1616 	return (NULL);
1617 }
1618 
1619 /*
1620  * Destroy a slab.
1621  */
1622 static void
1623 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1624 {
1625 	vmem_t *vmp = cp->cache_arena;
1626 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1627 
1628 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1629 	ASSERT(sp->slab_refcnt == 0);
1630 
1631 	if (cp->cache_flags & KMF_HASH) {
1632 		kmem_bufctl_t *bcp;
1633 		while ((bcp = sp->slab_head) != NULL) {
1634 			sp->slab_head = bcp->bc_next;
1635 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1636 		}
1637 		kmem_cache_free(kmem_slab_cache, sp);
1638 	}
1639 	vmem_free(vmp, slab, cp->cache_slabsize);
1640 }
1641 
1642 static void *
1643 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp)
1644 {
1645 	kmem_bufctl_t *bcp, **hash_bucket;
1646 	void *buf;
1647 
1648 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1649 	/*
1650 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1651 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1652 	 * slab is newly created (sp->slab_refcnt == 0).
1653 	 */
1654 	ASSERT((sp->slab_refcnt == 0) || (KMEM_SLAB_IS_PARTIAL(sp) &&
1655 	    (sp == avl_first(&cp->cache_partial_slabs))));
1656 	ASSERT(sp->slab_cache == cp);
1657 
1658 	cp->cache_slab_alloc++;
1659 	cp->cache_bufslab--;
1660 	sp->slab_refcnt++;
1661 
1662 	bcp = sp->slab_head;
1663 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1664 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1665 		if (sp->slab_refcnt == 1) {
1666 			ASSERT(sp->slab_chunks == 1);
1667 		} else {
1668 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1669 			avl_remove(&cp->cache_partial_slabs, sp);
1670 			sp->slab_later_count = 0; /* clear history */
1671 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1672 			sp->slab_stuck_offset = (uint32_t)-1;
1673 		}
1674 		list_insert_head(&cp->cache_complete_slabs, sp);
1675 		cp->cache_complete_slab_count++;
1676 	} else {
1677 		ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1678 		if (sp->slab_refcnt == 1) {
1679 			avl_add(&cp->cache_partial_slabs, sp);
1680 		} else {
1681 			/*
1682 			 * The slab is now more allocated than it was, so the
1683 			 * order remains unchanged.
1684 			 */
1685 			ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1686 		}
1687 	}
1688 
1689 	if (cp->cache_flags & KMF_HASH) {
1690 		/*
1691 		 * Add buffer to allocated-address hash table.
1692 		 */
1693 		buf = bcp->bc_addr;
1694 		hash_bucket = KMEM_HASH(cp, buf);
1695 		bcp->bc_next = *hash_bucket;
1696 		*hash_bucket = bcp;
1697 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1698 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1699 		}
1700 	} else {
1701 		buf = KMEM_BUF(cp, bcp);
1702 	}
1703 
1704 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1705 	return (buf);
1706 }
1707 
1708 /*
1709  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1710  */
1711 static void *
1712 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1713 {
1714 	kmem_slab_t *sp;
1715 	void *buf;
1716 	boolean_t test_destructor;
1717 
1718 	mutex_enter(&cp->cache_lock);
1719 	test_destructor = (cp->cache_slab_alloc == 0);
1720 	sp = avl_first(&cp->cache_partial_slabs);
1721 	if (sp == NULL) {
1722 		ASSERT(cp->cache_bufslab == 0);
1723 
1724 		/*
1725 		 * The freelist is empty.  Create a new slab.
1726 		 */
1727 		mutex_exit(&cp->cache_lock);
1728 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1729 			return (NULL);
1730 		}
1731 		mutex_enter(&cp->cache_lock);
1732 		cp->cache_slab_create++;
1733 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1734 			cp->cache_bufmax = cp->cache_buftotal;
1735 		cp->cache_bufslab += sp->slab_chunks;
1736 	}
1737 
1738 	buf = kmem_slab_alloc_impl(cp, sp);
1739 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1740 	    (cp->cache_complete_slab_count +
1741 	    avl_numnodes(&cp->cache_partial_slabs) +
1742 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1743 	mutex_exit(&cp->cache_lock);
1744 
1745 	if (test_destructor && cp->cache_destructor != NULL) {
1746 		/*
1747 		 * On the first kmem_slab_alloc(), assert that it is valid to
1748 		 * call the destructor on a newly constructed object without any
1749 		 * client involvement.
1750 		 */
1751 		if ((cp->cache_constructor == NULL) ||
1752 		    cp->cache_constructor(buf, cp->cache_private,
1753 		    kmflag) == 0) {
1754 			cp->cache_destructor(buf, cp->cache_private);
1755 		}
1756 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1757 		    cp->cache_bufsize);
1758 		if (cp->cache_flags & KMF_DEADBEEF) {
1759 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1760 		}
1761 	}
1762 
1763 	return (buf);
1764 }
1765 
1766 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1767 
1768 /*
1769  * Free a raw (unconstructed) buffer to cp's slab layer.
1770  */
1771 static void
1772 kmem_slab_free(kmem_cache_t *cp, void *buf)
1773 {
1774 	kmem_slab_t *sp;
1775 	kmem_bufctl_t *bcp, **prev_bcpp;
1776 
1777 	ASSERT(buf != NULL);
1778 
1779 	mutex_enter(&cp->cache_lock);
1780 	cp->cache_slab_free++;
1781 
1782 	if (cp->cache_flags & KMF_HASH) {
1783 		/*
1784 		 * Look up buffer in allocated-address hash table.
1785 		 */
1786 		prev_bcpp = KMEM_HASH(cp, buf);
1787 		while ((bcp = *prev_bcpp) != NULL) {
1788 			if (bcp->bc_addr == buf) {
1789 				*prev_bcpp = bcp->bc_next;
1790 				sp = bcp->bc_slab;
1791 				break;
1792 			}
1793 			cp->cache_lookup_depth++;
1794 			prev_bcpp = &bcp->bc_next;
1795 		}
1796 	} else {
1797 		bcp = KMEM_BUFCTL(cp, buf);
1798 		sp = KMEM_SLAB(cp, buf);
1799 	}
1800 
1801 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1802 		mutex_exit(&cp->cache_lock);
1803 		kmem_error(KMERR_BADADDR, cp, buf);
1804 		return;
1805 	}
1806 
1807 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1808 		/*
1809 		 * If this is the buffer that prevented the consolidator from
1810 		 * clearing the slab, we can reset the slab flags now that the
1811 		 * buffer is freed. (It makes sense to do this in
1812 		 * kmem_cache_free(), where the client gives up ownership of the
1813 		 * buffer, but on the hot path the test is too expensive.)
1814 		 */
1815 		kmem_slab_move_yes(cp, sp, buf);
1816 	}
1817 
1818 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1819 		if (cp->cache_flags & KMF_CONTENTS)
1820 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1821 			    kmem_log_enter(kmem_content_log, buf,
1822 			    cp->cache_contents);
1823 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1824 	}
1825 
1826 	bcp->bc_next = sp->slab_head;
1827 	sp->slab_head = bcp;
1828 
1829 	cp->cache_bufslab++;
1830 	ASSERT(sp->slab_refcnt >= 1);
1831 
1832 	if (--sp->slab_refcnt == 0) {
1833 		/*
1834 		 * There are no outstanding allocations from this slab,
1835 		 * so we can reclaim the memory.
1836 		 */
1837 		if (sp->slab_chunks == 1) {
1838 			list_remove(&cp->cache_complete_slabs, sp);
1839 			cp->cache_complete_slab_count--;
1840 		} else {
1841 			avl_remove(&cp->cache_partial_slabs, sp);
1842 		}
1843 
1844 		cp->cache_buftotal -= sp->slab_chunks;
1845 		cp->cache_bufslab -= sp->slab_chunks;
1846 		/*
1847 		 * Defer releasing the slab to the virtual memory subsystem
1848 		 * while there is a pending move callback, since we guarantee
1849 		 * that buffers passed to the move callback have only been
1850 		 * touched by kmem or by the client itself. Since the memory
1851 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1852 		 * set at least one of the two lowest order bits, the client can
1853 		 * test those bits in the move callback to determine whether or
1854 		 * not it knows about the buffer (assuming that the client also
1855 		 * sets one of those low order bits whenever it frees a buffer).
1856 		 */
1857 		if (cp->cache_defrag == NULL ||
1858 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1859 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1860 			cp->cache_slab_destroy++;
1861 			mutex_exit(&cp->cache_lock);
1862 			kmem_slab_destroy(cp, sp);
1863 		} else {
1864 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1865 			/*
1866 			 * Slabs are inserted at both ends of the deadlist to
1867 			 * distinguish between slabs freed while move callbacks
1868 			 * are pending (list head) and a slab freed while the
1869 			 * lock is dropped in kmem_move_buffers() (list tail) so
1870 			 * that in both cases slab_destroy() is called from the
1871 			 * right context.
1872 			 */
1873 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1874 				list_insert_tail(deadlist, sp);
1875 			} else {
1876 				list_insert_head(deadlist, sp);
1877 			}
1878 			cp->cache_defrag->kmd_deadcount++;
1879 			mutex_exit(&cp->cache_lock);
1880 		}
1881 		return;
1882 	}
1883 
1884 	if (bcp->bc_next == NULL) {
1885 		/* Transition the slab from completely allocated to partial. */
1886 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1887 		ASSERT(sp->slab_chunks > 1);
1888 		list_remove(&cp->cache_complete_slabs, sp);
1889 		cp->cache_complete_slab_count--;
1890 		avl_add(&cp->cache_partial_slabs, sp);
1891 	} else {
1892 #ifdef	DEBUG
1893 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1894 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1895 		} else {
1896 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1897 		}
1898 #else
1899 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1900 #endif
1901 	}
1902 
1903 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1904 	    (cp->cache_complete_slab_count +
1905 	    avl_numnodes(&cp->cache_partial_slabs) +
1906 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1907 	mutex_exit(&cp->cache_lock);
1908 }
1909 
1910 /*
1911  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1912  */
1913 static int
1914 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1915     caddr_t caller)
1916 {
1917 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1918 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1919 	uint32_t mtbf;
1920 
1921 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1922 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1923 		return (-1);
1924 	}
1925 
1926 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1927 
1928 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1929 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1930 		return (-1);
1931 	}
1932 
1933 	if (cp->cache_flags & KMF_DEADBEEF) {
1934 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1935 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1936 				kmem_error(KMERR_MODIFIED, cp, buf);
1937 				return (-1);
1938 			}
1939 			if (cp->cache_constructor != NULL)
1940 				*(uint64_t *)buf = btp->bt_redzone;
1941 			else
1942 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1943 		} else {
1944 			construct = 1;
1945 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1946 			    KMEM_UNINITIALIZED_PATTERN, buf,
1947 			    cp->cache_verify)) {
1948 				kmem_error(KMERR_MODIFIED, cp, buf);
1949 				return (-1);
1950 			}
1951 		}
1952 	}
1953 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1954 
1955 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1956 	    gethrtime() % mtbf == 0 &&
1957 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1958 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1959 		if (!construct && cp->cache_destructor != NULL)
1960 			cp->cache_destructor(buf, cp->cache_private);
1961 	} else {
1962 		mtbf = 0;
1963 	}
1964 
1965 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1966 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1967 		atomic_add_64(&cp->cache_alloc_fail, 1);
1968 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1969 		if (cp->cache_flags & KMF_DEADBEEF)
1970 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1971 		kmem_slab_free(cp, buf);
1972 		return (1);
1973 	}
1974 
1975 	if (cp->cache_flags & KMF_AUDIT) {
1976 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1977 	}
1978 
1979 	if ((cp->cache_flags & KMF_LITE) &&
1980 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1981 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1982 	}
1983 
1984 	return (0);
1985 }
1986 
1987 static int
1988 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1989 {
1990 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1991 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1992 	kmem_slab_t *sp;
1993 
1994 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1995 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1996 			kmem_error(KMERR_DUPFREE, cp, buf);
1997 			return (-1);
1998 		}
1999 		sp = kmem_findslab(cp, buf);
2000 		if (sp == NULL || sp->slab_cache != cp)
2001 			kmem_error(KMERR_BADADDR, cp, buf);
2002 		else
2003 			kmem_error(KMERR_REDZONE, cp, buf);
2004 		return (-1);
2005 	}
2006 
2007 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2008 
2009 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2010 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2011 		return (-1);
2012 	}
2013 
2014 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2015 		kmem_error(KMERR_REDZONE, cp, buf);
2016 		return (-1);
2017 	}
2018 
2019 	if (cp->cache_flags & KMF_AUDIT) {
2020 		if (cp->cache_flags & KMF_CONTENTS)
2021 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2022 			    buf, cp->cache_contents);
2023 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2024 	}
2025 
2026 	if ((cp->cache_flags & KMF_LITE) &&
2027 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2028 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2029 	}
2030 
2031 	if (cp->cache_flags & KMF_DEADBEEF) {
2032 		if (cp->cache_flags & KMF_LITE)
2033 			btp->bt_redzone = *(uint64_t *)buf;
2034 		else if (cp->cache_destructor != NULL)
2035 			cp->cache_destructor(buf, cp->cache_private);
2036 
2037 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2038 	}
2039 
2040 	return (0);
2041 }
2042 
2043 /*
2044  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2045  */
2046 static void
2047 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2048 {
2049 	int round;
2050 
2051 	ASSERT(!list_link_active(&cp->cache_link) ||
2052 	    taskq_member(kmem_taskq, curthread));
2053 
2054 	for (round = 0; round < nrounds; round++) {
2055 		void *buf = mp->mag_round[round];
2056 
2057 		if (cp->cache_flags & KMF_DEADBEEF) {
2058 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2059 			    cp->cache_verify) != NULL) {
2060 				kmem_error(KMERR_MODIFIED, cp, buf);
2061 				continue;
2062 			}
2063 			if ((cp->cache_flags & KMF_LITE) &&
2064 			    cp->cache_destructor != NULL) {
2065 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2066 				*(uint64_t *)buf = btp->bt_redzone;
2067 				cp->cache_destructor(buf, cp->cache_private);
2068 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2069 			}
2070 		} else if (cp->cache_destructor != NULL) {
2071 			cp->cache_destructor(buf, cp->cache_private);
2072 		}
2073 
2074 		kmem_slab_free(cp, buf);
2075 	}
2076 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2077 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2078 }
2079 
2080 /*
2081  * Allocate a magazine from the depot.
2082  */
2083 static kmem_magazine_t *
2084 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2085 {
2086 	kmem_magazine_t *mp;
2087 
2088 	/*
2089 	 * If we can't get the depot lock without contention,
2090 	 * update our contention count.  We use the depot
2091 	 * contention rate to determine whether we need to
2092 	 * increase the magazine size for better scalability.
2093 	 */
2094 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2095 		mutex_enter(&cp->cache_depot_lock);
2096 		cp->cache_depot_contention++;
2097 	}
2098 
2099 	if ((mp = mlp->ml_list) != NULL) {
2100 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2101 		mlp->ml_list = mp->mag_next;
2102 		if (--mlp->ml_total < mlp->ml_min)
2103 			mlp->ml_min = mlp->ml_total;
2104 		mlp->ml_alloc++;
2105 	}
2106 
2107 	mutex_exit(&cp->cache_depot_lock);
2108 
2109 	return (mp);
2110 }
2111 
2112 /*
2113  * Free a magazine to the depot.
2114  */
2115 static void
2116 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2117 {
2118 	mutex_enter(&cp->cache_depot_lock);
2119 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2120 	mp->mag_next = mlp->ml_list;
2121 	mlp->ml_list = mp;
2122 	mlp->ml_total++;
2123 	mutex_exit(&cp->cache_depot_lock);
2124 }
2125 
2126 /*
2127  * Update the working set statistics for cp's depot.
2128  */
2129 static void
2130 kmem_depot_ws_update(kmem_cache_t *cp)
2131 {
2132 	mutex_enter(&cp->cache_depot_lock);
2133 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2134 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2135 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2136 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2137 	mutex_exit(&cp->cache_depot_lock);
2138 }
2139 
2140 /*
2141  * Reap all magazines that have fallen out of the depot's working set.
2142  */
2143 static void
2144 kmem_depot_ws_reap(kmem_cache_t *cp)
2145 {
2146 	long reap;
2147 	kmem_magazine_t *mp;
2148 
2149 	ASSERT(!list_link_active(&cp->cache_link) ||
2150 	    taskq_member(kmem_taskq, curthread));
2151 
2152 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2153 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2154 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2155 
2156 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2157 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2158 		kmem_magazine_destroy(cp, mp, 0);
2159 }
2160 
2161 static void
2162 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2163 {
2164 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2165 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2166 	ASSERT(ccp->cc_magsize > 0);
2167 
2168 	ccp->cc_ploaded = ccp->cc_loaded;
2169 	ccp->cc_prounds = ccp->cc_rounds;
2170 	ccp->cc_loaded = mp;
2171 	ccp->cc_rounds = rounds;
2172 }
2173 
2174 /*
2175  * Allocate a constructed object from cache cp.
2176  */
2177 void *
2178 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2179 {
2180 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2181 	kmem_magazine_t *fmp;
2182 	void *buf;
2183 
2184 	mutex_enter(&ccp->cc_lock);
2185 	for (;;) {
2186 		/*
2187 		 * If there's an object available in the current CPU's
2188 		 * loaded magazine, just take it and return.
2189 		 */
2190 		if (ccp->cc_rounds > 0) {
2191 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2192 			ccp->cc_alloc++;
2193 			mutex_exit(&ccp->cc_lock);
2194 			if ((ccp->cc_flags & KMF_BUFTAG) &&
2195 			    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2196 			    caller()) != 0) {
2197 				if (kmflag & KM_NOSLEEP)
2198 					return (NULL);
2199 				mutex_enter(&ccp->cc_lock);
2200 				continue;
2201 			}
2202 			return (buf);
2203 		}
2204 
2205 		/*
2206 		 * The loaded magazine is empty.  If the previously loaded
2207 		 * magazine was full, exchange them and try again.
2208 		 */
2209 		if (ccp->cc_prounds > 0) {
2210 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2211 			continue;
2212 		}
2213 
2214 		/*
2215 		 * If the magazine layer is disabled, break out now.
2216 		 */
2217 		if (ccp->cc_magsize == 0)
2218 			break;
2219 
2220 		/*
2221 		 * Try to get a full magazine from the depot.
2222 		 */
2223 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2224 		if (fmp != NULL) {
2225 			if (ccp->cc_ploaded != NULL)
2226 				kmem_depot_free(cp, &cp->cache_empty,
2227 				    ccp->cc_ploaded);
2228 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2229 			continue;
2230 		}
2231 
2232 		/*
2233 		 * There are no full magazines in the depot,
2234 		 * so fall through to the slab layer.
2235 		 */
2236 		break;
2237 	}
2238 	mutex_exit(&ccp->cc_lock);
2239 
2240 	/*
2241 	 * We couldn't allocate a constructed object from the magazine layer,
2242 	 * so get a raw buffer from the slab layer and apply its constructor.
2243 	 */
2244 	buf = kmem_slab_alloc(cp, kmflag);
2245 
2246 	if (buf == NULL)
2247 		return (NULL);
2248 
2249 	if (cp->cache_flags & KMF_BUFTAG) {
2250 		/*
2251 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2252 		 */
2253 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2254 		if (rc != 0) {
2255 			if (kmflag & KM_NOSLEEP)
2256 				return (NULL);
2257 			/*
2258 			 * kmem_cache_alloc_debug() detected corruption
2259 			 * but didn't panic (kmem_panic <= 0). We should not be
2260 			 * here because the constructor failed (indicated by a
2261 			 * return code of 1). Try again.
2262 			 */
2263 			ASSERT(rc == -1);
2264 			return (kmem_cache_alloc(cp, kmflag));
2265 		}
2266 		return (buf);
2267 	}
2268 
2269 	if (cp->cache_constructor != NULL &&
2270 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2271 		atomic_add_64(&cp->cache_alloc_fail, 1);
2272 		kmem_slab_free(cp, buf);
2273 		return (NULL);
2274 	}
2275 
2276 	return (buf);
2277 }
2278 
2279 /*
2280  * The freed argument tells whether or not kmem_cache_free_debug() has already
2281  * been called so that we can avoid the duplicate free error. For example, a
2282  * buffer on a magazine has already been freed by the client but is still
2283  * constructed.
2284  */
2285 static void
2286 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2287 {
2288 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2289 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2290 			return;
2291 
2292 	/*
2293 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2294 	 * kmem_cache_free_debug() will have already applied the destructor.
2295 	 */
2296 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2297 	    cp->cache_destructor != NULL) {
2298 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2299 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2300 			*(uint64_t *)buf = btp->bt_redzone;
2301 			cp->cache_destructor(buf, cp->cache_private);
2302 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2303 		} else {
2304 			cp->cache_destructor(buf, cp->cache_private);
2305 		}
2306 	}
2307 
2308 	kmem_slab_free(cp, buf);
2309 }
2310 
2311 /*
2312  * Free a constructed object to cache cp.
2313  */
2314 void
2315 kmem_cache_free(kmem_cache_t *cp, void *buf)
2316 {
2317 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2318 	kmem_magazine_t *emp;
2319 	kmem_magtype_t *mtp;
2320 
2321 	/*
2322 	 * The client must not free either of the buffers passed to the move
2323 	 * callback function.
2324 	 */
2325 	ASSERT(cp->cache_defrag == NULL ||
2326 	    cp->cache_defrag->kmd_thread != curthread ||
2327 	    (buf != cp->cache_defrag->kmd_from_buf &&
2328 	    buf != cp->cache_defrag->kmd_to_buf));
2329 
2330 	if (ccp->cc_flags & KMF_BUFTAG)
2331 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2332 			return;
2333 
2334 	mutex_enter(&ccp->cc_lock);
2335 	for (;;) {
2336 		/*
2337 		 * If there's a slot available in the current CPU's
2338 		 * loaded magazine, just put the object there and return.
2339 		 */
2340 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2341 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2342 			ccp->cc_free++;
2343 			mutex_exit(&ccp->cc_lock);
2344 			return;
2345 		}
2346 
2347 		/*
2348 		 * The loaded magazine is full.  If the previously loaded
2349 		 * magazine was empty, exchange them and try again.
2350 		 */
2351 		if (ccp->cc_prounds == 0) {
2352 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2353 			continue;
2354 		}
2355 
2356 		/*
2357 		 * If the magazine layer is disabled, break out now.
2358 		 */
2359 		if (ccp->cc_magsize == 0)
2360 			break;
2361 
2362 		/*
2363 		 * Try to get an empty magazine from the depot.
2364 		 */
2365 		emp = kmem_depot_alloc(cp, &cp->cache_empty);
2366 		if (emp != NULL) {
2367 			if (ccp->cc_ploaded != NULL)
2368 				kmem_depot_free(cp, &cp->cache_full,
2369 				    ccp->cc_ploaded);
2370 			kmem_cpu_reload(ccp, emp, 0);
2371 			continue;
2372 		}
2373 
2374 		/*
2375 		 * There are no empty magazines in the depot,
2376 		 * so try to allocate a new one.  We must drop all locks
2377 		 * across kmem_cache_alloc() because lower layers may
2378 		 * attempt to allocate from this cache.
2379 		 */
2380 		mtp = cp->cache_magtype;
2381 		mutex_exit(&ccp->cc_lock);
2382 		emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2383 		mutex_enter(&ccp->cc_lock);
2384 
2385 		if (emp != NULL) {
2386 			/*
2387 			 * We successfully allocated an empty magazine.
2388 			 * However, we had to drop ccp->cc_lock to do it,
2389 			 * so the cache's magazine size may have changed.
2390 			 * If so, free the magazine and try again.
2391 			 */
2392 			if (ccp->cc_magsize != mtp->mt_magsize) {
2393 				mutex_exit(&ccp->cc_lock);
2394 				kmem_cache_free(mtp->mt_cache, emp);
2395 				mutex_enter(&ccp->cc_lock);
2396 				continue;
2397 			}
2398 
2399 			/*
2400 			 * We got a magazine of the right size.  Add it to
2401 			 * the depot and try the whole dance again.
2402 			 */
2403 			kmem_depot_free(cp, &cp->cache_empty, emp);
2404 			continue;
2405 		}
2406 
2407 		/*
2408 		 * We couldn't allocate an empty magazine,
2409 		 * so fall through to the slab layer.
2410 		 */
2411 		break;
2412 	}
2413 	mutex_exit(&ccp->cc_lock);
2414 
2415 	/*
2416 	 * We couldn't free our constructed object to the magazine layer,
2417 	 * so apply its destructor and free it to the slab layer.
2418 	 */
2419 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2420 }
2421 
2422 void *
2423 kmem_zalloc(size_t size, int kmflag)
2424 {
2425 	size_t index;
2426 	void *buf;
2427 
2428 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2429 		kmem_cache_t *cp = kmem_alloc_table[index];
2430 		buf = kmem_cache_alloc(cp, kmflag);
2431 		if (buf != NULL) {
2432 			if (cp->cache_flags & KMF_BUFTAG) {
2433 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2434 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2435 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2436 
2437 				if (cp->cache_flags & KMF_LITE) {
2438 					KMEM_BUFTAG_LITE_ENTER(btp,
2439 					    kmem_lite_count, caller());
2440 				}
2441 			}
2442 			bzero(buf, size);
2443 		}
2444 	} else {
2445 		buf = kmem_alloc(size, kmflag);
2446 		if (buf != NULL)
2447 			bzero(buf, size);
2448 	}
2449 	return (buf);
2450 }
2451 
2452 void *
2453 kmem_alloc(size_t size, int kmflag)
2454 {
2455 	size_t index;
2456 	kmem_cache_t *cp;
2457 	void *buf;
2458 
2459 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2460 		cp = kmem_alloc_table[index];
2461 		/* fall through to kmem_cache_alloc() */
2462 
2463 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2464 	    kmem_big_alloc_table_max) {
2465 		cp = kmem_big_alloc_table[index];
2466 		/* fall through to kmem_cache_alloc() */
2467 
2468 	} else {
2469 		if (size == 0)
2470 			return (NULL);
2471 
2472 		buf = vmem_alloc(kmem_oversize_arena, size,
2473 		    kmflag & KM_VMFLAGS);
2474 		if (buf == NULL)
2475 			kmem_log_event(kmem_failure_log, NULL, NULL,
2476 			    (void *)size);
2477 		return (buf);
2478 	}
2479 
2480 	buf = kmem_cache_alloc(cp, kmflag);
2481 	if ((cp->cache_flags & KMF_BUFTAG) && buf != NULL) {
2482 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2483 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2484 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2485 
2486 		if (cp->cache_flags & KMF_LITE) {
2487 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2488 		}
2489 	}
2490 	return (buf);
2491 }
2492 
2493 void
2494 kmem_free(void *buf, size_t size)
2495 {
2496 	size_t index;
2497 	kmem_cache_t *cp;
2498 
2499 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2500 		cp = kmem_alloc_table[index];
2501 		/* fall through to kmem_cache_free() */
2502 
2503 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2504 	    kmem_big_alloc_table_max) {
2505 		cp = kmem_big_alloc_table[index];
2506 		/* fall through to kmem_cache_free() */
2507 
2508 	} else {
2509 		if (buf == NULL && size == 0)
2510 			return;
2511 		vmem_free(kmem_oversize_arena, buf, size);
2512 		return;
2513 	}
2514 
2515 	if (cp->cache_flags & KMF_BUFTAG) {
2516 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2517 		uint32_t *ip = (uint32_t *)btp;
2518 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2519 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2520 				kmem_error(KMERR_DUPFREE, cp, buf);
2521 				return;
2522 			}
2523 			if (KMEM_SIZE_VALID(ip[1])) {
2524 				ip[0] = KMEM_SIZE_ENCODE(size);
2525 				kmem_error(KMERR_BADSIZE, cp, buf);
2526 			} else {
2527 				kmem_error(KMERR_REDZONE, cp, buf);
2528 			}
2529 			return;
2530 		}
2531 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2532 			kmem_error(KMERR_REDZONE, cp, buf);
2533 			return;
2534 		}
2535 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
2536 		if (cp->cache_flags & KMF_LITE) {
2537 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2538 			    caller());
2539 		}
2540 	}
2541 	kmem_cache_free(cp, buf);
2542 }
2543 
2544 void *
2545 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2546 {
2547 	size_t realsize = size + vmp->vm_quantum;
2548 	void *addr;
2549 
2550 	/*
2551 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2552 	 * vm_quantum will cause integer wraparound.  Check for this, and
2553 	 * blow off the firewall page in this case.  Note that such a
2554 	 * giant allocation (the entire kernel address space) can never
2555 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
2556 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2557 	 * corresponding check in kmem_firewall_va_free().
2558 	 */
2559 	if (realsize < size)
2560 		realsize = size;
2561 
2562 	/*
2563 	 * While boot still owns resource management, make sure that this
2564 	 * redzone virtual address allocation is properly accounted for in
2565 	 * OBPs "virtual-memory" "available" lists because we're
2566 	 * effectively claiming them for a red zone.  If we don't do this,
2567 	 * the available lists become too fragmented and too large for the
2568 	 * current boot/kernel memory list interface.
2569 	 */
2570 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
2571 
2572 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
2573 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
2574 
2575 	return (addr);
2576 }
2577 
2578 void
2579 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2580 {
2581 	ASSERT((kvseg.s_base == NULL ?
2582 	    va_to_pfn((char *)addr + size) :
2583 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
2584 
2585 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2586 }
2587 
2588 /*
2589  * Try to allocate at least `size' bytes of memory without sleeping or
2590  * panicking. Return actual allocated size in `asize'. If allocation failed,
2591  * try final allocation with sleep or panic allowed.
2592  */
2593 void *
2594 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
2595 {
2596 	void *p;
2597 
2598 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2599 	do {
2600 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
2601 		if (p != NULL)
2602 			return (p);
2603 		*asize += KMEM_ALIGN;
2604 	} while (*asize <= PAGESIZE);
2605 
2606 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2607 	return (kmem_alloc(*asize, kmflag));
2608 }
2609 
2610 /*
2611  * Reclaim all unused memory from a cache.
2612  */
2613 static void
2614 kmem_cache_reap(kmem_cache_t *cp)
2615 {
2616 	ASSERT(taskq_member(kmem_taskq, curthread));
2617 
2618 	/*
2619 	 * Ask the cache's owner to free some memory if possible.
2620 	 * The idea is to handle things like the inode cache, which
2621 	 * typically sits on a bunch of memory that it doesn't truly
2622 	 * *need*.  Reclaim policy is entirely up to the owner; this
2623 	 * callback is just an advisory plea for help.
2624 	 */
2625 	if (cp->cache_reclaim != NULL) {
2626 		long delta;
2627 
2628 		/*
2629 		 * Reclaimed memory should be reapable (not included in the
2630 		 * depot's working set).
2631 		 */
2632 		delta = cp->cache_full.ml_total;
2633 		cp->cache_reclaim(cp->cache_private);
2634 		delta = cp->cache_full.ml_total - delta;
2635 		if (delta > 0) {
2636 			mutex_enter(&cp->cache_depot_lock);
2637 			cp->cache_full.ml_reaplimit += delta;
2638 			cp->cache_full.ml_min += delta;
2639 			mutex_exit(&cp->cache_depot_lock);
2640 		}
2641 	}
2642 
2643 	kmem_depot_ws_reap(cp);
2644 
2645 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
2646 		kmem_cache_defrag(cp);
2647 	}
2648 }
2649 
2650 static void
2651 kmem_reap_timeout(void *flag_arg)
2652 {
2653 	uint32_t *flag = (uint32_t *)flag_arg;
2654 
2655 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
2656 	*flag = 0;
2657 }
2658 
2659 static void
2660 kmem_reap_done(void *flag)
2661 {
2662 	(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
2663 }
2664 
2665 static void
2666 kmem_reap_start(void *flag)
2667 {
2668 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
2669 
2670 	if (flag == &kmem_reaping) {
2671 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
2672 		/*
2673 		 * if we have segkp under heap, reap segkp cache.
2674 		 */
2675 		if (segkp_fromheap)
2676 			segkp_cache_free();
2677 	}
2678 	else
2679 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
2680 
2681 	/*
2682 	 * We use taskq_dispatch() to schedule a timeout to clear
2683 	 * the flag so that kmem_reap() becomes self-throttling:
2684 	 * we won't reap again until the current reap completes *and*
2685 	 * at least kmem_reap_interval ticks have elapsed.
2686 	 */
2687 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
2688 		kmem_reap_done(flag);
2689 }
2690 
2691 static void
2692 kmem_reap_common(void *flag_arg)
2693 {
2694 	uint32_t *flag = (uint32_t *)flag_arg;
2695 
2696 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
2697 	    cas32(flag, 0, 1) != 0)
2698 		return;
2699 
2700 	/*
2701 	 * It may not be kosher to do memory allocation when a reap is called
2702 	 * is called (for example, if vmem_populate() is in the call chain).
2703 	 * So we start the reap going with a TQ_NOALLOC dispatch.  If the
2704 	 * dispatch fails, we reset the flag, and the next reap will try again.
2705 	 */
2706 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
2707 		*flag = 0;
2708 }
2709 
2710 /*
2711  * Reclaim all unused memory from all caches.  Called from the VM system
2712  * when memory gets tight.
2713  */
2714 void
2715 kmem_reap(void)
2716 {
2717 	kmem_reap_common(&kmem_reaping);
2718 }
2719 
2720 /*
2721  * Reclaim all unused memory from identifier arenas, called when a vmem
2722  * arena not back by memory is exhausted.  Since reaping memory-backed caches
2723  * cannot help with identifier exhaustion, we avoid both a large amount of
2724  * work and unwanted side-effects from reclaim callbacks.
2725  */
2726 void
2727 kmem_reap_idspace(void)
2728 {
2729 	kmem_reap_common(&kmem_reaping_idspace);
2730 }
2731 
2732 /*
2733  * Purge all magazines from a cache and set its magazine limit to zero.
2734  * All calls are serialized by the kmem_taskq lock, except for the final
2735  * call from kmem_cache_destroy().
2736  */
2737 static void
2738 kmem_cache_magazine_purge(kmem_cache_t *cp)
2739 {
2740 	kmem_cpu_cache_t *ccp;
2741 	kmem_magazine_t *mp, *pmp;
2742 	int rounds, prounds, cpu_seqid;
2743 
2744 	ASSERT(!list_link_active(&cp->cache_link) ||
2745 	    taskq_member(kmem_taskq, curthread));
2746 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
2747 
2748 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
2749 		ccp = &cp->cache_cpu[cpu_seqid];
2750 
2751 		mutex_enter(&ccp->cc_lock);
2752 		mp = ccp->cc_loaded;
2753 		pmp = ccp->cc_ploaded;
2754 		rounds = ccp->cc_rounds;
2755 		prounds = ccp->cc_prounds;
2756 		ccp->cc_loaded = NULL;
2757 		ccp->cc_ploaded = NULL;
2758 		ccp->cc_rounds = -1;
2759 		ccp->cc_prounds = -1;
2760 		ccp->cc_magsize = 0;
2761 		mutex_exit(&ccp->cc_lock);
2762 
2763 		if (mp)
2764 			kmem_magazine_destroy(cp, mp, rounds);
2765 		if (pmp)
2766 			kmem_magazine_destroy(cp, pmp, prounds);
2767 	}
2768 
2769 	/*
2770 	 * Updating the working set statistics twice in a row has the
2771 	 * effect of setting the working set size to zero, so everything
2772 	 * is eligible for reaping.
2773 	 */
2774 	kmem_depot_ws_update(cp);
2775 	kmem_depot_ws_update(cp);
2776 
2777 	kmem_depot_ws_reap(cp);
2778 }
2779 
2780 /*
2781  * Enable per-cpu magazines on a cache.
2782  */
2783 static void
2784 kmem_cache_magazine_enable(kmem_cache_t *cp)
2785 {
2786 	int cpu_seqid;
2787 
2788 	if (cp->cache_flags & KMF_NOMAGAZINE)
2789 		return;
2790 
2791 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
2792 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2793 		mutex_enter(&ccp->cc_lock);
2794 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2795 		mutex_exit(&ccp->cc_lock);
2796 	}
2797 
2798 }
2799 
2800 /*
2801  * Reap (almost) everything right now.  See kmem_cache_magazine_purge()
2802  * for explanation of the back-to-back kmem_depot_ws_update() calls.
2803  */
2804 void
2805 kmem_cache_reap_now(kmem_cache_t *cp)
2806 {
2807 	ASSERT(list_link_active(&cp->cache_link));
2808 
2809 	kmem_depot_ws_update(cp);
2810 	kmem_depot_ws_update(cp);
2811 
2812 	(void) taskq_dispatch(kmem_taskq,
2813 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
2814 	taskq_wait(kmem_taskq);
2815 }
2816 
2817 /*
2818  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2819  * provide a higher transfer rate with the depot, while smaller magazines
2820  * reduce memory consumption.  Magazine resizing is an expensive operation;
2821  * it should not be done frequently.
2822  *
2823  * Changes to the magazine size are serialized by the kmem_taskq lock.
2824  *
2825  * Note: at present this only grows the magazine size.  It might be useful
2826  * to allow shrinkage too.
2827  */
2828 static void
2829 kmem_cache_magazine_resize(kmem_cache_t *cp)
2830 {
2831 	kmem_magtype_t *mtp = cp->cache_magtype;
2832 
2833 	ASSERT(taskq_member(kmem_taskq, curthread));
2834 
2835 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2836 		kmem_cache_magazine_purge(cp);
2837 		mutex_enter(&cp->cache_depot_lock);
2838 		cp->cache_magtype = ++mtp;
2839 		cp->cache_depot_contention_prev =
2840 		    cp->cache_depot_contention + INT_MAX;
2841 		mutex_exit(&cp->cache_depot_lock);
2842 		kmem_cache_magazine_enable(cp);
2843 	}
2844 }
2845 
2846 /*
2847  * Rescale a cache's hash table, so that the table size is roughly the
2848  * cache size.  We want the average lookup time to be extremely small.
2849  */
2850 static void
2851 kmem_hash_rescale(kmem_cache_t *cp)
2852 {
2853 	kmem_bufctl_t **old_table, **new_table, *bcp;
2854 	size_t old_size, new_size, h;
2855 
2856 	ASSERT(taskq_member(kmem_taskq, curthread));
2857 
2858 	new_size = MAX(KMEM_HASH_INITIAL,
2859 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2860 	old_size = cp->cache_hash_mask + 1;
2861 
2862 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2863 		return;
2864 
2865 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
2866 	    VM_NOSLEEP);
2867 	if (new_table == NULL)
2868 		return;
2869 	bzero(new_table, new_size * sizeof (void *));
2870 
2871 	mutex_enter(&cp->cache_lock);
2872 
2873 	old_size = cp->cache_hash_mask + 1;
2874 	old_table = cp->cache_hash_table;
2875 
2876 	cp->cache_hash_mask = new_size - 1;
2877 	cp->cache_hash_table = new_table;
2878 	cp->cache_rescale++;
2879 
2880 	for (h = 0; h < old_size; h++) {
2881 		bcp = old_table[h];
2882 		while (bcp != NULL) {
2883 			void *addr = bcp->bc_addr;
2884 			kmem_bufctl_t *next_bcp = bcp->bc_next;
2885 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
2886 			bcp->bc_next = *hash_bucket;
2887 			*hash_bucket = bcp;
2888 			bcp = next_bcp;
2889 		}
2890 	}
2891 
2892 	mutex_exit(&cp->cache_lock);
2893 
2894 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
2895 }
2896 
2897 /*
2898  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
2899  * update, magazine resizing, and slab consolidation.
2900  */
2901 static void
2902 kmem_cache_update(kmem_cache_t *cp)
2903 {
2904 	int need_hash_rescale = 0;
2905 	int need_magazine_resize = 0;
2906 
2907 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
2908 
2909 	/*
2910 	 * If the cache has become much larger or smaller than its hash table,
2911 	 * fire off a request to rescale the hash table.
2912 	 */
2913 	mutex_enter(&cp->cache_lock);
2914 
2915 	if ((cp->cache_flags & KMF_HASH) &&
2916 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2917 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2918 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
2919 		need_hash_rescale = 1;
2920 
2921 	mutex_exit(&cp->cache_lock);
2922 
2923 	/*
2924 	 * Update the depot working set statistics.
2925 	 */
2926 	kmem_depot_ws_update(cp);
2927 
2928 	/*
2929 	 * If there's a lot of contention in the depot,
2930 	 * increase the magazine size.
2931 	 */
2932 	mutex_enter(&cp->cache_depot_lock);
2933 
2934 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2935 	    (int)(cp->cache_depot_contention -
2936 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
2937 		need_magazine_resize = 1;
2938 
2939 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2940 
2941 	mutex_exit(&cp->cache_depot_lock);
2942 
2943 	if (need_hash_rescale)
2944 		(void) taskq_dispatch(kmem_taskq,
2945 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
2946 
2947 	if (need_magazine_resize)
2948 		(void) taskq_dispatch(kmem_taskq,
2949 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
2950 
2951 	if (cp->cache_defrag != NULL)
2952 		(void) taskq_dispatch(kmem_taskq,
2953 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
2954 }
2955 
2956 static void kmem_update(void *);
2957 
2958 static void
2959 kmem_update_timeout(void *dummy)
2960 {
2961 
2962 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
2963 }
2964 
2965 static void
2966 kmem_update(void *dummy)
2967 {
2968 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
2969 
2970 	/*
2971 	 * We use taskq_dispatch() to reschedule the timeout so that
2972 	 * kmem_update() becomes self-throttling: it won't schedule
2973 	 * new tasks until all previous tasks have completed.
2974 	 */
2975 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
2976 		kmem_update_timeout(NULL);
2977 }
2978 
2979 static int
2980 kmem_cache_kstat_update(kstat_t *ksp, int rw)
2981 {
2982 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
2983 	kmem_cache_t *cp = ksp->ks_private;
2984 	uint64_t cpu_buf_avail;
2985 	uint64_t buf_avail = 0;
2986 	int cpu_seqid;
2987 
2988 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
2989 
2990 	if (rw == KSTAT_WRITE)
2991 		return (EACCES);
2992 
2993 	mutex_enter(&cp->cache_lock);
2994 
2995 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
2996 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
2997 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
2998 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
2999 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3000 
3001 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3002 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3003 
3004 		mutex_enter(&ccp->cc_lock);
3005 
3006 		cpu_buf_avail = 0;
3007 		if (ccp->cc_rounds > 0)
3008 			cpu_buf_avail += ccp->cc_rounds;
3009 		if (ccp->cc_prounds > 0)
3010 			cpu_buf_avail += ccp->cc_prounds;
3011 
3012 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3013 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3014 		buf_avail			+= cpu_buf_avail;
3015 
3016 		mutex_exit(&ccp->cc_lock);
3017 	}
3018 
3019 	mutex_enter(&cp->cache_depot_lock);
3020 
3021 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3022 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3023 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3024 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3025 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3026 	kmcp->kmc_magazine_size.value.ui64	=
3027 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3028 	    0 : cp->cache_magtype->mt_magsize;
3029 
3030 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3031 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3032 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3033 
3034 	mutex_exit(&cp->cache_depot_lock);
3035 
3036 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3037 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3038 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3039 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3040 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3041 	buf_avail += cp->cache_bufslab;
3042 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3043 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3044 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3045 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3046 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3047 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3048 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3049 	    cp->cache_hash_mask + 1 : 0;
3050 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3051 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3052 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3053 
3054 	if (cp->cache_defrag == NULL) {
3055 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3056 		kmcp->kmc_move_yes.value.ui64		= 0;
3057 		kmcp->kmc_move_no.value.ui64		= 0;
3058 		kmcp->kmc_move_later.value.ui64		= 0;
3059 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3060 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3061 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3062 	} else {
3063 		kmem_defrag_t *kd = cp->cache_defrag;
3064 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3065 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3066 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3067 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3068 		kmcp->kmc_move_dont_need.value.ui64	= kd->kmd_dont_need;
3069 		kmcp->kmc_move_dont_know.value.ui64	= kd->kmd_dont_know;
3070 		kmcp->kmc_move_hunt_found.value.ui64	= kd->kmd_hunt_found;
3071 	}
3072 
3073 	mutex_exit(&cp->cache_lock);
3074 	return (0);
3075 }
3076 
3077 /*
3078  * Return a named statistic about a particular cache.
3079  * This shouldn't be called very often, so it's currently designed for
3080  * simplicity (leverages existing kstat support) rather than efficiency.
3081  */
3082 uint64_t
3083 kmem_cache_stat(kmem_cache_t *cp, char *name)
3084 {
3085 	int i;
3086 	kstat_t *ksp = cp->cache_kstat;
3087 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3088 	uint64_t value = 0;
3089 
3090 	if (ksp != NULL) {
3091 		mutex_enter(&kmem_cache_kstat_lock);
3092 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3093 		for (i = 0; i < ksp->ks_ndata; i++) {
3094 			if (strcmp(knp[i].name, name) == 0) {
3095 				value = knp[i].value.ui64;
3096 				break;
3097 			}
3098 		}
3099 		mutex_exit(&kmem_cache_kstat_lock);
3100 	}
3101 	return (value);
3102 }
3103 
3104 /*
3105  * Return an estimate of currently available kernel heap memory.
3106  * On 32-bit systems, physical memory may exceed virtual memory,
3107  * we just truncate the result at 1GB.
3108  */
3109 size_t
3110 kmem_avail(void)
3111 {
3112 	spgcnt_t rmem = availrmem - tune.t_minarmem;
3113 	spgcnt_t fmem = freemem - minfree;
3114 
3115 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3116 	    1 << (30 - PAGESHIFT))));
3117 }
3118 
3119 /*
3120  * Return the maximum amount of memory that is (in theory) allocatable
3121  * from the heap. This may be used as an estimate only since there
3122  * is no guarentee this space will still be available when an allocation
3123  * request is made, nor that the space may be allocated in one big request
3124  * due to kernel heap fragmentation.
3125  */
3126 size_t
3127 kmem_maxavail(void)
3128 {
3129 	spgcnt_t pmem = availrmem - tune.t_minarmem;
3130 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3131 
3132 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3133 }
3134 
3135 /*
3136  * Indicate whether memory-intensive kmem debugging is enabled.
3137  */
3138 int
3139 kmem_debugging(void)
3140 {
3141 	return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3142 }
3143 
3144 /* binning function, sorts finely at the two extremes */
3145 #define	KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift)				\
3146 	((((sp)->slab_refcnt <= (binshift)) ||				\
3147 	    (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift)))	\
3148 	    ? -(sp)->slab_refcnt					\
3149 	    : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3150 
3151 /*
3152  * Minimizing the number of partial slabs on the freelist minimizes
3153  * fragmentation (the ratio of unused buffers held by the slab layer). There are
3154  * two ways to get a slab off of the freelist: 1) free all the buffers on the
3155  * slab, and 2) allocate all the buffers on the slab. It follows that we want
3156  * the most-used slabs at the front of the list where they have the best chance
3157  * of being completely allocated, and the least-used slabs at a safe distance
3158  * from the front to improve the odds that the few remaining buffers will all be
3159  * freed before another allocation can tie up the slab. For that reason a slab
3160  * with a higher slab_refcnt sorts less than than a slab with a lower
3161  * slab_refcnt.
3162  *
3163  * However, if a slab has at least one buffer that is deemed unfreeable, we
3164  * would rather have that slab at the front of the list regardless of
3165  * slab_refcnt, since even one unfreeable buffer makes the entire slab
3166  * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3167  * callback, the slab is marked unfreeable for as long as it remains on the
3168  * freelist.
3169  */
3170 static int
3171 kmem_partial_slab_cmp(const void *p0, const void *p1)
3172 {
3173 	const kmem_cache_t *cp;
3174 	const kmem_slab_t *s0 = p0;
3175 	const kmem_slab_t *s1 = p1;
3176 	int w0, w1;
3177 	size_t binshift;
3178 
3179 	ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3180 	ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3181 	ASSERT(s0->slab_cache == s1->slab_cache);
3182 	cp = s1->slab_cache;
3183 	ASSERT(MUTEX_HELD(&cp->cache_lock));
3184 	binshift = cp->cache_partial_binshift;
3185 
3186 	/* weight of first slab */
3187 	w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3188 	if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3189 		w0 -= cp->cache_maxchunks;
3190 	}
3191 
3192 	/* weight of second slab */
3193 	w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3194 	if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3195 		w1 -= cp->cache_maxchunks;
3196 	}
3197 
3198 	if (w0 < w1)
3199 		return (-1);
3200 	if (w0 > w1)
3201 		return (1);
3202 
3203 	/* compare pointer values */
3204 	if ((uintptr_t)s0 < (uintptr_t)s1)
3205 		return (-1);
3206 	if ((uintptr_t)s0 > (uintptr_t)s1)
3207 		return (1);
3208 
3209 	return (0);
3210 }
3211 
3212 /*
3213  * It must be valid to call the destructor (if any) on a newly created object.
3214  * That is, the constructor (if any) must leave the object in a valid state for
3215  * the destructor.
3216  */
3217 kmem_cache_t *
3218 kmem_cache_create(
3219 	char *name,		/* descriptive name for this cache */
3220 	size_t bufsize,		/* size of the objects it manages */
3221 	size_t align,		/* required object alignment */
3222 	int (*constructor)(void *, void *, int), /* object constructor */
3223 	void (*destructor)(void *, void *),	/* object destructor */
3224 	void (*reclaim)(void *), /* memory reclaim callback */
3225 	void *private,		/* pass-thru arg for constr/destr/reclaim */
3226 	vmem_t *vmp,		/* vmem source for slab allocation */
3227 	int cflags)		/* cache creation flags */
3228 {
3229 	int cpu_seqid;
3230 	size_t chunksize;
3231 	kmem_cache_t *cp;
3232 	kmem_magtype_t *mtp;
3233 	size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3234 
3235 #ifdef	DEBUG
3236 	/*
3237 	 * Cache names should conform to the rules for valid C identifiers
3238 	 */
3239 	if (!strident_valid(name)) {
3240 		cmn_err(CE_CONT,
3241 		    "kmem_cache_create: '%s' is an invalid cache name\n"
3242 		    "cache names must conform to the rules for "
3243 		    "C identifiers\n", name);
3244 	}
3245 #endif	/* DEBUG */
3246 
3247 	if (vmp == NULL)
3248 		vmp = kmem_default_arena;
3249 
3250 	/*
3251 	 * If this kmem cache has an identifier vmem arena as its source, mark
3252 	 * it such to allow kmem_reap_idspace().
3253 	 */
3254 	ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
3255 	if (vmp->vm_cflags & VMC_IDENTIFIER)
3256 		cflags |= KMC_IDENTIFIER;
3257 
3258 	/*
3259 	 * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
3260 	 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3261 	 * false sharing of per-CPU data.
3262 	 */
3263 	cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3264 	    P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3265 	bzero(cp, csize);
3266 	list_link_init(&cp->cache_link);
3267 
3268 	if (align == 0)
3269 		align = KMEM_ALIGN;
3270 
3271 	/*
3272 	 * If we're not at least KMEM_ALIGN aligned, we can't use free
3273 	 * memory to hold bufctl information (because we can't safely
3274 	 * perform word loads and stores on it).
3275 	 */
3276 	if (align < KMEM_ALIGN)
3277 		cflags |= KMC_NOTOUCH;
3278 
3279 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum)
3280 		panic("kmem_cache_create: bad alignment %lu", align);
3281 
3282 	mutex_enter(&kmem_flags_lock);
3283 	if (kmem_flags & KMF_RANDOMIZE)
3284 		kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3285 		    KMF_RANDOMIZE;
3286 	cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3287 	mutex_exit(&kmem_flags_lock);
3288 
3289 	/*
3290 	 * Make sure all the various flags are reasonable.
3291 	 */
3292 	ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3293 
3294 	if (cp->cache_flags & KMF_LITE) {
3295 		if (bufsize >= kmem_lite_minsize &&
3296 		    align <= kmem_lite_maxalign &&
3297 		    P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3298 			cp->cache_flags |= KMF_BUFTAG;
3299 			cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3300 		} else {
3301 			cp->cache_flags &= ~KMF_DEBUG;
3302 		}
3303 	}
3304 
3305 	if (cp->cache_flags & KMF_DEADBEEF)
3306 		cp->cache_flags |= KMF_REDZONE;
3307 
3308 	if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3309 		cp->cache_flags |= KMF_NOMAGAZINE;
3310 
3311 	if (cflags & KMC_NODEBUG)
3312 		cp->cache_flags &= ~KMF_DEBUG;
3313 
3314 	if (cflags & KMC_NOTOUCH)
3315 		cp->cache_flags &= ~KMF_TOUCH;
3316 
3317 	if (cflags & KMC_NOHASH)
3318 		cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3319 
3320 	if (cflags & KMC_NOMAGAZINE)
3321 		cp->cache_flags |= KMF_NOMAGAZINE;
3322 
3323 	if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3324 		cp->cache_flags |= KMF_REDZONE;
3325 
3326 	if (!(cp->cache_flags & KMF_AUDIT))
3327 		cp->cache_flags &= ~KMF_CONTENTS;
3328 
3329 	if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3330 	    !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3331 		cp->cache_flags |= KMF_FIREWALL;
3332 
3333 	if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3334 		cp->cache_flags &= ~KMF_FIREWALL;
3335 
3336 	if (cp->cache_flags & KMF_FIREWALL) {
3337 		cp->cache_flags &= ~KMF_BUFTAG;
3338 		cp->cache_flags |= KMF_NOMAGAZINE;
3339 		ASSERT(vmp == kmem_default_arena);
3340 		vmp = kmem_firewall_arena;
3341 	}
3342 
3343 	/*
3344 	 * Set cache properties.
3345 	 */
3346 	(void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3347 	strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3348 	cp->cache_bufsize = bufsize;
3349 	cp->cache_align = align;
3350 	cp->cache_constructor = constructor;
3351 	cp->cache_destructor = destructor;
3352 	cp->cache_reclaim = reclaim;
3353 	cp->cache_private = private;
3354 	cp->cache_arena = vmp;
3355 	cp->cache_cflags = cflags;
3356 
3357 	/*
3358 	 * Determine the chunk size.
3359 	 */
3360 	chunksize = bufsize;
3361 
3362 	if (align >= KMEM_ALIGN) {
3363 		chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3364 		cp->cache_bufctl = chunksize - KMEM_ALIGN;
3365 	}
3366 
3367 	if (cp->cache_flags & KMF_BUFTAG) {
3368 		cp->cache_bufctl = chunksize;
3369 		cp->cache_buftag = chunksize;
3370 		if (cp->cache_flags & KMF_LITE)
3371 			chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3372 		else
3373 			chunksize += sizeof (kmem_buftag_t);
3374 	}
3375 
3376 	if (cp->cache_flags & KMF_DEADBEEF) {
3377 		cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3378 		if (cp->cache_flags & KMF_LITE)
3379 			cp->cache_verify = sizeof (uint64_t);
3380 	}
3381 
3382 	cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3383 
3384 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3385 
3386 	/*
3387 	 * Now that we know the chunk size, determine the optimal slab size.
3388 	 */
3389 	if (vmp == kmem_firewall_arena) {
3390 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3391 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
3392 		cp->cache_maxcolor = cp->cache_mincolor;
3393 		cp->cache_flags |= KMF_HASH;
3394 		ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3395 	} else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3396 	    !(cp->cache_flags & KMF_AUDIT) &&
3397 	    chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3398 		cp->cache_slabsize = vmp->vm_quantum;
3399 		cp->cache_mincolor = 0;
3400 		cp->cache_maxcolor =
3401 		    (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3402 		ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3403 		ASSERT(!(cp->cache_flags & KMF_AUDIT));
3404 	} else {
3405 		size_t chunks, bestfit, waste, slabsize;
3406 		size_t minwaste = LONG_MAX;
3407 
3408 		for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3409 			slabsize = P2ROUNDUP(chunksize * chunks,
3410 			    vmp->vm_quantum);
3411 			chunks = slabsize / chunksize;
3412 			waste = (slabsize % chunksize) / chunks;
3413 			if (waste < minwaste) {
3414 				minwaste = waste;
3415 				bestfit = slabsize;
3416 			}
3417 		}
3418 		if (cflags & KMC_QCACHE)
3419 			bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3420 		cp->cache_slabsize = bestfit;
3421 		cp->cache_mincolor = 0;
3422 		cp->cache_maxcolor = bestfit % chunksize;
3423 		cp->cache_flags |= KMF_HASH;
3424 	}
3425 
3426 	cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3427 	cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3428 
3429 	if (cp->cache_flags & KMF_HASH) {
3430 		ASSERT(!(cflags & KMC_NOHASH));
3431 		cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3432 		    kmem_bufctl_audit_cache : kmem_bufctl_cache;
3433 	}
3434 
3435 	if (cp->cache_maxcolor >= vmp->vm_quantum)
3436 		cp->cache_maxcolor = vmp->vm_quantum - 1;
3437 
3438 	cp->cache_color = cp->cache_mincolor;
3439 
3440 	/*
3441 	 * Initialize the rest of the slab layer.
3442 	 */
3443 	mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3444 
3445 	avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3446 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3447 	/* LINTED: E_TRUE_LOGICAL_EXPR */
3448 	ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3449 	/* reuse partial slab AVL linkage for complete slab list linkage */
3450 	list_create(&cp->cache_complete_slabs,
3451 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3452 
3453 	if (cp->cache_flags & KMF_HASH) {
3454 		cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3455 		    KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3456 		bzero(cp->cache_hash_table,
3457 		    KMEM_HASH_INITIAL * sizeof (void *));
3458 		cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3459 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3460 	}
3461 
3462 	/*
3463 	 * Initialize the depot.
3464 	 */
3465 	mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3466 
3467 	for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3468 		continue;
3469 
3470 	cp->cache_magtype = mtp;
3471 
3472 	/*
3473 	 * Initialize the CPU layer.
3474 	 */
3475 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3476 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3477 		mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3478 		ccp->cc_flags = cp->cache_flags;
3479 		ccp->cc_rounds = -1;
3480 		ccp->cc_prounds = -1;
3481 	}
3482 
3483 	/*
3484 	 * Create the cache's kstats.
3485 	 */
3486 	if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
3487 	    "kmem_cache", KSTAT_TYPE_NAMED,
3488 	    sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
3489 	    KSTAT_FLAG_VIRTUAL)) != NULL) {
3490 		cp->cache_kstat->ks_data = &kmem_cache_kstat;
3491 		cp->cache_kstat->ks_update = kmem_cache_kstat_update;
3492 		cp->cache_kstat->ks_private = cp;
3493 		cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
3494 		kstat_install(cp->cache_kstat);
3495 	}
3496 
3497 	/*
3498 	 * Add the cache to the global list.  This makes it visible
3499 	 * to kmem_update(), so the cache must be ready for business.
3500 	 */
3501 	mutex_enter(&kmem_cache_lock);
3502 	list_insert_tail(&kmem_caches, cp);
3503 	mutex_exit(&kmem_cache_lock);
3504 
3505 	if (kmem_ready)
3506 		kmem_cache_magazine_enable(cp);
3507 
3508 	return (cp);
3509 }
3510 
3511 static int
3512 kmem_move_cmp(const void *buf, const void *p)
3513 {
3514 	const kmem_move_t *