xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision 686031ed)
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_reap; /* number of kmem_cache_reap() calls */
865 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
866 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
867 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
868 	kstat_named_t	kmc_move_yes;
869 	kstat_named_t	kmc_move_no;
870 	kstat_named_t	kmc_move_later;
871 	kstat_named_t	kmc_move_dont_need;
872 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
873 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
874 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
875 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
876 } kmem_cache_kstat = {
877 	{ "buf_size",		KSTAT_DATA_UINT64 },
878 	{ "align",		KSTAT_DATA_UINT64 },
879 	{ "chunk_size",		KSTAT_DATA_UINT64 },
880 	{ "slab_size",		KSTAT_DATA_UINT64 },
881 	{ "alloc",		KSTAT_DATA_UINT64 },
882 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
883 	{ "free",		KSTAT_DATA_UINT64 },
884 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
885 	{ "depot_free",		KSTAT_DATA_UINT64 },
886 	{ "depot_contention",	KSTAT_DATA_UINT64 },
887 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
888 	{ "slab_free",		KSTAT_DATA_UINT64 },
889 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
890 	{ "buf_avail",		KSTAT_DATA_UINT64 },
891 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
892 	{ "buf_total",		KSTAT_DATA_UINT64 },
893 	{ "buf_max",		KSTAT_DATA_UINT64 },
894 	{ "slab_create",	KSTAT_DATA_UINT64 },
895 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
896 	{ "vmem_source",	KSTAT_DATA_UINT64 },
897 	{ "hash_size",		KSTAT_DATA_UINT64 },
898 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
899 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
900 	{ "full_magazines",	KSTAT_DATA_UINT64 },
901 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
902 	{ "magazine_size",	KSTAT_DATA_UINT64 },
903 	{ "reap",		KSTAT_DATA_UINT64 },
904 	{ "defrag",		KSTAT_DATA_UINT64 },
905 	{ "scan",		KSTAT_DATA_UINT64 },
906 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
907 	{ "move_yes",		KSTAT_DATA_UINT64 },
908 	{ "move_no",		KSTAT_DATA_UINT64 },
909 	{ "move_later",		KSTAT_DATA_UINT64 },
910 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
911 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
912 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
913 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
914 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
915 };
916 
917 static kmutex_t kmem_cache_kstat_lock;
918 
919 /*
920  * The default set of caches to back kmem_alloc().
921  * These sizes should be reevaluated periodically.
922  *
923  * We want allocations that are multiples of the coherency granularity
924  * (64 bytes) to be satisfied from a cache which is a multiple of 64
925  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
926  * the next kmem_cache_size greater than or equal to it must be a
927  * multiple of 64.
928  *
929  * We split the table into two sections:  size <= 4k and size > 4k.  This
930  * saves a lot of space and cache footprint in our cache tables.
931  */
932 static const int kmem_alloc_sizes[] = {
933 	1 * 8,
934 	2 * 8,
935 	3 * 8,
936 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
937 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
938 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
939 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
940 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
941 	P2ALIGN(8192 / 7, 64),
942 	P2ALIGN(8192 / 6, 64),
943 	P2ALIGN(8192 / 5, 64),
944 	P2ALIGN(8192 / 4, 64),
945 	P2ALIGN(8192 / 3, 64),
946 	P2ALIGN(8192 / 2, 64),
947 };
948 
949 static const int kmem_big_alloc_sizes[] = {
950 	2 * 4096,	3 * 4096,
951 	2 * 8192,	3 * 8192,
952 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
953 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
954 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
955 	16 * 8192
956 };
957 
958 #define	KMEM_MAXBUF		4096
959 #define	KMEM_BIG_MAXBUF_32BIT	32768
960 #define	KMEM_BIG_MAXBUF		131072
961 
962 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
963 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
964 
965 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
966 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
967 
968 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
969 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
970 
971 static kmem_magtype_t kmem_magtype[] = {
972 	{ 1,	8,	3200,	65536	},
973 	{ 3,	16,	256,	32768	},
974 	{ 7,	32,	64,	16384	},
975 	{ 15,	64,	0,	8192	},
976 	{ 31,	64,	0,	4096	},
977 	{ 47,	64,	0,	2048	},
978 	{ 63,	64,	0,	1024	},
979 	{ 95,	64,	0,	512	},
980 	{ 143,	64,	0,	0	},
981 };
982 
983 static uint32_t kmem_reaping;
984 static uint32_t kmem_reaping_idspace;
985 
986 /*
987  * kmem tunables
988  */
989 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
990 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
991 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
992 int kmem_panic = 1;		/* whether to panic on error */
993 int kmem_logging = 1;		/* kmem_log_enter() override */
994 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
995 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
996 size_t kmem_content_log_size;	/* content log size [2% of memory] */
997 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
998 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
999 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1000 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1001 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1002 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1003 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1004 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1005 
1006 #ifdef _LP64
1007 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1008 #else
1009 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1010 #endif
1011 
1012 #ifdef DEBUG
1013 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1014 #else
1015 int kmem_flags = 0;
1016 #endif
1017 int kmem_ready;
1018 
1019 static kmem_cache_t	*kmem_slab_cache;
1020 static kmem_cache_t	*kmem_bufctl_cache;
1021 static kmem_cache_t	*kmem_bufctl_audit_cache;
1022 
1023 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1024 static list_t		kmem_caches;
1025 
1026 static taskq_t		*kmem_taskq;
1027 static kmutex_t		kmem_flags_lock;
1028 static vmem_t		*kmem_metadata_arena;
1029 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1030 static vmem_t		*kmem_cache_arena;
1031 static vmem_t		*kmem_hash_arena;
1032 static vmem_t		*kmem_log_arena;
1033 static vmem_t		*kmem_oversize_arena;
1034 static vmem_t		*kmem_va_arena;
1035 static vmem_t		*kmem_default_arena;
1036 static vmem_t		*kmem_firewall_va_arena;
1037 static vmem_t		*kmem_firewall_arena;
1038 
1039 /*
1040  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1041  * turned on when DEBUG is also defined.
1042  */
1043 #ifdef	DEBUG
1044 #define	KMEM_STATS
1045 #endif	/* DEBUG */
1046 
1047 #ifdef	KMEM_STATS
1048 #define	KMEM_STAT_ADD(stat)			((stat)++)
1049 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1050 #else
1051 #define	KMEM_STAT_ADD(stat)			/* nothing */
1052 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1053 #endif	/* KMEM_STATS */
1054 
1055 /*
1056  * kmem slab consolidator thresholds (tunables)
1057  */
1058 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1059 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1060 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1061 /*
1062  * Maximum number of slabs from which to move buffers during a single
1063  * maintenance interval while the system is not low on memory.
1064  */
1065 size_t kmem_reclaim_max_slabs = 1;
1066 /*
1067  * Number of slabs to scan backwards from the end of the partial slab list
1068  * when searching for buffers to relocate.
1069  */
1070 size_t kmem_reclaim_scan_range = 12;
1071 
1072 #ifdef	KMEM_STATS
1073 static struct {
1074 	uint64_t kms_callbacks;
1075 	uint64_t kms_yes;
1076 	uint64_t kms_no;
1077 	uint64_t kms_later;
1078 	uint64_t kms_dont_need;
1079 	uint64_t kms_dont_know;
1080 	uint64_t kms_hunt_found_mag;
1081 	uint64_t kms_hunt_found_slab;
1082 	uint64_t kms_hunt_alloc_fail;
1083 	uint64_t kms_hunt_lucky;
1084 	uint64_t kms_notify;
1085 	uint64_t kms_notify_callbacks;
1086 	uint64_t kms_disbelief;
1087 	uint64_t kms_already_pending;
1088 	uint64_t kms_callback_alloc_fail;
1089 	uint64_t kms_callback_taskq_fail;
1090 	uint64_t kms_endscan_slab_dead;
1091 	uint64_t kms_endscan_slab_destroyed;
1092 	uint64_t kms_endscan_nomem;
1093 	uint64_t kms_endscan_refcnt_changed;
1094 	uint64_t kms_endscan_nomove_changed;
1095 	uint64_t kms_endscan_freelist;
1096 	uint64_t kms_avl_update;
1097 	uint64_t kms_avl_noupdate;
1098 	uint64_t kms_no_longer_reclaimable;
1099 	uint64_t kms_notify_no_longer_reclaimable;
1100 	uint64_t kms_notify_slab_dead;
1101 	uint64_t kms_notify_slab_destroyed;
1102 	uint64_t kms_alloc_fail;
1103 	uint64_t kms_constructor_fail;
1104 	uint64_t kms_dead_slabs_freed;
1105 	uint64_t kms_defrags;
1106 	uint64_t kms_scans;
1107 	uint64_t kms_scan_depot_ws_reaps;
1108 	uint64_t kms_debug_reaps;
1109 	uint64_t kms_debug_scans;
1110 } kmem_move_stats;
1111 #endif	/* KMEM_STATS */
1112 
1113 /* consolidator knobs */
1114 static boolean_t kmem_move_noreap;
1115 static boolean_t kmem_move_blocked;
1116 static boolean_t kmem_move_fulltilt;
1117 static boolean_t kmem_move_any_partial;
1118 
1119 #ifdef	DEBUG
1120 /*
1121  * kmem consolidator debug tunables:
1122  * Ensure code coverage by occasionally running the consolidator even when the
1123  * caches are not fragmented (they may never be). These intervals are mean time
1124  * in cache maintenance intervals (kmem_cache_update).
1125  */
1126 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1127 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1128 #endif	/* DEBUG */
1129 
1130 static kmem_cache_t	*kmem_defrag_cache;
1131 static kmem_cache_t	*kmem_move_cache;
1132 static taskq_t		*kmem_move_taskq;
1133 
1134 static void kmem_cache_scan(kmem_cache_t *);
1135 static void kmem_cache_defrag(kmem_cache_t *);
1136 
1137 
1138 kmem_log_header_t	*kmem_transaction_log;
1139 kmem_log_header_t	*kmem_content_log;
1140 kmem_log_header_t	*kmem_failure_log;
1141 kmem_log_header_t	*kmem_slab_log;
1142 
1143 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1144 
1145 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1146 	if ((count) > 0) {						\
1147 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1148 		pc_t *_e;						\
1149 		/* memmove() the old entries down one notch */		\
1150 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1151 			*_e = *(_e - 1);				\
1152 		*_s = (uintptr_t)(caller);				\
1153 	}
1154 
1155 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1156 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1157 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1158 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1159 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1160 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1161 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1162 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1163 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1164 
1165 struct {
1166 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1167 	int		kmp_error;	/* type of kmem error */
1168 	void		*kmp_buffer;	/* buffer that induced panic */
1169 	void		*kmp_realbuf;	/* real start address for buffer */
1170 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1171 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1172 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1173 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1174 } kmem_panic_info;
1175 
1176 
1177 static void
1178 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1179 {
1180 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1181 	uint64_t *buf = buf_arg;
1182 
1183 	while (buf < bufend)
1184 		*buf++ = pattern;
1185 }
1186 
1187 static void *
1188 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1189 {
1190 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1191 	uint64_t *buf;
1192 
1193 	for (buf = buf_arg; buf < bufend; buf++)
1194 		if (*buf != pattern)
1195 			return (buf);
1196 	return (NULL);
1197 }
1198 
1199 static void *
1200 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1201 {
1202 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1203 	uint64_t *buf;
1204 
1205 	for (buf = buf_arg; buf < bufend; buf++) {
1206 		if (*buf != old) {
1207 			copy_pattern(old, buf_arg,
1208 			    (char *)buf - (char *)buf_arg);
1209 			return (buf);
1210 		}
1211 		*buf = new;
1212 	}
1213 
1214 	return (NULL);
1215 }
1216 
1217 static void
1218 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1219 {
1220 	kmem_cache_t *cp;
1221 
1222 	mutex_enter(&kmem_cache_lock);
1223 	for (cp = list_head(&kmem_caches); cp != NULL;
1224 	    cp = list_next(&kmem_caches, cp))
1225 		if (tq != NULL)
1226 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1227 			    tqflag);
1228 		else
1229 			func(cp);
1230 	mutex_exit(&kmem_cache_lock);
1231 }
1232 
1233 static void
1234 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1235 {
1236 	kmem_cache_t *cp;
1237 
1238 	mutex_enter(&kmem_cache_lock);
1239 	for (cp = list_head(&kmem_caches); cp != NULL;
1240 	    cp = list_next(&kmem_caches, cp)) {
1241 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1242 			continue;
1243 		if (tq != NULL)
1244 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1245 			    tqflag);
1246 		else
1247 			func(cp);
1248 	}
1249 	mutex_exit(&kmem_cache_lock);
1250 }
1251 
1252 /*
1253  * Debugging support.  Given a buffer address, find its slab.
1254  */
1255 static kmem_slab_t *
1256 kmem_findslab(kmem_cache_t *cp, void *buf)
1257 {
1258 	kmem_slab_t *sp;
1259 
1260 	mutex_enter(&cp->cache_lock);
1261 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1262 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1263 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1264 			mutex_exit(&cp->cache_lock);
1265 			return (sp);
1266 		}
1267 	}
1268 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1269 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1270 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1271 			mutex_exit(&cp->cache_lock);
1272 			return (sp);
1273 		}
1274 	}
1275 	mutex_exit(&cp->cache_lock);
1276 
1277 	return (NULL);
1278 }
1279 
1280 static void
1281 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1282 {
1283 	kmem_buftag_t *btp = NULL;
1284 	kmem_bufctl_t *bcp = NULL;
1285 	kmem_cache_t *cp = cparg;
1286 	kmem_slab_t *sp;
1287 	uint64_t *off;
1288 	void *buf = bufarg;
1289 
1290 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1291 
1292 	kmem_panic_info.kmp_timestamp = gethrtime();
1293 
1294 	sp = kmem_findslab(cp, buf);
1295 	if (sp == NULL) {
1296 		for (cp = list_tail(&kmem_caches); cp != NULL;
1297 		    cp = list_prev(&kmem_caches, cp)) {
1298 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1299 				break;
1300 		}
1301 	}
1302 
1303 	if (sp == NULL) {
1304 		cp = NULL;
1305 		error = KMERR_BADADDR;
1306 	} else {
1307 		if (cp != cparg)
1308 			error = KMERR_BADCACHE;
1309 		else
1310 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1311 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1312 		if (buf != bufarg)
1313 			error = KMERR_BADBASE;
1314 		if (cp->cache_flags & KMF_BUFTAG)
1315 			btp = KMEM_BUFTAG(cp, buf);
1316 		if (cp->cache_flags & KMF_HASH) {
1317 			mutex_enter(&cp->cache_lock);
1318 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1319 				if (bcp->bc_addr == buf)
1320 					break;
1321 			mutex_exit(&cp->cache_lock);
1322 			if (bcp == NULL && btp != NULL)
1323 				bcp = btp->bt_bufctl;
1324 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1325 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1326 			    bcp->bc_addr != buf) {
1327 				error = KMERR_BADBUFCTL;
1328 				bcp = NULL;
1329 			}
1330 		}
1331 	}
1332 
1333 	kmem_panic_info.kmp_error = error;
1334 	kmem_panic_info.kmp_buffer = bufarg;
1335 	kmem_panic_info.kmp_realbuf = buf;
1336 	kmem_panic_info.kmp_cache = cparg;
1337 	kmem_panic_info.kmp_realcache = cp;
1338 	kmem_panic_info.kmp_slab = sp;
1339 	kmem_panic_info.kmp_bufctl = bcp;
1340 
1341 	printf("kernel memory allocator: ");
1342 
1343 	switch (error) {
1344 
1345 	case KMERR_MODIFIED:
1346 		printf("buffer modified after being freed\n");
1347 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1348 		if (off == NULL)	/* shouldn't happen */
1349 			off = buf;
1350 		printf("modification occurred at offset 0x%lx "
1351 		    "(0x%llx replaced by 0x%llx)\n",
1352 		    (uintptr_t)off - (uintptr_t)buf,
1353 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1354 		break;
1355 
1356 	case KMERR_REDZONE:
1357 		printf("redzone violation: write past end of buffer\n");
1358 		break;
1359 
1360 	case KMERR_BADADDR:
1361 		printf("invalid free: buffer not in cache\n");
1362 		break;
1363 
1364 	case KMERR_DUPFREE:
1365 		printf("duplicate free: buffer freed twice\n");
1366 		break;
1367 
1368 	case KMERR_BADBUFTAG:
1369 		printf("boundary tag corrupted\n");
1370 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1371 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1372 		    KMEM_BUFTAG_FREE);
1373 		break;
1374 
1375 	case KMERR_BADBUFCTL:
1376 		printf("bufctl corrupted\n");
1377 		break;
1378 
1379 	case KMERR_BADCACHE:
1380 		printf("buffer freed to wrong cache\n");
1381 		printf("buffer was allocated from %s,\n", cp->cache_name);
1382 		printf("caller attempting free to %s.\n", cparg->cache_name);
1383 		break;
1384 
1385 	case KMERR_BADSIZE:
1386 		printf("bad free: free size (%u) != alloc size (%u)\n",
1387 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1388 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1389 		break;
1390 
1391 	case KMERR_BADBASE:
1392 		printf("bad free: free address (%p) != alloc address (%p)\n",
1393 		    bufarg, buf);
1394 		break;
1395 	}
1396 
1397 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1398 	    bufarg, (void *)bcp, cparg->cache_name);
1399 
1400 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1401 	    error != KMERR_BADBUFCTL) {
1402 		int d;
1403 		timestruc_t ts;
1404 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1405 
1406 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1407 		printf("previous transaction on buffer %p:\n", buf);
1408 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1409 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1410 		    (void *)sp, cp->cache_name);
1411 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1412 			ulong_t off;
1413 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1414 			printf("%s+%lx\n", sym ? sym : "?", off);
1415 		}
1416 	}
1417 	if (kmem_panic > 0)
1418 		panic("kernel heap corruption detected");
1419 	if (kmem_panic == 0)
1420 		debug_enter(NULL);
1421 	kmem_logging = 1;	/* resume logging */
1422 }
1423 
1424 static kmem_log_header_t *
1425 kmem_log_init(size_t logsize)
1426 {
1427 	kmem_log_header_t *lhp;
1428 	int nchunks = 4 * max_ncpus;
1429 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1430 	int i;
1431 
1432 	/*
1433 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1434 	 * to prevent false sharing of cache lines.
1435 	 */
1436 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1437 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1438 	    NULL, NULL, VM_SLEEP);
1439 	bzero(lhp, lhsize);
1440 
1441 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1442 	lhp->lh_nchunks = nchunks;
1443 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1444 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1445 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1446 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1447 	    nchunks * sizeof (int), VM_SLEEP);
1448 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1449 
1450 	for (i = 0; i < max_ncpus; i++) {
1451 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1452 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1453 		clhp->clh_chunk = i;
1454 	}
1455 
1456 	for (i = max_ncpus; i < nchunks; i++)
1457 		lhp->lh_free[i] = i;
1458 
1459 	lhp->lh_head = max_ncpus;
1460 	lhp->lh_tail = 0;
1461 
1462 	return (lhp);
1463 }
1464 
1465 static void *
1466 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1467 {
1468 	void *logspace;
1469 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1470 
1471 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1472 		return (NULL);
1473 
1474 	mutex_enter(&clhp->clh_lock);
1475 	clhp->clh_hits++;
1476 	if (size > clhp->clh_avail) {
1477 		mutex_enter(&lhp->lh_lock);
1478 		lhp->lh_hits++;
1479 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1480 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1481 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1482 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1483 		clhp->clh_current = lhp->lh_base +
1484 		    clhp->clh_chunk * lhp->lh_chunksize;
1485 		clhp->clh_avail = lhp->lh_chunksize;
1486 		if (size > lhp->lh_chunksize)
1487 			size = lhp->lh_chunksize;
1488 		mutex_exit(&lhp->lh_lock);
1489 	}
1490 	logspace = clhp->clh_current;
1491 	clhp->clh_current += size;
1492 	clhp->clh_avail -= size;
1493 	bcopy(data, logspace, size);
1494 	mutex_exit(&clhp->clh_lock);
1495 	return (logspace);
1496 }
1497 
1498 #define	KMEM_AUDIT(lp, cp, bcp)						\
1499 {									\
1500 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1501 	_bcp->bc_timestamp = gethrtime();				\
1502 	_bcp->bc_thread = curthread;					\
1503 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1504 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1505 }
1506 
1507 static void
1508 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1509 	kmem_slab_t *sp, void *addr)
1510 {
1511 	kmem_bufctl_audit_t bca;
1512 
1513 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1514 	bca.bc_addr = addr;
1515 	bca.bc_slab = sp;
1516 	bca.bc_cache = cp;
1517 	KMEM_AUDIT(lp, cp, &bca);
1518 }
1519 
1520 /*
1521  * Create a new slab for cache cp.
1522  */
1523 static kmem_slab_t *
1524 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1525 {
1526 	size_t slabsize = cp->cache_slabsize;
1527 	size_t chunksize = cp->cache_chunksize;
1528 	int cache_flags = cp->cache_flags;
1529 	size_t color, chunks;
1530 	char *buf, *slab;
1531 	kmem_slab_t *sp;
1532 	kmem_bufctl_t *bcp;
1533 	vmem_t *vmp = cp->cache_arena;
1534 
1535 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1536 
1537 	color = cp->cache_color + cp->cache_align;
1538 	if (color > cp->cache_maxcolor)
1539 		color = cp->cache_mincolor;
1540 	cp->cache_color = color;
1541 
1542 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1543 
1544 	if (slab == NULL)
1545 		goto vmem_alloc_failure;
1546 
1547 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1548 
1549 	/*
1550 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1551 	 * consolidator depends (for correctness) on slabs being initialized
1552 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1553 	 * clients to distinguish uninitialized memory from known objects).
1554 	 */
1555 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1556 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1557 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1558 
1559 	if (cache_flags & KMF_HASH) {
1560 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1561 			goto slab_alloc_failure;
1562 		chunks = (slabsize - color) / chunksize;
1563 	} else {
1564 		sp = KMEM_SLAB(cp, slab);
1565 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1566 	}
1567 
1568 	sp->slab_cache	= cp;
1569 	sp->slab_head	= NULL;
1570 	sp->slab_refcnt	= 0;
1571 	sp->slab_base	= buf = slab + color;
1572 	sp->slab_chunks	= chunks;
1573 	sp->slab_stuck_offset = (uint32_t)-1;
1574 	sp->slab_later_count = 0;
1575 	sp->slab_flags = 0;
1576 
1577 	ASSERT(chunks > 0);
1578 	while (chunks-- != 0) {
1579 		if (cache_flags & KMF_HASH) {
1580 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1581 			if (bcp == NULL)
1582 				goto bufctl_alloc_failure;
1583 			if (cache_flags & KMF_AUDIT) {
1584 				kmem_bufctl_audit_t *bcap =
1585 				    (kmem_bufctl_audit_t *)bcp;
1586 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1587 				bcap->bc_cache = cp;
1588 			}
1589 			bcp->bc_addr = buf;
1590 			bcp->bc_slab = sp;
1591 		} else {
1592 			bcp = KMEM_BUFCTL(cp, buf);
1593 		}
1594 		if (cache_flags & KMF_BUFTAG) {
1595 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1596 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1597 			btp->bt_bufctl = bcp;
1598 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1599 			if (cache_flags & KMF_DEADBEEF) {
1600 				copy_pattern(KMEM_FREE_PATTERN, buf,
1601 				    cp->cache_verify);
1602 			}
1603 		}
1604 		bcp->bc_next = sp->slab_head;
1605 		sp->slab_head = bcp;
1606 		buf += chunksize;
1607 	}
1608 
1609 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1610 
1611 	return (sp);
1612 
1613 bufctl_alloc_failure:
1614 
1615 	while ((bcp = sp->slab_head) != NULL) {
1616 		sp->slab_head = bcp->bc_next;
1617 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1618 	}
1619 	kmem_cache_free(kmem_slab_cache, sp);
1620 
1621 slab_alloc_failure:
1622 
1623 	vmem_free(vmp, slab, slabsize);
1624 
1625 vmem_alloc_failure:
1626 
1627 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1628 	atomic_add_64(&cp->cache_alloc_fail, 1);
1629 
1630 	return (NULL);
1631 }
1632 
1633 /*
1634  * Destroy a slab.
1635  */
1636 static void
1637 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1638 {
1639 	vmem_t *vmp = cp->cache_arena;
1640 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1641 
1642 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1643 	ASSERT(sp->slab_refcnt == 0);
1644 
1645 	if (cp->cache_flags & KMF_HASH) {
1646 		kmem_bufctl_t *bcp;
1647 		while ((bcp = sp->slab_head) != NULL) {
1648 			sp->slab_head = bcp->bc_next;
1649 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1650 		}
1651 		kmem_cache_free(kmem_slab_cache, sp);
1652 	}
1653 	vmem_free(vmp, slab, cp->cache_slabsize);
1654 }
1655 
1656 static void *
1657 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp)
1658 {
1659 	kmem_bufctl_t *bcp, **hash_bucket;
1660 	void *buf;
1661 
1662 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1663 	/*
1664 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1665 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1666 	 * slab is newly created (sp->slab_refcnt == 0).
1667 	 */
1668 	ASSERT((sp->slab_refcnt == 0) || (KMEM_SLAB_IS_PARTIAL(sp) &&
1669 	    (sp == avl_first(&cp->cache_partial_slabs))));
1670 	ASSERT(sp->slab_cache == cp);
1671 
1672 	cp->cache_slab_alloc++;
1673 	cp->cache_bufslab--;
1674 	sp->slab_refcnt++;
1675 
1676 	bcp = sp->slab_head;
1677 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1678 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1679 		if (sp->slab_refcnt == 1) {
1680 			ASSERT(sp->slab_chunks == 1);
1681 		} else {
1682 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1683 			avl_remove(&cp->cache_partial_slabs, sp);
1684 			sp->slab_later_count = 0; /* clear history */
1685 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1686 			sp->slab_stuck_offset = (uint32_t)-1;
1687 		}
1688 		list_insert_head(&cp->cache_complete_slabs, sp);
1689 		cp->cache_complete_slab_count++;
1690 	} else {
1691 		ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1692 		if (sp->slab_refcnt == 1) {
1693 			avl_add(&cp->cache_partial_slabs, sp);
1694 		} else {
1695 			/*
1696 			 * The slab is now more allocated than it was, so the
1697 			 * order remains unchanged.
1698 			 */
1699 			ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1700 		}
1701 	}
1702 
1703 	if (cp->cache_flags & KMF_HASH) {
1704 		/*
1705 		 * Add buffer to allocated-address hash table.
1706 		 */
1707 		buf = bcp->bc_addr;
1708 		hash_bucket = KMEM_HASH(cp, buf);
1709 		bcp->bc_next = *hash_bucket;
1710 		*hash_bucket = bcp;
1711 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1712 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1713 		}
1714 	} else {
1715 		buf = KMEM_BUF(cp, bcp);
1716 	}
1717 
1718 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1719 	return (buf);
1720 }
1721 
1722 /*
1723  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1724  */
1725 static void *
1726 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1727 {
1728 	kmem_slab_t *sp;
1729 	void *buf;
1730 	boolean_t test_destructor;
1731 
1732 	mutex_enter(&cp->cache_lock);
1733 	test_destructor = (cp->cache_slab_alloc == 0);
1734 	sp = avl_first(&cp->cache_partial_slabs);
1735 	if (sp == NULL) {
1736 		ASSERT(cp->cache_bufslab == 0);
1737 
1738 		/*
1739 		 * The freelist is empty.  Create a new slab.
1740 		 */
1741 		mutex_exit(&cp->cache_lock);
1742 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1743 			return (NULL);
1744 		}
1745 		mutex_enter(&cp->cache_lock);
1746 		cp->cache_slab_create++;
1747 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1748 			cp->cache_bufmax = cp->cache_buftotal;
1749 		cp->cache_bufslab += sp->slab_chunks;
1750 	}
1751 
1752 	buf = kmem_slab_alloc_impl(cp, sp);
1753 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1754 	    (cp->cache_complete_slab_count +
1755 	    avl_numnodes(&cp->cache_partial_slabs) +
1756 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1757 	mutex_exit(&cp->cache_lock);
1758 
1759 	if (test_destructor && cp->cache_destructor != NULL) {
1760 		/*
1761 		 * On the first kmem_slab_alloc(), assert that it is valid to
1762 		 * call the destructor on a newly constructed object without any
1763 		 * client involvement.
1764 		 */
1765 		if ((cp->cache_constructor == NULL) ||
1766 		    cp->cache_constructor(buf, cp->cache_private,
1767 		    kmflag) == 0) {
1768 			cp->cache_destructor(buf, cp->cache_private);
1769 		}
1770 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1771 		    cp->cache_bufsize);
1772 		if (cp->cache_flags & KMF_DEADBEEF) {
1773 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1774 		}
1775 	}
1776 
1777 	return (buf);
1778 }
1779 
1780 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1781 
1782 /*
1783  * Free a raw (unconstructed) buffer to cp's slab layer.
1784  */
1785 static void
1786 kmem_slab_free(kmem_cache_t *cp, void *buf)
1787 {
1788 	kmem_slab_t *sp;
1789 	kmem_bufctl_t *bcp, **prev_bcpp;
1790 
1791 	ASSERT(buf != NULL);
1792 
1793 	mutex_enter(&cp->cache_lock);
1794 	cp->cache_slab_free++;
1795 
1796 	if (cp->cache_flags & KMF_HASH) {
1797 		/*
1798 		 * Look up buffer in allocated-address hash table.
1799 		 */
1800 		prev_bcpp = KMEM_HASH(cp, buf);
1801 		while ((bcp = *prev_bcpp) != NULL) {
1802 			if (bcp->bc_addr == buf) {
1803 				*prev_bcpp = bcp->bc_next;
1804 				sp = bcp->bc_slab;
1805 				break;
1806 			}
1807 			cp->cache_lookup_depth++;
1808 			prev_bcpp = &bcp->bc_next;
1809 		}
1810 	} else {
1811 		bcp = KMEM_BUFCTL(cp, buf);
1812 		sp = KMEM_SLAB(cp, buf);
1813 	}
1814 
1815 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1816 		mutex_exit(&cp->cache_lock);
1817 		kmem_error(KMERR_BADADDR, cp, buf);
1818 		return;
1819 	}
1820 
1821 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1822 		/*
1823 		 * If this is the buffer that prevented the consolidator from
1824 		 * clearing the slab, we can reset the slab flags now that the
1825 		 * buffer is freed. (It makes sense to do this in
1826 		 * kmem_cache_free(), where the client gives up ownership of the
1827 		 * buffer, but on the hot path the test is too expensive.)
1828 		 */
1829 		kmem_slab_move_yes(cp, sp, buf);
1830 	}
1831 
1832 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1833 		if (cp->cache_flags & KMF_CONTENTS)
1834 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1835 			    kmem_log_enter(kmem_content_log, buf,
1836 			    cp->cache_contents);
1837 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1838 	}
1839 
1840 	bcp->bc_next = sp->slab_head;
1841 	sp->slab_head = bcp;
1842 
1843 	cp->cache_bufslab++;
1844 	ASSERT(sp->slab_refcnt >= 1);
1845 
1846 	if (--sp->slab_refcnt == 0) {
1847 		/*
1848 		 * There are no outstanding allocations from this slab,
1849 		 * so we can reclaim the memory.
1850 		 */
1851 		if (sp->slab_chunks == 1) {
1852 			list_remove(&cp->cache_complete_slabs, sp);
1853 			cp->cache_complete_slab_count--;
1854 		} else {
1855 			avl_remove(&cp->cache_partial_slabs, sp);
1856 		}
1857 
1858 		cp->cache_buftotal -= sp->slab_chunks;
1859 		cp->cache_bufslab -= sp->slab_chunks;
1860 		/*
1861 		 * Defer releasing the slab to the virtual memory subsystem
1862 		 * while there is a pending move callback, since we guarantee
1863 		 * that buffers passed to the move callback have only been
1864 		 * touched by kmem or by the client itself. Since the memory
1865 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1866 		 * set at least one of the two lowest order bits, the client can
1867 		 * test those bits in the move callback to determine whether or
1868 		 * not it knows about the buffer (assuming that the client also
1869 		 * sets one of those low order bits whenever it frees a buffer).
1870 		 */
1871 		if (cp->cache_defrag == NULL ||
1872 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1873 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1874 			cp->cache_slab_destroy++;
1875 			mutex_exit(&cp->cache_lock);
1876 			kmem_slab_destroy(cp, sp);
1877 		} else {
1878 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1879 			/*
1880 			 * Slabs are inserted at both ends of the deadlist to
1881 			 * distinguish between slabs freed while move callbacks
1882 			 * are pending (list head) and a slab freed while the
1883 			 * lock is dropped in kmem_move_buffers() (list tail) so
1884 			 * that in both cases slab_destroy() is called from the
1885 			 * right context.
1886 			 */
1887 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1888 				list_insert_tail(deadlist, sp);
1889 			} else {
1890 				list_insert_head(deadlist, sp);
1891 			}
1892 			cp->cache_defrag->kmd_deadcount++;
1893 			mutex_exit(&cp->cache_lock);
1894 		}
1895 		return;
1896 	}
1897 
1898 	if (bcp->bc_next == NULL) {
1899 		/* Transition the slab from completely allocated to partial. */
1900 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1901 		ASSERT(sp->slab_chunks > 1);
1902 		list_remove(&cp->cache_complete_slabs, sp);
1903 		cp->cache_complete_slab_count--;
1904 		avl_add(&cp->cache_partial_slabs, sp);
1905 	} else {
1906 #ifdef	DEBUG
1907 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1908 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1909 		} else {
1910 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1911 		}
1912 #else
1913 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1914 #endif
1915 	}
1916 
1917 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1918 	    (cp->cache_complete_slab_count +
1919 	    avl_numnodes(&cp->cache_partial_slabs) +
1920 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1921 	mutex_exit(&cp->cache_lock);
1922 }
1923 
1924 /*
1925  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1926  */
1927 static int
1928 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1929     caddr_t caller)
1930 {
1931 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1932 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1933 	uint32_t mtbf;
1934 
1935 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1936 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1937 		return (-1);
1938 	}
1939 
1940 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1941 
1942 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1943 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1944 		return (-1);
1945 	}
1946 
1947 	if (cp->cache_flags & KMF_DEADBEEF) {
1948 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1949 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1950 				kmem_error(KMERR_MODIFIED, cp, buf);
1951 				return (-1);
1952 			}
1953 			if (cp->cache_constructor != NULL)
1954 				*(uint64_t *)buf = btp->bt_redzone;
1955 			else
1956 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1957 		} else {
1958 			construct = 1;
1959 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1960 			    KMEM_UNINITIALIZED_PATTERN, buf,
1961 			    cp->cache_verify)) {
1962 				kmem_error(KMERR_MODIFIED, cp, buf);
1963 				return (-1);
1964 			}
1965 		}
1966 	}
1967 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1968 
1969 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1970 	    gethrtime() % mtbf == 0 &&
1971 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1972 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1973 		if (!construct && cp->cache_destructor != NULL)
1974 			cp->cache_destructor(buf, cp->cache_private);
1975 	} else {
1976 		mtbf = 0;
1977 	}
1978 
1979 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1980 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1981 		atomic_add_64(&cp->cache_alloc_fail, 1);
1982 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1983 		if (cp->cache_flags & KMF_DEADBEEF)
1984 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1985 		kmem_slab_free(cp, buf);
1986 		return (1);
1987 	}
1988 
1989 	if (cp->cache_flags & KMF_AUDIT) {
1990 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1991 	}
1992 
1993 	if ((cp->cache_flags & KMF_LITE) &&
1994 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1995 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1996 	}
1997 
1998 	return (0);
1999 }
2000 
2001 static int
2002 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2003 {
2004 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2005 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2006 	kmem_slab_t *sp;
2007 
2008 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2009 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2010 			kmem_error(KMERR_DUPFREE, cp, buf);
2011 			return (-1);
2012 		}
2013 		sp = kmem_findslab(cp, buf);
2014 		if (sp == NULL || sp->slab_cache != cp)
2015 			kmem_error(KMERR_BADADDR, cp, buf);
2016 		else
2017 			kmem_error(KMERR_REDZONE, cp, buf);
2018 		return (-1);
2019 	}
2020 
2021 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2022 
2023 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2024 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2025 		return (-1);
2026 	}
2027 
2028 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2029 		kmem_error(KMERR_REDZONE, cp, buf);
2030 		return (-1);
2031 	}
2032 
2033 	if (cp->cache_flags & KMF_AUDIT) {
2034 		if (cp->cache_flags & KMF_CONTENTS)
2035 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2036 			    buf, cp->cache_contents);
2037 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2038 	}
2039 
2040 	if ((cp->cache_flags & KMF_LITE) &&
2041 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2042 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2043 	}
2044 
2045 	if (cp->cache_flags & KMF_DEADBEEF) {
2046 		if (cp->cache_flags & KMF_LITE)
2047 			btp->bt_redzone = *(uint64_t *)buf;
2048 		else if (cp->cache_destructor != NULL)
2049 			cp->cache_destructor(buf, cp->cache_private);
2050 
2051 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2052 	}
2053 
2054 	return (0);
2055 }
2056 
2057 /*
2058  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2059  */
2060 static void
2061 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2062 {
2063 	int round;
2064 
2065 	ASSERT(!list_link_active(&cp->cache_link) ||
2066 	    taskq_member(kmem_taskq, curthread));
2067 
2068 	for (round = 0; round < nrounds; round++) {
2069 		void *buf = mp->mag_round[round];
2070 
2071 		if (cp->cache_flags & KMF_DEADBEEF) {
2072 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2073 			    cp->cache_verify) != NULL) {
2074 				kmem_error(KMERR_MODIFIED, cp, buf);
2075 				continue;
2076 			}
2077 			if ((cp->cache_flags & KMF_LITE) &&
2078 			    cp->cache_destructor != NULL) {
2079 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2080 				*(uint64_t *)buf = btp->bt_redzone;
2081 				cp->cache_destructor(buf, cp->cache_private);
2082 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2083 			}
2084 		} else if (cp->cache_destructor != NULL) {
2085 			cp->cache_destructor(buf, cp->cache_private);
2086 		}
2087 
2088 		kmem_slab_free(cp, buf);
2089 	}
2090 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2091 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2092 }
2093 
2094 /*
2095  * Allocate a magazine from the depot.
2096  */
2097 static kmem_magazine_t *
2098 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2099 {
2100 	kmem_magazine_t *mp;
2101 
2102 	/*
2103 	 * If we can't get the depot lock without contention,
2104 	 * update our contention count.  We use the depot
2105 	 * contention rate to determine whether we need to
2106 	 * increase the magazine size for better scalability.
2107 	 */
2108 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2109 		mutex_enter(&cp->cache_depot_lock);
2110 		cp->cache_depot_contention++;
2111 	}
2112 
2113 	if ((mp = mlp->ml_list) != NULL) {
2114 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2115 		mlp->ml_list = mp->mag_next;
2116 		if (--mlp->ml_total < mlp->ml_min)
2117 			mlp->ml_min = mlp->ml_total;
2118 		mlp->ml_alloc++;
2119 	}
2120 
2121 	mutex_exit(&cp->cache_depot_lock);
2122 
2123 	return (mp);
2124 }
2125 
2126 /*
2127  * Free a magazine to the depot.
2128  */
2129 static void
2130 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2131 {
2132 	mutex_enter(&cp->cache_depot_lock);
2133 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2134 	mp->mag_next = mlp->ml_list;
2135 	mlp->ml_list = mp;
2136 	mlp->ml_total++;
2137 	mutex_exit(&cp->cache_depot_lock);
2138 }
2139 
2140 /*
2141  * Update the working set statistics for cp's depot.
2142  */
2143 static void
2144 kmem_depot_ws_update(kmem_cache_t *cp)
2145 {
2146 	mutex_enter(&cp->cache_depot_lock);
2147 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2148 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2149 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2150 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2151 	mutex_exit(&cp->cache_depot_lock);
2152 }
2153 
2154 /*
2155  * Reap all magazines that have fallen out of the depot's working set.
2156  */
2157 static void
2158 kmem_depot_ws_reap(kmem_cache_t *cp)
2159 {
2160 	long reap;
2161 	kmem_magazine_t *mp;
2162 
2163 	ASSERT(!list_link_active(&cp->cache_link) ||
2164 	    taskq_member(kmem_taskq, curthread));
2165 
2166 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2167 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2168 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2169 
2170 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2171 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2172 		kmem_magazine_destroy(cp, mp, 0);
2173 }
2174 
2175 static void
2176 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2177 {
2178 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2179 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2180 	ASSERT(ccp->cc_magsize > 0);
2181 
2182 	ccp->cc_ploaded = ccp->cc_loaded;
2183 	ccp->cc_prounds = ccp->cc_rounds;
2184 	ccp->cc_loaded = mp;
2185 	ccp->cc_rounds = rounds;
2186 }
2187 
2188 /*
2189  * Allocate a constructed object from cache cp.
2190  */
2191 void *
2192 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2193 {
2194 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2195 	kmem_magazine_t *fmp;
2196 	void *buf;
2197 
2198 	mutex_enter(&ccp->cc_lock);
2199 	for (;;) {
2200 		/*
2201 		 * If there's an object available in the current CPU's
2202 		 * loaded magazine, just take it and return.
2203 		 */
2204 		if (ccp->cc_rounds > 0) {
2205 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2206 			ccp->cc_alloc++;
2207 			mutex_exit(&ccp->cc_lock);
2208 			if ((ccp->cc_flags & KMF_BUFTAG) &&
2209 			    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2210 			    caller()) != 0) {
2211 				if (kmflag & KM_NOSLEEP)
2212 					return (NULL);
2213 				mutex_enter(&ccp->cc_lock);
2214 				continue;
2215 			}
2216 			return (buf);
2217 		}
2218 
2219 		/*
2220 		 * The loaded magazine is empty.  If the previously loaded
2221 		 * magazine was full, exchange them and try again.
2222 		 */
2223 		if (ccp->cc_prounds > 0) {
2224 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2225 			continue;
2226 		}
2227 
2228 		/*
2229 		 * If the magazine layer is disabled, break out now.
2230 		 */
2231 		if (ccp->cc_magsize == 0)
2232 			break;
2233 
2234 		/*
2235 		 * Try to get a full magazine from the depot.
2236 		 */
2237 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2238 		if (fmp != NULL) {
2239 			if (ccp->cc_ploaded != NULL)
2240 				kmem_depot_free(cp, &cp->cache_empty,
2241 				    ccp->cc_ploaded);
2242 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2243 			continue;
2244 		}
2245 
2246 		/*
2247 		 * There are no full magazines in the depot,
2248 		 * so fall through to the slab layer.
2249 		 */
2250 		break;
2251 	}
2252 	mutex_exit(&ccp->cc_lock);
2253 
2254 	/*
2255 	 * We couldn't allocate a constructed object from the magazine layer,
2256 	 * so get a raw buffer from the slab layer and apply its constructor.
2257 	 */
2258 	buf = kmem_slab_alloc(cp, kmflag);
2259 
2260 	if (buf == NULL)
2261 		return (NULL);
2262 
2263 	if (cp->cache_flags & KMF_BUFTAG) {
2264 		/*
2265 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2266 		 */
2267 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2268 		if (rc != 0) {
2269 			if (kmflag & KM_NOSLEEP)
2270 				return (NULL);
2271 			/*
2272 			 * kmem_cache_alloc_debug() detected corruption
2273 			 * but didn't panic (kmem_panic <= 0). We should not be
2274 			 * here because the constructor failed (indicated by a
2275 			 * return code of 1). Try again.
2276 			 */
2277 			ASSERT(rc == -1);
2278 			return (kmem_cache_alloc(cp, kmflag));
2279 		}
2280 		return (buf);
2281 	}
2282 
2283 	if (cp->cache_constructor != NULL &&
2284 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2285 		atomic_add_64(&cp->cache_alloc_fail, 1);
2286 		kmem_slab_free(cp, buf);
2287 		return (NULL);
2288 	}
2289 
2290 	return (buf);
2291 }
2292 
2293 /*
2294  * The freed argument tells whether or not kmem_cache_free_debug() has already
2295  * been called so that we can avoid the duplicate free error. For example, a
2296  * buffer on a magazine has already been freed by the client but is still
2297  * constructed.
2298  */
2299 static void
2300 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2301 {
2302 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2303 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2304 			return;
2305 
2306 	/*
2307 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2308 	 * kmem_cache_free_debug() will have already applied the destructor.
2309 	 */
2310 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2311 	    cp->cache_destructor != NULL) {
2312 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2313 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2314 			*(uint64_t *)buf = btp->bt_redzone;
2315 			cp->cache_destructor(buf, cp->cache_private);
2316 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2317 		} else {
2318 			cp->cache_destructor(buf, cp->cache_private);
2319 		}
2320 	}
2321 
2322 	kmem_slab_free(cp, buf);
2323 }
2324 
2325 /*
2326  * Free a constructed object to cache cp.
2327  */
2328 void
2329 kmem_cache_free(kmem_cache_t *cp, void *buf)
2330 {
2331 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2332 	kmem_magazine_t *emp;
2333 	kmem_magtype_t *mtp;
2334 
2335 	/*
2336 	 * The client must not free either of the buffers passed to the move
2337 	 * callback function.
2338 	 */
2339 	ASSERT(cp->cache_defrag == NULL ||
2340 	    cp->cache_defrag->kmd_thread != curthread ||
2341 	    (buf != cp->cache_defrag->kmd_from_buf &&
2342 	    buf != cp->cache_defrag->kmd_to_buf));
2343 
2344 	if (ccp->cc_flags & KMF_BUFTAG)
2345 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2346 			return;
2347 
2348 	mutex_enter(&ccp->cc_lock);
2349 	for (;;) {
2350 		/*
2351 		 * If there's a slot available in the current CPU's
2352 		 * loaded magazine, just put the object there and return.
2353 		 */
2354 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2355 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2356 			ccp->cc_free++;
2357 			mutex_exit(&ccp->cc_lock);
2358 			return;
2359 		}
2360 
2361 		/*
2362 		 * The loaded magazine is full.  If the previously loaded
2363 		 * magazine was empty, exchange them and try again.
2364 		 */
2365 		if (ccp->cc_prounds == 0) {
2366 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2367 			continue;
2368 		}
2369 
2370 		/*
2371 		 * If the magazine layer is disabled, break out now.
2372 		 */
2373 		if (ccp->cc_magsize == 0)
2374 			break;
2375 
2376 		/*
2377 		 * Try to get an empty magazine from the depot.
2378 		 */
2379 		emp = kmem_depot_alloc(cp, &cp->cache_empty);
2380 		if (emp != NULL) {
2381 			if (ccp->cc_ploaded != NULL)
2382 				kmem_depot_free(cp, &cp->cache_full,
2383 				    ccp->cc_ploaded);
2384 			kmem_cpu_reload(ccp, emp, 0);
2385 			continue;
2386 		}
2387 
2388 		/*
2389 		 * There are no empty magazines in the depot,
2390 		 * so try to allocate a new one.  We must drop all locks
2391 		 * across kmem_cache_alloc() because lower layers may
2392 		 * attempt to allocate from this cache.
2393 		 */
2394 		mtp = cp->cache_magtype;
2395 		mutex_exit(&ccp->cc_lock);
2396 		emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2397 		mutex_enter(&ccp->cc_lock);
2398 
2399 		if (emp != NULL) {
2400 			/*
2401 			 * We successfully allocated an empty magazine.
2402 			 * However, we had to drop ccp->cc_lock to do it,
2403 			 * so the cache's magazine size may have changed.
2404 			 * If so, free the magazine and try again.
2405 			 */
2406 			if (ccp->cc_magsize != mtp->mt_magsize) {
2407 				mutex_exit(&ccp->cc_lock);
2408 				kmem_cache_free(mtp->mt_cache, emp);
2409 				mutex_enter(&ccp->cc_lock);
2410 				continue;
2411 			}
2412 
2413 			/*
2414 			 * We got a magazine of the right size.  Add it to
2415 			 * the depot and try the whole dance again.
2416 			 */
2417 			kmem_depot_free(cp, &cp->cache_empty, emp);
2418 			continue;
2419 		}
2420 
2421 		/*
2422 		 * We couldn't allocate an empty magazine,
2423 		 * so fall through to the slab layer.
2424 		 */
2425 		break;
2426 	}
2427 	mutex_exit(&ccp->cc_lock);
2428 
2429 	/*
2430 	 * We couldn't free our constructed object to the magazine layer,
2431 	 * so apply its destructor and free it to the slab layer.
2432 	 */
2433 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2434 }
2435 
2436 void *
2437 kmem_zalloc(size_t size, int kmflag)
2438 {
2439 	size_t index;
2440 	void *buf;
2441 
2442 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2443 		kmem_cache_t *cp = kmem_alloc_table[index];
2444 		buf = kmem_cache_alloc(cp, kmflag);
2445 		if (buf != NULL) {
2446 			if (cp->cache_flags & KMF_BUFTAG) {
2447 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2448 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2449 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2450 
2451 				if (cp->cache_flags & KMF_LITE) {
2452 					KMEM_BUFTAG_LITE_ENTER(btp,
2453 					    kmem_lite_count, caller());
2454 				}
2455 			}
2456 			bzero(buf, size);
2457 		}
2458 	} else {
2459 		buf = kmem_alloc(size, kmflag);
2460 		if (buf != NULL)
2461 			bzero(buf, size);
2462 	}
2463 	return (buf);
2464 }
2465 
2466 void *
2467 kmem_alloc(size_t size, int kmflag)
2468 {
2469 	size_t index;
2470 	kmem_cache_t *cp;
2471 	void *buf;
2472 
2473 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2474 		cp = kmem_alloc_table[index];
2475 		/* fall through to kmem_cache_alloc() */
2476 
2477 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2478 	    kmem_big_alloc_table_max) {
2479 		cp = kmem_big_alloc_table[index];
2480 		/* fall through to kmem_cache_alloc() */
2481 
2482 	} else {
2483 		if (size == 0)
2484 			return (NULL);
2485 
2486 		buf = vmem_alloc(kmem_oversize_arena, size,
2487 		    kmflag & KM_VMFLAGS);
2488 		if (buf == NULL)
2489 			kmem_log_event(kmem_failure_log, NULL, NULL,
2490 			    (void *)size);
2491 		return (buf);
2492 	}
2493 
2494 	buf = kmem_cache_alloc(cp, kmflag);
2495 	if ((cp->cache_flags & KMF_BUFTAG) && buf != NULL) {
2496 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2497 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2498 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2499 
2500 		if (cp->cache_flags & KMF_LITE) {
2501 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2502 		}
2503 	}
2504 	return (buf);
2505 }
2506 
2507 void
2508 kmem_free(void *buf, size_t size)
2509 {
2510 	size_t index;
2511 	kmem_cache_t *cp;
2512 
2513 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2514 		cp = kmem_alloc_table[index];
2515 		/* fall through to kmem_cache_free() */
2516 
2517 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2518 	    kmem_big_alloc_table_max) {
2519 		cp = kmem_big_alloc_table[index];
2520 		/* fall through to kmem_cache_free() */
2521 
2522 	} else {
2523 		if (buf == NULL && size == 0)
2524 			return;
2525 		vmem_free(kmem_oversize_arena, buf, size);
2526 		return;
2527 	}
2528 
2529 	if (cp->cache_flags & KMF_BUFTAG) {
2530 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2531 		uint32_t *ip = (uint32_t *)btp;
2532 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2533 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2534 				kmem_error(KMERR_DUPFREE, cp, buf);
2535 				return;
2536 			}
2537 			if (KMEM_SIZE_VALID(ip[1])) {
2538 				ip[0] = KMEM_SIZE_ENCODE(size);
2539 				kmem_error(KMERR_BADSIZE, cp, buf);
2540 			} else {
2541 				kmem_error(KMERR_REDZONE, cp, buf);
2542 			}
2543 			return;
2544 		}
2545 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2546 			kmem_error(KMERR_REDZONE, cp, buf);
2547 			return;
2548 		}
2549 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
2550 		if (cp->cache_flags & KMF_LITE) {
2551 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2552 			    caller());
2553 		}
2554 	}
2555 	kmem_cache_free(cp, buf);
2556 }
2557 
2558 void *
2559 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2560 {
2561 	size_t realsize = size + vmp->vm_quantum;
2562 	void *addr;
2563 
2564 	/*
2565 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2566 	 * vm_quantum will cause integer wraparound.  Check for this, and
2567 	 * blow off the firewall page in this case.  Note that such a
2568 	 * giant allocation (the entire kernel address space) can never
2569 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
2570 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2571 	 * corresponding check in kmem_firewall_va_free().
2572 	 */
2573 	if (realsize < size)
2574 		realsize = size;
2575 
2576 	/*
2577 	 * While boot still owns resource management, make sure that this
2578 	 * redzone virtual address allocation is properly accounted for in
2579 	 * OBPs "virtual-memory" "available" lists because we're
2580 	 * effectively claiming them for a red zone.  If we don't do this,
2581 	 * the available lists become too fragmented and too large for the
2582 	 * current boot/kernel memory list interface.
2583 	 */
2584 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
2585 
2586 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
2587 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
2588 
2589 	return (addr);
2590 }
2591 
2592 void
2593 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2594 {
2595 	ASSERT((kvseg.s_base == NULL ?
2596 	    va_to_pfn((char *)addr + size) :
2597 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
2598 
2599 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2600 }
2601 
2602 /*
2603  * Try to allocate at least `size' bytes of memory without sleeping or
2604  * panicking. Return actual allocated size in `asize'. If allocation failed,
2605  * try final allocation with sleep or panic allowed.
2606  */
2607 void *
2608 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
2609 {
2610 	void *p;
2611 
2612 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2613 	do {
2614 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
2615 		if (p != NULL)
2616 			return (p);
2617 		*asize += KMEM_ALIGN;
2618 	} while (*asize <= PAGESIZE);
2619 
2620 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
2621 	return (kmem_alloc(*asize, kmflag));
2622 }
2623 
2624 /*
2625  * Reclaim all unused memory from a cache.
2626  */
2627 static void
2628 kmem_cache_reap(kmem_cache_t *cp)
2629 {
2630 	ASSERT(taskq_member(kmem_taskq, curthread));
2631 	cp->cache_reap++;
2632 
2633 	/*
2634 	 * Ask the cache's owner to free some memory if possible.
2635 	 * The idea is to handle things like the inode cache, which
2636 	 * typically sits on a bunch of memory that it doesn't truly
2637 	 * *need*.  Reclaim policy is entirely up to the owner; this
2638 	 * callback is just an advisory plea for help.
2639 	 */
2640 	if (cp->cache_reclaim != NULL) {
2641 		long delta;
2642 
2643 		/*
2644 		 * Reclaimed memory should be reapable (not included in the
2645 		 * depot's working set).
2646 		 */
2647 		delta = cp->cache_full.ml_total;
2648 		cp->cache_reclaim(cp->cache_private);
2649 		delta = cp->cache_full.ml_total - delta;
2650 		if (delta > 0) {
2651 			mutex_enter(&cp->cache_depot_lock);
2652 			cp->cache_full.ml_reaplimit += delta;
2653 			cp->cache_full.ml_min += delta;
2654 			mutex_exit(&cp->cache_depot_lock);
2655 		}
2656 	}
2657 
2658 	kmem_depot_ws_reap(cp);
2659 
2660 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
2661 		kmem_cache_defrag(cp);
2662 	}
2663 }
2664 
2665 static void
2666 kmem_reap_timeout(void *flag_arg)
2667 {
2668 	uint32_t *flag = (uint32_t *)flag_arg;
2669 
2670 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
2671 	*flag = 0;
2672 }
2673 
2674 static void
2675 kmem_reap_done(void *flag)
2676 {
2677 	(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
2678 }
2679 
2680 static void
2681 kmem_reap_start(void *flag)
2682 {
2683 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
2684 
2685 	if (flag == &kmem_reaping) {
2686 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
2687 		/*
2688 		 * if we have segkp under heap, reap segkp cache.
2689 		 */
2690 		if (segkp_fromheap)
2691 			segkp_cache_free();
2692 	}
2693 	else
2694 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
2695 
2696 	/*
2697 	 * We use taskq_dispatch() to schedule a timeout to clear
2698 	 * the flag so that kmem_reap() becomes self-throttling:
2699 	 * we won't reap again until the current reap completes *and*
2700 	 * at least kmem_reap_interval ticks have elapsed.
2701 	 */
2702 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
2703 		kmem_reap_done(flag);
2704 }
2705 
2706 static void
2707 kmem_reap_common(void *flag_arg)
2708 {
2709 	uint32_t *flag = (uint32_t *)flag_arg;
2710 
2711 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
2712 	    cas32(flag, 0, 1) != 0)
2713 		return;
2714 
2715 	/*
2716 	 * It may not be kosher to do memory allocation when a reap is called
2717 	 * is called (for example, if vmem_populate() is in the call chain).
2718 	 * So we start the reap going with a TQ_NOALLOC dispatch.  If the
2719 	 * dispatch fails, we reset the flag, and the next reap will try again.
2720 	 */
2721 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
2722 		*flag = 0;
2723 }
2724 
2725 /*
2726  * Reclaim all unused memory from all caches.  Called from the VM system
2727  * when memory gets tight.
2728  */
2729 void
2730 kmem_reap(void)
2731 {
2732 	kmem_reap_common(&kmem_reaping);
2733 }
2734 
2735 /*
2736  * Reclaim all unused memory from identifier arenas, called when a vmem
2737  * arena not back by memory is exhausted.  Since reaping memory-backed caches
2738  * cannot help with identifier exhaustion, we avoid both a large amount of
2739  * work and unwanted side-effects from reclaim callbacks.
2740  */
2741 void
2742 kmem_reap_idspace(void)
2743 {
2744 	kmem_reap_common(&kmem_reaping_idspace);
2745 }
2746 
2747 /*
2748  * Purge all magazines from a cache and set its magazine limit to zero.
2749  * All calls are serialized by the kmem_taskq lock, except for the final
2750  * call from kmem_cache_destroy().
2751  */
2752 static void
2753 kmem_cache_magazine_purge(kmem_cache_t *cp)
2754 {
2755 	kmem_cpu_cache_t *ccp;
2756 	kmem_magazine_t *mp, *pmp;
2757 	int rounds, prounds, cpu_seqid;
2758 
2759 	ASSERT(!list_link_active(&cp->cache_link) ||
2760 	    taskq_member(kmem_taskq, curthread));
2761 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
2762 
2763 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
2764 		ccp = &cp->cache_cpu[cpu_seqid];
2765 
2766 		mutex_enter(&ccp->cc_lock);
2767 		mp = ccp->cc_loaded;
2768 		pmp = ccp->cc_ploaded;
2769 		rounds = ccp->cc_rounds;
2770 		prounds = ccp->cc_prounds;
2771 		ccp->cc_loaded = NULL;
2772 		ccp->cc_ploaded = NULL;
2773 		ccp->cc_rounds = -1;
2774 		ccp->cc_prounds = -1;
2775 		ccp->cc_magsize = 0;
2776 		mutex_exit(&ccp->cc_lock);
2777 
2778 		if (mp)
2779 			kmem_magazine_destroy(cp, mp, rounds);
2780 		if (pmp)
2781 			kmem_magazine_destroy(cp, pmp, prounds);
2782 	}
2783 
2784 	/*
2785 	 * Updating the working set statistics twice in a row has the
2786 	 * effect of setting the working set size to zero, so everything
2787 	 * is eligible for reaping.
2788 	 */
2789 	kmem_depot_ws_update(cp);
2790 	kmem_depot_ws_update(cp);
2791 
2792 	kmem_depot_ws_reap(cp);
2793 }
2794 
2795 /*
2796  * Enable per-cpu magazines on a cache.
2797  */
2798 static void
2799 kmem_cache_magazine_enable(kmem_cache_t *cp)
2800 {
2801 	int cpu_seqid;
2802 
2803 	if (cp->cache_flags & KMF_NOMAGAZINE)
2804 		return;
2805 
2806 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
2807 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2808 		mutex_enter(&ccp->cc_lock);
2809 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2810 		mutex_exit(&ccp->cc_lock);
2811 	}
2812 
2813 }
2814 
2815 /*
2816  * Reap (almost) everything right now.  See kmem_cache_magazine_purge()
2817  * for explanation of the back-to-back kmem_depot_ws_update() calls.
2818  */
2819 void
2820 kmem_cache_reap_now(kmem_cache_t *cp)
2821 {
2822 	ASSERT(list_link_active(&cp->cache_link));
2823 
2824 	kmem_depot_ws_update(cp);
2825 	kmem_depot_ws_update(cp);
2826 
2827 	(void) taskq_dispatch(kmem_taskq,
2828 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
2829 	taskq_wait(kmem_taskq);
2830 }
2831 
2832 /*
2833  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2834  * provide a higher transfer rate with the depot, while smaller magazines
2835  * reduce memory consumption.  Magazine resizing is an expensive operation;
2836  * it should not be done frequently.
2837  *
2838  * Changes to the magazine size are serialized by the kmem_taskq lock.
2839  *
2840  * Note: at present this only grows the magazine size.  It might be useful
2841  * to allow shrinkage too.
2842  */
2843 static void
2844 kmem_cache_magazine_resize(kmem_cache_t *cp)
2845 {
2846 	kmem_magtype_t *mtp = cp->cache_magtype;
2847 
2848 	ASSERT(taskq_member(kmem_taskq, curthread));
2849 
2850 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2851 		kmem_cache_magazine_purge(cp);
2852 		mutex_enter(&cp->cache_depot_lock);
2853 		cp->cache_magtype = ++mtp;
2854 		cp->cache_depot_contention_prev =
2855 		    cp->cache_depot_contention + INT_MAX;
2856 		mutex_exit(&cp->cache_depot_lock);
2857 		kmem_cache_magazine_enable(cp);
2858 	}
2859 }
2860 
2861 /*
2862  * Rescale a cache's hash table, so that the table size is roughly the
2863  * cache size.  We want the average lookup time to be extremely small.
2864  */
2865 static void
2866 kmem_hash_rescale(kmem_cache_t *cp)
2867 {
2868 	kmem_bufctl_t **old_table, **new_table, *bcp;
2869 	size_t old_size, new_size, h;
2870 
2871 	ASSERT(taskq_member(kmem_taskq, curthread));
2872 
2873 	new_size = MAX(KMEM_HASH_INITIAL,
2874 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2875 	old_size = cp->cache_hash_mask + 1;
2876 
2877 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2878 		return;
2879 
2880 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
2881 	    VM_NOSLEEP);
2882 	if (new_table == NULL)
2883 		return;
2884 	bzero(new_table, new_size * sizeof (void *));
2885 
2886 	mutex_enter(&cp->cache_lock);
2887 
2888 	old_size = cp->cache_hash_mask + 1;
2889 	old_table = cp->cache_hash_table;
2890 
2891 	cp->cache_hash_mask = new_size - 1;
2892 	cp->cache_hash_table = new_table;
2893 	cp->cache_rescale++;
2894 
2895 	for (h = 0; h < old_size; h++) {
2896 		bcp = old_table[h];
2897 		while (bcp != NULL) {
2898 			void *addr = bcp->bc_addr;
2899 			kmem_bufctl_t *next_bcp = bcp->bc_next;
2900 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
2901 			bcp->bc_next = *hash_bucket;
2902 			*hash_bucket = bcp;
2903 			bcp = next_bcp;
2904 		}
2905 	}
2906 
2907 	mutex_exit(&cp->cache_lock);
2908 
2909 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
2910 }
2911 
2912 /*
2913  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
2914  * update, magazine resizing, and slab consolidation.
2915  */
2916 static void
2917 kmem_cache_update(kmem_cache_t *cp)
2918 {
2919 	int need_hash_rescale = 0;
2920 	int need_magazine_resize = 0;
2921 
2922 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
2923 
2924 	/*
2925 	 * If the cache has become much larger or smaller than its hash table,
2926 	 * fire off a request to rescale the hash table.
2927 	 */
2928 	mutex_enter(&cp->cache_lock);
2929 
2930 	if ((cp->cache_flags & KMF_HASH) &&
2931 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2932 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2933 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
2934 		need_hash_rescale = 1;
2935 
2936 	mutex_exit(&cp->cache_lock);
2937 
2938 	/*
2939 	 * Update the depot working set statistics.
2940 	 */
2941 	kmem_depot_ws_update(cp);
2942 
2943 	/*
2944 	 * If there's a lot of contention in the depot,
2945 	 * increase the magazine size.
2946 	 */
2947 	mutex_enter(&cp->cache_depot_lock);
2948 
2949 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2950 	    (int)(cp->cache_depot_contention -
2951 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
2952 		need_magazine_resize = 1;
2953 
2954 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2955 
2956 	mutex_exit(&cp->cache_depot_lock);
2957 
2958 	if (need_hash_rescale)
2959 		(void) taskq_dispatch(kmem_taskq,
2960 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
2961 
2962 	if (need_magazine_resize)
2963 		(void) taskq_dispatch(kmem_taskq,
2964 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
2965 
2966 	if (cp->cache_defrag != NULL)
2967 		(void) taskq_dispatch(kmem_taskq,
2968 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
2969 }
2970 
2971 static void kmem_update(void *);
2972 
2973 static void
2974 kmem_update_timeout(void *dummy)
2975 {
2976 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
2977 }
2978 
2979 static void
2980 kmem_update(void *dummy)
2981 {
2982 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
2983 
2984 	/*
2985 	 * We use taskq_dispatch() to reschedule the timeout so that
2986 	 * kmem_update() becomes self-throttling: it won't schedule
2987 	 * new tasks until all previous tasks have completed.
2988 	 */
2989 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
2990 		kmem_update_timeout(NULL);
2991 }
2992 
2993 static int
2994 kmem_cache_kstat_update(kstat_t *ksp, int rw)
2995 {
2996 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
2997 	kmem_cache_t *cp = ksp->ks_private;
2998 	uint64_t cpu_buf_avail;
2999 	uint64_t buf_avail = 0;
3000 	int cpu_seqid;
3001 	long reap;
3002 
3003 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3004 
3005 	if (rw == KSTAT_WRITE)
3006 		return (EACCES);
3007 
3008 	mutex_enter(&cp->cache_lock);
3009 
3010 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3011 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3012 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3013 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3014 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3015 
3016 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3017 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3018 
3019 		mutex_enter(&ccp->cc_lock);
3020 
3021 		cpu_buf_avail = 0;
3022 		if (ccp->cc_rounds > 0)
3023 			cpu_buf_avail += ccp->cc_rounds;
3024 		if (ccp->cc_prounds > 0)
3025 			cpu_buf_avail += ccp->cc_prounds;
3026 
3027 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3028 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3029 		buf_avail			+= cpu_buf_avail;
3030 
3031 		mutex_exit(&ccp->cc_lock);
3032 	}
3033 
3034 	mutex_enter(&cp->cache_depot_lock);
3035 
3036 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3037 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3038 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3039 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3040 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3041 	kmcp->kmc_magazine_size.value.ui64	=
3042 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3043 	    0 : cp->cache_magtype->mt_magsize;
3044 
3045 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3046 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3047 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3048 
3049 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3050 	reap = MIN(reap, cp->cache_full.ml_total);
3051 
3052 	mutex_exit(&cp->cache_depot_lock);
3053 
3054 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3055 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3056 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3057 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3058 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3059 	buf_avail += cp->cache_bufslab;
3060 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3061 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3062 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3063 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3064 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3065 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3066 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3067 	    cp->cache_hash_mask + 1 : 0;
3068 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3069 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3070 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3071 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3072 
3073 	if (cp->cache_defrag == NULL) {
3074 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3075 		kmcp->kmc_move_yes.value.ui64		= 0;
3076 		kmcp->kmc_move_no.value.ui64		= 0;
3077 		kmcp->kmc_move_later.value.ui64		= 0;
3078 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3079 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3080 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3081 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3082 		kmcp->kmc_defrag.value.ui64		= 0;
3083 		kmcp->kmc_scan.value.ui64		= 0;
3084 		kmcp->kmc_move_reclaimable.value.ui64	= 0;
3085 	} else {
3086 		int64_t reclaimable;
3087 
3088 		kmem_defrag_t *kd = cp->cache_defrag;
3089 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3090 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3091 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3092 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3093 		kmcp->kmc_move_dont_need.value.ui64	= kd->kmd_dont_need;
3094 		kmcp->kmc_move_dont_know.value.ui64	= kd->kmd_dont_know;
3095 		kmcp->kmc_move_hunt_found.value.ui64	= kd->kmd_hunt_found;
3096 		kmcp->kmc_move_slabs_freed.value.ui64	= kd->kmd_slabs_freed;
3097 		kmcp->kmc_defrag.value.ui64		= kd->kmd_defrags;
3098 		kmcp->kmc_scan.value.ui64		= kd->kmd_scans;
3099 
3100 		reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3101 		reclaimable = MAX(reclaimable, 0);
3102 		reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3103 		kmcp->kmc_move_reclaimable.value.ui64	= reclaimable;
3104 	}
3105 
3106 	mutex_exit(&cp->cache_lock);
3107 	return (0);
3108 }
3109 
3110 /*
3111  * Return a named statistic about a particular cache.
3112  * This shouldn't be called very often, so it's currently designed for
3113  * simplicity (leverages existing kstat support) rather than efficiency.
3114  */
3115 uint64_t
3116 kmem_cache_stat(kmem_cache_t *cp, char *name)
3117 {
3118 	int i;
3119 	kstat_t *ksp = cp->cache_kstat;
3120 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3121 	uint64_t value = 0;
3122 
3123 	if (ksp != NULL) {
3124 		mutex_enter(&kmem_cache_kstat_lock);
3125 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3126 		for (i = 0; i < ksp->ks_ndata; i++) {
3127 			if (strcmp(knp[i].name, name) == 0) {
3128 				value = knp[i].value.ui64;
3129 				break;
3130 			}
3131 		}
3132 		mutex_exit(&kmem_cache_kstat_lock);
3133 	}
3134 	return (value);
3135 }
3136 
3137 /*
3138  * Return an estimate of currently available kernel heap memory.
3139  * On 32-bit systems, physical memory may exceed virtual memory,
3140  * we just truncate the result at 1GB.
3141  */
3142 size_t
3143 kmem_avail(void)
3144 {
3145 	spgcnt_t rmem = availrmem - tune.t_minarmem;
3146 	spgcnt_t fmem = freemem - minfree;
3147 
3148 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3149 	    1 << (30 - PAGESHIFT))));
3150 }
3151 
3152 /*
3153  * Return the maximum amount of memory that is (in theory) allocatable
3154  * from the heap. This may be used as an estimate only since there
3155  * is no guarentee this space will still be available when an allocation
3156  * request is made, nor that the space may be allocated in one big request
3157  * due to kernel heap fragmentation.
3158  */
3159 size_t
3160 kmem_maxavail(void)
3161 {
3162 	spgcnt_t pmem = availrmem - tune.t_minarmem;
3163 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3164 
3165 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3166 }
3167 
3168 /*
3169  * Indicate whether memory-intensive kmem debugging is enabled.
3170  */
3171 int
3172 kmem_debugging(void)
3173 {
3174 	return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3175 }
3176 
3177 /* binning function, sorts finely at the two extremes */
3178 #define	KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift)				\
3179 	((((sp)->slab_refcnt <= (binshift)) ||				\
3180 	    (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift)))	\
3181 	    ? -(sp)->slab_refcnt					\
3182 	    : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3183 
3184 /*
3185  * Minimizing the number of partial slabs on the freelist minimizes
3186  * fragmentation (the ratio of unused buffers held by the slab layer). There are
3187  * two ways to get a slab off of the freelist: 1) free all the buffers on the
3188  * slab, and 2) allocate all the buffers on the slab. It follows that we want
3189  * the most-used slabs at the front of the list where they have the best chance
3190  * of being completely allocated, and the least-used slabs at a safe distance
3191  * from the front to improve the odds that the few remaining buffers will all be
3192  * freed before another allocation can tie up the slab. For that reason a slab
3193  * with a higher slab_refcnt sorts less than than a slab with a lower
3194  * slab_refcnt.
3195  *
3196  * However, if a slab has at least one buffer that is deemed unfreeable, we
3197  * would rather have that slab at the front of the list regardless of
3198  * slab_refcnt, since even one unfreeable buffer makes the entire slab
3199  * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3200  * callback, the slab is marked unfreeable for as long as it remains on the
3201  * freelist.
3202  */
3203 static int
3204 kmem_partial_slab_cmp(const void *p0, const void *p1)
3205 {
3206 	const kmem_cache_t *cp;
3207 	const kmem_slab_t *s0 = p0;
3208 	const kmem_slab_t *s1 = p1;
3209 	int w0, w1;
3210 	size_t binshift;
3211 
3212 	ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3213 	ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3214 	ASSERT(s0->slab_cache == s1->slab_cache);
3215 	cp = s1->slab_cache;
3216 	ASSERT(MUTEX_HELD(&cp->cache_lock));
3217 	binshift = cp->cache_partial_binshift;
3218 
3219 	/* weight of first slab */
3220 	w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3221 	if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3222 		w0 -= cp->cache_maxchunks;
3223 	}
3224 
3225 	/* weight of second slab */
3226 	w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3227 	if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3228 		w1 -= cp->cache_maxchunks;
3229 	}
3230 
3231 	if (w0 < w1)
3232 		return (-1);
3233 	if (w0 > w1)
3234 		return (1);
3235 
3236 	/* compare pointer values */
3237 	if ((uintptr_t)s0 < (uintptr_t)s1)
3238 		return (-1);
3239 	if ((uintptr_t)s0 > (uintptr_t)s1)
3240 		return (1);
3241 
3242 	return (0);
3243 }
3244 
3245 /*
3246  * It must be valid to call the destructor (if any) on a newly created object.
3247  * That is, the constructor (if any) must leave the object in a valid state for
3248  * the destructor.
3249  */
3250 kmem_cache_t *
3251 kmem_cache_create(
3252 	char *name,		/* descriptive name for this cache */
3253 	size_t bufsize,		/* size of the objects it manages */
3254 	size_t align,		/* required object alignment */
3255 	int (*constructor)(void *, void *, int), /* object constructor */
3256 	void (*destructor)(void *, void *),	/* object destructor */
3257 	void (*reclaim)(void *), /* memory reclaim callback */
3258 	void *private,		/* pass-thru arg for constr/destr/reclaim */
3259 	vmem_t *vmp,		/* vmem source for slab allocation */
3260 	int cflags)		/* cache creation flags */
3261 {
3262 	int cpu_seqid;
3263 	size_t chunksize;
3264 	kmem_cache_t *cp;
3265 	kmem_magtype_t *mtp;
3266 	size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3267 
3268 #ifdef	DEBUG
3269 	/*
3270 	 * Cache names should conform to the rules for valid C identifiers
3271 	 */
3272 	if (!strident_valid(name)) {
3273 		cmn_err(CE_CONT,
3274 		    "kmem_cache_create: '%s' is an invalid cache name\n"
3275 		    "cache names must conform to the rules for "
3276 		    "C identifiers\n", name);
3277 	}
3278 #endif	/* DEBUG */
3279 
3280 	if (vmp == NULL)
3281 		vmp = kmem_default_arena;
3282 
3283 	/*
3284 	 * If this kmem cache has an identifier vmem arena as its source, mark
3285 	 * it such to allow kmem_reap_idspace().
3286 	 */
3287 	ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
3288 	if (vmp->vm_cflags & VMC_IDENTIFIER)
3289 		cflags |= KMC_IDENTIFIER;
3290 
3291 	/*
3292 	 * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
3293 	 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3294 	 * false sharing of per-CPU data.
3295 	 */
3296 	cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3297 	    P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3298 	bzero(cp, csize);
3299 	list_link_init(&cp->cache_link);
3300 
3301 	if (align == 0)
3302 		align = KMEM_ALIGN;
3303 
3304 	/*
3305 	 * If we're not at least KMEM_ALIGN aligned, we can't use free
3306 	 * memory to hold bufctl information (because we can't safely
3307 	 * perform word loads and stores on it).
3308 	 */
3309 	if (align < KMEM_ALIGN)
3310 		cflags |= KMC_NOTOUCH;
3311 
3312 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum)
3313 		panic("kmem_cache_create: bad alignment %lu", align);
3314 
3315 	mutex_enter(&kmem_flags_lock);
3316 	if (kmem_flags & KMF_RANDOMIZE)
3317 		kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3318 		    KMF_RANDOMIZE;
3319 	cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3320 	mutex_exit(&kmem_flags_lock);
3321 
3322 	/*
3323 	 * Make sure all the various flags are reasonable.
3324 	 */
3325 	ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3326 
3327 	if (cp->cache_flags & KMF_LITE) {
3328 		if (bufsize >= kmem_lite_minsize &&
3329 		    align <= kmem_lite_maxalign &&
3330 		    P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3331 			cp->cache_flags |= KMF_BUFTAG;
3332 			cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3333 		} else {
3334 			cp->cache_flags &= ~KMF_DEBUG;
3335 		}
3336 	}
3337 
3338 	if (cp->cache_flags & KMF_DEADBEEF)
3339 		cp->cache_flags |= KMF_REDZONE;
3340 
3341 	if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3342 		cp->cache_flags |= KMF_NOMAGAZINE;
3343 
3344 	if (cflags & KMC_NODEBUG)
3345 		cp->cache_flags &= ~KMF_DEBUG;
3346 
3347 	if (cflags & KMC_NOTOUCH)
3348 		cp->cache_flags &= ~KMF_TOUCH;
3349 
3350 	if (cflags & KMC_NOHASH)
3351 		cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3352 
3353 	if (cflags & KMC_NOMAGAZINE)
3354 		cp->cache_flags |= KMF_NOMAGAZINE;
3355 
3356 	if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3357 		cp->cache_flags |= KMF_REDZONE;
3358 
3359 	if (!(cp->cache_flags & KMF_AUDIT))
3360 		cp->cache_flags &= ~KMF_CONTENTS;
3361 
3362 	if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3363 	    !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3364 		cp->cache_flags |= KMF_FIREWALL;
3365 
3366 	if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3367 		cp->cache_flags &= ~KMF_FIREWALL;
3368 
3369 	if (cp->cache_flags & KMF_FIREWALL) {
3370 		cp->cache_flags &= ~KMF_BUFTAG;
3371 		cp->cache_flags |= KMF_NOMAGAZINE;
3372 		ASSERT(vmp == kmem_default_arena);
3373 		vmp = kmem_firewall_arena;
3374 	}
3375 
3376 	/*
3377 	 * Set cache properties.
3378 	 */
3379 	(void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3380 	strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3381 	cp->cache_bufsize = bufsize;
3382 	cp->cache_align = align;
3383 	cp->cache_constructor = constructor;
3384 	cp->cache_destructor = destructor;
3385 	cp->cache_reclaim = reclaim;
3386 	cp->cache_private = private;
3387 	cp->cache_arena = vmp;
3388 	cp->cache_cflags = cflags;
3389 
3390 	/*
3391 	 * Determine the chunk size.
3392 	 */
3393 	chunksize = bufsize;
3394 
3395 	if (align >= KMEM_ALIGN) {
3396 		chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3397 		cp->cache_bufctl = chunksize - KMEM_ALIGN;
3398 	}
3399 
3400 	if (cp->cache_flags & KMF_BUFTAG) {
3401 		cp->cache_bufctl = chunksize;
3402 		cp->cache_buftag = chunksize;
3403 		if (cp->cache_flags & KMF_LITE)
3404 			chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3405 		else
3406 			chunksize += sizeof (kmem_buftag_t);
3407 	}
3408 
3409 	if (cp->cache_flags & KMF_DEADBEEF) {
3410 		cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3411 		if (cp->cache_flags & KMF_LITE)
3412 			cp->cache_verify = sizeof (uint64_t);
3413 	}
3414 
3415 	cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3416 
3417 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3418 
3419 	/*
3420 	 * Now that we know the chunk size, determine the optimal slab size.
3421 	 */
3422 	if (vmp == kmem_firewall_arena) {
3423 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3424 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
3425 		cp->cache_maxcolor = cp->cache_mincolor;
3426 		cp->cache_flags |= KMF_HASH;
3427 		ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3428 	} else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3429 	    !(cp->cache_flags & KMF_AUDIT) &&
3430 	    chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3431 		cp->cache_slabsize = vmp->vm_quantum;
3432 		cp->cache_mincolor = 0;
3433 		cp->cache_maxcolor =
3434 		    (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3435 		ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3436 		ASSERT(!(cp->cache_flags & KMF_AUDIT));
3437 	} else {
3438 		size_t chunks, bestfit, waste, slabsize;
3439 		size_t minwaste = LONG_MAX;
3440 
3441 		for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3442 			slabsize = P2ROUNDUP(chunksize * chunks,
3443 			    vmp->vm_quantum);
3444 			chunks = slabsize / chunksize;
3445 			waste = (slabsize % chunksize) / chunks;
3446 			if (waste < minwaste) {
3447 				minwaste = waste;
3448 				bestfit = slabsize;
3449 			}
3450 		}
3451 		if (cflags & KMC_QCACHE)
3452 			bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3453 		cp->cache_slabsize = bestfit;
3454 		cp->cache_mincolor = 0;
3455 		cp->cache_maxcolor = bestfit % chunksize;
3456 		cp->cache_flags |= KMF_HASH;
3457 	}
3458 
3459 	cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3460 	cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3461 
3462 	if (cp->cache_flags & KMF_HASH) {
3463 		ASSERT(!(cflags & KMC_NOHASH));
3464 		cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3465 		    kmem_bufctl_audit_cache : kmem_bufctl_cache;
3466 	}
3467 
3468 	if (cp->cache_maxcolor >= vmp->vm_quantum)
3469 		cp->cache_maxcolor = vmp->vm_quantum - 1;
3470 
3471 	cp->cache_color = cp->cache_mincolor;
3472 
3473 	/*
3474 	 * Initialize the rest of the slab layer.
3475 	 */
3476 	mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3477 
3478 	avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3479 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3480 	/* LINTED: E_TRUE_LOGICAL_EXPR */
3481 	ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3482 	/* reuse partial slab AVL linkage for complete slab list linkage */
3483 	list_create(&cp->cache_complete_slabs,
3484 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3485 
3486 	if (cp->cache_flags & KMF_HASH) {
3487 		cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3488 		    KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3489 		bzero(cp->cache_hash_table,
3490 		    KMEM_HASH_INITIAL * sizeof (void *));
3491 		cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3492