xref: /titanic_41/usr/src/uts/common/os/kmem.c (revision 86b2a80075734c8ecdc898b083ef73b81254096b)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
24  */
25 
26 /*
27  * Kernel memory allocator, as described in the following two papers and a
28  * statement about the consolidator:
29  *
30  * Jeff Bonwick,
31  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
32  * Proceedings of the Summer 1994 Usenix Conference.
33  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
34  *
35  * Jeff Bonwick and Jonathan Adams,
36  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
37  * Arbitrary Resources.
38  * Proceedings of the 2001 Usenix Conference.
39  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
40  *
41  * kmem Slab Consolidator Big Theory Statement:
42  *
43  * 1. Motivation
44  *
45  * As stated in Bonwick94, slabs provide the following advantages over other
46  * allocation structures in terms of memory fragmentation:
47  *
48  *  - Internal fragmentation (per-buffer wasted space) is minimal.
49  *  - Severe external fragmentation (unused buffers on the free list) is
50  *    unlikely.
51  *
52  * Segregating objects by size eliminates one source of external fragmentation,
53  * and according to Bonwick:
54  *
55  *   The other reason that slabs reduce external fragmentation is that all
56  *   objects in a slab are of the same type, so they have the same lifetime
57  *   distribution. The resulting segregation of short-lived and long-lived
58  *   objects at slab granularity reduces the likelihood of an entire page being
59  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
60  *
61  * While unlikely, severe external fragmentation remains possible. Clients that
62  * allocate both short- and long-lived objects from the same cache cannot
63  * anticipate the distribution of long-lived objects within the allocator's slab
64  * implementation. Even a small percentage of long-lived objects distributed
65  * randomly across many slabs can lead to a worst case scenario where the client
66  * frees the majority of its objects and the system gets back almost none of the
67  * slabs. Despite the client doing what it reasonably can to help the system
68  * reclaim memory, the allocator cannot shake free enough slabs because of
69  * lonely allocations stubbornly hanging on. Although the allocator is in a
70  * position to diagnose the fragmentation, there is nothing that the allocator
71  * by itself can do about it. It only takes a single allocated object to prevent
72  * an entire slab from being reclaimed, and any object handed out by
73  * kmem_cache_alloc() is by definition in the client's control. Conversely,
74  * although the client is in a position to move a long-lived object, it has no
75  * way of knowing if the object is causing fragmentation, and if so, where to
76  * move it. A solution necessarily requires further cooperation between the
77  * allocator and the client.
78  *
79  * 2. Move Callback
80  *
81  * The kmem slab consolidator therefore adds a move callback to the
82  * allocator/client interface, improving worst-case external fragmentation in
83  * kmem caches that supply a function to move objects from one memory location
84  * to another. In a situation of low memory kmem attempts to consolidate all of
85  * a cache's slabs at once; otherwise it works slowly to bring external
86  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
87  * thereby helping to avoid a low memory situation in the future.
88  *
89  * The callback has the following signature:
90  *
91  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
92  *
93  * It supplies the kmem client with two addresses: the allocated object that
94  * kmem wants to move and a buffer selected by kmem for the client to use as the
95  * copy destination. The callback is kmem's way of saying "Please get off of
96  * this buffer and use this one instead." kmem knows where it wants to move the
97  * object in order to best reduce fragmentation. All the client needs to know
98  * about the second argument (void *new) is that it is an allocated, constructed
99  * object ready to take the contents of the old object. When the move function
100  * is called, the system is likely to be low on memory, and the new object
101  * spares the client from having to worry about allocating memory for the
102  * requested move. The third argument supplies the size of the object, in case a
103  * single move function handles multiple caches whose objects differ only in
104  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
105  * user argument passed to the constructor, destructor, and reclaim functions is
106  * also passed to the move callback.
107  *
108  * 2.1 Setting the Move Callback
109  *
110  * The client sets the move callback after creating the cache and before
111  * allocating from it:
112  *
113  *	object_cache = kmem_cache_create(...);
114  *      kmem_cache_set_move(object_cache, object_move);
115  *
116  * 2.2 Move Callback Return Values
117  *
118  * Only the client knows about its own data and when is a good time to move it.
119  * The client is cooperating with kmem to return unused memory to the system,
120  * and kmem respectfully accepts this help at the client's convenience. When
121  * asked to move an object, the client can respond with any of the following:
122  *
123  *   typedef enum kmem_cbrc {
124  *           KMEM_CBRC_YES,
125  *           KMEM_CBRC_NO,
126  *           KMEM_CBRC_LATER,
127  *           KMEM_CBRC_DONT_NEED,
128  *           KMEM_CBRC_DONT_KNOW
129  *   } kmem_cbrc_t;
130  *
131  * The client must not explicitly kmem_cache_free() either of the objects passed
132  * to the callback, since kmem wants to free them directly to the slab layer
133  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
134  * objects to free:
135  *
136  *       YES: (Did it) The client moved the object, so kmem frees the old one.
137  *        NO: (Never) The client refused, so kmem frees the new object (the
138  *            unused copy destination). kmem also marks the slab of the old
139  *            object so as not to bother the client with further callbacks for
140  *            that object as long as the slab remains on the partial slab list.
141  *            (The system won't be getting the slab back as long as the
142  *            immovable object holds it hostage, so there's no point in moving
143  *            any of its objects.)
144  *     LATER: The client is using the object and cannot move it now, so kmem
145  *            frees the new object (the unused copy destination). kmem still
146  *            attempts to move other objects off the slab, since it expects to
147  *            succeed in clearing the slab in a later callback. The client
148  *            should use LATER instead of NO if the object is likely to become
149  *            movable very soon.
150  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
151  *            with the new object (the unused copy destination). This response
152  *            is the client's opportunity to be a model citizen and give back as
153  *            much as it can.
154  * DONT_KNOW: The client does not know about the object because
155  *            a) the client has just allocated the object and not yet put it
156  *               wherever it expects to find known objects
157  *            b) the client has removed the object from wherever it expects to
158  *               find known objects and is about to free it, or
159  *            c) the client has freed the object.
160  *            In all these cases (a, b, and c) kmem frees the new object (the
161  *            unused copy destination) and searches for the old object in the
162  *            magazine layer. If found, the object is removed from the magazine
163  *            layer and freed to the slab layer so it will no longer hold the
164  *            slab hostage.
165  *
166  * 2.3 Object States
167  *
168  * Neither kmem nor the client can be assumed to know the object's whereabouts
169  * at the time of the callback. An object belonging to a kmem cache may be in
170  * any of the following states:
171  *
172  * 1. Uninitialized on the slab
173  * 2. Allocated from the slab but not constructed (still uninitialized)
174  * 3. Allocated from the slab, constructed, but not yet ready for business
175  *    (not in a valid state for the move callback)
176  * 4. In use (valid and known to the client)
177  * 5. About to be freed (no longer in a valid state for the move callback)
178  * 6. Freed to a magazine (still constructed)
179  * 7. Allocated from a magazine, not yet ready for business (not in a valid
180  *    state for the move callback), and about to return to state #4
181  * 8. Deconstructed on a magazine that is about to be freed
182  * 9. Freed to the slab
183  *
184  * Since the move callback may be called at any time while the object is in any
185  * of the above states (except state #1), the client needs a safe way to
186  * determine whether or not it knows about the object. Specifically, the client
187  * needs to know whether or not the object is in state #4, the only state in
188  * which a move is valid. If the object is in any other state, the client should
189  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
190  * the object's fields.
191  *
192  * Note that although an object may be in state #4 when kmem initiates the move
193  * request, the object may no longer be in that state by the time kmem actually
194  * calls the move function. Not only does the client free objects
195  * asynchronously, kmem itself puts move requests on a queue where thay are
196  * pending until kmem processes them from another context. Also, objects freed
197  * to a magazine appear allocated from the point of view of the slab layer, so
198  * kmem may even initiate requests for objects in a state other than state #4.
199  *
200  * 2.3.1 Magazine Layer
201  *
202  * An important insight revealed by the states listed above is that the magazine
203  * layer is populated only by kmem_cache_free(). Magazines of constructed
204  * objects are never populated directly from the slab layer (which contains raw,
205  * unconstructed objects). Whenever an allocation request cannot be satisfied
206  * from the magazine layer, the magazines are bypassed and the request is
207  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
208  * the object constructor only when allocating from the slab layer, and only in
209  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
210  * the move callback. kmem does not preconstruct objects in anticipation of
211  * kmem_cache_alloc().
212  *
213  * 2.3.2 Object Constructor and Destructor
214  *
215  * If the client supplies a destructor, it must be valid to call the destructor
216  * on a newly created object (immediately after the constructor).
217  *
218  * 2.4 Recognizing Known Objects
219  *
220  * There is a simple test to determine safely whether or not the client knows
221  * about a given object in the move callback. It relies on the fact that kmem
222  * guarantees that the object of the move callback has only been touched by the
223  * client itself or else by kmem. kmem does this by ensuring that none of the
224  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
225  * callback is pending. When the last object on a slab is freed, if there is a
226  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
227  * slabs on that list until all pending callbacks are completed. That way,
228  * clients can be certain that the object of a move callback is in one of the
229  * states listed above, making it possible to distinguish known objects (in
230  * state #4) using the two low order bits of any pointer member (with the
231  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
232  * platforms).
233  *
234  * The test works as long as the client always transitions objects from state #4
235  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
236  * order bit of the client-designated pointer member. Since kmem only writes
237  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
238  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
239  * guaranteed to set at least one of the two low order bits. Therefore, given an
240  * object with a back pointer to a 'container_t *o_container', the client can
241  * test
242  *
243  *      container_t *container = object->o_container;
244  *      if ((uintptr_t)container & 0x3) {
245  *              return (KMEM_CBRC_DONT_KNOW);
246  *      }
247  *
248  * Typically, an object will have a pointer to some structure with a list or
249  * hash where objects from the cache are kept while in use. Assuming that the
250  * client has some way of knowing that the container structure is valid and will
251  * not go away during the move, and assuming that the structure includes a lock
252  * to protect whatever collection is used, then the client would continue as
253  * follows:
254  *
255  *	// Ensure that the container structure does not go away.
256  *      if (container_hold(container) == 0) {
257  *              return (KMEM_CBRC_DONT_KNOW);
258  *      }
259  *      mutex_enter(&container->c_objects_lock);
260  *      if (container != object->o_container) {
261  *              mutex_exit(&container->c_objects_lock);
262  *              container_rele(container);
263  *              return (KMEM_CBRC_DONT_KNOW);
264  *      }
265  *
266  * At this point the client knows that the object cannot be freed as long as
267  * c_objects_lock is held. Note that after acquiring the lock, the client must
268  * recheck the o_container pointer in case the object was removed just before
269  * acquiring the lock.
270  *
271  * When the client is about to free an object, it must first remove that object
272  * from the list, hash, or other structure where it is kept. At that time, to
273  * mark the object so it can be distinguished from the remaining, known objects,
274  * the client sets the designated low order bit:
275  *
276  *      mutex_enter(&container->c_objects_lock);
277  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
278  *      list_remove(&container->c_objects, object);
279  *      mutex_exit(&container->c_objects_lock);
280  *
281  * In the common case, the object is freed to the magazine layer, where it may
282  * be reused on a subsequent allocation without the overhead of calling the
283  * constructor. While in the magazine it appears allocated from the point of
284  * view of the slab layer, making it a candidate for the move callback. Most
285  * objects unrecognized by the client in the move callback fall into this
286  * category and are cheaply distinguished from known objects by the test
287  * described earlier. Since recognition is cheap for the client, and searching
288  * magazines is expensive for kmem, kmem defers searching until the client first
289  * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
290  * elsewhere does what it can to avoid bothering the client unnecessarily.
291  *
292  * Invalidating the designated pointer member before freeing the object marks
293  * the object to be avoided in the callback, and conversely, assigning a valid
294  * value to the designated pointer member after allocating the object makes the
295  * object fair game for the callback:
296  *
297  *      ... allocate object ...
298  *      ... set any initial state not set by the constructor ...
299  *
300  *      mutex_enter(&container->c_objects_lock);
301  *      list_insert_tail(&container->c_objects, object);
302  *      membar_producer();
303  *      object->o_container = container;
304  *      mutex_exit(&container->c_objects_lock);
305  *
306  * Note that everything else must be valid before setting o_container makes the
307  * object fair game for the move callback. The membar_producer() call ensures
308  * that all the object's state is written to memory before setting the pointer
309  * that transitions the object from state #3 or #7 (allocated, constructed, not
310  * yet in use) to state #4 (in use, valid). That's important because the move
311  * function has to check the validity of the pointer before it can safely
312  * acquire the lock protecting the collection where it expects to find known
313  * objects.
314  *
315  * This method of distinguishing known objects observes the usual symmetry:
316  * invalidating the designated pointer is the first thing the client does before
317  * freeing the object, and setting the designated pointer is the last thing the
318  * client does after allocating the object. Of course, the client is not
319  * required to use this method. Fundamentally, how the client recognizes known
320  * objects is completely up to the client, but this method is recommended as an
321  * efficient and safe way to take advantage of the guarantees made by kmem. If
322  * the entire object is arbitrary data without any markable bits from a suitable
323  * pointer member, then the client must find some other method, such as
324  * searching a hash table of known objects.
325  *
326  * 2.5 Preventing Objects From Moving
327  *
328  * Besides a way to distinguish known objects, the other thing that the client
329  * needs is a strategy to ensure that an object will not move while the client
330  * is actively using it. The details of satisfying this requirement tend to be
331  * highly cache-specific. It might seem that the same rules that let a client
332  * remove an object safely should also decide when an object can be moved
333  * safely. However, any object state that makes a removal attempt invalid is
334  * likely to be long-lasting for objects that the client does not expect to
335  * remove. kmem knows nothing about the object state and is equally likely (from
336  * the client's point of view) to request a move for any object in the cache,
337  * whether prepared for removal or not. Even a low percentage of objects stuck
338  * in place by unremovability will defeat the consolidator if the stuck objects
339  * are the same long-lived allocations likely to hold slabs hostage.
340  * Fundamentally, the consolidator is not aimed at common cases. Severe external
341  * fragmentation is a worst case scenario manifested as sparsely allocated
342  * slabs, by definition a low percentage of the cache's objects. When deciding
343  * what makes an object movable, keep in mind the goal of the consolidator: to
344  * bring worst-case external fragmentation within the limits guaranteed for
345  * internal fragmentation. Removability is a poor criterion if it is likely to
346  * exclude more than an insignificant percentage of objects for long periods of
347  * time.
348  *
349  * A tricky general solution exists, and it has the advantage of letting you
350  * move any object at almost any moment, practically eliminating the likelihood
351  * that an object can hold a slab hostage. However, if there is a cache-specific
352  * way to ensure that an object is not actively in use in the vast majority of
353  * cases, a simpler solution that leverages this cache-specific knowledge is
354  * preferred.
355  *
356  * 2.5.1 Cache-Specific Solution
357  *
358  * As an example of a cache-specific solution, the ZFS znode cache takes
359  * advantage of the fact that the vast majority of znodes are only being
360  * referenced from the DNLC. (A typical case might be a few hundred in active
361  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
362  * client has established that it recognizes the znode and can access its fields
363  * safely (using the method described earlier), it then tests whether the znode
364  * is referenced by anything other than the DNLC. If so, it assumes that the
365  * znode may be in active use and is unsafe to move, so it drops its locks and
366  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
367  * else znodes are used, no change is needed to protect against the possibility
368  * of the znode moving. The disadvantage is that it remains possible for an
369  * application to hold a znode slab hostage with an open file descriptor.
370  * However, this case ought to be rare and the consolidator has a way to deal
371  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
372  * object, kmem eventually stops believing it and treats the slab as if the
373  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
374  * then focus on getting it off of the partial slab list by allocating rather
375  * than freeing all of its objects. (Either way of getting a slab off the
376  * free list reduces fragmentation.)
377  *
378  * 2.5.2 General Solution
379  *
380  * The general solution, on the other hand, requires an explicit hold everywhere
381  * the object is used to prevent it from moving. To keep the client locking
382  * strategy as uncomplicated as possible, kmem guarantees the simplifying
383  * assumption that move callbacks are sequential, even across multiple caches.
384  * Internally, a global queue processed by a single thread supports all caches
385  * implementing the callback function. No matter how many caches supply a move
386  * function, the consolidator never moves more than one object at a time, so the
387  * client does not have to worry about tricky lock ordering involving several
388  * related objects from different kmem caches.
389  *
390  * The general solution implements the explicit hold as a read-write lock, which
391  * allows multiple readers to access an object from the cache simultaneously
392  * while a single writer is excluded from moving it. A single rwlock for the
393  * entire cache would lock out all threads from using any of the cache's objects
394  * even though only a single object is being moved, so to reduce contention,
395  * the client can fan out the single rwlock into an array of rwlocks hashed by
396  * the object address, making it probable that moving one object will not
397  * prevent other threads from using a different object. The rwlock cannot be a
398  * member of the object itself, because the possibility of the object moving
399  * makes it unsafe to access any of the object's fields until the lock is
400  * acquired.
401  *
402  * Assuming a small, fixed number of locks, it's possible that multiple objects
403  * will hash to the same lock. A thread that needs to use multiple objects in
404  * the same function may acquire the same lock multiple times. Since rwlocks are
405  * reentrant for readers, and since there is never more than a single writer at
406  * a time (assuming that the client acquires the lock as a writer only when
407  * moving an object inside the callback), there would seem to be no problem.
408  * However, a client locking multiple objects in the same function must handle
409  * one case of potential deadlock: Assume that thread A needs to prevent both
410  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
411  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
412  * same lock, that thread A will acquire the lock for object 1 as a reader
413  * before thread B sets the lock's write-wanted bit, preventing thread A from
414  * reacquiring the lock for object 2 as a reader. Unable to make forward
415  * progress, thread A will never release the lock for object 1, resulting in
416  * deadlock.
417  *
418  * There are two ways of avoiding the deadlock just described. The first is to
419  * use rw_tryenter() rather than rw_enter() in the callback function when
420  * attempting to acquire the lock as a writer. If tryenter discovers that the
421  * same object (or another object hashed to the same lock) is already in use, it
422  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
423  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
424  * since it allows a thread to acquire the lock as a reader in spite of a
425  * waiting writer. This second approach insists on moving the object now, no
426  * matter how many readers the move function must wait for in order to do so,
427  * and could delay the completion of the callback indefinitely (blocking
428  * callbacks to other clients). In practice, a less insistent callback using
429  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
430  * little reason to use anything else.
431  *
432  * Avoiding deadlock is not the only problem that an implementation using an
433  * explicit hold needs to solve. Locking the object in the first place (to
434  * prevent it from moving) remains a problem, since the object could move
435  * between the time you obtain a pointer to the object and the time you acquire
436  * the rwlock hashed to that pointer value. Therefore the client needs to
437  * recheck the value of the pointer after acquiring the lock, drop the lock if
438  * the value has changed, and try again. This requires a level of indirection:
439  * something that points to the object rather than the object itself, that the
440  * client can access safely while attempting to acquire the lock. (The object
441  * itself cannot be referenced safely because it can move at any time.)
442  * The following lock-acquisition function takes whatever is safe to reference
443  * (arg), follows its pointer to the object (using function f), and tries as
444  * often as necessary to acquire the hashed lock and verify that the object
445  * still has not moved:
446  *
447  *      object_t *
448  *      object_hold(object_f f, void *arg)
449  *      {
450  *              object_t *op;
451  *
452  *              op = f(arg);
453  *              if (op == NULL) {
454  *                      return (NULL);
455  *              }
456  *
457  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
458  *              while (op != f(arg)) {
459  *                      rw_exit(OBJECT_RWLOCK(op));
460  *                      op = f(arg);
461  *                      if (op == NULL) {
462  *                              break;
463  *                      }
464  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
465  *              }
466  *
467  *              return (op);
468  *      }
469  *
470  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
471  * lock reacquisition loop, while necessary, almost never executes. The function
472  * pointer f (used to obtain the object pointer from arg) has the following type
473  * definition:
474  *
475  *      typedef object_t *(*object_f)(void *arg);
476  *
477  * An object_f implementation is likely to be as simple as accessing a structure
478  * member:
479  *
480  *      object_t *
481  *      s_object(void *arg)
482  *      {
483  *              something_t *sp = arg;
484  *              return (sp->s_object);
485  *      }
486  *
487  * The flexibility of a function pointer allows the path to the object to be
488  * arbitrarily complex and also supports the notion that depending on where you
489  * are using the object, you may need to get it from someplace different.
490  *
491  * The function that releases the explicit hold is simpler because it does not
492  * have to worry about the object moving:
493  *
494  *      void
495  *      object_rele(object_t *op)
496  *      {
497  *              rw_exit(OBJECT_RWLOCK(op));
498  *      }
499  *
500  * The caller is spared these details so that obtaining and releasing an
501  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
502  * of object_hold() only needs to know that the returned object pointer is valid
503  * if not NULL and that the object will not move until released.
504  *
505  * Although object_hold() prevents an object from moving, it does not prevent it
506  * from being freed. The caller must take measures before calling object_hold()
507  * (afterwards is too late) to ensure that the held object cannot be freed. The
508  * caller must do so without accessing the unsafe object reference, so any lock
509  * or reference count used to ensure the continued existence of the object must
510  * live outside the object itself.
511  *
512  * Obtaining a new object is a special case where an explicit hold is impossible
513  * for the caller. Any function that returns a newly allocated object (either as
514  * a return value, or as an in-out paramter) must return it already held; after
515  * the caller gets it is too late, since the object cannot be safely accessed
516  * without the level of indirection described earlier. The following
517  * object_alloc() example uses the same code shown earlier to transition a new
518  * object into the state of being recognized (by the client) as a known object.
519  * The function must acquire the hold (rw_enter) before that state transition
520  * makes the object movable:
521  *
522  *      static object_t *
523  *      object_alloc(container_t *container)
524  *      {
525  *              object_t *object = kmem_cache_alloc(object_cache, 0);
526  *              ... set any initial state not set by the constructor ...
527  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
528  *              mutex_enter(&container->c_objects_lock);
529  *              list_insert_tail(&container->c_objects, object);
530  *              membar_producer();
531  *              object->o_container = container;
532  *              mutex_exit(&container->c_objects_lock);
533  *              return (object);
534  *      }
535  *
536  * Functions that implicitly acquire an object hold (any function that calls
537  * object_alloc() to supply an object for the caller) need to be carefully noted
538  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
539  * prevent all objects hashed to the affected rwlocks from ever being moved.
540  *
541  * The pointer to a held object can be hashed to the holding rwlock even after
542  * the object has been freed. Although it is possible to release the hold
543  * after freeing the object, you may decide to release the hold implicitly in
544  * whatever function frees the object, so as to release the hold as soon as
545  * possible, and for the sake of symmetry with the function that implicitly
546  * acquires the hold when it allocates the object. Here, object_free() releases
547  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
548  * matching pair with object_hold():
549  *
550  *      void
551  *      object_free(object_t *object)
552  *      {
553  *              container_t *container;
554  *
555  *              ASSERT(object_held(object));
556  *              container = object->o_container;
557  *              mutex_enter(&container->c_objects_lock);
558  *              object->o_container =
559  *                  (void *)((uintptr_t)object->o_container | 0x1);
560  *              list_remove(&container->c_objects, object);
561  *              mutex_exit(&container->c_objects_lock);
562  *              object_rele(object);
563  *              kmem_cache_free(object_cache, object);
564  *      }
565  *
566  * Note that object_free() cannot safely accept an object pointer as an argument
567  * unless the object is already held. Any function that calls object_free()
568  * needs to be carefully noted since it similarly forms a matching pair with
569  * object_hold().
570  *
571  * To complete the picture, the following callback function implements the
572  * general solution by moving objects only if they are currently unheld:
573  *
574  *      static kmem_cbrc_t
575  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
576  *      {
577  *              object_t *op = buf, *np = newbuf;
578  *              container_t *container;
579  *
580  *              container = op->o_container;
581  *              if ((uintptr_t)container & 0x3) {
582  *                      return (KMEM_CBRC_DONT_KNOW);
583  *              }
584  *
585  *	        // Ensure that the container structure does not go away.
586  *              if (container_hold(container) == 0) {
587  *                      return (KMEM_CBRC_DONT_KNOW);
588  *              }
589  *
590  *              mutex_enter(&container->c_objects_lock);
591  *              if (container != op->o_container) {
592  *                      mutex_exit(&container->c_objects_lock);
593  *                      container_rele(container);
594  *                      return (KMEM_CBRC_DONT_KNOW);
595  *              }
596  *
597  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
598  *                      mutex_exit(&container->c_objects_lock);
599  *                      container_rele(container);
600  *                      return (KMEM_CBRC_LATER);
601  *              }
602  *
603  *              object_move_impl(op, np); // critical section
604  *              rw_exit(OBJECT_RWLOCK(op));
605  *
606  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
607  *              list_link_replace(&op->o_link_node, &np->o_link_node);
608  *              mutex_exit(&container->c_objects_lock);
609  *              container_rele(container);
610  *              return (KMEM_CBRC_YES);
611  *      }
612  *
613  * Note that object_move() must invalidate the designated o_container pointer of
614  * the old object in the same way that object_free() does, since kmem will free
615  * the object in response to the KMEM_CBRC_YES return value.
616  *
617  * The lock order in object_move() differs from object_alloc(), which locks
618  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
619  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
620  * not a problem. Holding the lock on the object list in the example above
621  * through the entire callback not only prevents the object from going away, it
622  * also allows you to lock the list elsewhere and know that none of its elements
623  * will move during iteration.
624  *
625  * Adding an explicit hold everywhere an object from the cache is used is tricky
626  * and involves much more change to client code than a cache-specific solution
627  * that leverages existing state to decide whether or not an object is
628  * movable. However, this approach has the advantage that no object remains
629  * immovable for any significant length of time, making it extremely unlikely
630  * that long-lived allocations can continue holding slabs hostage; and it works
631  * for any cache.
632  *
633  * 3. Consolidator Implementation
634  *
635  * Once the client supplies a move function that a) recognizes known objects and
636  * b) avoids moving objects that are actively in use, the remaining work is up
637  * to the consolidator to decide which objects to move and when to issue
638  * callbacks.
639  *
640  * The consolidator relies on the fact that a cache's slabs are ordered by
641  * usage. Each slab has a fixed number of objects. Depending on the slab's
642  * "color" (the offset of the first object from the beginning of the slab;
643  * offsets are staggered to mitigate false sharing of cache lines) it is either
644  * the maximum number of objects per slab determined at cache creation time or
645  * else the number closest to the maximum that fits within the space remaining
646  * after the initial offset. A completely allocated slab may contribute some
647  * internal fragmentation (per-slab overhead) but no external fragmentation, so
648  * it is of no interest to the consolidator. At the other extreme, slabs whose
649  * objects have all been freed to the slab are released to the virtual memory
650  * (VM) subsystem (objects freed to magazines are still allocated as far as the
651  * slab is concerned). External fragmentation exists when there are slabs
652  * somewhere between these extremes. A partial slab has at least one but not all
653  * of its objects allocated. The more partial slabs, and the fewer allocated
654  * objects on each of them, the higher the fragmentation. Hence the
655  * consolidator's overall strategy is to reduce the number of partial slabs by
656  * moving allocated objects from the least allocated slabs to the most allocated
657  * slabs.
658  *
659  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
660  * slabs are kept separately in an unordered list. Since the majority of slabs
661  * tend to be completely allocated (a typical unfragmented cache may have
662  * thousands of complete slabs and only a single partial slab), separating
663  * complete slabs improves the efficiency of partial slab ordering, since the
664  * complete slabs do not affect the depth or balance of the AVL tree. This
665  * ordered sequence of partial slabs acts as a "free list" supplying objects for
666  * allocation requests.
667  *
668  * Objects are always allocated from the first partial slab in the free list,
669  * where the allocation is most likely to eliminate a partial slab (by
670  * completely allocating it). Conversely, when a single object from a completely
671  * allocated slab is freed to the slab, that slab is added to the front of the
672  * free list. Since most free list activity involves highly allocated slabs
673  * coming and going at the front of the list, slabs tend naturally toward the
674  * ideal order: highly allocated at the front, sparsely allocated at the back.
675  * Slabs with few allocated objects are likely to become completely free if they
676  * keep a safe distance away from the front of the free list. Slab misorders
677  * interfere with the natural tendency of slabs to become completely free or
678  * completely allocated. For example, a slab with a single allocated object
679  * needs only a single free to escape the cache; its natural desire is
680  * frustrated when it finds itself at the front of the list where a second
681  * allocation happens just before the free could have released it. Another slab
682  * with all but one object allocated might have supplied the buffer instead, so
683  * that both (as opposed to neither) of the slabs would have been taken off the
684  * free list.
685  *
686  * Although slabs tend naturally toward the ideal order, misorders allowed by a
687  * simple list implementation defeat the consolidator's strategy of merging
688  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
689  * needs another way to fix misorders to optimize its callback strategy. One
690  * approach is to periodically scan a limited number of slabs, advancing a
691  * marker to hold the current scan position, and to move extreme misorders to
692  * the front or back of the free list and to the front or back of the current
693  * scan range. By making consecutive scan ranges overlap by one slab, the least
694  * allocated slab in the current range can be carried along from the end of one
695  * scan to the start of the next.
696  *
697  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
698  * task, however. Since most of the cache's activity is in the magazine layer,
699  * and allocations from the slab layer represent only a startup cost, the
700  * overhead of maintaining a balanced tree is not a significant concern compared
701  * to the opportunity of reducing complexity by eliminating the partial slab
702  * scanner just described. The overhead of an AVL tree is minimized by
703  * maintaining only partial slabs in the tree and keeping completely allocated
704  * slabs separately in a list. To avoid increasing the size of the slab
705  * structure the AVL linkage pointers are reused for the slab's list linkage,
706  * since the slab will always be either partial or complete, never stored both
707  * ways at the same time. To further minimize the overhead of the AVL tree the
708  * compare function that orders partial slabs by usage divides the range of
709  * allocated object counts into bins such that counts within the same bin are
710  * considered equal. Binning partial slabs makes it less likely that allocating
711  * or freeing a single object will change the slab's order, requiring a tree
712  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
713  * requiring some rebalancing of the tree). Allocation counts closest to
714  * completely free and completely allocated are left unbinned (finely sorted) to
715  * better support the consolidator's strategy of merging slabs at either
716  * extreme.
717  *
718  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
719  *
720  * The consolidator piggybacks on the kmem maintenance thread and is called on
721  * the same interval as kmem_cache_update(), once per cache every fifteen
722  * seconds. kmem maintains a running count of unallocated objects in the slab
723  * layer (cache_bufslab). The consolidator checks whether that number exceeds
724  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
725  * there is a significant number of slabs in the cache (arbitrarily a minimum
726  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
727  * working set are included in the assessment, and magazines in the depot are
728  * reaped if those objects would lift cache_bufslab above the fragmentation
729  * threshold. Once the consolidator decides that a cache is fragmented, it looks
730  * for a candidate slab to reclaim, starting at the end of the partial slab free
731  * list and scanning backwards. At first the consolidator is choosy: only a slab
732  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
733  * single allocated object, regardless of percentage). If there is difficulty
734  * finding a candidate slab, kmem raises the allocation threshold incrementally,
735  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
736  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
737  * even in the worst case of every slab in the cache being almost 7/8 allocated.
738  * The threshold can also be lowered incrementally when candidate slabs are easy
739  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
740  * is no longer fragmented.
741  *
742  * 3.2 Generating Callbacks
743  *
744  * Once an eligible slab is chosen, a callback is generated for every allocated
745  * object on the slab, in the hope that the client will move everything off the
746  * slab and make it reclaimable. Objects selected as move destinations are
747  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
748  * order (most allocated at the front, least allocated at the back) and a
749  * cooperative client, the consolidator will succeed in removing slabs from both
750  * ends of the free list, completely allocating on the one hand and completely
751  * freeing on the other. Objects selected as move destinations are allocated in
752  * the kmem maintenance thread where move requests are enqueued. A separate
753  * callback thread removes pending callbacks from the queue and calls the
754  * client. The separate thread ensures that client code (the move function) does
755  * not interfere with internal kmem maintenance tasks. A map of pending
756  * callbacks keyed by object address (the object to be moved) is checked to
757  * ensure that duplicate callbacks are not generated for the same object.
758  * Allocating the move destination (the object to move to) prevents subsequent
759  * callbacks from selecting the same destination as an earlier pending callback.
760  *
761  * Move requests can also be generated by kmem_cache_reap() when the system is
762  * desperate for memory and by kmem_cache_move_notify(), called by the client to
763  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
764  * The map of pending callbacks is protected by the same lock that protects the
765  * slab layer.
766  *
767  * When the system is desperate for memory, kmem does not bother to determine
768  * whether or not the cache exceeds the fragmentation threshold, but tries to
769  * consolidate as many slabs as possible. Normally, the consolidator chews
770  * slowly, one sparsely allocated slab at a time during each maintenance
771  * interval that the cache is fragmented. When desperate, the consolidator
772  * starts at the last partial slab and enqueues callbacks for every allocated
773  * object on every partial slab, working backwards until it reaches the first
774  * partial slab. The first partial slab, meanwhile, advances in pace with the
775  * consolidator as allocations to supply move destinations for the enqueued
776  * callbacks use up the highly allocated slabs at the front of the free list.
777  * Ideally, the overgrown free list collapses like an accordion, starting at
778  * both ends and ending at the center with a single partial slab.
779  *
780  * 3.3 Client Responses
781  *
782  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
783  * marks the slab that supplied the stuck object non-reclaimable and moves it to
784  * front of the free list. The slab remains marked as long as it remains on the
785  * free list, and it appears more allocated to the partial slab compare function
786  * than any unmarked slab, no matter how many of its objects are allocated.
787  * Since even one immovable object ties up the entire slab, the goal is to
788  * completely allocate any slab that cannot be completely freed. kmem does not
789  * bother generating callbacks to move objects from a marked slab unless the
790  * system is desperate.
791  *
792  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
793  * slab. If the client responds LATER too many times, kmem disbelieves and
794  * treats the response as a NO. The count is cleared when the slab is taken off
795  * the partial slab list or when the client moves one of the slab's objects.
796  *
797  * 4. Observability
798  *
799  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
800  * the ::kmem_slabs dcmd. For a complete description of the command, enter
801  * '::help kmem_slabs' at the mdb prompt.
802  */
803 
804 #include <sys/kmem_impl.h>
805 #include <sys/vmem_impl.h>
806 #include <sys/param.h>
807 #include <sys/sysmacros.h>
808 #include <sys/vm.h>
809 #include <sys/proc.h>
810 #include <sys/tuneable.h>
811 #include <sys/systm.h>
812 #include <sys/cmn_err.h>
813 #include <sys/debug.h>
814 #include <sys/sdt.h>
815 #include <sys/mutex.h>
816 #include <sys/bitmap.h>
817 #include <sys/atomic.h>
818 #include <sys/kobj.h>
819 #include <sys/disp.h>
820 #include <vm/seg_kmem.h>
821 #include <sys/log.h>
822 #include <sys/callb.h>
823 #include <sys/taskq.h>
824 #include <sys/modctl.h>
825 #include <sys/reboot.h>
826 #include <sys/id32.h>
827 #include <sys/zone.h>
828 #include <sys/netstack.h>
829 #ifdef	DEBUG
830 #include <sys/random.h>
831 #endif
832 
833 extern void streams_msg_init(void);
834 extern int segkp_fromheap;
835 extern void segkp_cache_free(void);
836 extern int callout_init_done;
837 
838 struct kmem_cache_kstat {
839 	kstat_named_t	kmc_buf_size;
840 	kstat_named_t	kmc_align;
841 	kstat_named_t	kmc_chunk_size;
842 	kstat_named_t	kmc_slab_size;
843 	kstat_named_t	kmc_alloc;
844 	kstat_named_t	kmc_alloc_fail;
845 	kstat_named_t	kmc_free;
846 	kstat_named_t	kmc_depot_alloc;
847 	kstat_named_t	kmc_depot_free;
848 	kstat_named_t	kmc_depot_contention;
849 	kstat_named_t	kmc_slab_alloc;
850 	kstat_named_t	kmc_slab_free;
851 	kstat_named_t	kmc_buf_constructed;
852 	kstat_named_t	kmc_buf_avail;
853 	kstat_named_t	kmc_buf_inuse;
854 	kstat_named_t	kmc_buf_total;
855 	kstat_named_t	kmc_buf_max;
856 	kstat_named_t	kmc_slab_create;
857 	kstat_named_t	kmc_slab_destroy;
858 	kstat_named_t	kmc_vmem_source;
859 	kstat_named_t	kmc_hash_size;
860 	kstat_named_t	kmc_hash_lookup_depth;
861 	kstat_named_t	kmc_hash_rescale;
862 	kstat_named_t	kmc_full_magazines;
863 	kstat_named_t	kmc_empty_magazines;
864 	kstat_named_t	kmc_magazine_size;
865 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
866 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
867 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
868 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
869 	kstat_named_t	kmc_move_yes;
870 	kstat_named_t	kmc_move_no;
871 	kstat_named_t	kmc_move_later;
872 	kstat_named_t	kmc_move_dont_need;
873 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
874 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
875 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
876 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
877 } kmem_cache_kstat = {
878 	{ "buf_size",		KSTAT_DATA_UINT64 },
879 	{ "align",		KSTAT_DATA_UINT64 },
880 	{ "chunk_size",		KSTAT_DATA_UINT64 },
881 	{ "slab_size",		KSTAT_DATA_UINT64 },
882 	{ "alloc",		KSTAT_DATA_UINT64 },
883 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
884 	{ "free",		KSTAT_DATA_UINT64 },
885 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
886 	{ "depot_free",		KSTAT_DATA_UINT64 },
887 	{ "depot_contention",	KSTAT_DATA_UINT64 },
888 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
889 	{ "slab_free",		KSTAT_DATA_UINT64 },
890 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
891 	{ "buf_avail",		KSTAT_DATA_UINT64 },
892 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
893 	{ "buf_total",		KSTAT_DATA_UINT64 },
894 	{ "buf_max",		KSTAT_DATA_UINT64 },
895 	{ "slab_create",	KSTAT_DATA_UINT64 },
896 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
897 	{ "vmem_source",	KSTAT_DATA_UINT64 },
898 	{ "hash_size",		KSTAT_DATA_UINT64 },
899 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
900 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
901 	{ "full_magazines",	KSTAT_DATA_UINT64 },
902 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
903 	{ "magazine_size",	KSTAT_DATA_UINT64 },
904 	{ "reap",		KSTAT_DATA_UINT64 },
905 	{ "defrag",		KSTAT_DATA_UINT64 },
906 	{ "scan",		KSTAT_DATA_UINT64 },
907 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
908 	{ "move_yes",		KSTAT_DATA_UINT64 },
909 	{ "move_no",		KSTAT_DATA_UINT64 },
910 	{ "move_later",		KSTAT_DATA_UINT64 },
911 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
912 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
913 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
914 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
915 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
916 };
917 
918 static kmutex_t kmem_cache_kstat_lock;
919 
920 /*
921  * The default set of caches to back kmem_alloc().
922  * These sizes should be reevaluated periodically.
923  *
924  * We want allocations that are multiples of the coherency granularity
925  * (64 bytes) to be satisfied from a cache which is a multiple of 64
926  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
927  * the next kmem_cache_size greater than or equal to it must be a
928  * multiple of 64.
929  *
930  * We split the table into two sections:  size <= 4k and size > 4k.  This
931  * saves a lot of space and cache footprint in our cache tables.
932  */
933 static const int kmem_alloc_sizes[] = {
934 	1 * 8,
935 	2 * 8,
936 	3 * 8,
937 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
938 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
939 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
940 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
941 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
942 	P2ALIGN(8192 / 7, 64),
943 	P2ALIGN(8192 / 6, 64),
944 	P2ALIGN(8192 / 5, 64),
945 	P2ALIGN(8192 / 4, 64),
946 	P2ALIGN(8192 / 3, 64),
947 	P2ALIGN(8192 / 2, 64),
948 };
949 
950 static const int kmem_big_alloc_sizes[] = {
951 	2 * 4096,	3 * 4096,
952 	2 * 8192,	3 * 8192,
953 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
954 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
955 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
956 	16 * 8192
957 };
958 
959 #define	KMEM_MAXBUF		4096
960 #define	KMEM_BIG_MAXBUF_32BIT	32768
961 #define	KMEM_BIG_MAXBUF		131072
962 
963 #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
964 #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
965 
966 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
967 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
968 
969 #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
970 static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
971 
972 static kmem_magtype_t kmem_magtype[] = {
973 	{ 1,	8,	3200,	65536	},
974 	{ 3,	16,	256,	32768	},
975 	{ 7,	32,	64,	16384	},
976 	{ 15,	64,	0,	8192	},
977 	{ 31,	64,	0,	4096	},
978 	{ 47,	64,	0,	2048	},
979 	{ 63,	64,	0,	1024	},
980 	{ 95,	64,	0,	512	},
981 	{ 143,	64,	0,	0	},
982 };
983 
984 static uint32_t kmem_reaping;
985 static uint32_t kmem_reaping_idspace;
986 
987 /*
988  * kmem tunables
989  */
990 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
991 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
992 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
993 int kmem_panic = 1;		/* whether to panic on error */
994 int kmem_logging = 1;		/* kmem_log_enter() override */
995 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
996 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
997 size_t kmem_content_log_size;	/* content log size [2% of memory] */
998 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
999 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
1000 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1001 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
1002 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1003 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
1004 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
1005 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
1006 
1007 #ifdef _LP64
1008 size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1009 #else
1010 size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1011 #endif
1012 
1013 #ifdef DEBUG
1014 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1015 #else
1016 int kmem_flags = 0;
1017 #endif
1018 int kmem_ready;
1019 
1020 static kmem_cache_t	*kmem_slab_cache;
1021 static kmem_cache_t	*kmem_bufctl_cache;
1022 static kmem_cache_t	*kmem_bufctl_audit_cache;
1023 
1024 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1025 static list_t		kmem_caches;
1026 
1027 static taskq_t		*kmem_taskq;
1028 static kmutex_t		kmem_flags_lock;
1029 static vmem_t		*kmem_metadata_arena;
1030 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
1031 static vmem_t		*kmem_cache_arena;
1032 static vmem_t		*kmem_hash_arena;
1033 static vmem_t		*kmem_log_arena;
1034 static vmem_t		*kmem_oversize_arena;
1035 static vmem_t		*kmem_va_arena;
1036 static vmem_t		*kmem_default_arena;
1037 static vmem_t		*kmem_firewall_va_arena;
1038 static vmem_t		*kmem_firewall_arena;
1039 
1040 /*
1041  * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1042  * turned on when DEBUG is also defined.
1043  */
1044 #ifdef	DEBUG
1045 #define	KMEM_STATS
1046 #endif	/* DEBUG */
1047 
1048 #ifdef	KMEM_STATS
1049 #define	KMEM_STAT_ADD(stat)			((stat)++)
1050 #define	KMEM_STAT_COND_ADD(cond, stat)		((void) (!(cond) || (stat)++))
1051 #else
1052 #define	KMEM_STAT_ADD(stat)			/* nothing */
1053 #define	KMEM_STAT_COND_ADD(cond, stat)		/* nothing */
1054 #endif	/* KMEM_STATS */
1055 
1056 /*
1057  * kmem slab consolidator thresholds (tunables)
1058  */
1059 size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1060 size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1061 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1062 /*
1063  * Maximum number of slabs from which to move buffers during a single
1064  * maintenance interval while the system is not low on memory.
1065  */
1066 size_t kmem_reclaim_max_slabs = 1;
1067 /*
1068  * Number of slabs to scan backwards from the end of the partial slab list
1069  * when searching for buffers to relocate.
1070  */
1071 size_t kmem_reclaim_scan_range = 12;
1072 
1073 #ifdef	KMEM_STATS
1074 static struct {
1075 	uint64_t kms_callbacks;
1076 	uint64_t kms_yes;
1077 	uint64_t kms_no;
1078 	uint64_t kms_later;
1079 	uint64_t kms_dont_need;
1080 	uint64_t kms_dont_know;
1081 	uint64_t kms_hunt_found_mag;
1082 	uint64_t kms_hunt_found_slab;
1083 	uint64_t kms_hunt_alloc_fail;
1084 	uint64_t kms_hunt_lucky;
1085 	uint64_t kms_notify;
1086 	uint64_t kms_notify_callbacks;
1087 	uint64_t kms_disbelief;
1088 	uint64_t kms_already_pending;
1089 	uint64_t kms_callback_alloc_fail;
1090 	uint64_t kms_callback_taskq_fail;
1091 	uint64_t kms_endscan_slab_dead;
1092 	uint64_t kms_endscan_slab_destroyed;
1093 	uint64_t kms_endscan_nomem;
1094 	uint64_t kms_endscan_refcnt_changed;
1095 	uint64_t kms_endscan_nomove_changed;
1096 	uint64_t kms_endscan_freelist;
1097 	uint64_t kms_avl_update;
1098 	uint64_t kms_avl_noupdate;
1099 	uint64_t kms_no_longer_reclaimable;
1100 	uint64_t kms_notify_no_longer_reclaimable;
1101 	uint64_t kms_notify_slab_dead;
1102 	uint64_t kms_notify_slab_destroyed;
1103 	uint64_t kms_alloc_fail;
1104 	uint64_t kms_constructor_fail;
1105 	uint64_t kms_dead_slabs_freed;
1106 	uint64_t kms_defrags;
1107 	uint64_t kms_scans;
1108 	uint64_t kms_scan_depot_ws_reaps;
1109 	uint64_t kms_debug_reaps;
1110 	uint64_t kms_debug_scans;
1111 } kmem_move_stats;
1112 #endif	/* KMEM_STATS */
1113 
1114 /* consolidator knobs */
1115 static boolean_t kmem_move_noreap;
1116 static boolean_t kmem_move_blocked;
1117 static boolean_t kmem_move_fulltilt;
1118 static boolean_t kmem_move_any_partial;
1119 
1120 #ifdef	DEBUG
1121 /*
1122  * kmem consolidator debug tunables:
1123  * Ensure code coverage by occasionally running the consolidator even when the
1124  * caches are not fragmented (they may never be). These intervals are mean time
1125  * in cache maintenance intervals (kmem_cache_update).
1126  */
1127 uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1128 uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1129 #endif	/* DEBUG */
1130 
1131 static kmem_cache_t	*kmem_defrag_cache;
1132 static kmem_cache_t	*kmem_move_cache;
1133 static taskq_t		*kmem_move_taskq;
1134 
1135 static void kmem_cache_scan(kmem_cache_t *);
1136 static void kmem_cache_defrag(kmem_cache_t *);
1137 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1138 
1139 
1140 kmem_log_header_t	*kmem_transaction_log;
1141 kmem_log_header_t	*kmem_content_log;
1142 kmem_log_header_t	*kmem_failure_log;
1143 kmem_log_header_t	*kmem_slab_log;
1144 
1145 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1146 
1147 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
1148 	if ((count) > 0) {						\
1149 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
1150 		pc_t *_e;						\
1151 		/* memmove() the old entries down one notch */		\
1152 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
1153 			*_e = *(_e - 1);				\
1154 		*_s = (uintptr_t)(caller);				\
1155 	}
1156 
1157 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
1158 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
1159 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
1160 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
1161 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
1162 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
1163 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
1164 #define	KMERR_BADSIZE	7	/* alloc size != free size */
1165 #define	KMERR_BADBASE	8	/* buffer base address wrong */
1166 
1167 struct {
1168 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
1169 	int		kmp_error;	/* type of kmem error */
1170 	void		*kmp_buffer;	/* buffer that induced panic */
1171 	void		*kmp_realbuf;	/* real start address for buffer */
1172 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
1173 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
1174 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
1175 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
1176 } kmem_panic_info;
1177 
1178 
1179 static void
copy_pattern(uint64_t pattern,void * buf_arg,size_t size)1180 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1181 {
1182 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1183 	uint64_t *buf = buf_arg;
1184 
1185 	while (buf < bufend)
1186 		*buf++ = pattern;
1187 }
1188 
1189 static void *
verify_pattern(uint64_t pattern,void * buf_arg,size_t size)1190 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1191 {
1192 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1193 	uint64_t *buf;
1194 
1195 	for (buf = buf_arg; buf < bufend; buf++)
1196 		if (*buf != pattern)
1197 			return (buf);
1198 	return (NULL);
1199 }
1200 
1201 static void *
verify_and_copy_pattern(uint64_t old,uint64_t new,void * buf_arg,size_t size)1202 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1203 {
1204 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1205 	uint64_t *buf;
1206 
1207 	for (buf = buf_arg; buf < bufend; buf++) {
1208 		if (*buf != old) {
1209 			copy_pattern(old, buf_arg,
1210 			    (char *)buf - (char *)buf_arg);
1211 			return (buf);
1212 		}
1213 		*buf = new;
1214 	}
1215 
1216 	return (NULL);
1217 }
1218 
1219 static void
kmem_cache_applyall(void (* func)(kmem_cache_t *),taskq_t * tq,int tqflag)1220 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1221 {
1222 	kmem_cache_t *cp;
1223 
1224 	mutex_enter(&kmem_cache_lock);
1225 	for (cp = list_head(&kmem_caches); cp != NULL;
1226 	    cp = list_next(&kmem_caches, cp))
1227 		if (tq != NULL)
1228 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1229 			    tqflag);
1230 		else
1231 			func(cp);
1232 	mutex_exit(&kmem_cache_lock);
1233 }
1234 
1235 static void
kmem_cache_applyall_id(void (* func)(kmem_cache_t *),taskq_t * tq,int tqflag)1236 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1237 {
1238 	kmem_cache_t *cp;
1239 
1240 	mutex_enter(&kmem_cache_lock);
1241 	for (cp = list_head(&kmem_caches); cp != NULL;
1242 	    cp = list_next(&kmem_caches, cp)) {
1243 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
1244 			continue;
1245 		if (tq != NULL)
1246 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
1247 			    tqflag);
1248 		else
1249 			func(cp);
1250 	}
1251 	mutex_exit(&kmem_cache_lock);
1252 }
1253 
1254 /*
1255  * Debugging support.  Given a buffer address, find its slab.
1256  */
1257 static kmem_slab_t *
kmem_findslab(kmem_cache_t * cp,void * buf)1258 kmem_findslab(kmem_cache_t *cp, void *buf)
1259 {
1260 	kmem_slab_t *sp;
1261 
1262 	mutex_enter(&cp->cache_lock);
1263 	for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1264 	    sp = list_next(&cp->cache_complete_slabs, sp)) {
1265 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1266 			mutex_exit(&cp->cache_lock);
1267 			return (sp);
1268 		}
1269 	}
1270 	for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1271 	    sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1272 		if (KMEM_SLAB_MEMBER(sp, buf)) {
1273 			mutex_exit(&cp->cache_lock);
1274 			return (sp);
1275 		}
1276 	}
1277 	mutex_exit(&cp->cache_lock);
1278 
1279 	return (NULL);
1280 }
1281 
1282 static void
kmem_error(int error,kmem_cache_t * cparg,void * bufarg)1283 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1284 {
1285 	kmem_buftag_t *btp = NULL;
1286 	kmem_bufctl_t *bcp = NULL;
1287 	kmem_cache_t *cp = cparg;
1288 	kmem_slab_t *sp;
1289 	uint64_t *off;
1290 	void *buf = bufarg;
1291 
1292 	kmem_logging = 0;	/* stop logging when a bad thing happens */
1293 
1294 	kmem_panic_info.kmp_timestamp = gethrtime();
1295 
1296 	sp = kmem_findslab(cp, buf);
1297 	if (sp == NULL) {
1298 		for (cp = list_tail(&kmem_caches); cp != NULL;
1299 		    cp = list_prev(&kmem_caches, cp)) {
1300 			if ((sp = kmem_findslab(cp, buf)) != NULL)
1301 				break;
1302 		}
1303 	}
1304 
1305 	if (sp == NULL) {
1306 		cp = NULL;
1307 		error = KMERR_BADADDR;
1308 	} else {
1309 		if (cp != cparg)
1310 			error = KMERR_BADCACHE;
1311 		else
1312 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1313 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1314 		if (buf != bufarg)
1315 			error = KMERR_BADBASE;
1316 		if (cp->cache_flags & KMF_BUFTAG)
1317 			btp = KMEM_BUFTAG(cp, buf);
1318 		if (cp->cache_flags & KMF_HASH) {
1319 			mutex_enter(&cp->cache_lock);
1320 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1321 				if (bcp->bc_addr == buf)
1322 					break;
1323 			mutex_exit(&cp->cache_lock);
1324 			if (bcp == NULL && btp != NULL)
1325 				bcp = btp->bt_bufctl;
1326 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1327 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1328 			    bcp->bc_addr != buf) {
1329 				error = KMERR_BADBUFCTL;
1330 				bcp = NULL;
1331 			}
1332 		}
1333 	}
1334 
1335 	kmem_panic_info.kmp_error = error;
1336 	kmem_panic_info.kmp_buffer = bufarg;
1337 	kmem_panic_info.kmp_realbuf = buf;
1338 	kmem_panic_info.kmp_cache = cparg;
1339 	kmem_panic_info.kmp_realcache = cp;
1340 	kmem_panic_info.kmp_slab = sp;
1341 	kmem_panic_info.kmp_bufctl = bcp;
1342 
1343 	printf("kernel memory allocator: ");
1344 
1345 	switch (error) {
1346 
1347 	case KMERR_MODIFIED:
1348 		printf("buffer modified after being freed\n");
1349 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1350 		if (off == NULL)	/* shouldn't happen */
1351 			off = buf;
1352 		printf("modification occurred at offset 0x%lx "
1353 		    "(0x%llx replaced by 0x%llx)\n",
1354 		    (uintptr_t)off - (uintptr_t)buf,
1355 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1356 		break;
1357 
1358 	case KMERR_REDZONE:
1359 		printf("redzone violation: write past end of buffer\n");
1360 		break;
1361 
1362 	case KMERR_BADADDR:
1363 		printf("invalid free: buffer not in cache\n");
1364 		break;
1365 
1366 	case KMERR_DUPFREE:
1367 		printf("duplicate free: buffer freed twice\n");
1368 		break;
1369 
1370 	case KMERR_BADBUFTAG:
1371 		printf("boundary tag corrupted\n");
1372 		printf("bcp ^ bxstat = %lx, should be %lx\n",
1373 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1374 		    KMEM_BUFTAG_FREE);
1375 		break;
1376 
1377 	case KMERR_BADBUFCTL:
1378 		printf("bufctl corrupted\n");
1379 		break;
1380 
1381 	case KMERR_BADCACHE:
1382 		printf("buffer freed to wrong cache\n");
1383 		printf("buffer was allocated from %s,\n", cp->cache_name);
1384 		printf("caller attempting free to %s.\n", cparg->cache_name);
1385 		break;
1386 
1387 	case KMERR_BADSIZE:
1388 		printf("bad free: free size (%u) != alloc size (%u)\n",
1389 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1390 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1391 		break;
1392 
1393 	case KMERR_BADBASE:
1394 		printf("bad free: free address (%p) != alloc address (%p)\n",
1395 		    bufarg, buf);
1396 		break;
1397 	}
1398 
1399 	printf("buffer=%p  bufctl=%p  cache: %s\n",
1400 	    bufarg, (void *)bcp, cparg->cache_name);
1401 
1402 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1403 	    error != KMERR_BADBUFCTL) {
1404 		int d;
1405 		timestruc_t ts;
1406 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1407 
1408 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1409 		printf("previous transaction on buffer %p:\n", buf);
1410 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1411 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1412 		    (void *)sp, cp->cache_name);
1413 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1414 			ulong_t off;
1415 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1416 			printf("%s+%lx\n", sym ? sym : "?", off);
1417 		}
1418 	}
1419 	if (kmem_panic > 0)
1420 		panic("kernel heap corruption detected");
1421 	if (kmem_panic == 0)
1422 		debug_enter(NULL);
1423 	kmem_logging = 1;	/* resume logging */
1424 }
1425 
1426 static kmem_log_header_t *
kmem_log_init(size_t logsize)1427 kmem_log_init(size_t logsize)
1428 {
1429 	kmem_log_header_t *lhp;
1430 	int nchunks = 4 * max_ncpus;
1431 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1432 	int i;
1433 
1434 	/*
1435 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1436 	 * to prevent false sharing of cache lines.
1437 	 */
1438 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1439 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1440 	    NULL, NULL, VM_SLEEP);
1441 	bzero(lhp, lhsize);
1442 
1443 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1444 	lhp->lh_nchunks = nchunks;
1445 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1446 	lhp->lh_base = vmem_alloc(kmem_log_arena,
1447 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
1448 	lhp->lh_free = vmem_alloc(kmem_log_arena,
1449 	    nchunks * sizeof (int), VM_SLEEP);
1450 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1451 
1452 	for (i = 0; i < max_ncpus; i++) {
1453 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1454 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1455 		clhp->clh_chunk = i;
1456 	}
1457 
1458 	for (i = max_ncpus; i < nchunks; i++)
1459 		lhp->lh_free[i] = i;
1460 
1461 	lhp->lh_head = max_ncpus;
1462 	lhp->lh_tail = 0;
1463 
1464 	return (lhp);
1465 }
1466 
1467 static void *
kmem_log_enter(kmem_log_header_t * lhp,void * data,size_t size)1468 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1469 {
1470 	void *logspace;
1471 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1472 
1473 	if (lhp == NULL || kmem_logging == 0 || panicstr)
1474 		return (NULL);
1475 
1476 	mutex_enter(&clhp->clh_lock);
1477 	clhp->clh_hits++;
1478 	if (size > clhp->clh_avail) {
1479 		mutex_enter(&lhp->lh_lock);
1480 		lhp->lh_hits++;
1481 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1482 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1483 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1484 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1485 		clhp->clh_current = lhp->lh_base +
1486 		    clhp->clh_chunk * lhp->lh_chunksize;
1487 		clhp->clh_avail = lhp->lh_chunksize;
1488 		if (size > lhp->lh_chunksize)
1489 			size = lhp->lh_chunksize;
1490 		mutex_exit(&lhp->lh_lock);
1491 	}
1492 	logspace = clhp->clh_current;
1493 	clhp->clh_current += size;
1494 	clhp->clh_avail -= size;
1495 	bcopy(data, logspace, size);
1496 	mutex_exit(&clhp->clh_lock);
1497 	return (logspace);
1498 }
1499 
1500 #define	KMEM_AUDIT(lp, cp, bcp)						\
1501 {									\
1502 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
1503 	_bcp->bc_timestamp = gethrtime();				\
1504 	_bcp->bc_thread = curthread;					\
1505 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
1506 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
1507 }
1508 
1509 static void
kmem_log_event(kmem_log_header_t * lp,kmem_cache_t * cp,kmem_slab_t * sp,void * addr)1510 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1511 	kmem_slab_t *sp, void *addr)
1512 {
1513 	kmem_bufctl_audit_t bca;
1514 
1515 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
1516 	bca.bc_addr = addr;
1517 	bca.bc_slab = sp;
1518 	bca.bc_cache = cp;
1519 	KMEM_AUDIT(lp, cp, &bca);
1520 }
1521 
1522 /*
1523  * Create a new slab for cache cp.
1524  */
1525 static kmem_slab_t *
kmem_slab_create(kmem_cache_t * cp,int kmflag)1526 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1527 {
1528 	size_t slabsize = cp->cache_slabsize;
1529 	size_t chunksize = cp->cache_chunksize;
1530 	int cache_flags = cp->cache_flags;
1531 	size_t color, chunks;
1532 	char *buf, *slab;
1533 	kmem_slab_t *sp;
1534 	kmem_bufctl_t *bcp;
1535 	vmem_t *vmp = cp->cache_arena;
1536 
1537 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1538 
1539 	color = cp->cache_color + cp->cache_align;
1540 	if (color > cp->cache_maxcolor)
1541 		color = cp->cache_mincolor;
1542 	cp->cache_color = color;
1543 
1544 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1545 
1546 	if (slab == NULL)
1547 		goto vmem_alloc_failure;
1548 
1549 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1550 
1551 	/*
1552 	 * Reverify what was already checked in kmem_cache_set_move(), since the
1553 	 * consolidator depends (for correctness) on slabs being initialized
1554 	 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1555 	 * clients to distinguish uninitialized memory from known objects).
1556 	 */
1557 	ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1558 	if (!(cp->cache_cflags & KMC_NOTOUCH))
1559 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1560 
1561 	if (cache_flags & KMF_HASH) {
1562 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1563 			goto slab_alloc_failure;
1564 		chunks = (slabsize - color) / chunksize;
1565 	} else {
1566 		sp = KMEM_SLAB(cp, slab);
1567 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1568 	}
1569 
1570 	sp->slab_cache	= cp;
1571 	sp->slab_head	= NULL;
1572 	sp->slab_refcnt	= 0;
1573 	sp->slab_base	= buf = slab + color;
1574 	sp->slab_chunks	= chunks;
1575 	sp->slab_stuck_offset = (uint32_t)-1;
1576 	sp->slab_later_count = 0;
1577 	sp->slab_flags = 0;
1578 
1579 	ASSERT(chunks > 0);
1580 	while (chunks-- != 0) {
1581 		if (cache_flags & KMF_HASH) {
1582 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1583 			if (bcp == NULL)
1584 				goto bufctl_alloc_failure;
1585 			if (cache_flags & KMF_AUDIT) {
1586 				kmem_bufctl_audit_t *bcap =
1587 				    (kmem_bufctl_audit_t *)bcp;
1588 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
1589 				bcap->bc_cache = cp;
1590 			}
1591 			bcp->bc_addr = buf;
1592 			bcp->bc_slab = sp;
1593 		} else {
1594 			bcp = KMEM_BUFCTL(cp, buf);
1595 		}
1596 		if (cache_flags & KMF_BUFTAG) {
1597 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1598 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1599 			btp->bt_bufctl = bcp;
1600 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1601 			if (cache_flags & KMF_DEADBEEF) {
1602 				copy_pattern(KMEM_FREE_PATTERN, buf,
1603 				    cp->cache_verify);
1604 			}
1605 		}
1606 		bcp->bc_next = sp->slab_head;
1607 		sp->slab_head = bcp;
1608 		buf += chunksize;
1609 	}
1610 
1611 	kmem_log_event(kmem_slab_log, cp, sp, slab);
1612 
1613 	return (sp);
1614 
1615 bufctl_alloc_failure:
1616 
1617 	while ((bcp = sp->slab_head) != NULL) {
1618 		sp->slab_head = bcp->bc_next;
1619 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
1620 	}
1621 	kmem_cache_free(kmem_slab_cache, sp);
1622 
1623 slab_alloc_failure:
1624 
1625 	vmem_free(vmp, slab, slabsize);
1626 
1627 vmem_alloc_failure:
1628 
1629 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1630 	atomic_inc_64(&cp->cache_alloc_fail);
1631 
1632 	return (NULL);
1633 }
1634 
1635 /*
1636  * Destroy a slab.
1637  */
1638 static void
kmem_slab_destroy(kmem_cache_t * cp,kmem_slab_t * sp)1639 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1640 {
1641 	vmem_t *vmp = cp->cache_arena;
1642 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1643 
1644 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1645 	ASSERT(sp->slab_refcnt == 0);
1646 
1647 	if (cp->cache_flags & KMF_HASH) {
1648 		kmem_bufctl_t *bcp;
1649 		while ((bcp = sp->slab_head) != NULL) {
1650 			sp->slab_head = bcp->bc_next;
1651 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
1652 		}
1653 		kmem_cache_free(kmem_slab_cache, sp);
1654 	}
1655 	vmem_free(vmp, slab, cp->cache_slabsize);
1656 }
1657 
1658 static void *
kmem_slab_alloc_impl(kmem_cache_t * cp,kmem_slab_t * sp,boolean_t prefill)1659 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1660 {
1661 	kmem_bufctl_t *bcp, **hash_bucket;
1662 	void *buf;
1663 	boolean_t new_slab = (sp->slab_refcnt == 0);
1664 
1665 	ASSERT(MUTEX_HELD(&cp->cache_lock));
1666 	/*
1667 	 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1668 	 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1669 	 * slab is newly created.
1670 	 */
1671 	ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1672 	    (sp == avl_first(&cp->cache_partial_slabs))));
1673 	ASSERT(sp->slab_cache == cp);
1674 
1675 	cp->cache_slab_alloc++;
1676 	cp->cache_bufslab--;
1677 	sp->slab_refcnt++;
1678 
1679 	bcp = sp->slab_head;
1680 	sp->slab_head = bcp->bc_next;
1681 
1682 	if (cp->cache_flags & KMF_HASH) {
1683 		/*
1684 		 * Add buffer to allocated-address hash table.
1685 		 */
1686 		buf = bcp->bc_addr;
1687 		hash_bucket = KMEM_HASH(cp, buf);
1688 		bcp->bc_next = *hash_bucket;
1689 		*hash_bucket = bcp;
1690 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1691 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1692 		}
1693 	} else {
1694 		buf = KMEM_BUF(cp, bcp);
1695 	}
1696 
1697 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1698 
1699 	if (sp->slab_head == NULL) {
1700 		ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1701 		if (new_slab) {
1702 			ASSERT(sp->slab_chunks == 1);
1703 		} else {
1704 			ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1705 			avl_remove(&cp->cache_partial_slabs, sp);
1706 			sp->slab_later_count = 0; /* clear history */
1707 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1708 			sp->slab_stuck_offset = (uint32_t)-1;
1709 		}
1710 		list_insert_head(&cp->cache_complete_slabs, sp);
1711 		cp->cache_complete_slab_count++;
1712 		return (buf);
1713 	}
1714 
1715 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1716 	/*
1717 	 * Peek to see if the magazine layer is enabled before
1718 	 * we prefill.  We're not holding the cpu cache lock,
1719 	 * so the peek could be wrong, but there's no harm in it.
1720 	 */
1721 	if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1722 	    (KMEM_CPU_CACHE(cp)->cc_magsize != 0))  {
1723 		kmem_slab_prefill(cp, sp);
1724 		return (buf);
1725 	}
1726 
1727 	if (new_slab) {
1728 		avl_add(&cp->cache_partial_slabs, sp);
1729 		return (buf);
1730 	}
1731 
1732 	/*
1733 	 * The slab is now more allocated than it was, so the
1734 	 * order remains unchanged.
1735 	 */
1736 	ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1737 	return (buf);
1738 }
1739 
1740 /*
1741  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1742  */
1743 static void *
kmem_slab_alloc(kmem_cache_t * cp,int kmflag)1744 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1745 {
1746 	kmem_slab_t *sp;
1747 	void *buf;
1748 	boolean_t test_destructor;
1749 
1750 	mutex_enter(&cp->cache_lock);
1751 	test_destructor = (cp->cache_slab_alloc == 0);
1752 	sp = avl_first(&cp->cache_partial_slabs);
1753 	if (sp == NULL) {
1754 		ASSERT(cp->cache_bufslab == 0);
1755 
1756 		/*
1757 		 * The freelist is empty.  Create a new slab.
1758 		 */
1759 		mutex_exit(&cp->cache_lock);
1760 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1761 			return (NULL);
1762 		}
1763 		mutex_enter(&cp->cache_lock);
1764 		cp->cache_slab_create++;
1765 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1766 			cp->cache_bufmax = cp->cache_buftotal;
1767 		cp->cache_bufslab += sp->slab_chunks;
1768 	}
1769 
1770 	buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1771 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1772 	    (cp->cache_complete_slab_count +
1773 	    avl_numnodes(&cp->cache_partial_slabs) +
1774 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1775 	mutex_exit(&cp->cache_lock);
1776 
1777 	if (test_destructor && cp->cache_destructor != NULL) {
1778 		/*
1779 		 * On the first kmem_slab_alloc(), assert that it is valid to
1780 		 * call the destructor on a newly constructed object without any
1781 		 * client involvement.
1782 		 */
1783 		if ((cp->cache_constructor == NULL) ||
1784 		    cp->cache_constructor(buf, cp->cache_private,
1785 		    kmflag) == 0) {
1786 			cp->cache_destructor(buf, cp->cache_private);
1787 		}
1788 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1789 		    cp->cache_bufsize);
1790 		if (cp->cache_flags & KMF_DEADBEEF) {
1791 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1792 		}
1793 	}
1794 
1795 	return (buf);
1796 }
1797 
1798 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1799 
1800 /*
1801  * Free a raw (unconstructed) buffer to cp's slab layer.
1802  */
1803 static void
kmem_slab_free(kmem_cache_t * cp,void * buf)1804 kmem_slab_free(kmem_cache_t *cp, void *buf)
1805 {
1806 	kmem_slab_t *sp;
1807 	kmem_bufctl_t *bcp, **prev_bcpp;
1808 
1809 	ASSERT(buf != NULL);
1810 
1811 	mutex_enter(&cp->cache_lock);
1812 	cp->cache_slab_free++;
1813 
1814 	if (cp->cache_flags & KMF_HASH) {
1815 		/*
1816 		 * Look up buffer in allocated-address hash table.
1817 		 */
1818 		prev_bcpp = KMEM_HASH(cp, buf);
1819 		while ((bcp = *prev_bcpp) != NULL) {
1820 			if (bcp->bc_addr == buf) {
1821 				*prev_bcpp = bcp->bc_next;
1822 				sp = bcp->bc_slab;
1823 				break;
1824 			}
1825 			cp->cache_lookup_depth++;
1826 			prev_bcpp = &bcp->bc_next;
1827 		}
1828 	} else {
1829 		bcp = KMEM_BUFCTL(cp, buf);
1830 		sp = KMEM_SLAB(cp, buf);
1831 	}
1832 
1833 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1834 		mutex_exit(&cp->cache_lock);
1835 		kmem_error(KMERR_BADADDR, cp, buf);
1836 		return;
1837 	}
1838 
1839 	if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1840 		/*
1841 		 * If this is the buffer that prevented the consolidator from
1842 		 * clearing the slab, we can reset the slab flags now that the
1843 		 * buffer is freed. (It makes sense to do this in
1844 		 * kmem_cache_free(), where the client gives up ownership of the
1845 		 * buffer, but on the hot path the test is too expensive.)
1846 		 */
1847 		kmem_slab_move_yes(cp, sp, buf);
1848 	}
1849 
1850 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1851 		if (cp->cache_flags & KMF_CONTENTS)
1852 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
1853 			    kmem_log_enter(kmem_content_log, buf,
1854 			    cp->cache_contents);
1855 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1856 	}
1857 
1858 	bcp->bc_next = sp->slab_head;
1859 	sp->slab_head = bcp;
1860 
1861 	cp->cache_bufslab++;
1862 	ASSERT(sp->slab_refcnt >= 1);
1863 
1864 	if (--sp->slab_refcnt == 0) {
1865 		/*
1866 		 * There are no outstanding allocations from this slab,
1867 		 * so we can reclaim the memory.
1868 		 */
1869 		if (sp->slab_chunks == 1) {
1870 			list_remove(&cp->cache_complete_slabs, sp);
1871 			cp->cache_complete_slab_count--;
1872 		} else {
1873 			avl_remove(&cp->cache_partial_slabs, sp);
1874 		}
1875 
1876 		cp->cache_buftotal -= sp->slab_chunks;
1877 		cp->cache_bufslab -= sp->slab_chunks;
1878 		/*
1879 		 * Defer releasing the slab to the virtual memory subsystem
1880 		 * while there is a pending move callback, since we guarantee
1881 		 * that buffers passed to the move callback have only been
1882 		 * touched by kmem or by the client itself. Since the memory
1883 		 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1884 		 * set at least one of the two lowest order bits, the client can
1885 		 * test those bits in the move callback to determine whether or
1886 		 * not it knows about the buffer (assuming that the client also
1887 		 * sets one of those low order bits whenever it frees a buffer).
1888 		 */
1889 		if (cp->cache_defrag == NULL ||
1890 		    (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1891 		    !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1892 			cp->cache_slab_destroy++;
1893 			mutex_exit(&cp->cache_lock);
1894 			kmem_slab_destroy(cp, sp);
1895 		} else {
1896 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1897 			/*
1898 			 * Slabs are inserted at both ends of the deadlist to
1899 			 * distinguish between slabs freed while move callbacks
1900 			 * are pending (list head) and a slab freed while the
1901 			 * lock is dropped in kmem_move_buffers() (list tail) so
1902 			 * that in both cases slab_destroy() is called from the
1903 			 * right context.
1904 			 */
1905 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1906 				list_insert_tail(deadlist, sp);
1907 			} else {
1908 				list_insert_head(deadlist, sp);
1909 			}
1910 			cp->cache_defrag->kmd_deadcount++;
1911 			mutex_exit(&cp->cache_lock);
1912 		}
1913 		return;
1914 	}
1915 
1916 	if (bcp->bc_next == NULL) {
1917 		/* Transition the slab from completely allocated to partial. */
1918 		ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1919 		ASSERT(sp->slab_chunks > 1);
1920 		list_remove(&cp->cache_complete_slabs, sp);
1921 		cp->cache_complete_slab_count--;
1922 		avl_add(&cp->cache_partial_slabs, sp);
1923 	} else {
1924 #ifdef	DEBUG
1925 		if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1926 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1927 		} else {
1928 			KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1929 		}
1930 #else
1931 		(void) avl_update_gt(&cp->cache_partial_slabs, sp);
1932 #endif
1933 	}
1934 
1935 	ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1936 	    (cp->cache_complete_slab_count +
1937 	    avl_numnodes(&cp->cache_partial_slabs) +
1938 	    (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1939 	mutex_exit(&cp->cache_lock);
1940 }
1941 
1942 /*
1943  * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1944  */
1945 static int
kmem_cache_alloc_debug(kmem_cache_t * cp,void * buf,int kmflag,int construct,caddr_t caller)1946 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1947     caddr_t caller)
1948 {
1949 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1950 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1951 	uint32_t mtbf;
1952 
1953 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1954 		kmem_error(KMERR_BADBUFTAG, cp, buf);
1955 		return (-1);
1956 	}
1957 
1958 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1959 
1960 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1961 		kmem_error(KMERR_BADBUFCTL, cp, buf);
1962 		return (-1);
1963 	}
1964 
1965 	if (cp->cache_flags & KMF_DEADBEEF) {
1966 		if (!construct && (cp->cache_flags & KMF_LITE)) {
1967 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1968 				kmem_error(KMERR_MODIFIED, cp, buf);
1969 				return (-1);
1970 			}
1971 			if (cp->cache_constructor != NULL)
1972 				*(uint64_t *)buf = btp->bt_redzone;
1973 			else
1974 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1975 		} else {
1976 			construct = 1;
1977 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1978 			    KMEM_UNINITIALIZED_PATTERN, buf,
1979 			    cp->cache_verify)) {
1980 				kmem_error(KMERR_MODIFIED, cp, buf);
1981 				return (-1);
1982 			}
1983 		}
1984 	}
1985 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
1986 
1987 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1988 	    gethrtime() % mtbf == 0 &&
1989 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1990 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1991 		if (!construct && cp->cache_destructor != NULL)
1992 			cp->cache_destructor(buf, cp->cache_private);
1993 	} else {
1994 		mtbf = 0;
1995 	}
1996 
1997 	if (mtbf || (construct && cp->cache_constructor != NULL &&
1998 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1999 		atomic_inc_64(&cp->cache_alloc_fail);
2000 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2001 		if (cp->cache_flags & KMF_DEADBEEF)
2002 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2003 		kmem_slab_free(cp, buf);
2004 		return (1);
2005 	}
2006 
2007 	if (cp->cache_flags & KMF_AUDIT) {
2008 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2009 	}
2010 
2011 	if ((cp->cache_flags & KMF_LITE) &&
2012 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2013 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2014 	}
2015 
2016 	return (0);
2017 }
2018 
2019 static int
kmem_cache_free_debug(kmem_cache_t * cp,void * buf,caddr_t caller)2020 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2021 {
2022 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2023 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2024 	kmem_slab_t *sp;
2025 
2026 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2027 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2028 			kmem_error(KMERR_DUPFREE, cp, buf);
2029 			return (-1);
2030 		}
2031 		sp = kmem_findslab(cp, buf);
2032 		if (sp == NULL || sp->slab_cache != cp)
2033 			kmem_error(KMERR_BADADDR, cp, buf);
2034 		else
2035 			kmem_error(KMERR_REDZONE, cp, buf);
2036 		return (-1);
2037 	}
2038 
2039 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2040 
2041 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2042 		kmem_error(KMERR_BADBUFCTL, cp, buf);
2043 		return (-1);
2044 	}
2045 
2046 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2047 		kmem_error(KMERR_REDZONE, cp, buf);
2048 		return (-1);
2049 	}
2050 
2051 	if (cp->cache_flags & KMF_AUDIT) {
2052 		if (cp->cache_flags & KMF_CONTENTS)
2053 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
2054 			    buf, cp->cache_contents);
2055 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2056 	}
2057 
2058 	if ((cp->cache_flags & KMF_LITE) &&
2059 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2060 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2061 	}
2062 
2063 	if (cp->cache_flags & KMF_DEADBEEF) {
2064 		if (cp->cache_flags & KMF_LITE)
2065 			btp->bt_redzone = *(uint64_t *)buf;
2066 		else if (cp->cache_destructor != NULL)
2067 			cp->cache_destructor(buf, cp->cache_private);
2068 
2069 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2070 	}
2071 
2072 	return (0);
2073 }
2074 
2075 /*
2076  * Free each object in magazine mp to cp's slab layer, and free mp itself.
2077  */
2078 static void
kmem_magazine_destroy(kmem_cache_t * cp,kmem_magazine_t * mp,int nrounds)2079 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2080 {
2081 	int round;
2082 
2083 	ASSERT(!list_link_active(&cp->cache_link) ||
2084 	    taskq_member(kmem_taskq, curthread));
2085 
2086 	for (round = 0; round < nrounds; round++) {
2087 		void *buf = mp->mag_round[round];
2088 
2089 		if (cp->cache_flags & KMF_DEADBEEF) {
2090 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
2091 			    cp->cache_verify) != NULL) {
2092 				kmem_error(KMERR_MODIFIED, cp, buf);
2093 				continue;
2094 			}
2095 			if ((cp->cache_flags & KMF_LITE) &&
2096 			    cp->cache_destructor != NULL) {
2097 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2098 				*(uint64_t *)buf = btp->bt_redzone;
2099 				cp->cache_destructor(buf, cp->cache_private);
2100 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
2101 			}
2102 		} else if (cp->cache_destructor != NULL) {
2103 			cp->cache_destructor(buf, cp->cache_private);
2104 		}
2105 
2106 		kmem_slab_free(cp, buf);
2107 	}
2108 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2109 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2110 }
2111 
2112 /*
2113  * Allocate a magazine from the depot.
2114  */
2115 static kmem_magazine_t *
kmem_depot_alloc(kmem_cache_t * cp,kmem_maglist_t * mlp)2116 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2117 {
2118 	kmem_magazine_t *mp;
2119 
2120 	/*
2121 	 * If we can't get the depot lock without contention,
2122 	 * update our contention count.  We use the depot
2123 	 * contention rate to determine whether we need to
2124 	 * increase the magazine size for better scalability.
2125 	 */
2126 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
2127 		mutex_enter(&cp->cache_depot_lock);
2128 		cp->cache_depot_contention++;
2129 	}
2130 
2131 	if ((mp = mlp->ml_list) != NULL) {
2132 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2133 		mlp->ml_list = mp->mag_next;
2134 		if (--mlp->ml_total < mlp->ml_min)
2135 			mlp->ml_min = mlp->ml_total;
2136 		mlp->ml_alloc++;
2137 	}
2138 
2139 	mutex_exit(&cp->cache_depot_lock);
2140 
2141 	return (mp);
2142 }
2143 
2144 /*
2145  * Free a magazine to the depot.
2146  */
2147 static void
kmem_depot_free(kmem_cache_t * cp,kmem_maglist_t * mlp,kmem_magazine_t * mp)2148 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2149 {
2150 	mutex_enter(&cp->cache_depot_lock);
2151 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2152 	mp->mag_next = mlp->ml_list;
2153 	mlp->ml_list = mp;
2154 	mlp->ml_total++;
2155 	mutex_exit(&cp->cache_depot_lock);
2156 }
2157 
2158 /*
2159  * Update the working set statistics for cp's depot.
2160  */
2161 static void
kmem_depot_ws_update(kmem_cache_t * cp)2162 kmem_depot_ws_update(kmem_cache_t *cp)
2163 {
2164 	mutex_enter(&cp->cache_depot_lock);
2165 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2166 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2167 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2168 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2169 	mutex_exit(&cp->cache_depot_lock);
2170 }
2171 
2172 /*
2173  * Set the working set statistics for cp's depot to zero.  (Everything is
2174  * eligible for reaping.)
2175  */
2176 static void
kmem_depot_ws_zero(kmem_cache_t * cp)2177 kmem_depot_ws_zero(kmem_cache_t *cp)
2178 {
2179 	mutex_enter(&cp->cache_depot_lock);
2180 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2181 	cp->cache_full.ml_min = cp->cache_full.ml_total;
2182 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2183 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2184 	mutex_exit(&cp->cache_depot_lock);
2185 }
2186 
2187 /*
2188  * Reap all magazines that have fallen out of the depot's working set.
2189  */
2190 static void
kmem_depot_ws_reap(kmem_cache_t * cp)2191 kmem_depot_ws_reap(kmem_cache_t *cp)
2192 {
2193 	long reap;
2194 	kmem_magazine_t *mp;
2195 
2196 	ASSERT(!list_link_active(&cp->cache_link) ||
2197 	    taskq_member(kmem_taskq, curthread));
2198 
2199 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2200 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
2201 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2202 
2203 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2204 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
2205 		kmem_magazine_destroy(cp, mp, 0);
2206 }
2207 
2208 static void
kmem_cpu_reload(kmem_cpu_cache_t * ccp,kmem_magazine_t * mp,int rounds)2209 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2210 {
2211 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2212 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2213 	ASSERT(ccp->cc_magsize > 0);
2214 
2215 	ccp->cc_ploaded = ccp->cc_loaded;
2216 	ccp->cc_prounds = ccp->cc_rounds;
2217 	ccp->cc_loaded = mp;
2218 	ccp->cc_rounds = rounds;
2219 }
2220 
2221 /*
2222  * Intercept kmem alloc/free calls during crash dump in order to avoid
2223  * changing kmem state while memory is being saved to the dump device.
2224  * Otherwise, ::kmem_verify will report "corrupt buffers".  Note that
2225  * there are no locks because only one CPU calls kmem during a crash
2226  * dump. To enable this feature, first create the associated vmem
2227  * arena with VMC_DUMPSAFE.
2228  */
2229 static void *kmem_dump_start;	/* start of pre-reserved heap */
2230 static void *kmem_dump_end;	/* end of heap area */
2231 static void *kmem_dump_curr;	/* current free heap pointer */
2232 static size_t kmem_dump_size;	/* size of heap area */
2233 
2234 /* append to each buf created in the pre-reserved heap */
2235 typedef struct kmem_dumpctl {
2236 	void	*kdc_next;	/* cache dump free list linkage */
2237 } kmem_dumpctl_t;
2238 
2239 #define	KMEM_DUMPCTL(cp, buf)	\
2240 	((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2241 	    sizeof (void *)))
2242 
2243 /* set non zero for full report */
2244 uint_t kmem_dump_verbose = 0;
2245 
2246 /* stats for overize heap */
2247 uint_t kmem_dump_oversize_allocs = 0;
2248 uint_t kmem_dump_oversize_max = 0;
2249 
2250 static void
kmem_dumppr(char ** pp,char * e,const char * format,...)2251 kmem_dumppr(char **pp, char *e, const char *format, ...)
2252 {
2253 	char *p = *pp;
2254 
2255 	if (p < e) {
2256 		int n;
2257 		va_list ap;
2258 
2259 		va_start(ap, format);
2260 		n = vsnprintf(p, e - p, format, ap);
2261 		va_end(ap);
2262 		*pp = p + n;
2263 	}
2264 }
2265 
2266 /*
2267  * Called when dumpadm(1M) configures dump parameters.
2268  */
2269 void
kmem_dump_init(size_t size)2270 kmem_dump_init(size_t size)
2271 {
2272 	/* Our caller ensures size is always set. */
2273 	ASSERT3U(size, >, 0);
2274 
2275 	if (kmem_dump_start != NULL)
2276 		kmem_free(kmem_dump_start, kmem_dump_size);
2277 
2278 	kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2279 	kmem_dump_size = size;
2280 	kmem_dump_curr = kmem_dump_start;
2281 	kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2282 	copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2283 }
2284 
2285 /*
2286  * Set flag for each kmem_cache_t if is safe to use alternate dump
2287  * memory. Called just before panic crash dump starts. Set the flag
2288  * for the calling CPU.
2289  */
2290 void
kmem_dump_begin(void)2291 kmem_dump_begin(void)
2292 {
2293 	kmem_cache_t *cp;
2294 
2295 	ASSERT(panicstr != NULL);
2296 
2297 	for (cp = list_head(&kmem_caches); cp != NULL;
2298 	    cp = list_next(&kmem_caches, cp)) {
2299 		kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2300 
2301 		if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2302 			cp->cache_flags |= KMF_DUMPDIVERT;
2303 			ccp->cc_flags |= KMF_DUMPDIVERT;
2304 			ccp->cc_dump_rounds = ccp->cc_rounds;
2305 			ccp->cc_dump_prounds = ccp->cc_prounds;
2306 			ccp->cc_rounds = ccp->cc_prounds = -1;
2307 		} else {
2308 			cp->cache_flags |= KMF_DUMPUNSAFE;
2309 			ccp->cc_flags |= KMF_DUMPUNSAFE;
2310 		}
2311 	}
2312 }
2313 
2314 /*
2315  * finished dump intercept
2316  * print any warnings on the console
2317  * return verbose information to dumpsys() in the given buffer
2318  */
2319 size_t
kmem_dump_finish(char * buf,size_t size)2320 kmem_dump_finish(char *buf, size_t size)
2321 {
2322 	int percent = 0;
2323 	size_t used;
2324 	char *e = buf + size;
2325 	char *p = buf;
2326 
2327 	if (kmem_dump_curr == kmem_dump_end) {
2328 		cmn_err(CE_WARN, "exceeded kmem_dump space of %lu "
2329 		    "bytes: kmem state in dump may be inconsistent",
2330 		    kmem_dump_size);
2331 	}
2332 
2333 	if (kmem_dump_verbose == 0)
2334 		return (0);
2335 
2336 	used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2337 	percent = (used * 100) / kmem_dump_size;
2338 
2339 	kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2340 	kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2341 	kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2342 	kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2343 	    kmem_dump_oversize_allocs);
2344 	kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2345 	    kmem_dump_oversize_max);
2346 
2347 	/* return buffer size used */
2348 	if (p < e)
2349 		bzero(p, e - p);
2350 	return (p - buf);
2351 }
2352 
2353 /*
2354  * Allocate a constructed object from alternate dump memory.
2355  */
2356 void *
kmem_cache_alloc_dump(kmem_cache_t * cp,int kmflag)2357 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2358 {
2359 	void *buf;
2360 	void *curr;
2361 	char *bufend;
2362 
2363 	/* return a constructed object */
2364 	if ((buf = cp->cache_dump.kd_freelist) != NULL) {
2365 		cp->cache_dump.kd_freelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2366 		return (buf);
2367 	}
2368 
2369 	/* create a new constructed object */
2370 	curr = kmem_dump_curr;
2371 	buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2372 	bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2373 
2374 	/* hat layer objects cannot cross a page boundary */
2375 	if (cp->cache_align < PAGESIZE) {
2376 		char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2377 		if (bufend > page) {
2378 			bufend += page - (char *)buf;
2379 			buf = (void *)page;
2380 		}
2381 	}
2382 
2383 	/* fall back to normal alloc if reserved area is used up */
2384 	if (bufend > (char *)kmem_dump_end) {
2385 		kmem_dump_curr = kmem_dump_end;
2386 		cp->cache_dump.kd_alloc_fails++;
2387 		return (NULL);
2388 	}
2389 
2390 	/*
2391 	 * Must advance curr pointer before calling a constructor that
2392 	 * may also allocate memory.
2393 	 */
2394 	kmem_dump_curr = bufend;
2395 
2396 	/* run constructor */
2397 	if (cp->cache_constructor != NULL &&
2398 	    cp->cache_constructor(buf, cp->cache_private, kmflag)
2399 	    != 0) {
2400 #ifdef DEBUG
2401 		printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2402 		    cp->cache_name, (void *)cp);
2403 #endif
2404 		/* reset curr pointer iff no allocs were done */
2405 		if (kmem_dump_curr == bufend)
2406 			kmem_dump_curr = curr;
2407 
2408 		cp->cache_dump.kd_alloc_fails++;
2409 		/* fall back to normal alloc if the constructor fails */
2410 		return (NULL);
2411 	}
2412 
2413 	return (buf);
2414 }
2415 
2416 /*
2417  * Free a constructed object in alternate dump memory.
2418  */
2419 int
kmem_cache_free_dump(kmem_cache_t * cp,void * buf)2420 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2421 {
2422 	/* save constructed buffers for next time */
2423 	if ((char *)buf >= (char *)kmem_dump_start &&
2424 	    (char *)buf < (char *)kmem_dump_end) {
2425 		KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dump.kd_freelist;
2426 		cp->cache_dump.kd_freelist = buf;
2427 		return (0);
2428 	}
2429 
2430 	/* just drop buffers that were allocated before dump started */
2431 	if (kmem_dump_curr < kmem_dump_end)
2432 		return (0);
2433 
2434 	/* fall back to normal free if reserved area is used up */
2435 	return (1);
2436 }
2437 
2438 /*
2439  * Allocate a constructed object from cache cp.
2440  */
2441 void *
kmem_cache_alloc(kmem_cache_t * cp,int kmflag)2442 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2443 {
2444 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2445 	kmem_magazine_t *fmp;
2446 	void *buf;
2447 
2448 	mutex_enter(&ccp->cc_lock);
2449 	for (;;) {
2450 		/*
2451 		 * If there's an object available in the current CPU's
2452 		 * loaded magazine, just take it and return.
2453 		 */
2454 		if (ccp->cc_rounds > 0) {
2455 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2456 			ccp->cc_alloc++;
2457 			mutex_exit(&ccp->cc_lock);
2458 			if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2459 				if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2460 					ASSERT(!(ccp->cc_flags &
2461 					    KMF_DUMPDIVERT));
2462 					cp->cache_dump.kd_unsafe++;
2463 				}
2464 				if ((ccp->cc_flags & KMF_BUFTAG) &&
2465 				    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2466 				    caller()) != 0) {
2467 					if (kmflag & KM_NOSLEEP)
2468 						return (NULL);
2469 					mutex_enter(&ccp->cc_lock);
2470 					continue;
2471 				}
2472 			}
2473 			return (buf);
2474 		}
2475 
2476 		/*
2477 		 * The loaded magazine is empty.  If the previously loaded
2478 		 * magazine was full, exchange them and try again.
2479 		 */
2480 		if (ccp->cc_prounds > 0) {
2481 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2482 			continue;
2483 		}
2484 
2485 		/*
2486 		 * Return an alternate buffer at dump time to preserve
2487 		 * the heap.
2488 		 */
2489 		if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2490 			if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2491 				ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2492 				/* log it so that we can warn about it */
2493 				cp->cache_dump.kd_unsafe++;
2494 			} else {
2495 				if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2496 				    NULL) {
2497 					mutex_exit(&ccp->cc_lock);
2498 					return (buf);
2499 				}
2500 				break;		/* fall back to slab layer */
2501 			}
2502 		}
2503 
2504 		/*
2505 		 * If the magazine layer is disabled, break out now.
2506 		 */
2507 		if (ccp->cc_magsize == 0)
2508 			break;
2509 
2510 		/*
2511 		 * Try to get a full magazine from the depot.
2512 		 */
2513 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
2514 		if (fmp != NULL) {
2515 			if (ccp->cc_ploaded != NULL)
2516 				kmem_depot_free(cp, &cp->cache_empty,
2517 				    ccp->cc_ploaded);
2518 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2519 			continue;
2520 		}
2521 
2522 		/*
2523 		 * There are no full magazines in the depot,
2524 		 * so fall through to the slab layer.
2525 		 */
2526 		break;
2527 	}
2528 	mutex_exit(&ccp->cc_lock);
2529 
2530 	/*
2531 	 * We couldn't allocate a constructed object from the magazine layer,
2532 	 * so get a raw buffer from the slab layer and apply its constructor.
2533 	 */
2534 	buf = kmem_slab_alloc(cp, kmflag);
2535 
2536 	if (buf == NULL)
2537 		return (NULL);
2538 
2539 	if (cp->cache_flags & KMF_BUFTAG) {
2540 		/*
2541 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
2542 		 */
2543 		int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2544 		if (rc != 0) {
2545 			if (kmflag & KM_NOSLEEP)
2546 				return (NULL);
2547 			/*
2548 			 * kmem_cache_alloc_debug() detected corruption
2549 			 * but didn't panic (kmem_panic <= 0). We should not be
2550 			 * here because the constructor failed (indicated by a
2551 			 * return code of 1). Try again.
2552 			 */
2553 			ASSERT(rc == -1);
2554 			return (kmem_cache_alloc(cp, kmflag));
2555 		}
2556 		return (buf);
2557 	}
2558 
2559 	if (cp->cache_constructor != NULL &&
2560 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2561 		atomic_inc_64(&cp->cache_alloc_fail);
2562 		kmem_slab_free(cp, buf);
2563 		return (NULL);
2564 	}
2565 
2566 	return (buf);
2567 }
2568 
2569 /*
2570  * The freed argument tells whether or not kmem_cache_free_debug() has already
2571  * been called so that we can avoid the duplicate free error. For example, a
2572  * buffer on a magazine has already been freed by the client but is still
2573  * constructed.
2574  */
2575 static void
kmem_slab_free_constructed(kmem_cache_t * cp,void * buf,boolean_t freed)2576 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2577 {
2578 	if (!freed && (cp->cache_flags & KMF_BUFTAG))
2579 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2580 			return;
2581 
2582 	/*
2583 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2584 	 * kmem_cache_free_debug() will have already applied the destructor.
2585 	 */
2586 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2587 	    cp->cache_destructor != NULL) {
2588 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
2589 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2590 			*(uint64_t *)buf = btp->bt_redzone;
2591 			cp->cache_destructor(buf, cp->cache_private);
2592 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
2593 		} else {
2594 			cp->cache_destructor(buf, cp->cache_private);
2595 		}
2596 	}
2597 
2598 	kmem_slab_free(cp, buf);
2599 }
2600 
2601 /*
2602  * Used when there's no room to free a buffer to the per-CPU cache.
2603  * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2604  * caller should try freeing to the per-CPU cache again.
2605  * Note that we don't directly install the magazine in the cpu cache,
2606  * since its state may have changed wildly while the lock was dropped.
2607  */
2608 static int
kmem_cpucache_magazine_alloc(kmem_cpu_cache_t * ccp,kmem_cache_t * cp)2609 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2610 {
2611 	kmem_magazine_t *emp;
2612 	kmem_magtype_t *mtp;
2613 
2614 	ASSERT(MUTEX_HELD(&ccp->cc_lock));
2615 	ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2616 	    ((uint_t)ccp->cc_rounds == -1)) &&
2617 	    ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2618 	    ((uint_t)ccp->cc_prounds == -1)));
2619 
2620 	emp = kmem_depot_alloc(cp, &cp->cache_empty);
2621 	if (emp != NULL) {
2622 		if (ccp->cc_ploaded != NULL)
2623 			kmem_depot_free(cp, &cp->cache_full,
2624 			    ccp->cc_ploaded);
2625 		kmem_cpu_reload(ccp, emp, 0);
2626 		return (1);
2627 	}
2628 	/*
2629 	 * There are no empty magazines in the depot,
2630 	 * so try to allocate a new one.  We must drop all locks
2631 	 * across kmem_cache_alloc() because lower layers may
2632 	 * attempt to allocate from this cache.
2633 	 */
2634 	mtp = cp->cache_magtype;
2635 	mutex_exit(&ccp->cc_lock);
2636 	emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2637 	mutex_enter(&ccp->cc_lock);
2638 
2639 	if (emp != NULL) {
2640 		/*
2641 		 * We successfully allocated an empty magazine.
2642 		 * However, we had to drop ccp->cc_lock to do it,
2643 		 * so the cache's magazine size may have changed.
2644 		 * If so, free the magazine and try again.
2645 		 */
2646 		if (ccp->cc_magsize != mtp->mt_magsize) {
2647 			mutex_exit(&ccp->cc_lock);
2648 			kmem_cache_free(mtp->mt_cache, emp);
2649 			mutex_enter(&ccp->cc_lock);
2650 			return (1);
2651 		}
2652 
2653 		/*
2654 		 * We got a magazine of the right size.  Add it to
2655 		 * the depot and try the whole dance again.
2656 		 */
2657 		kmem_depot_free(cp, &cp->cache_empty, emp);
2658 		return (1);
2659 	}
2660 
2661 	/*
2662 	 * We couldn't allocate an empty magazine,
2663 	 * so fall through to the slab layer.
2664 	 */
2665 	return (0);
2666 }
2667 
2668 /*
2669  * Free a constructed object to cache cp.
2670  */
2671 void
kmem_cache_free(kmem_cache_t * cp,void * buf)2672 kmem_cache_free(kmem_cache_t *cp, void *buf)
2673 {
2674 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2675 
2676 	/*
2677 	 * The client must not free either of the buffers passed to the move
2678 	 * callback function.
2679 	 */
2680 	ASSERT(cp->cache_defrag == NULL ||
2681 	    cp->cache_defrag->kmd_thread != curthread ||
2682 	    (buf != cp->cache_defrag->kmd_from_buf &&
2683 	    buf != cp->cache_defrag->kmd_to_buf));
2684 
2685 	if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2686 		if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2687 			ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2688 			/* log it so that we can warn about it */
2689 			cp->cache_dump.kd_unsafe++;
2690 		} else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2691 			return;
2692 		}
2693 		if (ccp->cc_flags & KMF_BUFTAG) {
2694 			if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2695 				return;
2696 		}
2697 	}
2698 
2699 	mutex_enter(&ccp->cc_lock);
2700 	/*
2701 	 * Any changes to this logic should be reflected in kmem_slab_prefill()
2702 	 */
2703 	for (;;) {
2704 		/*
2705 		 * If there's a slot available in the current CPU's
2706 		 * loaded magazine, just put the object there and return.
2707 		 */
2708 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2709 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2710 			ccp->cc_free++;
2711 			mutex_exit(&ccp->cc_lock);
2712 			return;
2713 		}
2714 
2715 		/*
2716 		 * The loaded magazine is full.  If the previously loaded
2717 		 * magazine was empty, exchange them and try again.
2718 		 */
2719 		if (ccp->cc_prounds == 0) {
2720 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2721 			continue;
2722 		}
2723 
2724 		/*
2725 		 * If the magazine layer is disabled, break out now.
2726 		 */
2727 		if (ccp->cc_magsize == 0)
2728 			break;
2729 
2730 		if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2731 			/*
2732 			 * We couldn't free our constructed object to the
2733 			 * magazine layer, so apply its destructor and free it
2734 			 * to the slab layer.
2735 			 */
2736 			break;
2737 		}
2738 	}
2739 	mutex_exit(&ccp->cc_lock);
2740 	kmem_slab_free_constructed(cp, buf, B_TRUE);
2741 }
2742 
2743 static void
kmem_slab_prefill(kmem_cache_t * cp,kmem_slab_t * sp)2744 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2745 {
2746 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2747 	int cache_flags = cp->cache_flags;
2748 
2749 	kmem_bufctl_t *next, *head;
2750 	size_t nbufs;
2751 
2752 	/*
2753 	 * Completely allocate the newly created slab and put the pre-allocated
2754 	 * buffers in magazines. Any of the buffers that cannot be put in
2755 	 * magazines must be returned to the slab.
2756 	 */
2757 	ASSERT(MUTEX_HELD(&cp->cache_lock));
2758 	ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2759 	ASSERT(cp->cache_constructor == NULL);
2760 	ASSERT(sp->slab_cache == cp);
2761 	ASSERT(sp->slab_refcnt == 1);
2762 	ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2763 	ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2764 
2765 	head = sp->slab_head;
2766 	nbufs = (sp->slab_chunks - sp->slab_refcnt);
2767 	sp->slab_head = NULL;
2768 	sp->slab_refcnt += nbufs;
2769 	cp->cache_bufslab -= nbufs;
2770 	cp->cache_slab_alloc += nbufs;
2771 	list_insert_head(&cp->cache_complete_slabs, sp);
2772 	cp->cache_complete_slab_count++;
2773 	mutex_exit(&cp->cache_lock);
2774 	mutex_enter(&ccp->cc_lock);
2775 
2776 	while (head != NULL) {
2777 		void *buf = KMEM_BUF(cp, head);
2778 		/*
2779 		 * If there's a slot available in the current CPU's
2780 		 * loaded magazine, just put the object there and
2781 		 * continue.
2782 		 */
2783 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2784 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2785 			    buf;
2786 			ccp->cc_free++;
2787 			nbufs--;
2788 			head = head->bc_next;
2789 			continue;
2790 		}
2791 
2792 		/*
2793 		 * The loaded magazine is full.  If the previously
2794 		 * loaded magazine was empty, exchange them and try
2795 		 * again.
2796 		 */
2797 		if (ccp->cc_prounds == 0) {
2798 			kmem_cpu_reload(ccp, ccp->cc_ploaded,
2799 			    ccp->cc_prounds);
2800 			continue;
2801 		}
2802 
2803 		/*
2804 		 * If the magazine layer is disabled, break out now.
2805 		 */
2806 
2807 		if (ccp->cc_magsize == 0) {
2808 			break;
2809 		}
2810 
2811 		if (!kmem_cpucache_magazine_alloc(ccp, cp))
2812 			break;
2813 	}
2814 	mutex_exit(&ccp->cc_lock);
2815 	if (nbufs != 0) {
2816 		ASSERT(head != NULL);
2817 
2818 		/*
2819 		 * If there was a failure, return remaining objects to
2820 		 * the slab
2821 		 */
2822 		while (head != NULL) {
2823 			ASSERT(nbufs != 0);
2824 			next = head->bc_next;
2825 			head->bc_next = NULL;
2826 			kmem_slab_free(cp, KMEM_BUF(cp, head));
2827 			head = next;
2828 			nbufs--;
2829 		}
2830 	}
2831 	ASSERT(head == NULL);
2832 	ASSERT(nbufs == 0);
2833 	mutex_enter(&cp->cache_lock);
2834 }
2835 
2836 void *
kmem_zalloc(size_t size,int kmflag)2837 kmem_zalloc(size_t size, int kmflag)
2838 {
2839 	size_t index;
2840 	void *buf;
2841 
2842 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2843 		kmem_cache_t *cp = kmem_alloc_table[index];
2844 		buf = kmem_cache_alloc(cp, kmflag);
2845 		if (buf != NULL) {
2846 			if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2847 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2848 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2849 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2850 
2851 				if (cp->cache_flags & KMF_LITE) {
2852 					KMEM_BUFTAG_LITE_ENTER(btp,
2853 					    kmem_lite_count, caller());
2854 				}
2855 			}
2856 			bzero(buf, size);
2857 		}
2858 	} else {
2859 		buf = kmem_alloc(size, kmflag);
2860 		if (buf != NULL)
2861 			bzero(buf, size);
2862 	}
2863 	return (buf);
2864 }
2865 
2866 void *
kmem_alloc(size_t size,int kmflag)2867 kmem_alloc(size_t size, int kmflag)
2868 {
2869 	size_t index;
2870 	kmem_cache_t *cp;
2871 	void *buf;
2872 
2873 	if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2874 		cp = kmem_alloc_table[index];
2875 		/* fall through to kmem_cache_alloc() */
2876 
2877 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2878 	    kmem_big_alloc_table_max) {
2879 		cp = kmem_big_alloc_table[index];
2880 		/* fall through to kmem_cache_alloc() */
2881 
2882 	} else {
2883 		if (size == 0)
2884 			return (NULL);
2885 
2886 		buf = vmem_alloc(kmem_oversize_arena, size,
2887 		    kmflag & KM_VMFLAGS);
2888 		if (buf == NULL)
2889 			kmem_log_event(kmem_failure_log, NULL, NULL,
2890 			    (void *)size);
2891 		else if (KMEM_DUMP(kmem_slab_cache)) {
2892 			/* stats for dump intercept */
2893 			kmem_dump_oversize_allocs++;
2894 			if (size > kmem_dump_oversize_max)
2895 				kmem_dump_oversize_max = size;
2896 		}
2897 		return (buf);
2898 	}
2899 
2900 	buf = kmem_cache_alloc(cp, kmflag);
2901 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2902 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2903 		((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2904 		((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2905 
2906 		if (cp->cache_flags & KMF_LITE) {
2907 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2908 		}
2909 	}
2910 	return (buf);
2911 }
2912 
2913 void
kmem_free(void * buf,size_t size)2914 kmem_free(void *buf, size_t size)
2915 {
2916 	size_t index;
2917 	kmem_cache_t *cp;
2918 
2919 	if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2920 		cp = kmem_alloc_table[index];
2921 		/* fall through to kmem_cache_free() */
2922 
2923 	} else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2924 	    kmem_big_alloc_table_max) {
2925 		cp = kmem_big_alloc_table[index];
2926 		/* fall through to kmem_cache_free() */
2927 
2928 	} else {
2929 		EQUIV(buf == NULL, size == 0);
2930 		if (buf == NULL && size == 0)
2931 			return;
2932 		vmem_free(kmem_oversize_arena, buf, size);
2933 		return;
2934 	}
2935 
2936 	if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2937 		kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2938 		uint32_t *ip = (uint32_t *)btp;
2939 		if (ip[1] != KMEM_SIZE_ENCODE(size)) {
2940 			if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
2941 				kmem_error(KMERR_DUPFREE, cp, buf);
2942 				return;
2943 			}
2944 			if (KMEM_SIZE_VALID(ip[1])) {
2945 				ip[0] = KMEM_SIZE_ENCODE(size);
2946 				kmem_error(KMERR_BADSIZE, cp, buf);
2947 			} else {
2948 				kmem_error(KMERR_REDZONE, cp, buf);
2949 			}
2950 			return;
2951 		}
2952 		if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
2953 			kmem_error(KMERR_REDZONE, cp, buf);
2954 			return;
2955 		}
2956 		btp->bt_redzone = KMEM_REDZONE_PATTERN;
2957 		if (cp->cache_flags & KMF_LITE) {
2958 			KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
2959 			    caller());
2960 		}
2961 	}
2962 	kmem_cache_free(cp, buf);
2963 }
2964 
2965 void *
kmem_firewall_va_alloc(vmem_t * vmp,size_t size,int vmflag)2966 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2967 {
2968 	size_t realsize = size + vmp->vm_quantum;
2969 	void *addr;
2970 
2971 	/*
2972 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2973 	 * vm_quantum will cause integer wraparound.  Check for this, and
2974 	 * blow off the firewall page in this case.  Note that such a
2975 	 * giant allocation (the entire kernel address space) can never
2976 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
2977 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2978 	 * corresponding check in kmem_firewall_va_free().
2979 	 */
2980 	if (realsize < size)
2981 		realsize = size;
2982 
2983 	/*
2984 	 * While boot still owns resource management, make sure that this
2985 	 * redzone virtual address allocation is properly accounted for in
2986 	 * OBPs "virtual-memory" "available" lists because we're
2987 	 * effectively claiming them for a red zone.  If we don't do this,
2988 	 * the available lists become too fragmented and too large for the
2989 	 * current boot/kernel memory list interface.
2990 	 */
2991 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
2992 
2993 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
2994 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
2995 
2996 	return (addr);
2997 }
2998 
2999 void
kmem_firewall_va_free(vmem_t * vmp,void * addr,size_t size)3000 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3001 {
3002 	ASSERT((kvseg.s_base == NULL ?
3003 	    va_to_pfn((char *)addr + size) :
3004 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3005 
3006 	vmem_free(vmp, addr, size + vmp->vm_quantum);
3007 }
3008 
3009 /*
3010  * Try to allocate at least `size' bytes of memory without sleeping or
3011  * panicking. Return actual allocated size in `asize'. If allocation failed,
3012  * try final allocation with sleep or panic allowed.
3013  */
3014 void *
kmem_alloc_tryhard(size_t size,size_t * asize,int kmflag)3015 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3016 {
3017 	void *p;
3018 
3019 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3020 	do {
3021 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3022 		if (p != NULL)
3023 			return (p);
3024 		*asize += KMEM_ALIGN;
3025 	} while (*asize <= PAGESIZE);
3026 
3027 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
3028 	return (kmem_alloc(*asize, kmflag));
3029 }
3030 
3031 /*
3032  * Reclaim all unused memory from a cache.
3033  */
3034 static void
kmem_cache_reap(kmem_cache_t * cp)3035 kmem_cache_reap(kmem_cache_t *cp)
3036 {
3037 	ASSERT(taskq_member(kmem_taskq, curthread));
3038 	cp->cache_reap++;
3039 
3040 	/*
3041 	 * Ask the cache's owner to free some memory if possible.
3042 	 * The idea is to handle things like the inode cache, which
3043 	 * typically sits on a bunch of memory that it doesn't truly
3044 	 * *need*.  Reclaim policy is entirely up to the owner; this
3045 	 * callback is just an advisory plea for help.
3046 	 */
3047 	if (cp->cache_reclaim != NULL) {
3048 		long delta;
3049 
3050 		/*
3051 		 * Reclaimed memory should be reapable (not included in the
3052 		 * depot's working set).
3053 		 */
3054 		delta = cp->cache_full.ml_total;
3055 		cp->cache_reclaim(cp->cache_private);
3056 		delta = cp->cache_full.ml_total - delta;
3057 		if (delta > 0) {
3058 			mutex_enter(&cp->cache_depot_lock);
3059 			cp->cache_full.ml_reaplimit += delta;
3060 			cp->cache_full.ml_min += delta;
3061 			mutex_exit(&cp->cache_depot_lock);
3062 		}
3063 	}
3064 
3065 	kmem_depot_ws_reap(cp);
3066 
3067 	if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3068 		kmem_cache_defrag(cp);
3069 	}
3070 }
3071 
3072 static void
kmem_reap_timeout(void * flag_arg)3073 kmem_reap_timeout(void *flag_arg)
3074 {
3075 	uint32_t *flag = (uint32_t *)flag_arg;
3076 
3077 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3078 	*flag = 0;
3079 }
3080 
3081 static void
kmem_reap_done(void * flag)3082 kmem_reap_done(void *flag)
3083 {
3084 	if (!callout_init_done) {
3085 		/* can't schedule a timeout at this point */
3086 		kmem_reap_timeout(flag);
3087 	} else {
3088 		(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3089 	}
3090 }
3091 
3092 static void
kmem_reap_start(void * flag)3093 kmem_reap_start(void *flag)
3094 {
3095 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3096 
3097 	if (flag == &kmem_reaping) {
3098 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3099 		/*
3100 		 * if we have segkp under heap, reap segkp cache.
3101 		 */
3102 		if (segkp_fromheap)
3103 			segkp_cache_free();
3104 	}
3105 	else
3106 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3107 
3108 	/*
3109 	 * We use taskq_dispatch() to schedule a timeout to clear
3110 	 * the flag so that kmem_reap() becomes self-throttling:
3111 	 * we won't reap again until the current reap completes *and*
3112 	 * at least kmem_reap_interval ticks have elapsed.
3113 	 */
3114 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3115 		kmem_reap_done(flag);
3116 }
3117 
3118 static void
kmem_reap_common(void * flag_arg)3119 kmem_reap_common(void *flag_arg)
3120 {
3121 	uint32_t *flag = (uint32_t *)flag_arg;
3122 
3123 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3124 	    atomic_cas_32(flag, 0, 1) != 0)
3125 		return;
3126 
3127 	/*
3128 	 * It may not be kosher to do memory allocation when a reap is called
3129 	 * (for example, if vmem_populate() is in the call chain).  So we
3130 	 * start the reap going with a TQ_NOALLOC dispatch.  If the dispatch
3131 	 * fails, we reset the flag, and the next reap will try again.
3132 	 */
3133 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3134 		*flag = 0;
3135 }
3136 
3137 /*
3138  * Reclaim all unused memory from all caches.  Called from the VM system
3139  * when memory gets tight.
3140  */
3141 void
kmem_reap(void)3142 kmem_reap(void)
3143 {
3144 	kmem_reap_common(&kmem_reaping);
3145 }
3146 
3147 /*
3148  * Reclaim all unused memory from identifier arenas, called when a vmem
3149  * arena not back by memory is exhausted.  Since reaping memory-backed caches
3150  * cannot help with identifier exhaustion, we avoid both a large amount of
3151  * work and unwanted side-effects from reclaim callbacks.
3152  */
3153 void
kmem_reap_idspace(void)3154 kmem_reap_idspace(void)
3155 {
3156 	kmem_reap_common(&kmem_reaping_idspace);
3157 }
3158 
3159 /*
3160  * Purge all magazines from a cache and set its magazine limit to zero.
3161  * All calls are serialized by the kmem_taskq lock, except for the final
3162  * call from kmem_cache_destroy().
3163  */
3164 static void
kmem_cache_magazine_purge(kmem_cache_t * cp)3165 kmem_cache_magazine_purge(kmem_cache_t *cp)
3166 {
3167 	kmem_cpu_cache_t *ccp;
3168 	kmem_magazine_t *mp, *pmp;
3169 	int rounds, prounds, cpu_seqid;
3170 
3171 	ASSERT(!list_link_active(&cp->cache_link) ||
3172 	    taskq_member(kmem_taskq, curthread));
3173 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3174 
3175 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3176 		ccp = &cp->cache_cpu[cpu_seqid];
3177 
3178 		mutex_enter(&ccp->cc_lock);
3179 		mp = ccp->cc_loaded;
3180 		pmp = ccp->cc_ploaded;
3181 		rounds = ccp->cc_rounds;
3182 		prounds = ccp->cc_prounds;
3183 		ccp->cc_loaded = NULL;
3184 		ccp->cc_ploaded = NULL;
3185 		ccp->cc_rounds = -1;
3186 		ccp->cc_prounds = -1;
3187 		ccp->cc_magsize = 0;
3188 		mutex_exit(&ccp->cc_lock);
3189 
3190 		if (mp)
3191 			kmem_magazine_destroy(cp, mp, rounds);
3192 		if (pmp)
3193 			kmem_magazine_destroy(cp, pmp, prounds);
3194 	}
3195 
3196 	kmem_depot_ws_zero(cp);
3197 	kmem_depot_ws_reap(cp);
3198 }
3199 
3200 /*
3201  * Enable per-cpu magazines on a cache.
3202  */
3203 static void
kmem_cache_magazine_enable(kmem_cache_t * cp)3204 kmem_cache_magazine_enable(kmem_cache_t *cp)
3205 {
3206 	int cpu_seqid;
3207 
3208 	if (cp->cache_flags & KMF_NOMAGAZINE)
3209 		return;
3210 
3211 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3212 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3213 		mutex_enter(&ccp->cc_lock);
3214 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3215 		mutex_exit(&ccp->cc_lock);
3216 	}
3217 
3218 }
3219 
3220 /*
3221  * Reap (almost) everything right now.
3222  */
3223 void
kmem_cache_reap_now(kmem_cache_t * cp)3224 kmem_cache_reap_now(kmem_cache_t *cp)
3225 {
3226 	ASSERT(list_link_active(&cp->cache_link));
3227 
3228 	kmem_depot_ws_zero(cp);
3229 
3230 	(void) taskq_dispatch(kmem_taskq,
3231 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3232 	taskq_wait(kmem_taskq);
3233 }
3234 
3235 /*
3236  * Recompute a cache's magazine size.  The trade-off is that larger magazines
3237  * provide a higher transfer rate with the depot, while smaller magazines
3238  * reduce memory consumption.  Magazine resizing is an expensive operation;
3239  * it should not be done frequently.
3240  *
3241  * Changes to the magazine size are serialized by the kmem_taskq lock.
3242  *
3243  * Note: at present this only grows the magazine size.  It might be useful
3244  * to allow shrinkage too.
3245  */
3246 static void
kmem_cache_magazine_resize(kmem_cache_t * cp)3247 kmem_cache_magazine_resize(kmem_cache_t *cp)
3248 {
3249 	kmem_magtype_t *mtp = cp->cache_magtype;
3250 
3251 	ASSERT(taskq_member(kmem_taskq, curthread));
3252 
3253 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
3254 		kmem_cache_magazine_purge(cp);
3255 		mutex_enter(&cp->cache_depot_lock);
3256 		cp->cache_magtype = ++mtp;
3257 		cp->cache_depot_contention_prev =
3258 		    cp->cache_depot_contention + INT_MAX;
3259 		mutex_exit(&cp->cache_depot_lock);
3260 		kmem_cache_magazine_enable(cp);
3261 	}
3262 }
3263 
3264 /*
3265  * Rescale a cache's hash table, so that the table size is roughly the
3266  * cache size.  We want the average lookup time to be extremely small.
3267  */
3268 static void
kmem_hash_rescale(kmem_cache_t * cp)3269 kmem_hash_rescale(kmem_cache_t *cp)
3270 {
3271 	kmem_bufctl_t **old_table, **new_table, *bcp;
3272 	size_t old_size, new_size, h;
3273 
3274 	ASSERT(taskq_member(kmem_taskq, curthread));
3275 
3276 	new_size = MAX(KMEM_HASH_INITIAL,
3277 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3278 	old_size = cp->cache_hash_mask + 1;
3279 
3280 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3281 		return;
3282 
3283 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3284 	    VM_NOSLEEP);
3285 	if (new_table == NULL)
3286 		return;
3287 	bzero(new_table, new_size * sizeof (void *));
3288 
3289 	mutex_enter(&cp->cache_lock);
3290 
3291 	old_size = cp->cache_hash_mask + 1;
3292 	old_table = cp->cache_hash_table;
3293 
3294 	cp->cache_hash_mask = new_size - 1;
3295 	cp->cache_hash_table = new_table;
3296 	cp->cache_rescale++;
3297 
3298 	for (h = 0; h < old_size; h++) {
3299 		bcp = old_table[h];
3300 		while (bcp != NULL) {
3301 			void *addr = bcp->bc_addr;
3302 			kmem_bufctl_t *next_bcp = bcp->bc_next;
3303 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3304 			bcp->bc_next = *hash_bucket;
3305 			*hash_bucket = bcp;
3306 			bcp = next_bcp;
3307 		}
3308 	}
3309 
3310 	mutex_exit(&cp->cache_lock);
3311 
3312 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3313 }
3314 
3315 /*
3316  * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3317  * update, magazine resizing, and slab consolidation.
3318  */
3319 static void
kmem_cache_update(kmem_cache_t * cp)3320 kmem_cache_update(kmem_cache_t *cp)
3321 {
3322 	int need_hash_rescale = 0;
3323 	int need_magazine_resize = 0;
3324 
3325 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
3326 
3327 	/*
3328 	 * If the cache has become much larger or smaller than its hash table,
3329 	 * fire off a request to rescale the hash table.
3330 	 */
3331 	mutex_enter(&cp->cache_lock);
3332 
3333 	if ((cp->cache_flags & KMF_HASH) &&
3334 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3335 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3336 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3337 		need_hash_rescale = 1;
3338 
3339 	mutex_exit(&cp->cache_lock);
3340 
3341 	/*
3342 	 * Update the depot working set statistics.
3343 	 */
3344 	kmem_depot_ws_update(cp);
3345 
3346 	/*
3347 	 * If there's a lot of contention in the depot,
3348 	 * increase the magazine size.
3349 	 */
3350 	mutex_enter(&cp->cache_depot_lock);
3351 
3352 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3353 	    (int)(cp->cache_depot_contention -
3354 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
3355 		need_magazine_resize = 1;
3356 
3357 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
3358 
3359 	mutex_exit(&cp->cache_depot_lock);
3360 
3361 	if (need_hash_rescale)
3362 		(void) taskq_dispatch(kmem_taskq,
3363 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3364 
3365 	if (need_magazine_resize)
3366 		(void) taskq_dispatch(kmem_taskq,
3367 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3368 
3369 	if (cp->cache_defrag != NULL)
3370 		(void) taskq_dispatch(kmem_taskq,
3371 		    (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3372 }
3373 
3374 static void kmem_update(void *);
3375 
3376 static void
kmem_update_timeout(void * dummy)3377 kmem_update_timeout(void *dummy)
3378 {
3379 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
3380 }
3381 
3382 static void
kmem_update(void * dummy)3383 kmem_update(void *dummy)
3384 {
3385 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3386 
3387 	/*
3388 	 * We use taskq_dispatch() to reschedule the timeout so that
3389 	 * kmem_update() becomes self-throttling: it won't schedule
3390 	 * new tasks until all previous tasks have completed.
3391 	 */
3392 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3393 		kmem_update_timeout(NULL);
3394 }
3395 
3396 static int
kmem_cache_kstat_update(kstat_t * ksp,int rw)3397 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3398 {
3399 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3400 	kmem_cache_t *cp = ksp->ks_private;
3401 	uint64_t cpu_buf_avail;
3402 	uint64_t buf_avail = 0;
3403 	int cpu_seqid;
3404 	long reap;
3405 
3406 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3407 
3408 	if (rw == KSTAT_WRITE)
3409 		return (EACCES);
3410 
3411 	mutex_enter(&cp->cache_lock);
3412 
3413 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
3414 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
3415 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
3416 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
3417 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
3418 
3419 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3420 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3421 
3422 		mutex_enter(&ccp->cc_lock);
3423 
3424 		cpu_buf_avail = 0;
3425 		if (ccp->cc_rounds > 0)
3426 			cpu_buf_avail += ccp->cc_rounds;
3427 		if (ccp->cc_prounds > 0)
3428 			cpu_buf_avail += ccp->cc_prounds;
3429 
3430 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
3431 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
3432 		buf_avail			+= cpu_buf_avail;
3433 
3434 		mutex_exit(&ccp->cc_lock);
3435 	}
3436 
3437 	mutex_enter(&cp->cache_depot_lock);
3438 
3439 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
3440 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
3441 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
3442 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
3443 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
3444 	kmcp->kmc_magazine_size.value.ui64	=
3445 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
3446 	    0 : cp->cache_magtype->mt_magsize;
3447 
3448 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
3449 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
3450 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3451 
3452 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3453 	reap = MIN(reap, cp->cache_full.ml_total);
3454 
3455 	mutex_exit(&cp->cache_depot_lock);
3456 
3457 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
3458 	kmcp->kmc_align.value.ui64	= cp->cache_align;
3459 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
3460 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
3461 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3462 	buf_avail += cp->cache_bufslab;
3463 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
3464 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
3465 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
3466 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
3467 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
3468 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
3469 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
3470 	    cp->cache_hash_mask + 1 : 0;
3471 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
3472 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
3473 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
3474 	kmcp->kmc_reap.value.ui64	= cp->cache_reap;
3475 
3476 	if (cp->cache_defrag == NULL) {
3477 		kmcp->kmc_move_callbacks.value.ui64	= 0;
3478 		kmcp->kmc_move_yes.value.ui64		= 0;
3479 		kmcp->kmc_move_no.value.ui64		= 0;
3480 		kmcp->kmc_move_later.value.ui64		= 0;
3481 		kmcp->kmc_move_dont_need.value.ui64	= 0;
3482 		kmcp->kmc_move_dont_know.value.ui64	= 0;
3483 		kmcp->kmc_move_hunt_found.value.ui64	= 0;
3484 		kmcp->kmc_move_slabs_freed.value.ui64	= 0;
3485 		kmcp->kmc_defrag.value.ui64		= 0;
3486 		kmcp->kmc_scan.value.ui64		= 0;
3487 		kmcp->kmc_move_reclaimable.value.ui64	= 0;
3488 	} else {
3489 		int64_t reclaimable;
3490 
3491 		kmem_defrag_t *kd = cp->cache_defrag;
3492 		kmcp->kmc_move_callbacks.value.ui64	= kd->kmd_callbacks;
3493 		kmcp->kmc_move_yes.value.ui64		= kd->kmd_yes;
3494 		kmcp->kmc_move_no.value.ui64		= kd->kmd_no;
3495 		kmcp->kmc_move_later.value.ui64		= kd->kmd_later;
3496 		kmcp->kmc_move_dont_need.value.ui64	= kd->kmd_dont_need;
3497 		kmcp->kmc_move_dont_know.value.ui64	= kd->kmd_dont_know;
3498 		kmcp->kmc_move_hunt_found.value.ui64	= kd->kmd_hunt_found;
3499 		kmcp->kmc_move_slabs_freed.value.ui64	= kd->kmd_slabs_freed;
3500 		kmcp->kmc_defrag.value.ui64		= kd->kmd_defrags;
3501 		kmcp->kmc_scan.value.ui64		= kd->kmd_scans;
3502 
3503 		reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3504 		reclaimable = MAX(reclaimable, 0);
3505 		reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3506 		kmcp->kmc_move_reclaimable.value.ui64	= reclaimable;
3507 	}
3508 
3509 	mutex_exit(&cp->cache_lock);
3510 	return (0);
3511 }
3512 
3513 /*
3514  * Return a named statistic about a particular cache.
3515  * This shouldn't be called very often, so it's currently designed for
3516  * simplicity (leverages existing kstat support) rather than efficiency.
3517  */
3518 uint64_t
kmem_cache_stat(kmem_cache_t * cp,char * name)3519 kmem_cache_stat(kmem_cache_t *cp, char *name)
3520 {
3521 	int i;
3522 	kstat_t *ksp = cp->cache_kstat;
3523 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3524 	uint64_t value = 0;
3525 
3526 	if (ksp != NULL) {
3527 		mutex_enter(&kmem_cache_kstat_lock);
3528 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3529 		for (i = 0; i < ksp->ks_ndata; i++) {
3530 			if (strcmp(knp[i].name, name) == 0) {
3531 				value = knp[i].value.ui64;
3532 				break;
3533 			}
3534 		}
3535 		mutex_exit(&kmem_cache_kstat_lock);
3536 	}
3537 	return (value);
3538 }
3539 
3540 /*
3541  * Return an estimate of currently available kernel heap memory.
3542  * On 32-bit systems, physical memory may exceed virtual memory,
3543  * we just truncate the result at 1GB.
3544  */
3545 size_t
kmem_avail(void)3546 kmem_avail(void)
3547 {
3548 	spgcnt_t rmem = availrmem - tune.t_minarmem;
3549 	spgcnt_t fmem = freemem - minfree;
3550 
3551 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3552 	    1 << (30 - PAGESHIFT))));
3553 }
3554 
3555 /*
3556  * Return the maximum amount of memory that is (in theory) allocatable
3557  * from the heap. This may be used as an estimate only since there
3558  * is no guarentee this space will still be available when an allocation
3559  * request is made, nor that the space may be allocated in one big request
3560  * due to kernel heap fragmentation.
3561  */
3562 size_t
kmem_maxavail(void)3563 kmem_maxavail(void)
3564 {
3565 	spgcnt_t pmem = availrmem - tune.t_minarmem;
3566 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3567 
3568 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3569 }
3570 
3571 /*
3572  * Indicate whether memory-intensive kmem debugging is enabled.
3573  */
3574 int
kmem_debugging(void)3575 kmem_debugging(void)
3576 {
3577 	return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3578 }
3579 
3580 /* binning function, sorts finely at the two extremes */
3581 #define	KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift)				\
3582 	((((sp)->slab_refcnt <= (binshift)) ||				\
3583 	    (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift)))	\
3584 	    ? -(sp)->slab_refcnt					\
3585 	    : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3586 
3587 /*
3588  * Minimizing the number of partial slabs on the freelist minimizes
3589  * fragmentation (the ratio of unused buffers held by the slab layer). There are
3590  * two ways to get a slab off of the freelist: 1) free all the buffers on the
3591  * slab, and 2) allocate all the buffers on the slab. It follows that we want
3592  * the most-used slabs at the front of the list where they have the best chance
3593  * of being completely allocated, and the least-used slabs at a safe distance
3594  * from the front to improve the odds that the few remaining buffers will all be
3595  * freed before another allocation can tie up the slab. For that reason a slab
3596  * with a higher slab_refcnt sorts less than than a slab with a lower
3597  * slab_refcnt.
3598  *
3599  * However, if a slab has at least one buffer that is deemed unfreeable, we
3600  * would rather have that slab at the front of the list regardless of
3601  * slab_refcnt, since even one unfreeable buffer makes the entire slab
3602  * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3603  * callback, the slab is marked unfreeable for as long as it remains on the
3604  * freelist.
3605  */
3606 static int
kmem_partial_slab_cmp(const void * p0,const void * p1)3607 kmem_partial_slab_cmp(const void *p0, const void *p1)
3608 {
3609 	const kmem_cache_t *cp;
3610 	const kmem_slab_t *s0 = p0;
3611 	const kmem_slab_t *s1 = p1;
3612 	int w0, w1;
3613 	size_t binshift;
3614 
3615 	ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3616 	ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3617 	ASSERT(s0->slab_cache == s1->slab_cache);
3618 	cp = s1->slab_cache;
3619 	ASSERT(MUTEX_HELD(&cp->cache_lock));
3620 	binshift = cp->cache_partial_binshift;
3621 
3622 	/* weight of first slab */
3623 	w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3624 	if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3625 		w0 -= cp->cache_maxchunks;
3626 	}
3627 
3628 	/* weight of second slab */
3629 	w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3630 	if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3631 		w1 -= cp->cache_maxchunks;
3632 	}
3633 
3634 	if (w0 < w1)
3635 		return (-1);
3636 	if (w0 > w1)
3637 		return (1);
3638 
3639 	/* compare pointer values */
3640 	if ((uintptr_t)s0 < (uintptr_t)s1)
3641 		return (-1);
3642 	if ((uintptr_t)s0 > (uintptr_t)s1)
3643 		return (1);
3644 
3645 	return (0);
3646 }
3647 
3648 /*
3649  * It must be valid to call the destructor (if any) on a newly created object.
3650  * That is, the constructor (if any) must leave the object in a valid state for
3651  * the destructor.
3652  */
3653 kmem_cache_t *
kmem_cache_create(char * name,size_t bufsize,size_t align,int (* constructor)(void *,void *,int),void (* destructor)(void *,void *),void (* reclaim)(void *),void * private,vmem_t * vmp,int cflags)3654 kmem_cache_create(
3655 	char *name,		/* descriptive name for this cache */
3656 	size_t bufsize,		/* size of the objects it manages */
3657 	size_t align,		/* required object alignment */
3658 	int (*constructor)(void *, void *, int), /* object constructor */
3659 	void (*destructor)(void *, void *),	/* object destructor */
3660 	void (*reclaim)(void *), /* memory reclaim callback */
3661 	void *private,		/* pass-thru arg for constr/destr/reclaim */
3662 	vmem_t *vmp,		/* vmem source for slab allocation */
3663 	int cflags)		/* cache creation flags */
3664 {
3665 	int cpu_seqid;
3666 	size_t chunksize;
3667 	kmem_cache_t *cp;
3668 	kmem_magtype_t *mtp;
3669 	size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3670 
3671 #ifdef	DEBUG
3672 	/*
3673 	 * Cache names should conform to the rules for valid C identifiers
3674 	 */
3675 	if (!strident_valid(name)) {
3676 		cmn_err(CE_CONT,
3677 		    "kmem_cache_create: '%s' is an invalid cache name\n"
3678 		    "cache names must conform to the rules for "
3679 		    "C identifiers\n", name);
3680 	}
3681 #endif	/* DEBUG */
3682 
3683 	if (vmp == NULL)
3684 		vmp = kmem_default_arena;
3685 
3686 	/*
3687 	 * If this kmem cache has an identifier vmem arena as its source, mark
3688 	 * it such to allow kmem_reap_idspace().
3689 	 */
3690 	ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
3691 	if (vmp->vm_cflags & VMC_IDENTIFIER)
3692 		cflags |= KMC_IDENTIFIER;
3693 
3694 	/*
3695 	 * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
3696 	 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3697 	 * false sharing of per-CPU data.
3698 	 */
3699 	cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3700 	    P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3701 	bzero(cp, csize);
3702 	list_link_init(&cp->cache_link);
3703 
3704 	if (align == 0)
3705 		align = KMEM_ALIGN;
3706 
3707 	/*
3708 	 * If we're not at least KMEM_ALIGN aligned, we can't use free
3709 	 * memory to hold bufctl information (because we can't safely
3710 	 * perform word loads and stores on it).
3711 	 */
3712 	if (align < KMEM_ALIGN)
3713 		cflags |= KMC_NOTOUCH;
3714 
3715 	if (!ISP2(align) || align > vmp->vm_quantum)
3716 		panic("kmem_cache_create: bad alignment %lu", align);
3717 
3718 	mutex_enter(&kmem_flags_lock);
3719 	if (kmem_flags & KMF_RANDOMIZE)
3720 		kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3721 		    KMF_RANDOMIZE;
3722 	cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3723 	mutex_exit(&kmem_flags_lock);
3724 
3725 	/*
3726 	 * Make sure all the various flags are reasonable.
3727 	 */
3728 	ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3729 
3730 	if (cp->cache_flags & KMF_LITE) {
3731 		if (bufsize >= kmem_lite_minsize &&
3732 		    align <= kmem_lite_maxalign &&
3733 		    P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3734 			cp->cache_flags |= KMF_BUFTAG;
3735 			cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3736 		} else {
3737 			cp->cache_flags &= ~KMF_DEBUG;
3738 		}
3739 	}
3740 
3741 	if (cp->cache_flags & KMF_DEADBEEF)
3742 		cp->cache_flags |= KMF_REDZONE;
3743 
3744 	if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3745 		cp->cache_flags |= KMF_NOMAGAZINE;
3746 
3747 	if (cflags & KMC_NODEBUG)
3748 		cp->cache_flags &= ~KMF_DEBUG;
3749 
3750 	if (cflags & KMC_NOTOUCH)
3751 		cp->cache_flags &= ~KMF_TOUCH;
3752 
3753 	if (cflags & KMC_PREFILL)
3754 		cp->cache_flags |= KMF_PREFILL;
3755 
3756 	if (cflags & KMC_NOHASH)
3757 		cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3758 
3759 	if (cflags & KMC_NOMAGAZINE)
3760 		cp->cache_flags |= KMF_NOMAGAZINE;
3761 
3762 	if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3763 		cp->cache_flags |= KMF_REDZONE;
3764 
3765 	if (!(cp->cache_flags & KMF_AUDIT))
3766 		cp->cache_flags &= ~KMF_CONTENTS;
3767 
3768 	if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3769 	    !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3770 		cp->cache_flags |= KMF_FIREWALL;
3771 
3772 	if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3773 		cp->cache_flags &= ~KMF_FIREWALL;
3774 
3775 	if (cp->cache_flags & KMF_FIREWALL) {
3776 		cp->cache_flags &= ~KMF_BUFTAG;
3777 		cp->cache_flags |= KMF_NOMAGAZINE;
3778 		ASSERT(vmp == kmem_default_arena);
3779 		vmp = kmem_firewall_arena;
3780 	}
3781 
3782 	/*
3783 	 * Set cache properties.
3784 	 */
3785 	(void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3786 	strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3787 	cp->cache_bufsize = bufsize;
3788 	cp->cache_align = align;
3789 	cp->cache_constructor = constructor;
3790 	cp->cache_destructor = destructor;
3791 	cp->cache_reclaim = reclaim;
3792 	cp->cache_private = private;
3793 	cp->cache_arena = vmp;
3794 	cp->cache_cflags = cflags;
3795 
3796 	/*
3797 	 * Determine the chunk size.
3798 	 */
3799 	chunksize = bufsize;
3800 
3801 	if (align >= KMEM_ALIGN) {
3802 		chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3803 		cp->cache_bufctl = chunksize - KMEM_ALIGN;
3804 	}
3805 
3806 	if (cp->cache_flags & KMF_BUFTAG) {
3807 		cp->cache_bufctl = chunksize;
3808 		cp->cache_buftag = chunksize;
3809 		if (cp->cache_flags & KMF_LITE)
3810 			chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3811 		else
3812 			chunksize += sizeof (kmem_buftag_t);
3813 	}
3814 
3815 	if (cp->cache_flags & KMF_DEADBEEF) {
3816 		cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3817 		if (cp->cache_flags & KMF_LITE)
3818 			cp->cache_verify = sizeof (uint64_t);
3819 	}
3820 
3821 	cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3822 
3823 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3824 
3825 	/*
3826 	 * Now that we know the chunk size, determine the optimal slab size.
3827 	 */
3828 	if (vmp == kmem_firewall_arena) {
3829 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3830 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
3831 		cp->cache_maxcolor = cp->cache_mincolor;
3832 		cp->cache_flags |= KMF_HASH;
3833 		ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3834 	} else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3835 	    !(cp->cache_flags & KMF_AUDIT) &&
3836 	    chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3837 		cp->cache_slabsize = vmp->vm_quantum;
3838 		cp->cache_mincolor = 0;
3839 		cp->cache_maxcolor =
3840 		    (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3841 		ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3842 		ASSERT(!(cp->cache_flags & KMF_AUDIT));
3843 	} else {
3844 		size_t chunks, bestfit, waste, slabsize;
3845 		size_t minwaste = LONG_MAX;
3846 
3847 		for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3848 			slabsize = P2ROUNDUP(chunksize * chunks,
3849 			    vmp->vm_quantum);
3850 			chunks = slabsize / chunksize;
3851 			waste = (slabsize % chunksize) / chunks;
3852 			if (waste < minwaste) {
3853 				minwaste = waste;
3854 				bestfit = slabsize;
3855 			}
3856 		}
3857 		if (cflags & KMC_QCACHE)
3858 			bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3859 		cp->cache_slabsize = bestfit;
3860 		cp->cache_mincolor = 0;
3861 		cp->cache_maxcolor = bestfit % chunksize;
3862 		cp->cache_flags |= KMF_HASH;
3863 	}
3864 
3865 	cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3866 	cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3867 
3868 	/*
3869 	 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3870 	 * there is a constructor avoids some tricky issues with debug setup
3871 	 * that may be revisited later. We cannot allow prefill in a
3872 	 * metadata cache because of potential recursion.
3873 	 */
3874 	if (vmp == kmem_msb_arena ||
3875 	    cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3876 	    cp->cache_constructor != NULL)
3877 		cp->cache_flags &= ~KMF_PREFILL;
3878 
3879 	if (cp->cache_flags & KMF_HASH) {
3880 		ASSERT(!(cflags & KMC_NOHASH));
3881 		cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3882 		    kmem_bufctl_audit_cache : kmem_bufctl_cache;
3883 	}
3884 
3885 	if (cp->cache_maxcolor >= vmp->vm_quantum)
3886 		cp->cache_maxcolor = vmp->vm_quantum - 1;
3887 
3888 	cp->cache_color = cp->cache_mincolor;
3889 
3890 	/*
3891 	 * Initialize the rest of the slab layer.
3892 	 */
3893 	mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3894 
3895 	avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3896 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3897 	/* LINTED: E_TRUE_LOGICAL_EXPR */
3898 	ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3899 	/* reuse partial slab AVL linkage for complete slab list linkage */
3900 	list_create(&cp->cache_complete_slabs,
3901 	    sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3902 
3903 	if (cp->cache_flags & KMF_HASH) {
3904 		cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3905 		    KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3906 		bzero(cp->cache_hash_table,
3907 		    KMEM_HASH_INITIAL * sizeof (void *));
3908 		cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3909 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3910 	}
3911 
3912 	/*
3913 	 * Initialize the depot.
3914 	 */
3915 	mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3916 
3917 	for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3918 		continue;
3919 
3920 	cp->cache_magtype = mtp;
3921 
3922 	/*
3923 	 * Initialize the CPU layer.
3924 	 */
3925 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3926 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3927 		mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3928 		ccp->cc_flags = cp->cache_flags;
3929 		ccp->cc_rounds = -1;
3930 		ccp->cc_prounds = -1;
3931 	}
3932 
3933 	/*
3934 	 * Create the cache's kstats.
3935 	 */
3936 	if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
3937 	    "kmem_cache", KSTAT_TYPE_NAMED,
3938 	    sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
3939 	    KSTAT_FLAG_VIRTUAL)) != NULL) {
3940 		cp->cache_kstat->ks_data = &kmem_cache_kstat;
3941 		cp->cache_kstat->ks_update = kmem_cache_kstat_update;
3942 		cp->cache_kstat->ks_private = cp;
3943 		cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
3944 		kstat_install(cp->cache_kstat);
3945 	}
3946 
3947 	/*
3948 	 * Add the cache to the global list.  This makes it visible
3949 	 * to kmem_update(), so the cache must be ready for business.
3950 	 */
3951 	mutex_enter(&kmem_cache_lock);
3952 	list_insert_tail(&kmem_caches, cp);
3953 	mutex_exit(&kmem_cache_lock);
3954 
3955 	if (kmem_ready)
3956 		kmem_cache_magazine_enable(cp);
3957 
3958 	return (cp);
3959 }
3960 
3961 static int
kmem_move_cmp(const void * buf,const void * p)3962 kmem_move_cmp(const void *buf, const void *p)
3963 {
3964 	const kmem_move_t *kmm = p;
3965 	uintptr_t v1 = (uintptr_t)buf;
3966 	uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
3967 	return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
3968 }
3969 
3970 static void
kmem_reset_reclaim_threshold(kmem_defrag_t * kmd)3971 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
3972 {
3973 	kmd->kmd_reclaim_numer = 1;
3974 }
3975 
3976 /*
3977  * Initially, when choosing candidate slabs for buffers to move, we want to be
3978  * very selective and take only slabs that are less than
3979  * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
3980  * slabs, then we raise the allocation ceiling incrementally. The reclaim
3981  * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
3982  * longer fragmented.
3983  */
3984 static void
kmem_adjust_reclaim_threshold(kmem_defrag_t * kmd,int direction)3985 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
3986 {
3987 	if (direction > 0) {
3988 		/* make it easier to find a candidate slab */
3989 		if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
3990 			kmd->kmd_reclaim_numer++;
3991 		}
3992 	} else {
3993 		/* be more selective */
3994 		if (kmd->kmd_reclaim_numer > 1) {
3995 			kmd->kmd_reclaim_numer--;
3996 		}
3997 	}
3998 }
3999 
4000 void
kmem_cache_set_move(kmem_cache_t * cp,kmem_cbrc_t (* move)(void *,void *,size_t,void *))4001 kmem_cache_set_move(kmem_cache_t *cp,
4002     kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4003 {
4004 	kmem_defrag_t *defrag;
4005 
4006 	ASSERT(move != NULL);
4007 	/*
4008 	 * The consolidator does not support NOTOUCH caches because kmem cannot
4009 	 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4010 	 * a low order bit usable by clients to distinguish uninitialized memory
4011 	 * from known objects (see kmem_slab_create).
4012 	 */
4013 	ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4014 	ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4015 
4016 	/*
4017 	 * We should not be holding anyone's cache lock when calling
4018 	 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4019 	 * lock.
4020 	 */
4021 	defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4022 
4023 	mutex_enter(&cp->cache_lock);
4024 
4025 	if (KMEM_IS_MOVABLE(cp)) {
4026 		if (cp->cache_move == NULL) {
4027 			ASSERT(cp->cache_slab_alloc == 0);
4028 
4029 			cp->cache_defrag = defrag;
4030 			defrag = NULL; /* nothing to free */
4031 			bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4032 			avl_create(&cp->cache_defrag->kmd_moves_pending,
4033 			    kmem_move_cmp, sizeof (kmem_move_t),
4034 			    offsetof(kmem_move_t, kmm_entry));
4035 			/* LINTED: E_TRUE_LOGICAL_EXPR */
4036 			ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4037 			/* reuse the slab's AVL linkage for deadlist linkage */
4038 			list_create(&cp->cache_defrag->kmd_deadlist,
4039 			    sizeof (kmem_slab_t),
4040 			    offsetof(kmem_slab_t, slab_link));
4041 			kmem_reset_reclaim_threshold(cp->cache_defrag);
4042 		}
4043 		cp->cache_move = move;
4044 	}
4045 
4046 	mutex_exit(&cp->cache_lock);
4047 
4048 	if (defrag != NULL) {
4049 		kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4050 	}
4051 }
4052 
4053 void
kmem_cache_destroy(kmem_cache_t * cp)4054 kmem_cache_destroy(kmem_cache_t *cp)
4055 {
4056 	int cpu_seqid;
4057 
4058 	/*
4059 	 * Remove the cache from the global cache list so that no one else
4060 	 * can schedule tasks on its behalf, wait for any pending tasks to
4061 	 * complete, purge the cache, and then destroy it.
4062 	 */
4063 	mutex_enter(&kmem_cache_lock);
4064 	list_remove(&kmem_caches, cp);
4065 	mutex_exit(&kmem_cache_lock);
4066 
4067 	if (kmem_taskq != NULL)
4068 		taskq_wait(kmem_taskq);
4069 	if (kmem_move_taskq != NULL)
4070 		taskq_wait(kmem_move_taskq);
4071 
4072 	kmem_cache_magazine_purge(cp);
4073 
4074 	mutex_enter(&cp->cache_lock);
4075 	if (cp->cache_buftotal != 0)
4076 		cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4077 		    cp->cache_name, (void *)cp);
4078 	if (cp->cache_defrag != NULL) {
4079 		avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4080 		list_destroy(&cp->cache_defrag->kmd_deadlist);
4081 		kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4082 		cp->cache_defrag = NULL;
4083 	}
4084 	/*
4085 	 * The cache is now dead.  There should be no further activity.  We
4086 	 * enforce this by setting land mines in the constructor, destructor,
4087 	 * reclaim, and move routines that induce a kernel text fault if
4088 	 * invoked.
4089 	 */
4090 	cp->cache_constructor = (int (*)(void *, void *, int))1;
4091 	cp->cache_destructor = (void (*)(void *, void *))2;
4092 	cp->cache_reclaim = (void (*)(void *))3;
4093 	cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4094 	mutex_exit(&cp->cache_lock);
4095 
4096 	kstat_delete(cp->cache_kstat);
4097 
4098 	if (cp->cache_hash_table != NULL)
4099 		vmem_free(kmem_hash_arena, cp->cache_hash_table,
4100 		    (cp->cache_hash_mask + 1) * sizeof (void *));
4101 
4102 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4103 		mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4104 
4105 	mutex_destroy(&cp->cache_depot_lock);
4106 	mutex_destroy(&cp->cache_lock);
4107 
4108 	vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4109 }
4110 
4111 /*ARGSUSED*/
4112 static int
kmem_cpu_setup(cpu_setup_t what,int id,void * arg)4113 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4114 {
4115 	ASSERT(MUTEX_HELD(&cpu_lock));
4116 	if (what == CPU_UNCONFIG) {
4117 		kmem_cache_applyall(kmem_cache_magazine_purge,
4118 		    kmem_taskq, TQ_SLEEP);
4119 		kmem_cache_applyall(kmem_cache_magazine_enable,
4120 		    kmem_taskq, TQ_SLEEP);
4121 	}
4122 	return (0);
4123 }
4124 
4125 static void
kmem_alloc_caches_create(const int * array,size_t count,kmem_cache_t ** alloc_table,size_t maxbuf,uint_t shift)4126 kmem_alloc_caches_create(const int *array, size_t count,
4127     kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4128 {
4129 	char name[KMEM_CACHE_NAMELEN + 1];
4130 	size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4131 	size_t size = table_unit;
4132 	int i;
4133 
4134 	for (i = 0; i < count; i++) {
4135 		size_t cache_size = array[i];
4136 		size_t align = KMEM_ALIGN;
4137 		kmem_cache_t *cp;
4138 
4139 		/* if the table has an entry for maxbuf, we're done */
4140 		if (size > maxbuf)
4141 			break;
4142 
4143 		/* cache size must be a multiple of the table unit */
4144 		ASSERT(P2PHASE(cache_size, table_unit) == 0);
4145 
4146 		/*
4147 		 * If they allocate a multiple of the coherency granularity,
4148 		 * they get a coherency-granularity-aligned address.
4149 		 */
4150 		if (IS_P2ALIGNED(cache_size, 64))
4151 			align = 64;
4152 		if (IS_P2ALIGNED(cache_size, PAGESIZE))
4153 			align = PAGESIZE;
4154 		(void) snprintf(name, sizeof (name),
4155 		    "kmem_alloc_%lu", cache_size);
4156 		cp = kmem_cache_create(name, cache_size, align,
4157 		    NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4158 
4159 		while (size <= cache_size) {
4160 			alloc_table[(size - 1) >> shift] = cp;
4161 			size += table_unit;
4162 		}
4163 	}
4164 
4165 	ASSERT(size > maxbuf);		/* i.e. maxbuf <= max(cache_size) */
4166 }
4167 
4168 static void
kmem_cache_init(int pass,int use_large_pages)4169 kmem_cache_init(int pass, int use_large_pages)
4170 {
4171 	int i;
4172 	size_t maxbuf;
4173 	kmem_magtype_t *mtp;
4174 
4175 	for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4176 		char name[KMEM_CACHE_NAMELEN + 1];
4177 
4178 		mtp = &kmem_magtype[i];
4179 		(void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4180 		mtp->mt_cache = kmem_cache_create(name,
4181 		    (mtp->mt_magsize + 1) * sizeof (void *),
4182 		    mtp->mt_align, NULL, NULL, NULL, NULL,
4183 		    kmem_msb_arena, KMC_NOHASH);
4184 	}
4185 
4186 	kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4187 	    sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4188 	    kmem_msb_arena, KMC_NOHASH);
4189 
4190 	kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4191 	    sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4192 	    kmem_msb_arena, KMC_NOHASH);
4193 
4194 	kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4195 	    sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4196 	    kmem_msb_arena, KMC_NOHASH);
4197 
4198 	if (pass == 2) {
4199 		kmem_va_arena = vmem_create("kmem_va",
4200 		    NULL, 0, PAGESIZE,
4201 		    vmem_alloc, vmem_free, heap_arena,
4202 		    8 * PAGESIZE, VM_SLEEP);
4203 
4204 		if (use_large_pages) {
4205 			kmem_default_arena = vmem_xcreate("kmem_default",
4206 			    NULL, 0, PAGESIZE,
4207 			    segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4208 			    0, VMC_DUMPSAFE | VM_SLEEP);
4209 		} else {
4210 			kmem_default_arena = vmem_create("kmem_default",
4211 			    NULL, 0, PAGESIZE,
4212 			    segkmem_alloc, segkmem_free, kmem_va_arena,
4213 			    0, VMC_DUMPSAFE | VM_SLEEP);
4214 		}
4215 
4216 		/* Figure out what our maximum cache size is */
4217 		maxbuf = kmem_max_cached;
4218 		if (maxbuf <= KMEM_MAXBUF) {
4219 			maxbuf = 0;
4220 			kmem_max_cached = KMEM_MAXBUF;
4221 		} else {
4222 			size_t size = 0;
4223 			size_t max =
4224 			    sizeof (kmem_big_alloc_sizes) / sizeof (int);
4225 			/*
4226 			 * Round maxbuf up to an existing cache size.  If maxbuf
4227 			 * is larger than the largest cache, we truncate it to
4228 			 * the largest cache's size.
4229 			 */
4230 			for (i = 0; i < max; i++) {
4231 				size = kmem_big_alloc_sizes[i];
4232 				if (maxbuf <= size)
4233 					break;
4234 			}
4235 			kmem_max_cached = maxbuf = size;
4236 		}
4237 
4238 		/*
4239 		 * The big alloc table may not be completely overwritten, so
4240 		 * we clear out any stale cache pointers from the first pass.
4241 		 */
4242 		bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4243 	} else {
4244 		/*
4245 		 * During the first pass, the kmem_alloc_* caches
4246 		 * are treated as metadata.
4247 		 */
4248 		kmem_default_arena = kmem_msb_arena;
4249 		maxbuf = KMEM_BIG_MAXBUF_32BIT;
4250 	}
4251 
4252 	/*
4253 	 * Set up the default caches to back kmem_alloc()
4254 	 */
4255 	kmem_alloc_caches_create(
4256 	    kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4257 	    kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4258 
4259 	kmem_alloc_caches_create(
4260 	    kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4261 	    kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4262 
4263 	kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4264 }
4265 
4266 void
kmem_init(void)4267 kmem_init(void)
4268 {
4269 	kmem_cache_t *cp;
4270 	int old_kmem_flags = kmem_flags;
4271 	int use_large_pages = 0;
4272 	size_t maxverify, minfirewall;
4273 
4274 	kstat_init();
4275 
4276 	/*
4277 	 * Don't do firewalled allocations if the heap is less than 1TB
4278 	 * (i.e. on a 32-bit kernel)
4279 	 * The resulting VM_NEXTFIT allocations would create too much
4280 	 * fragmentation in a small heap.
4281 	 */
4282 #if defined(_LP64)
4283 	maxverify = minfirewall = PAGESIZE / 2;
4284 #else
4285 	maxverify = minfirewall = ULONG_MAX;
4286 #endif
4287 
4288 	/* LINTED */
4289 	ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4290 
4291 	list_create(&kmem_caches, sizeof (kmem_cache_t),
4292 	    offsetof(kmem_cache_t, cache_link));
4293 
4294 	kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4295 	    vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4296 	    VM_SLEEP | VMC_NO_QCACHE);
4297 
4298 	kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4299 	    PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4300 	    VMC_DUMPSAFE | VM_SLEEP);
4301 
4302 	kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4303 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4304 
4305 	kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4306 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4307 
4308 	kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4309 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4310 
4311 	kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4312 	    NULL, 0, PAGESIZE,
4313 	    kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4314 	    0, VM_SLEEP);
4315 
4316 	kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4317 	    segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4318 	    VMC_DUMPSAFE | VM_SLEEP);
4319 
4320 	/* temporary oversize arena for mod_read_system_file */
4321 	kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4322 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4323 
4324 	kmem_reap_interval = 15 * hz;
4325 
4326 	/*
4327 	 * Read /etc/system.  This is a chicken-and-egg problem because
4328 	 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4329 	 * needs to use the allocator.  The simplest solution is to create
4330 	 * all the standard kmem caches, read /etc/system, destroy all the
4331 	 * caches we just created, and then create them all again in light
4332 	 * of the (possibly) new kmem_flags and other kmem tunables.
4333 	 */
4334 	kmem_cache_init(1, 0);
4335 
4336 	mod_read_system_file(boothowto & RB_ASKNAME);
4337 
4338 	while ((cp = list_tail(&kmem_caches)) != NULL)
4339 		kmem_cache_destroy(cp);
4340 
4341 	vmem_destroy(kmem_oversize_arena);
4342 
4343 	if (old_kmem_flags & KMF_STICKY)
4344 		kmem_flags = old_kmem_flags;
4345 
4346 	if (!(kmem_flags & KMF_AUDIT))
4347 		vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4348 
4349 	if (kmem_maxverify == 0)
4350 		kmem_maxverify = maxverify;
4351 
4352 	if (kmem_minfirewall == 0)
4353 		kmem_minfirewall = minfirewall;
4354 
4355 	/*
4356 	 * give segkmem a chance to figure out if we are using large pages
4357 	 * for the kernel heap
4358 	 */
4359 	use_large_pages = segkmem_lpsetup();
4360 
4361 	/*
4362 	 * To protect against corruption, we keep the actual number of callers
4363 	 * KMF_LITE records seperate from the tunable.  We arbitrarily clamp
4364 	 * to 16, since the overhead for small buffers quickly gets out of
4365 	 * hand.
4366 	 *
4367 	 * The real limit would depend on the needs of the largest KMC_NOHASH
4368 	 * cache.
4369 	 */
4370 	kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4371 	kmem_lite_pcs = kmem_lite_count;
4372 
4373 	/*
4374 	 * Normally, we firewall oversized allocations when possible, but
4375 	 * if we are using large pages for kernel memory, and we don't have
4376 	 * any non-LITE debugging flags set, we want to allocate oversized
4377 	 * buffers from large pages, and so skip the firewalling.
4378 	 */
4379 	if (use_large_pages &&
4380 	    ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4381 		kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4382 		    PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4383 		    0, VMC_DUMPSAFE | VM_SLEEP);
4384 	} else {
4385 		kmem_oversize_arena = vmem_create("kmem_oversize",
4386 		    NULL, 0, PAGESIZE,
4387 		    segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4388 		    kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4389 		    VM_SLEEP);
4390 	}
4391 
4392 	kmem_cache_init(2, use_large_pages);
4393 
4394 	if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4395 		if (kmem_transaction_log_size == 0)
4396 			kmem_transaction_log_size = kmem_maxavail() / 50;
4397 		kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4398 	}
4399 
4400 	if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4401 		if (kmem_content_log_size == 0)
4402 			kmem_content_log_size = kmem_maxavail() / 50;
4403 		kmem_content_log = kmem_log_init(kmem_content_log_size);
4404 	}
4405 
4406 	kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4407 
4408 	kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4409 
4410 	/*
4411 	 * Initialize STREAMS message caches so allocb() is available.
4412 	 * This allows us to initialize the logging framework (cmn_err(9F),
4413 	 * strlog(9F), etc) so we can start recording messages.
4414 	 */
4415 	streams_msg_init();
4416 
4417 	/*
4418 	 * Initialize the ZSD framework in Zones so modules loaded henceforth
4419 	 * can register their callbacks.
4420 	 */
4421 	zone_zsd_init();
4422 
4423 	log_init();
4424 	taskq_init();
4425 
4426 	/*
4427 	 * Warn about invalid or dangerous values of kmem_flags.
4428 	 * Always warn about unsupported values.
4429 	 */
4430 	if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4431 	    KMF_CONTENTS | KMF_LITE)) != 0) ||
4432 	    ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4433 		cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4434 		    "See the Solaris Tunable Parameters Reference Manual.",
4435 		    kmem_flags);
4436 
4437 #ifdef DEBUG
4438 	if ((kmem_flags & KMF_DEBUG) == 0)
4439 		cmn_err(CE_NOTE, "kmem debugging disabled.");
4440 #else
4441 	/*
4442 	 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4443 	 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4444 	 * if KMF_AUDIT is set). We should warn the user about the performance
4445 	 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4446 	 * isn't set (since that disables AUDIT).
4447 	 */
4448 	if (!(kmem_flags & KMF_LITE) &&
4449 	    (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4450 		cmn_err(CE_WARN, "High-overhead kmem debugging features "
4451 		    "enabled (kmem_flags = 0x%x).  Performance degradation "
4452 		    "and large memory overhead possible. See the Solaris "
4453 		    "Tunable Parameters Reference Manual.", kmem_flags);
4454 #endif /* not DEBUG */
4455 
4456 	kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4457 
4458 	kmem_ready = 1;
4459 
4460 	/*
4461 	 * Initialize the platform-specific aligned/DMA memory allocator.
4462 	 */
4463 	ka_init();
4464 
4465 	/*
4466 	 * Initialize 32-bit ID cache.
4467 	 */
4468 	id32_init();
4469 
4470 	/*
4471 	 * Initialize the networking stack so modules loaded can
4472 	 * register their callbacks.
4473 	 */
4474 	netstack_init();
4475 }
4476 
4477 static void
kmem_move_init(void)4478 kmem_move_init(void)
4479 {
4480 	kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4481 	    sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4482 	    kmem_msb_arena, KMC_NOHASH);
4483 	kmem_move_cache = kmem_cache_create("kmem_move_cache",
4484 	    sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4485 	    kmem_msb_arena, KMC_NOHASH);
4486 
4487 	/*
4488 	 * kmem guarantees that move callbacks are sequential and that even
4489 	 * across multiple caches no two moves ever execute simultaneously.
4490 	 * Move callbacks are processed on a separate taskq so that client code
4491 	 * does not interfere with internal maintenance tasks.
4492 	 */
4493 	kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4494 	    minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4495 }
4496 
4497 void
kmem_thread_init(void)4498 kmem_thread_init(void)
4499 {
4500 	kmem_move_init();
4501 	kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4502 	    300, INT_MAX, TASKQ_PREPOPULATE);
4503 }
4504 
4505 void
kmem_mp_init(void)4506 kmem_mp_init(void)
4507 {
4508 	mutex_enter(&cpu_lock);
4509 	register_cpu_setup_func(kmem_cpu_setup, NULL);
4510 	mutex_exit(&cpu_lock);
4511 
4512 	kmem_update_timeout(NULL);
4513 
4514 	taskq_mp_init();
4515 }
4516 
4517 /*
4518  * Return the slab of the allocated buffer, or NULL if the buffer is not
4519  * allocated. This function may be called with a known slab address to determine
4520  * whether or not the buffer is allocated, or with a NULL slab address to obtain
4521  * an allocated buffer's slab.
4522  */
4523 static kmem_slab_t *
kmem_slab_allocated(kmem_cache_t * cp,kmem_slab_t * sp,void * buf)4524 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4525 {
4526 	kmem_bufctl_t *bcp, *bufbcp;
4527 
4528 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4529 	ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4530 
4531 	if (cp->cache_flags & KMF_HASH) {
4532 		for (bcp = *KMEM_HASH(cp, buf);
4533 		    (bcp != NULL) && (bcp->bc_addr != buf);
4534 		    bcp = bcp->bc_next) {
4535 			continue;
4536 		}
4537 		ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4538 		return (bcp == NULL ? NULL : bcp->bc_slab);
4539 	}
4540 
4541 	if (sp == NULL) {
4542 		sp = KMEM_SLAB(cp, buf);
4543 	}
4544 	bufbcp = KMEM_BUFCTL(cp, buf);
4545 	for (bcp = sp->slab_head;
4546 	    (bcp != NULL) && (bcp != bufbcp);
4547 	    bcp = bcp->bc_next) {
4548 		continue;
4549 	}
4550 	return (bcp == NULL ? sp : NULL);
4551 }
4552 
4553 static boolean_t
kmem_slab_is_reclaimable(kmem_cache_t * cp,kmem_slab_t * sp,int flags)4554 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4555 {
4556 	long refcnt = sp->slab_refcnt;
4557 
4558 	ASSERT(cp->cache_defrag != NULL);
4559 
4560 	/*
4561 	 * For code coverage we want to be able to move an object within the
4562 	 * same slab (the only partial slab) even if allocating the destination
4563 	 * buffer resulted in a completely allocated slab.
4564 	 */
4565 	if (flags & KMM_DEBUG) {
4566 		return ((flags & KMM_DESPERATE) ||
4567 		    ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4568 	}
4569 
4570 	/* If we're desperate, we don't care if the client said NO. */
4571 	if (flags & KMM_DESPERATE) {
4572 		return (refcnt < sp->slab_chunks); /* any partial */
4573 	}
4574 
4575 	if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4576 		return (B_FALSE);
4577 	}
4578 
4579 	if ((refcnt == 1) || kmem_move_any_partial) {
4580 		return (refcnt < sp->slab_chunks);
4581 	}
4582 
4583 	/*
4584 	 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4585 	 * slabs with a progressively higher percentage of used buffers can be
4586 	 * reclaimed until the cache as a whole is no longer fragmented.
4587 	 *
4588 	 *	sp->slab_refcnt   kmd_reclaim_numer
4589 	 *	--------------- < ------------------
4590 	 *	sp->slab_chunks   KMEM_VOID_FRACTION
4591 	 */
4592 	return ((refcnt * KMEM_VOID_FRACTION) <
4593 	    (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4594 }
4595 
4596 static void *
kmem_hunt_mag(kmem_cache_t * cp,kmem_magazine_t * m,int n,void * buf,void * tbuf)4597 kmem_hunt_mag(kmem_cache_t *cp, kmem_magazine_t *m, int n, void *buf,
4598     void *tbuf)
4599 {
4600 	int i;		/* magazine round index */
4601 
4602 	for (i = 0; i < n; i++) {
4603 		if (buf == m->mag_round[i]) {
4604 			if (cp->cache_flags & KMF_BUFTAG) {
4605 				(void) kmem_cache_free_debug(cp, tbuf,
4606 				    caller());
4607 			}
4608 			m->mag_round[i] = tbuf;
4609 			return (buf);
4610 		}
4611 	}
4612 
4613 	return (NULL);
4614 }
4615 
4616 /*
4617  * Hunt the magazine layer for the given buffer. If found, the buffer is
4618  * removed from the magazine layer and returned, otherwise NULL is returned.
4619  * The state of the returned buffer is freed and constructed.
4620  */
4621 static void *
kmem_hunt_mags(kmem_cache_t * cp,void * buf)4622 kmem_hunt_mags(kmem_cache_t *cp, void *buf)
4623 {
4624 	kmem_cpu_cache_t *ccp;
4625 	kmem_magazine_t	*m;
4626 	int cpu_seqid;
4627 	int n;		/* magazine rounds */
4628 	void *tbuf;	/* temporary swap buffer */
4629 
4630 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4631 
4632 	/*
4633 	 * Allocated a buffer to swap with the one we hope to pull out of a
4634 	 * magazine when found.
4635 	 */
4636 	tbuf = kmem_cache_alloc(cp, KM_NOSLEEP);
4637 	if (tbuf == NULL) {
4638 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_alloc_fail);
4639 		return (NULL);
4640 	}
4641 	if (tbuf == buf) {
4642 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_lucky);
4643 		if (cp->cache_flags & KMF_BUFTAG) {
4644 			(void) kmem_cache_free_debug(cp, buf, caller());
4645 		}
4646 		return (buf);
4647 	}
4648 
4649 	/* Hunt the depot. */
4650 	mutex_enter(&cp->cache_depot_lock);
4651 	n = cp->cache_magtype->mt_magsize;
4652 	for (m = cp->cache_full.ml_list; m != NULL; m = m->mag_next) {
4653 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4654 			mutex_exit(&cp->cache_depot_lock);
4655 			return (buf);
4656 		}
4657 	}
4658 	mutex_exit(&cp->cache_depot_lock);
4659 
4660 	/* Hunt the per-CPU magazines. */
4661 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4662 		ccp = &cp->cache_cpu[cpu_seqid];
4663 
4664 		mutex_enter(&ccp->cc_lock);
4665 		m = ccp->cc_loaded;
4666 		n = ccp->cc_rounds;
4667 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4668 			mutex_exit(&ccp->cc_lock);
4669 			return (buf);
4670 		}
4671 		m = ccp->cc_ploaded;
4672 		n = ccp->cc_prounds;
4673 		if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4674 			mutex_exit(&ccp->cc_lock);
4675 			return (buf);
4676 		}
4677 		mutex_exit(&ccp->cc_lock);
4678 	}
4679 
4680 	kmem_cache_free(cp, tbuf);
4681 	return (NULL);
4682 }
4683 
4684 /*
4685  * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4686  * or when the buffer is freed.
4687  */
4688 static void
kmem_slab_move_yes(kmem_cache_t * cp,kmem_slab_t * sp,void * from_buf)4689 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4690 {
4691 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4692 	ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4693 
4694 	if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4695 		return;
4696 	}
4697 
4698 	if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4699 		if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4700 			avl_remove(&cp->cache_partial_slabs, sp);
4701 			sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4702 			sp->slab_stuck_offset = (uint32_t)-1;
4703 			avl_add(&cp->cache_partial_slabs, sp);
4704 		}
4705 	} else {
4706 		sp->slab_later_count = 0;
4707 		sp->slab_stuck_offset = (uint32_t)-1;
4708 	}
4709 }
4710 
4711 static void
kmem_slab_move_no(kmem_cache_t * cp,kmem_slab_t * sp,void * from_buf)4712 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4713 {
4714 	ASSERT(taskq_member(kmem_move_taskq, curthread));
4715 	ASSERT(MUTEX_HELD(&cp->cache_lock));
4716 	ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4717 
4718 	if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4719 		return;
4720 	}
4721 
4722 	avl_remove(&cp->cache_partial_slabs, sp);
4723 	sp->slab_later_count = 0;
4724 	sp->slab_flags |= KMEM_SLAB_NOMOVE;
4725 	sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4726 	avl_add(&cp->cache_partial_slabs, sp);
4727 }
4728 
4729 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4730 
4731 /*
4732  * The move callback takes two buffer addresses, the buffer to be moved, and a
4733  * newly allocated and constructed buffer selected by kmem as the destination.
4734  * It also takes the size of the buffer and an optional user argument specified
4735  * at cache creation time. kmem guarantees that the buffer to be moved has not
4736  * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4737  * guarantee the present whereabouts of the buffer to be moved, so it is up to
4738  * the client to safely determine whether or not it is still using the buffer.
4739  * The client must not free either of the buffers passed to the move callback,
4740  * since kmem wants to free them directly to the slab layer. The client response
4741  * tells kmem which of the two buffers to free:
4742  *
4743  * YES		kmem frees the old buffer (the move was successful)
4744  * NO		kmem frees the new buffer, marks the slab of the old buffer
4745  *              non-reclaimable to avoid bothering the client again
4746  * LATER	kmem frees the new buffer, increments slab_later_count
4747  * DONT_KNOW	kmem frees the new buffer, searches mags for the old buffer
4748  * DONT_NEED	kmem frees both the old buffer and the new buffer
4749  *
4750  * The pending callback argument now being processed contains both of the
4751  * buffers (old and new) passed to the move callback function, the slab of the
4752  * old buffer, and flags related to the move request, such as whether or not the
4753  * system was desperate for memory.
4754  *
4755  * Slabs are not freed while there is a pending callback, but instead are kept
4756  * on a deadlist, which is drained after the last callback completes. This means
4757  * that slabs are safe to access until kmem_move_end(), no matter how many of
4758  * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4759  * zero for as long as the slab remains on the deadlist and until the slab is
4760  * freed.
4761  */
4762 static void
kmem_move_buffer(kmem_move_t * callback)4763 kmem_move_buffer(kmem_move_t *callback)
4764 {
4765 	kmem_cbrc_t response;
4766 	kmem_slab_t *sp = callback->kmm_from_slab;
4767 	kmem_cache_t *cp = sp->slab_cache;
4768 	boolean_t free_on_slab;
4769 
4770 	ASSERT(taskq_member(kmem_move_taskq, curthread));
4771 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4772 	ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4773 
4774 	/*
4775 	 * The number of allocated buffers on the slab may have changed since we
4776 	 * last checked the slab's reclaimability (when the pending move was
4777 	 * enqueued), or the client may have responded NO when asked to move
4778 	 * another buffer on the same slab.
4779 	 */
4780 	if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4781 		KMEM_STAT_ADD(kmem_move_stats.kms_no_longer_reclaimable);
4782 		KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4783 		    kmem_move_stats.kms_notify_no_longer_reclaimable);
4784 		kmem_slab_free(cp, callback->kmm_to_buf);
4785 		kmem_move_end(cp, callback);
4786 		return;
4787 	}
4788 
4789 	/*
4790 	 * Hunting magazines is expensive, so we'll wait to do that until the
4791 	 * client responds KMEM_CBRC_DONT_KNOW. However, checking the slab layer
4792 	 * is cheap, so we might as well do that here in case we can avoid
4793 	 * bothering the client.
4794 	 */
4795 	mutex_enter(&cp->cache_lock);
4796 	free_on_slab = (kmem_slab_allocated(cp, sp,
4797 	    callback->kmm_from_buf) == NULL);
4798 	mutex_exit(&cp->cache_lock);
4799 
4800 	if (free_on_slab) {
4801 		KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_slab);
4802 		kmem_slab_free(cp, callback->kmm_to_buf);
4803 		kmem_move_end(cp, callback);
4804 		return;
4805 	}
4806 
4807 	if (cp->cache_flags & KMF_BUFTAG) {
4808 		/*
4809 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
4810 		 */
4811 		if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4812 		    KM_NOSLEEP, 1, caller()) != 0) {
4813 			KMEM_STAT_ADD(kmem_move_stats.kms_alloc_fail);
4814 			kmem_move_end(cp, callback);
4815 			return;
4816 		}
4817 	} else if (cp->cache_constructor != NULL &&
4818 	    cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4819 	    KM_NOSLEEP) != 0) {
4820 		atomic_inc_64(&cp->cache_alloc_fail);
4821 		KMEM_STAT_ADD(kmem_move_stats.kms_constructor_fail);
4822 		kmem_slab_free(cp, callback->kmm_to_buf);
4823 		kmem_move_end(cp, callback);
4824 		return;
4825 	}
4826 
4827 	KMEM_STAT_ADD(kmem_move_stats.kms_callbacks);
4828 	KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4829 	    kmem_move_stats.kms_notify_callbacks);
4830 	cp->cache_defrag->kmd_callbacks++;
4831 	cp->cache_defrag->kmd_thread = curthread;
4832 	cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4833 	cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4834 	DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4835 	    callback);
4836 
4837 	response = cp->cache_move(callback->kmm_from_buf,
4838 	    callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4839 
4840 	DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4841 	    callback, kmem_cbrc_t, response);
4842 	cp->cache_defrag->kmd_thread = NULL;
4843 	cp->cache_defrag->kmd_from_buf = NULL;
4844 	cp->cache_defrag->kmd_to_buf = NULL;
4845 
4846 	if (response == KMEM_CBRC_YES) {
4847 		KMEM_STAT_ADD(kmem_move_stats.kms_yes);
4848 		cp->cache_defrag->kmd_yes++;
4849 		kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4850 		/* slab safe to access until kmem_move_end() */
4851 		if (sp->slab_refcnt == 0)
4852 			cp->cache_defrag->kmd_slabs_freed++;
4853 		mutex_enter(&cp->cache_lock);
4854 		kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4855 		mutex_exit(&cp->cache_lock);
4856 		kmem_move_end(cp, callback);
4857 		return;
4858 	}
4859 
4860 	switch (response) {
4861 	case KMEM_CBRC_NO:
4862 		KMEM_STAT_ADD(kmem_move_stats.kms_no);
4863 		cp->cache_defrag->kmd_no++;
4864 		mutex_enter(&cp->cache_lock);
4865 		kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4866 		mutex_exit(&cp->cache_lock);
4867 		break;
4868 	case KMEM_CBRC_LATER:
4869 		KMEM_STAT_ADD(kmem_move_stats.kms_later);
4870 		cp->cache_defrag->kmd_later++;
4871 		mutex_enter(&cp->cache_lock);
4872 		if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4873 			mutex_exit(&cp->cache_lock);
4874 			break;
4875 		}
4876 
4877 		if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4878 			KMEM_STAT_ADD(kmem_move_stats.kms_disbelief);
4879 			kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4880 		} else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4881 			sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4882 			    callback->kmm_from_buf);
4883 		}
4884 		mutex_exit(&cp->cache_lock);
4885 		break;
4886 	case KMEM_CBRC_DONT_NEED:
4887 		KMEM_STAT_ADD(kmem_move_stats.kms_dont_need);
4888 		cp->cache_defrag->kmd_dont_need++;
4889 		kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4890 		if (sp->slab_refcnt == 0)
4891 			cp->cache_defrag->kmd_slabs_freed++;
4892 		mutex_enter(&cp->cache_lock);
4893 		kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4894 		mutex_exit(&cp->cache_lock);
4895 		break;
4896 	case KMEM_CBRC_DONT_KNOW:
4897 		KMEM_STAT_ADD(kmem_move_stats.kms_dont_know);
4898 		cp->cache_defrag->kmd_dont_know++;
4899 		if (kmem_hunt_mags(cp, callback->kmm_from_buf) != NULL) {
4900 			KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_mag);
4901 			cp->cache_defrag->kmd_hunt_found++;
4902 			kmem_slab_free_constructed(cp, callback->kmm_from_buf,
4903 			    B_TRUE);
4904 			if (sp->slab_refcnt == 0)
4905 				cp->cache_defrag->kmd_slabs_freed++;
4906 			mutex_enter(&cp->cache_lock);
4907 			kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4908 			mutex_exit(&cp->cache_lock);
4909 		}
4910 		break;
4911 	default:
4912 		panic("'%s' (%p) unexpected move callback response %d\n",
4913 		    cp->cache_name, (void *)cp, response);
4914 	}
4915 
4916 	kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4917 	kmem_move_end(cp, callback);
4918 }
4919 
4920 /* Return B_FALSE if there is insufficient memory for the move request. */
4921 static boolean_t
kmem_move_begin(kmem_cache_t * cp,kmem_slab_t * sp,void * buf,int flags)4922 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4923 {
4924 	void *to_buf;
4925 	avl_index_t index;
4926 	kmem_move_t *callback, *pending;
4927 	ulong_t n;
4928 
4929 	ASSERT(taskq_member(kmem_taskq, curthread));
4930 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4931 	ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4932 
4933 	callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4934 	if (callback == NULL) {
4935 		KMEM_STAT_ADD(kmem_move_stats.kms_callback_alloc_fail);
4936 		return (B_FALSE);
4937 	}
4938 
4939 	callback->kmm_from_slab = sp;
4940 	callback->kmm_from_buf = buf;
4941 	callback->kmm_flags = flags;
4942 
4943 	mutex_enter(&cp->cache_lock);
4944 
4945 	n = avl_numnodes(&cp->cache_partial_slabs);
4946 	if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
4947 		mutex_exit(&cp->cache_lock);
4948 		kmem_cache_free(kmem_move_cache, callback);
4949 		return (B_TRUE); /* there is no need for the move request */
4950 	}
4951 
4952 	pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
4953 	if (pending != NULL) {
4954 		/*
4955 		 * If the move is already pending and we're desperate now,
4956 		 * update the move flags.
4957 		 */
4958 		if (flags & KMM_DESPERATE) {
4959 			pending->kmm_flags |= KMM_DESPERATE;
4960 		}
4961 		mutex_exit(&cp->cache_lock);
4962 		KMEM_STAT_ADD(kmem_move_stats.kms_already_pending);
4963 		kmem_cache_free(kmem_move_cache, callback);
4964 		return (B_TRUE);
4965 	}
4966 
4967 	to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
4968 	    B_FALSE);
4969 	callback->kmm_to_buf = to_buf;
4970 	avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
4971 
4972 	mutex_exit(&cp->cache_lock);
4973 
4974 	if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
4975 	    callback, TQ_NOSLEEP)) {
4976 		KMEM_STAT_ADD(kmem_move_stats.kms_callback_taskq_fail);
4977 		mutex_enter(&cp->cache_lock);
4978 		avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4979 		mutex_exit(&cp->cache_lock);
4980 		kmem_slab_free(cp, to_buf);
4981 		kmem_cache_free(kmem_move_cache, callback);
4982 		return (B_FALSE);
4983 	}
4984 
4985 	return (B_TRUE);
4986 }
4987 
4988 static void
kmem_move_end(kmem_cache_t * cp,kmem_move_t * callback)4989 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
4990 {
4991 	avl_index_t index;
4992 
4993 	ASSERT(cp->cache_defrag != NULL);
4994 	ASSERT(taskq_member(kmem_move_taskq, curthread));
4995 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4996 
4997 	mutex_enter(&cp->cache_lock);
4998 	VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
4999 	    callback->kmm_from_buf, &index) != NULL);
5000 	avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5001 	if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
5002 		list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5003 		kmem_slab_t *sp;
5004 
5005 		/*
5006 		 * The last pending move completed. Release all slabs from the
5007 		 * front of the dead list except for any slab at the tail that
5008 		 * needs to be released from the context of kmem_move_buffers().
5009 		 * kmem deferred unmapping the buffers on these slabs in order
5010 		 * to guarantee that buffers passed to the move callback have
5011 		 * been touched only by kmem or by the client itself.
5012 		 */
5013 		while ((sp = list_remove_head(deadlist)) != NULL) {
5014 			if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
5015 				list_insert_tail(deadlist, sp);
5016 				break;
5017 			}
5018 			cp->cache_defrag->kmd_deadcount--;
5019 			cp->cache_slab_destroy++;
5020 			mutex_exit(&cp->cache_lock);
5021 			kmem_slab_destroy(cp, sp);
5022 			KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5023 			mutex_enter(&cp->cache_lock);
5024 		}
5025 	}
5026 	mutex_exit(&cp->cache_lock);
5027 	kmem_cache_free(kmem_move_cache, callback);
5028 }
5029 
5030 /*
5031  * Move buffers from least used slabs first by scanning backwards from the end
5032  * of the partial slab list. Scan at most max_scan candidate slabs and move
5033  * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
5034  * If desperate to reclaim memory, move buffers from any partial slab, otherwise
5035  * skip slabs with a ratio of allocated buffers at or above the current
5036  * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5037  * scan is aborted) so that the caller can adjust the reclaimability threshold
5038  * depending on how many reclaimable slabs it finds.
5039  *
5040  * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5041  * move request, since it is not valid for kmem_move_begin() to call
5042  * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5043  */
5044 static int
kmem_move_buffers(kmem_cache_t * cp,size_t max_scan,size_t max_slabs,int flags)5045 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5046     int flags)
5047 {
5048 	kmem_slab_t *sp;
5049 	void *buf;
5050 	int i, j; /* slab index, buffer index */
5051 	int s; /* reclaimable slabs */
5052 	int b; /* allocated (movable) buffers on reclaimable slab */
5053 	boolean_t success;
5054 	int refcnt;
5055 	int nomove;
5056 
5057 	ASSERT(taskq_member(kmem_taskq, curthread));
5058 	ASSERT(MUTEX_HELD(&cp->cache_lock));
5059 	ASSERT(kmem_move_cache != NULL);
5060 	ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5061 	ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5062 	    avl_numnodes(&cp->cache_partial_slabs) > 1);
5063 
5064 	if (kmem_move_blocked) {
5065 		return (0);
5066 	}
5067 
5068 	if (kmem_move_fulltilt) {
5069 		flags |= KMM_DESPERATE;
5070 	}
5071 
5072 	if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5073 		/*
5074 		 * Scan as many slabs as needed to find the desired number of
5075 		 * candidate slabs.
5076 		 */
5077 		max_scan = (size_t)-1;
5078 	}
5079 
5080 	if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5081 		/* Find as many candidate slabs as possible. */
5082 		max_slabs = (size_t)-1;
5083 	}
5084 
5085 	sp = avl_last(&cp->cache_partial_slabs);
5086 	ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5087 	for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5088 	    ((sp != avl_first(&cp->cache_partial_slabs)) ||
5089 	    (flags & KMM_DEBUG));
5090 	    sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5091 
5092 		if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5093 			continue;
5094 		}
5095 		s++;
5096 
5097 		/* Look for allocated buffers to move. */
5098 		for (j = 0, b = 0, buf = sp->slab_base;
5099 		    (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5100 		    buf = (((char *)buf) + cp->cache_chunksize), j++) {
5101 
5102 			if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5103 				continue;
5104 			}
5105 
5106 			b++;
5107 
5108 			/*
5109 			 * Prevent the slab from being destroyed while we drop
5110 			 * cache_lock and while the pending move is not yet
5111 			 * registered. Flag the pending move while
5112 			 * kmd_moves_pending may still be empty, since we can't
5113 			 * yet rely on a non-zero pending move count to prevent
5114 			 * the slab from being destroyed.
5115 			 */
5116 			ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5117 			sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5118 			/*
5119 			 * Recheck refcnt and nomove after reacquiring the lock,
5120 			 * since these control the order of partial slabs, and
5121 			 * we want to know if we can pick up the scan where we
5122 			 * left off.
5123 			 */
5124 			refcnt = sp->slab_refcnt;
5125 			nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5126 			mutex_exit(&cp->cache_lock);
5127 
5128 			success = kmem_move_begin(cp, sp, buf, flags);
5129 
5130 			/*
5131 			 * Now, before the lock is reacquired, kmem could
5132 			 * process all pending move requests and purge the
5133 			 * deadlist, so that upon reacquiring the lock, sp has
5134 			 * been remapped. Or, the client may free all the
5135 			 * objects on the slab while the pending moves are still
5136 			 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5137 			 * flag causes the slab to be put at the end of the
5138 			 * deadlist and prevents it from being destroyed, since
5139 			 * we plan to destroy it here after reacquiring the
5140 			 * lock.
5141 			 */
5142 			mutex_enter(&cp->cache_lock);
5143 			ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5144 			sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5145 
5146 			if (sp->slab_refcnt == 0) {
5147 				list_t *deadlist =
5148 				    &cp->cache_defrag->kmd_deadlist;
5149 				list_remove(deadlist, sp);
5150 
5151 				if (!avl_is_empty(
5152 				    &cp->cache_defrag->kmd_moves_pending)) {
5153 					/*
5154 					 * A pending move makes it unsafe to
5155 					 * destroy the slab, because even though
5156 					 * the move is no longer needed, the
5157 					 * context where that is determined
5158 					 * requires the slab to exist.
5159 					 * Fortunately, a pending move also
5160 					 * means we don't need to destroy the
5161 					 * slab here, since it will get
5162 					 * destroyed along with any other slabs
5163 					 * on the deadlist after the last
5164 					 * pending move completes.
5165 					 */
5166 					list_insert_head(deadlist, sp);
5167 					KMEM_STAT_ADD(kmem_move_stats.
5168 					    kms_endscan_slab_dead);
5169 					return (-1);
5170 				}
5171 
5172 				/*
5173 				 * Destroy the slab now if it was completely
5174 				 * freed while we dropped cache_lock and there
5175 				 * are no pending moves. Since slab_refcnt
5176 				 * cannot change once it reaches zero, no new
5177 				 * pending moves from that slab are possible.
5178 				 */
5179 				cp->cache_defrag->kmd_deadcount--;
5180 				cp->cache_slab_destroy++;
5181 				mutex_exit(&cp->cache_lock);
5182 				kmem_slab_destroy(cp, sp);
5183 				KMEM_STAT_ADD(kmem_move_stats.
5184 				    kms_dead_slabs_freed);
5185 				KMEM_STAT_ADD(kmem_move_stats.
5186 				    kms_endscan_slab_destroyed);
5187 				mutex_enter(&cp->cache_lock);
5188 				/*
5189 				 * Since we can't pick up the scan where we left
5190 				 * off, abort the scan and say nothing about the
5191 				 * number of reclaimable slabs.
5192 				 */
5193 				return (-1);
5194 			}
5195 
5196 			if (!success) {
5197 				/*
5198 				 * Abort the scan if there is not enough memory
5199 				 * for the request and say nothing about the
5200 				 * number of reclaimable slabs.
5201 				 */
5202 				KMEM_STAT_COND_ADD(s < max_slabs,
5203 				    kmem_move_stats.kms_endscan_nomem);
5204 				return (-1);
5205 			}
5206 
5207 			/*
5208 			 * The slab's position changed while the lock was
5209 			 * dropped, so we don't know where we are in the
5210 			 * sequence any more.
5211 			 */
5212 			if (sp->slab_refcnt != refcnt) {
5213 				/*
5214 				 * If this is a KMM_DEBUG move, the slab_refcnt
5215 				 * may have changed because we allocated a
5216 				 * destination buffer on the same slab. In that
5217 				 * case, we're not interested in counting it.
5218 				 */
5219 				KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5220 				    (s < max_slabs),
5221 				    kmem_move_stats.kms_endscan_refcnt_changed);
5222 				return (-1);
5223 			}
5224 			if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove) {
5225 				KMEM_STAT_COND_ADD(s < max_slabs,
5226 				    kmem_move_stats.kms_endscan_nomove_changed);
5227 				return (-1);
5228 			}
5229 
5230 			/*
5231 			 * Generating a move request allocates a destination
5232 			 * buffer from the slab layer, bumping the first partial
5233 			 * slab if it is completely allocated. If the current
5234 			 * slab becomes the first partial slab as a result, we
5235 			 * can't continue to scan backwards.
5236 			 *
5237 			 * If this is a KMM_DEBUG move and we allocated the
5238 			 * destination buffer from the last partial slab, then
5239 			 * the buffer we're moving is on the same slab and our
5240 			 * slab_refcnt has changed, causing us to return before
5241 			 * reaching here if there are no partial slabs left.
5242 			 */
5243 			ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5244 			if (sp == avl_first(&cp->cache_partial_slabs)) {
5245 				/*
5246 				 * We're not interested in a second KMM_DEBUG
5247 				 * move.
5248 				 */
5249 				goto end_scan;
5250 			}
5251 		}
5252 	}
5253 end_scan:
5254 
5255 	KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5256 	    (s < max_slabs) &&
5257 	    (sp == avl_first(&cp->cache_partial_slabs)),
5258 	    kmem_move_stats.kms_endscan_freelist);
5259 
5260 	return (s);
5261 }
5262 
5263 typedef struct kmem_move_notify_args {
5264 	kmem_cache_t *kmna_cache;
5265 	void *kmna_buf;
5266 } kmem_move_notify_args_t;
5267 
5268 static void
kmem_cache_move_notify_task(void * arg)5269 kmem_cache_move_notify_task(void *arg)
5270 {
5271 	kmem_move_notify_args_t *args = arg;
5272 	kmem_cache_t *cp = args->kmna_cache;
5273 	void *buf = args->kmna_buf;
5274 	kmem_slab_t *sp;
5275 
5276 	ASSERT(taskq_member(kmem_taskq, curthread));
5277 	ASSERT(list_link_active(&cp->cache_link));
5278 
5279 	kmem_free(args, sizeof (kmem_move_notify_args_t));
5280 	mutex_enter(&cp->cache_lock);
5281 	sp = kmem_slab_allocated(cp, NULL, buf);
5282 
5283 	/* Ignore the notification if the buffer is no longer allocated. */
5284 	if (sp == NULL) {
5285 		mutex_exit(&cp->cache_lock);
5286 		return;
5287 	}
5288 
5289 	/* Ignore the notification if there's no reason to move the buffer. */
5290 	if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5291 		/*
5292 		 * So far the notification is not ignored. Ignore the
5293 		 * notification if the slab is not marked by an earlier refusal
5294 		 * to move a buffer.
5295 		 */
5296 		if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5297 		    (sp->slab_later_count == 0)) {
5298 			mutex_exit(&cp->cache_lock);
5299 			return;
5300 		}
5301 
5302 		kmem_slab_move_yes(cp, sp, buf);
5303 		ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5304 		sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5305 		mutex_exit(&cp->cache_lock);
5306 		/* see kmem_move_buffers() about dropping the lock */
5307 		(void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5308 		mutex_enter(&cp->cache_lock);
5309 		ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5310 		sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5311 		if (sp->slab_refcnt == 0) {
5312 			list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5313 			list_remove(deadlist, sp);
5314 
5315 			if (!avl_is_empty(
5316 			    &cp->cache_defrag->kmd_moves_pending)) {
5317 				list_insert_head(deadlist, sp);
5318 				mutex_exit(&cp->cache_lock);
5319 				KMEM_STAT_ADD(kmem_move_stats.
5320 				    kms_notify_slab_dead);
5321 				return;
5322 			}
5323 
5324 			cp->cache_defrag->kmd_deadcount--;
5325 			cp->cache_slab_destroy++;
5326 			mutex_exit(&cp->cache_lock);
5327 			kmem_slab_destroy(cp, sp);
5328 			KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5329 			KMEM_STAT_ADD(kmem_move_stats.
5330 			    kms_notify_slab_destroyed);
5331 			return;
5332 		}
5333 	} else {
5334 		kmem_slab_move_yes(cp, sp, buf);
5335 	}
5336 	mutex_exit(&cp->cache_lock);
5337 }
5338 
5339 void
kmem_cache_move_notify(kmem_cache_t * cp,void * buf)5340 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5341 {
5342 	kmem_move_notify_args_t *args;
5343 
5344 	KMEM_STAT_ADD(kmem_move_stats.kms_notify);
5345 	args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5346 	if (args != NULL) {
5347 		args->kmna_cache = cp;
5348 		args->kmna_buf = buf;
5349 		if (!taskq_dispatch(kmem_taskq,
5350 		    (task_func_t *)kmem_cache_move_notify_task, args,
5351 		    TQ_NOSLEEP))
5352 			kmem_free(args, sizeof (kmem_move_notify_args_t));
5353 	}
5354 }
5355 
5356 static void
kmem_cache_defrag(kmem_cache_t * cp)5357 kmem_cache_defrag(kmem_cache_t *cp)
5358 {
5359 	size_t n;
5360 
5361 	ASSERT(cp->cache_defrag != NULL);
5362 
5363 	mutex_enter(&cp->cache_lock);
5364 	n = avl_numnodes(&cp->cache_partial_slabs);
5365 	if (n > 1) {
5366 		/* kmem_move_buffers() drops and reacquires cache_lock */
5367 		KMEM_STAT_ADD(kmem_move_stats.kms_defrags);
5368 		cp->cache_defrag->kmd_defrags++;
5369 		(void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5370 	}
5371 	mutex_exit(&cp->cache_lock);
5372 }
5373 
5374 /* Is this cache above the fragmentation threshold? */
5375 static boolean_t
kmem_cache_frag_threshold(kmem_cache_t * cp,uint64_t nfree)5376 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5377 {
5378 	/*
5379 	 *	nfree		kmem_frag_numer
5380 	 * ------------------ > ---------------
5381 	 * cp->cache_buftotal	kmem_frag_denom
5382 	 */
5383 	return ((nfree * kmem_frag_denom) >
5384 	    (cp->cache_buftotal * kmem_frag_numer));
5385 }
5386 
5387 static boolean_t
kmem_cache_is_fragmented(kmem_cache_t * cp,boolean_t * doreap)5388 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5389 {
5390 	boolean_t fragmented;
5391 	uint64_t nfree;
5392 
5393 	ASSERT(MUTEX_HELD(&cp->cache_lock));
5394 	*doreap = B_FALSE;
5395 
5396 	if (kmem_move_fulltilt) {
5397 		if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5398 			return (B_TRUE);
5399 		}
5400 	} else {
5401 		if ((cp->cache_complete_slab_count + avl_numnodes(
5402 		    &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5403 			return (B_FALSE);
5404 		}
5405 	}
5406 
5407 	nfree = cp->cache_bufslab;
5408 	fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5409 	    kmem_cache_frag_threshold(cp, nfree));
5410 
5411 	/*
5412 	 * Free buffers in the magazine layer appear allocated from the point of
5413 	 * view of the slab layer. We want to know if the slab layer would
5414 	 * appear fragmented if we included free buffers from magazines that
5415 	 * have fallen out of the working set.
5416 	 */
5417 	if (!fragmented) {
5418 		long reap;
5419 
5420 		mutex_enter(&cp->cache_depot_lock);
5421 		reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5422 		reap = MIN(reap, cp->cache_full.ml_total);
5423 		mutex_exit(&cp->cache_depot_lock);
5424 
5425 		nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5426 		if (kmem_cache_frag_threshold(cp, nfree)) {
5427 			*doreap = B_TRUE;
5428 		}
5429 	}
5430 
5431 	return (fragmented);
5432 }
5433 
5434 /* Called periodically from kmem_taskq */
5435 static void
kmem_cache_scan(kmem_cache_t * cp)5436 kmem_cache_scan(kmem_cache_t *cp)
5437 {
5438 	boolean_t reap = B_FALSE;
5439 	kmem_defrag_t *kmd;
5440 
5441 	ASSERT(taskq_member(kmem_taskq, curthread));
5442 
5443 	mutex_enter(&cp->cache_lock);
5444 
5445 	kmd = cp->cache_defrag;
5446 	if (kmd->kmd_consolidate > 0) {
5447 		kmd->kmd_consolidate--;
5448 		mutex_exit(&cp->cache_lock);
5449 		kmem_cache_reap(cp);
5450 		return;
5451 	}
5452 
5453 	if (kmem_cache_is_fragmented(cp, &reap)) {
5454 		size_t slabs_found;
5455 
5456 		/*
5457 		 * Consolidate reclaimable slabs from the end of the partial
5458 		 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5459 		 * reclaimable slabs). Keep track of how many candidate slabs we
5460 		 * looked for and how many we actually found so we can adjust
5461 		 * the definition of a candidate slab if we're having trouble
5462 		 * finding them.
5463 		 *
5464 		 * kmem_move_buffers() drops and reacquires cache_lock.
5465 		 */
5466 		KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5467 		kmd->kmd_scans++;
5468 		slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5469 		    kmem_reclaim_max_slabs, 0);
5470 		if (slabs_found >= 0) {
5471 			kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5472 			kmd->kmd_slabs_found += slabs_found;
5473 		}
5474 
5475 		if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5476 			kmd->kmd_tries = 0;
5477 
5478 			/*
5479 			 * If we had difficulty finding candidate slabs in
5480 			 * previous scans, adjust the threshold so that
5481 			 * candidates are easier to find.
5482 			 */
5483 			if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5484 				kmem_adjust_reclaim_threshold(kmd, -1);
5485 			} else if ((kmd->kmd_slabs_found * 2) <
5486 			    kmd->kmd_slabs_sought) {
5487 				kmem_adjust_reclaim_threshold(kmd, 1);
5488 			}
5489 			kmd->kmd_slabs_sought = 0;
5490 			kmd->kmd_slabs_found = 0;
5491 		}
5492 	} else {
5493 		kmem_reset_reclaim_threshold(cp->cache_defrag);
5494 #ifdef	DEBUG
5495 		if (!avl_is_empty(&cp->cache_partial_slabs)) {
5496 			/*
5497 			 * In a debug kernel we want the consolidator to
5498 			 * run occasionally even when there is plenty of
5499 			 * memory.
5500 			 */
5501 			uint16_t debug_rand;
5502 
5503 			(void) random_get_bytes((uint8_t *)&debug_rand, 2);
5504 			if (!kmem_move_noreap &&
5505 			    ((debug_rand % kmem_mtb_reap) == 0)) {
5506 				mutex_exit(&cp->cache_lock);
5507 				KMEM_STAT_ADD(kmem_move_stats.kms_debug_reaps);
5508 				kmem_cache_reap(cp);
5509 				return;
5510 			} else if ((debug_rand % kmem_mtb_move) == 0) {
5511 				KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5512 				KMEM_STAT_ADD(kmem_move_stats.kms_debug_scans);
5513 				kmd->kmd_scans++;
5514 				(void) kmem_move_buffers(cp,
5515 				    kmem_reclaim_scan_range, 1, KMM_DEBUG);
5516 			}
5517 		}
5518 #endif	/* DEBUG */
5519 	}
5520 
5521 	mutex_exit(&cp->cache_lock);
5522 
5523 	if (reap) {
5524 		KMEM_STAT_ADD(kmem_move_stats.kms_scan_depot_ws_reaps);
5525 		kmem_depot_ws_reap(cp);
5526 	}
5527 }
5528