xref: /illumos-gate/usr/src/uts/common/os/kmem.c (revision baf00aa8)
17c478bd9Sstevel@tonic-gate /*
27c478bd9Sstevel@tonic-gate  * CDDL HEADER START
37c478bd9Sstevel@tonic-gate  *
47c478bd9Sstevel@tonic-gate  * The contents of this file are subject to the terms of the
57d692464Sdp  * Common Development and Distribution License (the "License").
67d692464Sdp  * You may not use this file except in compliance with the License.
77c478bd9Sstevel@tonic-gate  *
87c478bd9Sstevel@tonic-gate  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
97c478bd9Sstevel@tonic-gate  * or http://www.opensolaris.org/os/licensing.
107c478bd9Sstevel@tonic-gate  * See the License for the specific language governing permissions
117c478bd9Sstevel@tonic-gate  * and limitations under the License.
127c478bd9Sstevel@tonic-gate  *
137c478bd9Sstevel@tonic-gate  * When distributing Covered Code, include this CDDL HEADER in each
147c478bd9Sstevel@tonic-gate  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
157c478bd9Sstevel@tonic-gate  * If applicable, add the following below this CDDL HEADER, with the
167c478bd9Sstevel@tonic-gate  * fields enclosed by brackets "[]" replaced with your own identifying
177c478bd9Sstevel@tonic-gate  * information: Portions Copyright [yyyy] [name of copyright owner]
187c478bd9Sstevel@tonic-gate  *
197c478bd9Sstevel@tonic-gate  * CDDL HEADER END
207c478bd9Sstevel@tonic-gate  */
217c478bd9Sstevel@tonic-gate /*
22b942e89bSDavid Valin  * Copyright (c) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
23929d5b43SMatthew Ahrens  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
240c833d64SJosef 'Jeff' Sipek  * Copyright 2015 Nexenta Systems, Inc.  All rights reserved.
2536a64e62STim Kordas  * Copyright 2018, Joyent, Inc.
26*baf00aa8SJoshua M. Clulow  * Copyright 2020 Oxide Computer Company
277c478bd9Sstevel@tonic-gate  */
287c478bd9Sstevel@tonic-gate 
297c478bd9Sstevel@tonic-gate /*
30b5fca8f8Stomee  * Kernel memory allocator, as described in the following two papers and a
31b5fca8f8Stomee  * statement about the consolidator:
327c478bd9Sstevel@tonic-gate  *
337c478bd9Sstevel@tonic-gate  * Jeff Bonwick,
347c478bd9Sstevel@tonic-gate  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
357c478bd9Sstevel@tonic-gate  * Proceedings of the Summer 1994 Usenix Conference.
367c478bd9Sstevel@tonic-gate  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
377c478bd9Sstevel@tonic-gate  *
387c478bd9Sstevel@tonic-gate  * Jeff Bonwick and Jonathan Adams,
397c478bd9Sstevel@tonic-gate  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
407c478bd9Sstevel@tonic-gate  * Arbitrary Resources.
417c478bd9Sstevel@tonic-gate  * Proceedings of the 2001 Usenix Conference.
427c478bd9Sstevel@tonic-gate  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
43b5fca8f8Stomee  *
44b5fca8f8Stomee  * kmem Slab Consolidator Big Theory Statement:
45b5fca8f8Stomee  *
46b5fca8f8Stomee  * 1. Motivation
47b5fca8f8Stomee  *
48b5fca8f8Stomee  * As stated in Bonwick94, slabs provide the following advantages over other
49b5fca8f8Stomee  * allocation structures in terms of memory fragmentation:
50b5fca8f8Stomee  *
51b5fca8f8Stomee  *  - Internal fragmentation (per-buffer wasted space) is minimal.
52b5fca8f8Stomee  *  - Severe external fragmentation (unused buffers on the free list) is
53b5fca8f8Stomee  *    unlikely.
54b5fca8f8Stomee  *
55b5fca8f8Stomee  * Segregating objects by size eliminates one source of external fragmentation,
56b5fca8f8Stomee  * and according to Bonwick:
57b5fca8f8Stomee  *
58b5fca8f8Stomee  *   The other reason that slabs reduce external fragmentation is that all
59b5fca8f8Stomee  *   objects in a slab are of the same type, so they have the same lifetime
60b5fca8f8Stomee  *   distribution. The resulting segregation of short-lived and long-lived
61b5fca8f8Stomee  *   objects at slab granularity reduces the likelihood of an entire page being
62b5fca8f8Stomee  *   held hostage due to a single long-lived allocation [Barrett93, Hanson90].
63b5fca8f8Stomee  *
64b5fca8f8Stomee  * While unlikely, severe external fragmentation remains possible. Clients that
65b5fca8f8Stomee  * allocate both short- and long-lived objects from the same cache cannot
66b5fca8f8Stomee  * anticipate the distribution of long-lived objects within the allocator's slab
67b5fca8f8Stomee  * implementation. Even a small percentage of long-lived objects distributed
68b5fca8f8Stomee  * randomly across many slabs can lead to a worst case scenario where the client
69b5fca8f8Stomee  * frees the majority of its objects and the system gets back almost none of the
70b5fca8f8Stomee  * slabs. Despite the client doing what it reasonably can to help the system
71b5fca8f8Stomee  * reclaim memory, the allocator cannot shake free enough slabs because of
72b5fca8f8Stomee  * lonely allocations stubbornly hanging on. Although the allocator is in a
73b5fca8f8Stomee  * position to diagnose the fragmentation, there is nothing that the allocator
74b5fca8f8Stomee  * by itself can do about it. It only takes a single allocated object to prevent
75b5fca8f8Stomee  * an entire slab from being reclaimed, and any object handed out by
76b5fca8f8Stomee  * kmem_cache_alloc() is by definition in the client's control. Conversely,
77b5fca8f8Stomee  * although the client is in a position to move a long-lived object, it has no
78b5fca8f8Stomee  * way of knowing if the object is causing fragmentation, and if so, where to
79b5fca8f8Stomee  * move it. A solution necessarily requires further cooperation between the
80b5fca8f8Stomee  * allocator and the client.
81b5fca8f8Stomee  *
82b5fca8f8Stomee  * 2. Move Callback
83b5fca8f8Stomee  *
84b5fca8f8Stomee  * The kmem slab consolidator therefore adds a move callback to the
85b5fca8f8Stomee  * allocator/client interface, improving worst-case external fragmentation in
86b5fca8f8Stomee  * kmem caches that supply a function to move objects from one memory location
87b5fca8f8Stomee  * to another. In a situation of low memory kmem attempts to consolidate all of
88b5fca8f8Stomee  * a cache's slabs at once; otherwise it works slowly to bring external
89b5fca8f8Stomee  * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
90b5fca8f8Stomee  * thereby helping to avoid a low memory situation in the future.
91b5fca8f8Stomee  *
92b5fca8f8Stomee  * The callback has the following signature:
93b5fca8f8Stomee  *
94b5fca8f8Stomee  *   kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
95b5fca8f8Stomee  *
96b5fca8f8Stomee  * It supplies the kmem client with two addresses: the allocated object that
97b5fca8f8Stomee  * kmem wants to move and a buffer selected by kmem for the client to use as the
98b5fca8f8Stomee  * copy destination. The callback is kmem's way of saying "Please get off of
99b5fca8f8Stomee  * this buffer and use this one instead." kmem knows where it wants to move the
100b5fca8f8Stomee  * object in order to best reduce fragmentation. All the client needs to know
101b5fca8f8Stomee  * about the second argument (void *new) is that it is an allocated, constructed
102b5fca8f8Stomee  * object ready to take the contents of the old object. When the move function
103b5fca8f8Stomee  * is called, the system is likely to be low on memory, and the new object
104b5fca8f8Stomee  * spares the client from having to worry about allocating memory for the
105b5fca8f8Stomee  * requested move. The third argument supplies the size of the object, in case a
106b5fca8f8Stomee  * single move function handles multiple caches whose objects differ only in
107b5fca8f8Stomee  * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
108b5fca8f8Stomee  * user argument passed to the constructor, destructor, and reclaim functions is
109b5fca8f8Stomee  * also passed to the move callback.
110b5fca8f8Stomee  *
111b5fca8f8Stomee  * 2.1 Setting the Move Callback
112b5fca8f8Stomee  *
113b5fca8f8Stomee  * The client sets the move callback after creating the cache and before
114b5fca8f8Stomee  * allocating from it:
115b5fca8f8Stomee  *
116b5fca8f8Stomee  *	object_cache = kmem_cache_create(...);
117b5fca8f8Stomee  *      kmem_cache_set_move(object_cache, object_move);
118b5fca8f8Stomee  *
119b5fca8f8Stomee  * 2.2 Move Callback Return Values
120b5fca8f8Stomee  *
121b5fca8f8Stomee  * Only the client knows about its own data and when is a good time to move it.
122b5fca8f8Stomee  * The client is cooperating with kmem to return unused memory to the system,
123b5fca8f8Stomee  * and kmem respectfully accepts this help at the client's convenience. When
124b5fca8f8Stomee  * asked to move an object, the client can respond with any of the following:
125b5fca8f8Stomee  *
126b5fca8f8Stomee  *   typedef enum kmem_cbrc {
127b5fca8f8Stomee  *           KMEM_CBRC_YES,
128b5fca8f8Stomee  *           KMEM_CBRC_NO,
129b5fca8f8Stomee  *           KMEM_CBRC_LATER,
130b5fca8f8Stomee  *           KMEM_CBRC_DONT_NEED,
131b5fca8f8Stomee  *           KMEM_CBRC_DONT_KNOW
132b5fca8f8Stomee  *   } kmem_cbrc_t;
133b5fca8f8Stomee  *
134b5fca8f8Stomee  * The client must not explicitly kmem_cache_free() either of the objects passed
135b5fca8f8Stomee  * to the callback, since kmem wants to free them directly to the slab layer
136b5fca8f8Stomee  * (bypassing the per-CPU magazine layer). The response tells kmem which of the
137b5fca8f8Stomee  * objects to free:
138b5fca8f8Stomee  *
139b5fca8f8Stomee  *       YES: (Did it) The client moved the object, so kmem frees the old one.
140b5fca8f8Stomee  *        NO: (Never) The client refused, so kmem frees the new object (the
141b5fca8f8Stomee  *            unused copy destination). kmem also marks the slab of the old
142b5fca8f8Stomee  *            object so as not to bother the client with further callbacks for
143b5fca8f8Stomee  *            that object as long as the slab remains on the partial slab list.
144b5fca8f8Stomee  *            (The system won't be getting the slab back as long as the
145b5fca8f8Stomee  *            immovable object holds it hostage, so there's no point in moving
146b5fca8f8Stomee  *            any of its objects.)
147b5fca8f8Stomee  *     LATER: The client is using the object and cannot move it now, so kmem
148b5fca8f8Stomee  *            frees the new object (the unused copy destination). kmem still
149b5fca8f8Stomee  *            attempts to move other objects off the slab, since it expects to
150b5fca8f8Stomee  *            succeed in clearing the slab in a later callback. The client
151b5fca8f8Stomee  *            should use LATER instead of NO if the object is likely to become
152b5fca8f8Stomee  *            movable very soon.
153b5fca8f8Stomee  * DONT_NEED: The client no longer needs the object, so kmem frees the old along
154b5fca8f8Stomee  *            with the new object (the unused copy destination). This response
155b5fca8f8Stomee  *            is the client's opportunity to be a model citizen and give back as
156b5fca8f8Stomee  *            much as it can.
157b5fca8f8Stomee  * DONT_KNOW: The client does not know about the object because
158b5fca8f8Stomee  *            a) the client has just allocated the object and not yet put it
159b5fca8f8Stomee  *               wherever it expects to find known objects
160b5fca8f8Stomee  *            b) the client has removed the object from wherever it expects to
161b5fca8f8Stomee  *               find known objects and is about to free it, or
162b5fca8f8Stomee  *            c) the client has freed the object.
163b5fca8f8Stomee  *            In all these cases (a, b, and c) kmem frees the new object (the
164d7db73d1SBryan Cantrill  *            unused copy destination).  In the first case, the object is in
165d7db73d1SBryan Cantrill  *            use and the correct action is that for LATER; in the latter two
166d7db73d1SBryan Cantrill  *            cases, we know that the object is either freed or about to be
167d7db73d1SBryan Cantrill  *            freed, in which case it is either already in a magazine or about
168d7db73d1SBryan Cantrill  *            to be in one.  In these cases, we know that the object will either
169d7db73d1SBryan Cantrill  *            be reallocated and reused, or it will end up in a full magazine
170d7db73d1SBryan Cantrill  *            that will be reaped (thereby liberating the slab).  Because it
171d7db73d1SBryan Cantrill  *            is prohibitively expensive to differentiate these cases, and
172d7db73d1SBryan Cantrill  *            because the defrag code is executed when we're low on memory
173d7db73d1SBryan Cantrill  *            (thereby biasing the system to reclaim full magazines) we treat
174d7db73d1SBryan Cantrill  *            all DONT_KNOW cases as LATER and rely on cache reaping to
175d7db73d1SBryan Cantrill  *            generally clean up full magazines.  While we take the same action
176d7db73d1SBryan Cantrill  *            for these cases, we maintain their semantic distinction:  if
177d7db73d1SBryan Cantrill  *            defragmentation is not occurring, it is useful to know if this
178d7db73d1SBryan Cantrill  *            is due to objects in use (LATER) or objects in an unknown state
179d7db73d1SBryan Cantrill  *            of transition (DONT_KNOW).
180b5fca8f8Stomee  *
181b5fca8f8Stomee  * 2.3 Object States
182b5fca8f8Stomee  *
183b5fca8f8Stomee  * Neither kmem nor the client can be assumed to know the object's whereabouts
184b5fca8f8Stomee  * at the time of the callback. An object belonging to a kmem cache may be in
185b5fca8f8Stomee  * any of the following states:
186b5fca8f8Stomee  *
187b5fca8f8Stomee  * 1. Uninitialized on the slab
188b5fca8f8Stomee  * 2. Allocated from the slab but not constructed (still uninitialized)
189b5fca8f8Stomee  * 3. Allocated from the slab, constructed, but not yet ready for business
190b5fca8f8Stomee  *    (not in a valid state for the move callback)
191b5fca8f8Stomee  * 4. In use (valid and known to the client)
192b5fca8f8Stomee  * 5. About to be freed (no longer in a valid state for the move callback)
193b5fca8f8Stomee  * 6. Freed to a magazine (still constructed)
194b5fca8f8Stomee  * 7. Allocated from a magazine, not yet ready for business (not in a valid
195b5fca8f8Stomee  *    state for the move callback), and about to return to state #4
196b5fca8f8Stomee  * 8. Deconstructed on a magazine that is about to be freed
197b5fca8f8Stomee  * 9. Freed to the slab
198b5fca8f8Stomee  *
199b5fca8f8Stomee  * Since the move callback may be called at any time while the object is in any
200b5fca8f8Stomee  * of the above states (except state #1), the client needs a safe way to
201b5fca8f8Stomee  * determine whether or not it knows about the object. Specifically, the client
202b5fca8f8Stomee  * needs to know whether or not the object is in state #4, the only state in
203b5fca8f8Stomee  * which a move is valid. If the object is in any other state, the client should
204b5fca8f8Stomee  * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
205b5fca8f8Stomee  * the object's fields.
206b5fca8f8Stomee  *
207b5fca8f8Stomee  * Note that although an object may be in state #4 when kmem initiates the move
208b5fca8f8Stomee  * request, the object may no longer be in that state by the time kmem actually
209b5fca8f8Stomee  * calls the move function. Not only does the client free objects
210b5fca8f8Stomee  * asynchronously, kmem itself puts move requests on a queue where thay are
211b5fca8f8Stomee  * pending until kmem processes them from another context. Also, objects freed
212b5fca8f8Stomee  * to a magazine appear allocated from the point of view of the slab layer, so
213b5fca8f8Stomee  * kmem may even initiate requests for objects in a state other than state #4.
214b5fca8f8Stomee  *
215b5fca8f8Stomee  * 2.3.1 Magazine Layer
216b5fca8f8Stomee  *
217b5fca8f8Stomee  * An important insight revealed by the states listed above is that the magazine
218b5fca8f8Stomee  * layer is populated only by kmem_cache_free(). Magazines of constructed
219b5fca8f8Stomee  * objects are never populated directly from the slab layer (which contains raw,
220b5fca8f8Stomee  * unconstructed objects). Whenever an allocation request cannot be satisfied
221b5fca8f8Stomee  * from the magazine layer, the magazines are bypassed and the request is
222b5fca8f8Stomee  * satisfied from the slab layer (creating a new slab if necessary). kmem calls
223b5fca8f8Stomee  * the object constructor only when allocating from the slab layer, and only in
224b5fca8f8Stomee  * response to kmem_cache_alloc() or to prepare the destination buffer passed in
225b5fca8f8Stomee  * the move callback. kmem does not preconstruct objects in anticipation of
226b5fca8f8Stomee  * kmem_cache_alloc().
227b5fca8f8Stomee  *
228b5fca8f8Stomee  * 2.3.2 Object Constructor and Destructor
229b5fca8f8Stomee  *
230b5fca8f8Stomee  * If the client supplies a destructor, it must be valid to call the destructor
231b5fca8f8Stomee  * on a newly created object (immediately after the constructor).
232b5fca8f8Stomee  *
233b5fca8f8Stomee  * 2.4 Recognizing Known Objects
234b5fca8f8Stomee  *
235b5fca8f8Stomee  * There is a simple test to determine safely whether or not the client knows
236b5fca8f8Stomee  * about a given object in the move callback. It relies on the fact that kmem
237b5fca8f8Stomee  * guarantees that the object of the move callback has only been touched by the
238b5fca8f8Stomee  * client itself or else by kmem. kmem does this by ensuring that none of the
239b5fca8f8Stomee  * cache's slabs are freed to the virtual memory (VM) subsystem while a move
240b5fca8f8Stomee  * callback is pending. When the last object on a slab is freed, if there is a
241b5fca8f8Stomee  * pending move, kmem puts the slab on a per-cache dead list and defers freeing
242b5fca8f8Stomee  * slabs on that list until all pending callbacks are completed. That way,
243b5fca8f8Stomee  * clients can be certain that the object of a move callback is in one of the
244b5fca8f8Stomee  * states listed above, making it possible to distinguish known objects (in
245b5fca8f8Stomee  * state #4) using the two low order bits of any pointer member (with the
246b5fca8f8Stomee  * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
247b5fca8f8Stomee  * platforms).
248b5fca8f8Stomee  *
249b5fca8f8Stomee  * The test works as long as the client always transitions objects from state #4
250b5fca8f8Stomee  * (known, in use) to state #5 (about to be freed, invalid) by setting the low
251b5fca8f8Stomee  * order bit of the client-designated pointer member. Since kmem only writes
252b5fca8f8Stomee  * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
253b5fca8f8Stomee  * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
254b5fca8f8Stomee  * guaranteed to set at least one of the two low order bits. Therefore, given an
255b5fca8f8Stomee  * object with a back pointer to a 'container_t *o_container', the client can
256b5fca8f8Stomee  * test
257b5fca8f8Stomee  *
258b5fca8f8Stomee  *      container_t *container = object->o_container;
259b5fca8f8Stomee  *      if ((uintptr_t)container & 0x3) {
260b5fca8f8Stomee  *              return (KMEM_CBRC_DONT_KNOW);
261b5fca8f8Stomee  *      }
262b5fca8f8Stomee  *
263b5fca8f8Stomee  * Typically, an object will have a pointer to some structure with a list or
264b5fca8f8Stomee  * hash where objects from the cache are kept while in use. Assuming that the
265b5fca8f8Stomee  * client has some way of knowing that the container structure is valid and will
266b5fca8f8Stomee  * not go away during the move, and assuming that the structure includes a lock
267b5fca8f8Stomee  * to protect whatever collection is used, then the client would continue as
268b5fca8f8Stomee  * follows:
269b5fca8f8Stomee  *
270b5fca8f8Stomee  *	// Ensure that the container structure does not go away.
271b5fca8f8Stomee  *      if (container_hold(container) == 0) {
272b5fca8f8Stomee  *              return (KMEM_CBRC_DONT_KNOW);
273b5fca8f8Stomee  *      }
274b5fca8f8Stomee  *      mutex_enter(&container->c_objects_lock);
275b5fca8f8Stomee  *      if (container != object->o_container) {
276b5fca8f8Stomee  *              mutex_exit(&container->c_objects_lock);
277b5fca8f8Stomee  *              container_rele(container);
278b5fca8f8Stomee  *              return (KMEM_CBRC_DONT_KNOW);
279b5fca8f8Stomee  *      }
280b5fca8f8Stomee  *
281b5fca8f8Stomee  * At this point the client knows that the object cannot be freed as long as
282b5fca8f8Stomee  * c_objects_lock is held. Note that after acquiring the lock, the client must
283b5fca8f8Stomee  * recheck the o_container pointer in case the object was removed just before
284b5fca8f8Stomee  * acquiring the lock.
285b5fca8f8Stomee  *
286b5fca8f8Stomee  * When the client is about to free an object, it must first remove that object
287b5fca8f8Stomee  * from the list, hash, or other structure where it is kept. At that time, to
288b5fca8f8Stomee  * mark the object so it can be distinguished from the remaining, known objects,
289b5fca8f8Stomee  * the client sets the designated low order bit:
290b5fca8f8Stomee  *
291b5fca8f8Stomee  *      mutex_enter(&container->c_objects_lock);
292b5fca8f8Stomee  *      object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
293b5fca8f8Stomee  *      list_remove(&container->c_objects, object);
294b5fca8f8Stomee  *      mutex_exit(&container->c_objects_lock);
295b5fca8f8Stomee  *
296b5fca8f8Stomee  * In the common case, the object is freed to the magazine layer, where it may
297b5fca8f8Stomee  * be reused on a subsequent allocation without the overhead of calling the
298b5fca8f8Stomee  * constructor. While in the magazine it appears allocated from the point of
299b5fca8f8Stomee  * view of the slab layer, making it a candidate for the move callback. Most
300b5fca8f8Stomee  * objects unrecognized by the client in the move callback fall into this
301b5fca8f8Stomee  * category and are cheaply distinguished from known objects by the test
302d7db73d1SBryan Cantrill  * described earlier. Because searching magazines is prohibitively expensive
303d7db73d1SBryan Cantrill  * for kmem, clients that do not mark freed objects (and therefore return
304d7db73d1SBryan Cantrill  * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
305d7db73d1SBryan Cantrill  * efficacy reduced.
306b5fca8f8Stomee  *
307b5fca8f8Stomee  * Invalidating the designated pointer member before freeing the object marks
308b5fca8f8Stomee  * the object to be avoided in the callback, and conversely, assigning a valid
309b5fca8f8Stomee  * value to the designated pointer member after allocating the object makes the
310b5fca8f8Stomee  * object fair game for the callback:
311b5fca8f8Stomee  *
312b5fca8f8Stomee  *      ... allocate object ...
313b5fca8f8Stomee  *      ... set any initial state not set by the constructor ...
314b5fca8f8Stomee  *
315b5fca8f8Stomee  *      mutex_enter(&container->c_objects_lock);
316b5fca8f8Stomee  *      list_insert_tail(&container->c_objects, object);
317b5fca8f8Stomee  *      membar_producer();
318b5fca8f8Stomee  *      object->o_container = container;
319b5fca8f8Stomee  *      mutex_exit(&container->c_objects_lock);
320b5fca8f8Stomee  *
321b5fca8f8Stomee  * Note that everything else must be valid before setting o_container makes the
322b5fca8f8Stomee  * object fair game for the move callback. The membar_producer() call ensures
323b5fca8f8Stomee  * that all the object's state is written to memory before setting the pointer
324b5fca8f8Stomee  * that transitions the object from state #3 or #7 (allocated, constructed, not
325b5fca8f8Stomee  * yet in use) to state #4 (in use, valid). That's important because the move
326b5fca8f8Stomee  * function has to check the validity of the pointer before it can safely
327b5fca8f8Stomee  * acquire the lock protecting the collection where it expects to find known
328b5fca8f8Stomee  * objects.
329b5fca8f8Stomee  *
330b5fca8f8Stomee  * This method of distinguishing known objects observes the usual symmetry:
331b5fca8f8Stomee  * invalidating the designated pointer is the first thing the client does before
332b5fca8f8Stomee  * freeing the object, and setting the designated pointer is the last thing the
333b5fca8f8Stomee  * client does after allocating the object. Of course, the client is not
334b5fca8f8Stomee  * required to use this method. Fundamentally, how the client recognizes known
335b5fca8f8Stomee  * objects is completely up to the client, but this method is recommended as an
336b5fca8f8Stomee  * efficient and safe way to take advantage of the guarantees made by kmem. If
337b5fca8f8Stomee  * the entire object is arbitrary data without any markable bits from a suitable
338b5fca8f8Stomee  * pointer member, then the client must find some other method, such as
339b5fca8f8Stomee  * searching a hash table of known objects.
340b5fca8f8Stomee  *
341b5fca8f8Stomee  * 2.5 Preventing Objects From Moving
342b5fca8f8Stomee  *
343b5fca8f8Stomee  * Besides a way to distinguish known objects, the other thing that the client
344b5fca8f8Stomee  * needs is a strategy to ensure that an object will not move while the client
345b5fca8f8Stomee  * is actively using it. The details of satisfying this requirement tend to be
346b5fca8f8Stomee  * highly cache-specific. It might seem that the same rules that let a client
347b5fca8f8Stomee  * remove an object safely should also decide when an object can be moved
348b5fca8f8Stomee  * safely. However, any object state that makes a removal attempt invalid is
349b5fca8f8Stomee  * likely to be long-lasting for objects that the client does not expect to
350b5fca8f8Stomee  * remove. kmem knows nothing about the object state and is equally likely (from
351b5fca8f8Stomee  * the client's point of view) to request a move for any object in the cache,
352b5fca8f8Stomee  * whether prepared for removal or not. Even a low percentage of objects stuck
353b5fca8f8Stomee  * in place by unremovability will defeat the consolidator if the stuck objects
354b5fca8f8Stomee  * are the same long-lived allocations likely to hold slabs hostage.
355b5fca8f8Stomee  * Fundamentally, the consolidator is not aimed at common cases. Severe external
356b5fca8f8Stomee  * fragmentation is a worst case scenario manifested as sparsely allocated
357b5fca8f8Stomee  * slabs, by definition a low percentage of the cache's objects. When deciding
358b5fca8f8Stomee  * what makes an object movable, keep in mind the goal of the consolidator: to
359b5fca8f8Stomee  * bring worst-case external fragmentation within the limits guaranteed for
360b5fca8f8Stomee  * internal fragmentation. Removability is a poor criterion if it is likely to
361b5fca8f8Stomee  * exclude more than an insignificant percentage of objects for long periods of
362b5fca8f8Stomee  * time.
363b5fca8f8Stomee  *
364b5fca8f8Stomee  * A tricky general solution exists, and it has the advantage of letting you
365b5fca8f8Stomee  * move any object at almost any moment, practically eliminating the likelihood
366b5fca8f8Stomee  * that an object can hold a slab hostage. However, if there is a cache-specific
367b5fca8f8Stomee  * way to ensure that an object is not actively in use in the vast majority of
368b5fca8f8Stomee  * cases, a simpler solution that leverages this cache-specific knowledge is
369b5fca8f8Stomee  * preferred.
370b5fca8f8Stomee  *
371b5fca8f8Stomee  * 2.5.1 Cache-Specific Solution
372b5fca8f8Stomee  *
373b5fca8f8Stomee  * As an example of a cache-specific solution, the ZFS znode cache takes
374b5fca8f8Stomee  * advantage of the fact that the vast majority of znodes are only being
375b5fca8f8Stomee  * referenced from the DNLC. (A typical case might be a few hundred in active
376b5fca8f8Stomee  * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
377b5fca8f8Stomee  * client has established that it recognizes the znode and can access its fields
378b5fca8f8Stomee  * safely (using the method described earlier), it then tests whether the znode
379b5fca8f8Stomee  * is referenced by anything other than the DNLC. If so, it assumes that the
380b5fca8f8Stomee  * znode may be in active use and is unsafe to move, so it drops its locks and
381b5fca8f8Stomee  * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
382b5fca8f8Stomee  * else znodes are used, no change is needed to protect against the possibility
383b5fca8f8Stomee  * of the znode moving. The disadvantage is that it remains possible for an
384b5fca8f8Stomee  * application to hold a znode slab hostage with an open file descriptor.
385b5fca8f8Stomee  * However, this case ought to be rare and the consolidator has a way to deal
386b5fca8f8Stomee  * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
387b5fca8f8Stomee  * object, kmem eventually stops believing it and treats the slab as if the
388b5fca8f8Stomee  * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
389b5fca8f8Stomee  * then focus on getting it off of the partial slab list by allocating rather
390b5fca8f8Stomee  * than freeing all of its objects. (Either way of getting a slab off the
391b5fca8f8Stomee  * free list reduces fragmentation.)
392b5fca8f8Stomee  *
393b5fca8f8Stomee  * 2.5.2 General Solution
394b5fca8f8Stomee  *
395b5fca8f8Stomee  * The general solution, on the other hand, requires an explicit hold everywhere
396b5fca8f8Stomee  * the object is used to prevent it from moving. To keep the client locking
397b5fca8f8Stomee  * strategy as uncomplicated as possible, kmem guarantees the simplifying
398b5fca8f8Stomee  * assumption that move callbacks are sequential, even across multiple caches.
399b5fca8f8Stomee  * Internally, a global queue processed by a single thread supports all caches
400b5fca8f8Stomee  * implementing the callback function. No matter how many caches supply a move
401b5fca8f8Stomee  * function, the consolidator never moves more than one object at a time, so the
402b5fca8f8Stomee  * client does not have to worry about tricky lock ordering involving several
403b5fca8f8Stomee  * related objects from different kmem caches.
404b5fca8f8Stomee  *
405b5fca8f8Stomee  * The general solution implements the explicit hold as a read-write lock, which
406b5fca8f8Stomee  * allows multiple readers to access an object from the cache simultaneously
407b5fca8f8Stomee  * while a single writer is excluded from moving it. A single rwlock for the
408b5fca8f8Stomee  * entire cache would lock out all threads from using any of the cache's objects
409b5fca8f8Stomee  * even though only a single object is being moved, so to reduce contention,
410b5fca8f8Stomee  * the client can fan out the single rwlock into an array of rwlocks hashed by
411b5fca8f8Stomee  * the object address, making it probable that moving one object will not
412b5fca8f8Stomee  * prevent other threads from using a different object. The rwlock cannot be a
413b5fca8f8Stomee  * member of the object itself, because the possibility of the object moving
414b5fca8f8Stomee  * makes it unsafe to access any of the object's fields until the lock is
415b5fca8f8Stomee  * acquired.
416b5fca8f8Stomee  *
417b5fca8f8Stomee  * Assuming a small, fixed number of locks, it's possible that multiple objects
418b5fca8f8Stomee  * will hash to the same lock. A thread that needs to use multiple objects in
419b5fca8f8Stomee  * the same function may acquire the same lock multiple times. Since rwlocks are
420b5fca8f8Stomee  * reentrant for readers, and since there is never more than a single writer at
421b5fca8f8Stomee  * a time (assuming that the client acquires the lock as a writer only when
422b5fca8f8Stomee  * moving an object inside the callback), there would seem to be no problem.
423b5fca8f8Stomee  * However, a client locking multiple objects in the same function must handle
424b5fca8f8Stomee  * one case of potential deadlock: Assume that thread A needs to prevent both
425b5fca8f8Stomee  * object 1 and object 2 from moving, and thread B, the callback, meanwhile
426b5fca8f8Stomee  * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
427b5fca8f8Stomee  * same lock, that thread A will acquire the lock for object 1 as a reader
428b5fca8f8Stomee  * before thread B sets the lock's write-wanted bit, preventing thread A from
429b5fca8f8Stomee  * reacquiring the lock for object 2 as a reader. Unable to make forward
430b5fca8f8Stomee  * progress, thread A will never release the lock for object 1, resulting in
431b5fca8f8Stomee  * deadlock.
432b5fca8f8Stomee  *
433b5fca8f8Stomee  * There are two ways of avoiding the deadlock just described. The first is to
434b5fca8f8Stomee  * use rw_tryenter() rather than rw_enter() in the callback function when
435b5fca8f8Stomee  * attempting to acquire the lock as a writer. If tryenter discovers that the
436b5fca8f8Stomee  * same object (or another object hashed to the same lock) is already in use, it
437b5fca8f8Stomee  * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
438b5fca8f8Stomee  * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
439b5fca8f8Stomee  * since it allows a thread to acquire the lock as a reader in spite of a
440b5fca8f8Stomee  * waiting writer. This second approach insists on moving the object now, no
441b5fca8f8Stomee  * matter how many readers the move function must wait for in order to do so,
442b5fca8f8Stomee  * and could delay the completion of the callback indefinitely (blocking
443b5fca8f8Stomee  * callbacks to other clients). In practice, a less insistent callback using
444b5fca8f8Stomee  * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
445b5fca8f8Stomee  * little reason to use anything else.
446b5fca8f8Stomee  *
447b5fca8f8Stomee  * Avoiding deadlock is not the only problem that an implementation using an
448b5fca8f8Stomee  * explicit hold needs to solve. Locking the object in the first place (to
449b5fca8f8Stomee  * prevent it from moving) remains a problem, since the object could move
450b5fca8f8Stomee  * between the time you obtain a pointer to the object and the time you acquire
451b5fca8f8Stomee  * the rwlock hashed to that pointer value. Therefore the client needs to
452b5fca8f8Stomee  * recheck the value of the pointer after acquiring the lock, drop the lock if
453b5fca8f8Stomee  * the value has changed, and try again. This requires a level of indirection:
454b5fca8f8Stomee  * something that points to the object rather than the object itself, that the
455b5fca8f8Stomee  * client can access safely while attempting to acquire the lock. (The object
456b5fca8f8Stomee  * itself cannot be referenced safely because it can move at any time.)
457b5fca8f8Stomee  * The following lock-acquisition function takes whatever is safe to reference
458b5fca8f8Stomee  * (arg), follows its pointer to the object (using function f), and tries as
459b5fca8f8Stomee  * often as necessary to acquire the hashed lock and verify that the object
460b5fca8f8Stomee  * still has not moved:
461b5fca8f8Stomee  *
462b5fca8f8Stomee  *      object_t *
463b5fca8f8Stomee  *      object_hold(object_f f, void *arg)
464b5fca8f8Stomee  *      {
465b5fca8f8Stomee  *              object_t *op;
466b5fca8f8Stomee  *
467b5fca8f8Stomee  *              op = f(arg);
468b5fca8f8Stomee  *              if (op == NULL) {
469b5fca8f8Stomee  *                      return (NULL);
470b5fca8f8Stomee  *              }
471b5fca8f8Stomee  *
472b5fca8f8Stomee  *              rw_enter(OBJECT_RWLOCK(op), RW_READER);
473b5fca8f8Stomee  *              while (op != f(arg)) {
474b5fca8f8Stomee  *                      rw_exit(OBJECT_RWLOCK(op));
475b5fca8f8Stomee  *                      op = f(arg);
476b5fca8f8Stomee  *                      if (op == NULL) {
477b5fca8f8Stomee  *                              break;
478b5fca8f8Stomee  *                      }
479b5fca8f8Stomee  *                      rw_enter(OBJECT_RWLOCK(op), RW_READER);
480b5fca8f8Stomee  *              }
481b5fca8f8Stomee  *
482b5fca8f8Stomee  *              return (op);
483b5fca8f8Stomee  *      }
484b5fca8f8Stomee  *
485b5fca8f8Stomee  * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
486b5fca8f8Stomee  * lock reacquisition loop, while necessary, almost never executes. The function
487b5fca8f8Stomee  * pointer f (used to obtain the object pointer from arg) has the following type
488b5fca8f8Stomee  * definition:
489b5fca8f8Stomee  *
490b5fca8f8Stomee  *      typedef object_t *(*object_f)(void *arg);
491b5fca8f8Stomee  *
492b5fca8f8Stomee  * An object_f implementation is likely to be as simple as accessing a structure
493b5fca8f8Stomee  * member:
494b5fca8f8Stomee  *
495b5fca8f8Stomee  *      object_t *
496b5fca8f8Stomee  *      s_object(void *arg)
497b5fca8f8Stomee  *      {
498b5fca8f8Stomee  *              something_t *sp = arg;
499b5fca8f8Stomee  *              return (sp->s_object);
500b5fca8f8Stomee  *      }
501b5fca8f8Stomee  *
502b5fca8f8Stomee  * The flexibility of a function pointer allows the path to the object to be
503b5fca8f8Stomee  * arbitrarily complex and also supports the notion that depending on where you
504b5fca8f8Stomee  * are using the object, you may need to get it from someplace different.
505b5fca8f8Stomee  *
506b5fca8f8Stomee  * The function that releases the explicit hold is simpler because it does not
507b5fca8f8Stomee  * have to worry about the object moving:
508b5fca8f8Stomee  *
509b5fca8f8Stomee  *      void
510b5fca8f8Stomee  *      object_rele(object_t *op)
511b5fca8f8Stomee  *      {
512b5fca8f8Stomee  *              rw_exit(OBJECT_RWLOCK(op));
513b5fca8f8Stomee  *      }
514b5fca8f8Stomee  *
515b5fca8f8Stomee  * The caller is spared these details so that obtaining and releasing an
516b5fca8f8Stomee  * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
517b5fca8f8Stomee  * of object_hold() only needs to know that the returned object pointer is valid
518b5fca8f8Stomee  * if not NULL and that the object will not move until released.
519b5fca8f8Stomee  *
520b5fca8f8Stomee  * Although object_hold() prevents an object from moving, it does not prevent it
521b5fca8f8Stomee  * from being freed. The caller must take measures before calling object_hold()
522b5fca8f8Stomee  * (afterwards is too late) to ensure that the held object cannot be freed. The
523b5fca8f8Stomee  * caller must do so without accessing the unsafe object reference, so any lock
524b5fca8f8Stomee  * or reference count used to ensure the continued existence of the object must
525b5fca8f8Stomee  * live outside the object itself.
526b5fca8f8Stomee  *
527b5fca8f8Stomee  * Obtaining a new object is a special case where an explicit hold is impossible
528b5fca8f8Stomee  * for the caller. Any function that returns a newly allocated object (either as
529b5fca8f8Stomee  * a return value, or as an in-out paramter) must return it already held; after
530b5fca8f8Stomee  * the caller gets it is too late, since the object cannot be safely accessed
531b5fca8f8Stomee  * without the level of indirection described earlier. The following
532b5fca8f8Stomee  * object_alloc() example uses the same code shown earlier to transition a new
533b5fca8f8Stomee  * object into the state of being recognized (by the client) as a known object.
534b5fca8f8Stomee  * The function must acquire the hold (rw_enter) before that state transition
535b5fca8f8Stomee  * makes the object movable:
536b5fca8f8Stomee  *
537b5fca8f8Stomee  *      static object_t *
538b5fca8f8Stomee  *      object_alloc(container_t *container)
539b5fca8f8Stomee  *      {
5404d4c4c43STom Erickson  *              object_t *object = kmem_cache_alloc(object_cache, 0);
541b5fca8f8Stomee  *              ... set any initial state not set by the constructor ...
542b5fca8f8Stomee  *              rw_enter(OBJECT_RWLOCK(object), RW_READER);
543b5fca8f8Stomee  *              mutex_enter(&container->c_objects_lock);
544b5fca8f8Stomee  *              list_insert_tail(&container->c_objects, object);
545b5fca8f8Stomee  *              membar_producer();
546b5fca8f8Stomee  *              object->o_container = container;
547b5fca8f8Stomee  *              mutex_exit(&container->c_objects_lock);
548b5fca8f8Stomee  *              return (object);
549b5fca8f8Stomee  *      }
550b5fca8f8Stomee  *
551b5fca8f8Stomee  * Functions that implicitly acquire an object hold (any function that calls
552b5fca8f8Stomee  * object_alloc() to supply an object for the caller) need to be carefully noted
553b5fca8f8Stomee  * so that the matching object_rele() is not neglected. Otherwise, leaked holds
554b5fca8f8Stomee  * prevent all objects hashed to the affected rwlocks from ever being moved.
555b5fca8f8Stomee  *
556b5fca8f8Stomee  * The pointer to a held object can be hashed to the holding rwlock even after
557b5fca8f8Stomee  * the object has been freed. Although it is possible to release the hold
558b5fca8f8Stomee  * after freeing the object, you may decide to release the hold implicitly in
559b5fca8f8Stomee  * whatever function frees the object, so as to release the hold as soon as
560b5fca8f8Stomee  * possible, and for the sake of symmetry with the function that implicitly
561b5fca8f8Stomee  * acquires the hold when it allocates the object. Here, object_free() releases
562b5fca8f8Stomee  * the hold acquired by object_alloc(). Its implicit object_rele() forms a
563b5fca8f8Stomee  * matching pair with object_hold():
564b5fca8f8Stomee  *
565b5fca8f8Stomee  *      void
566b5fca8f8Stomee  *      object_free(object_t *object)
567b5fca8f8Stomee  *      {
568b5fca8f8Stomee  *              container_t *container;
569b5fca8f8Stomee  *
570b5fca8f8Stomee  *              ASSERT(object_held(object));
571b5fca8f8Stomee  *              container = object->o_container;
572b5fca8f8Stomee  *              mutex_enter(&container->c_objects_lock);
573b5fca8f8Stomee  *              object->o_container =
574b5fca8f8Stomee  *                  (void *)((uintptr_t)object->o_container | 0x1);
575b5fca8f8Stomee  *              list_remove(&container->c_objects, object);
576b5fca8f8Stomee  *              mutex_exit(&container->c_objects_lock);
577b5fca8f8Stomee  *              object_rele(object);
578b5fca8f8Stomee  *              kmem_cache_free(object_cache, object);
579b5fca8f8Stomee  *      }
580b5fca8f8Stomee  *
581b5fca8f8Stomee  * Note that object_free() cannot safely accept an object pointer as an argument
582b5fca8f8Stomee  * unless the object is already held. Any function that calls object_free()
583b5fca8f8Stomee  * needs to be carefully noted since it similarly forms a matching pair with
584b5fca8f8Stomee  * object_hold().
585b5fca8f8Stomee  *
586b5fca8f8Stomee  * To complete the picture, the following callback function implements the
587b5fca8f8Stomee  * general solution by moving objects only if they are currently unheld:
588b5fca8f8Stomee  *
589b5fca8f8Stomee  *      static kmem_cbrc_t
590b5fca8f8Stomee  *      object_move(void *buf, void *newbuf, size_t size, void *arg)
591b5fca8f8Stomee  *      {
592b5fca8f8Stomee  *              object_t *op = buf, *np = newbuf;
593b5fca8f8Stomee  *              container_t *container;
594b5fca8f8Stomee  *
595b5fca8f8Stomee  *              container = op->o_container;
596b5fca8f8Stomee  *              if ((uintptr_t)container & 0x3) {
597b5fca8f8Stomee  *                      return (KMEM_CBRC_DONT_KNOW);
598b5fca8f8Stomee  *              }
599b5fca8f8Stomee  *
600b5fca8f8Stomee  *	        // Ensure that the container structure does not go away.
601b5fca8f8Stomee  *              if (container_hold(container) == 0) {
602b5fca8f8Stomee  *                      return (KMEM_CBRC_DONT_KNOW);
603b5fca8f8Stomee  *              }
604b5fca8f8Stomee  *
605b5fca8f8Stomee  *              mutex_enter(&container->c_objects_lock);
606b5fca8f8Stomee  *              if (container != op->o_container) {
607b5fca8f8Stomee  *                      mutex_exit(&container->c_objects_lock);
608b5fca8f8Stomee  *                      container_rele(container);
609b5fca8f8Stomee  *                      return (KMEM_CBRC_DONT_KNOW);
610b5fca8f8Stomee  *              }
611b5fca8f8Stomee  *
612b5fca8f8Stomee  *              if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
613b5fca8f8Stomee  *                      mutex_exit(&container->c_objects_lock);
614b5fca8f8Stomee  *                      container_rele(container);
615b5fca8f8Stomee  *                      return (KMEM_CBRC_LATER);
616b5fca8f8Stomee  *              }
617b5fca8f8Stomee  *
618b5fca8f8Stomee  *              object_move_impl(op, np); // critical section
619b5fca8f8Stomee  *              rw_exit(OBJECT_RWLOCK(op));
620b5fca8f8Stomee  *
621b5fca8f8Stomee  *              op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
622b5fca8f8Stomee  *              list_link_replace(&op->o_link_node, &np->o_link_node);
623b5fca8f8Stomee  *              mutex_exit(&container->c_objects_lock);
624b5fca8f8Stomee  *              container_rele(container);
625b5fca8f8Stomee  *              return (KMEM_CBRC_YES);
626b5fca8f8Stomee  *      }
627b5fca8f8Stomee  *
628b5fca8f8Stomee  * Note that object_move() must invalidate the designated o_container pointer of
629b5fca8f8Stomee  * the old object in the same way that object_free() does, since kmem will free
630b5fca8f8Stomee  * the object in response to the KMEM_CBRC_YES return value.
631b5fca8f8Stomee  *
632b5fca8f8Stomee  * The lock order in object_move() differs from object_alloc(), which locks
633b5fca8f8Stomee  * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
634b5fca8f8Stomee  * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
635b5fca8f8Stomee  * not a problem. Holding the lock on the object list in the example above
636b5fca8f8Stomee  * through the entire callback not only prevents the object from going away, it
637b5fca8f8Stomee  * also allows you to lock the list elsewhere and know that none of its elements
638b5fca8f8Stomee  * will move during iteration.
639b5fca8f8Stomee  *
640b5fca8f8Stomee  * Adding an explicit hold everywhere an object from the cache is used is tricky
641b5fca8f8Stomee  * and involves much more change to client code than a cache-specific solution
642b5fca8f8Stomee  * that leverages existing state to decide whether or not an object is
643b5fca8f8Stomee  * movable. However, this approach has the advantage that no object remains
644b5fca8f8Stomee  * immovable for any significant length of time, making it extremely unlikely
645b5fca8f8Stomee  * that long-lived allocations can continue holding slabs hostage; and it works
646b5fca8f8Stomee  * for any cache.
647b5fca8f8Stomee  *
648b5fca8f8Stomee  * 3. Consolidator Implementation
649b5fca8f8Stomee  *
650b5fca8f8Stomee  * Once the client supplies a move function that a) recognizes known objects and
651b5fca8f8Stomee  * b) avoids moving objects that are actively in use, the remaining work is up
652b5fca8f8Stomee  * to the consolidator to decide which objects to move and when to issue
653b5fca8f8Stomee  * callbacks.
654b5fca8f8Stomee  *
655b5fca8f8Stomee  * The consolidator relies on the fact that a cache's slabs are ordered by
656b5fca8f8Stomee  * usage. Each slab has a fixed number of objects. Depending on the slab's
657b5fca8f8Stomee  * "color" (the offset of the first object from the beginning of the slab;
658b5fca8f8Stomee  * offsets are staggered to mitigate false sharing of cache lines) it is either
659b5fca8f8Stomee  * the maximum number of objects per slab determined at cache creation time or
660b5fca8f8Stomee  * else the number closest to the maximum that fits within the space remaining
661b5fca8f8Stomee  * after the initial offset. A completely allocated slab may contribute some
662b5fca8f8Stomee  * internal fragmentation (per-slab overhead) but no external fragmentation, so
663b5fca8f8Stomee  * it is of no interest to the consolidator. At the other extreme, slabs whose
664b5fca8f8Stomee  * objects have all been freed to the slab are released to the virtual memory
665b5fca8f8Stomee  * (VM) subsystem (objects freed to magazines are still allocated as far as the
666b5fca8f8Stomee  * slab is concerned). External fragmentation exists when there are slabs
667b5fca8f8Stomee  * somewhere between these extremes. A partial slab has at least one but not all
668b5fca8f8Stomee  * of its objects allocated. The more partial slabs, and the fewer allocated
669b5fca8f8Stomee  * objects on each of them, the higher the fragmentation. Hence the
670b5fca8f8Stomee  * consolidator's overall strategy is to reduce the number of partial slabs by
671b5fca8f8Stomee  * moving allocated objects from the least allocated slabs to the most allocated
672b5fca8f8Stomee  * slabs.
673b5fca8f8Stomee  *
674b5fca8f8Stomee  * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
675b5fca8f8Stomee  * slabs are kept separately in an unordered list. Since the majority of slabs
676b5fca8f8Stomee  * tend to be completely allocated (a typical unfragmented cache may have
677b5fca8f8Stomee  * thousands of complete slabs and only a single partial slab), separating
678b5fca8f8Stomee  * complete slabs improves the efficiency of partial slab ordering, since the
679b5fca8f8Stomee  * complete slabs do not affect the depth or balance of the AVL tree. This
680b5fca8f8Stomee  * ordered sequence of partial slabs acts as a "free list" supplying objects for
681b5fca8f8Stomee  * allocation requests.
682b5fca8f8Stomee  *
683b5fca8f8Stomee  * Objects are always allocated from the first partial slab in the free list,
684b5fca8f8Stomee  * where the allocation is most likely to eliminate a partial slab (by
685b5fca8f8Stomee  * completely allocating it). Conversely, when a single object from a completely
686b5fca8f8Stomee  * allocated slab is freed to the slab, that slab is added to the front of the
687b5fca8f8Stomee  * free list. Since most free list activity involves highly allocated slabs
688b5fca8f8Stomee  * coming and going at the front of the list, slabs tend naturally toward the
689b5fca8f8Stomee  * ideal order: highly allocated at the front, sparsely allocated at the back.
690b5fca8f8Stomee  * Slabs with few allocated objects are likely to become completely free if they
691b5fca8f8Stomee  * keep a safe distance away from the front of the free list. Slab misorders
692b5fca8f8Stomee  * interfere with the natural tendency of slabs to become completely free or
693b5fca8f8Stomee  * completely allocated. For example, a slab with a single allocated object
694b5fca8f8Stomee  * needs only a single free to escape the cache; its natural desire is
695b5fca8f8Stomee  * frustrated when it finds itself at the front of the list where a second
696b5fca8f8Stomee  * allocation happens just before the free could have released it. Another slab
697b5fca8f8Stomee  * with all but one object allocated might have supplied the buffer instead, so
698b5fca8f8Stomee  * that both (as opposed to neither) of the slabs would have been taken off the
699b5fca8f8Stomee  * free list.
700b5fca8f8Stomee  *
701b5fca8f8Stomee  * Although slabs tend naturally toward the ideal order, misorders allowed by a
702b5fca8f8Stomee  * simple list implementation defeat the consolidator's strategy of merging
703b5fca8f8Stomee  * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
704b5fca8f8Stomee  * needs another way to fix misorders to optimize its callback strategy. One
705b5fca8f8Stomee  * approach is to periodically scan a limited number of slabs, advancing a
706b5fca8f8Stomee  * marker to hold the current scan position, and to move extreme misorders to
707b5fca8f8Stomee  * the front or back of the free list and to the front or back of the current
708b5fca8f8Stomee  * scan range. By making consecutive scan ranges overlap by one slab, the least
709b5fca8f8Stomee  * allocated slab in the current range can be carried along from the end of one
710b5fca8f8Stomee  * scan to the start of the next.
711b5fca8f8Stomee  *
712b5fca8f8Stomee  * Maintaining partial slabs in an AVL tree relieves kmem of this additional
713b5fca8f8Stomee  * task, however. Since most of the cache's activity is in the magazine layer,
714b5fca8f8Stomee  * and allocations from the slab layer represent only a startup cost, the
715b5fca8f8Stomee  * overhead of maintaining a balanced tree is not a significant concern compared
716b5fca8f8Stomee  * to the opportunity of reducing complexity by eliminating the partial slab
717b5fca8f8Stomee  * scanner just described. The overhead of an AVL tree is minimized by
718b5fca8f8Stomee  * maintaining only partial slabs in the tree and keeping completely allocated
719b5fca8f8Stomee  * slabs separately in a list. To avoid increasing the size of the slab
720b5fca8f8Stomee  * structure the AVL linkage pointers are reused for the slab's list linkage,
721b5fca8f8Stomee  * since the slab will always be either partial or complete, never stored both
722b5fca8f8Stomee  * ways at the same time. To further minimize the overhead of the AVL tree the
723b5fca8f8Stomee  * compare function that orders partial slabs by usage divides the range of
724b5fca8f8Stomee  * allocated object counts into bins such that counts within the same bin are
725b5fca8f8Stomee  * considered equal. Binning partial slabs makes it less likely that allocating
726b5fca8f8Stomee  * or freeing a single object will change the slab's order, requiring a tree
727b5fca8f8Stomee  * reinsertion (an avl_remove() followed by an avl_add(), both potentially
728b5fca8f8Stomee  * requiring some rebalancing of the tree). Allocation counts closest to
729b5fca8f8Stomee  * completely free and completely allocated are left unbinned (finely sorted) to
730b5fca8f8Stomee  * better support the consolidator's strategy of merging slabs at either
731b5fca8f8Stomee  * extreme.
732b5fca8f8Stomee  *
733b5fca8f8Stomee  * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
734b5fca8f8Stomee  *
735b5fca8f8Stomee  * The consolidator piggybacks on the kmem maintenance thread and is called on
736b5fca8f8Stomee  * the same interval as kmem_cache_update(), once per cache every fifteen
737b5fca8f8Stomee  * seconds. kmem maintains a running count of unallocated objects in the slab
738b5fca8f8Stomee  * layer (cache_bufslab). The consolidator checks whether that number exceeds
739b5fca8f8Stomee  * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
740b5fca8f8Stomee  * there is a significant number of slabs in the cache (arbitrarily a minimum
741b5fca8f8Stomee  * 101 total slabs). Unused objects that have fallen out of the magazine layer's
742b5fca8f8Stomee  * working set are included in the assessment, and magazines in the depot are
743b5fca8f8Stomee  * reaped if those objects would lift cache_bufslab above the fragmentation
744b5fca8f8Stomee  * threshold. Once the consolidator decides that a cache is fragmented, it looks
745b5fca8f8Stomee  * for a candidate slab to reclaim, starting at the end of the partial slab free
746b5fca8f8Stomee  * list and scanning backwards. At first the consolidator is choosy: only a slab
747b5fca8f8Stomee  * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
748b5fca8f8Stomee  * single allocated object, regardless of percentage). If there is difficulty
749b5fca8f8Stomee  * finding a candidate slab, kmem raises the allocation threshold incrementally,
750b5fca8f8Stomee  * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
751b5fca8f8Stomee  * external fragmentation (unused objects on the free list) below 12.5% (1/8),
752b5fca8f8Stomee  * even in the worst case of every slab in the cache being almost 7/8 allocated.
753b5fca8f8Stomee  * The threshold can also be lowered incrementally when candidate slabs are easy
754b5fca8f8Stomee  * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
755b5fca8f8Stomee  * is no longer fragmented.
756b5fca8f8Stomee  *
757b5fca8f8Stomee  * 3.2 Generating Callbacks
758b5fca8f8Stomee  *
759b5fca8f8Stomee  * Once an eligible slab is chosen, a callback is generated for every allocated
760b5fca8f8Stomee  * object on the slab, in the hope that the client will move everything off the
761b5fca8f8Stomee  * slab and make it reclaimable. Objects selected as move destinations are
762b5fca8f8Stomee  * chosen from slabs at the front of the free list. Assuming slabs in the ideal
763b5fca8f8Stomee  * order (most allocated at the front, least allocated at the back) and a
764b5fca8f8Stomee  * cooperative client, the consolidator will succeed in removing slabs from both
765b5fca8f8Stomee  * ends of the free list, completely allocating on the one hand and completely
766b5fca8f8Stomee  * freeing on the other. Objects selected as move destinations are allocated in
767b5fca8f8Stomee  * the kmem maintenance thread where move requests are enqueued. A separate
768b5fca8f8Stomee  * callback thread removes pending callbacks from the queue and calls the
769b5fca8f8Stomee  * client. The separate thread ensures that client code (the move function) does
770b5fca8f8Stomee  * not interfere with internal kmem maintenance tasks. A map of pending
771b5fca8f8Stomee  * callbacks keyed by object address (the object to be moved) is checked to
772b5fca8f8Stomee  * ensure that duplicate callbacks are not generated for the same object.
773b5fca8f8Stomee  * Allocating the move destination (the object to move to) prevents subsequent
774b5fca8f8Stomee  * callbacks from selecting the same destination as an earlier pending callback.
775b5fca8f8Stomee  *
776b5fca8f8Stomee  * Move requests can also be generated by kmem_cache_reap() when the system is
777b5fca8f8Stomee  * desperate for memory and by kmem_cache_move_notify(), called by the client to
778b5fca8f8Stomee  * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
779b5fca8f8Stomee  * The map of pending callbacks is protected by the same lock that protects the
780b5fca8f8Stomee  * slab layer.
781b5fca8f8Stomee  *
782b5fca8f8Stomee  * When the system is desperate for memory, kmem does not bother to determine
783b5fca8f8Stomee  * whether or not the cache exceeds the fragmentation threshold, but tries to
784b5fca8f8Stomee  * consolidate as many slabs as possible. Normally, the consolidator chews
785b5fca8f8Stomee  * slowly, one sparsely allocated slab at a time during each maintenance
786b5fca8f8Stomee  * interval that the cache is fragmented. When desperate, the consolidator
787b5fca8f8Stomee  * starts at the last partial slab and enqueues callbacks for every allocated
788b5fca8f8Stomee  * object on every partial slab, working backwards until it reaches the first
789b5fca8f8Stomee  * partial slab. The first partial slab, meanwhile, advances in pace with the
790b5fca8f8Stomee  * consolidator as allocations to supply move destinations for the enqueued
791b5fca8f8Stomee  * callbacks use up the highly allocated slabs at the front of the free list.
792b5fca8f8Stomee  * Ideally, the overgrown free list collapses like an accordion, starting at
793b5fca8f8Stomee  * both ends and ending at the center with a single partial slab.
794b5fca8f8Stomee  *
795b5fca8f8Stomee  * 3.3 Client Responses
796b5fca8f8Stomee  *
797b5fca8f8Stomee  * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
798b5fca8f8Stomee  * marks the slab that supplied the stuck object non-reclaimable and moves it to
799b5fca8f8Stomee  * front of the free list. The slab remains marked as long as it remains on the
800b5fca8f8Stomee  * free list, and it appears more allocated to the partial slab compare function
801b5fca8f8Stomee  * than any unmarked slab, no matter how many of its objects are allocated.
802b5fca8f8Stomee  * Since even one immovable object ties up the entire slab, the goal is to
803b5fca8f8Stomee  * completely allocate any slab that cannot be completely freed. kmem does not
804b5fca8f8Stomee  * bother generating callbacks to move objects from a marked slab unless the
805b5fca8f8Stomee  * system is desperate.
806b5fca8f8Stomee  *
807b5fca8f8Stomee  * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
808b5fca8f8Stomee  * slab. If the client responds LATER too many times, kmem disbelieves and
809b5fca8f8Stomee  * treats the response as a NO. The count is cleared when the slab is taken off
810b5fca8f8Stomee  * the partial slab list or when the client moves one of the slab's objects.
811b5fca8f8Stomee  *
812b5fca8f8Stomee  * 4. Observability
813b5fca8f8Stomee  *
814b5fca8f8Stomee  * A kmem cache's external fragmentation is best observed with 'mdb -k' using
815b5fca8f8Stomee  * the ::kmem_slabs dcmd. For a complete description of the command, enter
816b5fca8f8Stomee  * '::help kmem_slabs' at the mdb prompt.
8177c478bd9Sstevel@tonic-gate  */
8187c478bd9Sstevel@tonic-gate 
8197c478bd9Sstevel@tonic-gate #include <sys/kmem_impl.h>
8207c478bd9Sstevel@tonic-gate #include <sys/vmem_impl.h>
8217c478bd9Sstevel@tonic-gate #include <sys/param.h>
8227c478bd9Sstevel@tonic-gate #include <sys/sysmacros.h>
8237c478bd9Sstevel@tonic-gate #include <sys/vm.h>
8247c478bd9Sstevel@tonic-gate #include <sys/proc.h>
8257c478bd9Sstevel@tonic-gate #include <sys/tuneable.h>
8267c478bd9Sstevel@tonic-gate #include <sys/systm.h>
8277c478bd9Sstevel@tonic-gate #include <sys/cmn_err.h>
8287c478bd9Sstevel@tonic-gate #include <sys/debug.h>
829b5fca8f8Stomee #include <sys/sdt.h>
8307c478bd9Sstevel@tonic-gate #include <sys/mutex.h>
8317c478bd9Sstevel@tonic-gate #include <sys/bitmap.h>
8327c478bd9Sstevel@tonic-gate #include <sys/atomic.h>
8337c478bd9Sstevel@tonic-gate #include <sys/kobj.h>
8347c478bd9Sstevel@tonic-gate #include <sys/disp.h>
8357c478bd9Sstevel@tonic-gate #include <vm/seg_kmem.h>
8367c478bd9Sstevel@tonic-gate #include <sys/log.h>
8377c478bd9Sstevel@tonic-gate #include <sys/callb.h>
8387c478bd9Sstevel@tonic-gate #include <sys/taskq.h>
8397c478bd9Sstevel@tonic-gate #include <sys/modctl.h>
8407c478bd9Sstevel@tonic-gate #include <sys/reboot.h>
8417c478bd9Sstevel@tonic-gate #include <sys/id32.h>
8427c478bd9Sstevel@tonic-gate #include <sys/zone.h>
843f4b3ec61Sdh #include <sys/netstack.h>
844b5fca8f8Stomee #ifdef	DEBUG
845b5fca8f8Stomee #include <sys/random.h>
846b5fca8f8Stomee #endif
8477c478bd9Sstevel@tonic-gate 
8487c478bd9Sstevel@tonic-gate extern void streams_msg_init(void);
8497c478bd9Sstevel@tonic-gate extern int segkp_fromheap;
8507c478bd9Sstevel@tonic-gate extern void segkp_cache_free(void);
8516e00b116SPeter Telford extern int callout_init_done;
8527c478bd9Sstevel@tonic-gate 
8537c478bd9Sstevel@tonic-gate struct kmem_cache_kstat {
8547c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_size;
8557c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_align;
8567c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_chunk_size;
8577c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_slab_size;
8587c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_alloc;
8597c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_alloc_fail;
8607c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_free;
8617c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_depot_alloc;
8627c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_depot_free;
8637c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_depot_contention;
8647c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_slab_alloc;
8657c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_slab_free;
8667c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_constructed;
8677c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_avail;
8687c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_inuse;
8697c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_total;
8707c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_buf_max;
8717c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_slab_create;
8727c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_slab_destroy;
8737c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_vmem_source;
8747c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_hash_size;
8757c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_hash_lookup_depth;
8767c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_hash_rescale;
8777c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_full_magazines;
8787c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_empty_magazines;
8797c478bd9Sstevel@tonic-gate 	kstat_named_t	kmc_magazine_size;
880686031edSTom Erickson 	kstat_named_t	kmc_reap; /* number of kmem_cache_reap() calls */
881686031edSTom Erickson 	kstat_named_t	kmc_defrag; /* attempts to defrag all partial slabs */
882686031edSTom Erickson 	kstat_named_t	kmc_scan; /* attempts to defrag one partial slab */
883686031edSTom Erickson 	kstat_named_t	kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
884b5fca8f8Stomee 	kstat_named_t	kmc_move_yes;
885b5fca8f8Stomee 	kstat_named_t	kmc_move_no;
886b5fca8f8Stomee 	kstat_named_t	kmc_move_later;
887b5fca8f8Stomee 	kstat_named_t	kmc_move_dont_need;
888686031edSTom Erickson 	kstat_named_t	kmc_move_dont_know; /* obj unrecognized by client ... */
889686031edSTom Erickson 	kstat_named_t	kmc_move_hunt_found; /* ... but found in mag layer */
890686031edSTom Erickson 	kstat_named_t	kmc_move_slabs_freed; /* slabs freed by consolidator */
891686031edSTom Erickson 	kstat_named_t	kmc_move_reclaimable; /* buffers, if consolidator ran */
8927c478bd9Sstevel@tonic-gate } kmem_cache_kstat = {
8937c478bd9Sstevel@tonic-gate 	{ "buf_size",		KSTAT_DATA_UINT64 },
8947c478bd9Sstevel@tonic-gate 	{ "align",		KSTAT_DATA_UINT64 },
8957c478bd9Sstevel@tonic-gate 	{ "chunk_size",		KSTAT_DATA_UINT64 },
8967c478bd9Sstevel@tonic-gate 	{ "slab_size",		KSTAT_DATA_UINT64 },
8977c478bd9Sstevel@tonic-gate 	{ "alloc",		KSTAT_DATA_UINT64 },
8987c478bd9Sstevel@tonic-gate 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
8997c478bd9Sstevel@tonic-gate 	{ "free",		KSTAT_DATA_UINT64 },
9007c478bd9Sstevel@tonic-gate 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
9017c478bd9Sstevel@tonic-gate 	{ "depot_free",		KSTAT_DATA_UINT64 },
9027c478bd9Sstevel@tonic-gate 	{ "depot_contention",	KSTAT_DATA_UINT64 },
9037c478bd9Sstevel@tonic-gate 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
9047c478bd9Sstevel@tonic-gate 	{ "slab_free",		KSTAT_DATA_UINT64 },
9057c478bd9Sstevel@tonic-gate 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
9067c478bd9Sstevel@tonic-gate 	{ "buf_avail",		KSTAT_DATA_UINT64 },
9077c478bd9Sstevel@tonic-gate 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
9087c478bd9Sstevel@tonic-gate 	{ "buf_total",		KSTAT_DATA_UINT64 },
9097c478bd9Sstevel@tonic-gate 	{ "buf_max",		KSTAT_DATA_UINT64 },
9107c478bd9Sstevel@tonic-gate 	{ "slab_create",	KSTAT_DATA_UINT64 },
9117c478bd9Sstevel@tonic-gate 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
9127c478bd9Sstevel@tonic-gate 	{ "vmem_source",	KSTAT_DATA_UINT64 },
9137c478bd9Sstevel@tonic-gate 	{ "hash_size",		KSTAT_DATA_UINT64 },
9147c478bd9Sstevel@tonic-gate 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
9157c478bd9Sstevel@tonic-gate 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
9167c478bd9Sstevel@tonic-gate 	{ "full_magazines",	KSTAT_DATA_UINT64 },
9177c478bd9Sstevel@tonic-gate 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
9187c478bd9Sstevel@tonic-gate 	{ "magazine_size",	KSTAT_DATA_UINT64 },
919686031edSTom Erickson 	{ "reap",		KSTAT_DATA_UINT64 },
920686031edSTom Erickson 	{ "defrag",		KSTAT_DATA_UINT64 },
921686031edSTom Erickson 	{ "scan",		KSTAT_DATA_UINT64 },
922b5fca8f8Stomee 	{ "move_callbacks",	KSTAT_DATA_UINT64 },
923b5fca8f8Stomee 	{ "move_yes",		KSTAT_DATA_UINT64 },
924b5fca8f8Stomee 	{ "move_no",		KSTAT_DATA_UINT64 },
925b5fca8f8Stomee 	{ "move_later",		KSTAT_DATA_UINT64 },
926b5fca8f8Stomee 	{ "move_dont_need",	KSTAT_DATA_UINT64 },
927b5fca8f8Stomee 	{ "move_dont_know",	KSTAT_DATA_UINT64 },
928b5fca8f8Stomee 	{ "move_hunt_found",	KSTAT_DATA_UINT64 },
929686031edSTom Erickson 	{ "move_slabs_freed",	KSTAT_DATA_UINT64 },
930686031edSTom Erickson 	{ "move_reclaimable",	KSTAT_DATA_UINT64 },
9317c478bd9Sstevel@tonic-gate };
9327c478bd9Sstevel@tonic-gate 
9337c478bd9Sstevel@tonic-gate static kmutex_t kmem_cache_kstat_lock;
9347c478bd9Sstevel@tonic-gate 
9357c478bd9Sstevel@tonic-gate /*
9367c478bd9Sstevel@tonic-gate  * The default set of caches to back kmem_alloc().
9377c478bd9Sstevel@tonic-gate  * These sizes should be reevaluated periodically.
9387c478bd9Sstevel@tonic-gate  *
9397c478bd9Sstevel@tonic-gate  * We want allocations that are multiples of the coherency granularity
9407c478bd9Sstevel@tonic-gate  * (64 bytes) to be satisfied from a cache which is a multiple of 64
9417c478bd9Sstevel@tonic-gate  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
9427c478bd9Sstevel@tonic-gate  * the next kmem_cache_size greater than or equal to it must be a
9437c478bd9Sstevel@tonic-gate  * multiple of 64.
944dce01e3fSJonathan W Adams  *
945dce01e3fSJonathan W Adams  * We split the table into two sections:  size <= 4k and size > 4k.  This
946dce01e3fSJonathan W Adams  * saves a lot of space and cache footprint in our cache tables.
9477c478bd9Sstevel@tonic-gate  */
9487c478bd9Sstevel@tonic-gate static const int kmem_alloc_sizes[] = {
9497c478bd9Sstevel@tonic-gate 	1 * 8,
9507c478bd9Sstevel@tonic-gate 	2 * 8,
9517c478bd9Sstevel@tonic-gate 	3 * 8,
9527c478bd9Sstevel@tonic-gate 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
9537c478bd9Sstevel@tonic-gate 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
9547c478bd9Sstevel@tonic-gate 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
9557c478bd9Sstevel@tonic-gate 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
9567c478bd9Sstevel@tonic-gate 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
9577c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 7, 64),
9587c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 6, 64),
9597c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 5, 64),
9607c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 4, 64),
9617c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 3, 64),
9627c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 2, 64),
9637c478bd9Sstevel@tonic-gate };
9647c478bd9Sstevel@tonic-gate 
965dce01e3fSJonathan W Adams static const int kmem_big_alloc_sizes[] = {
966dce01e3fSJonathan W Adams 	2 * 4096,	3 * 4096,
967dce01e3fSJonathan W Adams 	2 * 8192,	3 * 8192,
968dce01e3fSJonathan W Adams 	4 * 8192,	5 * 8192,	6 * 8192,	7 * 8192,
969dce01e3fSJonathan W Adams 	8 * 8192,	9 * 8192,	10 * 8192,	11 * 8192,
970dce01e3fSJonathan W Adams 	12 * 8192,	13 * 8192,	14 * 8192,	15 * 8192,
971dce01e3fSJonathan W Adams 	16 * 8192
972dce01e3fSJonathan W Adams };
973dce01e3fSJonathan W Adams 
974dce01e3fSJonathan W Adams #define	KMEM_MAXBUF		4096
975dce01e3fSJonathan W Adams #define	KMEM_BIG_MAXBUF_32BIT	32768
976dce01e3fSJonathan W Adams #define	KMEM_BIG_MAXBUF		131072
977dce01e3fSJonathan W Adams 
978dce01e3fSJonathan W Adams #define	KMEM_BIG_MULTIPLE	4096	/* big_alloc_sizes must be a multiple */
979dce01e3fSJonathan W Adams #define	KMEM_BIG_SHIFT		12	/* lg(KMEM_BIG_MULTIPLE) */
9807c478bd9Sstevel@tonic-gate 
9817c478bd9Sstevel@tonic-gate static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
982dce01e3fSJonathan W Adams static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
983dce01e3fSJonathan W Adams 
984dce01e3fSJonathan W Adams #define	KMEM_ALLOC_TABLE_MAX	(KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
985dce01e3fSJonathan W Adams static size_t kmem_big_alloc_table_max = 0;	/* # of filled elements */
9867c478bd9Sstevel@tonic-gate 
9877c478bd9Sstevel@tonic-gate static kmem_magtype_t kmem_magtype[] = {
9887c478bd9Sstevel@tonic-gate 	{ 1,	8,	3200,	65536	},
9897c478bd9Sstevel@tonic-gate 	{ 3,	16,	256,	32768	},
9907c478bd9Sstevel@tonic-gate 	{ 7,	32,	64,	16384	},
9917c478bd9Sstevel@tonic-gate 	{ 15,	64,	0,	8192	},
9927c478bd9Sstevel@tonic-gate 	{ 31,	64,	0,	4096	},
9937c478bd9Sstevel@tonic-gate 	{ 47,	64,	0,	2048	},
9947c478bd9Sstevel@tonic-gate 	{ 63,	64,	0,	1024	},
9957c478bd9Sstevel@tonic-gate 	{ 95,	64,	0,	512	},
9967c478bd9Sstevel@tonic-gate 	{ 143,	64,	0,	0	},
9977c478bd9Sstevel@tonic-gate };
9987c478bd9Sstevel@tonic-gate 
9997c478bd9Sstevel@tonic-gate static uint32_t kmem_reaping;
10007c478bd9Sstevel@tonic-gate static uint32_t kmem_reaping_idspace;
10017c478bd9Sstevel@tonic-gate 
10027c478bd9Sstevel@tonic-gate /*
10037c478bd9Sstevel@tonic-gate  * kmem tunables
10047c478bd9Sstevel@tonic-gate  */
10057c478bd9Sstevel@tonic-gate clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
10067c478bd9Sstevel@tonic-gate int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
10077c478bd9Sstevel@tonic-gate pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
10087c478bd9Sstevel@tonic-gate int kmem_panic = 1;		/* whether to panic on error */
10097c478bd9Sstevel@tonic-gate int kmem_logging = 1;		/* kmem_log_enter() override */
10107c478bd9Sstevel@tonic-gate uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
10117c478bd9Sstevel@tonic-gate size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
10127c478bd9Sstevel@tonic-gate size_t kmem_content_log_size;	/* content log size [2% of memory] */
10137c478bd9Sstevel@tonic-gate size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
10147c478bd9Sstevel@tonic-gate size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
1015d1580181SBryan Cantrill size_t kmem_zerosized_log_size;	/* zero-sized log [4 pages per CPU] */
10167c478bd9Sstevel@tonic-gate size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
10177c478bd9Sstevel@tonic-gate size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
10187c478bd9Sstevel@tonic-gate size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
10197c478bd9Sstevel@tonic-gate int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
10207c478bd9Sstevel@tonic-gate size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
10217c478bd9Sstevel@tonic-gate size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
10227c478bd9Sstevel@tonic-gate 
1023d1580181SBryan Cantrill #ifdef DEBUG
1024d1580181SBryan Cantrill int kmem_warn_zerosized = 1;	/* whether to warn on zero-sized KM_SLEEP */
1025d1580181SBryan Cantrill #else
1026d1580181SBryan Cantrill int kmem_warn_zerosized = 0;	/* whether to warn on zero-sized KM_SLEEP */
1027d1580181SBryan Cantrill #endif
1028d1580181SBryan Cantrill 
1029d1580181SBryan Cantrill int kmem_panic_zerosized = 0;	/* whether to panic on zero-sized KM_SLEEP */
1030d1580181SBryan Cantrill 
1031dce01e3fSJonathan W Adams #ifdef _LP64
1032dce01e3fSJonathan W Adams size_t	kmem_max_cached = KMEM_BIG_MAXBUF;	/* maximum kmem_alloc cache */
1033dce01e3fSJonathan W Adams #else
1034dce01e3fSJonathan W Adams size_t	kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1035dce01e3fSJonathan W Adams #endif
1036dce01e3fSJonathan W Adams 
10377c478bd9Sstevel@tonic-gate #ifdef DEBUG
10387c478bd9Sstevel@tonic-gate int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
10397c478bd9Sstevel@tonic-gate #else
10407c478bd9Sstevel@tonic-gate int kmem_flags = 0;
10417c478bd9Sstevel@tonic-gate #endif
10427c478bd9Sstevel@tonic-gate int kmem_ready;
10437c478bd9Sstevel@tonic-gate 
10447c478bd9Sstevel@tonic-gate static kmem_cache_t	*kmem_slab_cache;
10457c478bd9Sstevel@tonic-gate static kmem_cache_t	*kmem_bufctl_cache;
10467c478bd9Sstevel@tonic-gate static kmem_cache_t	*kmem_bufctl_audit_cache;
10477c478bd9Sstevel@tonic-gate 
10487c478bd9Sstevel@tonic-gate static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
1049b5fca8f8Stomee static list_t		kmem_caches;
10507c478bd9Sstevel@tonic-gate 
10517c478bd9Sstevel@tonic-gate static taskq_t		*kmem_taskq;
10527c478bd9Sstevel@tonic-gate static kmutex_t		kmem_flags_lock;
10537c478bd9Sstevel@tonic-gate static vmem_t		*kmem_metadata_arena;
10547c478bd9Sstevel@tonic-gate static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
10557c478bd9Sstevel@tonic-gate static vmem_t		*kmem_cache_arena;
10567c478bd9Sstevel@tonic-gate static vmem_t		*kmem_hash_arena;
10577c478bd9Sstevel@tonic-gate static vmem_t		*kmem_log_arena;
10587c478bd9Sstevel@tonic-gate static vmem_t		*kmem_oversize_arena;
10597c478bd9Sstevel@tonic-gate static vmem_t		*kmem_va_arena;
10607c478bd9Sstevel@tonic-gate static vmem_t		*kmem_default_arena;
10617c478bd9Sstevel@tonic-gate static vmem_t		*kmem_firewall_va_arena;
10627c478bd9Sstevel@tonic-gate static vmem_t		*kmem_firewall_arena;
10637c478bd9Sstevel@tonic-gate 
1064d1580181SBryan Cantrill static int		kmem_zerosized;		/* # of zero-sized allocs */
1065d1580181SBryan Cantrill 
1066b5fca8f8Stomee /*
1067b5fca8f8Stomee  * kmem slab consolidator thresholds (tunables)
1068b5fca8f8Stomee  */
1069686031edSTom Erickson size_t kmem_frag_minslabs = 101;	/* minimum total slabs */
1070686031edSTom Erickson size_t kmem_frag_numer = 1;		/* free buffers (numerator) */
1071686031edSTom Erickson size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1072b5fca8f8Stomee /*
1073b5fca8f8Stomee  * Maximum number of slabs from which to move buffers during a single
1074b5fca8f8Stomee  * maintenance interval while the system is not low on memory.
1075b5fca8f8Stomee  */
1076686031edSTom Erickson size_t kmem_reclaim_max_slabs = 1;
1077b5fca8f8Stomee /*
1078b5fca8f8Stomee  * Number of slabs to scan backwards from the end of the partial slab list
1079b5fca8f8Stomee  * when searching for buffers to relocate.
1080b5fca8f8Stomee  */
1081686031edSTom Erickson size_t kmem_reclaim_scan_range = 12;
1082b5fca8f8Stomee 
1083b5fca8f8Stomee /* consolidator knobs */
1084929d5b43SMatthew Ahrens boolean_t kmem_move_noreap;
1085929d5b43SMatthew Ahrens boolean_t kmem_move_blocked;
1086929d5b43SMatthew Ahrens boolean_t kmem_move_fulltilt;
1087929d5b43SMatthew Ahrens boolean_t kmem_move_any_partial;
1088b5fca8f8Stomee 
1089b5fca8f8Stomee #ifdef	DEBUG
1090b5fca8f8Stomee /*
1091686031edSTom Erickson  * kmem consolidator debug tunables:
1092b5fca8f8Stomee  * Ensure code coverage by occasionally running the consolidator even when the
1093b5fca8f8Stomee  * caches are not fragmented (they may never be). These intervals are mean time
1094b5fca8f8Stomee  * in cache maintenance intervals (kmem_cache_update).
1095b5fca8f8Stomee  */
1096686031edSTom Erickson uint32_t kmem_mtb_move = 60;	/* defrag 1 slab (~15min) */
1097686031edSTom Erickson uint32_t kmem_mtb_reap = 1800;	/* defrag all slabs (~7.5hrs) */
1098b5fca8f8Stomee #endif	/* DEBUG */
1099