Print this page
8493 kmem_move taskq appears to be inducing significant system latency
Reviewed by: Robert Mustacchi <rm@joyent.com>
Reviewed by: Dan McDonald <danmcd@joyent.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/os/kmem.c
+++ new/usr/src/uts/common/os/kmem.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 1994, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 24 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
25 25 */
26 26
27 27 /*
28 28 * Kernel memory allocator, as described in the following two papers and a
29 29 * statement about the consolidator:
30 30 *
31 31 * Jeff Bonwick,
32 32 * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
33 33 * Proceedings of the Summer 1994 Usenix Conference.
34 34 * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
35 35 *
36 36 * Jeff Bonwick and Jonathan Adams,
37 37 * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
38 38 * Arbitrary Resources.
39 39 * Proceedings of the 2001 Usenix Conference.
40 40 * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
41 41 *
42 42 * kmem Slab Consolidator Big Theory Statement:
43 43 *
44 44 * 1. Motivation
45 45 *
46 46 * As stated in Bonwick94, slabs provide the following advantages over other
47 47 * allocation structures in terms of memory fragmentation:
48 48 *
49 49 * - Internal fragmentation (per-buffer wasted space) is minimal.
50 50 * - Severe external fragmentation (unused buffers on the free list) is
51 51 * unlikely.
52 52 *
53 53 * Segregating objects by size eliminates one source of external fragmentation,
54 54 * and according to Bonwick:
55 55 *
56 56 * The other reason that slabs reduce external fragmentation is that all
57 57 * objects in a slab are of the same type, so they have the same lifetime
58 58 * distribution. The resulting segregation of short-lived and long-lived
59 59 * objects at slab granularity reduces the likelihood of an entire page being
60 60 * held hostage due to a single long-lived allocation [Barrett93, Hanson90].
61 61 *
62 62 * While unlikely, severe external fragmentation remains possible. Clients that
63 63 * allocate both short- and long-lived objects from the same cache cannot
64 64 * anticipate the distribution of long-lived objects within the allocator's slab
65 65 * implementation. Even a small percentage of long-lived objects distributed
66 66 * randomly across many slabs can lead to a worst case scenario where the client
67 67 * frees the majority of its objects and the system gets back almost none of the
68 68 * slabs. Despite the client doing what it reasonably can to help the system
69 69 * reclaim memory, the allocator cannot shake free enough slabs because of
70 70 * lonely allocations stubbornly hanging on. Although the allocator is in a
71 71 * position to diagnose the fragmentation, there is nothing that the allocator
72 72 * by itself can do about it. It only takes a single allocated object to prevent
73 73 * an entire slab from being reclaimed, and any object handed out by
74 74 * kmem_cache_alloc() is by definition in the client's control. Conversely,
75 75 * although the client is in a position to move a long-lived object, it has no
76 76 * way of knowing if the object is causing fragmentation, and if so, where to
77 77 * move it. A solution necessarily requires further cooperation between the
78 78 * allocator and the client.
79 79 *
80 80 * 2. Move Callback
81 81 *
82 82 * The kmem slab consolidator therefore adds a move callback to the
83 83 * allocator/client interface, improving worst-case external fragmentation in
84 84 * kmem caches that supply a function to move objects from one memory location
85 85 * to another. In a situation of low memory kmem attempts to consolidate all of
86 86 * a cache's slabs at once; otherwise it works slowly to bring external
87 87 * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
88 88 * thereby helping to avoid a low memory situation in the future.
89 89 *
90 90 * The callback has the following signature:
91 91 *
92 92 * kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
93 93 *
94 94 * It supplies the kmem client with two addresses: the allocated object that
95 95 * kmem wants to move and a buffer selected by kmem for the client to use as the
96 96 * copy destination. The callback is kmem's way of saying "Please get off of
97 97 * this buffer and use this one instead." kmem knows where it wants to move the
98 98 * object in order to best reduce fragmentation. All the client needs to know
99 99 * about the second argument (void *new) is that it is an allocated, constructed
100 100 * object ready to take the contents of the old object. When the move function
101 101 * is called, the system is likely to be low on memory, and the new object
102 102 * spares the client from having to worry about allocating memory for the
103 103 * requested move. The third argument supplies the size of the object, in case a
104 104 * single move function handles multiple caches whose objects differ only in
105 105 * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
106 106 * user argument passed to the constructor, destructor, and reclaim functions is
107 107 * also passed to the move callback.
108 108 *
109 109 * 2.1 Setting the Move Callback
110 110 *
111 111 * The client sets the move callback after creating the cache and before
112 112 * allocating from it:
113 113 *
114 114 * object_cache = kmem_cache_create(...);
115 115 * kmem_cache_set_move(object_cache, object_move);
116 116 *
117 117 * 2.2 Move Callback Return Values
118 118 *
119 119 * Only the client knows about its own data and when is a good time to move it.
120 120 * The client is cooperating with kmem to return unused memory to the system,
121 121 * and kmem respectfully accepts this help at the client's convenience. When
122 122 * asked to move an object, the client can respond with any of the following:
123 123 *
124 124 * typedef enum kmem_cbrc {
125 125 * KMEM_CBRC_YES,
126 126 * KMEM_CBRC_NO,
127 127 * KMEM_CBRC_LATER,
128 128 * KMEM_CBRC_DONT_NEED,
129 129 * KMEM_CBRC_DONT_KNOW
130 130 * } kmem_cbrc_t;
131 131 *
132 132 * The client must not explicitly kmem_cache_free() either of the objects passed
133 133 * to the callback, since kmem wants to free them directly to the slab layer
134 134 * (bypassing the per-CPU magazine layer). The response tells kmem which of the
135 135 * objects to free:
136 136 *
137 137 * YES: (Did it) The client moved the object, so kmem frees the old one.
138 138 * NO: (Never) The client refused, so kmem frees the new object (the
139 139 * unused copy destination). kmem also marks the slab of the old
140 140 * object so as not to bother the client with further callbacks for
141 141 * that object as long as the slab remains on the partial slab list.
142 142 * (The system won't be getting the slab back as long as the
143 143 * immovable object holds it hostage, so there's no point in moving
144 144 * any of its objects.)
145 145 * LATER: The client is using the object and cannot move it now, so kmem
146 146 * frees the new object (the unused copy destination). kmem still
147 147 * attempts to move other objects off the slab, since it expects to
148 148 * succeed in clearing the slab in a later callback. The client
149 149 * should use LATER instead of NO if the object is likely to become
150 150 * movable very soon.
151 151 * DONT_NEED: The client no longer needs the object, so kmem frees the old along
|
↓ open down ↓ |
151 lines elided |
↑ open up ↑ |
152 152 * with the new object (the unused copy destination). This response
153 153 * is the client's opportunity to be a model citizen and give back as
154 154 * much as it can.
155 155 * DONT_KNOW: The client does not know about the object because
156 156 * a) the client has just allocated the object and not yet put it
157 157 * wherever it expects to find known objects
158 158 * b) the client has removed the object from wherever it expects to
159 159 * find known objects and is about to free it, or
160 160 * c) the client has freed the object.
161 161 * In all these cases (a, b, and c) kmem frees the new object (the
162 - * unused copy destination) and searches for the old object in the
163 - * magazine layer. If found, the object is removed from the magazine
164 - * layer and freed to the slab layer so it will no longer hold the
165 - * slab hostage.
162 + * unused copy destination). In the first case, the object is in
163 + * use and the correct action is that for LATER; in the latter two
164 + * cases, we know that the object is either freed or about to be
165 + * freed, in which case it is either already in a magazine or about
166 + * to be in one. In these cases, we know that the object will either
167 + * be reallocated and reused, or it will end up in a full magazine
168 + * that will be reaped (thereby liberating the slab). Because it
169 + * is prohibitively expensive to differentiate these cases, and
170 + * because the defrag code is executed when we're low on memory
171 + * (thereby biasing the system to reclaim full magazines) we treat
172 + * all DONT_KNOW cases as LATER and rely on cache reaping to
173 + * generally clean up full magazines. While we take the same action
174 + * for these cases, we maintain their semantic distinction: if
175 + * defragmentation is not occurring, it is useful to know if this
176 + * is due to objects in use (LATER) or objects in an unknown state
177 + * of transition (DONT_KNOW).
166 178 *
167 179 * 2.3 Object States
168 180 *
169 181 * Neither kmem nor the client can be assumed to know the object's whereabouts
170 182 * at the time of the callback. An object belonging to a kmem cache may be in
171 183 * any of the following states:
172 184 *
173 185 * 1. Uninitialized on the slab
174 186 * 2. Allocated from the slab but not constructed (still uninitialized)
175 187 * 3. Allocated from the slab, constructed, but not yet ready for business
176 188 * (not in a valid state for the move callback)
177 189 * 4. In use (valid and known to the client)
178 190 * 5. About to be freed (no longer in a valid state for the move callback)
179 191 * 6. Freed to a magazine (still constructed)
180 192 * 7. Allocated from a magazine, not yet ready for business (not in a valid
181 193 * state for the move callback), and about to return to state #4
182 194 * 8. Deconstructed on a magazine that is about to be freed
183 195 * 9. Freed to the slab
184 196 *
185 197 * Since the move callback may be called at any time while the object is in any
186 198 * of the above states (except state #1), the client needs a safe way to
187 199 * determine whether or not it knows about the object. Specifically, the client
188 200 * needs to know whether or not the object is in state #4, the only state in
189 201 * which a move is valid. If the object is in any other state, the client should
190 202 * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
191 203 * the object's fields.
192 204 *
193 205 * Note that although an object may be in state #4 when kmem initiates the move
194 206 * request, the object may no longer be in that state by the time kmem actually
195 207 * calls the move function. Not only does the client free objects
196 208 * asynchronously, kmem itself puts move requests on a queue where thay are
197 209 * pending until kmem processes them from another context. Also, objects freed
198 210 * to a magazine appear allocated from the point of view of the slab layer, so
199 211 * kmem may even initiate requests for objects in a state other than state #4.
200 212 *
201 213 * 2.3.1 Magazine Layer
202 214 *
203 215 * An important insight revealed by the states listed above is that the magazine
204 216 * layer is populated only by kmem_cache_free(). Magazines of constructed
205 217 * objects are never populated directly from the slab layer (which contains raw,
206 218 * unconstructed objects). Whenever an allocation request cannot be satisfied
207 219 * from the magazine layer, the magazines are bypassed and the request is
208 220 * satisfied from the slab layer (creating a new slab if necessary). kmem calls
209 221 * the object constructor only when allocating from the slab layer, and only in
210 222 * response to kmem_cache_alloc() or to prepare the destination buffer passed in
211 223 * the move callback. kmem does not preconstruct objects in anticipation of
212 224 * kmem_cache_alloc().
213 225 *
214 226 * 2.3.2 Object Constructor and Destructor
215 227 *
216 228 * If the client supplies a destructor, it must be valid to call the destructor
217 229 * on a newly created object (immediately after the constructor).
218 230 *
219 231 * 2.4 Recognizing Known Objects
220 232 *
221 233 * There is a simple test to determine safely whether or not the client knows
222 234 * about a given object in the move callback. It relies on the fact that kmem
223 235 * guarantees that the object of the move callback has only been touched by the
224 236 * client itself or else by kmem. kmem does this by ensuring that none of the
225 237 * cache's slabs are freed to the virtual memory (VM) subsystem while a move
226 238 * callback is pending. When the last object on a slab is freed, if there is a
227 239 * pending move, kmem puts the slab on a per-cache dead list and defers freeing
228 240 * slabs on that list until all pending callbacks are completed. That way,
229 241 * clients can be certain that the object of a move callback is in one of the
230 242 * states listed above, making it possible to distinguish known objects (in
231 243 * state #4) using the two low order bits of any pointer member (with the
232 244 * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
233 245 * platforms).
234 246 *
235 247 * The test works as long as the client always transitions objects from state #4
236 248 * (known, in use) to state #5 (about to be freed, invalid) by setting the low
237 249 * order bit of the client-designated pointer member. Since kmem only writes
238 250 * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
239 251 * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
240 252 * guaranteed to set at least one of the two low order bits. Therefore, given an
241 253 * object with a back pointer to a 'container_t *o_container', the client can
242 254 * test
243 255 *
244 256 * container_t *container = object->o_container;
245 257 * if ((uintptr_t)container & 0x3) {
246 258 * return (KMEM_CBRC_DONT_KNOW);
247 259 * }
248 260 *
249 261 * Typically, an object will have a pointer to some structure with a list or
250 262 * hash where objects from the cache are kept while in use. Assuming that the
251 263 * client has some way of knowing that the container structure is valid and will
252 264 * not go away during the move, and assuming that the structure includes a lock
253 265 * to protect whatever collection is used, then the client would continue as
254 266 * follows:
255 267 *
256 268 * // Ensure that the container structure does not go away.
257 269 * if (container_hold(container) == 0) {
258 270 * return (KMEM_CBRC_DONT_KNOW);
259 271 * }
260 272 * mutex_enter(&container->c_objects_lock);
261 273 * if (container != object->o_container) {
262 274 * mutex_exit(&container->c_objects_lock);
263 275 * container_rele(container);
264 276 * return (KMEM_CBRC_DONT_KNOW);
265 277 * }
266 278 *
267 279 * At this point the client knows that the object cannot be freed as long as
268 280 * c_objects_lock is held. Note that after acquiring the lock, the client must
269 281 * recheck the o_container pointer in case the object was removed just before
270 282 * acquiring the lock.
271 283 *
272 284 * When the client is about to free an object, it must first remove that object
273 285 * from the list, hash, or other structure where it is kept. At that time, to
274 286 * mark the object so it can be distinguished from the remaining, known objects,
275 287 * the client sets the designated low order bit:
276 288 *
277 289 * mutex_enter(&container->c_objects_lock);
|
↓ open down ↓ |
102 lines elided |
↑ open up ↑ |
278 290 * object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
279 291 * list_remove(&container->c_objects, object);
280 292 * mutex_exit(&container->c_objects_lock);
281 293 *
282 294 * In the common case, the object is freed to the magazine layer, where it may
283 295 * be reused on a subsequent allocation without the overhead of calling the
284 296 * constructor. While in the magazine it appears allocated from the point of
285 297 * view of the slab layer, making it a candidate for the move callback. Most
286 298 * objects unrecognized by the client in the move callback fall into this
287 299 * category and are cheaply distinguished from known objects by the test
288 - * described earlier. Since recognition is cheap for the client, and searching
289 - * magazines is expensive for kmem, kmem defers searching until the client first
290 - * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
291 - * elsewhere does what it can to avoid bothering the client unnecessarily.
300 + * described earlier. Because searching magazines is prohibitively expensive
301 + * for kmem, clients that do not mark freed objects (and therefore return
302 + * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
303 + * efficacy reduced.
292 304 *
293 305 * Invalidating the designated pointer member before freeing the object marks
294 306 * the object to be avoided in the callback, and conversely, assigning a valid
295 307 * value to the designated pointer member after allocating the object makes the
296 308 * object fair game for the callback:
297 309 *
298 310 * ... allocate object ...
299 311 * ... set any initial state not set by the constructor ...
300 312 *
301 313 * mutex_enter(&container->c_objects_lock);
302 314 * list_insert_tail(&container->c_objects, object);
303 315 * membar_producer();
304 316 * object->o_container = container;
305 317 * mutex_exit(&container->c_objects_lock);
306 318 *
307 319 * Note that everything else must be valid before setting o_container makes the
308 320 * object fair game for the move callback. The membar_producer() call ensures
309 321 * that all the object's state is written to memory before setting the pointer
310 322 * that transitions the object from state #3 or #7 (allocated, constructed, not
311 323 * yet in use) to state #4 (in use, valid). That's important because the move
312 324 * function has to check the validity of the pointer before it can safely
313 325 * acquire the lock protecting the collection where it expects to find known
314 326 * objects.
315 327 *
316 328 * This method of distinguishing known objects observes the usual symmetry:
317 329 * invalidating the designated pointer is the first thing the client does before
318 330 * freeing the object, and setting the designated pointer is the last thing the
319 331 * client does after allocating the object. Of course, the client is not
320 332 * required to use this method. Fundamentally, how the client recognizes known
321 333 * objects is completely up to the client, but this method is recommended as an
322 334 * efficient and safe way to take advantage of the guarantees made by kmem. If
323 335 * the entire object is arbitrary data without any markable bits from a suitable
324 336 * pointer member, then the client must find some other method, such as
325 337 * searching a hash table of known objects.
326 338 *
327 339 * 2.5 Preventing Objects From Moving
328 340 *
329 341 * Besides a way to distinguish known objects, the other thing that the client
330 342 * needs is a strategy to ensure that an object will not move while the client
331 343 * is actively using it. The details of satisfying this requirement tend to be
332 344 * highly cache-specific. It might seem that the same rules that let a client
333 345 * remove an object safely should also decide when an object can be moved
334 346 * safely. However, any object state that makes a removal attempt invalid is
335 347 * likely to be long-lasting for objects that the client does not expect to
336 348 * remove. kmem knows nothing about the object state and is equally likely (from
337 349 * the client's point of view) to request a move for any object in the cache,
338 350 * whether prepared for removal or not. Even a low percentage of objects stuck
339 351 * in place by unremovability will defeat the consolidator if the stuck objects
340 352 * are the same long-lived allocations likely to hold slabs hostage.
341 353 * Fundamentally, the consolidator is not aimed at common cases. Severe external
342 354 * fragmentation is a worst case scenario manifested as sparsely allocated
343 355 * slabs, by definition a low percentage of the cache's objects. When deciding
344 356 * what makes an object movable, keep in mind the goal of the consolidator: to
345 357 * bring worst-case external fragmentation within the limits guaranteed for
346 358 * internal fragmentation. Removability is a poor criterion if it is likely to
347 359 * exclude more than an insignificant percentage of objects for long periods of
348 360 * time.
349 361 *
350 362 * A tricky general solution exists, and it has the advantage of letting you
351 363 * move any object at almost any moment, practically eliminating the likelihood
352 364 * that an object can hold a slab hostage. However, if there is a cache-specific
353 365 * way to ensure that an object is not actively in use in the vast majority of
354 366 * cases, a simpler solution that leverages this cache-specific knowledge is
355 367 * preferred.
356 368 *
357 369 * 2.5.1 Cache-Specific Solution
358 370 *
359 371 * As an example of a cache-specific solution, the ZFS znode cache takes
360 372 * advantage of the fact that the vast majority of znodes are only being
361 373 * referenced from the DNLC. (A typical case might be a few hundred in active
362 374 * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
363 375 * client has established that it recognizes the znode and can access its fields
364 376 * safely (using the method described earlier), it then tests whether the znode
365 377 * is referenced by anything other than the DNLC. If so, it assumes that the
366 378 * znode may be in active use and is unsafe to move, so it drops its locks and
367 379 * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
368 380 * else znodes are used, no change is needed to protect against the possibility
369 381 * of the znode moving. The disadvantage is that it remains possible for an
370 382 * application to hold a znode slab hostage with an open file descriptor.
371 383 * However, this case ought to be rare and the consolidator has a way to deal
372 384 * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
373 385 * object, kmem eventually stops believing it and treats the slab as if the
374 386 * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
375 387 * then focus on getting it off of the partial slab list by allocating rather
376 388 * than freeing all of its objects. (Either way of getting a slab off the
377 389 * free list reduces fragmentation.)
378 390 *
379 391 * 2.5.2 General Solution
380 392 *
381 393 * The general solution, on the other hand, requires an explicit hold everywhere
382 394 * the object is used to prevent it from moving. To keep the client locking
383 395 * strategy as uncomplicated as possible, kmem guarantees the simplifying
384 396 * assumption that move callbacks are sequential, even across multiple caches.
385 397 * Internally, a global queue processed by a single thread supports all caches
386 398 * implementing the callback function. No matter how many caches supply a move
387 399 * function, the consolidator never moves more than one object at a time, so the
388 400 * client does not have to worry about tricky lock ordering involving several
389 401 * related objects from different kmem caches.
390 402 *
391 403 * The general solution implements the explicit hold as a read-write lock, which
392 404 * allows multiple readers to access an object from the cache simultaneously
393 405 * while a single writer is excluded from moving it. A single rwlock for the
394 406 * entire cache would lock out all threads from using any of the cache's objects
395 407 * even though only a single object is being moved, so to reduce contention,
396 408 * the client can fan out the single rwlock into an array of rwlocks hashed by
397 409 * the object address, making it probable that moving one object will not
398 410 * prevent other threads from using a different object. The rwlock cannot be a
399 411 * member of the object itself, because the possibility of the object moving
400 412 * makes it unsafe to access any of the object's fields until the lock is
401 413 * acquired.
402 414 *
403 415 * Assuming a small, fixed number of locks, it's possible that multiple objects
404 416 * will hash to the same lock. A thread that needs to use multiple objects in
405 417 * the same function may acquire the same lock multiple times. Since rwlocks are
406 418 * reentrant for readers, and since there is never more than a single writer at
407 419 * a time (assuming that the client acquires the lock as a writer only when
408 420 * moving an object inside the callback), there would seem to be no problem.
409 421 * However, a client locking multiple objects in the same function must handle
410 422 * one case of potential deadlock: Assume that thread A needs to prevent both
411 423 * object 1 and object 2 from moving, and thread B, the callback, meanwhile
412 424 * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
413 425 * same lock, that thread A will acquire the lock for object 1 as a reader
414 426 * before thread B sets the lock's write-wanted bit, preventing thread A from
415 427 * reacquiring the lock for object 2 as a reader. Unable to make forward
416 428 * progress, thread A will never release the lock for object 1, resulting in
417 429 * deadlock.
418 430 *
419 431 * There are two ways of avoiding the deadlock just described. The first is to
420 432 * use rw_tryenter() rather than rw_enter() in the callback function when
421 433 * attempting to acquire the lock as a writer. If tryenter discovers that the
422 434 * same object (or another object hashed to the same lock) is already in use, it
423 435 * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
424 436 * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
425 437 * since it allows a thread to acquire the lock as a reader in spite of a
426 438 * waiting writer. This second approach insists on moving the object now, no
427 439 * matter how many readers the move function must wait for in order to do so,
428 440 * and could delay the completion of the callback indefinitely (blocking
429 441 * callbacks to other clients). In practice, a less insistent callback using
430 442 * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
431 443 * little reason to use anything else.
432 444 *
433 445 * Avoiding deadlock is not the only problem that an implementation using an
434 446 * explicit hold needs to solve. Locking the object in the first place (to
435 447 * prevent it from moving) remains a problem, since the object could move
436 448 * between the time you obtain a pointer to the object and the time you acquire
437 449 * the rwlock hashed to that pointer value. Therefore the client needs to
438 450 * recheck the value of the pointer after acquiring the lock, drop the lock if
439 451 * the value has changed, and try again. This requires a level of indirection:
440 452 * something that points to the object rather than the object itself, that the
441 453 * client can access safely while attempting to acquire the lock. (The object
442 454 * itself cannot be referenced safely because it can move at any time.)
443 455 * The following lock-acquisition function takes whatever is safe to reference
444 456 * (arg), follows its pointer to the object (using function f), and tries as
445 457 * often as necessary to acquire the hashed lock and verify that the object
446 458 * still has not moved:
447 459 *
448 460 * object_t *
449 461 * object_hold(object_f f, void *arg)
450 462 * {
451 463 * object_t *op;
452 464 *
453 465 * op = f(arg);
454 466 * if (op == NULL) {
455 467 * return (NULL);
456 468 * }
457 469 *
458 470 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
459 471 * while (op != f(arg)) {
460 472 * rw_exit(OBJECT_RWLOCK(op));
461 473 * op = f(arg);
462 474 * if (op == NULL) {
463 475 * break;
464 476 * }
465 477 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
466 478 * }
467 479 *
468 480 * return (op);
469 481 * }
470 482 *
471 483 * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
472 484 * lock reacquisition loop, while necessary, almost never executes. The function
473 485 * pointer f (used to obtain the object pointer from arg) has the following type
474 486 * definition:
475 487 *
476 488 * typedef object_t *(*object_f)(void *arg);
477 489 *
478 490 * An object_f implementation is likely to be as simple as accessing a structure
479 491 * member:
480 492 *
481 493 * object_t *
482 494 * s_object(void *arg)
483 495 * {
484 496 * something_t *sp = arg;
485 497 * return (sp->s_object);
486 498 * }
487 499 *
488 500 * The flexibility of a function pointer allows the path to the object to be
489 501 * arbitrarily complex and also supports the notion that depending on where you
490 502 * are using the object, you may need to get it from someplace different.
491 503 *
492 504 * The function that releases the explicit hold is simpler because it does not
493 505 * have to worry about the object moving:
494 506 *
495 507 * void
496 508 * object_rele(object_t *op)
497 509 * {
498 510 * rw_exit(OBJECT_RWLOCK(op));
499 511 * }
500 512 *
501 513 * The caller is spared these details so that obtaining and releasing an
502 514 * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
503 515 * of object_hold() only needs to know that the returned object pointer is valid
504 516 * if not NULL and that the object will not move until released.
505 517 *
506 518 * Although object_hold() prevents an object from moving, it does not prevent it
507 519 * from being freed. The caller must take measures before calling object_hold()
508 520 * (afterwards is too late) to ensure that the held object cannot be freed. The
509 521 * caller must do so without accessing the unsafe object reference, so any lock
510 522 * or reference count used to ensure the continued existence of the object must
511 523 * live outside the object itself.
512 524 *
513 525 * Obtaining a new object is a special case where an explicit hold is impossible
514 526 * for the caller. Any function that returns a newly allocated object (either as
515 527 * a return value, or as an in-out paramter) must return it already held; after
516 528 * the caller gets it is too late, since the object cannot be safely accessed
517 529 * without the level of indirection described earlier. The following
518 530 * object_alloc() example uses the same code shown earlier to transition a new
519 531 * object into the state of being recognized (by the client) as a known object.
520 532 * The function must acquire the hold (rw_enter) before that state transition
521 533 * makes the object movable:
522 534 *
523 535 * static object_t *
524 536 * object_alloc(container_t *container)
525 537 * {
526 538 * object_t *object = kmem_cache_alloc(object_cache, 0);
527 539 * ... set any initial state not set by the constructor ...
528 540 * rw_enter(OBJECT_RWLOCK(object), RW_READER);
529 541 * mutex_enter(&container->c_objects_lock);
530 542 * list_insert_tail(&container->c_objects, object);
531 543 * membar_producer();
532 544 * object->o_container = container;
533 545 * mutex_exit(&container->c_objects_lock);
534 546 * return (object);
535 547 * }
536 548 *
537 549 * Functions that implicitly acquire an object hold (any function that calls
538 550 * object_alloc() to supply an object for the caller) need to be carefully noted
539 551 * so that the matching object_rele() is not neglected. Otherwise, leaked holds
540 552 * prevent all objects hashed to the affected rwlocks from ever being moved.
541 553 *
542 554 * The pointer to a held object can be hashed to the holding rwlock even after
543 555 * the object has been freed. Although it is possible to release the hold
544 556 * after freeing the object, you may decide to release the hold implicitly in
545 557 * whatever function frees the object, so as to release the hold as soon as
546 558 * possible, and for the sake of symmetry with the function that implicitly
547 559 * acquires the hold when it allocates the object. Here, object_free() releases
548 560 * the hold acquired by object_alloc(). Its implicit object_rele() forms a
549 561 * matching pair with object_hold():
550 562 *
551 563 * void
552 564 * object_free(object_t *object)
553 565 * {
554 566 * container_t *container;
555 567 *
556 568 * ASSERT(object_held(object));
557 569 * container = object->o_container;
558 570 * mutex_enter(&container->c_objects_lock);
559 571 * object->o_container =
560 572 * (void *)((uintptr_t)object->o_container | 0x1);
561 573 * list_remove(&container->c_objects, object);
562 574 * mutex_exit(&container->c_objects_lock);
563 575 * object_rele(object);
564 576 * kmem_cache_free(object_cache, object);
565 577 * }
566 578 *
567 579 * Note that object_free() cannot safely accept an object pointer as an argument
568 580 * unless the object is already held. Any function that calls object_free()
569 581 * needs to be carefully noted since it similarly forms a matching pair with
570 582 * object_hold().
571 583 *
572 584 * To complete the picture, the following callback function implements the
573 585 * general solution by moving objects only if they are currently unheld:
574 586 *
575 587 * static kmem_cbrc_t
576 588 * object_move(void *buf, void *newbuf, size_t size, void *arg)
577 589 * {
578 590 * object_t *op = buf, *np = newbuf;
579 591 * container_t *container;
580 592 *
581 593 * container = op->o_container;
582 594 * if ((uintptr_t)container & 0x3) {
583 595 * return (KMEM_CBRC_DONT_KNOW);
584 596 * }
585 597 *
586 598 * // Ensure that the container structure does not go away.
587 599 * if (container_hold(container) == 0) {
588 600 * return (KMEM_CBRC_DONT_KNOW);
589 601 * }
590 602 *
591 603 * mutex_enter(&container->c_objects_lock);
592 604 * if (container != op->o_container) {
593 605 * mutex_exit(&container->c_objects_lock);
594 606 * container_rele(container);
595 607 * return (KMEM_CBRC_DONT_KNOW);
596 608 * }
597 609 *
598 610 * if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
599 611 * mutex_exit(&container->c_objects_lock);
600 612 * container_rele(container);
601 613 * return (KMEM_CBRC_LATER);
602 614 * }
603 615 *
604 616 * object_move_impl(op, np); // critical section
605 617 * rw_exit(OBJECT_RWLOCK(op));
606 618 *
607 619 * op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
608 620 * list_link_replace(&op->o_link_node, &np->o_link_node);
609 621 * mutex_exit(&container->c_objects_lock);
610 622 * container_rele(container);
611 623 * return (KMEM_CBRC_YES);
612 624 * }
613 625 *
614 626 * Note that object_move() must invalidate the designated o_container pointer of
615 627 * the old object in the same way that object_free() does, since kmem will free
616 628 * the object in response to the KMEM_CBRC_YES return value.
617 629 *
618 630 * The lock order in object_move() differs from object_alloc(), which locks
619 631 * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
620 632 * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
621 633 * not a problem. Holding the lock on the object list in the example above
622 634 * through the entire callback not only prevents the object from going away, it
623 635 * also allows you to lock the list elsewhere and know that none of its elements
624 636 * will move during iteration.
625 637 *
626 638 * Adding an explicit hold everywhere an object from the cache is used is tricky
627 639 * and involves much more change to client code than a cache-specific solution
628 640 * that leverages existing state to decide whether or not an object is
629 641 * movable. However, this approach has the advantage that no object remains
630 642 * immovable for any significant length of time, making it extremely unlikely
631 643 * that long-lived allocations can continue holding slabs hostage; and it works
632 644 * for any cache.
633 645 *
634 646 * 3. Consolidator Implementation
635 647 *
636 648 * Once the client supplies a move function that a) recognizes known objects and
637 649 * b) avoids moving objects that are actively in use, the remaining work is up
638 650 * to the consolidator to decide which objects to move and when to issue
639 651 * callbacks.
640 652 *
641 653 * The consolidator relies on the fact that a cache's slabs are ordered by
642 654 * usage. Each slab has a fixed number of objects. Depending on the slab's
643 655 * "color" (the offset of the first object from the beginning of the slab;
644 656 * offsets are staggered to mitigate false sharing of cache lines) it is either
645 657 * the maximum number of objects per slab determined at cache creation time or
646 658 * else the number closest to the maximum that fits within the space remaining
647 659 * after the initial offset. A completely allocated slab may contribute some
648 660 * internal fragmentation (per-slab overhead) but no external fragmentation, so
649 661 * it is of no interest to the consolidator. At the other extreme, slabs whose
650 662 * objects have all been freed to the slab are released to the virtual memory
651 663 * (VM) subsystem (objects freed to magazines are still allocated as far as the
652 664 * slab is concerned). External fragmentation exists when there are slabs
653 665 * somewhere between these extremes. A partial slab has at least one but not all
654 666 * of its objects allocated. The more partial slabs, and the fewer allocated
655 667 * objects on each of them, the higher the fragmentation. Hence the
656 668 * consolidator's overall strategy is to reduce the number of partial slabs by
657 669 * moving allocated objects from the least allocated slabs to the most allocated
658 670 * slabs.
659 671 *
660 672 * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
661 673 * slabs are kept separately in an unordered list. Since the majority of slabs
662 674 * tend to be completely allocated (a typical unfragmented cache may have
663 675 * thousands of complete slabs and only a single partial slab), separating
664 676 * complete slabs improves the efficiency of partial slab ordering, since the
665 677 * complete slabs do not affect the depth or balance of the AVL tree. This
666 678 * ordered sequence of partial slabs acts as a "free list" supplying objects for
667 679 * allocation requests.
668 680 *
669 681 * Objects are always allocated from the first partial slab in the free list,
670 682 * where the allocation is most likely to eliminate a partial slab (by
671 683 * completely allocating it). Conversely, when a single object from a completely
672 684 * allocated slab is freed to the slab, that slab is added to the front of the
673 685 * free list. Since most free list activity involves highly allocated slabs
674 686 * coming and going at the front of the list, slabs tend naturally toward the
675 687 * ideal order: highly allocated at the front, sparsely allocated at the back.
676 688 * Slabs with few allocated objects are likely to become completely free if they
677 689 * keep a safe distance away from the front of the free list. Slab misorders
678 690 * interfere with the natural tendency of slabs to become completely free or
679 691 * completely allocated. For example, a slab with a single allocated object
680 692 * needs only a single free to escape the cache; its natural desire is
681 693 * frustrated when it finds itself at the front of the list where a second
682 694 * allocation happens just before the free could have released it. Another slab
683 695 * with all but one object allocated might have supplied the buffer instead, so
684 696 * that both (as opposed to neither) of the slabs would have been taken off the
685 697 * free list.
686 698 *
687 699 * Although slabs tend naturally toward the ideal order, misorders allowed by a
688 700 * simple list implementation defeat the consolidator's strategy of merging
689 701 * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
690 702 * needs another way to fix misorders to optimize its callback strategy. One
691 703 * approach is to periodically scan a limited number of slabs, advancing a
692 704 * marker to hold the current scan position, and to move extreme misorders to
693 705 * the front or back of the free list and to the front or back of the current
694 706 * scan range. By making consecutive scan ranges overlap by one slab, the least
695 707 * allocated slab in the current range can be carried along from the end of one
696 708 * scan to the start of the next.
697 709 *
698 710 * Maintaining partial slabs in an AVL tree relieves kmem of this additional
699 711 * task, however. Since most of the cache's activity is in the magazine layer,
700 712 * and allocations from the slab layer represent only a startup cost, the
701 713 * overhead of maintaining a balanced tree is not a significant concern compared
702 714 * to the opportunity of reducing complexity by eliminating the partial slab
703 715 * scanner just described. The overhead of an AVL tree is minimized by
704 716 * maintaining only partial slabs in the tree and keeping completely allocated
705 717 * slabs separately in a list. To avoid increasing the size of the slab
706 718 * structure the AVL linkage pointers are reused for the slab's list linkage,
707 719 * since the slab will always be either partial or complete, never stored both
708 720 * ways at the same time. To further minimize the overhead of the AVL tree the
709 721 * compare function that orders partial slabs by usage divides the range of
710 722 * allocated object counts into bins such that counts within the same bin are
711 723 * considered equal. Binning partial slabs makes it less likely that allocating
712 724 * or freeing a single object will change the slab's order, requiring a tree
713 725 * reinsertion (an avl_remove() followed by an avl_add(), both potentially
714 726 * requiring some rebalancing of the tree). Allocation counts closest to
715 727 * completely free and completely allocated are left unbinned (finely sorted) to
716 728 * better support the consolidator's strategy of merging slabs at either
717 729 * extreme.
718 730 *
719 731 * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
720 732 *
721 733 * The consolidator piggybacks on the kmem maintenance thread and is called on
722 734 * the same interval as kmem_cache_update(), once per cache every fifteen
723 735 * seconds. kmem maintains a running count of unallocated objects in the slab
724 736 * layer (cache_bufslab). The consolidator checks whether that number exceeds
725 737 * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
726 738 * there is a significant number of slabs in the cache (arbitrarily a minimum
727 739 * 101 total slabs). Unused objects that have fallen out of the magazine layer's
728 740 * working set are included in the assessment, and magazines in the depot are
729 741 * reaped if those objects would lift cache_bufslab above the fragmentation
730 742 * threshold. Once the consolidator decides that a cache is fragmented, it looks
731 743 * for a candidate slab to reclaim, starting at the end of the partial slab free
732 744 * list and scanning backwards. At first the consolidator is choosy: only a slab
733 745 * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
734 746 * single allocated object, regardless of percentage). If there is difficulty
735 747 * finding a candidate slab, kmem raises the allocation threshold incrementally,
736 748 * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
737 749 * external fragmentation (unused objects on the free list) below 12.5% (1/8),
738 750 * even in the worst case of every slab in the cache being almost 7/8 allocated.
739 751 * The threshold can also be lowered incrementally when candidate slabs are easy
740 752 * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
741 753 * is no longer fragmented.
742 754 *
743 755 * 3.2 Generating Callbacks
744 756 *
745 757 * Once an eligible slab is chosen, a callback is generated for every allocated
746 758 * object on the slab, in the hope that the client will move everything off the
747 759 * slab and make it reclaimable. Objects selected as move destinations are
748 760 * chosen from slabs at the front of the free list. Assuming slabs in the ideal
749 761 * order (most allocated at the front, least allocated at the back) and a
750 762 * cooperative client, the consolidator will succeed in removing slabs from both
751 763 * ends of the free list, completely allocating on the one hand and completely
752 764 * freeing on the other. Objects selected as move destinations are allocated in
753 765 * the kmem maintenance thread where move requests are enqueued. A separate
754 766 * callback thread removes pending callbacks from the queue and calls the
755 767 * client. The separate thread ensures that client code (the move function) does
756 768 * not interfere with internal kmem maintenance tasks. A map of pending
757 769 * callbacks keyed by object address (the object to be moved) is checked to
758 770 * ensure that duplicate callbacks are not generated for the same object.
759 771 * Allocating the move destination (the object to move to) prevents subsequent
760 772 * callbacks from selecting the same destination as an earlier pending callback.
761 773 *
762 774 * Move requests can also be generated by kmem_cache_reap() when the system is
763 775 * desperate for memory and by kmem_cache_move_notify(), called by the client to
764 776 * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
765 777 * The map of pending callbacks is protected by the same lock that protects the
766 778 * slab layer.
767 779 *
768 780 * When the system is desperate for memory, kmem does not bother to determine
769 781 * whether or not the cache exceeds the fragmentation threshold, but tries to
770 782 * consolidate as many slabs as possible. Normally, the consolidator chews
771 783 * slowly, one sparsely allocated slab at a time during each maintenance
772 784 * interval that the cache is fragmented. When desperate, the consolidator
773 785 * starts at the last partial slab and enqueues callbacks for every allocated
774 786 * object on every partial slab, working backwards until it reaches the first
775 787 * partial slab. The first partial slab, meanwhile, advances in pace with the
776 788 * consolidator as allocations to supply move destinations for the enqueued
777 789 * callbacks use up the highly allocated slabs at the front of the free list.
778 790 * Ideally, the overgrown free list collapses like an accordion, starting at
779 791 * both ends and ending at the center with a single partial slab.
780 792 *
781 793 * 3.3 Client Responses
782 794 *
783 795 * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
784 796 * marks the slab that supplied the stuck object non-reclaimable and moves it to
785 797 * front of the free list. The slab remains marked as long as it remains on the
786 798 * free list, and it appears more allocated to the partial slab compare function
787 799 * than any unmarked slab, no matter how many of its objects are allocated.
788 800 * Since even one immovable object ties up the entire slab, the goal is to
789 801 * completely allocate any slab that cannot be completely freed. kmem does not
790 802 * bother generating callbacks to move objects from a marked slab unless the
791 803 * system is desperate.
792 804 *
793 805 * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
794 806 * slab. If the client responds LATER too many times, kmem disbelieves and
795 807 * treats the response as a NO. The count is cleared when the slab is taken off
796 808 * the partial slab list or when the client moves one of the slab's objects.
797 809 *
798 810 * 4. Observability
799 811 *
800 812 * A kmem cache's external fragmentation is best observed with 'mdb -k' using
801 813 * the ::kmem_slabs dcmd. For a complete description of the command, enter
802 814 * '::help kmem_slabs' at the mdb prompt.
803 815 */
804 816
805 817 #include <sys/kmem_impl.h>
806 818 #include <sys/vmem_impl.h>
807 819 #include <sys/param.h>
808 820 #include <sys/sysmacros.h>
809 821 #include <sys/vm.h>
810 822 #include <sys/proc.h>
811 823 #include <sys/tuneable.h>
812 824 #include <sys/systm.h>
813 825 #include <sys/cmn_err.h>
814 826 #include <sys/debug.h>
815 827 #include <sys/sdt.h>
816 828 #include <sys/mutex.h>
817 829 #include <sys/bitmap.h>
818 830 #include <sys/atomic.h>
819 831 #include <sys/kobj.h>
820 832 #include <sys/disp.h>
821 833 #include <vm/seg_kmem.h>
822 834 #include <sys/log.h>
823 835 #include <sys/callb.h>
824 836 #include <sys/taskq.h>
825 837 #include <sys/modctl.h>
826 838 #include <sys/reboot.h>
827 839 #include <sys/id32.h>
828 840 #include <sys/zone.h>
829 841 #include <sys/netstack.h>
830 842 #ifdef DEBUG
831 843 #include <sys/random.h>
832 844 #endif
833 845
834 846 extern void streams_msg_init(void);
835 847 extern int segkp_fromheap;
836 848 extern void segkp_cache_free(void);
837 849 extern int callout_init_done;
838 850
839 851 struct kmem_cache_kstat {
840 852 kstat_named_t kmc_buf_size;
841 853 kstat_named_t kmc_align;
842 854 kstat_named_t kmc_chunk_size;
843 855 kstat_named_t kmc_slab_size;
844 856 kstat_named_t kmc_alloc;
845 857 kstat_named_t kmc_alloc_fail;
846 858 kstat_named_t kmc_free;
847 859 kstat_named_t kmc_depot_alloc;
848 860 kstat_named_t kmc_depot_free;
849 861 kstat_named_t kmc_depot_contention;
850 862 kstat_named_t kmc_slab_alloc;
851 863 kstat_named_t kmc_slab_free;
852 864 kstat_named_t kmc_buf_constructed;
853 865 kstat_named_t kmc_buf_avail;
854 866 kstat_named_t kmc_buf_inuse;
855 867 kstat_named_t kmc_buf_total;
856 868 kstat_named_t kmc_buf_max;
857 869 kstat_named_t kmc_slab_create;
858 870 kstat_named_t kmc_slab_destroy;
859 871 kstat_named_t kmc_vmem_source;
860 872 kstat_named_t kmc_hash_size;
861 873 kstat_named_t kmc_hash_lookup_depth;
862 874 kstat_named_t kmc_hash_rescale;
863 875 kstat_named_t kmc_full_magazines;
864 876 kstat_named_t kmc_empty_magazines;
865 877 kstat_named_t kmc_magazine_size;
866 878 kstat_named_t kmc_reap; /* number of kmem_cache_reap() calls */
867 879 kstat_named_t kmc_defrag; /* attempts to defrag all partial slabs */
868 880 kstat_named_t kmc_scan; /* attempts to defrag one partial slab */
869 881 kstat_named_t kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
870 882 kstat_named_t kmc_move_yes;
871 883 kstat_named_t kmc_move_no;
872 884 kstat_named_t kmc_move_later;
873 885 kstat_named_t kmc_move_dont_need;
874 886 kstat_named_t kmc_move_dont_know; /* obj unrecognized by client ... */
875 887 kstat_named_t kmc_move_hunt_found; /* ... but found in mag layer */
876 888 kstat_named_t kmc_move_slabs_freed; /* slabs freed by consolidator */
877 889 kstat_named_t kmc_move_reclaimable; /* buffers, if consolidator ran */
878 890 } kmem_cache_kstat = {
879 891 { "buf_size", KSTAT_DATA_UINT64 },
880 892 { "align", KSTAT_DATA_UINT64 },
881 893 { "chunk_size", KSTAT_DATA_UINT64 },
882 894 { "slab_size", KSTAT_DATA_UINT64 },
883 895 { "alloc", KSTAT_DATA_UINT64 },
884 896 { "alloc_fail", KSTAT_DATA_UINT64 },
885 897 { "free", KSTAT_DATA_UINT64 },
886 898 { "depot_alloc", KSTAT_DATA_UINT64 },
887 899 { "depot_free", KSTAT_DATA_UINT64 },
888 900 { "depot_contention", KSTAT_DATA_UINT64 },
889 901 { "slab_alloc", KSTAT_DATA_UINT64 },
890 902 { "slab_free", KSTAT_DATA_UINT64 },
891 903 { "buf_constructed", KSTAT_DATA_UINT64 },
892 904 { "buf_avail", KSTAT_DATA_UINT64 },
893 905 { "buf_inuse", KSTAT_DATA_UINT64 },
894 906 { "buf_total", KSTAT_DATA_UINT64 },
895 907 { "buf_max", KSTAT_DATA_UINT64 },
896 908 { "slab_create", KSTAT_DATA_UINT64 },
897 909 { "slab_destroy", KSTAT_DATA_UINT64 },
898 910 { "vmem_source", KSTAT_DATA_UINT64 },
899 911 { "hash_size", KSTAT_DATA_UINT64 },
900 912 { "hash_lookup_depth", KSTAT_DATA_UINT64 },
901 913 { "hash_rescale", KSTAT_DATA_UINT64 },
902 914 { "full_magazines", KSTAT_DATA_UINT64 },
903 915 { "empty_magazines", KSTAT_DATA_UINT64 },
904 916 { "magazine_size", KSTAT_DATA_UINT64 },
905 917 { "reap", KSTAT_DATA_UINT64 },
906 918 { "defrag", KSTAT_DATA_UINT64 },
907 919 { "scan", KSTAT_DATA_UINT64 },
908 920 { "move_callbacks", KSTAT_DATA_UINT64 },
909 921 { "move_yes", KSTAT_DATA_UINT64 },
910 922 { "move_no", KSTAT_DATA_UINT64 },
911 923 { "move_later", KSTAT_DATA_UINT64 },
912 924 { "move_dont_need", KSTAT_DATA_UINT64 },
913 925 { "move_dont_know", KSTAT_DATA_UINT64 },
914 926 { "move_hunt_found", KSTAT_DATA_UINT64 },
915 927 { "move_slabs_freed", KSTAT_DATA_UINT64 },
916 928 { "move_reclaimable", KSTAT_DATA_UINT64 },
917 929 };
918 930
919 931 static kmutex_t kmem_cache_kstat_lock;
920 932
921 933 /*
922 934 * The default set of caches to back kmem_alloc().
923 935 * These sizes should be reevaluated periodically.
924 936 *
925 937 * We want allocations that are multiples of the coherency granularity
926 938 * (64 bytes) to be satisfied from a cache which is a multiple of 64
927 939 * bytes, so that it will be 64-byte aligned. For all multiples of 64,
928 940 * the next kmem_cache_size greater than or equal to it must be a
929 941 * multiple of 64.
930 942 *
931 943 * We split the table into two sections: size <= 4k and size > 4k. This
932 944 * saves a lot of space and cache footprint in our cache tables.
933 945 */
934 946 static const int kmem_alloc_sizes[] = {
935 947 1 * 8,
936 948 2 * 8,
937 949 3 * 8,
938 950 4 * 8, 5 * 8, 6 * 8, 7 * 8,
939 951 4 * 16, 5 * 16, 6 * 16, 7 * 16,
940 952 4 * 32, 5 * 32, 6 * 32, 7 * 32,
941 953 4 * 64, 5 * 64, 6 * 64, 7 * 64,
942 954 4 * 128, 5 * 128, 6 * 128, 7 * 128,
943 955 P2ALIGN(8192 / 7, 64),
944 956 P2ALIGN(8192 / 6, 64),
945 957 P2ALIGN(8192 / 5, 64),
946 958 P2ALIGN(8192 / 4, 64),
947 959 P2ALIGN(8192 / 3, 64),
948 960 P2ALIGN(8192 / 2, 64),
949 961 };
950 962
951 963 static const int kmem_big_alloc_sizes[] = {
952 964 2 * 4096, 3 * 4096,
953 965 2 * 8192, 3 * 8192,
954 966 4 * 8192, 5 * 8192, 6 * 8192, 7 * 8192,
955 967 8 * 8192, 9 * 8192, 10 * 8192, 11 * 8192,
956 968 12 * 8192, 13 * 8192, 14 * 8192, 15 * 8192,
957 969 16 * 8192
958 970 };
959 971
960 972 #define KMEM_MAXBUF 4096
961 973 #define KMEM_BIG_MAXBUF_32BIT 32768
962 974 #define KMEM_BIG_MAXBUF 131072
963 975
964 976 #define KMEM_BIG_MULTIPLE 4096 /* big_alloc_sizes must be a multiple */
965 977 #define KMEM_BIG_SHIFT 12 /* lg(KMEM_BIG_MULTIPLE) */
966 978
967 979 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
968 980 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
969 981
970 982 #define KMEM_ALLOC_TABLE_MAX (KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
971 983 static size_t kmem_big_alloc_table_max = 0; /* # of filled elements */
972 984
973 985 static kmem_magtype_t kmem_magtype[] = {
974 986 { 1, 8, 3200, 65536 },
975 987 { 3, 16, 256, 32768 },
976 988 { 7, 32, 64, 16384 },
977 989 { 15, 64, 0, 8192 },
978 990 { 31, 64, 0, 4096 },
979 991 { 47, 64, 0, 2048 },
980 992 { 63, 64, 0, 1024 },
981 993 { 95, 64, 0, 512 },
982 994 { 143, 64, 0, 0 },
983 995 };
984 996
985 997 static uint32_t kmem_reaping;
986 998 static uint32_t kmem_reaping_idspace;
987 999
988 1000 /*
989 1001 * kmem tunables
990 1002 */
991 1003 clock_t kmem_reap_interval; /* cache reaping rate [15 * HZ ticks] */
992 1004 int kmem_depot_contention = 3; /* max failed tryenters per real interval */
993 1005 pgcnt_t kmem_reapahead = 0; /* start reaping N pages before pageout */
994 1006 int kmem_panic = 1; /* whether to panic on error */
995 1007 int kmem_logging = 1; /* kmem_log_enter() override */
996 1008 uint32_t kmem_mtbf = 0; /* mean time between failures [default: off] */
997 1009 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
998 1010 size_t kmem_content_log_size; /* content log size [2% of memory] */
999 1011 size_t kmem_failure_log_size; /* failure log [4 pages per CPU] */
1000 1012 size_t kmem_slab_log_size; /* slab create log [4 pages per CPU] */
1001 1013 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1002 1014 size_t kmem_lite_minsize = 0; /* minimum buffer size for KMF_LITE */
1003 1015 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1004 1016 int kmem_lite_pcs = 4; /* number of PCs to store in KMF_LITE mode */
1005 1017 size_t kmem_maxverify; /* maximum bytes to inspect in debug routines */
1006 1018 size_t kmem_minfirewall; /* hardware-enforced redzone threshold */
1007 1019
1008 1020 #ifdef _LP64
1009 1021 size_t kmem_max_cached = KMEM_BIG_MAXBUF; /* maximum kmem_alloc cache */
1010 1022 #else
1011 1023 size_t kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1012 1024 #endif
1013 1025
1014 1026 #ifdef DEBUG
1015 1027 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1016 1028 #else
1017 1029 int kmem_flags = 0;
1018 1030 #endif
1019 1031 int kmem_ready;
1020 1032
1021 1033 static kmem_cache_t *kmem_slab_cache;
1022 1034 static kmem_cache_t *kmem_bufctl_cache;
1023 1035 static kmem_cache_t *kmem_bufctl_audit_cache;
1024 1036
1025 1037 static kmutex_t kmem_cache_lock; /* inter-cache linkage only */
1026 1038 static list_t kmem_caches;
1027 1039
1028 1040 static taskq_t *kmem_taskq;
1029 1041 static kmutex_t kmem_flags_lock;
1030 1042 static vmem_t *kmem_metadata_arena;
1031 1043 static vmem_t *kmem_msb_arena; /* arena for metadata caches */
|
↓ open down ↓ |
730 lines elided |
↑ open up ↑ |
1032 1044 static vmem_t *kmem_cache_arena;
1033 1045 static vmem_t *kmem_hash_arena;
1034 1046 static vmem_t *kmem_log_arena;
1035 1047 static vmem_t *kmem_oversize_arena;
1036 1048 static vmem_t *kmem_va_arena;
1037 1049 static vmem_t *kmem_default_arena;
1038 1050 static vmem_t *kmem_firewall_va_arena;
1039 1051 static vmem_t *kmem_firewall_arena;
1040 1052
1041 1053 /*
1042 - * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1043 - * turned on when DEBUG is also defined.
1044 - */
1045 -#ifdef DEBUG
1046 -#define KMEM_STATS
1047 -#endif /* DEBUG */
1048 -
1049 -#ifdef KMEM_STATS
1050 -#define KMEM_STAT_ADD(stat) ((stat)++)
1051 -#define KMEM_STAT_COND_ADD(cond, stat) ((void) (!(cond) || (stat)++))
1052 -#else
1053 -#define KMEM_STAT_ADD(stat) /* nothing */
1054 -#define KMEM_STAT_COND_ADD(cond, stat) /* nothing */
1055 -#endif /* KMEM_STATS */
1056 -
1057 -/*
1058 1054 * kmem slab consolidator thresholds (tunables)
1059 1055 */
1060 1056 size_t kmem_frag_minslabs = 101; /* minimum total slabs */
1061 1057 size_t kmem_frag_numer = 1; /* free buffers (numerator) */
1062 1058 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1063 1059 /*
1064 1060 * Maximum number of slabs from which to move buffers during a single
1065 1061 * maintenance interval while the system is not low on memory.
1066 1062 */
1067 1063 size_t kmem_reclaim_max_slabs = 1;
1068 1064 /*
1069 1065 * Number of slabs to scan backwards from the end of the partial slab list
1070 1066 * when searching for buffers to relocate.
1071 1067 */
1072 1068 size_t kmem_reclaim_scan_range = 12;
1073 1069
1074 -#ifdef KMEM_STATS
1075 -static struct {
1076 - uint64_t kms_callbacks;
1077 - uint64_t kms_yes;
1078 - uint64_t kms_no;
1079 - uint64_t kms_later;
1080 - uint64_t kms_dont_need;
1081 - uint64_t kms_dont_know;
1082 - uint64_t kms_hunt_found_mag;
1083 - uint64_t kms_hunt_found_slab;
1084 - uint64_t kms_hunt_alloc_fail;
1085 - uint64_t kms_hunt_lucky;
1086 - uint64_t kms_notify;
1087 - uint64_t kms_notify_callbacks;
1088 - uint64_t kms_disbelief;
1089 - uint64_t kms_already_pending;
1090 - uint64_t kms_callback_alloc_fail;
1091 - uint64_t kms_callback_taskq_fail;
1092 - uint64_t kms_endscan_slab_dead;
1093 - uint64_t kms_endscan_slab_destroyed;
1094 - uint64_t kms_endscan_nomem;
1095 - uint64_t kms_endscan_refcnt_changed;
1096 - uint64_t kms_endscan_nomove_changed;
1097 - uint64_t kms_endscan_freelist;
1098 - uint64_t kms_avl_update;
1099 - uint64_t kms_avl_noupdate;
1100 - uint64_t kms_no_longer_reclaimable;
1101 - uint64_t kms_notify_no_longer_reclaimable;
1102 - uint64_t kms_notify_slab_dead;
1103 - uint64_t kms_notify_slab_destroyed;
1104 - uint64_t kms_alloc_fail;
1105 - uint64_t kms_constructor_fail;
1106 - uint64_t kms_dead_slabs_freed;
1107 - uint64_t kms_defrags;
1108 - uint64_t kms_scans;
1109 - uint64_t kms_scan_depot_ws_reaps;
1110 - uint64_t kms_debug_reaps;
1111 - uint64_t kms_debug_scans;
1112 -} kmem_move_stats;
1113 -#endif /* KMEM_STATS */
1114 -
1115 1070 /* consolidator knobs */
1116 1071 boolean_t kmem_move_noreap;
1117 1072 boolean_t kmem_move_blocked;
1118 1073 boolean_t kmem_move_fulltilt;
1119 1074 boolean_t kmem_move_any_partial;
1120 1075
1121 1076 #ifdef DEBUG
1122 1077 /*
1123 1078 * kmem consolidator debug tunables:
1124 1079 * Ensure code coverage by occasionally running the consolidator even when the
1125 1080 * caches are not fragmented (they may never be). These intervals are mean time
1126 1081 * in cache maintenance intervals (kmem_cache_update).
1127 1082 */
1128 1083 uint32_t kmem_mtb_move = 60; /* defrag 1 slab (~15min) */
1129 1084 uint32_t kmem_mtb_reap = 1800; /* defrag all slabs (~7.5hrs) */
1130 1085 #endif /* DEBUG */
1131 1086
1132 1087 static kmem_cache_t *kmem_defrag_cache;
1133 1088 static kmem_cache_t *kmem_move_cache;
1134 1089 static taskq_t *kmem_move_taskq;
1135 1090
1136 1091 static void kmem_cache_scan(kmem_cache_t *);
1137 1092 static void kmem_cache_defrag(kmem_cache_t *);
1138 1093 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1139 1094
1140 1095
1141 1096 kmem_log_header_t *kmem_transaction_log;
1142 1097 kmem_log_header_t *kmem_content_log;
1143 1098 kmem_log_header_t *kmem_failure_log;
1144 1099 kmem_log_header_t *kmem_slab_log;
1145 1100
1146 1101 static int kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1147 1102
1148 1103 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller) \
1149 1104 if ((count) > 0) { \
1150 1105 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1151 1106 pc_t *_e; \
1152 1107 /* memmove() the old entries down one notch */ \
1153 1108 for (_e = &_s[(count) - 1]; _e > _s; _e--) \
1154 1109 *_e = *(_e - 1); \
1155 1110 *_s = (uintptr_t)(caller); \
1156 1111 }
1157 1112
1158 1113 #define KMERR_MODIFIED 0 /* buffer modified while on freelist */
1159 1114 #define KMERR_REDZONE 1 /* redzone violation (write past end of buf) */
1160 1115 #define KMERR_DUPFREE 2 /* freed a buffer twice */
1161 1116 #define KMERR_BADADDR 3 /* freed a bad (unallocated) address */
1162 1117 #define KMERR_BADBUFTAG 4 /* buftag corrupted */
1163 1118 #define KMERR_BADBUFCTL 5 /* bufctl corrupted */
1164 1119 #define KMERR_BADCACHE 6 /* freed a buffer to the wrong cache */
1165 1120 #define KMERR_BADSIZE 7 /* alloc size != free size */
1166 1121 #define KMERR_BADBASE 8 /* buffer base address wrong */
1167 1122
1168 1123 struct {
1169 1124 hrtime_t kmp_timestamp; /* timestamp of panic */
1170 1125 int kmp_error; /* type of kmem error */
1171 1126 void *kmp_buffer; /* buffer that induced panic */
1172 1127 void *kmp_realbuf; /* real start address for buffer */
1173 1128 kmem_cache_t *kmp_cache; /* buffer's cache according to client */
1174 1129 kmem_cache_t *kmp_realcache; /* actual cache containing buffer */
1175 1130 kmem_slab_t *kmp_slab; /* slab accoring to kmem_findslab() */
1176 1131 kmem_bufctl_t *kmp_bufctl; /* bufctl */
1177 1132 } kmem_panic_info;
1178 1133
1179 1134
1180 1135 static void
1181 1136 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1182 1137 {
1183 1138 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1184 1139 uint64_t *buf = buf_arg;
1185 1140
1186 1141 while (buf < bufend)
1187 1142 *buf++ = pattern;
1188 1143 }
1189 1144
1190 1145 static void *
1191 1146 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1192 1147 {
1193 1148 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1194 1149 uint64_t *buf;
1195 1150
1196 1151 for (buf = buf_arg; buf < bufend; buf++)
1197 1152 if (*buf != pattern)
1198 1153 return (buf);
1199 1154 return (NULL);
1200 1155 }
1201 1156
1202 1157 static void *
1203 1158 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1204 1159 {
1205 1160 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1206 1161 uint64_t *buf;
1207 1162
1208 1163 for (buf = buf_arg; buf < bufend; buf++) {
1209 1164 if (*buf != old) {
1210 1165 copy_pattern(old, buf_arg,
1211 1166 (char *)buf - (char *)buf_arg);
1212 1167 return (buf);
1213 1168 }
1214 1169 *buf = new;
1215 1170 }
1216 1171
1217 1172 return (NULL);
1218 1173 }
1219 1174
1220 1175 static void
1221 1176 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1222 1177 {
1223 1178 kmem_cache_t *cp;
1224 1179
1225 1180 mutex_enter(&kmem_cache_lock);
1226 1181 for (cp = list_head(&kmem_caches); cp != NULL;
1227 1182 cp = list_next(&kmem_caches, cp))
1228 1183 if (tq != NULL)
1229 1184 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1230 1185 tqflag);
1231 1186 else
1232 1187 func(cp);
1233 1188 mutex_exit(&kmem_cache_lock);
1234 1189 }
1235 1190
1236 1191 static void
1237 1192 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1238 1193 {
1239 1194 kmem_cache_t *cp;
1240 1195
1241 1196 mutex_enter(&kmem_cache_lock);
1242 1197 for (cp = list_head(&kmem_caches); cp != NULL;
1243 1198 cp = list_next(&kmem_caches, cp)) {
1244 1199 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1245 1200 continue;
1246 1201 if (tq != NULL)
1247 1202 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1248 1203 tqflag);
1249 1204 else
1250 1205 func(cp);
1251 1206 }
1252 1207 mutex_exit(&kmem_cache_lock);
1253 1208 }
1254 1209
1255 1210 /*
1256 1211 * Debugging support. Given a buffer address, find its slab.
1257 1212 */
1258 1213 static kmem_slab_t *
1259 1214 kmem_findslab(kmem_cache_t *cp, void *buf)
1260 1215 {
1261 1216 kmem_slab_t *sp;
1262 1217
1263 1218 mutex_enter(&cp->cache_lock);
1264 1219 for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1265 1220 sp = list_next(&cp->cache_complete_slabs, sp)) {
1266 1221 if (KMEM_SLAB_MEMBER(sp, buf)) {
1267 1222 mutex_exit(&cp->cache_lock);
1268 1223 return (sp);
1269 1224 }
1270 1225 }
1271 1226 for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1272 1227 sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1273 1228 if (KMEM_SLAB_MEMBER(sp, buf)) {
1274 1229 mutex_exit(&cp->cache_lock);
1275 1230 return (sp);
1276 1231 }
1277 1232 }
1278 1233 mutex_exit(&cp->cache_lock);
1279 1234
1280 1235 return (NULL);
1281 1236 }
1282 1237
1283 1238 static void
1284 1239 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1285 1240 {
1286 1241 kmem_buftag_t *btp = NULL;
1287 1242 kmem_bufctl_t *bcp = NULL;
1288 1243 kmem_cache_t *cp = cparg;
1289 1244 kmem_slab_t *sp;
1290 1245 uint64_t *off;
1291 1246 void *buf = bufarg;
1292 1247
1293 1248 kmem_logging = 0; /* stop logging when a bad thing happens */
1294 1249
1295 1250 kmem_panic_info.kmp_timestamp = gethrtime();
1296 1251
1297 1252 sp = kmem_findslab(cp, buf);
1298 1253 if (sp == NULL) {
1299 1254 for (cp = list_tail(&kmem_caches); cp != NULL;
1300 1255 cp = list_prev(&kmem_caches, cp)) {
1301 1256 if ((sp = kmem_findslab(cp, buf)) != NULL)
1302 1257 break;
1303 1258 }
1304 1259 }
1305 1260
1306 1261 if (sp == NULL) {
1307 1262 cp = NULL;
1308 1263 error = KMERR_BADADDR;
1309 1264 } else {
1310 1265 if (cp != cparg)
1311 1266 error = KMERR_BADCACHE;
1312 1267 else
1313 1268 buf = (char *)bufarg - ((uintptr_t)bufarg -
1314 1269 (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1315 1270 if (buf != bufarg)
1316 1271 error = KMERR_BADBASE;
1317 1272 if (cp->cache_flags & KMF_BUFTAG)
1318 1273 btp = KMEM_BUFTAG(cp, buf);
1319 1274 if (cp->cache_flags & KMF_HASH) {
1320 1275 mutex_enter(&cp->cache_lock);
1321 1276 for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1322 1277 if (bcp->bc_addr == buf)
1323 1278 break;
1324 1279 mutex_exit(&cp->cache_lock);
1325 1280 if (bcp == NULL && btp != NULL)
1326 1281 bcp = btp->bt_bufctl;
1327 1282 if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1328 1283 NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1329 1284 bcp->bc_addr != buf) {
1330 1285 error = KMERR_BADBUFCTL;
1331 1286 bcp = NULL;
1332 1287 }
1333 1288 }
1334 1289 }
1335 1290
1336 1291 kmem_panic_info.kmp_error = error;
1337 1292 kmem_panic_info.kmp_buffer = bufarg;
1338 1293 kmem_panic_info.kmp_realbuf = buf;
1339 1294 kmem_panic_info.kmp_cache = cparg;
1340 1295 kmem_panic_info.kmp_realcache = cp;
1341 1296 kmem_panic_info.kmp_slab = sp;
1342 1297 kmem_panic_info.kmp_bufctl = bcp;
1343 1298
1344 1299 printf("kernel memory allocator: ");
1345 1300
1346 1301 switch (error) {
1347 1302
1348 1303 case KMERR_MODIFIED:
1349 1304 printf("buffer modified after being freed\n");
1350 1305 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1351 1306 if (off == NULL) /* shouldn't happen */
1352 1307 off = buf;
1353 1308 printf("modification occurred at offset 0x%lx "
1354 1309 "(0x%llx replaced by 0x%llx)\n",
1355 1310 (uintptr_t)off - (uintptr_t)buf,
1356 1311 (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1357 1312 break;
1358 1313
1359 1314 case KMERR_REDZONE:
1360 1315 printf("redzone violation: write past end of buffer\n");
1361 1316 break;
1362 1317
1363 1318 case KMERR_BADADDR:
1364 1319 printf("invalid free: buffer not in cache\n");
1365 1320 break;
1366 1321
1367 1322 case KMERR_DUPFREE:
1368 1323 printf("duplicate free: buffer freed twice\n");
1369 1324 break;
1370 1325
1371 1326 case KMERR_BADBUFTAG:
1372 1327 printf("boundary tag corrupted\n");
1373 1328 printf("bcp ^ bxstat = %lx, should be %lx\n",
1374 1329 (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1375 1330 KMEM_BUFTAG_FREE);
1376 1331 break;
1377 1332
1378 1333 case KMERR_BADBUFCTL:
1379 1334 printf("bufctl corrupted\n");
1380 1335 break;
1381 1336
1382 1337 case KMERR_BADCACHE:
1383 1338 printf("buffer freed to wrong cache\n");
1384 1339 printf("buffer was allocated from %s,\n", cp->cache_name);
1385 1340 printf("caller attempting free to %s.\n", cparg->cache_name);
1386 1341 break;
1387 1342
1388 1343 case KMERR_BADSIZE:
1389 1344 printf("bad free: free size (%u) != alloc size (%u)\n",
1390 1345 KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1391 1346 KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1392 1347 break;
1393 1348
1394 1349 case KMERR_BADBASE:
1395 1350 printf("bad free: free address (%p) != alloc address (%p)\n",
1396 1351 bufarg, buf);
1397 1352 break;
1398 1353 }
1399 1354
1400 1355 printf("buffer=%p bufctl=%p cache: %s\n",
1401 1356 bufarg, (void *)bcp, cparg->cache_name);
1402 1357
1403 1358 if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1404 1359 error != KMERR_BADBUFCTL) {
1405 1360 int d;
1406 1361 timestruc_t ts;
1407 1362 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1408 1363
1409 1364 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1410 1365 printf("previous transaction on buffer %p:\n", buf);
1411 1366 printf("thread=%p time=T-%ld.%09ld slab=%p cache: %s\n",
1412 1367 (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1413 1368 (void *)sp, cp->cache_name);
1414 1369 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1415 1370 ulong_t off;
1416 1371 char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1417 1372 printf("%s+%lx\n", sym ? sym : "?", off);
1418 1373 }
1419 1374 }
1420 1375 if (kmem_panic > 0)
1421 1376 panic("kernel heap corruption detected");
1422 1377 if (kmem_panic == 0)
1423 1378 debug_enter(NULL);
1424 1379 kmem_logging = 1; /* resume logging */
1425 1380 }
1426 1381
1427 1382 static kmem_log_header_t *
1428 1383 kmem_log_init(size_t logsize)
1429 1384 {
1430 1385 kmem_log_header_t *lhp;
1431 1386 int nchunks = 4 * max_ncpus;
1432 1387 size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1433 1388 int i;
1434 1389
1435 1390 /*
1436 1391 * Make sure that lhp->lh_cpu[] is nicely aligned
1437 1392 * to prevent false sharing of cache lines.
1438 1393 */
1439 1394 lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1440 1395 lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1441 1396 NULL, NULL, VM_SLEEP);
1442 1397 bzero(lhp, lhsize);
1443 1398
1444 1399 mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1445 1400 lhp->lh_nchunks = nchunks;
1446 1401 lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1447 1402 lhp->lh_base = vmem_alloc(kmem_log_arena,
1448 1403 lhp->lh_chunksize * nchunks, VM_SLEEP);
1449 1404 lhp->lh_free = vmem_alloc(kmem_log_arena,
1450 1405 nchunks * sizeof (int), VM_SLEEP);
1451 1406 bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1452 1407
1453 1408 for (i = 0; i < max_ncpus; i++) {
1454 1409 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1455 1410 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1456 1411 clhp->clh_chunk = i;
1457 1412 }
1458 1413
1459 1414 for (i = max_ncpus; i < nchunks; i++)
1460 1415 lhp->lh_free[i] = i;
1461 1416
1462 1417 lhp->lh_head = max_ncpus;
1463 1418 lhp->lh_tail = 0;
1464 1419
1465 1420 return (lhp);
1466 1421 }
1467 1422
1468 1423 static void *
1469 1424 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1470 1425 {
1471 1426 void *logspace;
1472 1427 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1473 1428
1474 1429 if (lhp == NULL || kmem_logging == 0 || panicstr)
1475 1430 return (NULL);
1476 1431
1477 1432 mutex_enter(&clhp->clh_lock);
1478 1433 clhp->clh_hits++;
1479 1434 if (size > clhp->clh_avail) {
1480 1435 mutex_enter(&lhp->lh_lock);
1481 1436 lhp->lh_hits++;
1482 1437 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1483 1438 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1484 1439 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1485 1440 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1486 1441 clhp->clh_current = lhp->lh_base +
1487 1442 clhp->clh_chunk * lhp->lh_chunksize;
1488 1443 clhp->clh_avail = lhp->lh_chunksize;
1489 1444 if (size > lhp->lh_chunksize)
1490 1445 size = lhp->lh_chunksize;
1491 1446 mutex_exit(&lhp->lh_lock);
1492 1447 }
1493 1448 logspace = clhp->clh_current;
1494 1449 clhp->clh_current += size;
1495 1450 clhp->clh_avail -= size;
1496 1451 bcopy(data, logspace, size);
1497 1452 mutex_exit(&clhp->clh_lock);
1498 1453 return (logspace);
1499 1454 }
1500 1455
1501 1456 #define KMEM_AUDIT(lp, cp, bcp) \
1502 1457 { \
1503 1458 kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp); \
1504 1459 _bcp->bc_timestamp = gethrtime(); \
1505 1460 _bcp->bc_thread = curthread; \
1506 1461 _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH); \
1507 1462 _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp)); \
1508 1463 }
1509 1464
1510 1465 static void
1511 1466 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1512 1467 kmem_slab_t *sp, void *addr)
1513 1468 {
1514 1469 kmem_bufctl_audit_t bca;
1515 1470
1516 1471 bzero(&bca, sizeof (kmem_bufctl_audit_t));
1517 1472 bca.bc_addr = addr;
1518 1473 bca.bc_slab = sp;
1519 1474 bca.bc_cache = cp;
1520 1475 KMEM_AUDIT(lp, cp, &bca);
1521 1476 }
1522 1477
1523 1478 /*
1524 1479 * Create a new slab for cache cp.
1525 1480 */
1526 1481 static kmem_slab_t *
1527 1482 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1528 1483 {
1529 1484 size_t slabsize = cp->cache_slabsize;
1530 1485 size_t chunksize = cp->cache_chunksize;
1531 1486 int cache_flags = cp->cache_flags;
1532 1487 size_t color, chunks;
1533 1488 char *buf, *slab;
1534 1489 kmem_slab_t *sp;
1535 1490 kmem_bufctl_t *bcp;
1536 1491 vmem_t *vmp = cp->cache_arena;
1537 1492
1538 1493 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1539 1494
1540 1495 color = cp->cache_color + cp->cache_align;
1541 1496 if (color > cp->cache_maxcolor)
1542 1497 color = cp->cache_mincolor;
1543 1498 cp->cache_color = color;
1544 1499
1545 1500 slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1546 1501
1547 1502 if (slab == NULL)
1548 1503 goto vmem_alloc_failure;
1549 1504
1550 1505 ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1551 1506
1552 1507 /*
1553 1508 * Reverify what was already checked in kmem_cache_set_move(), since the
1554 1509 * consolidator depends (for correctness) on slabs being initialized
1555 1510 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1556 1511 * clients to distinguish uninitialized memory from known objects).
1557 1512 */
1558 1513 ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1559 1514 if (!(cp->cache_cflags & KMC_NOTOUCH))
1560 1515 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1561 1516
1562 1517 if (cache_flags & KMF_HASH) {
1563 1518 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1564 1519 goto slab_alloc_failure;
1565 1520 chunks = (slabsize - color) / chunksize;
1566 1521 } else {
1567 1522 sp = KMEM_SLAB(cp, slab);
1568 1523 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1569 1524 }
1570 1525
1571 1526 sp->slab_cache = cp;
1572 1527 sp->slab_head = NULL;
1573 1528 sp->slab_refcnt = 0;
1574 1529 sp->slab_base = buf = slab + color;
1575 1530 sp->slab_chunks = chunks;
1576 1531 sp->slab_stuck_offset = (uint32_t)-1;
1577 1532 sp->slab_later_count = 0;
1578 1533 sp->slab_flags = 0;
1579 1534
1580 1535 ASSERT(chunks > 0);
1581 1536 while (chunks-- != 0) {
1582 1537 if (cache_flags & KMF_HASH) {
1583 1538 bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1584 1539 if (bcp == NULL)
1585 1540 goto bufctl_alloc_failure;
1586 1541 if (cache_flags & KMF_AUDIT) {
1587 1542 kmem_bufctl_audit_t *bcap =
1588 1543 (kmem_bufctl_audit_t *)bcp;
1589 1544 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1590 1545 bcap->bc_cache = cp;
1591 1546 }
1592 1547 bcp->bc_addr = buf;
1593 1548 bcp->bc_slab = sp;
1594 1549 } else {
1595 1550 bcp = KMEM_BUFCTL(cp, buf);
1596 1551 }
1597 1552 if (cache_flags & KMF_BUFTAG) {
1598 1553 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1599 1554 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1600 1555 btp->bt_bufctl = bcp;
1601 1556 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1602 1557 if (cache_flags & KMF_DEADBEEF) {
1603 1558 copy_pattern(KMEM_FREE_PATTERN, buf,
1604 1559 cp->cache_verify);
1605 1560 }
1606 1561 }
1607 1562 bcp->bc_next = sp->slab_head;
1608 1563 sp->slab_head = bcp;
1609 1564 buf += chunksize;
1610 1565 }
1611 1566
1612 1567 kmem_log_event(kmem_slab_log, cp, sp, slab);
1613 1568
1614 1569 return (sp);
1615 1570
1616 1571 bufctl_alloc_failure:
1617 1572
1618 1573 while ((bcp = sp->slab_head) != NULL) {
1619 1574 sp->slab_head = bcp->bc_next;
1620 1575 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1621 1576 }
1622 1577 kmem_cache_free(kmem_slab_cache, sp);
1623 1578
1624 1579 slab_alloc_failure:
1625 1580
1626 1581 vmem_free(vmp, slab, slabsize);
1627 1582
1628 1583 vmem_alloc_failure:
1629 1584
1630 1585 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1631 1586 atomic_inc_64(&cp->cache_alloc_fail);
1632 1587
1633 1588 return (NULL);
1634 1589 }
1635 1590
1636 1591 /*
1637 1592 * Destroy a slab.
1638 1593 */
1639 1594 static void
1640 1595 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1641 1596 {
1642 1597 vmem_t *vmp = cp->cache_arena;
1643 1598 void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1644 1599
1645 1600 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1646 1601 ASSERT(sp->slab_refcnt == 0);
1647 1602
1648 1603 if (cp->cache_flags & KMF_HASH) {
1649 1604 kmem_bufctl_t *bcp;
1650 1605 while ((bcp = sp->slab_head) != NULL) {
1651 1606 sp->slab_head = bcp->bc_next;
1652 1607 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1653 1608 }
1654 1609 kmem_cache_free(kmem_slab_cache, sp);
1655 1610 }
1656 1611 vmem_free(vmp, slab, cp->cache_slabsize);
1657 1612 }
1658 1613
1659 1614 static void *
1660 1615 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1661 1616 {
1662 1617 kmem_bufctl_t *bcp, **hash_bucket;
1663 1618 void *buf;
1664 1619 boolean_t new_slab = (sp->slab_refcnt == 0);
1665 1620
1666 1621 ASSERT(MUTEX_HELD(&cp->cache_lock));
1667 1622 /*
1668 1623 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1669 1624 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1670 1625 * slab is newly created.
1671 1626 */
1672 1627 ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1673 1628 (sp == avl_first(&cp->cache_partial_slabs))));
1674 1629 ASSERT(sp->slab_cache == cp);
1675 1630
1676 1631 cp->cache_slab_alloc++;
1677 1632 cp->cache_bufslab--;
1678 1633 sp->slab_refcnt++;
1679 1634
1680 1635 bcp = sp->slab_head;
1681 1636 sp->slab_head = bcp->bc_next;
1682 1637
1683 1638 if (cp->cache_flags & KMF_HASH) {
1684 1639 /*
1685 1640 * Add buffer to allocated-address hash table.
1686 1641 */
1687 1642 buf = bcp->bc_addr;
1688 1643 hash_bucket = KMEM_HASH(cp, buf);
1689 1644 bcp->bc_next = *hash_bucket;
1690 1645 *hash_bucket = bcp;
1691 1646 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1692 1647 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1693 1648 }
1694 1649 } else {
1695 1650 buf = KMEM_BUF(cp, bcp);
1696 1651 }
1697 1652
1698 1653 ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1699 1654
1700 1655 if (sp->slab_head == NULL) {
1701 1656 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1702 1657 if (new_slab) {
1703 1658 ASSERT(sp->slab_chunks == 1);
1704 1659 } else {
1705 1660 ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1706 1661 avl_remove(&cp->cache_partial_slabs, sp);
1707 1662 sp->slab_later_count = 0; /* clear history */
1708 1663 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1709 1664 sp->slab_stuck_offset = (uint32_t)-1;
1710 1665 }
1711 1666 list_insert_head(&cp->cache_complete_slabs, sp);
1712 1667 cp->cache_complete_slab_count++;
1713 1668 return (buf);
1714 1669 }
1715 1670
1716 1671 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1717 1672 /*
1718 1673 * Peek to see if the magazine layer is enabled before
1719 1674 * we prefill. We're not holding the cpu cache lock,
1720 1675 * so the peek could be wrong, but there's no harm in it.
1721 1676 */
1722 1677 if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1723 1678 (KMEM_CPU_CACHE(cp)->cc_magsize != 0)) {
1724 1679 kmem_slab_prefill(cp, sp);
1725 1680 return (buf);
1726 1681 }
1727 1682
1728 1683 if (new_slab) {
1729 1684 avl_add(&cp->cache_partial_slabs, sp);
1730 1685 return (buf);
1731 1686 }
1732 1687
1733 1688 /*
1734 1689 * The slab is now more allocated than it was, so the
1735 1690 * order remains unchanged.
1736 1691 */
1737 1692 ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1738 1693 return (buf);
1739 1694 }
1740 1695
1741 1696 /*
1742 1697 * Allocate a raw (unconstructed) buffer from cp's slab layer.
1743 1698 */
1744 1699 static void *
1745 1700 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1746 1701 {
1747 1702 kmem_slab_t *sp;
1748 1703 void *buf;
1749 1704 boolean_t test_destructor;
1750 1705
1751 1706 mutex_enter(&cp->cache_lock);
1752 1707 test_destructor = (cp->cache_slab_alloc == 0);
1753 1708 sp = avl_first(&cp->cache_partial_slabs);
1754 1709 if (sp == NULL) {
1755 1710 ASSERT(cp->cache_bufslab == 0);
1756 1711
1757 1712 /*
1758 1713 * The freelist is empty. Create a new slab.
1759 1714 */
1760 1715 mutex_exit(&cp->cache_lock);
1761 1716 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1762 1717 return (NULL);
1763 1718 }
1764 1719 mutex_enter(&cp->cache_lock);
1765 1720 cp->cache_slab_create++;
1766 1721 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1767 1722 cp->cache_bufmax = cp->cache_buftotal;
1768 1723 cp->cache_bufslab += sp->slab_chunks;
1769 1724 }
1770 1725
1771 1726 buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1772 1727 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1773 1728 (cp->cache_complete_slab_count +
1774 1729 avl_numnodes(&cp->cache_partial_slabs) +
1775 1730 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1776 1731 mutex_exit(&cp->cache_lock);
1777 1732
1778 1733 if (test_destructor && cp->cache_destructor != NULL) {
1779 1734 /*
1780 1735 * On the first kmem_slab_alloc(), assert that it is valid to
1781 1736 * call the destructor on a newly constructed object without any
1782 1737 * client involvement.
1783 1738 */
1784 1739 if ((cp->cache_constructor == NULL) ||
1785 1740 cp->cache_constructor(buf, cp->cache_private,
1786 1741 kmflag) == 0) {
1787 1742 cp->cache_destructor(buf, cp->cache_private);
1788 1743 }
1789 1744 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1790 1745 cp->cache_bufsize);
1791 1746 if (cp->cache_flags & KMF_DEADBEEF) {
1792 1747 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1793 1748 }
1794 1749 }
1795 1750
1796 1751 return (buf);
1797 1752 }
1798 1753
1799 1754 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1800 1755
1801 1756 /*
1802 1757 * Free a raw (unconstructed) buffer to cp's slab layer.
1803 1758 */
1804 1759 static void
1805 1760 kmem_slab_free(kmem_cache_t *cp, void *buf)
1806 1761 {
1807 1762 kmem_slab_t *sp;
1808 1763 kmem_bufctl_t *bcp, **prev_bcpp;
1809 1764
1810 1765 ASSERT(buf != NULL);
1811 1766
1812 1767 mutex_enter(&cp->cache_lock);
1813 1768 cp->cache_slab_free++;
1814 1769
1815 1770 if (cp->cache_flags & KMF_HASH) {
1816 1771 /*
1817 1772 * Look up buffer in allocated-address hash table.
1818 1773 */
1819 1774 prev_bcpp = KMEM_HASH(cp, buf);
1820 1775 while ((bcp = *prev_bcpp) != NULL) {
1821 1776 if (bcp->bc_addr == buf) {
1822 1777 *prev_bcpp = bcp->bc_next;
1823 1778 sp = bcp->bc_slab;
1824 1779 break;
1825 1780 }
1826 1781 cp->cache_lookup_depth++;
1827 1782 prev_bcpp = &bcp->bc_next;
1828 1783 }
1829 1784 } else {
1830 1785 bcp = KMEM_BUFCTL(cp, buf);
1831 1786 sp = KMEM_SLAB(cp, buf);
1832 1787 }
1833 1788
1834 1789 if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1835 1790 mutex_exit(&cp->cache_lock);
1836 1791 kmem_error(KMERR_BADADDR, cp, buf);
1837 1792 return;
1838 1793 }
1839 1794
1840 1795 if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1841 1796 /*
1842 1797 * If this is the buffer that prevented the consolidator from
1843 1798 * clearing the slab, we can reset the slab flags now that the
1844 1799 * buffer is freed. (It makes sense to do this in
1845 1800 * kmem_cache_free(), where the client gives up ownership of the
1846 1801 * buffer, but on the hot path the test is too expensive.)
1847 1802 */
1848 1803 kmem_slab_move_yes(cp, sp, buf);
1849 1804 }
1850 1805
1851 1806 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1852 1807 if (cp->cache_flags & KMF_CONTENTS)
1853 1808 ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1854 1809 kmem_log_enter(kmem_content_log, buf,
1855 1810 cp->cache_contents);
1856 1811 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1857 1812 }
1858 1813
1859 1814 bcp->bc_next = sp->slab_head;
1860 1815 sp->slab_head = bcp;
1861 1816
1862 1817 cp->cache_bufslab++;
1863 1818 ASSERT(sp->slab_refcnt >= 1);
1864 1819
1865 1820 if (--sp->slab_refcnt == 0) {
1866 1821 /*
1867 1822 * There are no outstanding allocations from this slab,
1868 1823 * so we can reclaim the memory.
1869 1824 */
1870 1825 if (sp->slab_chunks == 1) {
1871 1826 list_remove(&cp->cache_complete_slabs, sp);
1872 1827 cp->cache_complete_slab_count--;
1873 1828 } else {
1874 1829 avl_remove(&cp->cache_partial_slabs, sp);
1875 1830 }
1876 1831
1877 1832 cp->cache_buftotal -= sp->slab_chunks;
1878 1833 cp->cache_bufslab -= sp->slab_chunks;
1879 1834 /*
1880 1835 * Defer releasing the slab to the virtual memory subsystem
1881 1836 * while there is a pending move callback, since we guarantee
1882 1837 * that buffers passed to the move callback have only been
1883 1838 * touched by kmem or by the client itself. Since the memory
1884 1839 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1885 1840 * set at least one of the two lowest order bits, the client can
1886 1841 * test those bits in the move callback to determine whether or
1887 1842 * not it knows about the buffer (assuming that the client also
1888 1843 * sets one of those low order bits whenever it frees a buffer).
1889 1844 */
1890 1845 if (cp->cache_defrag == NULL ||
1891 1846 (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1892 1847 !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1893 1848 cp->cache_slab_destroy++;
1894 1849 mutex_exit(&cp->cache_lock);
1895 1850 kmem_slab_destroy(cp, sp);
1896 1851 } else {
1897 1852 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1898 1853 /*
1899 1854 * Slabs are inserted at both ends of the deadlist to
1900 1855 * distinguish between slabs freed while move callbacks
1901 1856 * are pending (list head) and a slab freed while the
1902 1857 * lock is dropped in kmem_move_buffers() (list tail) so
1903 1858 * that in both cases slab_destroy() is called from the
1904 1859 * right context.
1905 1860 */
1906 1861 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1907 1862 list_insert_tail(deadlist, sp);
1908 1863 } else {
1909 1864 list_insert_head(deadlist, sp);
1910 1865 }
1911 1866 cp->cache_defrag->kmd_deadcount++;
1912 1867 mutex_exit(&cp->cache_lock);
1913 1868 }
1914 1869 return;
|
↓ open down ↓ |
790 lines elided |
↑ open up ↑ |
1915 1870 }
1916 1871
1917 1872 if (bcp->bc_next == NULL) {
1918 1873 /* Transition the slab from completely allocated to partial. */
1919 1874 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1920 1875 ASSERT(sp->slab_chunks > 1);
1921 1876 list_remove(&cp->cache_complete_slabs, sp);
1922 1877 cp->cache_complete_slab_count--;
1923 1878 avl_add(&cp->cache_partial_slabs, sp);
1924 1879 } else {
1925 -#ifdef DEBUG
1926 - if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1927 - KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1928 - } else {
1929 - KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1930 - }
1931 -#else
1932 1880 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1933 -#endif
1934 1881 }
1935 1882
1936 1883 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1937 1884 (cp->cache_complete_slab_count +
1938 1885 avl_numnodes(&cp->cache_partial_slabs) +
1939 1886 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1940 1887 mutex_exit(&cp->cache_lock);
1941 1888 }
1942 1889
1943 1890 /*
1944 1891 * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1945 1892 */
1946 1893 static int
1947 1894 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1948 1895 caddr_t caller)
1949 1896 {
1950 1897 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1951 1898 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1952 1899 uint32_t mtbf;
1953 1900
1954 1901 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1955 1902 kmem_error(KMERR_BADBUFTAG, cp, buf);
1956 1903 return (-1);
1957 1904 }
1958 1905
1959 1906 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1960 1907
1961 1908 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1962 1909 kmem_error(KMERR_BADBUFCTL, cp, buf);
1963 1910 return (-1);
1964 1911 }
1965 1912
1966 1913 if (cp->cache_flags & KMF_DEADBEEF) {
1967 1914 if (!construct && (cp->cache_flags & KMF_LITE)) {
1968 1915 if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1969 1916 kmem_error(KMERR_MODIFIED, cp, buf);
1970 1917 return (-1);
1971 1918 }
1972 1919 if (cp->cache_constructor != NULL)
1973 1920 *(uint64_t *)buf = btp->bt_redzone;
1974 1921 else
1975 1922 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1976 1923 } else {
1977 1924 construct = 1;
1978 1925 if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1979 1926 KMEM_UNINITIALIZED_PATTERN, buf,
1980 1927 cp->cache_verify)) {
1981 1928 kmem_error(KMERR_MODIFIED, cp, buf);
1982 1929 return (-1);
1983 1930 }
1984 1931 }
1985 1932 }
1986 1933 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1987 1934
1988 1935 if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1989 1936 gethrtime() % mtbf == 0 &&
1990 1937 (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1991 1938 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1992 1939 if (!construct && cp->cache_destructor != NULL)
1993 1940 cp->cache_destructor(buf, cp->cache_private);
1994 1941 } else {
1995 1942 mtbf = 0;
1996 1943 }
1997 1944
1998 1945 if (mtbf || (construct && cp->cache_constructor != NULL &&
1999 1946 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
2000 1947 atomic_inc_64(&cp->cache_alloc_fail);
2001 1948 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2002 1949 if (cp->cache_flags & KMF_DEADBEEF)
2003 1950 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2004 1951 kmem_slab_free(cp, buf);
2005 1952 return (1);
2006 1953 }
2007 1954
2008 1955 if (cp->cache_flags & KMF_AUDIT) {
2009 1956 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2010 1957 }
2011 1958
2012 1959 if ((cp->cache_flags & KMF_LITE) &&
2013 1960 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2014 1961 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2015 1962 }
2016 1963
2017 1964 return (0);
2018 1965 }
2019 1966
2020 1967 static int
2021 1968 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2022 1969 {
2023 1970 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2024 1971 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2025 1972 kmem_slab_t *sp;
2026 1973
2027 1974 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2028 1975 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2029 1976 kmem_error(KMERR_DUPFREE, cp, buf);
2030 1977 return (-1);
2031 1978 }
2032 1979 sp = kmem_findslab(cp, buf);
2033 1980 if (sp == NULL || sp->slab_cache != cp)
2034 1981 kmem_error(KMERR_BADADDR, cp, buf);
2035 1982 else
2036 1983 kmem_error(KMERR_REDZONE, cp, buf);
2037 1984 return (-1);
2038 1985 }
2039 1986
2040 1987 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2041 1988
2042 1989 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2043 1990 kmem_error(KMERR_BADBUFCTL, cp, buf);
2044 1991 return (-1);
2045 1992 }
2046 1993
2047 1994 if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2048 1995 kmem_error(KMERR_REDZONE, cp, buf);
2049 1996 return (-1);
2050 1997 }
2051 1998
2052 1999 if (cp->cache_flags & KMF_AUDIT) {
2053 2000 if (cp->cache_flags & KMF_CONTENTS)
2054 2001 bcp->bc_contents = kmem_log_enter(kmem_content_log,
2055 2002 buf, cp->cache_contents);
2056 2003 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2057 2004 }
2058 2005
2059 2006 if ((cp->cache_flags & KMF_LITE) &&
2060 2007 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2061 2008 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2062 2009 }
2063 2010
2064 2011 if (cp->cache_flags & KMF_DEADBEEF) {
2065 2012 if (cp->cache_flags & KMF_LITE)
2066 2013 btp->bt_redzone = *(uint64_t *)buf;
2067 2014 else if (cp->cache_destructor != NULL)
2068 2015 cp->cache_destructor(buf, cp->cache_private);
2069 2016
2070 2017 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2071 2018 }
2072 2019
2073 2020 return (0);
2074 2021 }
2075 2022
2076 2023 /*
2077 2024 * Free each object in magazine mp to cp's slab layer, and free mp itself.
2078 2025 */
2079 2026 static void
2080 2027 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2081 2028 {
2082 2029 int round;
2083 2030
2084 2031 ASSERT(!list_link_active(&cp->cache_link) ||
2085 2032 taskq_member(kmem_taskq, curthread));
2086 2033
2087 2034 for (round = 0; round < nrounds; round++) {
2088 2035 void *buf = mp->mag_round[round];
2089 2036
2090 2037 if (cp->cache_flags & KMF_DEADBEEF) {
2091 2038 if (verify_pattern(KMEM_FREE_PATTERN, buf,
2092 2039 cp->cache_verify) != NULL) {
2093 2040 kmem_error(KMERR_MODIFIED, cp, buf);
2094 2041 continue;
2095 2042 }
2096 2043 if ((cp->cache_flags & KMF_LITE) &&
2097 2044 cp->cache_destructor != NULL) {
2098 2045 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2099 2046 *(uint64_t *)buf = btp->bt_redzone;
2100 2047 cp->cache_destructor(buf, cp->cache_private);
2101 2048 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2102 2049 }
2103 2050 } else if (cp->cache_destructor != NULL) {
2104 2051 cp->cache_destructor(buf, cp->cache_private);
2105 2052 }
2106 2053
2107 2054 kmem_slab_free(cp, buf);
2108 2055 }
2109 2056 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2110 2057 kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2111 2058 }
2112 2059
2113 2060 /*
2114 2061 * Allocate a magazine from the depot.
2115 2062 */
2116 2063 static kmem_magazine_t *
2117 2064 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2118 2065 {
2119 2066 kmem_magazine_t *mp;
2120 2067
2121 2068 /*
2122 2069 * If we can't get the depot lock without contention,
2123 2070 * update our contention count. We use the depot
2124 2071 * contention rate to determine whether we need to
2125 2072 * increase the magazine size for better scalability.
2126 2073 */
2127 2074 if (!mutex_tryenter(&cp->cache_depot_lock)) {
2128 2075 mutex_enter(&cp->cache_depot_lock);
2129 2076 cp->cache_depot_contention++;
2130 2077 }
2131 2078
2132 2079 if ((mp = mlp->ml_list) != NULL) {
2133 2080 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2134 2081 mlp->ml_list = mp->mag_next;
2135 2082 if (--mlp->ml_total < mlp->ml_min)
2136 2083 mlp->ml_min = mlp->ml_total;
2137 2084 mlp->ml_alloc++;
2138 2085 }
2139 2086
2140 2087 mutex_exit(&cp->cache_depot_lock);
2141 2088
2142 2089 return (mp);
2143 2090 }
2144 2091
2145 2092 /*
2146 2093 * Free a magazine to the depot.
2147 2094 */
2148 2095 static void
2149 2096 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2150 2097 {
2151 2098 mutex_enter(&cp->cache_depot_lock);
2152 2099 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2153 2100 mp->mag_next = mlp->ml_list;
2154 2101 mlp->ml_list = mp;
2155 2102 mlp->ml_total++;
2156 2103 mutex_exit(&cp->cache_depot_lock);
2157 2104 }
2158 2105
2159 2106 /*
2160 2107 * Update the working set statistics for cp's depot.
2161 2108 */
2162 2109 static void
2163 2110 kmem_depot_ws_update(kmem_cache_t *cp)
2164 2111 {
2165 2112 mutex_enter(&cp->cache_depot_lock);
2166 2113 cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2167 2114 cp->cache_full.ml_min = cp->cache_full.ml_total;
2168 2115 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2169 2116 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2170 2117 mutex_exit(&cp->cache_depot_lock);
2171 2118 }
2172 2119
2173 2120 /*
2174 2121 * Set the working set statistics for cp's depot to zero. (Everything is
2175 2122 * eligible for reaping.)
2176 2123 */
2177 2124 static void
2178 2125 kmem_depot_ws_zero(kmem_cache_t *cp)
2179 2126 {
2180 2127 mutex_enter(&cp->cache_depot_lock);
2181 2128 cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2182 2129 cp->cache_full.ml_min = cp->cache_full.ml_total;
2183 2130 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2184 2131 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2185 2132 mutex_exit(&cp->cache_depot_lock);
2186 2133 }
2187 2134
2188 2135 /*
2189 2136 * The number of bytes to reap before we call kpreempt(). The default (1MB)
2190 2137 * causes us to preempt reaping up to hundreds of times per second. Using a
2191 2138 * larger value (1GB) causes this to have virtually no effect.
2192 2139 */
2193 2140 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2194 2141
2195 2142 /*
2196 2143 * Reap all magazines that have fallen out of the depot's working set.
2197 2144 */
2198 2145 static void
2199 2146 kmem_depot_ws_reap(kmem_cache_t *cp)
2200 2147 {
2201 2148 size_t bytes = 0;
2202 2149 long reap;
2203 2150 kmem_magazine_t *mp;
2204 2151
2205 2152 ASSERT(!list_link_active(&cp->cache_link) ||
2206 2153 taskq_member(kmem_taskq, curthread));
2207 2154
2208 2155 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2209 2156 while (reap-- &&
2210 2157 (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2211 2158 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2212 2159 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2213 2160 if (bytes > kmem_reap_preempt_bytes) {
2214 2161 kpreempt(KPREEMPT_SYNC);
2215 2162 bytes = 0;
2216 2163 }
2217 2164 }
2218 2165
2219 2166 reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2220 2167 while (reap-- &&
2221 2168 (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2222 2169 kmem_magazine_destroy(cp, mp, 0);
2223 2170 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2224 2171 if (bytes > kmem_reap_preempt_bytes) {
2225 2172 kpreempt(KPREEMPT_SYNC);
2226 2173 bytes = 0;
2227 2174 }
2228 2175 }
2229 2176 }
2230 2177
2231 2178 static void
2232 2179 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2233 2180 {
2234 2181 ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2235 2182 (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2236 2183 ASSERT(ccp->cc_magsize > 0);
2237 2184
2238 2185 ccp->cc_ploaded = ccp->cc_loaded;
2239 2186 ccp->cc_prounds = ccp->cc_rounds;
2240 2187 ccp->cc_loaded = mp;
2241 2188 ccp->cc_rounds = rounds;
2242 2189 }
2243 2190
2244 2191 /*
2245 2192 * Intercept kmem alloc/free calls during crash dump in order to avoid
2246 2193 * changing kmem state while memory is being saved to the dump device.
2247 2194 * Otherwise, ::kmem_verify will report "corrupt buffers". Note that
2248 2195 * there are no locks because only one CPU calls kmem during a crash
2249 2196 * dump. To enable this feature, first create the associated vmem
2250 2197 * arena with VMC_DUMPSAFE.
2251 2198 */
2252 2199 static void *kmem_dump_start; /* start of pre-reserved heap */
2253 2200 static void *kmem_dump_end; /* end of heap area */
2254 2201 static void *kmem_dump_curr; /* current free heap pointer */
2255 2202 static size_t kmem_dump_size; /* size of heap area */
2256 2203
2257 2204 /* append to each buf created in the pre-reserved heap */
2258 2205 typedef struct kmem_dumpctl {
2259 2206 void *kdc_next; /* cache dump free list linkage */
2260 2207 } kmem_dumpctl_t;
2261 2208
2262 2209 #define KMEM_DUMPCTL(cp, buf) \
2263 2210 ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2264 2211 sizeof (void *)))
2265 2212
2266 2213 /* Keep some simple stats. */
2267 2214 #define KMEM_DUMP_LOGS (100)
2268 2215
2269 2216 typedef struct kmem_dump_log {
2270 2217 kmem_cache_t *kdl_cache;
2271 2218 uint_t kdl_allocs; /* # of dump allocations */
2272 2219 uint_t kdl_frees; /* # of dump frees */
2273 2220 uint_t kdl_alloc_fails; /* # of allocation failures */
2274 2221 uint_t kdl_free_nondump; /* # of non-dump frees */
2275 2222 uint_t kdl_unsafe; /* cache was used, but unsafe */
2276 2223 } kmem_dump_log_t;
2277 2224
2278 2225 static kmem_dump_log_t *kmem_dump_log;
2279 2226 static int kmem_dump_log_idx;
2280 2227
2281 2228 #define KDI_LOG(cp, stat) { \
2282 2229 kmem_dump_log_t *kdl; \
2283 2230 if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) { \
2284 2231 kdl->stat++; \
2285 2232 } else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) { \
2286 2233 kdl = &kmem_dump_log[kmem_dump_log_idx++]; \
2287 2234 kdl->stat++; \
2288 2235 kdl->kdl_cache = (cp); \
2289 2236 (cp)->cache_dumplog = kdl; \
2290 2237 } \
2291 2238 }
2292 2239
2293 2240 /* set non zero for full report */
2294 2241 uint_t kmem_dump_verbose = 0;
2295 2242
2296 2243 /* stats for overize heap */
2297 2244 uint_t kmem_dump_oversize_allocs = 0;
2298 2245 uint_t kmem_dump_oversize_max = 0;
2299 2246
2300 2247 static void
2301 2248 kmem_dumppr(char **pp, char *e, const char *format, ...)
2302 2249 {
2303 2250 char *p = *pp;
2304 2251
2305 2252 if (p < e) {
2306 2253 int n;
2307 2254 va_list ap;
2308 2255
2309 2256 va_start(ap, format);
2310 2257 n = vsnprintf(p, e - p, format, ap);
2311 2258 va_end(ap);
2312 2259 *pp = p + n;
2313 2260 }
2314 2261 }
2315 2262
2316 2263 /*
2317 2264 * Called when dumpadm(1M) configures dump parameters.
2318 2265 */
2319 2266 void
2320 2267 kmem_dump_init(size_t size)
2321 2268 {
2322 2269 if (kmem_dump_start != NULL)
2323 2270 kmem_free(kmem_dump_start, kmem_dump_size);
2324 2271
2325 2272 if (kmem_dump_log == NULL)
2326 2273 kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2327 2274 sizeof (kmem_dump_log_t), KM_SLEEP);
2328 2275
2329 2276 kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2330 2277
2331 2278 if (kmem_dump_start != NULL) {
2332 2279 kmem_dump_size = size;
2333 2280 kmem_dump_curr = kmem_dump_start;
2334 2281 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2335 2282 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2336 2283 } else {
2337 2284 kmem_dump_size = 0;
2338 2285 kmem_dump_curr = NULL;
2339 2286 kmem_dump_end = NULL;
2340 2287 }
2341 2288 }
2342 2289
2343 2290 /*
2344 2291 * Set flag for each kmem_cache_t if is safe to use alternate dump
2345 2292 * memory. Called just before panic crash dump starts. Set the flag
2346 2293 * for the calling CPU.
2347 2294 */
2348 2295 void
2349 2296 kmem_dump_begin(void)
2350 2297 {
2351 2298 ASSERT(panicstr != NULL);
2352 2299 if (kmem_dump_start != NULL) {
2353 2300 kmem_cache_t *cp;
2354 2301
2355 2302 for (cp = list_head(&kmem_caches); cp != NULL;
2356 2303 cp = list_next(&kmem_caches, cp)) {
2357 2304 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2358 2305
2359 2306 if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2360 2307 cp->cache_flags |= KMF_DUMPDIVERT;
2361 2308 ccp->cc_flags |= KMF_DUMPDIVERT;
2362 2309 ccp->cc_dump_rounds = ccp->cc_rounds;
2363 2310 ccp->cc_dump_prounds = ccp->cc_prounds;
2364 2311 ccp->cc_rounds = ccp->cc_prounds = -1;
2365 2312 } else {
2366 2313 cp->cache_flags |= KMF_DUMPUNSAFE;
2367 2314 ccp->cc_flags |= KMF_DUMPUNSAFE;
2368 2315 }
2369 2316 }
2370 2317 }
2371 2318 }
2372 2319
2373 2320 /*
2374 2321 * finished dump intercept
2375 2322 * print any warnings on the console
2376 2323 * return verbose information to dumpsys() in the given buffer
2377 2324 */
2378 2325 size_t
2379 2326 kmem_dump_finish(char *buf, size_t size)
2380 2327 {
2381 2328 int kdi_idx;
2382 2329 int kdi_end = kmem_dump_log_idx;
2383 2330 int percent = 0;
2384 2331 int header = 0;
2385 2332 int warn = 0;
2386 2333 size_t used;
2387 2334 kmem_cache_t *cp;
2388 2335 kmem_dump_log_t *kdl;
2389 2336 char *e = buf + size;
2390 2337 char *p = buf;
2391 2338
2392 2339 if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2393 2340 return (0);
2394 2341
2395 2342 used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2396 2343 percent = (used * 100) / kmem_dump_size;
2397 2344
2398 2345 kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2399 2346 kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2400 2347 kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2401 2348 kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2402 2349 kmem_dump_oversize_allocs);
2403 2350 kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2404 2351 kmem_dump_oversize_max);
2405 2352
2406 2353 for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2407 2354 kdl = &kmem_dump_log[kdi_idx];
2408 2355 cp = kdl->kdl_cache;
2409 2356 if (cp == NULL)
2410 2357 break;
2411 2358 if (kdl->kdl_alloc_fails)
2412 2359 ++warn;
2413 2360 if (header == 0) {
2414 2361 kmem_dumppr(&p, e,
2415 2362 "Cache Name,Allocs,Frees,Alloc Fails,"
2416 2363 "Nondump Frees,Unsafe Allocs/Frees\n");
2417 2364 header = 1;
2418 2365 }
2419 2366 kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2420 2367 cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2421 2368 kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2422 2369 kdl->kdl_unsafe);
2423 2370 }
2424 2371
2425 2372 /* return buffer size used */
2426 2373 if (p < e)
2427 2374 bzero(p, e - p);
2428 2375 return (p - buf);
2429 2376 }
2430 2377
2431 2378 /*
2432 2379 * Allocate a constructed object from alternate dump memory.
2433 2380 */
2434 2381 void *
2435 2382 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2436 2383 {
2437 2384 void *buf;
2438 2385 void *curr;
2439 2386 char *bufend;
2440 2387
2441 2388 /* return a constructed object */
2442 2389 if ((buf = cp->cache_dumpfreelist) != NULL) {
2443 2390 cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2444 2391 KDI_LOG(cp, kdl_allocs);
2445 2392 return (buf);
2446 2393 }
2447 2394
2448 2395 /* create a new constructed object */
2449 2396 curr = kmem_dump_curr;
2450 2397 buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2451 2398 bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2452 2399
2453 2400 /* hat layer objects cannot cross a page boundary */
2454 2401 if (cp->cache_align < PAGESIZE) {
2455 2402 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2456 2403 if (bufend > page) {
2457 2404 bufend += page - (char *)buf;
2458 2405 buf = (void *)page;
2459 2406 }
2460 2407 }
2461 2408
2462 2409 /* fall back to normal alloc if reserved area is used up */
2463 2410 if (bufend > (char *)kmem_dump_end) {
2464 2411 kmem_dump_curr = kmem_dump_end;
2465 2412 KDI_LOG(cp, kdl_alloc_fails);
2466 2413 return (NULL);
2467 2414 }
2468 2415
2469 2416 /*
2470 2417 * Must advance curr pointer before calling a constructor that
2471 2418 * may also allocate memory.
2472 2419 */
2473 2420 kmem_dump_curr = bufend;
2474 2421
2475 2422 /* run constructor */
2476 2423 if (cp->cache_constructor != NULL &&
2477 2424 cp->cache_constructor(buf, cp->cache_private, kmflag)
2478 2425 != 0) {
2479 2426 #ifdef DEBUG
2480 2427 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2481 2428 cp->cache_name, (void *)cp);
2482 2429 #endif
2483 2430 /* reset curr pointer iff no allocs were done */
2484 2431 if (kmem_dump_curr == bufend)
2485 2432 kmem_dump_curr = curr;
2486 2433
2487 2434 /* fall back to normal alloc if the constructor fails */
2488 2435 KDI_LOG(cp, kdl_alloc_fails);
2489 2436 return (NULL);
2490 2437 }
2491 2438
2492 2439 KDI_LOG(cp, kdl_allocs);
2493 2440 return (buf);
2494 2441 }
2495 2442
2496 2443 /*
2497 2444 * Free a constructed object in alternate dump memory.
2498 2445 */
2499 2446 int
2500 2447 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2501 2448 {
2502 2449 /* save constructed buffers for next time */
2503 2450 if ((char *)buf >= (char *)kmem_dump_start &&
2504 2451 (char *)buf < (char *)kmem_dump_end) {
2505 2452 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2506 2453 cp->cache_dumpfreelist = buf;
2507 2454 KDI_LOG(cp, kdl_frees);
2508 2455 return (0);
2509 2456 }
2510 2457
2511 2458 /* count all non-dump buf frees */
2512 2459 KDI_LOG(cp, kdl_free_nondump);
2513 2460
2514 2461 /* just drop buffers that were allocated before dump started */
2515 2462 if (kmem_dump_curr < kmem_dump_end)
2516 2463 return (0);
2517 2464
2518 2465 /* fall back to normal free if reserved area is used up */
2519 2466 return (1);
2520 2467 }
2521 2468
2522 2469 /*
2523 2470 * Allocate a constructed object from cache cp.
2524 2471 */
2525 2472 void *
2526 2473 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2527 2474 {
2528 2475 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2529 2476 kmem_magazine_t *fmp;
2530 2477 void *buf;
2531 2478
2532 2479 mutex_enter(&ccp->cc_lock);
2533 2480 for (;;) {
2534 2481 /*
2535 2482 * If there's an object available in the current CPU's
2536 2483 * loaded magazine, just take it and return.
2537 2484 */
2538 2485 if (ccp->cc_rounds > 0) {
2539 2486 buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2540 2487 ccp->cc_alloc++;
2541 2488 mutex_exit(&ccp->cc_lock);
2542 2489 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2543 2490 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2544 2491 ASSERT(!(ccp->cc_flags &
2545 2492 KMF_DUMPDIVERT));
2546 2493 KDI_LOG(cp, kdl_unsafe);
2547 2494 }
2548 2495 if ((ccp->cc_flags & KMF_BUFTAG) &&
2549 2496 kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2550 2497 caller()) != 0) {
2551 2498 if (kmflag & KM_NOSLEEP)
2552 2499 return (NULL);
2553 2500 mutex_enter(&ccp->cc_lock);
2554 2501 continue;
2555 2502 }
2556 2503 }
2557 2504 return (buf);
2558 2505 }
2559 2506
2560 2507 /*
2561 2508 * The loaded magazine is empty. If the previously loaded
2562 2509 * magazine was full, exchange them and try again.
2563 2510 */
2564 2511 if (ccp->cc_prounds > 0) {
2565 2512 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2566 2513 continue;
2567 2514 }
2568 2515
2569 2516 /*
2570 2517 * Return an alternate buffer at dump time to preserve
2571 2518 * the heap.
2572 2519 */
2573 2520 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2574 2521 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2575 2522 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2576 2523 /* log it so that we can warn about it */
2577 2524 KDI_LOG(cp, kdl_unsafe);
2578 2525 } else {
2579 2526 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2580 2527 NULL) {
2581 2528 mutex_exit(&ccp->cc_lock);
2582 2529 return (buf);
2583 2530 }
2584 2531 break; /* fall back to slab layer */
2585 2532 }
2586 2533 }
2587 2534
2588 2535 /*
2589 2536 * If the magazine layer is disabled, break out now.
2590 2537 */
2591 2538 if (ccp->cc_magsize == 0)
2592 2539 break;
2593 2540
2594 2541 /*
2595 2542 * Try to get a full magazine from the depot.
2596 2543 */
2597 2544 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2598 2545 if (fmp != NULL) {
2599 2546 if (ccp->cc_ploaded != NULL)
2600 2547 kmem_depot_free(cp, &cp->cache_empty,
2601 2548 ccp->cc_ploaded);
2602 2549 kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2603 2550 continue;
2604 2551 }
2605 2552
2606 2553 /*
2607 2554 * There are no full magazines in the depot,
2608 2555 * so fall through to the slab layer.
2609 2556 */
2610 2557 break;
2611 2558 }
2612 2559 mutex_exit(&ccp->cc_lock);
2613 2560
2614 2561 /*
2615 2562 * We couldn't allocate a constructed object from the magazine layer,
2616 2563 * so get a raw buffer from the slab layer and apply its constructor.
2617 2564 */
2618 2565 buf = kmem_slab_alloc(cp, kmflag);
2619 2566
2620 2567 if (buf == NULL)
2621 2568 return (NULL);
2622 2569
2623 2570 if (cp->cache_flags & KMF_BUFTAG) {
2624 2571 /*
2625 2572 * Make kmem_cache_alloc_debug() apply the constructor for us.
2626 2573 */
2627 2574 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2628 2575 if (rc != 0) {
2629 2576 if (kmflag & KM_NOSLEEP)
2630 2577 return (NULL);
2631 2578 /*
2632 2579 * kmem_cache_alloc_debug() detected corruption
2633 2580 * but didn't panic (kmem_panic <= 0). We should not be
2634 2581 * here because the constructor failed (indicated by a
2635 2582 * return code of 1). Try again.
2636 2583 */
2637 2584 ASSERT(rc == -1);
2638 2585 return (kmem_cache_alloc(cp, kmflag));
2639 2586 }
2640 2587 return (buf);
2641 2588 }
2642 2589
2643 2590 if (cp->cache_constructor != NULL &&
2644 2591 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2645 2592 atomic_inc_64(&cp->cache_alloc_fail);
2646 2593 kmem_slab_free(cp, buf);
2647 2594 return (NULL);
2648 2595 }
2649 2596
2650 2597 return (buf);
2651 2598 }
2652 2599
2653 2600 /*
2654 2601 * The freed argument tells whether or not kmem_cache_free_debug() has already
2655 2602 * been called so that we can avoid the duplicate free error. For example, a
2656 2603 * buffer on a magazine has already been freed by the client but is still
2657 2604 * constructed.
2658 2605 */
2659 2606 static void
2660 2607 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2661 2608 {
2662 2609 if (!freed && (cp->cache_flags & KMF_BUFTAG))
2663 2610 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2664 2611 return;
2665 2612
2666 2613 /*
2667 2614 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2668 2615 * kmem_cache_free_debug() will have already applied the destructor.
2669 2616 */
2670 2617 if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2671 2618 cp->cache_destructor != NULL) {
2672 2619 if (cp->cache_flags & KMF_DEADBEEF) { /* KMF_LITE implied */
2673 2620 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2674 2621 *(uint64_t *)buf = btp->bt_redzone;
2675 2622 cp->cache_destructor(buf, cp->cache_private);
2676 2623 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2677 2624 } else {
2678 2625 cp->cache_destructor(buf, cp->cache_private);
2679 2626 }
2680 2627 }
2681 2628
2682 2629 kmem_slab_free(cp, buf);
2683 2630 }
2684 2631
2685 2632 /*
2686 2633 * Used when there's no room to free a buffer to the per-CPU cache.
2687 2634 * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2688 2635 * caller should try freeing to the per-CPU cache again.
2689 2636 * Note that we don't directly install the magazine in the cpu cache,
2690 2637 * since its state may have changed wildly while the lock was dropped.
2691 2638 */
2692 2639 static int
2693 2640 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2694 2641 {
2695 2642 kmem_magazine_t *emp;
2696 2643 kmem_magtype_t *mtp;
2697 2644
2698 2645 ASSERT(MUTEX_HELD(&ccp->cc_lock));
2699 2646 ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2700 2647 ((uint_t)ccp->cc_rounds == -1)) &&
2701 2648 ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2702 2649 ((uint_t)ccp->cc_prounds == -1)));
2703 2650
2704 2651 emp = kmem_depot_alloc(cp, &cp->cache_empty);
2705 2652 if (emp != NULL) {
2706 2653 if (ccp->cc_ploaded != NULL)
2707 2654 kmem_depot_free(cp, &cp->cache_full,
2708 2655 ccp->cc_ploaded);
2709 2656 kmem_cpu_reload(ccp, emp, 0);
2710 2657 return (1);
2711 2658 }
2712 2659 /*
2713 2660 * There are no empty magazines in the depot,
2714 2661 * so try to allocate a new one. We must drop all locks
2715 2662 * across kmem_cache_alloc() because lower layers may
2716 2663 * attempt to allocate from this cache.
2717 2664 */
2718 2665 mtp = cp->cache_magtype;
2719 2666 mutex_exit(&ccp->cc_lock);
2720 2667 emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2721 2668 mutex_enter(&ccp->cc_lock);
2722 2669
2723 2670 if (emp != NULL) {
2724 2671 /*
2725 2672 * We successfully allocated an empty magazine.
2726 2673 * However, we had to drop ccp->cc_lock to do it,
2727 2674 * so the cache's magazine size may have changed.
2728 2675 * If so, free the magazine and try again.
2729 2676 */
2730 2677 if (ccp->cc_magsize != mtp->mt_magsize) {
2731 2678 mutex_exit(&ccp->cc_lock);
2732 2679 kmem_cache_free(mtp->mt_cache, emp);
2733 2680 mutex_enter(&ccp->cc_lock);
2734 2681 return (1);
2735 2682 }
2736 2683
2737 2684 /*
2738 2685 * We got a magazine of the right size. Add it to
2739 2686 * the depot and try the whole dance again.
2740 2687 */
2741 2688 kmem_depot_free(cp, &cp->cache_empty, emp);
2742 2689 return (1);
2743 2690 }
2744 2691
2745 2692 /*
2746 2693 * We couldn't allocate an empty magazine,
2747 2694 * so fall through to the slab layer.
2748 2695 */
2749 2696 return (0);
2750 2697 }
2751 2698
2752 2699 /*
2753 2700 * Free a constructed object to cache cp.
2754 2701 */
2755 2702 void
2756 2703 kmem_cache_free(kmem_cache_t *cp, void *buf)
2757 2704 {
2758 2705 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2759 2706
2760 2707 /*
2761 2708 * The client must not free either of the buffers passed to the move
2762 2709 * callback function.
2763 2710 */
2764 2711 ASSERT(cp->cache_defrag == NULL ||
2765 2712 cp->cache_defrag->kmd_thread != curthread ||
2766 2713 (buf != cp->cache_defrag->kmd_from_buf &&
2767 2714 buf != cp->cache_defrag->kmd_to_buf));
2768 2715
2769 2716 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2770 2717 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2771 2718 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2772 2719 /* log it so that we can warn about it */
2773 2720 KDI_LOG(cp, kdl_unsafe);
2774 2721 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2775 2722 return;
2776 2723 }
2777 2724 if (ccp->cc_flags & KMF_BUFTAG) {
2778 2725 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2779 2726 return;
2780 2727 }
2781 2728 }
2782 2729
2783 2730 mutex_enter(&ccp->cc_lock);
2784 2731 /*
2785 2732 * Any changes to this logic should be reflected in kmem_slab_prefill()
2786 2733 */
2787 2734 for (;;) {
2788 2735 /*
2789 2736 * If there's a slot available in the current CPU's
2790 2737 * loaded magazine, just put the object there and return.
2791 2738 */
2792 2739 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2793 2740 ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2794 2741 ccp->cc_free++;
2795 2742 mutex_exit(&ccp->cc_lock);
2796 2743 return;
2797 2744 }
2798 2745
2799 2746 /*
2800 2747 * The loaded magazine is full. If the previously loaded
2801 2748 * magazine was empty, exchange them and try again.
2802 2749 */
2803 2750 if (ccp->cc_prounds == 0) {
2804 2751 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2805 2752 continue;
2806 2753 }
2807 2754
2808 2755 /*
2809 2756 * If the magazine layer is disabled, break out now.
2810 2757 */
2811 2758 if (ccp->cc_magsize == 0)
2812 2759 break;
2813 2760
2814 2761 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2815 2762 /*
2816 2763 * We couldn't free our constructed object to the
2817 2764 * magazine layer, so apply its destructor and free it
2818 2765 * to the slab layer.
2819 2766 */
2820 2767 break;
2821 2768 }
2822 2769 }
2823 2770 mutex_exit(&ccp->cc_lock);
2824 2771 kmem_slab_free_constructed(cp, buf, B_TRUE);
2825 2772 }
2826 2773
2827 2774 static void
2828 2775 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2829 2776 {
2830 2777 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2831 2778 int cache_flags = cp->cache_flags;
2832 2779
2833 2780 kmem_bufctl_t *next, *head;
2834 2781 size_t nbufs;
2835 2782
2836 2783 /*
2837 2784 * Completely allocate the newly created slab and put the pre-allocated
2838 2785 * buffers in magazines. Any of the buffers that cannot be put in
2839 2786 * magazines must be returned to the slab.
2840 2787 */
2841 2788 ASSERT(MUTEX_HELD(&cp->cache_lock));
2842 2789 ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2843 2790 ASSERT(cp->cache_constructor == NULL);
2844 2791 ASSERT(sp->slab_cache == cp);
2845 2792 ASSERT(sp->slab_refcnt == 1);
2846 2793 ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2847 2794 ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2848 2795
2849 2796 head = sp->slab_head;
2850 2797 nbufs = (sp->slab_chunks - sp->slab_refcnt);
2851 2798 sp->slab_head = NULL;
2852 2799 sp->slab_refcnt += nbufs;
2853 2800 cp->cache_bufslab -= nbufs;
2854 2801 cp->cache_slab_alloc += nbufs;
2855 2802 list_insert_head(&cp->cache_complete_slabs, sp);
2856 2803 cp->cache_complete_slab_count++;
2857 2804 mutex_exit(&cp->cache_lock);
2858 2805 mutex_enter(&ccp->cc_lock);
2859 2806
2860 2807 while (head != NULL) {
2861 2808 void *buf = KMEM_BUF(cp, head);
2862 2809 /*
2863 2810 * If there's a slot available in the current CPU's
2864 2811 * loaded magazine, just put the object there and
2865 2812 * continue.
2866 2813 */
2867 2814 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2868 2815 ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2869 2816 buf;
2870 2817 ccp->cc_free++;
2871 2818 nbufs--;
2872 2819 head = head->bc_next;
2873 2820 continue;
2874 2821 }
2875 2822
2876 2823 /*
2877 2824 * The loaded magazine is full. If the previously
2878 2825 * loaded magazine was empty, exchange them and try
2879 2826 * again.
2880 2827 */
2881 2828 if (ccp->cc_prounds == 0) {
2882 2829 kmem_cpu_reload(ccp, ccp->cc_ploaded,
2883 2830 ccp->cc_prounds);
2884 2831 continue;
2885 2832 }
2886 2833
2887 2834 /*
2888 2835 * If the magazine layer is disabled, break out now.
2889 2836 */
2890 2837
2891 2838 if (ccp->cc_magsize == 0) {
2892 2839 break;
2893 2840 }
2894 2841
2895 2842 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2896 2843 break;
2897 2844 }
2898 2845 mutex_exit(&ccp->cc_lock);
2899 2846 if (nbufs != 0) {
2900 2847 ASSERT(head != NULL);
2901 2848
2902 2849 /*
2903 2850 * If there was a failure, return remaining objects to
2904 2851 * the slab
2905 2852 */
2906 2853 while (head != NULL) {
2907 2854 ASSERT(nbufs != 0);
2908 2855 next = head->bc_next;
2909 2856 head->bc_next = NULL;
2910 2857 kmem_slab_free(cp, KMEM_BUF(cp, head));
2911 2858 head = next;
2912 2859 nbufs--;
2913 2860 }
2914 2861 }
2915 2862 ASSERT(head == NULL);
2916 2863 ASSERT(nbufs == 0);
2917 2864 mutex_enter(&cp->cache_lock);
2918 2865 }
2919 2866
2920 2867 void *
2921 2868 kmem_zalloc(size_t size, int kmflag)
2922 2869 {
2923 2870 size_t index;
2924 2871 void *buf;
2925 2872
2926 2873 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2927 2874 kmem_cache_t *cp = kmem_alloc_table[index];
2928 2875 buf = kmem_cache_alloc(cp, kmflag);
2929 2876 if (buf != NULL) {
2930 2877 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2931 2878 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2932 2879 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2933 2880 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2934 2881
2935 2882 if (cp->cache_flags & KMF_LITE) {
2936 2883 KMEM_BUFTAG_LITE_ENTER(btp,
2937 2884 kmem_lite_count, caller());
2938 2885 }
2939 2886 }
2940 2887 bzero(buf, size);
2941 2888 }
2942 2889 } else {
2943 2890 buf = kmem_alloc(size, kmflag);
2944 2891 if (buf != NULL)
2945 2892 bzero(buf, size);
2946 2893 }
2947 2894 return (buf);
2948 2895 }
2949 2896
2950 2897 void *
2951 2898 kmem_alloc(size_t size, int kmflag)
2952 2899 {
2953 2900 size_t index;
2954 2901 kmem_cache_t *cp;
2955 2902 void *buf;
2956 2903
2957 2904 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2958 2905 cp = kmem_alloc_table[index];
2959 2906 /* fall through to kmem_cache_alloc() */
2960 2907
2961 2908 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2962 2909 kmem_big_alloc_table_max) {
2963 2910 cp = kmem_big_alloc_table[index];
2964 2911 /* fall through to kmem_cache_alloc() */
2965 2912
2966 2913 } else {
2967 2914 if (size == 0)
2968 2915 return (NULL);
2969 2916
2970 2917 buf = vmem_alloc(kmem_oversize_arena, size,
2971 2918 kmflag & KM_VMFLAGS);
2972 2919 if (buf == NULL)
2973 2920 kmem_log_event(kmem_failure_log, NULL, NULL,
2974 2921 (void *)size);
2975 2922 else if (KMEM_DUMP(kmem_slab_cache)) {
2976 2923 /* stats for dump intercept */
2977 2924 kmem_dump_oversize_allocs++;
2978 2925 if (size > kmem_dump_oversize_max)
2979 2926 kmem_dump_oversize_max = size;
2980 2927 }
2981 2928 return (buf);
2982 2929 }
2983 2930
2984 2931 buf = kmem_cache_alloc(cp, kmflag);
2985 2932 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2986 2933 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2987 2934 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2988 2935 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2989 2936
2990 2937 if (cp->cache_flags & KMF_LITE) {
2991 2938 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2992 2939 }
2993 2940 }
2994 2941 return (buf);
2995 2942 }
2996 2943
2997 2944 void
2998 2945 kmem_free(void *buf, size_t size)
2999 2946 {
3000 2947 size_t index;
3001 2948 kmem_cache_t *cp;
3002 2949
3003 2950 if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
3004 2951 cp = kmem_alloc_table[index];
3005 2952 /* fall through to kmem_cache_free() */
3006 2953
3007 2954 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
3008 2955 kmem_big_alloc_table_max) {
3009 2956 cp = kmem_big_alloc_table[index];
3010 2957 /* fall through to kmem_cache_free() */
3011 2958
3012 2959 } else {
3013 2960 EQUIV(buf == NULL, size == 0);
3014 2961 if (buf == NULL && size == 0)
3015 2962 return;
3016 2963 vmem_free(kmem_oversize_arena, buf, size);
3017 2964 return;
3018 2965 }
3019 2966
3020 2967 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
3021 2968 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
3022 2969 uint32_t *ip = (uint32_t *)btp;
3023 2970 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3024 2971 if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3025 2972 kmem_error(KMERR_DUPFREE, cp, buf);
3026 2973 return;
3027 2974 }
3028 2975 if (KMEM_SIZE_VALID(ip[1])) {
3029 2976 ip[0] = KMEM_SIZE_ENCODE(size);
3030 2977 kmem_error(KMERR_BADSIZE, cp, buf);
3031 2978 } else {
3032 2979 kmem_error(KMERR_REDZONE, cp, buf);
3033 2980 }
3034 2981 return;
3035 2982 }
3036 2983 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3037 2984 kmem_error(KMERR_REDZONE, cp, buf);
3038 2985 return;
3039 2986 }
3040 2987 btp->bt_redzone = KMEM_REDZONE_PATTERN;
3041 2988 if (cp->cache_flags & KMF_LITE) {
3042 2989 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3043 2990 caller());
3044 2991 }
3045 2992 }
3046 2993 kmem_cache_free(cp, buf);
3047 2994 }
3048 2995
3049 2996 void *
3050 2997 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3051 2998 {
3052 2999 size_t realsize = size + vmp->vm_quantum;
3053 3000 void *addr;
3054 3001
3055 3002 /*
3056 3003 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3057 3004 * vm_quantum will cause integer wraparound. Check for this, and
3058 3005 * blow off the firewall page in this case. Note that such a
3059 3006 * giant allocation (the entire kernel address space) can never
3060 3007 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3061 3008 * or sleep forever (VM_SLEEP). Thus, there is no need for a
3062 3009 * corresponding check in kmem_firewall_va_free().
3063 3010 */
3064 3011 if (realsize < size)
3065 3012 realsize = size;
3066 3013
3067 3014 /*
3068 3015 * While boot still owns resource management, make sure that this
3069 3016 * redzone virtual address allocation is properly accounted for in
3070 3017 * OBPs "virtual-memory" "available" lists because we're
3071 3018 * effectively claiming them for a red zone. If we don't do this,
3072 3019 * the available lists become too fragmented and too large for the
3073 3020 * current boot/kernel memory list interface.
3074 3021 */
3075 3022 addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3076 3023
3077 3024 if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3078 3025 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3079 3026
3080 3027 return (addr);
3081 3028 }
3082 3029
3083 3030 void
3084 3031 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3085 3032 {
3086 3033 ASSERT((kvseg.s_base == NULL ?
3087 3034 va_to_pfn((char *)addr + size) :
3088 3035 hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3089 3036
3090 3037 vmem_free(vmp, addr, size + vmp->vm_quantum);
3091 3038 }
3092 3039
3093 3040 /*
3094 3041 * Try to allocate at least `size' bytes of memory without sleeping or
3095 3042 * panicking. Return actual allocated size in `asize'. If allocation failed,
3096 3043 * try final allocation with sleep or panic allowed.
3097 3044 */
3098 3045 void *
3099 3046 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3100 3047 {
3101 3048 void *p;
3102 3049
3103 3050 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3104 3051 do {
3105 3052 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3106 3053 if (p != NULL)
3107 3054 return (p);
3108 3055 *asize += KMEM_ALIGN;
3109 3056 } while (*asize <= PAGESIZE);
3110 3057
3111 3058 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3112 3059 return (kmem_alloc(*asize, kmflag));
3113 3060 }
3114 3061
3115 3062 /*
3116 3063 * Reclaim all unused memory from a cache.
3117 3064 */
3118 3065 static void
3119 3066 kmem_cache_reap(kmem_cache_t *cp)
3120 3067 {
3121 3068 ASSERT(taskq_member(kmem_taskq, curthread));
3122 3069 cp->cache_reap++;
3123 3070
3124 3071 /*
3125 3072 * Ask the cache's owner to free some memory if possible.
3126 3073 * The idea is to handle things like the inode cache, which
3127 3074 * typically sits on a bunch of memory that it doesn't truly
3128 3075 * *need*. Reclaim policy is entirely up to the owner; this
3129 3076 * callback is just an advisory plea for help.
3130 3077 */
3131 3078 if (cp->cache_reclaim != NULL) {
3132 3079 long delta;
3133 3080
3134 3081 /*
3135 3082 * Reclaimed memory should be reapable (not included in the
3136 3083 * depot's working set).
3137 3084 */
3138 3085 delta = cp->cache_full.ml_total;
3139 3086 cp->cache_reclaim(cp->cache_private);
3140 3087 delta = cp->cache_full.ml_total - delta;
3141 3088 if (delta > 0) {
3142 3089 mutex_enter(&cp->cache_depot_lock);
3143 3090 cp->cache_full.ml_reaplimit += delta;
3144 3091 cp->cache_full.ml_min += delta;
3145 3092 mutex_exit(&cp->cache_depot_lock);
3146 3093 }
3147 3094 }
3148 3095
3149 3096 kmem_depot_ws_reap(cp);
3150 3097
3151 3098 if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3152 3099 kmem_cache_defrag(cp);
3153 3100 }
3154 3101 }
3155 3102
3156 3103 static void
3157 3104 kmem_reap_timeout(void *flag_arg)
3158 3105 {
3159 3106 uint32_t *flag = (uint32_t *)flag_arg;
3160 3107
3161 3108 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3162 3109 *flag = 0;
3163 3110 }
3164 3111
3165 3112 static void
3166 3113 kmem_reap_done(void *flag)
3167 3114 {
3168 3115 if (!callout_init_done) {
3169 3116 /* can't schedule a timeout at this point */
3170 3117 kmem_reap_timeout(flag);
3171 3118 } else {
3172 3119 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3173 3120 }
3174 3121 }
3175 3122
3176 3123 static void
3177 3124 kmem_reap_start(void *flag)
3178 3125 {
3179 3126 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3180 3127
3181 3128 if (flag == &kmem_reaping) {
3182 3129 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3183 3130 /*
3184 3131 * if we have segkp under heap, reap segkp cache.
3185 3132 */
3186 3133 if (segkp_fromheap)
3187 3134 segkp_cache_free();
3188 3135 }
3189 3136 else
3190 3137 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3191 3138
3192 3139 /*
3193 3140 * We use taskq_dispatch() to schedule a timeout to clear
3194 3141 * the flag so that kmem_reap() becomes self-throttling:
3195 3142 * we won't reap again until the current reap completes *and*
3196 3143 * at least kmem_reap_interval ticks have elapsed.
3197 3144 */
3198 3145 if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3199 3146 kmem_reap_done(flag);
3200 3147 }
3201 3148
3202 3149 static void
3203 3150 kmem_reap_common(void *flag_arg)
3204 3151 {
3205 3152 uint32_t *flag = (uint32_t *)flag_arg;
3206 3153
3207 3154 if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3208 3155 atomic_cas_32(flag, 0, 1) != 0)
3209 3156 return;
3210 3157
3211 3158 /*
3212 3159 * It may not be kosher to do memory allocation when a reap is called
3213 3160 * (for example, if vmem_populate() is in the call chain). So we
3214 3161 * start the reap going with a TQ_NOALLOC dispatch. If the dispatch
3215 3162 * fails, we reset the flag, and the next reap will try again.
3216 3163 */
3217 3164 if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3218 3165 *flag = 0;
3219 3166 }
3220 3167
3221 3168 /*
3222 3169 * Reclaim all unused memory from all caches. Called from the VM system
3223 3170 * when memory gets tight.
3224 3171 */
3225 3172 void
3226 3173 kmem_reap(void)
3227 3174 {
3228 3175 kmem_reap_common(&kmem_reaping);
3229 3176 }
3230 3177
3231 3178 /*
3232 3179 * Reclaim all unused memory from identifier arenas, called when a vmem
3233 3180 * arena not back by memory is exhausted. Since reaping memory-backed caches
3234 3181 * cannot help with identifier exhaustion, we avoid both a large amount of
3235 3182 * work and unwanted side-effects from reclaim callbacks.
3236 3183 */
3237 3184 void
3238 3185 kmem_reap_idspace(void)
3239 3186 {
3240 3187 kmem_reap_common(&kmem_reaping_idspace);
3241 3188 }
3242 3189
3243 3190 /*
3244 3191 * Purge all magazines from a cache and set its magazine limit to zero.
3245 3192 * All calls are serialized by the kmem_taskq lock, except for the final
3246 3193 * call from kmem_cache_destroy().
3247 3194 */
3248 3195 static void
3249 3196 kmem_cache_magazine_purge(kmem_cache_t *cp)
3250 3197 {
3251 3198 kmem_cpu_cache_t *ccp;
3252 3199 kmem_magazine_t *mp, *pmp;
3253 3200 int rounds, prounds, cpu_seqid;
3254 3201
3255 3202 ASSERT(!list_link_active(&cp->cache_link) ||
3256 3203 taskq_member(kmem_taskq, curthread));
3257 3204 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3258 3205
3259 3206 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3260 3207 ccp = &cp->cache_cpu[cpu_seqid];
3261 3208
3262 3209 mutex_enter(&ccp->cc_lock);
3263 3210 mp = ccp->cc_loaded;
3264 3211 pmp = ccp->cc_ploaded;
3265 3212 rounds = ccp->cc_rounds;
3266 3213 prounds = ccp->cc_prounds;
3267 3214 ccp->cc_loaded = NULL;
3268 3215 ccp->cc_ploaded = NULL;
3269 3216 ccp->cc_rounds = -1;
3270 3217 ccp->cc_prounds = -1;
3271 3218 ccp->cc_magsize = 0;
3272 3219 mutex_exit(&ccp->cc_lock);
3273 3220
3274 3221 if (mp)
3275 3222 kmem_magazine_destroy(cp, mp, rounds);
3276 3223 if (pmp)
3277 3224 kmem_magazine_destroy(cp, pmp, prounds);
3278 3225 }
3279 3226
3280 3227 kmem_depot_ws_zero(cp);
3281 3228 kmem_depot_ws_reap(cp);
3282 3229 }
3283 3230
3284 3231 /*
3285 3232 * Enable per-cpu magazines on a cache.
3286 3233 */
3287 3234 static void
3288 3235 kmem_cache_magazine_enable(kmem_cache_t *cp)
3289 3236 {
3290 3237 int cpu_seqid;
3291 3238
3292 3239 if (cp->cache_flags & KMF_NOMAGAZINE)
3293 3240 return;
3294 3241
3295 3242 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3296 3243 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3297 3244 mutex_enter(&ccp->cc_lock);
3298 3245 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3299 3246 mutex_exit(&ccp->cc_lock);
3300 3247 }
3301 3248
3302 3249 }
3303 3250
3304 3251 /*
3305 3252 * Reap (almost) everything right now.
3306 3253 */
3307 3254 void
3308 3255 kmem_cache_reap_now(kmem_cache_t *cp)
3309 3256 {
3310 3257 ASSERT(list_link_active(&cp->cache_link));
3311 3258
3312 3259 kmem_depot_ws_zero(cp);
3313 3260
3314 3261 (void) taskq_dispatch(kmem_taskq,
3315 3262 (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3316 3263 taskq_wait(kmem_taskq);
3317 3264 }
3318 3265
3319 3266 /*
3320 3267 * Recompute a cache's magazine size. The trade-off is that larger magazines
3321 3268 * provide a higher transfer rate with the depot, while smaller magazines
3322 3269 * reduce memory consumption. Magazine resizing is an expensive operation;
3323 3270 * it should not be done frequently.
3324 3271 *
3325 3272 * Changes to the magazine size are serialized by the kmem_taskq lock.
3326 3273 *
3327 3274 * Note: at present this only grows the magazine size. It might be useful
3328 3275 * to allow shrinkage too.
3329 3276 */
3330 3277 static void
3331 3278 kmem_cache_magazine_resize(kmem_cache_t *cp)
3332 3279 {
3333 3280 kmem_magtype_t *mtp = cp->cache_magtype;
3334 3281
3335 3282 ASSERT(taskq_member(kmem_taskq, curthread));
3336 3283
3337 3284 if (cp->cache_chunksize < mtp->mt_maxbuf) {
3338 3285 kmem_cache_magazine_purge(cp);
3339 3286 mutex_enter(&cp->cache_depot_lock);
3340 3287 cp->cache_magtype = ++mtp;
3341 3288 cp->cache_depot_contention_prev =
3342 3289 cp->cache_depot_contention + INT_MAX;
3343 3290 mutex_exit(&cp->cache_depot_lock);
3344 3291 kmem_cache_magazine_enable(cp);
3345 3292 }
3346 3293 }
3347 3294
3348 3295 /*
3349 3296 * Rescale a cache's hash table, so that the table size is roughly the
3350 3297 * cache size. We want the average lookup time to be extremely small.
3351 3298 */
3352 3299 static void
3353 3300 kmem_hash_rescale(kmem_cache_t *cp)
3354 3301 {
3355 3302 kmem_bufctl_t **old_table, **new_table, *bcp;
3356 3303 size_t old_size, new_size, h;
3357 3304
3358 3305 ASSERT(taskq_member(kmem_taskq, curthread));
3359 3306
3360 3307 new_size = MAX(KMEM_HASH_INITIAL,
3361 3308 1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3362 3309 old_size = cp->cache_hash_mask + 1;
3363 3310
3364 3311 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3365 3312 return;
3366 3313
3367 3314 new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3368 3315 VM_NOSLEEP);
3369 3316 if (new_table == NULL)
3370 3317 return;
3371 3318 bzero(new_table, new_size * sizeof (void *));
3372 3319
3373 3320 mutex_enter(&cp->cache_lock);
3374 3321
3375 3322 old_size = cp->cache_hash_mask + 1;
3376 3323 old_table = cp->cache_hash_table;
3377 3324
3378 3325 cp->cache_hash_mask = new_size - 1;
3379 3326 cp->cache_hash_table = new_table;
3380 3327 cp->cache_rescale++;
3381 3328
3382 3329 for (h = 0; h < old_size; h++) {
3383 3330 bcp = old_table[h];
3384 3331 while (bcp != NULL) {
3385 3332 void *addr = bcp->bc_addr;
3386 3333 kmem_bufctl_t *next_bcp = bcp->bc_next;
3387 3334 kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3388 3335 bcp->bc_next = *hash_bucket;
3389 3336 *hash_bucket = bcp;
3390 3337 bcp = next_bcp;
3391 3338 }
3392 3339 }
3393 3340
3394 3341 mutex_exit(&cp->cache_lock);
3395 3342
3396 3343 vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3397 3344 }
3398 3345
3399 3346 /*
3400 3347 * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3401 3348 * update, magazine resizing, and slab consolidation.
3402 3349 */
3403 3350 static void
3404 3351 kmem_cache_update(kmem_cache_t *cp)
3405 3352 {
3406 3353 int need_hash_rescale = 0;
3407 3354 int need_magazine_resize = 0;
3408 3355
3409 3356 ASSERT(MUTEX_HELD(&kmem_cache_lock));
3410 3357
3411 3358 /*
3412 3359 * If the cache has become much larger or smaller than its hash table,
3413 3360 * fire off a request to rescale the hash table.
3414 3361 */
3415 3362 mutex_enter(&cp->cache_lock);
3416 3363
3417 3364 if ((cp->cache_flags & KMF_HASH) &&
3418 3365 (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3419 3366 (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3420 3367 cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3421 3368 need_hash_rescale = 1;
3422 3369
3423 3370 mutex_exit(&cp->cache_lock);
3424 3371
3425 3372 /*
3426 3373 * Update the depot working set statistics.
3427 3374 */
3428 3375 kmem_depot_ws_update(cp);
3429 3376
3430 3377 /*
3431 3378 * If there's a lot of contention in the depot,
3432 3379 * increase the magazine size.
3433 3380 */
3434 3381 mutex_enter(&cp->cache_depot_lock);
3435 3382
3436 3383 if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3437 3384 (int)(cp->cache_depot_contention -
3438 3385 cp->cache_depot_contention_prev) > kmem_depot_contention)
3439 3386 need_magazine_resize = 1;
3440 3387
3441 3388 cp->cache_depot_contention_prev = cp->cache_depot_contention;
3442 3389
3443 3390 mutex_exit(&cp->cache_depot_lock);
3444 3391
3445 3392 if (need_hash_rescale)
3446 3393 (void) taskq_dispatch(kmem_taskq,
3447 3394 (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3448 3395
3449 3396 if (need_magazine_resize)
3450 3397 (void) taskq_dispatch(kmem_taskq,
3451 3398 (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3452 3399
3453 3400 if (cp->cache_defrag != NULL)
3454 3401 (void) taskq_dispatch(kmem_taskq,
3455 3402 (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3456 3403 }
3457 3404
3458 3405 static void kmem_update(void *);
3459 3406
3460 3407 static void
3461 3408 kmem_update_timeout(void *dummy)
3462 3409 {
3463 3410 (void) timeout(kmem_update, dummy, kmem_reap_interval);
3464 3411 }
3465 3412
3466 3413 static void
3467 3414 kmem_update(void *dummy)
3468 3415 {
3469 3416 kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3470 3417
3471 3418 /*
3472 3419 * We use taskq_dispatch() to reschedule the timeout so that
3473 3420 * kmem_update() becomes self-throttling: it won't schedule
3474 3421 * new tasks until all previous tasks have completed.
3475 3422 */
3476 3423 if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3477 3424 kmem_update_timeout(NULL);
3478 3425 }
3479 3426
3480 3427 static int
3481 3428 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3482 3429 {
3483 3430 struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3484 3431 kmem_cache_t *cp = ksp->ks_private;
3485 3432 uint64_t cpu_buf_avail;
3486 3433 uint64_t buf_avail = 0;
3487 3434 int cpu_seqid;
3488 3435 long reap;
3489 3436
3490 3437 ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3491 3438
3492 3439 if (rw == KSTAT_WRITE)
3493 3440 return (EACCES);
3494 3441
3495 3442 mutex_enter(&cp->cache_lock);
3496 3443
3497 3444 kmcp->kmc_alloc_fail.value.ui64 = cp->cache_alloc_fail;
3498 3445 kmcp->kmc_alloc.value.ui64 = cp->cache_slab_alloc;
3499 3446 kmcp->kmc_free.value.ui64 = cp->cache_slab_free;
3500 3447 kmcp->kmc_slab_alloc.value.ui64 = cp->cache_slab_alloc;
3501 3448 kmcp->kmc_slab_free.value.ui64 = cp->cache_slab_free;
3502 3449
3503 3450 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3504 3451 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3505 3452
3506 3453 mutex_enter(&ccp->cc_lock);
3507 3454
3508 3455 cpu_buf_avail = 0;
3509 3456 if (ccp->cc_rounds > 0)
3510 3457 cpu_buf_avail += ccp->cc_rounds;
3511 3458 if (ccp->cc_prounds > 0)
3512 3459 cpu_buf_avail += ccp->cc_prounds;
3513 3460
3514 3461 kmcp->kmc_alloc.value.ui64 += ccp->cc_alloc;
3515 3462 kmcp->kmc_free.value.ui64 += ccp->cc_free;
3516 3463 buf_avail += cpu_buf_avail;
3517 3464
3518 3465 mutex_exit(&ccp->cc_lock);
3519 3466 }
3520 3467
3521 3468 mutex_enter(&cp->cache_depot_lock);
3522 3469
3523 3470 kmcp->kmc_depot_alloc.value.ui64 = cp->cache_full.ml_alloc;
3524 3471 kmcp->kmc_depot_free.value.ui64 = cp->cache_empty.ml_alloc;
3525 3472 kmcp->kmc_depot_contention.value.ui64 = cp->cache_depot_contention;
3526 3473 kmcp->kmc_full_magazines.value.ui64 = cp->cache_full.ml_total;
3527 3474 kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3528 3475 kmcp->kmc_magazine_size.value.ui64 =
3529 3476 (cp->cache_flags & KMF_NOMAGAZINE) ?
3530 3477 0 : cp->cache_magtype->mt_magsize;
3531 3478
3532 3479 kmcp->kmc_alloc.value.ui64 += cp->cache_full.ml_alloc;
3533 3480 kmcp->kmc_free.value.ui64 += cp->cache_empty.ml_alloc;
3534 3481 buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3535 3482
3536 3483 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3537 3484 reap = MIN(reap, cp->cache_full.ml_total);
3538 3485
3539 3486 mutex_exit(&cp->cache_depot_lock);
3540 3487
3541 3488 kmcp->kmc_buf_size.value.ui64 = cp->cache_bufsize;
3542 3489 kmcp->kmc_align.value.ui64 = cp->cache_align;
3543 3490 kmcp->kmc_chunk_size.value.ui64 = cp->cache_chunksize;
3544 3491 kmcp->kmc_slab_size.value.ui64 = cp->cache_slabsize;
3545 3492 kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3546 3493 buf_avail += cp->cache_bufslab;
3547 3494 kmcp->kmc_buf_avail.value.ui64 = buf_avail;
3548 3495 kmcp->kmc_buf_inuse.value.ui64 = cp->cache_buftotal - buf_avail;
3549 3496 kmcp->kmc_buf_total.value.ui64 = cp->cache_buftotal;
3550 3497 kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3551 3498 kmcp->kmc_slab_create.value.ui64 = cp->cache_slab_create;
3552 3499 kmcp->kmc_slab_destroy.value.ui64 = cp->cache_slab_destroy;
3553 3500 kmcp->kmc_hash_size.value.ui64 = (cp->cache_flags & KMF_HASH) ?
3554 3501 cp->cache_hash_mask + 1 : 0;
3555 3502 kmcp->kmc_hash_lookup_depth.value.ui64 = cp->cache_lookup_depth;
3556 3503 kmcp->kmc_hash_rescale.value.ui64 = cp->cache_rescale;
3557 3504 kmcp->kmc_vmem_source.value.ui64 = cp->cache_arena->vm_id;
3558 3505 kmcp->kmc_reap.value.ui64 = cp->cache_reap;
3559 3506
3560 3507 if (cp->cache_defrag == NULL) {
3561 3508 kmcp->kmc_move_callbacks.value.ui64 = 0;
3562 3509 kmcp->kmc_move_yes.value.ui64 = 0;
3563 3510 kmcp->kmc_move_no.value.ui64 = 0;
3564 3511 kmcp->kmc_move_later.value.ui64 = 0;
3565 3512 kmcp->kmc_move_dont_need.value.ui64 = 0;
3566 3513 kmcp->kmc_move_dont_know.value.ui64 = 0;
3567 3514 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3568 3515 kmcp->kmc_move_slabs_freed.value.ui64 = 0;
3569 3516 kmcp->kmc_defrag.value.ui64 = 0;
3570 3517 kmcp->kmc_scan.value.ui64 = 0;
3571 3518 kmcp->kmc_move_reclaimable.value.ui64 = 0;
|
↓ open down ↓ |
1628 lines elided |
↑ open up ↑ |
3572 3519 } else {
3573 3520 int64_t reclaimable;
3574 3521
3575 3522 kmem_defrag_t *kd = cp->cache_defrag;
3576 3523 kmcp->kmc_move_callbacks.value.ui64 = kd->kmd_callbacks;
3577 3524 kmcp->kmc_move_yes.value.ui64 = kd->kmd_yes;
3578 3525 kmcp->kmc_move_no.value.ui64 = kd->kmd_no;
3579 3526 kmcp->kmc_move_later.value.ui64 = kd->kmd_later;
3580 3527 kmcp->kmc_move_dont_need.value.ui64 = kd->kmd_dont_need;
3581 3528 kmcp->kmc_move_dont_know.value.ui64 = kd->kmd_dont_know;
3582 - kmcp->kmc_move_hunt_found.value.ui64 = kd->kmd_hunt_found;
3529 + kmcp->kmc_move_hunt_found.value.ui64 = 0;
3583 3530 kmcp->kmc_move_slabs_freed.value.ui64 = kd->kmd_slabs_freed;
3584 3531 kmcp->kmc_defrag.value.ui64 = kd->kmd_defrags;
3585 3532 kmcp->kmc_scan.value.ui64 = kd->kmd_scans;
3586 3533
3587 3534 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3588 3535 reclaimable = MAX(reclaimable, 0);
3589 3536 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3590 3537 kmcp->kmc_move_reclaimable.value.ui64 = reclaimable;
3591 3538 }
3592 3539
3593 3540 mutex_exit(&cp->cache_lock);
3594 3541 return (0);
3595 3542 }
3596 3543
3597 3544 /*
3598 3545 * Return a named statistic about a particular cache.
3599 3546 * This shouldn't be called very often, so it's currently designed for
3600 3547 * simplicity (leverages existing kstat support) rather than efficiency.
3601 3548 */
3602 3549 uint64_t
3603 3550 kmem_cache_stat(kmem_cache_t *cp, char *name)
3604 3551 {
3605 3552 int i;
3606 3553 kstat_t *ksp = cp->cache_kstat;
3607 3554 kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3608 3555 uint64_t value = 0;
3609 3556
3610 3557 if (ksp != NULL) {
3611 3558 mutex_enter(&kmem_cache_kstat_lock);
3612 3559 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3613 3560 for (i = 0; i < ksp->ks_ndata; i++) {
3614 3561 if (strcmp(knp[i].name, name) == 0) {
3615 3562 value = knp[i].value.ui64;
3616 3563 break;
3617 3564 }
3618 3565 }
3619 3566 mutex_exit(&kmem_cache_kstat_lock);
3620 3567 }
3621 3568 return (value);
3622 3569 }
3623 3570
3624 3571 /*
3625 3572 * Return an estimate of currently available kernel heap memory.
3626 3573 * On 32-bit systems, physical memory may exceed virtual memory,
3627 3574 * we just truncate the result at 1GB.
3628 3575 */
3629 3576 size_t
3630 3577 kmem_avail(void)
3631 3578 {
3632 3579 spgcnt_t rmem = availrmem - tune.t_minarmem;
3633 3580 spgcnt_t fmem = freemem - minfree;
3634 3581
3635 3582 return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3636 3583 1 << (30 - PAGESHIFT))));
3637 3584 }
3638 3585
3639 3586 /*
3640 3587 * Return the maximum amount of memory that is (in theory) allocatable
3641 3588 * from the heap. This may be used as an estimate only since there
3642 3589 * is no guarentee this space will still be available when an allocation
3643 3590 * request is made, nor that the space may be allocated in one big request
3644 3591 * due to kernel heap fragmentation.
3645 3592 */
3646 3593 size_t
3647 3594 kmem_maxavail(void)
3648 3595 {
3649 3596 spgcnt_t pmem = availrmem - tune.t_minarmem;
3650 3597 spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3651 3598
3652 3599 return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3653 3600 }
3654 3601
3655 3602 /*
3656 3603 * Indicate whether memory-intensive kmem debugging is enabled.
3657 3604 */
3658 3605 int
3659 3606 kmem_debugging(void)
3660 3607 {
3661 3608 return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3662 3609 }
3663 3610
3664 3611 /* binning function, sorts finely at the two extremes */
3665 3612 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift) \
3666 3613 ((((sp)->slab_refcnt <= (binshift)) || \
3667 3614 (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift))) \
3668 3615 ? -(sp)->slab_refcnt \
3669 3616 : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3670 3617
3671 3618 /*
3672 3619 * Minimizing the number of partial slabs on the freelist minimizes
3673 3620 * fragmentation (the ratio of unused buffers held by the slab layer). There are
3674 3621 * two ways to get a slab off of the freelist: 1) free all the buffers on the
3675 3622 * slab, and 2) allocate all the buffers on the slab. It follows that we want
3676 3623 * the most-used slabs at the front of the list where they have the best chance
3677 3624 * of being completely allocated, and the least-used slabs at a safe distance
3678 3625 * from the front to improve the odds that the few remaining buffers will all be
3679 3626 * freed before another allocation can tie up the slab. For that reason a slab
3680 3627 * with a higher slab_refcnt sorts less than than a slab with a lower
3681 3628 * slab_refcnt.
3682 3629 *
3683 3630 * However, if a slab has at least one buffer that is deemed unfreeable, we
3684 3631 * would rather have that slab at the front of the list regardless of
3685 3632 * slab_refcnt, since even one unfreeable buffer makes the entire slab
3686 3633 * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3687 3634 * callback, the slab is marked unfreeable for as long as it remains on the
3688 3635 * freelist.
3689 3636 */
3690 3637 static int
3691 3638 kmem_partial_slab_cmp(const void *p0, const void *p1)
3692 3639 {
3693 3640 const kmem_cache_t *cp;
3694 3641 const kmem_slab_t *s0 = p0;
3695 3642 const kmem_slab_t *s1 = p1;
3696 3643 int w0, w1;
3697 3644 size_t binshift;
3698 3645
3699 3646 ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3700 3647 ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3701 3648 ASSERT(s0->slab_cache == s1->slab_cache);
3702 3649 cp = s1->slab_cache;
3703 3650 ASSERT(MUTEX_HELD(&cp->cache_lock));
3704 3651 binshift = cp->cache_partial_binshift;
3705 3652
3706 3653 /* weight of first slab */
3707 3654 w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3708 3655 if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3709 3656 w0 -= cp->cache_maxchunks;
3710 3657 }
3711 3658
3712 3659 /* weight of second slab */
3713 3660 w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3714 3661 if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3715 3662 w1 -= cp->cache_maxchunks;
3716 3663 }
3717 3664
3718 3665 if (w0 < w1)
3719 3666 return (-1);
3720 3667 if (w0 > w1)
3721 3668 return (1);
3722 3669
3723 3670 /* compare pointer values */
3724 3671 if ((uintptr_t)s0 < (uintptr_t)s1)
3725 3672 return (-1);
3726 3673 if ((uintptr_t)s0 > (uintptr_t)s1)
3727 3674 return (1);
3728 3675
3729 3676 return (0);
3730 3677 }
3731 3678
3732 3679 /*
3733 3680 * It must be valid to call the destructor (if any) on a newly created object.
3734 3681 * That is, the constructor (if any) must leave the object in a valid state for
3735 3682 * the destructor.
3736 3683 */
3737 3684 kmem_cache_t *
3738 3685 kmem_cache_create(
3739 3686 char *name, /* descriptive name for this cache */
3740 3687 size_t bufsize, /* size of the objects it manages */
3741 3688 size_t align, /* required object alignment */
3742 3689 int (*constructor)(void *, void *, int), /* object constructor */
3743 3690 void (*destructor)(void *, void *), /* object destructor */
3744 3691 void (*reclaim)(void *), /* memory reclaim callback */
3745 3692 void *private, /* pass-thru arg for constr/destr/reclaim */
3746 3693 vmem_t *vmp, /* vmem source for slab allocation */
3747 3694 int cflags) /* cache creation flags */
3748 3695 {
3749 3696 int cpu_seqid;
3750 3697 size_t chunksize;
3751 3698 kmem_cache_t *cp;
3752 3699 kmem_magtype_t *mtp;
3753 3700 size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3754 3701
3755 3702 #ifdef DEBUG
3756 3703 /*
3757 3704 * Cache names should conform to the rules for valid C identifiers
3758 3705 */
3759 3706 if (!strident_valid(name)) {
3760 3707 cmn_err(CE_CONT,
3761 3708 "kmem_cache_create: '%s' is an invalid cache name\n"
3762 3709 "cache names must conform to the rules for "
3763 3710 "C identifiers\n", name);
3764 3711 }
3765 3712 #endif /* DEBUG */
3766 3713
3767 3714 if (vmp == NULL)
3768 3715 vmp = kmem_default_arena;
3769 3716
3770 3717 /*
3771 3718 * If this kmem cache has an identifier vmem arena as its source, mark
3772 3719 * it such to allow kmem_reap_idspace().
3773 3720 */
3774 3721 ASSERT(!(cflags & KMC_IDENTIFIER)); /* consumer should not set this */
3775 3722 if (vmp->vm_cflags & VMC_IDENTIFIER)
3776 3723 cflags |= KMC_IDENTIFIER;
3777 3724
3778 3725 /*
3779 3726 * Get a kmem_cache structure. We arrange that cp->cache_cpu[]
3780 3727 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3781 3728 * false sharing of per-CPU data.
3782 3729 */
3783 3730 cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3784 3731 P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3785 3732 bzero(cp, csize);
3786 3733 list_link_init(&cp->cache_link);
3787 3734
3788 3735 if (align == 0)
3789 3736 align = KMEM_ALIGN;
3790 3737
3791 3738 /*
3792 3739 * If we're not at least KMEM_ALIGN aligned, we can't use free
3793 3740 * memory to hold bufctl information (because we can't safely
3794 3741 * perform word loads and stores on it).
3795 3742 */
3796 3743 if (align < KMEM_ALIGN)
3797 3744 cflags |= KMC_NOTOUCH;
3798 3745
3799 3746 if (!ISP2(align) || align > vmp->vm_quantum)
3800 3747 panic("kmem_cache_create: bad alignment %lu", align);
3801 3748
3802 3749 mutex_enter(&kmem_flags_lock);
3803 3750 if (kmem_flags & KMF_RANDOMIZE)
3804 3751 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3805 3752 KMF_RANDOMIZE;
3806 3753 cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3807 3754 mutex_exit(&kmem_flags_lock);
3808 3755
3809 3756 /*
3810 3757 * Make sure all the various flags are reasonable.
3811 3758 */
3812 3759 ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3813 3760
3814 3761 if (cp->cache_flags & KMF_LITE) {
3815 3762 if (bufsize >= kmem_lite_minsize &&
3816 3763 align <= kmem_lite_maxalign &&
3817 3764 P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3818 3765 cp->cache_flags |= KMF_BUFTAG;
3819 3766 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3820 3767 } else {
3821 3768 cp->cache_flags &= ~KMF_DEBUG;
3822 3769 }
3823 3770 }
3824 3771
3825 3772 if (cp->cache_flags & KMF_DEADBEEF)
3826 3773 cp->cache_flags |= KMF_REDZONE;
3827 3774
3828 3775 if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3829 3776 cp->cache_flags |= KMF_NOMAGAZINE;
3830 3777
3831 3778 if (cflags & KMC_NODEBUG)
3832 3779 cp->cache_flags &= ~KMF_DEBUG;
3833 3780
3834 3781 if (cflags & KMC_NOTOUCH)
3835 3782 cp->cache_flags &= ~KMF_TOUCH;
3836 3783
3837 3784 if (cflags & KMC_PREFILL)
3838 3785 cp->cache_flags |= KMF_PREFILL;
3839 3786
3840 3787 if (cflags & KMC_NOHASH)
3841 3788 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3842 3789
3843 3790 if (cflags & KMC_NOMAGAZINE)
3844 3791 cp->cache_flags |= KMF_NOMAGAZINE;
3845 3792
3846 3793 if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3847 3794 cp->cache_flags |= KMF_REDZONE;
3848 3795
3849 3796 if (!(cp->cache_flags & KMF_AUDIT))
3850 3797 cp->cache_flags &= ~KMF_CONTENTS;
3851 3798
3852 3799 if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3853 3800 !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3854 3801 cp->cache_flags |= KMF_FIREWALL;
3855 3802
3856 3803 if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3857 3804 cp->cache_flags &= ~KMF_FIREWALL;
3858 3805
3859 3806 if (cp->cache_flags & KMF_FIREWALL) {
3860 3807 cp->cache_flags &= ~KMF_BUFTAG;
3861 3808 cp->cache_flags |= KMF_NOMAGAZINE;
3862 3809 ASSERT(vmp == kmem_default_arena);
3863 3810 vmp = kmem_firewall_arena;
3864 3811 }
3865 3812
3866 3813 /*
3867 3814 * Set cache properties.
3868 3815 */
3869 3816 (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3870 3817 strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3871 3818 cp->cache_bufsize = bufsize;
3872 3819 cp->cache_align = align;
3873 3820 cp->cache_constructor = constructor;
3874 3821 cp->cache_destructor = destructor;
3875 3822 cp->cache_reclaim = reclaim;
3876 3823 cp->cache_private = private;
3877 3824 cp->cache_arena = vmp;
3878 3825 cp->cache_cflags = cflags;
3879 3826
3880 3827 /*
3881 3828 * Determine the chunk size.
3882 3829 */
3883 3830 chunksize = bufsize;
3884 3831
3885 3832 if (align >= KMEM_ALIGN) {
3886 3833 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3887 3834 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3888 3835 }
3889 3836
3890 3837 if (cp->cache_flags & KMF_BUFTAG) {
3891 3838 cp->cache_bufctl = chunksize;
3892 3839 cp->cache_buftag = chunksize;
3893 3840 if (cp->cache_flags & KMF_LITE)
3894 3841 chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3895 3842 else
3896 3843 chunksize += sizeof (kmem_buftag_t);
3897 3844 }
3898 3845
3899 3846 if (cp->cache_flags & KMF_DEADBEEF) {
3900 3847 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3901 3848 if (cp->cache_flags & KMF_LITE)
3902 3849 cp->cache_verify = sizeof (uint64_t);
3903 3850 }
3904 3851
3905 3852 cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3906 3853
3907 3854 cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3908 3855
3909 3856 /*
3910 3857 * Now that we know the chunk size, determine the optimal slab size.
3911 3858 */
3912 3859 if (vmp == kmem_firewall_arena) {
3913 3860 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3914 3861 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3915 3862 cp->cache_maxcolor = cp->cache_mincolor;
3916 3863 cp->cache_flags |= KMF_HASH;
3917 3864 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3918 3865 } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3919 3866 !(cp->cache_flags & KMF_AUDIT) &&
3920 3867 chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3921 3868 cp->cache_slabsize = vmp->vm_quantum;
3922 3869 cp->cache_mincolor = 0;
3923 3870 cp->cache_maxcolor =
3924 3871 (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3925 3872 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3926 3873 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3927 3874 } else {
3928 3875 size_t chunks, bestfit, waste, slabsize;
3929 3876 size_t minwaste = LONG_MAX;
3930 3877
3931 3878 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3932 3879 slabsize = P2ROUNDUP(chunksize * chunks,
3933 3880 vmp->vm_quantum);
3934 3881 chunks = slabsize / chunksize;
3935 3882 waste = (slabsize % chunksize) / chunks;
3936 3883 if (waste < minwaste) {
3937 3884 minwaste = waste;
3938 3885 bestfit = slabsize;
3939 3886 }
3940 3887 }
3941 3888 if (cflags & KMC_QCACHE)
3942 3889 bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3943 3890 cp->cache_slabsize = bestfit;
3944 3891 cp->cache_mincolor = 0;
3945 3892 cp->cache_maxcolor = bestfit % chunksize;
3946 3893 cp->cache_flags |= KMF_HASH;
3947 3894 }
3948 3895
3949 3896 cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3950 3897 cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3951 3898
3952 3899 /*
3953 3900 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3954 3901 * there is a constructor avoids some tricky issues with debug setup
3955 3902 * that may be revisited later. We cannot allow prefill in a
3956 3903 * metadata cache because of potential recursion.
3957 3904 */
3958 3905 if (vmp == kmem_msb_arena ||
3959 3906 cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3960 3907 cp->cache_constructor != NULL)
3961 3908 cp->cache_flags &= ~KMF_PREFILL;
3962 3909
3963 3910 if (cp->cache_flags & KMF_HASH) {
3964 3911 ASSERT(!(cflags & KMC_NOHASH));
3965 3912 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3966 3913 kmem_bufctl_audit_cache : kmem_bufctl_cache;
3967 3914 }
3968 3915
3969 3916 if (cp->cache_maxcolor >= vmp->vm_quantum)
3970 3917 cp->cache_maxcolor = vmp->vm_quantum - 1;
3971 3918
3972 3919 cp->cache_color = cp->cache_mincolor;
3973 3920
3974 3921 /*
3975 3922 * Initialize the rest of the slab layer.
3976 3923 */
3977 3924 mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3978 3925
3979 3926 avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3980 3927 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3981 3928 /* LINTED: E_TRUE_LOGICAL_EXPR */
3982 3929 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3983 3930 /* reuse partial slab AVL linkage for complete slab list linkage */
3984 3931 list_create(&cp->cache_complete_slabs,
3985 3932 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3986 3933
3987 3934 if (cp->cache_flags & KMF_HASH) {
3988 3935 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3989 3936 KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3990 3937 bzero(cp->cache_hash_table,
3991 3938 KMEM_HASH_INITIAL * sizeof (void *));
3992 3939 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3993 3940 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3994 3941 }
3995 3942
3996 3943 /*
3997 3944 * Initialize the depot.
3998 3945 */
3999 3946 mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
4000 3947
4001 3948 for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
4002 3949 continue;
4003 3950
4004 3951 cp->cache_magtype = mtp;
4005 3952
4006 3953 /*
4007 3954 * Initialize the CPU layer.
4008 3955 */
4009 3956 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4010 3957 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
4011 3958 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
4012 3959 ccp->cc_flags = cp->cache_flags;
4013 3960 ccp->cc_rounds = -1;
4014 3961 ccp->cc_prounds = -1;
4015 3962 }
4016 3963
4017 3964 /*
4018 3965 * Create the cache's kstats.
4019 3966 */
4020 3967 if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
4021 3968 "kmem_cache", KSTAT_TYPE_NAMED,
4022 3969 sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
4023 3970 KSTAT_FLAG_VIRTUAL)) != NULL) {
4024 3971 cp->cache_kstat->ks_data = &kmem_cache_kstat;
4025 3972 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
4026 3973 cp->cache_kstat->ks_private = cp;
4027 3974 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
4028 3975 kstat_install(cp->cache_kstat);
4029 3976 }
4030 3977
4031 3978 /*
4032 3979 * Add the cache to the global list. This makes it visible
4033 3980 * to kmem_update(), so the cache must be ready for business.
4034 3981 */
4035 3982 mutex_enter(&kmem_cache_lock);
4036 3983 list_insert_tail(&kmem_caches, cp);
4037 3984 mutex_exit(&kmem_cache_lock);
4038 3985
4039 3986 if (kmem_ready)
4040 3987 kmem_cache_magazine_enable(cp);
4041 3988
4042 3989 return (cp);
4043 3990 }
4044 3991
4045 3992 static int
4046 3993 kmem_move_cmp(const void *buf, const void *p)
4047 3994 {
4048 3995 const kmem_move_t *kmm = p;
4049 3996 uintptr_t v1 = (uintptr_t)buf;
4050 3997 uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
4051 3998 return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
4052 3999 }
4053 4000
4054 4001 static void
4055 4002 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4056 4003 {
4057 4004 kmd->kmd_reclaim_numer = 1;
4058 4005 }
4059 4006
4060 4007 /*
4061 4008 * Initially, when choosing candidate slabs for buffers to move, we want to be
4062 4009 * very selective and take only slabs that are less than
4063 4010 * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4064 4011 * slabs, then we raise the allocation ceiling incrementally. The reclaim
4065 4012 * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4066 4013 * longer fragmented.
4067 4014 */
4068 4015 static void
4069 4016 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4070 4017 {
4071 4018 if (direction > 0) {
4072 4019 /* make it easier to find a candidate slab */
4073 4020 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4074 4021 kmd->kmd_reclaim_numer++;
4075 4022 }
4076 4023 } else {
4077 4024 /* be more selective */
4078 4025 if (kmd->kmd_reclaim_numer > 1) {
4079 4026 kmd->kmd_reclaim_numer--;
4080 4027 }
4081 4028 }
4082 4029 }
4083 4030
4084 4031 void
4085 4032 kmem_cache_set_move(kmem_cache_t *cp,
4086 4033 kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4087 4034 {
4088 4035 kmem_defrag_t *defrag;
4089 4036
4090 4037 ASSERT(move != NULL);
4091 4038 /*
4092 4039 * The consolidator does not support NOTOUCH caches because kmem cannot
4093 4040 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4094 4041 * a low order bit usable by clients to distinguish uninitialized memory
4095 4042 * from known objects (see kmem_slab_create).
4096 4043 */
4097 4044 ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4098 4045 ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4099 4046
4100 4047 /*
4101 4048 * We should not be holding anyone's cache lock when calling
4102 4049 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4103 4050 * lock.
4104 4051 */
4105 4052 defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4106 4053
4107 4054 mutex_enter(&cp->cache_lock);
4108 4055
4109 4056 if (KMEM_IS_MOVABLE(cp)) {
4110 4057 if (cp->cache_move == NULL) {
4111 4058 ASSERT(cp->cache_slab_alloc == 0);
4112 4059
4113 4060 cp->cache_defrag = defrag;
4114 4061 defrag = NULL; /* nothing to free */
4115 4062 bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4116 4063 avl_create(&cp->cache_defrag->kmd_moves_pending,
4117 4064 kmem_move_cmp, sizeof (kmem_move_t),
4118 4065 offsetof(kmem_move_t, kmm_entry));
4119 4066 /* LINTED: E_TRUE_LOGICAL_EXPR */
4120 4067 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4121 4068 /* reuse the slab's AVL linkage for deadlist linkage */
4122 4069 list_create(&cp->cache_defrag->kmd_deadlist,
4123 4070 sizeof (kmem_slab_t),
4124 4071 offsetof(kmem_slab_t, slab_link));
4125 4072 kmem_reset_reclaim_threshold(cp->cache_defrag);
4126 4073 }
4127 4074 cp->cache_move = move;
4128 4075 }
4129 4076
4130 4077 mutex_exit(&cp->cache_lock);
4131 4078
4132 4079 if (defrag != NULL) {
4133 4080 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4134 4081 }
4135 4082 }
4136 4083
4137 4084 void
4138 4085 kmem_cache_destroy(kmem_cache_t *cp)
4139 4086 {
4140 4087 int cpu_seqid;
4141 4088
4142 4089 /*
|
↓ open down ↓ |
550 lines elided |
↑ open up ↑ |
4143 4090 * Remove the cache from the global cache list so that no one else
4144 4091 * can schedule tasks on its behalf, wait for any pending tasks to
4145 4092 * complete, purge the cache, and then destroy it.
4146 4093 */
4147 4094 mutex_enter(&kmem_cache_lock);
4148 4095 list_remove(&kmem_caches, cp);
4149 4096 mutex_exit(&kmem_cache_lock);
4150 4097
4151 4098 if (kmem_taskq != NULL)
4152 4099 taskq_wait(kmem_taskq);
4153 - if (kmem_move_taskq != NULL)
4100 +
4101 + if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4154 4102 taskq_wait(kmem_move_taskq);
4155 4103
4156 4104 kmem_cache_magazine_purge(cp);
4157 4105
4158 4106 mutex_enter(&cp->cache_lock);
4159 4107 if (cp->cache_buftotal != 0)
4160 4108 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4161 4109 cp->cache_name, (void *)cp);
4162 4110 if (cp->cache_defrag != NULL) {
4163 4111 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4164 4112 list_destroy(&cp->cache_defrag->kmd_deadlist);
4165 4113 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4166 4114 cp->cache_defrag = NULL;
4167 4115 }
4168 4116 /*
4169 4117 * The cache is now dead. There should be no further activity. We
4170 4118 * enforce this by setting land mines in the constructor, destructor,
4171 4119 * reclaim, and move routines that induce a kernel text fault if
4172 4120 * invoked.
4173 4121 */
4174 4122 cp->cache_constructor = (int (*)(void *, void *, int))1;
4175 4123 cp->cache_destructor = (void (*)(void *, void *))2;
4176 4124 cp->cache_reclaim = (void (*)(void *))3;
4177 4125 cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4178 4126 mutex_exit(&cp->cache_lock);
4179 4127
4180 4128 kstat_delete(cp->cache_kstat);
4181 4129
4182 4130 if (cp->cache_hash_table != NULL)
4183 4131 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4184 4132 (cp->cache_hash_mask + 1) * sizeof (void *));
4185 4133
4186 4134 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4187 4135 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4188 4136
4189 4137 mutex_destroy(&cp->cache_depot_lock);
4190 4138 mutex_destroy(&cp->cache_lock);
4191 4139
4192 4140 vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4193 4141 }
4194 4142
4195 4143 /*ARGSUSED*/
4196 4144 static int
4197 4145 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4198 4146 {
4199 4147 ASSERT(MUTEX_HELD(&cpu_lock));
4200 4148 if (what == CPU_UNCONFIG) {
4201 4149 kmem_cache_applyall(kmem_cache_magazine_purge,
4202 4150 kmem_taskq, TQ_SLEEP);
4203 4151 kmem_cache_applyall(kmem_cache_magazine_enable,
4204 4152 kmem_taskq, TQ_SLEEP);
4205 4153 }
4206 4154 return (0);
4207 4155 }
4208 4156
4209 4157 static void
4210 4158 kmem_alloc_caches_create(const int *array, size_t count,
4211 4159 kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4212 4160 {
4213 4161 char name[KMEM_CACHE_NAMELEN + 1];
4214 4162 size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4215 4163 size_t size = table_unit;
4216 4164 int i;
4217 4165
4218 4166 for (i = 0; i < count; i++) {
4219 4167 size_t cache_size = array[i];
4220 4168 size_t align = KMEM_ALIGN;
4221 4169 kmem_cache_t *cp;
4222 4170
4223 4171 /* if the table has an entry for maxbuf, we're done */
4224 4172 if (size > maxbuf)
4225 4173 break;
4226 4174
4227 4175 /* cache size must be a multiple of the table unit */
4228 4176 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4229 4177
4230 4178 /*
4231 4179 * If they allocate a multiple of the coherency granularity,
4232 4180 * they get a coherency-granularity-aligned address.
4233 4181 */
4234 4182 if (IS_P2ALIGNED(cache_size, 64))
4235 4183 align = 64;
4236 4184 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4237 4185 align = PAGESIZE;
4238 4186 (void) snprintf(name, sizeof (name),
4239 4187 "kmem_alloc_%lu", cache_size);
4240 4188 cp = kmem_cache_create(name, cache_size, align,
4241 4189 NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4242 4190
4243 4191 while (size <= cache_size) {
4244 4192 alloc_table[(size - 1) >> shift] = cp;
4245 4193 size += table_unit;
4246 4194 }
4247 4195 }
4248 4196
4249 4197 ASSERT(size > maxbuf); /* i.e. maxbuf <= max(cache_size) */
4250 4198 }
4251 4199
4252 4200 static void
4253 4201 kmem_cache_init(int pass, int use_large_pages)
4254 4202 {
4255 4203 int i;
4256 4204 size_t maxbuf;
4257 4205 kmem_magtype_t *mtp;
4258 4206
4259 4207 for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4260 4208 char name[KMEM_CACHE_NAMELEN + 1];
4261 4209
4262 4210 mtp = &kmem_magtype[i];
4263 4211 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4264 4212 mtp->mt_cache = kmem_cache_create(name,
4265 4213 (mtp->mt_magsize + 1) * sizeof (void *),
4266 4214 mtp->mt_align, NULL, NULL, NULL, NULL,
4267 4215 kmem_msb_arena, KMC_NOHASH);
4268 4216 }
4269 4217
4270 4218 kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4271 4219 sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4272 4220 kmem_msb_arena, KMC_NOHASH);
4273 4221
4274 4222 kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4275 4223 sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4276 4224 kmem_msb_arena, KMC_NOHASH);
4277 4225
4278 4226 kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4279 4227 sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4280 4228 kmem_msb_arena, KMC_NOHASH);
4281 4229
4282 4230 if (pass == 2) {
4283 4231 kmem_va_arena = vmem_create("kmem_va",
4284 4232 NULL, 0, PAGESIZE,
4285 4233 vmem_alloc, vmem_free, heap_arena,
4286 4234 8 * PAGESIZE, VM_SLEEP);
4287 4235
4288 4236 if (use_large_pages) {
4289 4237 kmem_default_arena = vmem_xcreate("kmem_default",
4290 4238 NULL, 0, PAGESIZE,
4291 4239 segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4292 4240 0, VMC_DUMPSAFE | VM_SLEEP);
4293 4241 } else {
4294 4242 kmem_default_arena = vmem_create("kmem_default",
4295 4243 NULL, 0, PAGESIZE,
4296 4244 segkmem_alloc, segkmem_free, kmem_va_arena,
4297 4245 0, VMC_DUMPSAFE | VM_SLEEP);
4298 4246 }
4299 4247
4300 4248 /* Figure out what our maximum cache size is */
4301 4249 maxbuf = kmem_max_cached;
4302 4250 if (maxbuf <= KMEM_MAXBUF) {
4303 4251 maxbuf = 0;
4304 4252 kmem_max_cached = KMEM_MAXBUF;
4305 4253 } else {
4306 4254 size_t size = 0;
4307 4255 size_t max =
4308 4256 sizeof (kmem_big_alloc_sizes) / sizeof (int);
4309 4257 /*
4310 4258 * Round maxbuf up to an existing cache size. If maxbuf
4311 4259 * is larger than the largest cache, we truncate it to
4312 4260 * the largest cache's size.
4313 4261 */
4314 4262 for (i = 0; i < max; i++) {
4315 4263 size = kmem_big_alloc_sizes[i];
4316 4264 if (maxbuf <= size)
4317 4265 break;
4318 4266 }
4319 4267 kmem_max_cached = maxbuf = size;
4320 4268 }
4321 4269
4322 4270 /*
4323 4271 * The big alloc table may not be completely overwritten, so
4324 4272 * we clear out any stale cache pointers from the first pass.
4325 4273 */
4326 4274 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4327 4275 } else {
4328 4276 /*
4329 4277 * During the first pass, the kmem_alloc_* caches
4330 4278 * are treated as metadata.
4331 4279 */
4332 4280 kmem_default_arena = kmem_msb_arena;
4333 4281 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4334 4282 }
4335 4283
4336 4284 /*
4337 4285 * Set up the default caches to back kmem_alloc()
4338 4286 */
4339 4287 kmem_alloc_caches_create(
4340 4288 kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4341 4289 kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4342 4290
4343 4291 kmem_alloc_caches_create(
4344 4292 kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4345 4293 kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4346 4294
4347 4295 kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4348 4296 }
4349 4297
4350 4298 void
4351 4299 kmem_init(void)
4352 4300 {
4353 4301 kmem_cache_t *cp;
4354 4302 int old_kmem_flags = kmem_flags;
4355 4303 int use_large_pages = 0;
4356 4304 size_t maxverify, minfirewall;
4357 4305
4358 4306 kstat_init();
4359 4307
4360 4308 /*
4361 4309 * Don't do firewalled allocations if the heap is less than 1TB
4362 4310 * (i.e. on a 32-bit kernel)
4363 4311 * The resulting VM_NEXTFIT allocations would create too much
4364 4312 * fragmentation in a small heap.
4365 4313 */
4366 4314 #if defined(_LP64)
4367 4315 maxverify = minfirewall = PAGESIZE / 2;
4368 4316 #else
4369 4317 maxverify = minfirewall = ULONG_MAX;
4370 4318 #endif
4371 4319
4372 4320 /* LINTED */
4373 4321 ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4374 4322
4375 4323 list_create(&kmem_caches, sizeof (kmem_cache_t),
4376 4324 offsetof(kmem_cache_t, cache_link));
4377 4325
4378 4326 kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4379 4327 vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4380 4328 VM_SLEEP | VMC_NO_QCACHE);
4381 4329
4382 4330 kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4383 4331 PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4384 4332 VMC_DUMPSAFE | VM_SLEEP);
4385 4333
4386 4334 kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4387 4335 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4388 4336
4389 4337 kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4390 4338 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4391 4339
4392 4340 kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4393 4341 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4394 4342
4395 4343 kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4396 4344 NULL, 0, PAGESIZE,
4397 4345 kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4398 4346 0, VM_SLEEP);
4399 4347
4400 4348 kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4401 4349 segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4402 4350 VMC_DUMPSAFE | VM_SLEEP);
4403 4351
4404 4352 /* temporary oversize arena for mod_read_system_file */
4405 4353 kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4406 4354 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4407 4355
4408 4356 kmem_reap_interval = 15 * hz;
4409 4357
4410 4358 /*
4411 4359 * Read /etc/system. This is a chicken-and-egg problem because
4412 4360 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4413 4361 * needs to use the allocator. The simplest solution is to create
4414 4362 * all the standard kmem caches, read /etc/system, destroy all the
4415 4363 * caches we just created, and then create them all again in light
4416 4364 * of the (possibly) new kmem_flags and other kmem tunables.
4417 4365 */
4418 4366 kmem_cache_init(1, 0);
4419 4367
4420 4368 mod_read_system_file(boothowto & RB_ASKNAME);
4421 4369
4422 4370 while ((cp = list_tail(&kmem_caches)) != NULL)
4423 4371 kmem_cache_destroy(cp);
4424 4372
4425 4373 vmem_destroy(kmem_oversize_arena);
4426 4374
4427 4375 if (old_kmem_flags & KMF_STICKY)
4428 4376 kmem_flags = old_kmem_flags;
4429 4377
4430 4378 if (!(kmem_flags & KMF_AUDIT))
4431 4379 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4432 4380
4433 4381 if (kmem_maxverify == 0)
4434 4382 kmem_maxverify = maxverify;
4435 4383
4436 4384 if (kmem_minfirewall == 0)
4437 4385 kmem_minfirewall = minfirewall;
4438 4386
4439 4387 /*
4440 4388 * give segkmem a chance to figure out if we are using large pages
4441 4389 * for the kernel heap
4442 4390 */
4443 4391 use_large_pages = segkmem_lpsetup();
4444 4392
4445 4393 /*
4446 4394 * To protect against corruption, we keep the actual number of callers
4447 4395 * KMF_LITE records seperate from the tunable. We arbitrarily clamp
4448 4396 * to 16, since the overhead for small buffers quickly gets out of
4449 4397 * hand.
4450 4398 *
4451 4399 * The real limit would depend on the needs of the largest KMC_NOHASH
4452 4400 * cache.
4453 4401 */
4454 4402 kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4455 4403 kmem_lite_pcs = kmem_lite_count;
4456 4404
4457 4405 /*
4458 4406 * Normally, we firewall oversized allocations when possible, but
4459 4407 * if we are using large pages for kernel memory, and we don't have
4460 4408 * any non-LITE debugging flags set, we want to allocate oversized
4461 4409 * buffers from large pages, and so skip the firewalling.
4462 4410 */
4463 4411 if (use_large_pages &&
4464 4412 ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4465 4413 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4466 4414 PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4467 4415 0, VMC_DUMPSAFE | VM_SLEEP);
4468 4416 } else {
4469 4417 kmem_oversize_arena = vmem_create("kmem_oversize",
4470 4418 NULL, 0, PAGESIZE,
4471 4419 segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4472 4420 kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4473 4421 VM_SLEEP);
4474 4422 }
4475 4423
4476 4424 kmem_cache_init(2, use_large_pages);
4477 4425
4478 4426 if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4479 4427 if (kmem_transaction_log_size == 0)
4480 4428 kmem_transaction_log_size = kmem_maxavail() / 50;
4481 4429 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4482 4430 }
4483 4431
4484 4432 if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4485 4433 if (kmem_content_log_size == 0)
4486 4434 kmem_content_log_size = kmem_maxavail() / 50;
4487 4435 kmem_content_log = kmem_log_init(kmem_content_log_size);
4488 4436 }
4489 4437
4490 4438 kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4491 4439
4492 4440 kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4493 4441
4494 4442 /*
4495 4443 * Initialize STREAMS message caches so allocb() is available.
4496 4444 * This allows us to initialize the logging framework (cmn_err(9F),
4497 4445 * strlog(9F), etc) so we can start recording messages.
4498 4446 */
4499 4447 streams_msg_init();
4500 4448
4501 4449 /*
4502 4450 * Initialize the ZSD framework in Zones so modules loaded henceforth
4503 4451 * can register their callbacks.
4504 4452 */
4505 4453 zone_zsd_init();
4506 4454
4507 4455 log_init();
4508 4456 taskq_init();
4509 4457
4510 4458 /*
4511 4459 * Warn about invalid or dangerous values of kmem_flags.
4512 4460 * Always warn about unsupported values.
4513 4461 */
4514 4462 if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4515 4463 KMF_CONTENTS | KMF_LITE)) != 0) ||
4516 4464 ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4517 4465 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4518 4466 "See the Solaris Tunable Parameters Reference Manual.",
4519 4467 kmem_flags);
4520 4468
4521 4469 #ifdef DEBUG
4522 4470 if ((kmem_flags & KMF_DEBUG) == 0)
4523 4471 cmn_err(CE_NOTE, "kmem debugging disabled.");
4524 4472 #else
4525 4473 /*
4526 4474 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4527 4475 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4528 4476 * if KMF_AUDIT is set). We should warn the user about the performance
4529 4477 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4530 4478 * isn't set (since that disables AUDIT).
4531 4479 */
4532 4480 if (!(kmem_flags & KMF_LITE) &&
4533 4481 (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4534 4482 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4535 4483 "enabled (kmem_flags = 0x%x). Performance degradation "
4536 4484 "and large memory overhead possible. See the Solaris "
4537 4485 "Tunable Parameters Reference Manual.", kmem_flags);
4538 4486 #endif /* not DEBUG */
4539 4487
4540 4488 kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4541 4489
4542 4490 kmem_ready = 1;
4543 4491
4544 4492 /*
4545 4493 * Initialize the platform-specific aligned/DMA memory allocator.
4546 4494 */
4547 4495 ka_init();
4548 4496
4549 4497 /*
4550 4498 * Initialize 32-bit ID cache.
4551 4499 */
4552 4500 id32_init();
4553 4501
4554 4502 /*
4555 4503 * Initialize the networking stack so modules loaded can
4556 4504 * register their callbacks.
4557 4505 */
4558 4506 netstack_init();
4559 4507 }
4560 4508
4561 4509 static void
4562 4510 kmem_move_init(void)
4563 4511 {
4564 4512 kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4565 4513 sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4566 4514 kmem_msb_arena, KMC_NOHASH);
4567 4515 kmem_move_cache = kmem_cache_create("kmem_move_cache",
4568 4516 sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4569 4517 kmem_msb_arena, KMC_NOHASH);
4570 4518
4571 4519 /*
4572 4520 * kmem guarantees that move callbacks are sequential and that even
4573 4521 * across multiple caches no two moves ever execute simultaneously.
4574 4522 * Move callbacks are processed on a separate taskq so that client code
4575 4523 * does not interfere with internal maintenance tasks.
4576 4524 */
4577 4525 kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4578 4526 minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4579 4527 }
4580 4528
4581 4529 void
4582 4530 kmem_thread_init(void)
4583 4531 {
4584 4532 kmem_move_init();
4585 4533 kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4586 4534 300, INT_MAX, TASKQ_PREPOPULATE);
4587 4535 }
4588 4536
4589 4537 void
4590 4538 kmem_mp_init(void)
4591 4539 {
4592 4540 mutex_enter(&cpu_lock);
4593 4541 register_cpu_setup_func(kmem_cpu_setup, NULL);
4594 4542 mutex_exit(&cpu_lock);
4595 4543
4596 4544 kmem_update_timeout(NULL);
4597 4545
4598 4546 taskq_mp_init();
4599 4547 }
4600 4548
4601 4549 /*
4602 4550 * Return the slab of the allocated buffer, or NULL if the buffer is not
4603 4551 * allocated. This function may be called with a known slab address to determine
4604 4552 * whether or not the buffer is allocated, or with a NULL slab address to obtain
4605 4553 * an allocated buffer's slab.
4606 4554 */
4607 4555 static kmem_slab_t *
4608 4556 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4609 4557 {
4610 4558 kmem_bufctl_t *bcp, *bufbcp;
4611 4559
4612 4560 ASSERT(MUTEX_HELD(&cp->cache_lock));
4613 4561 ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4614 4562
4615 4563 if (cp->cache_flags & KMF_HASH) {
4616 4564 for (bcp = *KMEM_HASH(cp, buf);
4617 4565 (bcp != NULL) && (bcp->bc_addr != buf);
4618 4566 bcp = bcp->bc_next) {
4619 4567 continue;
4620 4568 }
4621 4569 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4622 4570 return (bcp == NULL ? NULL : bcp->bc_slab);
4623 4571 }
4624 4572
4625 4573 if (sp == NULL) {
4626 4574 sp = KMEM_SLAB(cp, buf);
4627 4575 }
4628 4576 bufbcp = KMEM_BUFCTL(cp, buf);
4629 4577 for (bcp = sp->slab_head;
4630 4578 (bcp != NULL) && (bcp != bufbcp);
4631 4579 bcp = bcp->bc_next) {
4632 4580 continue;
4633 4581 }
4634 4582 return (bcp == NULL ? sp : NULL);
4635 4583 }
4636 4584
4637 4585 static boolean_t
4638 4586 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4639 4587 {
4640 4588 long refcnt = sp->slab_refcnt;
4641 4589
4642 4590 ASSERT(cp->cache_defrag != NULL);
4643 4591
4644 4592 /*
4645 4593 * For code coverage we want to be able to move an object within the
4646 4594 * same slab (the only partial slab) even if allocating the destination
4647 4595 * buffer resulted in a completely allocated slab.
4648 4596 */
4649 4597 if (flags & KMM_DEBUG) {
4650 4598 return ((flags & KMM_DESPERATE) ||
4651 4599 ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4652 4600 }
4653 4601
4654 4602 /* If we're desperate, we don't care if the client said NO. */
4655 4603 if (flags & KMM_DESPERATE) {
4656 4604 return (refcnt < sp->slab_chunks); /* any partial */
4657 4605 }
4658 4606
4659 4607 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4660 4608 return (B_FALSE);
4661 4609 }
4662 4610
4663 4611 if ((refcnt == 1) || kmem_move_any_partial) {
4664 4612 return (refcnt < sp->slab_chunks);
4665 4613 }
4666 4614
4667 4615 /*
4668 4616 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4669 4617 * slabs with a progressively higher percentage of used buffers can be
|
↓ open down ↓ |
506 lines elided |
↑ open up ↑ |
4670 4618 * reclaimed until the cache as a whole is no longer fragmented.
4671 4619 *
4672 4620 * sp->slab_refcnt kmd_reclaim_numer
4673 4621 * --------------- < ------------------
4674 4622 * sp->slab_chunks KMEM_VOID_FRACTION
4675 4623 */
4676 4624 return ((refcnt * KMEM_VOID_FRACTION) <
4677 4625 (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4678 4626 }
4679 4627
4680 -static void *
4681 -kmem_hunt_mag(kmem_cache_t *cp, kmem_magazine_t *m, int n, void *buf,
4682 - void *tbuf)
4683 -{
4684 - int i; /* magazine round index */
4685 -
4686 - for (i = 0; i < n; i++) {
4687 - if (buf == m->mag_round[i]) {
4688 - if (cp->cache_flags & KMF_BUFTAG) {
4689 - (void) kmem_cache_free_debug(cp, tbuf,
4690 - caller());
4691 - }
4692 - m->mag_round[i] = tbuf;
4693 - return (buf);
4694 - }
4695 - }
4696 -
4697 - return (NULL);
4698 -}
4699 -
4700 4628 /*
4701 - * Hunt the magazine layer for the given buffer. If found, the buffer is
4702 - * removed from the magazine layer and returned, otherwise NULL is returned.
4703 - * The state of the returned buffer is freed and constructed.
4704 - */
4705 -static void *
4706 -kmem_hunt_mags(kmem_cache_t *cp, void *buf)
4707 -{
4708 - kmem_cpu_cache_t *ccp;
4709 - kmem_magazine_t *m;
4710 - int cpu_seqid;
4711 - int n; /* magazine rounds */
4712 - void *tbuf; /* temporary swap buffer */
4713 -
4714 - ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4715 -
4716 - /*
4717 - * Allocated a buffer to swap with the one we hope to pull out of a
4718 - * magazine when found.
4719 - */
4720 - tbuf = kmem_cache_alloc(cp, KM_NOSLEEP);
4721 - if (tbuf == NULL) {
4722 - KMEM_STAT_ADD(kmem_move_stats.kms_hunt_alloc_fail);
4723 - return (NULL);
4724 - }
4725 - if (tbuf == buf) {
4726 - KMEM_STAT_ADD(kmem_move_stats.kms_hunt_lucky);
4727 - if (cp->cache_flags & KMF_BUFTAG) {
4728 - (void) kmem_cache_free_debug(cp, buf, caller());
4729 - }
4730 - return (buf);
4731 - }
4732 -
4733 - /* Hunt the depot. */
4734 - mutex_enter(&cp->cache_depot_lock);
4735 - n = cp->cache_magtype->mt_magsize;
4736 - for (m = cp->cache_full.ml_list; m != NULL; m = m->mag_next) {
4737 - if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4738 - mutex_exit(&cp->cache_depot_lock);
4739 - return (buf);
4740 - }
4741 - }
4742 - mutex_exit(&cp->cache_depot_lock);
4743 -
4744 - /* Hunt the per-CPU magazines. */
4745 - for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4746 - ccp = &cp->cache_cpu[cpu_seqid];
4747 -
4748 - mutex_enter(&ccp->cc_lock);
4749 - m = ccp->cc_loaded;
4750 - n = ccp->cc_rounds;
4751 - if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4752 - mutex_exit(&ccp->cc_lock);
4753 - return (buf);
4754 - }
4755 - m = ccp->cc_ploaded;
4756 - n = ccp->cc_prounds;
4757 - if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4758 - mutex_exit(&ccp->cc_lock);
4759 - return (buf);
4760 - }
4761 - mutex_exit(&ccp->cc_lock);
4762 - }
4763 -
4764 - kmem_cache_free(cp, tbuf);
4765 - return (NULL);
4766 -}
4767 -
4768 -/*
4769 4629 * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4770 4630 * or when the buffer is freed.
4771 4631 */
4772 4632 static void
4773 4633 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4774 4634 {
4775 4635 ASSERT(MUTEX_HELD(&cp->cache_lock));
4776 4636 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4777 4637
4778 4638 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4779 4639 return;
4780 4640 }
4781 4641
4782 4642 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4783 4643 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4784 4644 avl_remove(&cp->cache_partial_slabs, sp);
4785 4645 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4786 4646 sp->slab_stuck_offset = (uint32_t)-1;
4787 4647 avl_add(&cp->cache_partial_slabs, sp);
4788 4648 }
4789 4649 } else {
4790 4650 sp->slab_later_count = 0;
4791 4651 sp->slab_stuck_offset = (uint32_t)-1;
4792 4652 }
4793 4653 }
4794 4654
4795 4655 static void
4796 4656 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4797 4657 {
4798 4658 ASSERT(taskq_member(kmem_move_taskq, curthread));
4799 4659 ASSERT(MUTEX_HELD(&cp->cache_lock));
4800 4660 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4801 4661
4802 4662 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4803 4663 return;
4804 4664 }
4805 4665
4806 4666 avl_remove(&cp->cache_partial_slabs, sp);
4807 4667 sp->slab_later_count = 0;
4808 4668 sp->slab_flags |= KMEM_SLAB_NOMOVE;
4809 4669 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4810 4670 avl_add(&cp->cache_partial_slabs, sp);
4811 4671 }
4812 4672
4813 4673 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4814 4674
4815 4675 /*
4816 4676 * The move callback takes two buffer addresses, the buffer to be moved, and a
4817 4677 * newly allocated and constructed buffer selected by kmem as the destination.
4818 4678 * It also takes the size of the buffer and an optional user argument specified
4819 4679 * at cache creation time. kmem guarantees that the buffer to be moved has not
4820 4680 * been unmapped by the virtual memory subsystem. Beyond that, it cannot
|
↓ open down ↓ |
42 lines elided |
↑ open up ↑ |
4821 4681 * guarantee the present whereabouts of the buffer to be moved, so it is up to
4822 4682 * the client to safely determine whether or not it is still using the buffer.
4823 4683 * The client must not free either of the buffers passed to the move callback,
4824 4684 * since kmem wants to free them directly to the slab layer. The client response
4825 4685 * tells kmem which of the two buffers to free:
4826 4686 *
4827 4687 * YES kmem frees the old buffer (the move was successful)
4828 4688 * NO kmem frees the new buffer, marks the slab of the old buffer
4829 4689 * non-reclaimable to avoid bothering the client again
4830 4690 * LATER kmem frees the new buffer, increments slab_later_count
4831 - * DONT_KNOW kmem frees the new buffer, searches mags for the old buffer
4691 + * DONT_KNOW kmem frees the new buffer
4832 4692 * DONT_NEED kmem frees both the old buffer and the new buffer
4833 4693 *
4834 4694 * The pending callback argument now being processed contains both of the
4835 4695 * buffers (old and new) passed to the move callback function, the slab of the
4836 4696 * old buffer, and flags related to the move request, such as whether or not the
4837 4697 * system was desperate for memory.
4838 4698 *
4839 4699 * Slabs are not freed while there is a pending callback, but instead are kept
4840 4700 * on a deadlist, which is drained after the last callback completes. This means
4841 4701 * that slabs are safe to access until kmem_move_end(), no matter how many of
4842 4702 * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4843 4703 * zero for as long as the slab remains on the deadlist and until the slab is
4844 4704 * freed.
4845 4705 */
4846 4706 static void
4847 4707 kmem_move_buffer(kmem_move_t *callback)
4848 4708 {
4849 4709 kmem_cbrc_t response;
4850 4710 kmem_slab_t *sp = callback->kmm_from_slab;
4851 4711 kmem_cache_t *cp = sp->slab_cache;
4852 4712 boolean_t free_on_slab;
4853 4713
4854 4714 ASSERT(taskq_member(kmem_move_taskq, curthread));
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
4855 4715 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4856 4716 ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4857 4717
4858 4718 /*
4859 4719 * The number of allocated buffers on the slab may have changed since we
4860 4720 * last checked the slab's reclaimability (when the pending move was
4861 4721 * enqueued), or the client may have responded NO when asked to move
4862 4722 * another buffer on the same slab.
4863 4723 */
4864 4724 if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4865 - KMEM_STAT_ADD(kmem_move_stats.kms_no_longer_reclaimable);
4866 - KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4867 - kmem_move_stats.kms_notify_no_longer_reclaimable);
4868 4725 kmem_slab_free(cp, callback->kmm_to_buf);
4869 4726 kmem_move_end(cp, callback);
4870 4727 return;
4871 4728 }
4872 4729
4873 4730 /*
4874 - * Hunting magazines is expensive, so we'll wait to do that until the
4875 - * client responds KMEM_CBRC_DONT_KNOW. However, checking the slab layer
4876 - * is cheap, so we might as well do that here in case we can avoid
4877 - * bothering the client.
4731 + * Checking the slab layer is easy, so we might as well do that here
4732 + * in case we can avoid bothering the client.
4878 4733 */
4879 4734 mutex_enter(&cp->cache_lock);
4880 4735 free_on_slab = (kmem_slab_allocated(cp, sp,
4881 4736 callback->kmm_from_buf) == NULL);
4882 4737 mutex_exit(&cp->cache_lock);
4883 4738
4884 4739 if (free_on_slab) {
4885 - KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_slab);
4886 4740 kmem_slab_free(cp, callback->kmm_to_buf);
4887 4741 kmem_move_end(cp, callback);
4888 4742 return;
4889 4743 }
4890 4744
4891 4745 if (cp->cache_flags & KMF_BUFTAG) {
4892 4746 /*
4893 4747 * Make kmem_cache_alloc_debug() apply the constructor for us.
4894 4748 */
4895 4749 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4896 4750 KM_NOSLEEP, 1, caller()) != 0) {
4897 - KMEM_STAT_ADD(kmem_move_stats.kms_alloc_fail);
4898 4751 kmem_move_end(cp, callback);
4899 4752 return;
4900 4753 }
4901 4754 } else if (cp->cache_constructor != NULL &&
4902 4755 cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4903 4756 KM_NOSLEEP) != 0) {
4904 4757 atomic_inc_64(&cp->cache_alloc_fail);
4905 - KMEM_STAT_ADD(kmem_move_stats.kms_constructor_fail);
4906 4758 kmem_slab_free(cp, callback->kmm_to_buf);
4907 4759 kmem_move_end(cp, callback);
4908 4760 return;
4909 4761 }
4910 4762
4911 - KMEM_STAT_ADD(kmem_move_stats.kms_callbacks);
4912 - KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4913 - kmem_move_stats.kms_notify_callbacks);
4914 4763 cp->cache_defrag->kmd_callbacks++;
4915 4764 cp->cache_defrag->kmd_thread = curthread;
4916 4765 cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4917 4766 cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4918 4767 DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4919 4768 callback);
4920 4769
4921 4770 response = cp->cache_move(callback->kmm_from_buf,
4922 4771 callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4923 4772
4924 4773 DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4925 4774 callback, kmem_cbrc_t, response);
4926 4775 cp->cache_defrag->kmd_thread = NULL;
4927 4776 cp->cache_defrag->kmd_from_buf = NULL;
4928 4777 cp->cache_defrag->kmd_to_buf = NULL;
4929 4778
4930 4779 if (response == KMEM_CBRC_YES) {
4931 - KMEM_STAT_ADD(kmem_move_stats.kms_yes);
4932 4780 cp->cache_defrag->kmd_yes++;
4933 4781 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4934 4782 /* slab safe to access until kmem_move_end() */
4935 4783 if (sp->slab_refcnt == 0)
4936 4784 cp->cache_defrag->kmd_slabs_freed++;
4937 4785 mutex_enter(&cp->cache_lock);
4938 4786 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4939 4787 mutex_exit(&cp->cache_lock);
4940 4788 kmem_move_end(cp, callback);
4941 4789 return;
4942 4790 }
4943 4791
4944 4792 switch (response) {
4945 4793 case KMEM_CBRC_NO:
4946 - KMEM_STAT_ADD(kmem_move_stats.kms_no);
4947 4794 cp->cache_defrag->kmd_no++;
4948 4795 mutex_enter(&cp->cache_lock);
4949 4796 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4950 4797 mutex_exit(&cp->cache_lock);
4951 4798 break;
4952 4799 case KMEM_CBRC_LATER:
4953 - KMEM_STAT_ADD(kmem_move_stats.kms_later);
4954 4800 cp->cache_defrag->kmd_later++;
4955 4801 mutex_enter(&cp->cache_lock);
4956 4802 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4957 4803 mutex_exit(&cp->cache_lock);
4958 4804 break;
4959 4805 }
4960 4806
4961 4807 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4962 - KMEM_STAT_ADD(kmem_move_stats.kms_disbelief);
4963 4808 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4964 4809 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4965 4810 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4966 4811 callback->kmm_from_buf);
4967 4812 }
4968 4813 mutex_exit(&cp->cache_lock);
4969 4814 break;
4970 4815 case KMEM_CBRC_DONT_NEED:
4971 - KMEM_STAT_ADD(kmem_move_stats.kms_dont_need);
4972 4816 cp->cache_defrag->kmd_dont_need++;
4973 4817 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4974 4818 if (sp->slab_refcnt == 0)
4975 4819 cp->cache_defrag->kmd_slabs_freed++;
4976 4820 mutex_enter(&cp->cache_lock);
4977 4821 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4978 4822 mutex_exit(&cp->cache_lock);
4979 4823 break;
4980 4824 case KMEM_CBRC_DONT_KNOW:
4981 - KMEM_STAT_ADD(kmem_move_stats.kms_dont_know);
4825 + /*
4826 + * If we don't know if we can move this buffer or not, we'll
4827 + * just assume that we can't: if the buffer is in fact free,
4828 + * then it is sitting in one of the per-CPU magazines or in
4829 + * a full magazine in the depot layer. Either way, because
4830 + * defrag is induced in the same logic that reaps a cache,
4831 + * it's likely that full magazines will be returned to the
4832 + * system soon (thereby accomplishing what we're trying to
4833 + * accomplish here: return those magazines to their slabs).
4834 + * Given this, any work that we might do now to locate a buffer
4835 + * in a magazine is wasted (and expensive!) work; we bump
4836 + * a counter in this case and otherwise assume that we can't
4837 + * move it.
4838 + */
4982 4839 cp->cache_defrag->kmd_dont_know++;
4983 - if (kmem_hunt_mags(cp, callback->kmm_from_buf) != NULL) {
4984 - KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_mag);
4985 - cp->cache_defrag->kmd_hunt_found++;
4986 - kmem_slab_free_constructed(cp, callback->kmm_from_buf,
4987 - B_TRUE);
4988 - if (sp->slab_refcnt == 0)
4989 - cp->cache_defrag->kmd_slabs_freed++;
4990 - mutex_enter(&cp->cache_lock);
4991 - kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4992 - mutex_exit(&cp->cache_lock);
4993 - }
4994 4840 break;
4995 4841 default:
4996 4842 panic("'%s' (%p) unexpected move callback response %d\n",
4997 4843 cp->cache_name, (void *)cp, response);
4998 4844 }
4999 4845
5000 4846 kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
5001 4847 kmem_move_end(cp, callback);
5002 4848 }
5003 4849
5004 4850 /* Return B_FALSE if there is insufficient memory for the move request. */
5005 4851 static boolean_t
5006 4852 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
5007 4853 {
|
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
5008 4854 void *to_buf;
5009 4855 avl_index_t index;
5010 4856 kmem_move_t *callback, *pending;
5011 4857 ulong_t n;
5012 4858
5013 4859 ASSERT(taskq_member(kmem_taskq, curthread));
5014 4860 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
5015 4861 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5016 4862
5017 4863 callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
5018 - if (callback == NULL) {
5019 - KMEM_STAT_ADD(kmem_move_stats.kms_callback_alloc_fail);
4864 +
4865 + if (callback == NULL)
5020 4866 return (B_FALSE);
5021 - }
5022 4867
5023 4868 callback->kmm_from_slab = sp;
5024 4869 callback->kmm_from_buf = buf;
5025 4870 callback->kmm_flags = flags;
5026 4871
5027 4872 mutex_enter(&cp->cache_lock);
5028 4873
5029 4874 n = avl_numnodes(&cp->cache_partial_slabs);
5030 4875 if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
5031 4876 mutex_exit(&cp->cache_lock);
5032 4877 kmem_cache_free(kmem_move_cache, callback);
5033 4878 return (B_TRUE); /* there is no need for the move request */
5034 4879 }
5035 4880
|
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
5036 4881 pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
5037 4882 if (pending != NULL) {
5038 4883 /*
5039 4884 * If the move is already pending and we're desperate now,
5040 4885 * update the move flags.
5041 4886 */
5042 4887 if (flags & KMM_DESPERATE) {
5043 4888 pending->kmm_flags |= KMM_DESPERATE;
5044 4889 }
5045 4890 mutex_exit(&cp->cache_lock);
5046 - KMEM_STAT_ADD(kmem_move_stats.kms_already_pending);
5047 4891 kmem_cache_free(kmem_move_cache, callback);
5048 4892 return (B_TRUE);
5049 4893 }
5050 4894
5051 4895 to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
5052 4896 B_FALSE);
5053 4897 callback->kmm_to_buf = to_buf;
5054 4898 avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
5055 4899
5056 4900 mutex_exit(&cp->cache_lock);
5057 4901
5058 4902 if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
5059 4903 callback, TQ_NOSLEEP)) {
5060 - KMEM_STAT_ADD(kmem_move_stats.kms_callback_taskq_fail);
5061 4904 mutex_enter(&cp->cache_lock);
5062 4905 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5063 4906 mutex_exit(&cp->cache_lock);
5064 4907 kmem_slab_free(cp, to_buf);
5065 4908 kmem_cache_free(kmem_move_cache, callback);
5066 4909 return (B_FALSE);
5067 4910 }
5068 4911
5069 4912 return (B_TRUE);
5070 4913 }
5071 4914
5072 4915 static void
5073 4916 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
5074 4917 {
5075 4918 avl_index_t index;
5076 4919
5077 4920 ASSERT(cp->cache_defrag != NULL);
5078 4921 ASSERT(taskq_member(kmem_move_taskq, curthread));
5079 4922 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
5080 4923
5081 4924 mutex_enter(&cp->cache_lock);
5082 4925 VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
5083 4926 callback->kmm_from_buf, &index) != NULL);
5084 4927 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5085 4928 if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
5086 4929 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5087 4930 kmem_slab_t *sp;
5088 4931
5089 4932 /*
5090 4933 * The last pending move completed. Release all slabs from the
5091 4934 * front of the dead list except for any slab at the tail that
5092 4935 * needs to be released from the context of kmem_move_buffers().
5093 4936 * kmem deferred unmapping the buffers on these slabs in order
5094 4937 * to guarantee that buffers passed to the move callback have
5095 4938 * been touched only by kmem or by the client itself.
|
↓ open down ↓ |
25 lines elided |
↑ open up ↑ |
5096 4939 */
5097 4940 while ((sp = list_remove_head(deadlist)) != NULL) {
5098 4941 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
5099 4942 list_insert_tail(deadlist, sp);
5100 4943 break;
5101 4944 }
5102 4945 cp->cache_defrag->kmd_deadcount--;
5103 4946 cp->cache_slab_destroy++;
5104 4947 mutex_exit(&cp->cache_lock);
5105 4948 kmem_slab_destroy(cp, sp);
5106 - KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5107 4949 mutex_enter(&cp->cache_lock);
5108 4950 }
5109 4951 }
5110 4952 mutex_exit(&cp->cache_lock);
5111 4953 kmem_cache_free(kmem_move_cache, callback);
5112 4954 }
5113 4955
5114 4956 /*
5115 4957 * Move buffers from least used slabs first by scanning backwards from the end
5116 4958 * of the partial slab list. Scan at most max_scan candidate slabs and move
5117 4959 * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
5118 4960 * If desperate to reclaim memory, move buffers from any partial slab, otherwise
5119 4961 * skip slabs with a ratio of allocated buffers at or above the current
5120 4962 * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5121 4963 * scan is aborted) so that the caller can adjust the reclaimability threshold
5122 4964 * depending on how many reclaimable slabs it finds.
5123 4965 *
5124 4966 * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5125 4967 * move request, since it is not valid for kmem_move_begin() to call
5126 4968 * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5127 4969 */
5128 4970 static int
5129 4971 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5130 4972 int flags)
5131 4973 {
5132 4974 kmem_slab_t *sp;
5133 4975 void *buf;
5134 4976 int i, j; /* slab index, buffer index */
5135 4977 int s; /* reclaimable slabs */
5136 4978 int b; /* allocated (movable) buffers on reclaimable slab */
5137 4979 boolean_t success;
5138 4980 int refcnt;
5139 4981 int nomove;
5140 4982
5141 4983 ASSERT(taskq_member(kmem_taskq, curthread));
5142 4984 ASSERT(MUTEX_HELD(&cp->cache_lock));
5143 4985 ASSERT(kmem_move_cache != NULL);
5144 4986 ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5145 4987 ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5146 4988 avl_numnodes(&cp->cache_partial_slabs) > 1);
5147 4989
5148 4990 if (kmem_move_blocked) {
5149 4991 return (0);
5150 4992 }
5151 4993
5152 4994 if (kmem_move_fulltilt) {
5153 4995 flags |= KMM_DESPERATE;
5154 4996 }
5155 4997
5156 4998 if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5157 4999 /*
5158 5000 * Scan as many slabs as needed to find the desired number of
5159 5001 * candidate slabs.
5160 5002 */
5161 5003 max_scan = (size_t)-1;
5162 5004 }
5163 5005
5164 5006 if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5165 5007 /* Find as many candidate slabs as possible. */
5166 5008 max_slabs = (size_t)-1;
5167 5009 }
5168 5010
5169 5011 sp = avl_last(&cp->cache_partial_slabs);
5170 5012 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5171 5013 for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5172 5014 ((sp != avl_first(&cp->cache_partial_slabs)) ||
5173 5015 (flags & KMM_DEBUG));
5174 5016 sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5175 5017
5176 5018 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5177 5019 continue;
5178 5020 }
5179 5021 s++;
5180 5022
5181 5023 /* Look for allocated buffers to move. */
5182 5024 for (j = 0, b = 0, buf = sp->slab_base;
5183 5025 (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5184 5026 buf = (((char *)buf) + cp->cache_chunksize), j++) {
5185 5027
5186 5028 if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5187 5029 continue;
5188 5030 }
5189 5031
5190 5032 b++;
5191 5033
5192 5034 /*
5193 5035 * Prevent the slab from being destroyed while we drop
5194 5036 * cache_lock and while the pending move is not yet
5195 5037 * registered. Flag the pending move while
5196 5038 * kmd_moves_pending may still be empty, since we can't
5197 5039 * yet rely on a non-zero pending move count to prevent
5198 5040 * the slab from being destroyed.
5199 5041 */
5200 5042 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5201 5043 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5202 5044 /*
5203 5045 * Recheck refcnt and nomove after reacquiring the lock,
5204 5046 * since these control the order of partial slabs, and
5205 5047 * we want to know if we can pick up the scan where we
5206 5048 * left off.
5207 5049 */
5208 5050 refcnt = sp->slab_refcnt;
5209 5051 nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5210 5052 mutex_exit(&cp->cache_lock);
5211 5053
5212 5054 success = kmem_move_begin(cp, sp, buf, flags);
5213 5055
5214 5056 /*
5215 5057 * Now, before the lock is reacquired, kmem could
5216 5058 * process all pending move requests and purge the
5217 5059 * deadlist, so that upon reacquiring the lock, sp has
5218 5060 * been remapped. Or, the client may free all the
5219 5061 * objects on the slab while the pending moves are still
5220 5062 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5221 5063 * flag causes the slab to be put at the end of the
5222 5064 * deadlist and prevents it from being destroyed, since
5223 5065 * we plan to destroy it here after reacquiring the
5224 5066 * lock.
5225 5067 */
5226 5068 mutex_enter(&cp->cache_lock);
5227 5069 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5228 5070 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5229 5071
5230 5072 if (sp->slab_refcnt == 0) {
5231 5073 list_t *deadlist =
5232 5074 &cp->cache_defrag->kmd_deadlist;
5233 5075 list_remove(deadlist, sp);
5234 5076
5235 5077 if (!avl_is_empty(
5236 5078 &cp->cache_defrag->kmd_moves_pending)) {
5237 5079 /*
5238 5080 * A pending move makes it unsafe to
5239 5081 * destroy the slab, because even though
5240 5082 * the move is no longer needed, the
|
↓ open down ↓ |
124 lines elided |
↑ open up ↑ |
5241 5083 * context where that is determined
5242 5084 * requires the slab to exist.
5243 5085 * Fortunately, a pending move also
5244 5086 * means we don't need to destroy the
5245 5087 * slab here, since it will get
5246 5088 * destroyed along with any other slabs
5247 5089 * on the deadlist after the last
5248 5090 * pending move completes.
5249 5091 */
5250 5092 list_insert_head(deadlist, sp);
5251 - KMEM_STAT_ADD(kmem_move_stats.
5252 - kms_endscan_slab_dead);
5253 5093 return (-1);
5254 5094 }
5255 5095
5256 5096 /*
5257 5097 * Destroy the slab now if it was completely
5258 5098 * freed while we dropped cache_lock and there
5259 5099 * are no pending moves. Since slab_refcnt
5260 5100 * cannot change once it reaches zero, no new
5261 5101 * pending moves from that slab are possible.
5262 5102 */
5263 5103 cp->cache_defrag->kmd_deadcount--;
5264 5104 cp->cache_slab_destroy++;
5265 5105 mutex_exit(&cp->cache_lock);
5266 5106 kmem_slab_destroy(cp, sp);
5267 - KMEM_STAT_ADD(kmem_move_stats.
5268 - kms_dead_slabs_freed);
5269 - KMEM_STAT_ADD(kmem_move_stats.
5270 - kms_endscan_slab_destroyed);
5271 5107 mutex_enter(&cp->cache_lock);
5272 5108 /*
5273 5109 * Since we can't pick up the scan where we left
5274 5110 * off, abort the scan and say nothing about the
5275 5111 * number of reclaimable slabs.
5276 5112 */
5277 5113 return (-1);
5278 5114 }
5279 5115
5280 5116 if (!success) {
5281 5117 /*
5282 5118 * Abort the scan if there is not enough memory
5283 5119 * for the request and say nothing about the
5284 5120 * number of reclaimable slabs.
5285 5121 */
5286 - KMEM_STAT_COND_ADD(s < max_slabs,
5287 - kmem_move_stats.kms_endscan_nomem);
5288 5122 return (-1);
5289 5123 }
5290 5124
5291 5125 /*
5292 5126 * The slab's position changed while the lock was
5293 5127 * dropped, so we don't know where we are in the
5294 5128 * sequence any more.
5295 5129 */
5296 5130 if (sp->slab_refcnt != refcnt) {
5297 5131 /*
5298 5132 * If this is a KMM_DEBUG move, the slab_refcnt
5299 5133 * may have changed because we allocated a
5300 5134 * destination buffer on the same slab. In that
5301 5135 * case, we're not interested in counting it.
5302 5136 */
5303 - KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5304 - (s < max_slabs),
5305 - kmem_move_stats.kms_endscan_refcnt_changed);
5306 5137 return (-1);
5307 5138 }
5308 - if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove) {
5309 - KMEM_STAT_COND_ADD(s < max_slabs,
5310 - kmem_move_stats.kms_endscan_nomove_changed);
5139 + if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove)
5311 5140 return (-1);
5312 - }
5313 5141
5314 5142 /*
5315 5143 * Generating a move request allocates a destination
5316 5144 * buffer from the slab layer, bumping the first partial
5317 5145 * slab if it is completely allocated. If the current
5318 5146 * slab becomes the first partial slab as a result, we
5319 5147 * can't continue to scan backwards.
5320 5148 *
5321 5149 * If this is a KMM_DEBUG move and we allocated the
5322 5150 * destination buffer from the last partial slab, then
5323 5151 * the buffer we're moving is on the same slab and our
5324 5152 * slab_refcnt has changed, causing us to return before
5325 5153 * reaching here if there are no partial slabs left.
5326 5154 */
5327 5155 ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5328 5156 if (sp == avl_first(&cp->cache_partial_slabs)) {
|
↓ open down ↓ |
6 lines elided |
↑ open up ↑ |
5329 5157 /*
5330 5158 * We're not interested in a second KMM_DEBUG
5331 5159 * move.
5332 5160 */
5333 5161 goto end_scan;
5334 5162 }
5335 5163 }
5336 5164 }
5337 5165 end_scan:
5338 5166
5339 - KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5340 - (s < max_slabs) &&
5341 - (sp == avl_first(&cp->cache_partial_slabs)),
5342 - kmem_move_stats.kms_endscan_freelist);
5343 -
5344 5167 return (s);
5345 5168 }
5346 5169
5347 5170 typedef struct kmem_move_notify_args {
5348 5171 kmem_cache_t *kmna_cache;
5349 5172 void *kmna_buf;
5350 5173 } kmem_move_notify_args_t;
5351 5174
5352 5175 static void
5353 5176 kmem_cache_move_notify_task(void *arg)
5354 5177 {
5355 5178 kmem_move_notify_args_t *args = arg;
5356 5179 kmem_cache_t *cp = args->kmna_cache;
5357 5180 void *buf = args->kmna_buf;
5358 5181 kmem_slab_t *sp;
5359 5182
5360 5183 ASSERT(taskq_member(kmem_taskq, curthread));
5361 5184 ASSERT(list_link_active(&cp->cache_link));
5362 5185
5363 5186 kmem_free(args, sizeof (kmem_move_notify_args_t));
5364 5187 mutex_enter(&cp->cache_lock);
5365 5188 sp = kmem_slab_allocated(cp, NULL, buf);
5366 5189
5367 5190 /* Ignore the notification if the buffer is no longer allocated. */
5368 5191 if (sp == NULL) {
5369 5192 mutex_exit(&cp->cache_lock);
5370 5193 return;
5371 5194 }
5372 5195
5373 5196 /* Ignore the notification if there's no reason to move the buffer. */
5374 5197 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5375 5198 /*
5376 5199 * So far the notification is not ignored. Ignore the
5377 5200 * notification if the slab is not marked by an earlier refusal
5378 5201 * to move a buffer.
5379 5202 */
5380 5203 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5381 5204 (sp->slab_later_count == 0)) {
5382 5205 mutex_exit(&cp->cache_lock);
5383 5206 return;
5384 5207 }
5385 5208
5386 5209 kmem_slab_move_yes(cp, sp, buf);
5387 5210 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5388 5211 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5389 5212 mutex_exit(&cp->cache_lock);
5390 5213 /* see kmem_move_buffers() about dropping the lock */
5391 5214 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5392 5215 mutex_enter(&cp->cache_lock);
|
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
5393 5216 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5394 5217 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5395 5218 if (sp->slab_refcnt == 0) {
5396 5219 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5397 5220 list_remove(deadlist, sp);
5398 5221
5399 5222 if (!avl_is_empty(
5400 5223 &cp->cache_defrag->kmd_moves_pending)) {
5401 5224 list_insert_head(deadlist, sp);
5402 5225 mutex_exit(&cp->cache_lock);
5403 - KMEM_STAT_ADD(kmem_move_stats.
5404 - kms_notify_slab_dead);
5405 5226 return;
5406 5227 }
5407 5228
5408 5229 cp->cache_defrag->kmd_deadcount--;
5409 5230 cp->cache_slab_destroy++;
5410 5231 mutex_exit(&cp->cache_lock);
5411 5232 kmem_slab_destroy(cp, sp);
5412 - KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5413 - KMEM_STAT_ADD(kmem_move_stats.
5414 - kms_notify_slab_destroyed);
5415 5233 return;
5416 5234 }
5417 5235 } else {
5418 5236 kmem_slab_move_yes(cp, sp, buf);
5419 5237 }
5420 5238 mutex_exit(&cp->cache_lock);
5421 5239 }
5422 5240
5423 5241 void
5424 5242 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5425 5243 {
5426 5244 kmem_move_notify_args_t *args;
5427 5245
5428 - KMEM_STAT_ADD(kmem_move_stats.kms_notify);
5429 5246 args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5430 5247 if (args != NULL) {
5431 5248 args->kmna_cache = cp;
5432 5249 args->kmna_buf = buf;
5433 5250 if (!taskq_dispatch(kmem_taskq,
5434 5251 (task_func_t *)kmem_cache_move_notify_task, args,
5435 5252 TQ_NOSLEEP))
5436 5253 kmem_free(args, sizeof (kmem_move_notify_args_t));
5437 5254 }
5438 5255 }
5439 5256
5440 5257 static void
|
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
5441 5258 kmem_cache_defrag(kmem_cache_t *cp)
5442 5259 {
5443 5260 size_t n;
5444 5261
5445 5262 ASSERT(cp->cache_defrag != NULL);
5446 5263
5447 5264 mutex_enter(&cp->cache_lock);
5448 5265 n = avl_numnodes(&cp->cache_partial_slabs);
5449 5266 if (n > 1) {
5450 5267 /* kmem_move_buffers() drops and reacquires cache_lock */
5451 - KMEM_STAT_ADD(kmem_move_stats.kms_defrags);
5452 5268 cp->cache_defrag->kmd_defrags++;
5453 5269 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5454 5270 }
5455 5271 mutex_exit(&cp->cache_lock);
5456 5272 }
5457 5273
5458 5274 /* Is this cache above the fragmentation threshold? */
5459 5275 static boolean_t
5460 5276 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5461 5277 {
5462 5278 /*
5463 5279 * nfree kmem_frag_numer
5464 5280 * ------------------ > ---------------
5465 5281 * cp->cache_buftotal kmem_frag_denom
5466 5282 */
5467 5283 return ((nfree * kmem_frag_denom) >
5468 5284 (cp->cache_buftotal * kmem_frag_numer));
5469 5285 }
5470 5286
5471 5287 static boolean_t
5472 5288 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5473 5289 {
5474 5290 boolean_t fragmented;
5475 5291 uint64_t nfree;
5476 5292
5477 5293 ASSERT(MUTEX_HELD(&cp->cache_lock));
5478 5294 *doreap = B_FALSE;
5479 5295
5480 5296 if (kmem_move_fulltilt) {
5481 5297 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5482 5298 return (B_TRUE);
5483 5299 }
5484 5300 } else {
5485 5301 if ((cp->cache_complete_slab_count + avl_numnodes(
5486 5302 &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5487 5303 return (B_FALSE);
5488 5304 }
5489 5305 }
5490 5306
5491 5307 nfree = cp->cache_bufslab;
5492 5308 fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5493 5309 kmem_cache_frag_threshold(cp, nfree));
5494 5310
5495 5311 /*
5496 5312 * Free buffers in the magazine layer appear allocated from the point of
5497 5313 * view of the slab layer. We want to know if the slab layer would
5498 5314 * appear fragmented if we included free buffers from magazines that
5499 5315 * have fallen out of the working set.
5500 5316 */
5501 5317 if (!fragmented) {
5502 5318 long reap;
5503 5319
5504 5320 mutex_enter(&cp->cache_depot_lock);
5505 5321 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5506 5322 reap = MIN(reap, cp->cache_full.ml_total);
5507 5323 mutex_exit(&cp->cache_depot_lock);
5508 5324
5509 5325 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5510 5326 if (kmem_cache_frag_threshold(cp, nfree)) {
5511 5327 *doreap = B_TRUE;
5512 5328 }
5513 5329 }
5514 5330
5515 5331 return (fragmented);
5516 5332 }
5517 5333
5518 5334 /* Called periodically from kmem_taskq */
5519 5335 static void
5520 5336 kmem_cache_scan(kmem_cache_t *cp)
5521 5337 {
5522 5338 boolean_t reap = B_FALSE;
5523 5339 kmem_defrag_t *kmd;
5524 5340
5525 5341 ASSERT(taskq_member(kmem_taskq, curthread));
5526 5342
5527 5343 mutex_enter(&cp->cache_lock);
5528 5344
5529 5345 kmd = cp->cache_defrag;
5530 5346 if (kmd->kmd_consolidate > 0) {
5531 5347 kmd->kmd_consolidate--;
5532 5348 mutex_exit(&cp->cache_lock);
5533 5349 kmem_cache_reap(cp);
5534 5350 return;
5535 5351 }
5536 5352
5537 5353 if (kmem_cache_is_fragmented(cp, &reap)) {
5538 5354 size_t slabs_found;
5539 5355
|
↓ open down ↓ |
78 lines elided |
↑ open up ↑ |
5540 5356 /*
5541 5357 * Consolidate reclaimable slabs from the end of the partial
5542 5358 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5543 5359 * reclaimable slabs). Keep track of how many candidate slabs we
5544 5360 * looked for and how many we actually found so we can adjust
5545 5361 * the definition of a candidate slab if we're having trouble
5546 5362 * finding them.
5547 5363 *
5548 5364 * kmem_move_buffers() drops and reacquires cache_lock.
5549 5365 */
5550 - KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5551 5366 kmd->kmd_scans++;
5552 5367 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5553 5368 kmem_reclaim_max_slabs, 0);
5554 5369 if (slabs_found >= 0) {
5555 5370 kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5556 5371 kmd->kmd_slabs_found += slabs_found;
5557 5372 }
5558 5373
5559 5374 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5560 5375 kmd->kmd_tries = 0;
5561 5376
5562 5377 /*
5563 5378 * If we had difficulty finding candidate slabs in
5564 5379 * previous scans, adjust the threshold so that
5565 5380 * candidates are easier to find.
5566 5381 */
5567 5382 if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5568 5383 kmem_adjust_reclaim_threshold(kmd, -1);
5569 5384 } else if ((kmd->kmd_slabs_found * 2) <
5570 5385 kmd->kmd_slabs_sought) {
5571 5386 kmem_adjust_reclaim_threshold(kmd, 1);
5572 5387 }
5573 5388 kmd->kmd_slabs_sought = 0;
5574 5389 kmd->kmd_slabs_found = 0;
5575 5390 }
5576 5391 } else {
5577 5392 kmem_reset_reclaim_threshold(cp->cache_defrag);
5578 5393 #ifdef DEBUG
5579 5394 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5580 5395 /*
|
↓ open down ↓ |
20 lines elided |
↑ open up ↑ |
5581 5396 * In a debug kernel we want the consolidator to
5582 5397 * run occasionally even when there is plenty of
5583 5398 * memory.
5584 5399 */
5585 5400 uint16_t debug_rand;
5586 5401
5587 5402 (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5588 5403 if (!kmem_move_noreap &&
5589 5404 ((debug_rand % kmem_mtb_reap) == 0)) {
5590 5405 mutex_exit(&cp->cache_lock);
5591 - KMEM_STAT_ADD(kmem_move_stats.kms_debug_reaps);
5592 5406 kmem_cache_reap(cp);
5593 5407 return;
5594 5408 } else if ((debug_rand % kmem_mtb_move) == 0) {
5595 - KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5596 - KMEM_STAT_ADD(kmem_move_stats.kms_debug_scans);
5597 5409 kmd->kmd_scans++;
5598 5410 (void) kmem_move_buffers(cp,
5599 5411 kmem_reclaim_scan_range, 1, KMM_DEBUG);
5600 5412 }
5601 5413 }
5602 5414 #endif /* DEBUG */
5603 5415 }
5604 5416
5605 5417 mutex_exit(&cp->cache_lock);
5606 5418
5607 - if (reap) {
5608 - KMEM_STAT_ADD(kmem_move_stats.kms_scan_depot_ws_reaps);
5419 + if (reap)
5609 5420 kmem_depot_ws_reap(cp);
5610 - }
5611 5421 }
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX