Print this page
OS-4162 system appears (mostly) hung on kmem_move taskq
| 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, 2016 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
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
162 162 * unused copy destination) and searches for the old object in the
163 163 * magazine layer. If found, the object is removed from the magazine
164 164 * layer and freed to the slab layer so it will no longer hold the
165 165 * slab hostage.
166 166 *
167 167 * 2.3 Object States
168 168 *
169 169 * Neither kmem nor the client can be assumed to know the object's whereabouts
170 170 * at the time of the callback. An object belonging to a kmem cache may be in
171 171 * any of the following states:
172 172 *
173 173 * 1. Uninitialized on the slab
174 174 * 2. Allocated from the slab but not constructed (still uninitialized)
175 175 * 3. Allocated from the slab, constructed, but not yet ready for business
176 176 * (not in a valid state for the move callback)
177 177 * 4. In use (valid and known to the client)
178 178 * 5. About to be freed (no longer in a valid state for the move callback)
179 179 * 6. Freed to a magazine (still constructed)
180 180 * 7. Allocated from a magazine, not yet ready for business (not in a valid
181 181 * state for the move callback), and about to return to state #4
182 182 * 8. Deconstructed on a magazine that is about to be freed
183 183 * 9. Freed to the slab
184 184 *
185 185 * Since the move callback may be called at any time while the object is in any
186 186 * of the above states (except state #1), the client needs a safe way to
187 187 * determine whether or not it knows about the object. Specifically, the client
188 188 * needs to know whether or not the object is in state #4, the only state in
189 189 * which a move is valid. If the object is in any other state, the client should
190 190 * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
191 191 * the object's fields.
192 192 *
193 193 * Note that although an object may be in state #4 when kmem initiates the move
194 194 * request, the object may no longer be in that state by the time kmem actually
195 195 * calls the move function. Not only does the client free objects
196 196 * asynchronously, kmem itself puts move requests on a queue where thay are
197 197 * pending until kmem processes them from another context. Also, objects freed
198 198 * to a magazine appear allocated from the point of view of the slab layer, so
199 199 * kmem may even initiate requests for objects in a state other than state #4.
200 200 *
201 201 * 2.3.1 Magazine Layer
202 202 *
203 203 * An important insight revealed by the states listed above is that the magazine
204 204 * layer is populated only by kmem_cache_free(). Magazines of constructed
205 205 * objects are never populated directly from the slab layer (which contains raw,
206 206 * unconstructed objects). Whenever an allocation request cannot be satisfied
207 207 * from the magazine layer, the magazines are bypassed and the request is
208 208 * satisfied from the slab layer (creating a new slab if necessary). kmem calls
209 209 * the object constructor only when allocating from the slab layer, and only in
210 210 * response to kmem_cache_alloc() or to prepare the destination buffer passed in
211 211 * the move callback. kmem does not preconstruct objects in anticipation of
212 212 * kmem_cache_alloc().
213 213 *
214 214 * 2.3.2 Object Constructor and Destructor
215 215 *
216 216 * If the client supplies a destructor, it must be valid to call the destructor
217 217 * on a newly created object (immediately after the constructor).
218 218 *
219 219 * 2.4 Recognizing Known Objects
220 220 *
221 221 * There is a simple test to determine safely whether or not the client knows
222 222 * about a given object in the move callback. It relies on the fact that kmem
223 223 * guarantees that the object of the move callback has only been touched by the
224 224 * client itself or else by kmem. kmem does this by ensuring that none of the
225 225 * cache's slabs are freed to the virtual memory (VM) subsystem while a move
226 226 * callback is pending. When the last object on a slab is freed, if there is a
227 227 * pending move, kmem puts the slab on a per-cache dead list and defers freeing
228 228 * slabs on that list until all pending callbacks are completed. That way,
229 229 * clients can be certain that the object of a move callback is in one of the
230 230 * states listed above, making it possible to distinguish known objects (in
231 231 * state #4) using the two low order bits of any pointer member (with the
232 232 * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
233 233 * platforms).
234 234 *
235 235 * The test works as long as the client always transitions objects from state #4
236 236 * (known, in use) to state #5 (about to be freed, invalid) by setting the low
237 237 * order bit of the client-designated pointer member. Since kmem only writes
238 238 * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
239 239 * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
240 240 * guaranteed to set at least one of the two low order bits. Therefore, given an
241 241 * object with a back pointer to a 'container_t *o_container', the client can
242 242 * test
243 243 *
244 244 * container_t *container = object->o_container;
245 245 * if ((uintptr_t)container & 0x3) {
246 246 * return (KMEM_CBRC_DONT_KNOW);
247 247 * }
248 248 *
249 249 * Typically, an object will have a pointer to some structure with a list or
250 250 * hash where objects from the cache are kept while in use. Assuming that the
251 251 * client has some way of knowing that the container structure is valid and will
252 252 * not go away during the move, and assuming that the structure includes a lock
253 253 * to protect whatever collection is used, then the client would continue as
254 254 * follows:
255 255 *
256 256 * // Ensure that the container structure does not go away.
257 257 * if (container_hold(container) == 0) {
258 258 * return (KMEM_CBRC_DONT_KNOW);
259 259 * }
260 260 * mutex_enter(&container->c_objects_lock);
261 261 * if (container != object->o_container) {
262 262 * mutex_exit(&container->c_objects_lock);
263 263 * container_rele(container);
264 264 * return (KMEM_CBRC_DONT_KNOW);
265 265 * }
266 266 *
267 267 * At this point the client knows that the object cannot be freed as long as
268 268 * c_objects_lock is held. Note that after acquiring the lock, the client must
269 269 * recheck the o_container pointer in case the object was removed just before
270 270 * acquiring the lock.
271 271 *
272 272 * When the client is about to free an object, it must first remove that object
273 273 * from the list, hash, or other structure where it is kept. At that time, to
274 274 * mark the object so it can be distinguished from the remaining, known objects,
275 275 * the client sets the designated low order bit:
276 276 *
277 277 * mutex_enter(&container->c_objects_lock);
278 278 * object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
279 279 * list_remove(&container->c_objects, object);
280 280 * mutex_exit(&container->c_objects_lock);
281 281 *
282 282 * In the common case, the object is freed to the magazine layer, where it may
283 283 * be reused on a subsequent allocation without the overhead of calling the
284 284 * constructor. While in the magazine it appears allocated from the point of
285 285 * view of the slab layer, making it a candidate for the move callback. Most
286 286 * objects unrecognized by the client in the move callback fall into this
287 287 * category and are cheaply distinguished from known objects by the test
288 288 * described earlier. Since recognition is cheap for the client, and searching
289 289 * magazines is expensive for kmem, kmem defers searching until the client first
290 290 * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
291 291 * elsewhere does what it can to avoid bothering the client unnecessarily.
292 292 *
293 293 * Invalidating the designated pointer member before freeing the object marks
294 294 * the object to be avoided in the callback, and conversely, assigning a valid
295 295 * value to the designated pointer member after allocating the object makes the
296 296 * object fair game for the callback:
297 297 *
298 298 * ... allocate object ...
299 299 * ... set any initial state not set by the constructor ...
300 300 *
301 301 * mutex_enter(&container->c_objects_lock);
302 302 * list_insert_tail(&container->c_objects, object);
303 303 * membar_producer();
304 304 * object->o_container = container;
305 305 * mutex_exit(&container->c_objects_lock);
306 306 *
307 307 * Note that everything else must be valid before setting o_container makes the
308 308 * object fair game for the move callback. The membar_producer() call ensures
309 309 * that all the object's state is written to memory before setting the pointer
310 310 * that transitions the object from state #3 or #7 (allocated, constructed, not
311 311 * yet in use) to state #4 (in use, valid). That's important because the move
312 312 * function has to check the validity of the pointer before it can safely
313 313 * acquire the lock protecting the collection where it expects to find known
314 314 * objects.
315 315 *
316 316 * This method of distinguishing known objects observes the usual symmetry:
317 317 * invalidating the designated pointer is the first thing the client does before
318 318 * freeing the object, and setting the designated pointer is the last thing the
319 319 * client does after allocating the object. Of course, the client is not
320 320 * required to use this method. Fundamentally, how the client recognizes known
321 321 * objects is completely up to the client, but this method is recommended as an
322 322 * efficient and safe way to take advantage of the guarantees made by kmem. If
323 323 * the entire object is arbitrary data without any markable bits from a suitable
324 324 * pointer member, then the client must find some other method, such as
325 325 * searching a hash table of known objects.
326 326 *
327 327 * 2.5 Preventing Objects From Moving
328 328 *
329 329 * Besides a way to distinguish known objects, the other thing that the client
330 330 * needs is a strategy to ensure that an object will not move while the client
331 331 * is actively using it. The details of satisfying this requirement tend to be
332 332 * highly cache-specific. It might seem that the same rules that let a client
333 333 * remove an object safely should also decide when an object can be moved
334 334 * safely. However, any object state that makes a removal attempt invalid is
335 335 * likely to be long-lasting for objects that the client does not expect to
336 336 * remove. kmem knows nothing about the object state and is equally likely (from
337 337 * the client's point of view) to request a move for any object in the cache,
338 338 * whether prepared for removal or not. Even a low percentage of objects stuck
339 339 * in place by unremovability will defeat the consolidator if the stuck objects
340 340 * are the same long-lived allocations likely to hold slabs hostage.
341 341 * Fundamentally, the consolidator is not aimed at common cases. Severe external
342 342 * fragmentation is a worst case scenario manifested as sparsely allocated
343 343 * slabs, by definition a low percentage of the cache's objects. When deciding
344 344 * what makes an object movable, keep in mind the goal of the consolidator: to
345 345 * bring worst-case external fragmentation within the limits guaranteed for
346 346 * internal fragmentation. Removability is a poor criterion if it is likely to
347 347 * exclude more than an insignificant percentage of objects for long periods of
348 348 * time.
349 349 *
350 350 * A tricky general solution exists, and it has the advantage of letting you
351 351 * move any object at almost any moment, practically eliminating the likelihood
352 352 * that an object can hold a slab hostage. However, if there is a cache-specific
353 353 * way to ensure that an object is not actively in use in the vast majority of
354 354 * cases, a simpler solution that leverages this cache-specific knowledge is
355 355 * preferred.
356 356 *
357 357 * 2.5.1 Cache-Specific Solution
358 358 *
359 359 * As an example of a cache-specific solution, the ZFS znode cache takes
360 360 * advantage of the fact that the vast majority of znodes are only being
361 361 * referenced from the DNLC. (A typical case might be a few hundred in active
362 362 * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
363 363 * client has established that it recognizes the znode and can access its fields
364 364 * safely (using the method described earlier), it then tests whether the znode
365 365 * is referenced by anything other than the DNLC. If so, it assumes that the
366 366 * znode may be in active use and is unsafe to move, so it drops its locks and
367 367 * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
368 368 * else znodes are used, no change is needed to protect against the possibility
369 369 * of the znode moving. The disadvantage is that it remains possible for an
370 370 * application to hold a znode slab hostage with an open file descriptor.
371 371 * However, this case ought to be rare and the consolidator has a way to deal
372 372 * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
373 373 * object, kmem eventually stops believing it and treats the slab as if the
374 374 * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
375 375 * then focus on getting it off of the partial slab list by allocating rather
376 376 * than freeing all of its objects. (Either way of getting a slab off the
377 377 * free list reduces fragmentation.)
378 378 *
379 379 * 2.5.2 General Solution
380 380 *
381 381 * The general solution, on the other hand, requires an explicit hold everywhere
382 382 * the object is used to prevent it from moving. To keep the client locking
383 383 * strategy as uncomplicated as possible, kmem guarantees the simplifying
384 384 * assumption that move callbacks are sequential, even across multiple caches.
385 385 * Internally, a global queue processed by a single thread supports all caches
386 386 * implementing the callback function. No matter how many caches supply a move
387 387 * function, the consolidator never moves more than one object at a time, so the
388 388 * client does not have to worry about tricky lock ordering involving several
389 389 * related objects from different kmem caches.
390 390 *
391 391 * The general solution implements the explicit hold as a read-write lock, which
392 392 * allows multiple readers to access an object from the cache simultaneously
393 393 * while a single writer is excluded from moving it. A single rwlock for the
394 394 * entire cache would lock out all threads from using any of the cache's objects
395 395 * even though only a single object is being moved, so to reduce contention,
396 396 * the client can fan out the single rwlock into an array of rwlocks hashed by
397 397 * the object address, making it probable that moving one object will not
398 398 * prevent other threads from using a different object. The rwlock cannot be a
399 399 * member of the object itself, because the possibility of the object moving
400 400 * makes it unsafe to access any of the object's fields until the lock is
401 401 * acquired.
402 402 *
403 403 * Assuming a small, fixed number of locks, it's possible that multiple objects
404 404 * will hash to the same lock. A thread that needs to use multiple objects in
405 405 * the same function may acquire the same lock multiple times. Since rwlocks are
406 406 * reentrant for readers, and since there is never more than a single writer at
407 407 * a time (assuming that the client acquires the lock as a writer only when
408 408 * moving an object inside the callback), there would seem to be no problem.
409 409 * However, a client locking multiple objects in the same function must handle
410 410 * one case of potential deadlock: Assume that thread A needs to prevent both
411 411 * object 1 and object 2 from moving, and thread B, the callback, meanwhile
412 412 * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
413 413 * same lock, that thread A will acquire the lock for object 1 as a reader
414 414 * before thread B sets the lock's write-wanted bit, preventing thread A from
415 415 * reacquiring the lock for object 2 as a reader. Unable to make forward
416 416 * progress, thread A will never release the lock for object 1, resulting in
417 417 * deadlock.
418 418 *
419 419 * There are two ways of avoiding the deadlock just described. The first is to
420 420 * use rw_tryenter() rather than rw_enter() in the callback function when
421 421 * attempting to acquire the lock as a writer. If tryenter discovers that the
422 422 * same object (or another object hashed to the same lock) is already in use, it
423 423 * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
424 424 * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
425 425 * since it allows a thread to acquire the lock as a reader in spite of a
426 426 * waiting writer. This second approach insists on moving the object now, no
427 427 * matter how many readers the move function must wait for in order to do so,
428 428 * and could delay the completion of the callback indefinitely (blocking
429 429 * callbacks to other clients). In practice, a less insistent callback using
430 430 * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
431 431 * little reason to use anything else.
432 432 *
433 433 * Avoiding deadlock is not the only problem that an implementation using an
434 434 * explicit hold needs to solve. Locking the object in the first place (to
435 435 * prevent it from moving) remains a problem, since the object could move
436 436 * between the time you obtain a pointer to the object and the time you acquire
437 437 * the rwlock hashed to that pointer value. Therefore the client needs to
438 438 * recheck the value of the pointer after acquiring the lock, drop the lock if
439 439 * the value has changed, and try again. This requires a level of indirection:
440 440 * something that points to the object rather than the object itself, that the
441 441 * client can access safely while attempting to acquire the lock. (The object
442 442 * itself cannot be referenced safely because it can move at any time.)
443 443 * The following lock-acquisition function takes whatever is safe to reference
444 444 * (arg), follows its pointer to the object (using function f), and tries as
445 445 * often as necessary to acquire the hashed lock and verify that the object
446 446 * still has not moved:
447 447 *
448 448 * object_t *
449 449 * object_hold(object_f f, void *arg)
450 450 * {
451 451 * object_t *op;
452 452 *
453 453 * op = f(arg);
454 454 * if (op == NULL) {
455 455 * return (NULL);
456 456 * }
457 457 *
458 458 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
459 459 * while (op != f(arg)) {
460 460 * rw_exit(OBJECT_RWLOCK(op));
461 461 * op = f(arg);
462 462 * if (op == NULL) {
463 463 * break;
464 464 * }
465 465 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
466 466 * }
467 467 *
468 468 * return (op);
469 469 * }
470 470 *
471 471 * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
472 472 * lock reacquisition loop, while necessary, almost never executes. The function
473 473 * pointer f (used to obtain the object pointer from arg) has the following type
474 474 * definition:
475 475 *
476 476 * typedef object_t *(*object_f)(void *arg);
477 477 *
478 478 * An object_f implementation is likely to be as simple as accessing a structure
479 479 * member:
480 480 *
481 481 * object_t *
482 482 * s_object(void *arg)
483 483 * {
484 484 * something_t *sp = arg;
485 485 * return (sp->s_object);
486 486 * }
487 487 *
488 488 * The flexibility of a function pointer allows the path to the object to be
489 489 * arbitrarily complex and also supports the notion that depending on where you
490 490 * are using the object, you may need to get it from someplace different.
491 491 *
492 492 * The function that releases the explicit hold is simpler because it does not
493 493 * have to worry about the object moving:
494 494 *
495 495 * void
496 496 * object_rele(object_t *op)
497 497 * {
498 498 * rw_exit(OBJECT_RWLOCK(op));
499 499 * }
500 500 *
501 501 * The caller is spared these details so that obtaining and releasing an
502 502 * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
503 503 * of object_hold() only needs to know that the returned object pointer is valid
504 504 * if not NULL and that the object will not move until released.
505 505 *
506 506 * Although object_hold() prevents an object from moving, it does not prevent it
507 507 * from being freed. The caller must take measures before calling object_hold()
508 508 * (afterwards is too late) to ensure that the held object cannot be freed. The
509 509 * caller must do so without accessing the unsafe object reference, so any lock
510 510 * or reference count used to ensure the continued existence of the object must
511 511 * live outside the object itself.
512 512 *
513 513 * Obtaining a new object is a special case where an explicit hold is impossible
514 514 * for the caller. Any function that returns a newly allocated object (either as
515 515 * a return value, or as an in-out paramter) must return it already held; after
516 516 * the caller gets it is too late, since the object cannot be safely accessed
517 517 * without the level of indirection described earlier. The following
518 518 * object_alloc() example uses the same code shown earlier to transition a new
519 519 * object into the state of being recognized (by the client) as a known object.
520 520 * The function must acquire the hold (rw_enter) before that state transition
521 521 * makes the object movable:
522 522 *
523 523 * static object_t *
524 524 * object_alloc(container_t *container)
525 525 * {
526 526 * object_t *object = kmem_cache_alloc(object_cache, 0);
527 527 * ... set any initial state not set by the constructor ...
528 528 * rw_enter(OBJECT_RWLOCK(object), RW_READER);
529 529 * mutex_enter(&container->c_objects_lock);
530 530 * list_insert_tail(&container->c_objects, object);
531 531 * membar_producer();
532 532 * object->o_container = container;
533 533 * mutex_exit(&container->c_objects_lock);
534 534 * return (object);
535 535 * }
536 536 *
537 537 * Functions that implicitly acquire an object hold (any function that calls
538 538 * object_alloc() to supply an object for the caller) need to be carefully noted
539 539 * so that the matching object_rele() is not neglected. Otherwise, leaked holds
540 540 * prevent all objects hashed to the affected rwlocks from ever being moved.
541 541 *
542 542 * The pointer to a held object can be hashed to the holding rwlock even after
543 543 * the object has been freed. Although it is possible to release the hold
544 544 * after freeing the object, you may decide to release the hold implicitly in
545 545 * whatever function frees the object, so as to release the hold as soon as
546 546 * possible, and for the sake of symmetry with the function that implicitly
547 547 * acquires the hold when it allocates the object. Here, object_free() releases
548 548 * the hold acquired by object_alloc(). Its implicit object_rele() forms a
549 549 * matching pair with object_hold():
550 550 *
551 551 * void
552 552 * object_free(object_t *object)
553 553 * {
554 554 * container_t *container;
555 555 *
556 556 * ASSERT(object_held(object));
557 557 * container = object->o_container;
558 558 * mutex_enter(&container->c_objects_lock);
559 559 * object->o_container =
560 560 * (void *)((uintptr_t)object->o_container | 0x1);
561 561 * list_remove(&container->c_objects, object);
562 562 * mutex_exit(&container->c_objects_lock);
563 563 * object_rele(object);
564 564 * kmem_cache_free(object_cache, object);
565 565 * }
566 566 *
567 567 * Note that object_free() cannot safely accept an object pointer as an argument
568 568 * unless the object is already held. Any function that calls object_free()
569 569 * needs to be carefully noted since it similarly forms a matching pair with
570 570 * object_hold().
571 571 *
572 572 * To complete the picture, the following callback function implements the
573 573 * general solution by moving objects only if they are currently unheld:
574 574 *
575 575 * static kmem_cbrc_t
576 576 * object_move(void *buf, void *newbuf, size_t size, void *arg)
577 577 * {
578 578 * object_t *op = buf, *np = newbuf;
579 579 * container_t *container;
580 580 *
581 581 * container = op->o_container;
582 582 * if ((uintptr_t)container & 0x3) {
583 583 * return (KMEM_CBRC_DONT_KNOW);
584 584 * }
585 585 *
586 586 * // Ensure that the container structure does not go away.
587 587 * if (container_hold(container) == 0) {
588 588 * return (KMEM_CBRC_DONT_KNOW);
589 589 * }
590 590 *
591 591 * mutex_enter(&container->c_objects_lock);
592 592 * if (container != op->o_container) {
593 593 * mutex_exit(&container->c_objects_lock);
594 594 * container_rele(container);
595 595 * return (KMEM_CBRC_DONT_KNOW);
596 596 * }
597 597 *
598 598 * if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
599 599 * mutex_exit(&container->c_objects_lock);
600 600 * container_rele(container);
601 601 * return (KMEM_CBRC_LATER);
602 602 * }
603 603 *
604 604 * object_move_impl(op, np); // critical section
605 605 * rw_exit(OBJECT_RWLOCK(op));
606 606 *
607 607 * op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
608 608 * list_link_replace(&op->o_link_node, &np->o_link_node);
609 609 * mutex_exit(&container->c_objects_lock);
610 610 * container_rele(container);
611 611 * return (KMEM_CBRC_YES);
612 612 * }
613 613 *
614 614 * Note that object_move() must invalidate the designated o_container pointer of
615 615 * the old object in the same way that object_free() does, since kmem will free
616 616 * the object in response to the KMEM_CBRC_YES return value.
617 617 *
618 618 * The lock order in object_move() differs from object_alloc(), which locks
619 619 * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
620 620 * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
621 621 * not a problem. Holding the lock on the object list in the example above
622 622 * through the entire callback not only prevents the object from going away, it
623 623 * also allows you to lock the list elsewhere and know that none of its elements
624 624 * will move during iteration.
625 625 *
626 626 * Adding an explicit hold everywhere an object from the cache is used is tricky
627 627 * and involves much more change to client code than a cache-specific solution
628 628 * that leverages existing state to decide whether or not an object is
629 629 * movable. However, this approach has the advantage that no object remains
630 630 * immovable for any significant length of time, making it extremely unlikely
631 631 * that long-lived allocations can continue holding slabs hostage; and it works
632 632 * for any cache.
633 633 *
634 634 * 3. Consolidator Implementation
635 635 *
636 636 * Once the client supplies a move function that a) recognizes known objects and
637 637 * b) avoids moving objects that are actively in use, the remaining work is up
638 638 * to the consolidator to decide which objects to move and when to issue
639 639 * callbacks.
640 640 *
641 641 * The consolidator relies on the fact that a cache's slabs are ordered by
642 642 * usage. Each slab has a fixed number of objects. Depending on the slab's
643 643 * "color" (the offset of the first object from the beginning of the slab;
644 644 * offsets are staggered to mitigate false sharing of cache lines) it is either
645 645 * the maximum number of objects per slab determined at cache creation time or
646 646 * else the number closest to the maximum that fits within the space remaining
647 647 * after the initial offset. A completely allocated slab may contribute some
648 648 * internal fragmentation (per-slab overhead) but no external fragmentation, so
649 649 * it is of no interest to the consolidator. At the other extreme, slabs whose
650 650 * objects have all been freed to the slab are released to the virtual memory
651 651 * (VM) subsystem (objects freed to magazines are still allocated as far as the
652 652 * slab is concerned). External fragmentation exists when there are slabs
653 653 * somewhere between these extremes. A partial slab has at least one but not all
654 654 * of its objects allocated. The more partial slabs, and the fewer allocated
655 655 * objects on each of them, the higher the fragmentation. Hence the
656 656 * consolidator's overall strategy is to reduce the number of partial slabs by
657 657 * moving allocated objects from the least allocated slabs to the most allocated
658 658 * slabs.
659 659 *
660 660 * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
661 661 * slabs are kept separately in an unordered list. Since the majority of slabs
662 662 * tend to be completely allocated (a typical unfragmented cache may have
663 663 * thousands of complete slabs and only a single partial slab), separating
664 664 * complete slabs improves the efficiency of partial slab ordering, since the
665 665 * complete slabs do not affect the depth or balance of the AVL tree. This
666 666 * ordered sequence of partial slabs acts as a "free list" supplying objects for
667 667 * allocation requests.
668 668 *
669 669 * Objects are always allocated from the first partial slab in the free list,
670 670 * where the allocation is most likely to eliminate a partial slab (by
671 671 * completely allocating it). Conversely, when a single object from a completely
672 672 * allocated slab is freed to the slab, that slab is added to the front of the
673 673 * free list. Since most free list activity involves highly allocated slabs
674 674 * coming and going at the front of the list, slabs tend naturally toward the
675 675 * ideal order: highly allocated at the front, sparsely allocated at the back.
676 676 * Slabs with few allocated objects are likely to become completely free if they
677 677 * keep a safe distance away from the front of the free list. Slab misorders
678 678 * interfere with the natural tendency of slabs to become completely free or
679 679 * completely allocated. For example, a slab with a single allocated object
680 680 * needs only a single free to escape the cache; its natural desire is
681 681 * frustrated when it finds itself at the front of the list where a second
682 682 * allocation happens just before the free could have released it. Another slab
683 683 * with all but one object allocated might have supplied the buffer instead, so
684 684 * that both (as opposed to neither) of the slabs would have been taken off the
685 685 * free list.
686 686 *
687 687 * Although slabs tend naturally toward the ideal order, misorders allowed by a
688 688 * simple list implementation defeat the consolidator's strategy of merging
689 689 * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
690 690 * needs another way to fix misorders to optimize its callback strategy. One
691 691 * approach is to periodically scan a limited number of slabs, advancing a
692 692 * marker to hold the current scan position, and to move extreme misorders to
693 693 * the front or back of the free list and to the front or back of the current
694 694 * scan range. By making consecutive scan ranges overlap by one slab, the least
695 695 * allocated slab in the current range can be carried along from the end of one
696 696 * scan to the start of the next.
697 697 *
698 698 * Maintaining partial slabs in an AVL tree relieves kmem of this additional
699 699 * task, however. Since most of the cache's activity is in the magazine layer,
700 700 * and allocations from the slab layer represent only a startup cost, the
701 701 * overhead of maintaining a balanced tree is not a significant concern compared
702 702 * to the opportunity of reducing complexity by eliminating the partial slab
703 703 * scanner just described. The overhead of an AVL tree is minimized by
704 704 * maintaining only partial slabs in the tree and keeping completely allocated
705 705 * slabs separately in a list. To avoid increasing the size of the slab
706 706 * structure the AVL linkage pointers are reused for the slab's list linkage,
707 707 * since the slab will always be either partial or complete, never stored both
708 708 * ways at the same time. To further minimize the overhead of the AVL tree the
709 709 * compare function that orders partial slabs by usage divides the range of
710 710 * allocated object counts into bins such that counts within the same bin are
711 711 * considered equal. Binning partial slabs makes it less likely that allocating
712 712 * or freeing a single object will change the slab's order, requiring a tree
713 713 * reinsertion (an avl_remove() followed by an avl_add(), both potentially
714 714 * requiring some rebalancing of the tree). Allocation counts closest to
715 715 * completely free and completely allocated are left unbinned (finely sorted) to
716 716 * better support the consolidator's strategy of merging slabs at either
717 717 * extreme.
718 718 *
719 719 * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
720 720 *
721 721 * The consolidator piggybacks on the kmem maintenance thread and is called on
722 722 * the same interval as kmem_cache_update(), once per cache every fifteen
723 723 * seconds. kmem maintains a running count of unallocated objects in the slab
724 724 * layer (cache_bufslab). The consolidator checks whether that number exceeds
725 725 * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
726 726 * there is a significant number of slabs in the cache (arbitrarily a minimum
727 727 * 101 total slabs). Unused objects that have fallen out of the magazine layer's
728 728 * working set are included in the assessment, and magazines in the depot are
729 729 * reaped if those objects would lift cache_bufslab above the fragmentation
730 730 * threshold. Once the consolidator decides that a cache is fragmented, it looks
731 731 * for a candidate slab to reclaim, starting at the end of the partial slab free
732 732 * list and scanning backwards. At first the consolidator is choosy: only a slab
733 733 * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
734 734 * single allocated object, regardless of percentage). If there is difficulty
735 735 * finding a candidate slab, kmem raises the allocation threshold incrementally,
736 736 * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
737 737 * external fragmentation (unused objects on the free list) below 12.5% (1/8),
738 738 * even in the worst case of every slab in the cache being almost 7/8 allocated.
739 739 * The threshold can also be lowered incrementally when candidate slabs are easy
740 740 * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
741 741 * is no longer fragmented.
742 742 *
743 743 * 3.2 Generating Callbacks
744 744 *
745 745 * Once an eligible slab is chosen, a callback is generated for every allocated
746 746 * object on the slab, in the hope that the client will move everything off the
747 747 * slab and make it reclaimable. Objects selected as move destinations are
748 748 * chosen from slabs at the front of the free list. Assuming slabs in the ideal
749 749 * order (most allocated at the front, least allocated at the back) and a
750 750 * cooperative client, the consolidator will succeed in removing slabs from both
751 751 * ends of the free list, completely allocating on the one hand and completely
752 752 * freeing on the other. Objects selected as move destinations are allocated in
753 753 * the kmem maintenance thread where move requests are enqueued. A separate
754 754 * callback thread removes pending callbacks from the queue and calls the
755 755 * client. The separate thread ensures that client code (the move function) does
756 756 * not interfere with internal kmem maintenance tasks. A map of pending
757 757 * callbacks keyed by object address (the object to be moved) is checked to
758 758 * ensure that duplicate callbacks are not generated for the same object.
759 759 * Allocating the move destination (the object to move to) prevents subsequent
760 760 * callbacks from selecting the same destination as an earlier pending callback.
761 761 *
762 762 * Move requests can also be generated by kmem_cache_reap() when the system is
763 763 * desperate for memory and by kmem_cache_move_notify(), called by the client to
764 764 * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
765 765 * The map of pending callbacks is protected by the same lock that protects the
766 766 * slab layer.
767 767 *
768 768 * When the system is desperate for memory, kmem does not bother to determine
769 769 * whether or not the cache exceeds the fragmentation threshold, but tries to
770 770 * consolidate as many slabs as possible. Normally, the consolidator chews
771 771 * slowly, one sparsely allocated slab at a time during each maintenance
772 772 * interval that the cache is fragmented. When desperate, the consolidator
773 773 * starts at the last partial slab and enqueues callbacks for every allocated
774 774 * object on every partial slab, working backwards until it reaches the first
775 775 * partial slab. The first partial slab, meanwhile, advances in pace with the
776 776 * consolidator as allocations to supply move destinations for the enqueued
777 777 * callbacks use up the highly allocated slabs at the front of the free list.
778 778 * Ideally, the overgrown free list collapses like an accordion, starting at
779 779 * both ends and ending at the center with a single partial slab.
780 780 *
781 781 * 3.3 Client Responses
782 782 *
783 783 * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
784 784 * marks the slab that supplied the stuck object non-reclaimable and moves it to
785 785 * front of the free list. The slab remains marked as long as it remains on the
786 786 * free list, and it appears more allocated to the partial slab compare function
787 787 * than any unmarked slab, no matter how many of its objects are allocated.
788 788 * Since even one immovable object ties up the entire slab, the goal is to
789 789 * completely allocate any slab that cannot be completely freed. kmem does not
790 790 * bother generating callbacks to move objects from a marked slab unless the
791 791 * system is desperate.
792 792 *
793 793 * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
794 794 * slab. If the client responds LATER too many times, kmem disbelieves and
795 795 * treats the response as a NO. The count is cleared when the slab is taken off
796 796 * the partial slab list or when the client moves one of the slab's objects.
797 797 *
798 798 * 4. Observability
799 799 *
800 800 * A kmem cache's external fragmentation is best observed with 'mdb -k' using
801 801 * the ::kmem_slabs dcmd. For a complete description of the command, enter
802 802 * '::help kmem_slabs' at the mdb prompt.
803 803 */
804 804
805 805 #include <sys/kmem_impl.h>
806 806 #include <sys/vmem_impl.h>
807 807 #include <sys/param.h>
808 808 #include <sys/sysmacros.h>
809 809 #include <sys/vm.h>
810 810 #include <sys/proc.h>
811 811 #include <sys/tuneable.h>
812 812 #include <sys/systm.h>
813 813 #include <sys/cmn_err.h>
814 814 #include <sys/debug.h>
815 815 #include <sys/sdt.h>
816 816 #include <sys/mutex.h>
817 817 #include <sys/bitmap.h>
818 818 #include <sys/atomic.h>
819 819 #include <sys/kobj.h>
820 820 #include <sys/disp.h>
821 821 #include <vm/seg_kmem.h>
822 822 #include <sys/log.h>
823 823 #include <sys/callb.h>
824 824 #include <sys/taskq.h>
825 825 #include <sys/modctl.h>
826 826 #include <sys/reboot.h>
827 827 #include <sys/id32.h>
828 828 #include <sys/zone.h>
829 829 #include <sys/netstack.h>
830 830 #ifdef DEBUG
831 831 #include <sys/random.h>
832 832 #endif
833 833
834 834 extern void streams_msg_init(void);
835 835 extern int segkp_fromheap;
836 836 extern void segkp_cache_free(void);
837 837 extern int callout_init_done;
838 838
839 839 struct kmem_cache_kstat {
840 840 kstat_named_t kmc_buf_size;
841 841 kstat_named_t kmc_align;
842 842 kstat_named_t kmc_chunk_size;
843 843 kstat_named_t kmc_slab_size;
844 844 kstat_named_t kmc_alloc;
845 845 kstat_named_t kmc_alloc_fail;
846 846 kstat_named_t kmc_free;
847 847 kstat_named_t kmc_depot_alloc;
848 848 kstat_named_t kmc_depot_free;
849 849 kstat_named_t kmc_depot_contention;
850 850 kstat_named_t kmc_slab_alloc;
851 851 kstat_named_t kmc_slab_free;
852 852 kstat_named_t kmc_buf_constructed;
853 853 kstat_named_t kmc_buf_avail;
854 854 kstat_named_t kmc_buf_inuse;
855 855 kstat_named_t kmc_buf_total;
856 856 kstat_named_t kmc_buf_max;
857 857 kstat_named_t kmc_slab_create;
858 858 kstat_named_t kmc_slab_destroy;
859 859 kstat_named_t kmc_vmem_source;
860 860 kstat_named_t kmc_hash_size;
861 861 kstat_named_t kmc_hash_lookup_depth;
862 862 kstat_named_t kmc_hash_rescale;
863 863 kstat_named_t kmc_full_magazines;
864 864 kstat_named_t kmc_empty_magazines;
865 865 kstat_named_t kmc_magazine_size;
866 866 kstat_named_t kmc_reap; /* number of kmem_cache_reap() calls */
867 867 kstat_named_t kmc_defrag; /* attempts to defrag all partial slabs */
868 868 kstat_named_t kmc_scan; /* attempts to defrag one partial slab */
869 869 kstat_named_t kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
870 870 kstat_named_t kmc_move_yes;
871 871 kstat_named_t kmc_move_no;
872 872 kstat_named_t kmc_move_later;
873 873 kstat_named_t kmc_move_dont_need;
874 874 kstat_named_t kmc_move_dont_know; /* obj unrecognized by client ... */
875 875 kstat_named_t kmc_move_hunt_found; /* ... but found in mag layer */
876 876 kstat_named_t kmc_move_slabs_freed; /* slabs freed by consolidator */
877 877 kstat_named_t kmc_move_reclaimable; /* buffers, if consolidator ran */
878 878 } kmem_cache_kstat = {
879 879 { "buf_size", KSTAT_DATA_UINT64 },
880 880 { "align", KSTAT_DATA_UINT64 },
881 881 { "chunk_size", KSTAT_DATA_UINT64 },
882 882 { "slab_size", KSTAT_DATA_UINT64 },
883 883 { "alloc", KSTAT_DATA_UINT64 },
884 884 { "alloc_fail", KSTAT_DATA_UINT64 },
885 885 { "free", KSTAT_DATA_UINT64 },
886 886 { "depot_alloc", KSTAT_DATA_UINT64 },
887 887 { "depot_free", KSTAT_DATA_UINT64 },
888 888 { "depot_contention", KSTAT_DATA_UINT64 },
889 889 { "slab_alloc", KSTAT_DATA_UINT64 },
890 890 { "slab_free", KSTAT_DATA_UINT64 },
891 891 { "buf_constructed", KSTAT_DATA_UINT64 },
892 892 { "buf_avail", KSTAT_DATA_UINT64 },
893 893 { "buf_inuse", KSTAT_DATA_UINT64 },
894 894 { "buf_total", KSTAT_DATA_UINT64 },
895 895 { "buf_max", KSTAT_DATA_UINT64 },
896 896 { "slab_create", KSTAT_DATA_UINT64 },
897 897 { "slab_destroy", KSTAT_DATA_UINT64 },
898 898 { "vmem_source", KSTAT_DATA_UINT64 },
899 899 { "hash_size", KSTAT_DATA_UINT64 },
900 900 { "hash_lookup_depth", KSTAT_DATA_UINT64 },
901 901 { "hash_rescale", KSTAT_DATA_UINT64 },
902 902 { "full_magazines", KSTAT_DATA_UINT64 },
903 903 { "empty_magazines", KSTAT_DATA_UINT64 },
904 904 { "magazine_size", KSTAT_DATA_UINT64 },
905 905 { "reap", KSTAT_DATA_UINT64 },
906 906 { "defrag", KSTAT_DATA_UINT64 },
907 907 { "scan", KSTAT_DATA_UINT64 },
908 908 { "move_callbacks", KSTAT_DATA_UINT64 },
909 909 { "move_yes", KSTAT_DATA_UINT64 },
910 910 { "move_no", KSTAT_DATA_UINT64 },
911 911 { "move_later", KSTAT_DATA_UINT64 },
912 912 { "move_dont_need", KSTAT_DATA_UINT64 },
913 913 { "move_dont_know", KSTAT_DATA_UINT64 },
914 914 { "move_hunt_found", KSTAT_DATA_UINT64 },
915 915 { "move_slabs_freed", KSTAT_DATA_UINT64 },
916 916 { "move_reclaimable", KSTAT_DATA_UINT64 },
917 917 };
918 918
919 919 static kmutex_t kmem_cache_kstat_lock;
920 920
921 921 /*
922 922 * The default set of caches to back kmem_alloc().
923 923 * These sizes should be reevaluated periodically.
924 924 *
925 925 * We want allocations that are multiples of the coherency granularity
926 926 * (64 bytes) to be satisfied from a cache which is a multiple of 64
927 927 * bytes, so that it will be 64-byte aligned. For all multiples of 64,
928 928 * the next kmem_cache_size greater than or equal to it must be a
929 929 * multiple of 64.
930 930 *
931 931 * We split the table into two sections: size <= 4k and size > 4k. This
932 932 * saves a lot of space and cache footprint in our cache tables.
933 933 */
934 934 static const int kmem_alloc_sizes[] = {
935 935 1 * 8,
936 936 2 * 8,
937 937 3 * 8,
938 938 4 * 8, 5 * 8, 6 * 8, 7 * 8,
939 939 4 * 16, 5 * 16, 6 * 16, 7 * 16,
940 940 4 * 32, 5 * 32, 6 * 32, 7 * 32,
941 941 4 * 64, 5 * 64, 6 * 64, 7 * 64,
942 942 4 * 128, 5 * 128, 6 * 128, 7 * 128,
943 943 P2ALIGN(8192 / 7, 64),
944 944 P2ALIGN(8192 / 6, 64),
945 945 P2ALIGN(8192 / 5, 64),
946 946 P2ALIGN(8192 / 4, 64),
947 947 P2ALIGN(8192 / 3, 64),
948 948 P2ALIGN(8192 / 2, 64),
949 949 };
950 950
951 951 static const int kmem_big_alloc_sizes[] = {
952 952 2 * 4096, 3 * 4096,
953 953 2 * 8192, 3 * 8192,
954 954 4 * 8192, 5 * 8192, 6 * 8192, 7 * 8192,
955 955 8 * 8192, 9 * 8192, 10 * 8192, 11 * 8192,
956 956 12 * 8192, 13 * 8192, 14 * 8192, 15 * 8192,
957 957 16 * 8192
958 958 };
959 959
960 960 #define KMEM_MAXBUF 4096
961 961 #define KMEM_BIG_MAXBUF_32BIT 32768
962 962 #define KMEM_BIG_MAXBUF 131072
963 963
964 964 #define KMEM_BIG_MULTIPLE 4096 /* big_alloc_sizes must be a multiple */
965 965 #define KMEM_BIG_SHIFT 12 /* lg(KMEM_BIG_MULTIPLE) */
966 966
967 967 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
968 968 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
969 969
970 970 #define KMEM_ALLOC_TABLE_MAX (KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
971 971 static size_t kmem_big_alloc_table_max = 0; /* # of filled elements */
972 972
973 973 static kmem_magtype_t kmem_magtype[] = {
974 974 { 1, 8, 3200, 65536 },
975 975 { 3, 16, 256, 32768 },
976 976 { 7, 32, 64, 16384 },
977 977 { 15, 64, 0, 8192 },
978 978 { 31, 64, 0, 4096 },
979 979 { 47, 64, 0, 2048 },
980 980 { 63, 64, 0, 1024 },
981 981 { 95, 64, 0, 512 },
982 982 { 143, 64, 0, 0 },
983 983 };
984 984
985 985 static uint32_t kmem_reaping;
986 986 static uint32_t kmem_reaping_idspace;
987 987
988 988 /*
989 989 * kmem tunables
990 990 */
991 991 clock_t kmem_reap_interval; /* cache reaping rate [15 * HZ ticks] */
992 992 int kmem_depot_contention = 3; /* max failed tryenters per real interval */
993 993 pgcnt_t kmem_reapahead = 0; /* start reaping N pages before pageout */
994 994 int kmem_panic = 1; /* whether to panic on error */
995 995 int kmem_logging = 1; /* kmem_log_enter() override */
996 996 uint32_t kmem_mtbf = 0; /* mean time between failures [default: off] */
997 997 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
998 998 size_t kmem_content_log_size; /* content log size [2% of memory] */
999 999 size_t kmem_failure_log_size; /* failure log [4 pages per CPU] */
1000 1000 size_t kmem_slab_log_size; /* slab create log [4 pages per CPU] */
1001 1001 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1002 1002 size_t kmem_lite_minsize = 0; /* minimum buffer size for KMF_LITE */
1003 1003 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1004 1004 int kmem_lite_pcs = 4; /* number of PCs to store in KMF_LITE mode */
1005 1005 size_t kmem_maxverify; /* maximum bytes to inspect in debug routines */
1006 1006 size_t kmem_minfirewall; /* hardware-enforced redzone threshold */
1007 1007
1008 1008 #ifdef _LP64
1009 1009 size_t kmem_max_cached = KMEM_BIG_MAXBUF; /* maximum kmem_alloc cache */
1010 1010 #else
1011 1011 size_t kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1012 1012 #endif
1013 1013
1014 1014 #ifdef DEBUG
1015 1015 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1016 1016 #else
1017 1017 int kmem_flags = 0;
1018 1018 #endif
1019 1019 int kmem_ready;
1020 1020
1021 1021 static kmem_cache_t *kmem_slab_cache;
1022 1022 static kmem_cache_t *kmem_bufctl_cache;
1023 1023 static kmem_cache_t *kmem_bufctl_audit_cache;
1024 1024
1025 1025 static kmutex_t kmem_cache_lock; /* inter-cache linkage only */
1026 1026 static list_t kmem_caches;
1027 1027
1028 1028 static taskq_t *kmem_taskq;
1029 1029 static kmutex_t kmem_flags_lock;
1030 1030 static vmem_t *kmem_metadata_arena;
1031 1031 static vmem_t *kmem_msb_arena; /* arena for metadata caches */
1032 1032 static vmem_t *kmem_cache_arena;
1033 1033 static vmem_t *kmem_hash_arena;
1034 1034 static vmem_t *kmem_log_arena;
1035 1035 static vmem_t *kmem_oversize_arena;
1036 1036 static vmem_t *kmem_va_arena;
1037 1037 static vmem_t *kmem_default_arena;
1038 1038 static vmem_t *kmem_firewall_va_arena;
1039 1039 static vmem_t *kmem_firewall_arena;
1040 1040
1041 1041 /*
1042 1042 * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1043 1043 * turned on when DEBUG is also defined.
1044 1044 */
1045 1045 #ifdef DEBUG
1046 1046 #define KMEM_STATS
1047 1047 #endif /* DEBUG */
1048 1048
1049 1049 #ifdef KMEM_STATS
1050 1050 #define KMEM_STAT_ADD(stat) ((stat)++)
1051 1051 #define KMEM_STAT_COND_ADD(cond, stat) ((void) (!(cond) || (stat)++))
1052 1052 #else
1053 1053 #define KMEM_STAT_ADD(stat) /* nothing */
1054 1054 #define KMEM_STAT_COND_ADD(cond, stat) /* nothing */
1055 1055 #endif /* KMEM_STATS */
1056 1056
1057 1057 /*
1058 1058 * kmem slab consolidator thresholds (tunables)
1059 1059 */
1060 1060 size_t kmem_frag_minslabs = 101; /* minimum total slabs */
1061 1061 size_t kmem_frag_numer = 1; /* free buffers (numerator) */
1062 1062 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1063 1063 /*
1064 1064 * Maximum number of slabs from which to move buffers during a single
1065 1065 * maintenance interval while the system is not low on memory.
1066 1066 */
1067 1067 size_t kmem_reclaim_max_slabs = 1;
1068 1068 /*
1069 1069 * Number of slabs to scan backwards from the end of the partial slab list
1070 1070 * when searching for buffers to relocate.
1071 1071 */
1072 1072 size_t kmem_reclaim_scan_range = 12;
1073 1073
1074 1074 #ifdef KMEM_STATS
1075 1075 static struct {
1076 1076 uint64_t kms_callbacks;
1077 1077 uint64_t kms_yes;
1078 1078 uint64_t kms_no;
1079 1079 uint64_t kms_later;
1080 1080 uint64_t kms_dont_need;
1081 1081 uint64_t kms_dont_know;
1082 1082 uint64_t kms_hunt_found_mag;
1083 1083 uint64_t kms_hunt_found_slab;
1084 1084 uint64_t kms_hunt_alloc_fail;
1085 1085 uint64_t kms_hunt_lucky;
1086 1086 uint64_t kms_notify;
1087 1087 uint64_t kms_notify_callbacks;
1088 1088 uint64_t kms_disbelief;
1089 1089 uint64_t kms_already_pending;
1090 1090 uint64_t kms_callback_alloc_fail;
1091 1091 uint64_t kms_callback_taskq_fail;
1092 1092 uint64_t kms_endscan_slab_dead;
1093 1093 uint64_t kms_endscan_slab_destroyed;
1094 1094 uint64_t kms_endscan_nomem;
1095 1095 uint64_t kms_endscan_refcnt_changed;
1096 1096 uint64_t kms_endscan_nomove_changed;
1097 1097 uint64_t kms_endscan_freelist;
1098 1098 uint64_t kms_avl_update;
1099 1099 uint64_t kms_avl_noupdate;
1100 1100 uint64_t kms_no_longer_reclaimable;
1101 1101 uint64_t kms_notify_no_longer_reclaimable;
1102 1102 uint64_t kms_notify_slab_dead;
1103 1103 uint64_t kms_notify_slab_destroyed;
1104 1104 uint64_t kms_alloc_fail;
1105 1105 uint64_t kms_constructor_fail;
1106 1106 uint64_t kms_dead_slabs_freed;
1107 1107 uint64_t kms_defrags;
1108 1108 uint64_t kms_scans;
1109 1109 uint64_t kms_scan_depot_ws_reaps;
1110 1110 uint64_t kms_debug_reaps;
1111 1111 uint64_t kms_debug_scans;
1112 1112 } kmem_move_stats;
1113 1113 #endif /* KMEM_STATS */
1114 1114
1115 1115 /* consolidator knobs */
1116 1116 static boolean_t kmem_move_noreap;
1117 1117 static boolean_t kmem_move_blocked;
1118 1118 static boolean_t kmem_move_fulltilt;
1119 1119 static boolean_t kmem_move_any_partial;
1120 1120
1121 1121 #ifdef DEBUG
1122 1122 /*
1123 1123 * kmem consolidator debug tunables:
1124 1124 * Ensure code coverage by occasionally running the consolidator even when the
1125 1125 * caches are not fragmented (they may never be). These intervals are mean time
1126 1126 * in cache maintenance intervals (kmem_cache_update).
1127 1127 */
1128 1128 uint32_t kmem_mtb_move = 60; /* defrag 1 slab (~15min) */
1129 1129 uint32_t kmem_mtb_reap = 1800; /* defrag all slabs (~7.5hrs) */
1130 1130 #endif /* DEBUG */
1131 1131
1132 1132 static kmem_cache_t *kmem_defrag_cache;
1133 1133 static kmem_cache_t *kmem_move_cache;
1134 1134 static taskq_t *kmem_move_taskq;
1135 1135
1136 1136 static void kmem_cache_scan(kmem_cache_t *);
1137 1137 static void kmem_cache_defrag(kmem_cache_t *);
1138 1138 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1139 1139
1140 1140
1141 1141 kmem_log_header_t *kmem_transaction_log;
1142 1142 kmem_log_header_t *kmem_content_log;
1143 1143 kmem_log_header_t *kmem_failure_log;
1144 1144 kmem_log_header_t *kmem_slab_log;
1145 1145
1146 1146 static int kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1147 1147
1148 1148 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller) \
1149 1149 if ((count) > 0) { \
1150 1150 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1151 1151 pc_t *_e; \
1152 1152 /* memmove() the old entries down one notch */ \
1153 1153 for (_e = &_s[(count) - 1]; _e > _s; _e--) \
1154 1154 *_e = *(_e - 1); \
1155 1155 *_s = (uintptr_t)(caller); \
1156 1156 }
1157 1157
1158 1158 #define KMERR_MODIFIED 0 /* buffer modified while on freelist */
1159 1159 #define KMERR_REDZONE 1 /* redzone violation (write past end of buf) */
1160 1160 #define KMERR_DUPFREE 2 /* freed a buffer twice */
1161 1161 #define KMERR_BADADDR 3 /* freed a bad (unallocated) address */
1162 1162 #define KMERR_BADBUFTAG 4 /* buftag corrupted */
1163 1163 #define KMERR_BADBUFCTL 5 /* bufctl corrupted */
1164 1164 #define KMERR_BADCACHE 6 /* freed a buffer to the wrong cache */
1165 1165 #define KMERR_BADSIZE 7 /* alloc size != free size */
1166 1166 #define KMERR_BADBASE 8 /* buffer base address wrong */
1167 1167
1168 1168 struct {
1169 1169 hrtime_t kmp_timestamp; /* timestamp of panic */
1170 1170 int kmp_error; /* type of kmem error */
1171 1171 void *kmp_buffer; /* buffer that induced panic */
1172 1172 void *kmp_realbuf; /* real start address for buffer */
1173 1173 kmem_cache_t *kmp_cache; /* buffer's cache according to client */
1174 1174 kmem_cache_t *kmp_realcache; /* actual cache containing buffer */
1175 1175 kmem_slab_t *kmp_slab; /* slab accoring to kmem_findslab() */
1176 1176 kmem_bufctl_t *kmp_bufctl; /* bufctl */
1177 1177 } kmem_panic_info;
1178 1178
1179 1179
1180 1180 static void
1181 1181 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1182 1182 {
1183 1183 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1184 1184 uint64_t *buf = buf_arg;
1185 1185
1186 1186 while (buf < bufend)
1187 1187 *buf++ = pattern;
1188 1188 }
1189 1189
1190 1190 static void *
1191 1191 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1192 1192 {
1193 1193 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1194 1194 uint64_t *buf;
1195 1195
1196 1196 for (buf = buf_arg; buf < bufend; buf++)
1197 1197 if (*buf != pattern)
1198 1198 return (buf);
1199 1199 return (NULL);
1200 1200 }
1201 1201
1202 1202 static void *
1203 1203 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1204 1204 {
1205 1205 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1206 1206 uint64_t *buf;
1207 1207
1208 1208 for (buf = buf_arg; buf < bufend; buf++) {
1209 1209 if (*buf != old) {
1210 1210 copy_pattern(old, buf_arg,
1211 1211 (char *)buf - (char *)buf_arg);
1212 1212 return (buf);
1213 1213 }
1214 1214 *buf = new;
1215 1215 }
1216 1216
1217 1217 return (NULL);
1218 1218 }
1219 1219
1220 1220 static void
1221 1221 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1222 1222 {
1223 1223 kmem_cache_t *cp;
1224 1224
1225 1225 mutex_enter(&kmem_cache_lock);
1226 1226 for (cp = list_head(&kmem_caches); cp != NULL;
1227 1227 cp = list_next(&kmem_caches, cp))
1228 1228 if (tq != NULL)
1229 1229 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1230 1230 tqflag);
1231 1231 else
1232 1232 func(cp);
1233 1233 mutex_exit(&kmem_cache_lock);
1234 1234 }
1235 1235
1236 1236 static void
1237 1237 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1238 1238 {
1239 1239 kmem_cache_t *cp;
1240 1240
1241 1241 mutex_enter(&kmem_cache_lock);
1242 1242 for (cp = list_head(&kmem_caches); cp != NULL;
1243 1243 cp = list_next(&kmem_caches, cp)) {
1244 1244 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1245 1245 continue;
1246 1246 if (tq != NULL)
1247 1247 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1248 1248 tqflag);
1249 1249 else
1250 1250 func(cp);
1251 1251 }
1252 1252 mutex_exit(&kmem_cache_lock);
1253 1253 }
1254 1254
1255 1255 /*
1256 1256 * Debugging support. Given a buffer address, find its slab.
1257 1257 */
1258 1258 static kmem_slab_t *
1259 1259 kmem_findslab(kmem_cache_t *cp, void *buf)
1260 1260 {
1261 1261 kmem_slab_t *sp;
1262 1262
1263 1263 mutex_enter(&cp->cache_lock);
1264 1264 for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1265 1265 sp = list_next(&cp->cache_complete_slabs, sp)) {
1266 1266 if (KMEM_SLAB_MEMBER(sp, buf)) {
1267 1267 mutex_exit(&cp->cache_lock);
1268 1268 return (sp);
1269 1269 }
1270 1270 }
1271 1271 for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1272 1272 sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1273 1273 if (KMEM_SLAB_MEMBER(sp, buf)) {
1274 1274 mutex_exit(&cp->cache_lock);
1275 1275 return (sp);
1276 1276 }
1277 1277 }
1278 1278 mutex_exit(&cp->cache_lock);
1279 1279
1280 1280 return (NULL);
1281 1281 }
1282 1282
1283 1283 static void
1284 1284 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1285 1285 {
1286 1286 kmem_buftag_t *btp = NULL;
1287 1287 kmem_bufctl_t *bcp = NULL;
1288 1288 kmem_cache_t *cp = cparg;
1289 1289 kmem_slab_t *sp;
1290 1290 uint64_t *off;
1291 1291 void *buf = bufarg;
1292 1292
1293 1293 kmem_logging = 0; /* stop logging when a bad thing happens */
1294 1294
1295 1295 kmem_panic_info.kmp_timestamp = gethrtime();
1296 1296
1297 1297 sp = kmem_findslab(cp, buf);
1298 1298 if (sp == NULL) {
1299 1299 for (cp = list_tail(&kmem_caches); cp != NULL;
1300 1300 cp = list_prev(&kmem_caches, cp)) {
1301 1301 if ((sp = kmem_findslab(cp, buf)) != NULL)
1302 1302 break;
1303 1303 }
1304 1304 }
1305 1305
1306 1306 if (sp == NULL) {
1307 1307 cp = NULL;
1308 1308 error = KMERR_BADADDR;
1309 1309 } else {
1310 1310 if (cp != cparg)
1311 1311 error = KMERR_BADCACHE;
1312 1312 else
1313 1313 buf = (char *)bufarg - ((uintptr_t)bufarg -
1314 1314 (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1315 1315 if (buf != bufarg)
1316 1316 error = KMERR_BADBASE;
1317 1317 if (cp->cache_flags & KMF_BUFTAG)
1318 1318 btp = KMEM_BUFTAG(cp, buf);
1319 1319 if (cp->cache_flags & KMF_HASH) {
1320 1320 mutex_enter(&cp->cache_lock);
1321 1321 for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1322 1322 if (bcp->bc_addr == buf)
1323 1323 break;
1324 1324 mutex_exit(&cp->cache_lock);
1325 1325 if (bcp == NULL && btp != NULL)
1326 1326 bcp = btp->bt_bufctl;
1327 1327 if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1328 1328 NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1329 1329 bcp->bc_addr != buf) {
1330 1330 error = KMERR_BADBUFCTL;
1331 1331 bcp = NULL;
1332 1332 }
1333 1333 }
1334 1334 }
1335 1335
1336 1336 kmem_panic_info.kmp_error = error;
1337 1337 kmem_panic_info.kmp_buffer = bufarg;
1338 1338 kmem_panic_info.kmp_realbuf = buf;
1339 1339 kmem_panic_info.kmp_cache = cparg;
1340 1340 kmem_panic_info.kmp_realcache = cp;
1341 1341 kmem_panic_info.kmp_slab = sp;
1342 1342 kmem_panic_info.kmp_bufctl = bcp;
1343 1343
1344 1344 printf("kernel memory allocator: ");
1345 1345
1346 1346 switch (error) {
1347 1347
1348 1348 case KMERR_MODIFIED:
1349 1349 printf("buffer modified after being freed\n");
1350 1350 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1351 1351 if (off == NULL) /* shouldn't happen */
1352 1352 off = buf;
1353 1353 printf("modification occurred at offset 0x%lx "
1354 1354 "(0x%llx replaced by 0x%llx)\n",
1355 1355 (uintptr_t)off - (uintptr_t)buf,
1356 1356 (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1357 1357 break;
1358 1358
1359 1359 case KMERR_REDZONE:
1360 1360 printf("redzone violation: write past end of buffer\n");
1361 1361 break;
1362 1362
1363 1363 case KMERR_BADADDR:
1364 1364 printf("invalid free: buffer not in cache\n");
1365 1365 break;
1366 1366
1367 1367 case KMERR_DUPFREE:
1368 1368 printf("duplicate free: buffer freed twice\n");
1369 1369 break;
1370 1370
1371 1371 case KMERR_BADBUFTAG:
1372 1372 printf("boundary tag corrupted\n");
1373 1373 printf("bcp ^ bxstat = %lx, should be %lx\n",
1374 1374 (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1375 1375 KMEM_BUFTAG_FREE);
1376 1376 break;
1377 1377
1378 1378 case KMERR_BADBUFCTL:
1379 1379 printf("bufctl corrupted\n");
1380 1380 break;
1381 1381
1382 1382 case KMERR_BADCACHE:
1383 1383 printf("buffer freed to wrong cache\n");
1384 1384 printf("buffer was allocated from %s,\n", cp->cache_name);
1385 1385 printf("caller attempting free to %s.\n", cparg->cache_name);
1386 1386 break;
1387 1387
1388 1388 case KMERR_BADSIZE:
1389 1389 printf("bad free: free size (%u) != alloc size (%u)\n",
1390 1390 KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1391 1391 KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1392 1392 break;
1393 1393
1394 1394 case KMERR_BADBASE:
1395 1395 printf("bad free: free address (%p) != alloc address (%p)\n",
1396 1396 bufarg, buf);
1397 1397 break;
1398 1398 }
1399 1399
1400 1400 printf("buffer=%p bufctl=%p cache: %s\n",
1401 1401 bufarg, (void *)bcp, cparg->cache_name);
1402 1402
1403 1403 if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1404 1404 error != KMERR_BADBUFCTL) {
1405 1405 int d;
1406 1406 timestruc_t ts;
1407 1407 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1408 1408
1409 1409 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1410 1410 printf("previous transaction on buffer %p:\n", buf);
1411 1411 printf("thread=%p time=T-%ld.%09ld slab=%p cache: %s\n",
1412 1412 (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1413 1413 (void *)sp, cp->cache_name);
1414 1414 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1415 1415 ulong_t off;
1416 1416 char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1417 1417 printf("%s+%lx\n", sym ? sym : "?", off);
1418 1418 }
1419 1419 }
1420 1420 if (kmem_panic > 0)
1421 1421 panic("kernel heap corruption detected");
1422 1422 if (kmem_panic == 0)
1423 1423 debug_enter(NULL);
1424 1424 kmem_logging = 1; /* resume logging */
1425 1425 }
1426 1426
1427 1427 static kmem_log_header_t *
1428 1428 kmem_log_init(size_t logsize)
1429 1429 {
1430 1430 kmem_log_header_t *lhp;
1431 1431 int nchunks = 4 * max_ncpus;
1432 1432 size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1433 1433 int i;
1434 1434
1435 1435 /*
1436 1436 * Make sure that lhp->lh_cpu[] is nicely aligned
1437 1437 * to prevent false sharing of cache lines.
1438 1438 */
1439 1439 lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1440 1440 lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1441 1441 NULL, NULL, VM_SLEEP);
1442 1442 bzero(lhp, lhsize);
1443 1443
1444 1444 mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1445 1445 lhp->lh_nchunks = nchunks;
1446 1446 lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1447 1447 lhp->lh_base = vmem_alloc(kmem_log_arena,
1448 1448 lhp->lh_chunksize * nchunks, VM_SLEEP);
1449 1449 lhp->lh_free = vmem_alloc(kmem_log_arena,
1450 1450 nchunks * sizeof (int), VM_SLEEP);
1451 1451 bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1452 1452
1453 1453 for (i = 0; i < max_ncpus; i++) {
1454 1454 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1455 1455 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1456 1456 clhp->clh_chunk = i;
1457 1457 }
1458 1458
1459 1459 for (i = max_ncpus; i < nchunks; i++)
1460 1460 lhp->lh_free[i] = i;
1461 1461
1462 1462 lhp->lh_head = max_ncpus;
1463 1463 lhp->lh_tail = 0;
1464 1464
1465 1465 return (lhp);
1466 1466 }
1467 1467
1468 1468 static void *
1469 1469 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1470 1470 {
1471 1471 void *logspace;
1472 1472 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1473 1473
1474 1474 if (lhp == NULL || kmem_logging == 0 || panicstr)
1475 1475 return (NULL);
1476 1476
1477 1477 mutex_enter(&clhp->clh_lock);
1478 1478 clhp->clh_hits++;
1479 1479 if (size > clhp->clh_avail) {
1480 1480 mutex_enter(&lhp->lh_lock);
1481 1481 lhp->lh_hits++;
1482 1482 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1483 1483 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1484 1484 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1485 1485 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1486 1486 clhp->clh_current = lhp->lh_base +
1487 1487 clhp->clh_chunk * lhp->lh_chunksize;
1488 1488 clhp->clh_avail = lhp->lh_chunksize;
1489 1489 if (size > lhp->lh_chunksize)
1490 1490 size = lhp->lh_chunksize;
1491 1491 mutex_exit(&lhp->lh_lock);
1492 1492 }
1493 1493 logspace = clhp->clh_current;
1494 1494 clhp->clh_current += size;
1495 1495 clhp->clh_avail -= size;
1496 1496 bcopy(data, logspace, size);
1497 1497 mutex_exit(&clhp->clh_lock);
1498 1498 return (logspace);
1499 1499 }
1500 1500
1501 1501 #define KMEM_AUDIT(lp, cp, bcp) \
1502 1502 { \
1503 1503 kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp); \
1504 1504 _bcp->bc_timestamp = gethrtime(); \
1505 1505 _bcp->bc_thread = curthread; \
1506 1506 _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH); \
1507 1507 _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp)); \
1508 1508 }
1509 1509
1510 1510 static void
1511 1511 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1512 1512 kmem_slab_t *sp, void *addr)
1513 1513 {
1514 1514 kmem_bufctl_audit_t bca;
1515 1515
1516 1516 bzero(&bca, sizeof (kmem_bufctl_audit_t));
1517 1517 bca.bc_addr = addr;
1518 1518 bca.bc_slab = sp;
1519 1519 bca.bc_cache = cp;
1520 1520 KMEM_AUDIT(lp, cp, &bca);
1521 1521 }
1522 1522
1523 1523 /*
1524 1524 * Create a new slab for cache cp.
1525 1525 */
1526 1526 static kmem_slab_t *
1527 1527 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1528 1528 {
1529 1529 size_t slabsize = cp->cache_slabsize;
1530 1530 size_t chunksize = cp->cache_chunksize;
1531 1531 int cache_flags = cp->cache_flags;
1532 1532 size_t color, chunks;
1533 1533 char *buf, *slab;
1534 1534 kmem_slab_t *sp;
1535 1535 kmem_bufctl_t *bcp;
1536 1536 vmem_t *vmp = cp->cache_arena;
1537 1537
1538 1538 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1539 1539
1540 1540 color = cp->cache_color + cp->cache_align;
1541 1541 if (color > cp->cache_maxcolor)
1542 1542 color = cp->cache_mincolor;
1543 1543 cp->cache_color = color;
1544 1544
1545 1545 slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1546 1546
1547 1547 if (slab == NULL)
1548 1548 goto vmem_alloc_failure;
1549 1549
1550 1550 ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1551 1551
1552 1552 /*
1553 1553 * Reverify what was already checked in kmem_cache_set_move(), since the
1554 1554 * consolidator depends (for correctness) on slabs being initialized
1555 1555 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1556 1556 * clients to distinguish uninitialized memory from known objects).
1557 1557 */
1558 1558 ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1559 1559 if (!(cp->cache_cflags & KMC_NOTOUCH))
1560 1560 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1561 1561
1562 1562 if (cache_flags & KMF_HASH) {
1563 1563 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1564 1564 goto slab_alloc_failure;
1565 1565 chunks = (slabsize - color) / chunksize;
1566 1566 } else {
1567 1567 sp = KMEM_SLAB(cp, slab);
1568 1568 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1569 1569 }
1570 1570
1571 1571 sp->slab_cache = cp;
1572 1572 sp->slab_head = NULL;
1573 1573 sp->slab_refcnt = 0;
1574 1574 sp->slab_base = buf = slab + color;
1575 1575 sp->slab_chunks = chunks;
1576 1576 sp->slab_stuck_offset = (uint32_t)-1;
1577 1577 sp->slab_later_count = 0;
1578 1578 sp->slab_flags = 0;
1579 1579
1580 1580 ASSERT(chunks > 0);
1581 1581 while (chunks-- != 0) {
1582 1582 if (cache_flags & KMF_HASH) {
1583 1583 bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1584 1584 if (bcp == NULL)
1585 1585 goto bufctl_alloc_failure;
1586 1586 if (cache_flags & KMF_AUDIT) {
1587 1587 kmem_bufctl_audit_t *bcap =
1588 1588 (kmem_bufctl_audit_t *)bcp;
1589 1589 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1590 1590 bcap->bc_cache = cp;
1591 1591 }
1592 1592 bcp->bc_addr = buf;
1593 1593 bcp->bc_slab = sp;
1594 1594 } else {
1595 1595 bcp = KMEM_BUFCTL(cp, buf);
1596 1596 }
1597 1597 if (cache_flags & KMF_BUFTAG) {
1598 1598 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1599 1599 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1600 1600 btp->bt_bufctl = bcp;
1601 1601 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1602 1602 if (cache_flags & KMF_DEADBEEF) {
1603 1603 copy_pattern(KMEM_FREE_PATTERN, buf,
1604 1604 cp->cache_verify);
1605 1605 }
1606 1606 }
1607 1607 bcp->bc_next = sp->slab_head;
1608 1608 sp->slab_head = bcp;
1609 1609 buf += chunksize;
1610 1610 }
1611 1611
1612 1612 kmem_log_event(kmem_slab_log, cp, sp, slab);
1613 1613
1614 1614 return (sp);
1615 1615
1616 1616 bufctl_alloc_failure:
1617 1617
1618 1618 while ((bcp = sp->slab_head) != NULL) {
1619 1619 sp->slab_head = bcp->bc_next;
1620 1620 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1621 1621 }
1622 1622 kmem_cache_free(kmem_slab_cache, sp);
1623 1623
1624 1624 slab_alloc_failure:
1625 1625
1626 1626 vmem_free(vmp, slab, slabsize);
1627 1627
1628 1628 vmem_alloc_failure:
1629 1629
1630 1630 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1631 1631 atomic_inc_64(&cp->cache_alloc_fail);
1632 1632
1633 1633 return (NULL);
1634 1634 }
1635 1635
1636 1636 /*
1637 1637 * Destroy a slab.
1638 1638 */
1639 1639 static void
1640 1640 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1641 1641 {
1642 1642 vmem_t *vmp = cp->cache_arena;
1643 1643 void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1644 1644
1645 1645 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1646 1646 ASSERT(sp->slab_refcnt == 0);
1647 1647
1648 1648 if (cp->cache_flags & KMF_HASH) {
1649 1649 kmem_bufctl_t *bcp;
1650 1650 while ((bcp = sp->slab_head) != NULL) {
1651 1651 sp->slab_head = bcp->bc_next;
1652 1652 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1653 1653 }
1654 1654 kmem_cache_free(kmem_slab_cache, sp);
1655 1655 }
1656 1656 vmem_free(vmp, slab, cp->cache_slabsize);
1657 1657 }
1658 1658
1659 1659 static void *
1660 1660 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1661 1661 {
1662 1662 kmem_bufctl_t *bcp, **hash_bucket;
1663 1663 void *buf;
1664 1664 boolean_t new_slab = (sp->slab_refcnt == 0);
1665 1665
1666 1666 ASSERT(MUTEX_HELD(&cp->cache_lock));
1667 1667 /*
1668 1668 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1669 1669 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1670 1670 * slab is newly created.
1671 1671 */
1672 1672 ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1673 1673 (sp == avl_first(&cp->cache_partial_slabs))));
1674 1674 ASSERT(sp->slab_cache == cp);
1675 1675
1676 1676 cp->cache_slab_alloc++;
1677 1677 cp->cache_bufslab--;
1678 1678 sp->slab_refcnt++;
1679 1679
1680 1680 bcp = sp->slab_head;
1681 1681 sp->slab_head = bcp->bc_next;
1682 1682
1683 1683 if (cp->cache_flags & KMF_HASH) {
1684 1684 /*
1685 1685 * Add buffer to allocated-address hash table.
1686 1686 */
1687 1687 buf = bcp->bc_addr;
1688 1688 hash_bucket = KMEM_HASH(cp, buf);
1689 1689 bcp->bc_next = *hash_bucket;
1690 1690 *hash_bucket = bcp;
1691 1691 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1692 1692 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1693 1693 }
1694 1694 } else {
1695 1695 buf = KMEM_BUF(cp, bcp);
1696 1696 }
1697 1697
1698 1698 ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1699 1699
1700 1700 if (sp->slab_head == NULL) {
1701 1701 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1702 1702 if (new_slab) {
1703 1703 ASSERT(sp->slab_chunks == 1);
1704 1704 } else {
1705 1705 ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1706 1706 avl_remove(&cp->cache_partial_slabs, sp);
1707 1707 sp->slab_later_count = 0; /* clear history */
1708 1708 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1709 1709 sp->slab_stuck_offset = (uint32_t)-1;
1710 1710 }
1711 1711 list_insert_head(&cp->cache_complete_slabs, sp);
1712 1712 cp->cache_complete_slab_count++;
1713 1713 return (buf);
1714 1714 }
1715 1715
1716 1716 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1717 1717 /*
1718 1718 * Peek to see if the magazine layer is enabled before
1719 1719 * we prefill. We're not holding the cpu cache lock,
1720 1720 * so the peek could be wrong, but there's no harm in it.
1721 1721 */
1722 1722 if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1723 1723 (KMEM_CPU_CACHE(cp)->cc_magsize != 0)) {
1724 1724 kmem_slab_prefill(cp, sp);
1725 1725 return (buf);
1726 1726 }
1727 1727
1728 1728 if (new_slab) {
1729 1729 avl_add(&cp->cache_partial_slabs, sp);
1730 1730 return (buf);
1731 1731 }
1732 1732
1733 1733 /*
1734 1734 * The slab is now more allocated than it was, so the
1735 1735 * order remains unchanged.
1736 1736 */
1737 1737 ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1738 1738 return (buf);
1739 1739 }
1740 1740
1741 1741 /*
1742 1742 * Allocate a raw (unconstructed) buffer from cp's slab layer.
1743 1743 */
1744 1744 static void *
1745 1745 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1746 1746 {
1747 1747 kmem_slab_t *sp;
1748 1748 void *buf;
1749 1749 boolean_t test_destructor;
1750 1750
1751 1751 mutex_enter(&cp->cache_lock);
1752 1752 test_destructor = (cp->cache_slab_alloc == 0);
1753 1753 sp = avl_first(&cp->cache_partial_slabs);
1754 1754 if (sp == NULL) {
1755 1755 ASSERT(cp->cache_bufslab == 0);
1756 1756
1757 1757 /*
1758 1758 * The freelist is empty. Create a new slab.
1759 1759 */
1760 1760 mutex_exit(&cp->cache_lock);
1761 1761 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1762 1762 return (NULL);
1763 1763 }
1764 1764 mutex_enter(&cp->cache_lock);
1765 1765 cp->cache_slab_create++;
1766 1766 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1767 1767 cp->cache_bufmax = cp->cache_buftotal;
1768 1768 cp->cache_bufslab += sp->slab_chunks;
1769 1769 }
1770 1770
1771 1771 buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1772 1772 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1773 1773 (cp->cache_complete_slab_count +
1774 1774 avl_numnodes(&cp->cache_partial_slabs) +
1775 1775 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1776 1776 mutex_exit(&cp->cache_lock);
1777 1777
1778 1778 if (test_destructor && cp->cache_destructor != NULL) {
1779 1779 /*
1780 1780 * On the first kmem_slab_alloc(), assert that it is valid to
1781 1781 * call the destructor on a newly constructed object without any
1782 1782 * client involvement.
1783 1783 */
1784 1784 if ((cp->cache_constructor == NULL) ||
1785 1785 cp->cache_constructor(buf, cp->cache_private,
1786 1786 kmflag) == 0) {
1787 1787 cp->cache_destructor(buf, cp->cache_private);
1788 1788 }
1789 1789 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1790 1790 cp->cache_bufsize);
1791 1791 if (cp->cache_flags & KMF_DEADBEEF) {
1792 1792 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1793 1793 }
1794 1794 }
1795 1795
1796 1796 return (buf);
1797 1797 }
1798 1798
1799 1799 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1800 1800
1801 1801 /*
1802 1802 * Free a raw (unconstructed) buffer to cp's slab layer.
1803 1803 */
1804 1804 static void
1805 1805 kmem_slab_free(kmem_cache_t *cp, void *buf)
1806 1806 {
1807 1807 kmem_slab_t *sp;
1808 1808 kmem_bufctl_t *bcp, **prev_bcpp;
1809 1809
1810 1810 ASSERT(buf != NULL);
1811 1811
1812 1812 mutex_enter(&cp->cache_lock);
1813 1813 cp->cache_slab_free++;
1814 1814
1815 1815 if (cp->cache_flags & KMF_HASH) {
1816 1816 /*
1817 1817 * Look up buffer in allocated-address hash table.
1818 1818 */
1819 1819 prev_bcpp = KMEM_HASH(cp, buf);
1820 1820 while ((bcp = *prev_bcpp) != NULL) {
1821 1821 if (bcp->bc_addr == buf) {
1822 1822 *prev_bcpp = bcp->bc_next;
1823 1823 sp = bcp->bc_slab;
1824 1824 break;
1825 1825 }
1826 1826 cp->cache_lookup_depth++;
1827 1827 prev_bcpp = &bcp->bc_next;
1828 1828 }
1829 1829 } else {
1830 1830 bcp = KMEM_BUFCTL(cp, buf);
1831 1831 sp = KMEM_SLAB(cp, buf);
1832 1832 }
1833 1833
1834 1834 if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1835 1835 mutex_exit(&cp->cache_lock);
1836 1836 kmem_error(KMERR_BADADDR, cp, buf);
1837 1837 return;
1838 1838 }
1839 1839
1840 1840 if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1841 1841 /*
1842 1842 * If this is the buffer that prevented the consolidator from
1843 1843 * clearing the slab, we can reset the slab flags now that the
1844 1844 * buffer is freed. (It makes sense to do this in
1845 1845 * kmem_cache_free(), where the client gives up ownership of the
1846 1846 * buffer, but on the hot path the test is too expensive.)
1847 1847 */
1848 1848 kmem_slab_move_yes(cp, sp, buf);
1849 1849 }
1850 1850
1851 1851 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1852 1852 if (cp->cache_flags & KMF_CONTENTS)
1853 1853 ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1854 1854 kmem_log_enter(kmem_content_log, buf,
1855 1855 cp->cache_contents);
1856 1856 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1857 1857 }
1858 1858
1859 1859 bcp->bc_next = sp->slab_head;
1860 1860 sp->slab_head = bcp;
1861 1861
1862 1862 cp->cache_bufslab++;
1863 1863 ASSERT(sp->slab_refcnt >= 1);
1864 1864
1865 1865 if (--sp->slab_refcnt == 0) {
1866 1866 /*
1867 1867 * There are no outstanding allocations from this slab,
1868 1868 * so we can reclaim the memory.
1869 1869 */
1870 1870 if (sp->slab_chunks == 1) {
1871 1871 list_remove(&cp->cache_complete_slabs, sp);
1872 1872 cp->cache_complete_slab_count--;
1873 1873 } else {
1874 1874 avl_remove(&cp->cache_partial_slabs, sp);
1875 1875 }
1876 1876
1877 1877 cp->cache_buftotal -= sp->slab_chunks;
1878 1878 cp->cache_bufslab -= sp->slab_chunks;
1879 1879 /*
1880 1880 * Defer releasing the slab to the virtual memory subsystem
1881 1881 * while there is a pending move callback, since we guarantee
1882 1882 * that buffers passed to the move callback have only been
1883 1883 * touched by kmem or by the client itself. Since the memory
1884 1884 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1885 1885 * set at least one of the two lowest order bits, the client can
1886 1886 * test those bits in the move callback to determine whether or
1887 1887 * not it knows about the buffer (assuming that the client also
1888 1888 * sets one of those low order bits whenever it frees a buffer).
1889 1889 */
1890 1890 if (cp->cache_defrag == NULL ||
1891 1891 (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1892 1892 !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1893 1893 cp->cache_slab_destroy++;
1894 1894 mutex_exit(&cp->cache_lock);
1895 1895 kmem_slab_destroy(cp, sp);
1896 1896 } else {
1897 1897 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1898 1898 /*
1899 1899 * Slabs are inserted at both ends of the deadlist to
1900 1900 * distinguish between slabs freed while move callbacks
1901 1901 * are pending (list head) and a slab freed while the
1902 1902 * lock is dropped in kmem_move_buffers() (list tail) so
1903 1903 * that in both cases slab_destroy() is called from the
1904 1904 * right context.
1905 1905 */
1906 1906 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1907 1907 list_insert_tail(deadlist, sp);
1908 1908 } else {
1909 1909 list_insert_head(deadlist, sp);
1910 1910 }
1911 1911 cp->cache_defrag->kmd_deadcount++;
1912 1912 mutex_exit(&cp->cache_lock);
1913 1913 }
1914 1914 return;
1915 1915 }
1916 1916
1917 1917 if (bcp->bc_next == NULL) {
1918 1918 /* Transition the slab from completely allocated to partial. */
1919 1919 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1920 1920 ASSERT(sp->slab_chunks > 1);
1921 1921 list_remove(&cp->cache_complete_slabs, sp);
1922 1922 cp->cache_complete_slab_count--;
1923 1923 avl_add(&cp->cache_partial_slabs, sp);
1924 1924 } else {
1925 1925 #ifdef DEBUG
1926 1926 if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1927 1927 KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1928 1928 } else {
1929 1929 KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1930 1930 }
1931 1931 #else
1932 1932 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1933 1933 #endif
1934 1934 }
1935 1935
1936 1936 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1937 1937 (cp->cache_complete_slab_count +
1938 1938 avl_numnodes(&cp->cache_partial_slabs) +
1939 1939 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1940 1940 mutex_exit(&cp->cache_lock);
1941 1941 }
1942 1942
1943 1943 /*
1944 1944 * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1945 1945 */
1946 1946 static int
1947 1947 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1948 1948 caddr_t caller)
1949 1949 {
1950 1950 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1951 1951 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1952 1952 uint32_t mtbf;
1953 1953
1954 1954 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1955 1955 kmem_error(KMERR_BADBUFTAG, cp, buf);
1956 1956 return (-1);
1957 1957 }
1958 1958
1959 1959 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1960 1960
1961 1961 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1962 1962 kmem_error(KMERR_BADBUFCTL, cp, buf);
1963 1963 return (-1);
1964 1964 }
1965 1965
1966 1966 if (cp->cache_flags & KMF_DEADBEEF) {
1967 1967 if (!construct && (cp->cache_flags & KMF_LITE)) {
1968 1968 if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1969 1969 kmem_error(KMERR_MODIFIED, cp, buf);
1970 1970 return (-1);
1971 1971 }
1972 1972 if (cp->cache_constructor != NULL)
1973 1973 *(uint64_t *)buf = btp->bt_redzone;
1974 1974 else
1975 1975 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1976 1976 } else {
1977 1977 construct = 1;
1978 1978 if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1979 1979 KMEM_UNINITIALIZED_PATTERN, buf,
1980 1980 cp->cache_verify)) {
1981 1981 kmem_error(KMERR_MODIFIED, cp, buf);
1982 1982 return (-1);
1983 1983 }
1984 1984 }
1985 1985 }
1986 1986 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1987 1987
1988 1988 if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1989 1989 gethrtime() % mtbf == 0 &&
1990 1990 (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1991 1991 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1992 1992 if (!construct && cp->cache_destructor != NULL)
1993 1993 cp->cache_destructor(buf, cp->cache_private);
1994 1994 } else {
1995 1995 mtbf = 0;
1996 1996 }
1997 1997
1998 1998 if (mtbf || (construct && cp->cache_constructor != NULL &&
1999 1999 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
2000 2000 atomic_inc_64(&cp->cache_alloc_fail);
2001 2001 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2002 2002 if (cp->cache_flags & KMF_DEADBEEF)
2003 2003 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2004 2004 kmem_slab_free(cp, buf);
2005 2005 return (1);
2006 2006 }
2007 2007
2008 2008 if (cp->cache_flags & KMF_AUDIT) {
2009 2009 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2010 2010 }
2011 2011
2012 2012 if ((cp->cache_flags & KMF_LITE) &&
2013 2013 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2014 2014 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2015 2015 }
2016 2016
2017 2017 return (0);
2018 2018 }
2019 2019
2020 2020 static int
2021 2021 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
2022 2022 {
2023 2023 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2024 2024 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
2025 2025 kmem_slab_t *sp;
2026 2026
2027 2027 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
2028 2028 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
2029 2029 kmem_error(KMERR_DUPFREE, cp, buf);
2030 2030 return (-1);
2031 2031 }
2032 2032 sp = kmem_findslab(cp, buf);
2033 2033 if (sp == NULL || sp->slab_cache != cp)
2034 2034 kmem_error(KMERR_BADADDR, cp, buf);
2035 2035 else
2036 2036 kmem_error(KMERR_REDZONE, cp, buf);
2037 2037 return (-1);
2038 2038 }
2039 2039
2040 2040 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2041 2041
2042 2042 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2043 2043 kmem_error(KMERR_BADBUFCTL, cp, buf);
2044 2044 return (-1);
2045 2045 }
2046 2046
2047 2047 if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2048 2048 kmem_error(KMERR_REDZONE, cp, buf);
2049 2049 return (-1);
2050 2050 }
2051 2051
2052 2052 if (cp->cache_flags & KMF_AUDIT) {
2053 2053 if (cp->cache_flags & KMF_CONTENTS)
2054 2054 bcp->bc_contents = kmem_log_enter(kmem_content_log,
2055 2055 buf, cp->cache_contents);
2056 2056 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2057 2057 }
2058 2058
2059 2059 if ((cp->cache_flags & KMF_LITE) &&
2060 2060 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2061 2061 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2062 2062 }
2063 2063
2064 2064 if (cp->cache_flags & KMF_DEADBEEF) {
2065 2065 if (cp->cache_flags & KMF_LITE)
2066 2066 btp->bt_redzone = *(uint64_t *)buf;
2067 2067 else if (cp->cache_destructor != NULL)
2068 2068 cp->cache_destructor(buf, cp->cache_private);
2069 2069
2070 2070 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2071 2071 }
2072 2072
2073 2073 return (0);
2074 2074 }
2075 2075
2076 2076 /*
2077 2077 * Free each object in magazine mp to cp's slab layer, and free mp itself.
2078 2078 */
2079 2079 static void
2080 2080 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2081 2081 {
2082 2082 int round;
2083 2083
2084 2084 ASSERT(!list_link_active(&cp->cache_link) ||
2085 2085 taskq_member(kmem_taskq, curthread));
2086 2086
2087 2087 for (round = 0; round < nrounds; round++) {
2088 2088 void *buf = mp->mag_round[round];
2089 2089
2090 2090 if (cp->cache_flags & KMF_DEADBEEF) {
2091 2091 if (verify_pattern(KMEM_FREE_PATTERN, buf,
2092 2092 cp->cache_verify) != NULL) {
2093 2093 kmem_error(KMERR_MODIFIED, cp, buf);
2094 2094 continue;
2095 2095 }
2096 2096 if ((cp->cache_flags & KMF_LITE) &&
2097 2097 cp->cache_destructor != NULL) {
2098 2098 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2099 2099 *(uint64_t *)buf = btp->bt_redzone;
2100 2100 cp->cache_destructor(buf, cp->cache_private);
2101 2101 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2102 2102 }
2103 2103 } else if (cp->cache_destructor != NULL) {
2104 2104 cp->cache_destructor(buf, cp->cache_private);
2105 2105 }
2106 2106
2107 2107 kmem_slab_free(cp, buf);
2108 2108 }
2109 2109 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2110 2110 kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2111 2111 }
2112 2112
2113 2113 /*
2114 2114 * Allocate a magazine from the depot.
2115 2115 */
2116 2116 static kmem_magazine_t *
2117 2117 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2118 2118 {
2119 2119 kmem_magazine_t *mp;
2120 2120
2121 2121 /*
2122 2122 * If we can't get the depot lock without contention,
2123 2123 * update our contention count. We use the depot
2124 2124 * contention rate to determine whether we need to
2125 2125 * increase the magazine size for better scalability.
2126 2126 */
2127 2127 if (!mutex_tryenter(&cp->cache_depot_lock)) {
2128 2128 mutex_enter(&cp->cache_depot_lock);
2129 2129 cp->cache_depot_contention++;
2130 2130 }
2131 2131
2132 2132 if ((mp = mlp->ml_list) != NULL) {
2133 2133 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2134 2134 mlp->ml_list = mp->mag_next;
2135 2135 if (--mlp->ml_total < mlp->ml_min)
2136 2136 mlp->ml_min = mlp->ml_total;
2137 2137 mlp->ml_alloc++;
2138 2138 }
2139 2139
2140 2140 mutex_exit(&cp->cache_depot_lock);
2141 2141
2142 2142 return (mp);
2143 2143 }
2144 2144
2145 2145 /*
2146 2146 * Free a magazine to the depot.
2147 2147 */
2148 2148 static void
2149 2149 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2150 2150 {
2151 2151 mutex_enter(&cp->cache_depot_lock);
2152 2152 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2153 2153 mp->mag_next = mlp->ml_list;
2154 2154 mlp->ml_list = mp;
2155 2155 mlp->ml_total++;
2156 2156 mutex_exit(&cp->cache_depot_lock);
2157 2157 }
2158 2158
2159 2159 /*
2160 2160 * Update the working set statistics for cp's depot.
2161 2161 */
2162 2162 static void
2163 2163 kmem_depot_ws_update(kmem_cache_t *cp)
2164 2164 {
2165 2165 mutex_enter(&cp->cache_depot_lock);
2166 2166 cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2167 2167 cp->cache_full.ml_min = cp->cache_full.ml_total;
2168 2168 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2169 2169 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2170 2170 mutex_exit(&cp->cache_depot_lock);
2171 2171 }
2172 2172
2173 2173 /*
2174 2174 * Set the working set statistics for cp's depot to zero. (Everything is
2175 2175 * eligible for reaping.)
2176 2176 */
2177 2177 static void
2178 2178 kmem_depot_ws_zero(kmem_cache_t *cp)
2179 2179 {
2180 2180 mutex_enter(&cp->cache_depot_lock);
2181 2181 cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2182 2182 cp->cache_full.ml_min = cp->cache_full.ml_total;
2183 2183 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2184 2184 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2185 2185 mutex_exit(&cp->cache_depot_lock);
2186 2186 }
2187 2187
2188 2188 /*
2189 2189 * The number of bytes to reap before we call kpreempt(). The default (1MB)
2190 2190 * causes us to preempt reaping up to hundreds of times per second. Using a
2191 2191 * larger value (1GB) causes this to have virtually no effect.
2192 2192 */
2193 2193 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2194 2194
2195 2195 /*
2196 2196 * Reap all magazines that have fallen out of the depot's working set.
2197 2197 */
2198 2198 static void
2199 2199 kmem_depot_ws_reap(kmem_cache_t *cp)
2200 2200 {
2201 2201 size_t bytes = 0;
2202 2202 long reap;
2203 2203 kmem_magazine_t *mp;
2204 2204
2205 2205 ASSERT(!list_link_active(&cp->cache_link) ||
2206 2206 taskq_member(kmem_taskq, curthread));
2207 2207
2208 2208 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2209 2209 while (reap-- &&
2210 2210 (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2211 2211 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2212 2212 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2213 2213 if (bytes > kmem_reap_preempt_bytes) {
2214 2214 kpreempt(KPREEMPT_SYNC);
2215 2215 bytes = 0;
2216 2216 }
2217 2217 }
2218 2218
2219 2219 reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2220 2220 while (reap-- &&
2221 2221 (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2222 2222 kmem_magazine_destroy(cp, mp, 0);
2223 2223 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2224 2224 if (bytes > kmem_reap_preempt_bytes) {
2225 2225 kpreempt(KPREEMPT_SYNC);
2226 2226 bytes = 0;
2227 2227 }
2228 2228 }
2229 2229 }
2230 2230
2231 2231 static void
2232 2232 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2233 2233 {
2234 2234 ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2235 2235 (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2236 2236 ASSERT(ccp->cc_magsize > 0);
2237 2237
2238 2238 ccp->cc_ploaded = ccp->cc_loaded;
2239 2239 ccp->cc_prounds = ccp->cc_rounds;
2240 2240 ccp->cc_loaded = mp;
2241 2241 ccp->cc_rounds = rounds;
2242 2242 }
2243 2243
2244 2244 /*
2245 2245 * Intercept kmem alloc/free calls during crash dump in order to avoid
2246 2246 * changing kmem state while memory is being saved to the dump device.
2247 2247 * Otherwise, ::kmem_verify will report "corrupt buffers". Note that
2248 2248 * there are no locks because only one CPU calls kmem during a crash
2249 2249 * dump. To enable this feature, first create the associated vmem
2250 2250 * arena with VMC_DUMPSAFE.
2251 2251 */
2252 2252 static void *kmem_dump_start; /* start of pre-reserved heap */
2253 2253 static void *kmem_dump_end; /* end of heap area */
2254 2254 static void *kmem_dump_curr; /* current free heap pointer */
2255 2255 static size_t kmem_dump_size; /* size of heap area */
2256 2256
2257 2257 /* append to each buf created in the pre-reserved heap */
2258 2258 typedef struct kmem_dumpctl {
2259 2259 void *kdc_next; /* cache dump free list linkage */
2260 2260 } kmem_dumpctl_t;
2261 2261
2262 2262 #define KMEM_DUMPCTL(cp, buf) \
2263 2263 ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2264 2264 sizeof (void *)))
2265 2265
2266 2266 /* Keep some simple stats. */
2267 2267 #define KMEM_DUMP_LOGS (100)
2268 2268
2269 2269 typedef struct kmem_dump_log {
2270 2270 kmem_cache_t *kdl_cache;
2271 2271 uint_t kdl_allocs; /* # of dump allocations */
2272 2272 uint_t kdl_frees; /* # of dump frees */
2273 2273 uint_t kdl_alloc_fails; /* # of allocation failures */
2274 2274 uint_t kdl_free_nondump; /* # of non-dump frees */
2275 2275 uint_t kdl_unsafe; /* cache was used, but unsafe */
2276 2276 } kmem_dump_log_t;
2277 2277
2278 2278 static kmem_dump_log_t *kmem_dump_log;
2279 2279 static int kmem_dump_log_idx;
2280 2280
2281 2281 #define KDI_LOG(cp, stat) { \
2282 2282 kmem_dump_log_t *kdl; \
2283 2283 if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) { \
2284 2284 kdl->stat++; \
2285 2285 } else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) { \
2286 2286 kdl = &kmem_dump_log[kmem_dump_log_idx++]; \
2287 2287 kdl->stat++; \
2288 2288 kdl->kdl_cache = (cp); \
2289 2289 (cp)->cache_dumplog = kdl; \
2290 2290 } \
2291 2291 }
2292 2292
2293 2293 /* set non zero for full report */
2294 2294 uint_t kmem_dump_verbose = 0;
2295 2295
2296 2296 /* stats for overize heap */
2297 2297 uint_t kmem_dump_oversize_allocs = 0;
2298 2298 uint_t kmem_dump_oversize_max = 0;
2299 2299
2300 2300 static void
2301 2301 kmem_dumppr(char **pp, char *e, const char *format, ...)
2302 2302 {
2303 2303 char *p = *pp;
2304 2304
2305 2305 if (p < e) {
2306 2306 int n;
2307 2307 va_list ap;
2308 2308
2309 2309 va_start(ap, format);
2310 2310 n = vsnprintf(p, e - p, format, ap);
2311 2311 va_end(ap);
2312 2312 *pp = p + n;
2313 2313 }
2314 2314 }
2315 2315
2316 2316 /*
2317 2317 * Called when dumpadm(1M) configures dump parameters.
2318 2318 */
2319 2319 void
2320 2320 kmem_dump_init(size_t size)
2321 2321 {
2322 2322 if (kmem_dump_start != NULL)
2323 2323 kmem_free(kmem_dump_start, kmem_dump_size);
2324 2324
2325 2325 if (kmem_dump_log == NULL)
2326 2326 kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2327 2327 sizeof (kmem_dump_log_t), KM_SLEEP);
2328 2328
2329 2329 kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2330 2330
2331 2331 if (kmem_dump_start != NULL) {
2332 2332 kmem_dump_size = size;
2333 2333 kmem_dump_curr = kmem_dump_start;
2334 2334 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2335 2335 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2336 2336 } else {
2337 2337 kmem_dump_size = 0;
2338 2338 kmem_dump_curr = NULL;
2339 2339 kmem_dump_end = NULL;
2340 2340 }
2341 2341 }
2342 2342
2343 2343 /*
2344 2344 * Set flag for each kmem_cache_t if is safe to use alternate dump
2345 2345 * memory. Called just before panic crash dump starts. Set the flag
2346 2346 * for the calling CPU.
2347 2347 */
2348 2348 void
2349 2349 kmem_dump_begin(void)
2350 2350 {
2351 2351 ASSERT(panicstr != NULL);
2352 2352 if (kmem_dump_start != NULL) {
2353 2353 kmem_cache_t *cp;
2354 2354
2355 2355 for (cp = list_head(&kmem_caches); cp != NULL;
2356 2356 cp = list_next(&kmem_caches, cp)) {
2357 2357 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2358 2358
2359 2359 if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2360 2360 cp->cache_flags |= KMF_DUMPDIVERT;
2361 2361 ccp->cc_flags |= KMF_DUMPDIVERT;
2362 2362 ccp->cc_dump_rounds = ccp->cc_rounds;
2363 2363 ccp->cc_dump_prounds = ccp->cc_prounds;
2364 2364 ccp->cc_rounds = ccp->cc_prounds = -1;
2365 2365 } else {
2366 2366 cp->cache_flags |= KMF_DUMPUNSAFE;
2367 2367 ccp->cc_flags |= KMF_DUMPUNSAFE;
2368 2368 }
2369 2369 }
2370 2370 }
2371 2371 }
2372 2372
2373 2373 /*
2374 2374 * finished dump intercept
2375 2375 * print any warnings on the console
2376 2376 * return verbose information to dumpsys() in the given buffer
2377 2377 */
2378 2378 size_t
2379 2379 kmem_dump_finish(char *buf, size_t size)
2380 2380 {
2381 2381 int kdi_idx;
2382 2382 int kdi_end = kmem_dump_log_idx;
2383 2383 int percent = 0;
2384 2384 int header = 0;
2385 2385 int warn = 0;
2386 2386 size_t used;
2387 2387 kmem_cache_t *cp;
2388 2388 kmem_dump_log_t *kdl;
2389 2389 char *e = buf + size;
2390 2390 char *p = buf;
2391 2391
2392 2392 if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2393 2393 return (0);
2394 2394
2395 2395 used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2396 2396 percent = (used * 100) / kmem_dump_size;
2397 2397
2398 2398 kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2399 2399 kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2400 2400 kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2401 2401 kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2402 2402 kmem_dump_oversize_allocs);
2403 2403 kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2404 2404 kmem_dump_oversize_max);
2405 2405
2406 2406 for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2407 2407 kdl = &kmem_dump_log[kdi_idx];
2408 2408 cp = kdl->kdl_cache;
2409 2409 if (cp == NULL)
2410 2410 break;
2411 2411 if (kdl->kdl_alloc_fails)
2412 2412 ++warn;
2413 2413 if (header == 0) {
2414 2414 kmem_dumppr(&p, e,
2415 2415 "Cache Name,Allocs,Frees,Alloc Fails,"
2416 2416 "Nondump Frees,Unsafe Allocs/Frees\n");
2417 2417 header = 1;
2418 2418 }
2419 2419 kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2420 2420 cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2421 2421 kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2422 2422 kdl->kdl_unsafe);
2423 2423 }
2424 2424
2425 2425 /* return buffer size used */
2426 2426 if (p < e)
2427 2427 bzero(p, e - p);
2428 2428 return (p - buf);
2429 2429 }
2430 2430
2431 2431 /*
2432 2432 * Allocate a constructed object from alternate dump memory.
2433 2433 */
2434 2434 void *
2435 2435 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2436 2436 {
2437 2437 void *buf;
2438 2438 void *curr;
2439 2439 char *bufend;
2440 2440
2441 2441 /* return a constructed object */
2442 2442 if ((buf = cp->cache_dumpfreelist) != NULL) {
2443 2443 cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2444 2444 KDI_LOG(cp, kdl_allocs);
2445 2445 return (buf);
2446 2446 }
2447 2447
2448 2448 /* create a new constructed object */
2449 2449 curr = kmem_dump_curr;
2450 2450 buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2451 2451 bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2452 2452
2453 2453 /* hat layer objects cannot cross a page boundary */
2454 2454 if (cp->cache_align < PAGESIZE) {
2455 2455 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2456 2456 if (bufend > page) {
2457 2457 bufend += page - (char *)buf;
2458 2458 buf = (void *)page;
2459 2459 }
2460 2460 }
2461 2461
2462 2462 /* fall back to normal alloc if reserved area is used up */
2463 2463 if (bufend > (char *)kmem_dump_end) {
2464 2464 kmem_dump_curr = kmem_dump_end;
2465 2465 KDI_LOG(cp, kdl_alloc_fails);
2466 2466 return (NULL);
2467 2467 }
2468 2468
2469 2469 /*
2470 2470 * Must advance curr pointer before calling a constructor that
2471 2471 * may also allocate memory.
2472 2472 */
2473 2473 kmem_dump_curr = bufend;
2474 2474
2475 2475 /* run constructor */
2476 2476 if (cp->cache_constructor != NULL &&
2477 2477 cp->cache_constructor(buf, cp->cache_private, kmflag)
2478 2478 != 0) {
2479 2479 #ifdef DEBUG
2480 2480 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2481 2481 cp->cache_name, (void *)cp);
2482 2482 #endif
2483 2483 /* reset curr pointer iff no allocs were done */
2484 2484 if (kmem_dump_curr == bufend)
2485 2485 kmem_dump_curr = curr;
2486 2486
2487 2487 /* fall back to normal alloc if the constructor fails */
2488 2488 KDI_LOG(cp, kdl_alloc_fails);
2489 2489 return (NULL);
2490 2490 }
2491 2491
2492 2492 KDI_LOG(cp, kdl_allocs);
2493 2493 return (buf);
2494 2494 }
2495 2495
2496 2496 /*
2497 2497 * Free a constructed object in alternate dump memory.
2498 2498 */
2499 2499 int
2500 2500 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2501 2501 {
2502 2502 /* save constructed buffers for next time */
2503 2503 if ((char *)buf >= (char *)kmem_dump_start &&
2504 2504 (char *)buf < (char *)kmem_dump_end) {
2505 2505 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2506 2506 cp->cache_dumpfreelist = buf;
2507 2507 KDI_LOG(cp, kdl_frees);
2508 2508 return (0);
2509 2509 }
2510 2510
2511 2511 /* count all non-dump buf frees */
2512 2512 KDI_LOG(cp, kdl_free_nondump);
2513 2513
2514 2514 /* just drop buffers that were allocated before dump started */
2515 2515 if (kmem_dump_curr < kmem_dump_end)
2516 2516 return (0);
2517 2517
2518 2518 /* fall back to normal free if reserved area is used up */
2519 2519 return (1);
2520 2520 }
2521 2521
2522 2522 /*
2523 2523 * Allocate a constructed object from cache cp.
2524 2524 */
2525 2525 void *
2526 2526 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2527 2527 {
2528 2528 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2529 2529 kmem_magazine_t *fmp;
2530 2530 void *buf;
2531 2531
2532 2532 mutex_enter(&ccp->cc_lock);
2533 2533 for (;;) {
2534 2534 /*
2535 2535 * If there's an object available in the current CPU's
2536 2536 * loaded magazine, just take it and return.
2537 2537 */
2538 2538 if (ccp->cc_rounds > 0) {
2539 2539 buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2540 2540 ccp->cc_alloc++;
2541 2541 mutex_exit(&ccp->cc_lock);
2542 2542 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2543 2543 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2544 2544 ASSERT(!(ccp->cc_flags &
2545 2545 KMF_DUMPDIVERT));
2546 2546 KDI_LOG(cp, kdl_unsafe);
2547 2547 }
2548 2548 if ((ccp->cc_flags & KMF_BUFTAG) &&
2549 2549 kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2550 2550 caller()) != 0) {
2551 2551 if (kmflag & KM_NOSLEEP)
2552 2552 return (NULL);
2553 2553 mutex_enter(&ccp->cc_lock);
2554 2554 continue;
2555 2555 }
2556 2556 }
2557 2557 return (buf);
2558 2558 }
2559 2559
2560 2560 /*
2561 2561 * The loaded magazine is empty. If the previously loaded
2562 2562 * magazine was full, exchange them and try again.
2563 2563 */
2564 2564 if (ccp->cc_prounds > 0) {
2565 2565 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2566 2566 continue;
2567 2567 }
2568 2568
2569 2569 /*
2570 2570 * Return an alternate buffer at dump time to preserve
2571 2571 * the heap.
2572 2572 */
2573 2573 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2574 2574 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2575 2575 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2576 2576 /* log it so that we can warn about it */
2577 2577 KDI_LOG(cp, kdl_unsafe);
2578 2578 } else {
2579 2579 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2580 2580 NULL) {
2581 2581 mutex_exit(&ccp->cc_lock);
2582 2582 return (buf);
2583 2583 }
2584 2584 break; /* fall back to slab layer */
2585 2585 }
2586 2586 }
2587 2587
2588 2588 /*
2589 2589 * If the magazine layer is disabled, break out now.
2590 2590 */
2591 2591 if (ccp->cc_magsize == 0)
2592 2592 break;
2593 2593
2594 2594 /*
2595 2595 * Try to get a full magazine from the depot.
2596 2596 */
2597 2597 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2598 2598 if (fmp != NULL) {
2599 2599 if (ccp->cc_ploaded != NULL)
2600 2600 kmem_depot_free(cp, &cp->cache_empty,
2601 2601 ccp->cc_ploaded);
2602 2602 kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2603 2603 continue;
2604 2604 }
2605 2605
2606 2606 /*
2607 2607 * There are no full magazines in the depot,
2608 2608 * so fall through to the slab layer.
2609 2609 */
2610 2610 break;
2611 2611 }
2612 2612 mutex_exit(&ccp->cc_lock);
2613 2613
2614 2614 /*
2615 2615 * We couldn't allocate a constructed object from the magazine layer,
2616 2616 * so get a raw buffer from the slab layer and apply its constructor.
2617 2617 */
2618 2618 buf = kmem_slab_alloc(cp, kmflag);
2619 2619
2620 2620 if (buf == NULL)
2621 2621 return (NULL);
2622 2622
2623 2623 if (cp->cache_flags & KMF_BUFTAG) {
2624 2624 /*
2625 2625 * Make kmem_cache_alloc_debug() apply the constructor for us.
2626 2626 */
2627 2627 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2628 2628 if (rc != 0) {
2629 2629 if (kmflag & KM_NOSLEEP)
2630 2630 return (NULL);
2631 2631 /*
2632 2632 * kmem_cache_alloc_debug() detected corruption
2633 2633 * but didn't panic (kmem_panic <= 0). We should not be
2634 2634 * here because the constructor failed (indicated by a
2635 2635 * return code of 1). Try again.
2636 2636 */
2637 2637 ASSERT(rc == -1);
2638 2638 return (kmem_cache_alloc(cp, kmflag));
2639 2639 }
2640 2640 return (buf);
2641 2641 }
2642 2642
2643 2643 if (cp->cache_constructor != NULL &&
2644 2644 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2645 2645 atomic_inc_64(&cp->cache_alloc_fail);
2646 2646 kmem_slab_free(cp, buf);
2647 2647 return (NULL);
2648 2648 }
2649 2649
2650 2650 return (buf);
2651 2651 }
2652 2652
2653 2653 /*
2654 2654 * The freed argument tells whether or not kmem_cache_free_debug() has already
2655 2655 * been called so that we can avoid the duplicate free error. For example, a
2656 2656 * buffer on a magazine has already been freed by the client but is still
2657 2657 * constructed.
2658 2658 */
2659 2659 static void
2660 2660 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2661 2661 {
2662 2662 if (!freed && (cp->cache_flags & KMF_BUFTAG))
2663 2663 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2664 2664 return;
2665 2665
2666 2666 /*
2667 2667 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2668 2668 * kmem_cache_free_debug() will have already applied the destructor.
2669 2669 */
2670 2670 if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2671 2671 cp->cache_destructor != NULL) {
2672 2672 if (cp->cache_flags & KMF_DEADBEEF) { /* KMF_LITE implied */
2673 2673 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2674 2674 *(uint64_t *)buf = btp->bt_redzone;
2675 2675 cp->cache_destructor(buf, cp->cache_private);
2676 2676 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2677 2677 } else {
2678 2678 cp->cache_destructor(buf, cp->cache_private);
2679 2679 }
2680 2680 }
2681 2681
2682 2682 kmem_slab_free(cp, buf);
2683 2683 }
2684 2684
2685 2685 /*
2686 2686 * Used when there's no room to free a buffer to the per-CPU cache.
2687 2687 * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2688 2688 * caller should try freeing to the per-CPU cache again.
2689 2689 * Note that we don't directly install the magazine in the cpu cache,
2690 2690 * since its state may have changed wildly while the lock was dropped.
2691 2691 */
2692 2692 static int
2693 2693 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2694 2694 {
2695 2695 kmem_magazine_t *emp;
2696 2696 kmem_magtype_t *mtp;
2697 2697
2698 2698 ASSERT(MUTEX_HELD(&ccp->cc_lock));
2699 2699 ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2700 2700 ((uint_t)ccp->cc_rounds == -1)) &&
2701 2701 ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2702 2702 ((uint_t)ccp->cc_prounds == -1)));
2703 2703
2704 2704 emp = kmem_depot_alloc(cp, &cp->cache_empty);
2705 2705 if (emp != NULL) {
2706 2706 if (ccp->cc_ploaded != NULL)
2707 2707 kmem_depot_free(cp, &cp->cache_full,
2708 2708 ccp->cc_ploaded);
2709 2709 kmem_cpu_reload(ccp, emp, 0);
2710 2710 return (1);
2711 2711 }
2712 2712 /*
2713 2713 * There are no empty magazines in the depot,
2714 2714 * so try to allocate a new one. We must drop all locks
2715 2715 * across kmem_cache_alloc() because lower layers may
2716 2716 * attempt to allocate from this cache.
2717 2717 */
2718 2718 mtp = cp->cache_magtype;
2719 2719 mutex_exit(&ccp->cc_lock);
2720 2720 emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2721 2721 mutex_enter(&ccp->cc_lock);
2722 2722
2723 2723 if (emp != NULL) {
2724 2724 /*
2725 2725 * We successfully allocated an empty magazine.
2726 2726 * However, we had to drop ccp->cc_lock to do it,
2727 2727 * so the cache's magazine size may have changed.
2728 2728 * If so, free the magazine and try again.
2729 2729 */
2730 2730 if (ccp->cc_magsize != mtp->mt_magsize) {
2731 2731 mutex_exit(&ccp->cc_lock);
2732 2732 kmem_cache_free(mtp->mt_cache, emp);
2733 2733 mutex_enter(&ccp->cc_lock);
2734 2734 return (1);
2735 2735 }
2736 2736
2737 2737 /*
2738 2738 * We got a magazine of the right size. Add it to
2739 2739 * the depot and try the whole dance again.
2740 2740 */
2741 2741 kmem_depot_free(cp, &cp->cache_empty, emp);
2742 2742 return (1);
2743 2743 }
2744 2744
2745 2745 /*
2746 2746 * We couldn't allocate an empty magazine,
2747 2747 * so fall through to the slab layer.
2748 2748 */
2749 2749 return (0);
2750 2750 }
2751 2751
2752 2752 /*
2753 2753 * Free a constructed object to cache cp.
2754 2754 */
2755 2755 void
2756 2756 kmem_cache_free(kmem_cache_t *cp, void *buf)
2757 2757 {
2758 2758 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2759 2759
2760 2760 /*
2761 2761 * The client must not free either of the buffers passed to the move
2762 2762 * callback function.
2763 2763 */
2764 2764 ASSERT(cp->cache_defrag == NULL ||
2765 2765 cp->cache_defrag->kmd_thread != curthread ||
2766 2766 (buf != cp->cache_defrag->kmd_from_buf &&
2767 2767 buf != cp->cache_defrag->kmd_to_buf));
2768 2768
2769 2769 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2770 2770 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2771 2771 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2772 2772 /* log it so that we can warn about it */
2773 2773 KDI_LOG(cp, kdl_unsafe);
2774 2774 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2775 2775 return;
2776 2776 }
2777 2777 if (ccp->cc_flags & KMF_BUFTAG) {
2778 2778 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2779 2779 return;
2780 2780 }
2781 2781 }
2782 2782
2783 2783 mutex_enter(&ccp->cc_lock);
2784 2784 /*
2785 2785 * Any changes to this logic should be reflected in kmem_slab_prefill()
2786 2786 */
2787 2787 for (;;) {
2788 2788 /*
2789 2789 * If there's a slot available in the current CPU's
2790 2790 * loaded magazine, just put the object there and return.
2791 2791 */
2792 2792 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2793 2793 ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2794 2794 ccp->cc_free++;
2795 2795 mutex_exit(&ccp->cc_lock);
2796 2796 return;
2797 2797 }
2798 2798
2799 2799 /*
2800 2800 * The loaded magazine is full. If the previously loaded
2801 2801 * magazine was empty, exchange them and try again.
2802 2802 */
2803 2803 if (ccp->cc_prounds == 0) {
2804 2804 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2805 2805 continue;
2806 2806 }
2807 2807
2808 2808 /*
2809 2809 * If the magazine layer is disabled, break out now.
2810 2810 */
2811 2811 if (ccp->cc_magsize == 0)
2812 2812 break;
2813 2813
2814 2814 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2815 2815 /*
2816 2816 * We couldn't free our constructed object to the
2817 2817 * magazine layer, so apply its destructor and free it
2818 2818 * to the slab layer.
2819 2819 */
2820 2820 break;
2821 2821 }
2822 2822 }
2823 2823 mutex_exit(&ccp->cc_lock);
2824 2824 kmem_slab_free_constructed(cp, buf, B_TRUE);
2825 2825 }
2826 2826
2827 2827 static void
2828 2828 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2829 2829 {
2830 2830 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2831 2831 int cache_flags = cp->cache_flags;
2832 2832
2833 2833 kmem_bufctl_t *next, *head;
2834 2834 size_t nbufs;
2835 2835
2836 2836 /*
2837 2837 * Completely allocate the newly created slab and put the pre-allocated
2838 2838 * buffers in magazines. Any of the buffers that cannot be put in
2839 2839 * magazines must be returned to the slab.
2840 2840 */
2841 2841 ASSERT(MUTEX_HELD(&cp->cache_lock));
2842 2842 ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2843 2843 ASSERT(cp->cache_constructor == NULL);
2844 2844 ASSERT(sp->slab_cache == cp);
2845 2845 ASSERT(sp->slab_refcnt == 1);
2846 2846 ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2847 2847 ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2848 2848
2849 2849 head = sp->slab_head;
2850 2850 nbufs = (sp->slab_chunks - sp->slab_refcnt);
2851 2851 sp->slab_head = NULL;
2852 2852 sp->slab_refcnt += nbufs;
2853 2853 cp->cache_bufslab -= nbufs;
2854 2854 cp->cache_slab_alloc += nbufs;
2855 2855 list_insert_head(&cp->cache_complete_slabs, sp);
2856 2856 cp->cache_complete_slab_count++;
2857 2857 mutex_exit(&cp->cache_lock);
2858 2858 mutex_enter(&ccp->cc_lock);
2859 2859
2860 2860 while (head != NULL) {
2861 2861 void *buf = KMEM_BUF(cp, head);
2862 2862 /*
2863 2863 * If there's a slot available in the current CPU's
2864 2864 * loaded magazine, just put the object there and
2865 2865 * continue.
2866 2866 */
2867 2867 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2868 2868 ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2869 2869 buf;
2870 2870 ccp->cc_free++;
2871 2871 nbufs--;
2872 2872 head = head->bc_next;
2873 2873 continue;
2874 2874 }
2875 2875
2876 2876 /*
2877 2877 * The loaded magazine is full. If the previously
2878 2878 * loaded magazine was empty, exchange them and try
2879 2879 * again.
2880 2880 */
2881 2881 if (ccp->cc_prounds == 0) {
2882 2882 kmem_cpu_reload(ccp, ccp->cc_ploaded,
2883 2883 ccp->cc_prounds);
2884 2884 continue;
2885 2885 }
2886 2886
2887 2887 /*
2888 2888 * If the magazine layer is disabled, break out now.
2889 2889 */
2890 2890
2891 2891 if (ccp->cc_magsize == 0) {
2892 2892 break;
2893 2893 }
2894 2894
2895 2895 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2896 2896 break;
2897 2897 }
2898 2898 mutex_exit(&ccp->cc_lock);
2899 2899 if (nbufs != 0) {
2900 2900 ASSERT(head != NULL);
2901 2901
2902 2902 /*
2903 2903 * If there was a failure, return remaining objects to
2904 2904 * the slab
2905 2905 */
2906 2906 while (head != NULL) {
2907 2907 ASSERT(nbufs != 0);
2908 2908 next = head->bc_next;
2909 2909 head->bc_next = NULL;
2910 2910 kmem_slab_free(cp, KMEM_BUF(cp, head));
2911 2911 head = next;
2912 2912 nbufs--;
2913 2913 }
2914 2914 }
2915 2915 ASSERT(head == NULL);
2916 2916 ASSERT(nbufs == 0);
2917 2917 mutex_enter(&cp->cache_lock);
2918 2918 }
2919 2919
2920 2920 void *
2921 2921 kmem_zalloc(size_t size, int kmflag)
2922 2922 {
2923 2923 size_t index;
2924 2924 void *buf;
2925 2925
2926 2926 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2927 2927 kmem_cache_t *cp = kmem_alloc_table[index];
2928 2928 buf = kmem_cache_alloc(cp, kmflag);
2929 2929 if (buf != NULL) {
2930 2930 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2931 2931 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2932 2932 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2933 2933 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2934 2934
2935 2935 if (cp->cache_flags & KMF_LITE) {
2936 2936 KMEM_BUFTAG_LITE_ENTER(btp,
2937 2937 kmem_lite_count, caller());
2938 2938 }
2939 2939 }
2940 2940 bzero(buf, size);
2941 2941 }
2942 2942 } else {
2943 2943 buf = kmem_alloc(size, kmflag);
2944 2944 if (buf != NULL)
2945 2945 bzero(buf, size);
2946 2946 }
2947 2947 return (buf);
2948 2948 }
2949 2949
2950 2950 void *
2951 2951 kmem_alloc(size_t size, int kmflag)
2952 2952 {
2953 2953 size_t index;
2954 2954 kmem_cache_t *cp;
2955 2955 void *buf;
2956 2956
2957 2957 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2958 2958 cp = kmem_alloc_table[index];
2959 2959 /* fall through to kmem_cache_alloc() */
2960 2960
2961 2961 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2962 2962 kmem_big_alloc_table_max) {
2963 2963 cp = kmem_big_alloc_table[index];
2964 2964 /* fall through to kmem_cache_alloc() */
2965 2965
2966 2966 } else {
2967 2967 if (size == 0)
2968 2968 return (NULL);
2969 2969
2970 2970 buf = vmem_alloc(kmem_oversize_arena, size,
2971 2971 kmflag & KM_VMFLAGS);
2972 2972 if (buf == NULL)
2973 2973 kmem_log_event(kmem_failure_log, NULL, NULL,
2974 2974 (void *)size);
2975 2975 else if (KMEM_DUMP(kmem_slab_cache)) {
2976 2976 /* stats for dump intercept */
2977 2977 kmem_dump_oversize_allocs++;
2978 2978 if (size > kmem_dump_oversize_max)
2979 2979 kmem_dump_oversize_max = size;
2980 2980 }
2981 2981 return (buf);
2982 2982 }
2983 2983
2984 2984 buf = kmem_cache_alloc(cp, kmflag);
2985 2985 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2986 2986 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2987 2987 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2988 2988 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2989 2989
2990 2990 if (cp->cache_flags & KMF_LITE) {
2991 2991 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2992 2992 }
2993 2993 }
2994 2994 return (buf);
2995 2995 }
2996 2996
2997 2997 void
2998 2998 kmem_free(void *buf, size_t size)
2999 2999 {
3000 3000 size_t index;
3001 3001 kmem_cache_t *cp;
3002 3002
3003 3003 if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
3004 3004 cp = kmem_alloc_table[index];
3005 3005 /* fall through to kmem_cache_free() */
3006 3006
3007 3007 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
3008 3008 kmem_big_alloc_table_max) {
3009 3009 cp = kmem_big_alloc_table[index];
3010 3010 /* fall through to kmem_cache_free() */
3011 3011
3012 3012 } else {
3013 3013 EQUIV(buf == NULL, size == 0);
3014 3014 if (buf == NULL && size == 0)
3015 3015 return;
3016 3016 vmem_free(kmem_oversize_arena, buf, size);
3017 3017 return;
3018 3018 }
3019 3019
3020 3020 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
3021 3021 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
3022 3022 uint32_t *ip = (uint32_t *)btp;
3023 3023 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3024 3024 if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3025 3025 kmem_error(KMERR_DUPFREE, cp, buf);
3026 3026 return;
3027 3027 }
3028 3028 if (KMEM_SIZE_VALID(ip[1])) {
3029 3029 ip[0] = KMEM_SIZE_ENCODE(size);
3030 3030 kmem_error(KMERR_BADSIZE, cp, buf);
3031 3031 } else {
3032 3032 kmem_error(KMERR_REDZONE, cp, buf);
3033 3033 }
3034 3034 return;
3035 3035 }
3036 3036 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3037 3037 kmem_error(KMERR_REDZONE, cp, buf);
3038 3038 return;
3039 3039 }
3040 3040 btp->bt_redzone = KMEM_REDZONE_PATTERN;
3041 3041 if (cp->cache_flags & KMF_LITE) {
3042 3042 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3043 3043 caller());
3044 3044 }
3045 3045 }
3046 3046 kmem_cache_free(cp, buf);
3047 3047 }
3048 3048
3049 3049 void *
3050 3050 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3051 3051 {
3052 3052 size_t realsize = size + vmp->vm_quantum;
3053 3053 void *addr;
3054 3054
3055 3055 /*
3056 3056 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3057 3057 * vm_quantum will cause integer wraparound. Check for this, and
3058 3058 * blow off the firewall page in this case. Note that such a
3059 3059 * giant allocation (the entire kernel address space) can never
3060 3060 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3061 3061 * or sleep forever (VM_SLEEP). Thus, there is no need for a
3062 3062 * corresponding check in kmem_firewall_va_free().
3063 3063 */
3064 3064 if (realsize < size)
3065 3065 realsize = size;
3066 3066
3067 3067 /*
3068 3068 * While boot still owns resource management, make sure that this
3069 3069 * redzone virtual address allocation is properly accounted for in
3070 3070 * OBPs "virtual-memory" "available" lists because we're
3071 3071 * effectively claiming them for a red zone. If we don't do this,
3072 3072 * the available lists become too fragmented and too large for the
3073 3073 * current boot/kernel memory list interface.
3074 3074 */
3075 3075 addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3076 3076
3077 3077 if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3078 3078 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3079 3079
3080 3080 return (addr);
3081 3081 }
3082 3082
3083 3083 void
3084 3084 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3085 3085 {
3086 3086 ASSERT((kvseg.s_base == NULL ?
3087 3087 va_to_pfn((char *)addr + size) :
3088 3088 hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3089 3089
3090 3090 vmem_free(vmp, addr, size + vmp->vm_quantum);
3091 3091 }
3092 3092
3093 3093 /*
3094 3094 * Try to allocate at least `size' bytes of memory without sleeping or
3095 3095 * panicking. Return actual allocated size in `asize'. If allocation failed,
3096 3096 * try final allocation with sleep or panic allowed.
3097 3097 */
3098 3098 void *
3099 3099 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3100 3100 {
3101 3101 void *p;
3102 3102
3103 3103 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3104 3104 do {
3105 3105 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3106 3106 if (p != NULL)
3107 3107 return (p);
3108 3108 *asize += KMEM_ALIGN;
3109 3109 } while (*asize <= PAGESIZE);
3110 3110
3111 3111 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3112 3112 return (kmem_alloc(*asize, kmflag));
3113 3113 }
3114 3114
3115 3115 /*
3116 3116 * Reclaim all unused memory from a cache.
3117 3117 */
3118 3118 static void
3119 3119 kmem_cache_reap(kmem_cache_t *cp)
3120 3120 {
3121 3121 ASSERT(taskq_member(kmem_taskq, curthread));
3122 3122 cp->cache_reap++;
3123 3123
3124 3124 /*
3125 3125 * Ask the cache's owner to free some memory if possible.
3126 3126 * The idea is to handle things like the inode cache, which
3127 3127 * typically sits on a bunch of memory that it doesn't truly
3128 3128 * *need*. Reclaim policy is entirely up to the owner; this
3129 3129 * callback is just an advisory plea for help.
3130 3130 */
3131 3131 if (cp->cache_reclaim != NULL) {
3132 3132 long delta;
3133 3133
3134 3134 /*
3135 3135 * Reclaimed memory should be reapable (not included in the
3136 3136 * depot's working set).
3137 3137 */
3138 3138 delta = cp->cache_full.ml_total;
3139 3139 cp->cache_reclaim(cp->cache_private);
3140 3140 delta = cp->cache_full.ml_total - delta;
3141 3141 if (delta > 0) {
3142 3142 mutex_enter(&cp->cache_depot_lock);
3143 3143 cp->cache_full.ml_reaplimit += delta;
3144 3144 cp->cache_full.ml_min += delta;
3145 3145 mutex_exit(&cp->cache_depot_lock);
3146 3146 }
3147 3147 }
3148 3148
3149 3149 kmem_depot_ws_reap(cp);
3150 3150
3151 3151 if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3152 3152 kmem_cache_defrag(cp);
3153 3153 }
3154 3154 }
3155 3155
3156 3156 static void
3157 3157 kmem_reap_timeout(void *flag_arg)
3158 3158 {
3159 3159 uint32_t *flag = (uint32_t *)flag_arg;
3160 3160
3161 3161 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3162 3162 *flag = 0;
3163 3163 }
3164 3164
3165 3165 static void
3166 3166 kmem_reap_done(void *flag)
3167 3167 {
3168 3168 if (!callout_init_done) {
3169 3169 /* can't schedule a timeout at this point */
3170 3170 kmem_reap_timeout(flag);
3171 3171 } else {
3172 3172 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3173 3173 }
3174 3174 }
3175 3175
3176 3176 static void
3177 3177 kmem_reap_start(void *flag)
3178 3178 {
3179 3179 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3180 3180
3181 3181 if (flag == &kmem_reaping) {
3182 3182 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3183 3183 /*
3184 3184 * if we have segkp under heap, reap segkp cache.
3185 3185 */
3186 3186 if (segkp_fromheap)
3187 3187 segkp_cache_free();
3188 3188 }
3189 3189 else
3190 3190 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3191 3191
3192 3192 /*
3193 3193 * We use taskq_dispatch() to schedule a timeout to clear
3194 3194 * the flag so that kmem_reap() becomes self-throttling:
3195 3195 * we won't reap again until the current reap completes *and*
3196 3196 * at least kmem_reap_interval ticks have elapsed.
3197 3197 */
3198 3198 if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3199 3199 kmem_reap_done(flag);
3200 3200 }
3201 3201
3202 3202 static void
3203 3203 kmem_reap_common(void *flag_arg)
3204 3204 {
3205 3205 uint32_t *flag = (uint32_t *)flag_arg;
3206 3206
3207 3207 if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3208 3208 atomic_cas_32(flag, 0, 1) != 0)
3209 3209 return;
3210 3210
3211 3211 /*
3212 3212 * It may not be kosher to do memory allocation when a reap is called
3213 3213 * (for example, if vmem_populate() is in the call chain). So we
3214 3214 * start the reap going with a TQ_NOALLOC dispatch. If the dispatch
3215 3215 * fails, we reset the flag, and the next reap will try again.
3216 3216 */
3217 3217 if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3218 3218 *flag = 0;
3219 3219 }
3220 3220
3221 3221 /*
3222 3222 * Reclaim all unused memory from all caches. Called from the VM system
3223 3223 * when memory gets tight.
3224 3224 */
3225 3225 void
3226 3226 kmem_reap(void)
3227 3227 {
3228 3228 kmem_reap_common(&kmem_reaping);
3229 3229 }
3230 3230
3231 3231 /*
3232 3232 * Reclaim all unused memory from identifier arenas, called when a vmem
3233 3233 * arena not back by memory is exhausted. Since reaping memory-backed caches
3234 3234 * cannot help with identifier exhaustion, we avoid both a large amount of
3235 3235 * work and unwanted side-effects from reclaim callbacks.
3236 3236 */
3237 3237 void
3238 3238 kmem_reap_idspace(void)
3239 3239 {
3240 3240 kmem_reap_common(&kmem_reaping_idspace);
3241 3241 }
3242 3242
3243 3243 /*
3244 3244 * Purge all magazines from a cache and set its magazine limit to zero.
3245 3245 * All calls are serialized by the kmem_taskq lock, except for the final
3246 3246 * call from kmem_cache_destroy().
3247 3247 */
3248 3248 static void
3249 3249 kmem_cache_magazine_purge(kmem_cache_t *cp)
3250 3250 {
3251 3251 kmem_cpu_cache_t *ccp;
3252 3252 kmem_magazine_t *mp, *pmp;
3253 3253 int rounds, prounds, cpu_seqid;
3254 3254
3255 3255 ASSERT(!list_link_active(&cp->cache_link) ||
3256 3256 taskq_member(kmem_taskq, curthread));
3257 3257 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3258 3258
3259 3259 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3260 3260 ccp = &cp->cache_cpu[cpu_seqid];
3261 3261
3262 3262 mutex_enter(&ccp->cc_lock);
3263 3263 mp = ccp->cc_loaded;
3264 3264 pmp = ccp->cc_ploaded;
3265 3265 rounds = ccp->cc_rounds;
3266 3266 prounds = ccp->cc_prounds;
3267 3267 ccp->cc_loaded = NULL;
3268 3268 ccp->cc_ploaded = NULL;
3269 3269 ccp->cc_rounds = -1;
3270 3270 ccp->cc_prounds = -1;
3271 3271 ccp->cc_magsize = 0;
3272 3272 mutex_exit(&ccp->cc_lock);
3273 3273
3274 3274 if (mp)
3275 3275 kmem_magazine_destroy(cp, mp, rounds);
3276 3276 if (pmp)
3277 3277 kmem_magazine_destroy(cp, pmp, prounds);
3278 3278 }
3279 3279
3280 3280 kmem_depot_ws_zero(cp);
3281 3281 kmem_depot_ws_reap(cp);
3282 3282 }
3283 3283
3284 3284 /*
3285 3285 * Enable per-cpu magazines on a cache.
3286 3286 */
3287 3287 static void
3288 3288 kmem_cache_magazine_enable(kmem_cache_t *cp)
3289 3289 {
3290 3290 int cpu_seqid;
3291 3291
3292 3292 if (cp->cache_flags & KMF_NOMAGAZINE)
3293 3293 return;
3294 3294
3295 3295 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3296 3296 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3297 3297 mutex_enter(&ccp->cc_lock);
3298 3298 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3299 3299 mutex_exit(&ccp->cc_lock);
3300 3300 }
3301 3301
3302 3302 }
3303 3303
3304 3304 /*
3305 3305 * Reap (almost) everything right now.
3306 3306 */
3307 3307 void
3308 3308 kmem_cache_reap_now(kmem_cache_t *cp)
3309 3309 {
3310 3310 ASSERT(list_link_active(&cp->cache_link));
3311 3311
3312 3312 kmem_depot_ws_zero(cp);
3313 3313
3314 3314 (void) taskq_dispatch(kmem_taskq,
3315 3315 (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3316 3316 taskq_wait(kmem_taskq);
3317 3317 }
3318 3318
3319 3319 /*
3320 3320 * Recompute a cache's magazine size. The trade-off is that larger magazines
3321 3321 * provide a higher transfer rate with the depot, while smaller magazines
3322 3322 * reduce memory consumption. Magazine resizing is an expensive operation;
3323 3323 * it should not be done frequently.
3324 3324 *
3325 3325 * Changes to the magazine size are serialized by the kmem_taskq lock.
3326 3326 *
3327 3327 * Note: at present this only grows the magazine size. It might be useful
3328 3328 * to allow shrinkage too.
3329 3329 */
3330 3330 static void
3331 3331 kmem_cache_magazine_resize(kmem_cache_t *cp)
3332 3332 {
3333 3333 kmem_magtype_t *mtp = cp->cache_magtype;
3334 3334
3335 3335 ASSERT(taskq_member(kmem_taskq, curthread));
3336 3336
3337 3337 if (cp->cache_chunksize < mtp->mt_maxbuf) {
3338 3338 kmem_cache_magazine_purge(cp);
3339 3339 mutex_enter(&cp->cache_depot_lock);
3340 3340 cp->cache_magtype = ++mtp;
3341 3341 cp->cache_depot_contention_prev =
3342 3342 cp->cache_depot_contention + INT_MAX;
3343 3343 mutex_exit(&cp->cache_depot_lock);
3344 3344 kmem_cache_magazine_enable(cp);
3345 3345 }
3346 3346 }
3347 3347
3348 3348 /*
3349 3349 * Rescale a cache's hash table, so that the table size is roughly the
3350 3350 * cache size. We want the average lookup time to be extremely small.
3351 3351 */
3352 3352 static void
3353 3353 kmem_hash_rescale(kmem_cache_t *cp)
3354 3354 {
3355 3355 kmem_bufctl_t **old_table, **new_table, *bcp;
3356 3356 size_t old_size, new_size, h;
3357 3357
3358 3358 ASSERT(taskq_member(kmem_taskq, curthread));
3359 3359
3360 3360 new_size = MAX(KMEM_HASH_INITIAL,
3361 3361 1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3362 3362 old_size = cp->cache_hash_mask + 1;
3363 3363
3364 3364 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3365 3365 return;
3366 3366
3367 3367 new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3368 3368 VM_NOSLEEP);
3369 3369 if (new_table == NULL)
3370 3370 return;
3371 3371 bzero(new_table, new_size * sizeof (void *));
3372 3372
3373 3373 mutex_enter(&cp->cache_lock);
3374 3374
3375 3375 old_size = cp->cache_hash_mask + 1;
3376 3376 old_table = cp->cache_hash_table;
3377 3377
3378 3378 cp->cache_hash_mask = new_size - 1;
3379 3379 cp->cache_hash_table = new_table;
3380 3380 cp->cache_rescale++;
3381 3381
3382 3382 for (h = 0; h < old_size; h++) {
3383 3383 bcp = old_table[h];
3384 3384 while (bcp != NULL) {
3385 3385 void *addr = bcp->bc_addr;
3386 3386 kmem_bufctl_t *next_bcp = bcp->bc_next;
3387 3387 kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3388 3388 bcp->bc_next = *hash_bucket;
3389 3389 *hash_bucket = bcp;
3390 3390 bcp = next_bcp;
3391 3391 }
3392 3392 }
3393 3393
3394 3394 mutex_exit(&cp->cache_lock);
3395 3395
3396 3396 vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3397 3397 }
3398 3398
3399 3399 /*
3400 3400 * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3401 3401 * update, magazine resizing, and slab consolidation.
3402 3402 */
3403 3403 static void
3404 3404 kmem_cache_update(kmem_cache_t *cp)
3405 3405 {
3406 3406 int need_hash_rescale = 0;
3407 3407 int need_magazine_resize = 0;
3408 3408
3409 3409 ASSERT(MUTEX_HELD(&kmem_cache_lock));
3410 3410
3411 3411 /*
3412 3412 * If the cache has become much larger or smaller than its hash table,
3413 3413 * fire off a request to rescale the hash table.
3414 3414 */
3415 3415 mutex_enter(&cp->cache_lock);
3416 3416
3417 3417 if ((cp->cache_flags & KMF_HASH) &&
3418 3418 (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3419 3419 (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3420 3420 cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3421 3421 need_hash_rescale = 1;
3422 3422
3423 3423 mutex_exit(&cp->cache_lock);
3424 3424
3425 3425 /*
3426 3426 * Update the depot working set statistics.
3427 3427 */
3428 3428 kmem_depot_ws_update(cp);
3429 3429
3430 3430 /*
3431 3431 * If there's a lot of contention in the depot,
3432 3432 * increase the magazine size.
3433 3433 */
3434 3434 mutex_enter(&cp->cache_depot_lock);
3435 3435
3436 3436 if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3437 3437 (int)(cp->cache_depot_contention -
3438 3438 cp->cache_depot_contention_prev) > kmem_depot_contention)
3439 3439 need_magazine_resize = 1;
3440 3440
3441 3441 cp->cache_depot_contention_prev = cp->cache_depot_contention;
3442 3442
3443 3443 mutex_exit(&cp->cache_depot_lock);
3444 3444
3445 3445 if (need_hash_rescale)
3446 3446 (void) taskq_dispatch(kmem_taskq,
3447 3447 (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3448 3448
3449 3449 if (need_magazine_resize)
3450 3450 (void) taskq_dispatch(kmem_taskq,
3451 3451 (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3452 3452
3453 3453 if (cp->cache_defrag != NULL)
3454 3454 (void) taskq_dispatch(kmem_taskq,
3455 3455 (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3456 3456 }
3457 3457
3458 3458 static void kmem_update(void *);
3459 3459
3460 3460 static void
3461 3461 kmem_update_timeout(void *dummy)
3462 3462 {
3463 3463 (void) timeout(kmem_update, dummy, kmem_reap_interval);
3464 3464 }
3465 3465
3466 3466 static void
3467 3467 kmem_update(void *dummy)
3468 3468 {
3469 3469 kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3470 3470
3471 3471 /*
3472 3472 * We use taskq_dispatch() to reschedule the timeout so that
3473 3473 * kmem_update() becomes self-throttling: it won't schedule
3474 3474 * new tasks until all previous tasks have completed.
3475 3475 */
3476 3476 if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3477 3477 kmem_update_timeout(NULL);
3478 3478 }
3479 3479
3480 3480 static int
3481 3481 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3482 3482 {
3483 3483 struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3484 3484 kmem_cache_t *cp = ksp->ks_private;
3485 3485 uint64_t cpu_buf_avail;
3486 3486 uint64_t buf_avail = 0;
3487 3487 int cpu_seqid;
3488 3488 long reap;
3489 3489
3490 3490 ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3491 3491
3492 3492 if (rw == KSTAT_WRITE)
3493 3493 return (EACCES);
3494 3494
3495 3495 mutex_enter(&cp->cache_lock);
3496 3496
3497 3497 kmcp->kmc_alloc_fail.value.ui64 = cp->cache_alloc_fail;
3498 3498 kmcp->kmc_alloc.value.ui64 = cp->cache_slab_alloc;
3499 3499 kmcp->kmc_free.value.ui64 = cp->cache_slab_free;
3500 3500 kmcp->kmc_slab_alloc.value.ui64 = cp->cache_slab_alloc;
3501 3501 kmcp->kmc_slab_free.value.ui64 = cp->cache_slab_free;
3502 3502
3503 3503 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3504 3504 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3505 3505
3506 3506 mutex_enter(&ccp->cc_lock);
3507 3507
3508 3508 cpu_buf_avail = 0;
3509 3509 if (ccp->cc_rounds > 0)
3510 3510 cpu_buf_avail += ccp->cc_rounds;
3511 3511 if (ccp->cc_prounds > 0)
3512 3512 cpu_buf_avail += ccp->cc_prounds;
3513 3513
3514 3514 kmcp->kmc_alloc.value.ui64 += ccp->cc_alloc;
3515 3515 kmcp->kmc_free.value.ui64 += ccp->cc_free;
3516 3516 buf_avail += cpu_buf_avail;
3517 3517
3518 3518 mutex_exit(&ccp->cc_lock);
3519 3519 }
3520 3520
3521 3521 mutex_enter(&cp->cache_depot_lock);
3522 3522
3523 3523 kmcp->kmc_depot_alloc.value.ui64 = cp->cache_full.ml_alloc;
3524 3524 kmcp->kmc_depot_free.value.ui64 = cp->cache_empty.ml_alloc;
3525 3525 kmcp->kmc_depot_contention.value.ui64 = cp->cache_depot_contention;
3526 3526 kmcp->kmc_full_magazines.value.ui64 = cp->cache_full.ml_total;
3527 3527 kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3528 3528 kmcp->kmc_magazine_size.value.ui64 =
3529 3529 (cp->cache_flags & KMF_NOMAGAZINE) ?
3530 3530 0 : cp->cache_magtype->mt_magsize;
3531 3531
3532 3532 kmcp->kmc_alloc.value.ui64 += cp->cache_full.ml_alloc;
3533 3533 kmcp->kmc_free.value.ui64 += cp->cache_empty.ml_alloc;
3534 3534 buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3535 3535
3536 3536 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3537 3537 reap = MIN(reap, cp->cache_full.ml_total);
3538 3538
3539 3539 mutex_exit(&cp->cache_depot_lock);
3540 3540
3541 3541 kmcp->kmc_buf_size.value.ui64 = cp->cache_bufsize;
3542 3542 kmcp->kmc_align.value.ui64 = cp->cache_align;
3543 3543 kmcp->kmc_chunk_size.value.ui64 = cp->cache_chunksize;
3544 3544 kmcp->kmc_slab_size.value.ui64 = cp->cache_slabsize;
3545 3545 kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3546 3546 buf_avail += cp->cache_bufslab;
3547 3547 kmcp->kmc_buf_avail.value.ui64 = buf_avail;
3548 3548 kmcp->kmc_buf_inuse.value.ui64 = cp->cache_buftotal - buf_avail;
3549 3549 kmcp->kmc_buf_total.value.ui64 = cp->cache_buftotal;
3550 3550 kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3551 3551 kmcp->kmc_slab_create.value.ui64 = cp->cache_slab_create;
3552 3552 kmcp->kmc_slab_destroy.value.ui64 = cp->cache_slab_destroy;
3553 3553 kmcp->kmc_hash_size.value.ui64 = (cp->cache_flags & KMF_HASH) ?
3554 3554 cp->cache_hash_mask + 1 : 0;
3555 3555 kmcp->kmc_hash_lookup_depth.value.ui64 = cp->cache_lookup_depth;
3556 3556 kmcp->kmc_hash_rescale.value.ui64 = cp->cache_rescale;
3557 3557 kmcp->kmc_vmem_source.value.ui64 = cp->cache_arena->vm_id;
3558 3558 kmcp->kmc_reap.value.ui64 = cp->cache_reap;
3559 3559
3560 3560 if (cp->cache_defrag == NULL) {
3561 3561 kmcp->kmc_move_callbacks.value.ui64 = 0;
3562 3562 kmcp->kmc_move_yes.value.ui64 = 0;
3563 3563 kmcp->kmc_move_no.value.ui64 = 0;
3564 3564 kmcp->kmc_move_later.value.ui64 = 0;
3565 3565 kmcp->kmc_move_dont_need.value.ui64 = 0;
3566 3566 kmcp->kmc_move_dont_know.value.ui64 = 0;
3567 3567 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3568 3568 kmcp->kmc_move_slabs_freed.value.ui64 = 0;
3569 3569 kmcp->kmc_defrag.value.ui64 = 0;
3570 3570 kmcp->kmc_scan.value.ui64 = 0;
3571 3571 kmcp->kmc_move_reclaimable.value.ui64 = 0;
3572 3572 } else {
3573 3573 int64_t reclaimable;
3574 3574
3575 3575 kmem_defrag_t *kd = cp->cache_defrag;
3576 3576 kmcp->kmc_move_callbacks.value.ui64 = kd->kmd_callbacks;
3577 3577 kmcp->kmc_move_yes.value.ui64 = kd->kmd_yes;
3578 3578 kmcp->kmc_move_no.value.ui64 = kd->kmd_no;
3579 3579 kmcp->kmc_move_later.value.ui64 = kd->kmd_later;
3580 3580 kmcp->kmc_move_dont_need.value.ui64 = kd->kmd_dont_need;
3581 3581 kmcp->kmc_move_dont_know.value.ui64 = kd->kmd_dont_know;
3582 3582 kmcp->kmc_move_hunt_found.value.ui64 = kd->kmd_hunt_found;
3583 3583 kmcp->kmc_move_slabs_freed.value.ui64 = kd->kmd_slabs_freed;
3584 3584 kmcp->kmc_defrag.value.ui64 = kd->kmd_defrags;
3585 3585 kmcp->kmc_scan.value.ui64 = kd->kmd_scans;
3586 3586
3587 3587 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3588 3588 reclaimable = MAX(reclaimable, 0);
3589 3589 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3590 3590 kmcp->kmc_move_reclaimable.value.ui64 = reclaimable;
3591 3591 }
3592 3592
3593 3593 mutex_exit(&cp->cache_lock);
3594 3594 return (0);
3595 3595 }
3596 3596
3597 3597 /*
3598 3598 * Return a named statistic about a particular cache.
3599 3599 * This shouldn't be called very often, so it's currently designed for
3600 3600 * simplicity (leverages existing kstat support) rather than efficiency.
3601 3601 */
3602 3602 uint64_t
3603 3603 kmem_cache_stat(kmem_cache_t *cp, char *name)
3604 3604 {
3605 3605 int i;
3606 3606 kstat_t *ksp = cp->cache_kstat;
3607 3607 kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3608 3608 uint64_t value = 0;
3609 3609
3610 3610 if (ksp != NULL) {
3611 3611 mutex_enter(&kmem_cache_kstat_lock);
3612 3612 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3613 3613 for (i = 0; i < ksp->ks_ndata; i++) {
3614 3614 if (strcmp(knp[i].name, name) == 0) {
3615 3615 value = knp[i].value.ui64;
3616 3616 break;
3617 3617 }
3618 3618 }
3619 3619 mutex_exit(&kmem_cache_kstat_lock);
3620 3620 }
3621 3621 return (value);
3622 3622 }
3623 3623
3624 3624 /*
3625 3625 * Return an estimate of currently available kernel heap memory.
3626 3626 * On 32-bit systems, physical memory may exceed virtual memory,
3627 3627 * we just truncate the result at 1GB.
3628 3628 */
3629 3629 size_t
3630 3630 kmem_avail(void)
3631 3631 {
3632 3632 spgcnt_t rmem = availrmem - tune.t_minarmem;
3633 3633 spgcnt_t fmem = freemem - minfree;
3634 3634
3635 3635 return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3636 3636 1 << (30 - PAGESHIFT))));
3637 3637 }
3638 3638
3639 3639 /*
3640 3640 * Return the maximum amount of memory that is (in theory) allocatable
3641 3641 * from the heap. This may be used as an estimate only since there
3642 3642 * is no guarentee this space will still be available when an allocation
3643 3643 * request is made, nor that the space may be allocated in one big request
3644 3644 * due to kernel heap fragmentation.
3645 3645 */
3646 3646 size_t
3647 3647 kmem_maxavail(void)
3648 3648 {
3649 3649 spgcnt_t pmem = availrmem - tune.t_minarmem;
3650 3650 spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3651 3651
3652 3652 return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3653 3653 }
3654 3654
3655 3655 /*
3656 3656 * Indicate whether memory-intensive kmem debugging is enabled.
3657 3657 */
3658 3658 int
3659 3659 kmem_debugging(void)
3660 3660 {
3661 3661 return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3662 3662 }
3663 3663
3664 3664 /* binning function, sorts finely at the two extremes */
3665 3665 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift) \
3666 3666 ((((sp)->slab_refcnt <= (binshift)) || \
3667 3667 (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift))) \
3668 3668 ? -(sp)->slab_refcnt \
3669 3669 : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3670 3670
3671 3671 /*
3672 3672 * Minimizing the number of partial slabs on the freelist minimizes
3673 3673 * fragmentation (the ratio of unused buffers held by the slab layer). There are
3674 3674 * two ways to get a slab off of the freelist: 1) free all the buffers on the
3675 3675 * slab, and 2) allocate all the buffers on the slab. It follows that we want
3676 3676 * the most-used slabs at the front of the list where they have the best chance
3677 3677 * of being completely allocated, and the least-used slabs at a safe distance
3678 3678 * from the front to improve the odds that the few remaining buffers will all be
3679 3679 * freed before another allocation can tie up the slab. For that reason a slab
3680 3680 * with a higher slab_refcnt sorts less than than a slab with a lower
3681 3681 * slab_refcnt.
3682 3682 *
3683 3683 * However, if a slab has at least one buffer that is deemed unfreeable, we
3684 3684 * would rather have that slab at the front of the list regardless of
3685 3685 * slab_refcnt, since even one unfreeable buffer makes the entire slab
3686 3686 * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3687 3687 * callback, the slab is marked unfreeable for as long as it remains on the
3688 3688 * freelist.
3689 3689 */
3690 3690 static int
3691 3691 kmem_partial_slab_cmp(const void *p0, const void *p1)
3692 3692 {
3693 3693 const kmem_cache_t *cp;
3694 3694 const kmem_slab_t *s0 = p0;
3695 3695 const kmem_slab_t *s1 = p1;
3696 3696 int w0, w1;
3697 3697 size_t binshift;
3698 3698
3699 3699 ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3700 3700 ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3701 3701 ASSERT(s0->slab_cache == s1->slab_cache);
3702 3702 cp = s1->slab_cache;
3703 3703 ASSERT(MUTEX_HELD(&cp->cache_lock));
3704 3704 binshift = cp->cache_partial_binshift;
3705 3705
3706 3706 /* weight of first slab */
3707 3707 w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3708 3708 if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3709 3709 w0 -= cp->cache_maxchunks;
3710 3710 }
3711 3711
3712 3712 /* weight of second slab */
3713 3713 w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3714 3714 if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3715 3715 w1 -= cp->cache_maxchunks;
3716 3716 }
3717 3717
3718 3718 if (w0 < w1)
3719 3719 return (-1);
3720 3720 if (w0 > w1)
3721 3721 return (1);
3722 3722
3723 3723 /* compare pointer values */
3724 3724 if ((uintptr_t)s0 < (uintptr_t)s1)
3725 3725 return (-1);
3726 3726 if ((uintptr_t)s0 > (uintptr_t)s1)
3727 3727 return (1);
3728 3728
3729 3729 return (0);
3730 3730 }
3731 3731
3732 3732 /*
3733 3733 * It must be valid to call the destructor (if any) on a newly created object.
3734 3734 * That is, the constructor (if any) must leave the object in a valid state for
3735 3735 * the destructor.
3736 3736 */
3737 3737 kmem_cache_t *
3738 3738 kmem_cache_create(
3739 3739 char *name, /* descriptive name for this cache */
3740 3740 size_t bufsize, /* size of the objects it manages */
3741 3741 size_t align, /* required object alignment */
3742 3742 int (*constructor)(void *, void *, int), /* object constructor */
3743 3743 void (*destructor)(void *, void *), /* object destructor */
3744 3744 void (*reclaim)(void *), /* memory reclaim callback */
3745 3745 void *private, /* pass-thru arg for constr/destr/reclaim */
3746 3746 vmem_t *vmp, /* vmem source for slab allocation */
3747 3747 int cflags) /* cache creation flags */
3748 3748 {
3749 3749 int cpu_seqid;
3750 3750 size_t chunksize;
3751 3751 kmem_cache_t *cp;
3752 3752 kmem_magtype_t *mtp;
3753 3753 size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3754 3754
3755 3755 #ifdef DEBUG
3756 3756 /*
3757 3757 * Cache names should conform to the rules for valid C identifiers
3758 3758 */
3759 3759 if (!strident_valid(name)) {
3760 3760 cmn_err(CE_CONT,
3761 3761 "kmem_cache_create: '%s' is an invalid cache name\n"
3762 3762 "cache names must conform to the rules for "
3763 3763 "C identifiers\n", name);
3764 3764 }
3765 3765 #endif /* DEBUG */
3766 3766
3767 3767 if (vmp == NULL)
3768 3768 vmp = kmem_default_arena;
3769 3769
3770 3770 /*
3771 3771 * If this kmem cache has an identifier vmem arena as its source, mark
3772 3772 * it such to allow kmem_reap_idspace().
3773 3773 */
3774 3774 ASSERT(!(cflags & KMC_IDENTIFIER)); /* consumer should not set this */
3775 3775 if (vmp->vm_cflags & VMC_IDENTIFIER)
3776 3776 cflags |= KMC_IDENTIFIER;
3777 3777
3778 3778 /*
3779 3779 * Get a kmem_cache structure. We arrange that cp->cache_cpu[]
3780 3780 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3781 3781 * false sharing of per-CPU data.
3782 3782 */
3783 3783 cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3784 3784 P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3785 3785 bzero(cp, csize);
3786 3786 list_link_init(&cp->cache_link);
3787 3787
3788 3788 if (align == 0)
3789 3789 align = KMEM_ALIGN;
3790 3790
3791 3791 /*
3792 3792 * If we're not at least KMEM_ALIGN aligned, we can't use free
3793 3793 * memory to hold bufctl information (because we can't safely
3794 3794 * perform word loads and stores on it).
3795 3795 */
3796 3796 if (align < KMEM_ALIGN)
3797 3797 cflags |= KMC_NOTOUCH;
3798 3798
3799 3799 if (!ISP2(align) || align > vmp->vm_quantum)
3800 3800 panic("kmem_cache_create: bad alignment %lu", align);
3801 3801
3802 3802 mutex_enter(&kmem_flags_lock);
3803 3803 if (kmem_flags & KMF_RANDOMIZE)
3804 3804 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3805 3805 KMF_RANDOMIZE;
3806 3806 cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3807 3807 mutex_exit(&kmem_flags_lock);
3808 3808
3809 3809 /*
3810 3810 * Make sure all the various flags are reasonable.
3811 3811 */
3812 3812 ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3813 3813
3814 3814 if (cp->cache_flags & KMF_LITE) {
3815 3815 if (bufsize >= kmem_lite_minsize &&
3816 3816 align <= kmem_lite_maxalign &&
3817 3817 P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3818 3818 cp->cache_flags |= KMF_BUFTAG;
3819 3819 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3820 3820 } else {
3821 3821 cp->cache_flags &= ~KMF_DEBUG;
3822 3822 }
3823 3823 }
3824 3824
3825 3825 if (cp->cache_flags & KMF_DEADBEEF)
3826 3826 cp->cache_flags |= KMF_REDZONE;
3827 3827
3828 3828 if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3829 3829 cp->cache_flags |= KMF_NOMAGAZINE;
3830 3830
3831 3831 if (cflags & KMC_NODEBUG)
3832 3832 cp->cache_flags &= ~KMF_DEBUG;
3833 3833
3834 3834 if (cflags & KMC_NOTOUCH)
3835 3835 cp->cache_flags &= ~KMF_TOUCH;
3836 3836
3837 3837 if (cflags & KMC_PREFILL)
3838 3838 cp->cache_flags |= KMF_PREFILL;
3839 3839
3840 3840 if (cflags & KMC_NOHASH)
3841 3841 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3842 3842
3843 3843 if (cflags & KMC_NOMAGAZINE)
3844 3844 cp->cache_flags |= KMF_NOMAGAZINE;
3845 3845
3846 3846 if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3847 3847 cp->cache_flags |= KMF_REDZONE;
3848 3848
3849 3849 if (!(cp->cache_flags & KMF_AUDIT))
3850 3850 cp->cache_flags &= ~KMF_CONTENTS;
3851 3851
3852 3852 if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3853 3853 !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3854 3854 cp->cache_flags |= KMF_FIREWALL;
3855 3855
3856 3856 if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3857 3857 cp->cache_flags &= ~KMF_FIREWALL;
3858 3858
3859 3859 if (cp->cache_flags & KMF_FIREWALL) {
3860 3860 cp->cache_flags &= ~KMF_BUFTAG;
3861 3861 cp->cache_flags |= KMF_NOMAGAZINE;
3862 3862 ASSERT(vmp == kmem_default_arena);
3863 3863 vmp = kmem_firewall_arena;
3864 3864 }
3865 3865
3866 3866 /*
3867 3867 * Set cache properties.
3868 3868 */
3869 3869 (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3870 3870 strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3871 3871 cp->cache_bufsize = bufsize;
3872 3872 cp->cache_align = align;
3873 3873 cp->cache_constructor = constructor;
3874 3874 cp->cache_destructor = destructor;
3875 3875 cp->cache_reclaim = reclaim;
3876 3876 cp->cache_private = private;
3877 3877 cp->cache_arena = vmp;
3878 3878 cp->cache_cflags = cflags;
3879 3879
3880 3880 /*
3881 3881 * Determine the chunk size.
3882 3882 */
3883 3883 chunksize = bufsize;
3884 3884
3885 3885 if (align >= KMEM_ALIGN) {
3886 3886 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3887 3887 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3888 3888 }
3889 3889
3890 3890 if (cp->cache_flags & KMF_BUFTAG) {
3891 3891 cp->cache_bufctl = chunksize;
3892 3892 cp->cache_buftag = chunksize;
3893 3893 if (cp->cache_flags & KMF_LITE)
3894 3894 chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3895 3895 else
3896 3896 chunksize += sizeof (kmem_buftag_t);
3897 3897 }
3898 3898
3899 3899 if (cp->cache_flags & KMF_DEADBEEF) {
3900 3900 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3901 3901 if (cp->cache_flags & KMF_LITE)
3902 3902 cp->cache_verify = sizeof (uint64_t);
3903 3903 }
3904 3904
3905 3905 cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3906 3906
3907 3907 cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3908 3908
3909 3909 /*
3910 3910 * Now that we know the chunk size, determine the optimal slab size.
3911 3911 */
3912 3912 if (vmp == kmem_firewall_arena) {
3913 3913 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3914 3914 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3915 3915 cp->cache_maxcolor = cp->cache_mincolor;
3916 3916 cp->cache_flags |= KMF_HASH;
3917 3917 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3918 3918 } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3919 3919 !(cp->cache_flags & KMF_AUDIT) &&
3920 3920 chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3921 3921 cp->cache_slabsize = vmp->vm_quantum;
3922 3922 cp->cache_mincolor = 0;
3923 3923 cp->cache_maxcolor =
3924 3924 (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3925 3925 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3926 3926 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3927 3927 } else {
3928 3928 size_t chunks, bestfit, waste, slabsize;
3929 3929 size_t minwaste = LONG_MAX;
3930 3930
3931 3931 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3932 3932 slabsize = P2ROUNDUP(chunksize * chunks,
3933 3933 vmp->vm_quantum);
3934 3934 chunks = slabsize / chunksize;
3935 3935 waste = (slabsize % chunksize) / chunks;
3936 3936 if (waste < minwaste) {
3937 3937 minwaste = waste;
3938 3938 bestfit = slabsize;
3939 3939 }
3940 3940 }
3941 3941 if (cflags & KMC_QCACHE)
3942 3942 bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3943 3943 cp->cache_slabsize = bestfit;
3944 3944 cp->cache_mincolor = 0;
3945 3945 cp->cache_maxcolor = bestfit % chunksize;
3946 3946 cp->cache_flags |= KMF_HASH;
3947 3947 }
3948 3948
3949 3949 cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3950 3950 cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3951 3951
3952 3952 /*
3953 3953 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3954 3954 * there is a constructor avoids some tricky issues with debug setup
3955 3955 * that may be revisited later. We cannot allow prefill in a
3956 3956 * metadata cache because of potential recursion.
3957 3957 */
3958 3958 if (vmp == kmem_msb_arena ||
3959 3959 cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3960 3960 cp->cache_constructor != NULL)
3961 3961 cp->cache_flags &= ~KMF_PREFILL;
3962 3962
3963 3963 if (cp->cache_flags & KMF_HASH) {
3964 3964 ASSERT(!(cflags & KMC_NOHASH));
3965 3965 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3966 3966 kmem_bufctl_audit_cache : kmem_bufctl_cache;
3967 3967 }
3968 3968
3969 3969 if (cp->cache_maxcolor >= vmp->vm_quantum)
3970 3970 cp->cache_maxcolor = vmp->vm_quantum - 1;
3971 3971
3972 3972 cp->cache_color = cp->cache_mincolor;
3973 3973
3974 3974 /*
3975 3975 * Initialize the rest of the slab layer.
3976 3976 */
3977 3977 mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3978 3978
3979 3979 avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3980 3980 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3981 3981 /* LINTED: E_TRUE_LOGICAL_EXPR */
3982 3982 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3983 3983 /* reuse partial slab AVL linkage for complete slab list linkage */
3984 3984 list_create(&cp->cache_complete_slabs,
3985 3985 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3986 3986
3987 3987 if (cp->cache_flags & KMF_HASH) {
3988 3988 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3989 3989 KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3990 3990 bzero(cp->cache_hash_table,
3991 3991 KMEM_HASH_INITIAL * sizeof (void *));
3992 3992 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3993 3993 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3994 3994 }
3995 3995
3996 3996 /*
3997 3997 * Initialize the depot.
3998 3998 */
3999 3999 mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
4000 4000
4001 4001 for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
4002 4002 continue;
4003 4003
4004 4004 cp->cache_magtype = mtp;
4005 4005
4006 4006 /*
4007 4007 * Initialize the CPU layer.
4008 4008 */
4009 4009 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4010 4010 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
4011 4011 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
4012 4012 ccp->cc_flags = cp->cache_flags;
4013 4013 ccp->cc_rounds = -1;
4014 4014 ccp->cc_prounds = -1;
4015 4015 }
4016 4016
4017 4017 /*
4018 4018 * Create the cache's kstats.
4019 4019 */
4020 4020 if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
4021 4021 "kmem_cache", KSTAT_TYPE_NAMED,
4022 4022 sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
4023 4023 KSTAT_FLAG_VIRTUAL)) != NULL) {
4024 4024 cp->cache_kstat->ks_data = &kmem_cache_kstat;
4025 4025 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
4026 4026 cp->cache_kstat->ks_private = cp;
4027 4027 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
4028 4028 kstat_install(cp->cache_kstat);
4029 4029 }
4030 4030
4031 4031 /*
4032 4032 * Add the cache to the global list. This makes it visible
4033 4033 * to kmem_update(), so the cache must be ready for business.
4034 4034 */
4035 4035 mutex_enter(&kmem_cache_lock);
4036 4036 list_insert_tail(&kmem_caches, cp);
4037 4037 mutex_exit(&kmem_cache_lock);
4038 4038
4039 4039 if (kmem_ready)
4040 4040 kmem_cache_magazine_enable(cp);
4041 4041
4042 4042 return (cp);
4043 4043 }
4044 4044
4045 4045 static int
4046 4046 kmem_move_cmp(const void *buf, const void *p)
4047 4047 {
4048 4048 const kmem_move_t *kmm = p;
4049 4049 uintptr_t v1 = (uintptr_t)buf;
4050 4050 uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
4051 4051 return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
4052 4052 }
4053 4053
4054 4054 static void
4055 4055 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4056 4056 {
4057 4057 kmd->kmd_reclaim_numer = 1;
4058 4058 }
4059 4059
4060 4060 /*
4061 4061 * Initially, when choosing candidate slabs for buffers to move, we want to be
4062 4062 * very selective and take only slabs that are less than
4063 4063 * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4064 4064 * slabs, then we raise the allocation ceiling incrementally. The reclaim
4065 4065 * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4066 4066 * longer fragmented.
4067 4067 */
4068 4068 static void
4069 4069 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4070 4070 {
4071 4071 if (direction > 0) {
4072 4072 /* make it easier to find a candidate slab */
4073 4073 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4074 4074 kmd->kmd_reclaim_numer++;
4075 4075 }
4076 4076 } else {
4077 4077 /* be more selective */
4078 4078 if (kmd->kmd_reclaim_numer > 1) {
4079 4079 kmd->kmd_reclaim_numer--;
4080 4080 }
4081 4081 }
4082 4082 }
4083 4083
4084 4084 void
4085 4085 kmem_cache_set_move(kmem_cache_t *cp,
4086 4086 kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4087 4087 {
4088 4088 kmem_defrag_t *defrag;
4089 4089
4090 4090 ASSERT(move != NULL);
4091 4091 /*
4092 4092 * The consolidator does not support NOTOUCH caches because kmem cannot
4093 4093 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4094 4094 * a low order bit usable by clients to distinguish uninitialized memory
4095 4095 * from known objects (see kmem_slab_create).
4096 4096 */
4097 4097 ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4098 4098 ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4099 4099
4100 4100 /*
4101 4101 * We should not be holding anyone's cache lock when calling
4102 4102 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4103 4103 * lock.
4104 4104 */
4105 4105 defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4106 4106
4107 4107 mutex_enter(&cp->cache_lock);
4108 4108
4109 4109 if (KMEM_IS_MOVABLE(cp)) {
4110 4110 if (cp->cache_move == NULL) {
4111 4111 ASSERT(cp->cache_slab_alloc == 0);
4112 4112
4113 4113 cp->cache_defrag = defrag;
4114 4114 defrag = NULL; /* nothing to free */
4115 4115 bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4116 4116 avl_create(&cp->cache_defrag->kmd_moves_pending,
4117 4117 kmem_move_cmp, sizeof (kmem_move_t),
4118 4118 offsetof(kmem_move_t, kmm_entry));
4119 4119 /* LINTED: E_TRUE_LOGICAL_EXPR */
4120 4120 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4121 4121 /* reuse the slab's AVL linkage for deadlist linkage */
4122 4122 list_create(&cp->cache_defrag->kmd_deadlist,
4123 4123 sizeof (kmem_slab_t),
4124 4124 offsetof(kmem_slab_t, slab_link));
4125 4125 kmem_reset_reclaim_threshold(cp->cache_defrag);
4126 4126 }
4127 4127 cp->cache_move = move;
4128 4128 }
4129 4129
4130 4130 mutex_exit(&cp->cache_lock);
4131 4131
4132 4132 if (defrag != NULL) {
4133 4133 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4134 4134 }
4135 4135 }
4136 4136
4137 4137 void
4138 4138 kmem_cache_destroy(kmem_cache_t *cp)
4139 4139 {
4140 4140 int cpu_seqid;
4141 4141
4142 4142 /*
|
↓ open down ↓ |
4142 lines elided |
↑ open up ↑ |
4143 4143 * Remove the cache from the global cache list so that no one else
4144 4144 * can schedule tasks on its behalf, wait for any pending tasks to
4145 4145 * complete, purge the cache, and then destroy it.
4146 4146 */
4147 4147 mutex_enter(&kmem_cache_lock);
4148 4148 list_remove(&kmem_caches, cp);
4149 4149 mutex_exit(&kmem_cache_lock);
4150 4150
4151 4151 if (kmem_taskq != NULL)
4152 4152 taskq_wait(kmem_taskq);
4153 - if (kmem_move_taskq != NULL)
4153 +
4154 + if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4154 4155 taskq_wait(kmem_move_taskq);
4155 4156
4156 4157 kmem_cache_magazine_purge(cp);
4157 4158
4158 4159 mutex_enter(&cp->cache_lock);
4159 4160 if (cp->cache_buftotal != 0)
4160 4161 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4161 4162 cp->cache_name, (void *)cp);
4162 4163 if (cp->cache_defrag != NULL) {
4163 4164 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4164 4165 list_destroy(&cp->cache_defrag->kmd_deadlist);
4165 4166 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4166 4167 cp->cache_defrag = NULL;
4167 4168 }
4168 4169 /*
4169 4170 * The cache is now dead. There should be no further activity. We
4170 4171 * enforce this by setting land mines in the constructor, destructor,
4171 4172 * reclaim, and move routines that induce a kernel text fault if
4172 4173 * invoked.
4173 4174 */
4174 4175 cp->cache_constructor = (int (*)(void *, void *, int))1;
4175 4176 cp->cache_destructor = (void (*)(void *, void *))2;
4176 4177 cp->cache_reclaim = (void (*)(void *))3;
4177 4178 cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4178 4179 mutex_exit(&cp->cache_lock);
4179 4180
4180 4181 kstat_delete(cp->cache_kstat);
4181 4182
4182 4183 if (cp->cache_hash_table != NULL)
4183 4184 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4184 4185 (cp->cache_hash_mask + 1) * sizeof (void *));
4185 4186
4186 4187 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4187 4188 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4188 4189
4189 4190 mutex_destroy(&cp->cache_depot_lock);
4190 4191 mutex_destroy(&cp->cache_lock);
4191 4192
4192 4193 vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4193 4194 }
4194 4195
4195 4196 /*ARGSUSED*/
4196 4197 static int
4197 4198 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4198 4199 {
4199 4200 ASSERT(MUTEX_HELD(&cpu_lock));
4200 4201 if (what == CPU_UNCONFIG) {
4201 4202 kmem_cache_applyall(kmem_cache_magazine_purge,
4202 4203 kmem_taskq, TQ_SLEEP);
4203 4204 kmem_cache_applyall(kmem_cache_magazine_enable,
4204 4205 kmem_taskq, TQ_SLEEP);
4205 4206 }
4206 4207 return (0);
4207 4208 }
4208 4209
4209 4210 static void
4210 4211 kmem_alloc_caches_create(const int *array, size_t count,
4211 4212 kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4212 4213 {
4213 4214 char name[KMEM_CACHE_NAMELEN + 1];
4214 4215 size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4215 4216 size_t size = table_unit;
4216 4217 int i;
4217 4218
4218 4219 for (i = 0; i < count; i++) {
4219 4220 size_t cache_size = array[i];
4220 4221 size_t align = KMEM_ALIGN;
4221 4222 kmem_cache_t *cp;
4222 4223
4223 4224 /* if the table has an entry for maxbuf, we're done */
4224 4225 if (size > maxbuf)
4225 4226 break;
4226 4227
4227 4228 /* cache size must be a multiple of the table unit */
4228 4229 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4229 4230
4230 4231 /*
4231 4232 * If they allocate a multiple of the coherency granularity,
4232 4233 * they get a coherency-granularity-aligned address.
4233 4234 */
4234 4235 if (IS_P2ALIGNED(cache_size, 64))
4235 4236 align = 64;
4236 4237 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4237 4238 align = PAGESIZE;
4238 4239 (void) snprintf(name, sizeof (name),
4239 4240 "kmem_alloc_%lu", cache_size);
4240 4241 cp = kmem_cache_create(name, cache_size, align,
4241 4242 NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4242 4243
4243 4244 while (size <= cache_size) {
4244 4245 alloc_table[(size - 1) >> shift] = cp;
4245 4246 size += table_unit;
4246 4247 }
4247 4248 }
4248 4249
4249 4250 ASSERT(size > maxbuf); /* i.e. maxbuf <= max(cache_size) */
4250 4251 }
4251 4252
4252 4253 static void
4253 4254 kmem_cache_init(int pass, int use_large_pages)
4254 4255 {
4255 4256 int i;
4256 4257 size_t maxbuf;
4257 4258 kmem_magtype_t *mtp;
4258 4259
4259 4260 for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4260 4261 char name[KMEM_CACHE_NAMELEN + 1];
4261 4262
4262 4263 mtp = &kmem_magtype[i];
4263 4264 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4264 4265 mtp->mt_cache = kmem_cache_create(name,
4265 4266 (mtp->mt_magsize + 1) * sizeof (void *),
4266 4267 mtp->mt_align, NULL, NULL, NULL, NULL,
4267 4268 kmem_msb_arena, KMC_NOHASH);
4268 4269 }
4269 4270
4270 4271 kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4271 4272 sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4272 4273 kmem_msb_arena, KMC_NOHASH);
4273 4274
4274 4275 kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4275 4276 sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4276 4277 kmem_msb_arena, KMC_NOHASH);
4277 4278
4278 4279 kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4279 4280 sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4280 4281 kmem_msb_arena, KMC_NOHASH);
4281 4282
4282 4283 if (pass == 2) {
4283 4284 kmem_va_arena = vmem_create("kmem_va",
4284 4285 NULL, 0, PAGESIZE,
4285 4286 vmem_alloc, vmem_free, heap_arena,
4286 4287 8 * PAGESIZE, VM_SLEEP);
4287 4288
4288 4289 if (use_large_pages) {
4289 4290 kmem_default_arena = vmem_xcreate("kmem_default",
4290 4291 NULL, 0, PAGESIZE,
4291 4292 segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4292 4293 0, VMC_DUMPSAFE | VM_SLEEP);
4293 4294 } else {
4294 4295 kmem_default_arena = vmem_create("kmem_default",
4295 4296 NULL, 0, PAGESIZE,
4296 4297 segkmem_alloc, segkmem_free, kmem_va_arena,
4297 4298 0, VMC_DUMPSAFE | VM_SLEEP);
4298 4299 }
4299 4300
4300 4301 /* Figure out what our maximum cache size is */
4301 4302 maxbuf = kmem_max_cached;
4302 4303 if (maxbuf <= KMEM_MAXBUF) {
4303 4304 maxbuf = 0;
4304 4305 kmem_max_cached = KMEM_MAXBUF;
4305 4306 } else {
4306 4307 size_t size = 0;
4307 4308 size_t max =
4308 4309 sizeof (kmem_big_alloc_sizes) / sizeof (int);
4309 4310 /*
4310 4311 * Round maxbuf up to an existing cache size. If maxbuf
4311 4312 * is larger than the largest cache, we truncate it to
4312 4313 * the largest cache's size.
4313 4314 */
4314 4315 for (i = 0; i < max; i++) {
4315 4316 size = kmem_big_alloc_sizes[i];
4316 4317 if (maxbuf <= size)
4317 4318 break;
4318 4319 }
4319 4320 kmem_max_cached = maxbuf = size;
4320 4321 }
4321 4322
4322 4323 /*
4323 4324 * The big alloc table may not be completely overwritten, so
4324 4325 * we clear out any stale cache pointers from the first pass.
4325 4326 */
4326 4327 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4327 4328 } else {
4328 4329 /*
4329 4330 * During the first pass, the kmem_alloc_* caches
4330 4331 * are treated as metadata.
4331 4332 */
4332 4333 kmem_default_arena = kmem_msb_arena;
4333 4334 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4334 4335 }
4335 4336
4336 4337 /*
4337 4338 * Set up the default caches to back kmem_alloc()
4338 4339 */
4339 4340 kmem_alloc_caches_create(
4340 4341 kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4341 4342 kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4342 4343
4343 4344 kmem_alloc_caches_create(
4344 4345 kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4345 4346 kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4346 4347
4347 4348 kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4348 4349 }
4349 4350
4350 4351 void
4351 4352 kmem_init(void)
4352 4353 {
4353 4354 kmem_cache_t *cp;
4354 4355 int old_kmem_flags = kmem_flags;
4355 4356 int use_large_pages = 0;
4356 4357 size_t maxverify, minfirewall;
4357 4358
4358 4359 kstat_init();
4359 4360
4360 4361 /*
4361 4362 * Don't do firewalled allocations if the heap is less than 1TB
4362 4363 * (i.e. on a 32-bit kernel)
4363 4364 * The resulting VM_NEXTFIT allocations would create too much
4364 4365 * fragmentation in a small heap.
4365 4366 */
4366 4367 #if defined(_LP64)
4367 4368 maxverify = minfirewall = PAGESIZE / 2;
4368 4369 #else
4369 4370 maxverify = minfirewall = ULONG_MAX;
4370 4371 #endif
4371 4372
4372 4373 /* LINTED */
4373 4374 ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4374 4375
4375 4376 list_create(&kmem_caches, sizeof (kmem_cache_t),
4376 4377 offsetof(kmem_cache_t, cache_link));
4377 4378
4378 4379 kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4379 4380 vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4380 4381 VM_SLEEP | VMC_NO_QCACHE);
4381 4382
4382 4383 kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4383 4384 PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4384 4385 VMC_DUMPSAFE | VM_SLEEP);
4385 4386
4386 4387 kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4387 4388 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4388 4389
4389 4390 kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4390 4391 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4391 4392
4392 4393 kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4393 4394 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4394 4395
4395 4396 kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4396 4397 NULL, 0, PAGESIZE,
4397 4398 kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4398 4399 0, VM_SLEEP);
4399 4400
4400 4401 kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4401 4402 segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4402 4403 VMC_DUMPSAFE | VM_SLEEP);
4403 4404
4404 4405 /* temporary oversize arena for mod_read_system_file */
4405 4406 kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4406 4407 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4407 4408
4408 4409 kmem_reap_interval = 15 * hz;
4409 4410
4410 4411 /*
4411 4412 * Read /etc/system. This is a chicken-and-egg problem because
4412 4413 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4413 4414 * needs to use the allocator. The simplest solution is to create
4414 4415 * all the standard kmem caches, read /etc/system, destroy all the
4415 4416 * caches we just created, and then create them all again in light
4416 4417 * of the (possibly) new kmem_flags and other kmem tunables.
4417 4418 */
4418 4419 kmem_cache_init(1, 0);
4419 4420
4420 4421 mod_read_system_file(boothowto & RB_ASKNAME);
4421 4422
4422 4423 while ((cp = list_tail(&kmem_caches)) != NULL)
4423 4424 kmem_cache_destroy(cp);
4424 4425
4425 4426 vmem_destroy(kmem_oversize_arena);
4426 4427
4427 4428 if (old_kmem_flags & KMF_STICKY)
4428 4429 kmem_flags = old_kmem_flags;
4429 4430
4430 4431 if (!(kmem_flags & KMF_AUDIT))
4431 4432 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4432 4433
4433 4434 if (kmem_maxverify == 0)
4434 4435 kmem_maxverify = maxverify;
4435 4436
4436 4437 if (kmem_minfirewall == 0)
4437 4438 kmem_minfirewall = minfirewall;
4438 4439
4439 4440 /*
4440 4441 * give segkmem a chance to figure out if we are using large pages
4441 4442 * for the kernel heap
4442 4443 */
4443 4444 use_large_pages = segkmem_lpsetup();
4444 4445
4445 4446 /*
4446 4447 * To protect against corruption, we keep the actual number of callers
4447 4448 * KMF_LITE records seperate from the tunable. We arbitrarily clamp
4448 4449 * to 16, since the overhead for small buffers quickly gets out of
4449 4450 * hand.
4450 4451 *
4451 4452 * The real limit would depend on the needs of the largest KMC_NOHASH
4452 4453 * cache.
4453 4454 */
4454 4455 kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4455 4456 kmem_lite_pcs = kmem_lite_count;
4456 4457
4457 4458 /*
4458 4459 * Normally, we firewall oversized allocations when possible, but
4459 4460 * if we are using large pages for kernel memory, and we don't have
4460 4461 * any non-LITE debugging flags set, we want to allocate oversized
4461 4462 * buffers from large pages, and so skip the firewalling.
4462 4463 */
4463 4464 if (use_large_pages &&
4464 4465 ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4465 4466 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4466 4467 PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4467 4468 0, VMC_DUMPSAFE | VM_SLEEP);
4468 4469 } else {
4469 4470 kmem_oversize_arena = vmem_create("kmem_oversize",
4470 4471 NULL, 0, PAGESIZE,
4471 4472 segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4472 4473 kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4473 4474 VM_SLEEP);
4474 4475 }
4475 4476
4476 4477 kmem_cache_init(2, use_large_pages);
4477 4478
4478 4479 if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4479 4480 if (kmem_transaction_log_size == 0)
4480 4481 kmem_transaction_log_size = kmem_maxavail() / 50;
4481 4482 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4482 4483 }
4483 4484
4484 4485 if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4485 4486 if (kmem_content_log_size == 0)
4486 4487 kmem_content_log_size = kmem_maxavail() / 50;
4487 4488 kmem_content_log = kmem_log_init(kmem_content_log_size);
4488 4489 }
4489 4490
4490 4491 kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4491 4492
4492 4493 kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4493 4494
4494 4495 /*
4495 4496 * Initialize STREAMS message caches so allocb() is available.
4496 4497 * This allows us to initialize the logging framework (cmn_err(9F),
4497 4498 * strlog(9F), etc) so we can start recording messages.
4498 4499 */
4499 4500 streams_msg_init();
4500 4501
4501 4502 /*
4502 4503 * Initialize the ZSD framework in Zones so modules loaded henceforth
4503 4504 * can register their callbacks.
4504 4505 */
4505 4506 zone_zsd_init();
4506 4507
4507 4508 log_init();
4508 4509 taskq_init();
4509 4510
4510 4511 /*
4511 4512 * Warn about invalid or dangerous values of kmem_flags.
4512 4513 * Always warn about unsupported values.
4513 4514 */
4514 4515 if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4515 4516 KMF_CONTENTS | KMF_LITE)) != 0) ||
4516 4517 ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4517 4518 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4518 4519 "See the Solaris Tunable Parameters Reference Manual.",
4519 4520 kmem_flags);
4520 4521
4521 4522 #ifdef DEBUG
4522 4523 if ((kmem_flags & KMF_DEBUG) == 0)
4523 4524 cmn_err(CE_NOTE, "kmem debugging disabled.");
4524 4525 #else
4525 4526 /*
4526 4527 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4527 4528 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4528 4529 * if KMF_AUDIT is set). We should warn the user about the performance
4529 4530 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4530 4531 * isn't set (since that disables AUDIT).
4531 4532 */
4532 4533 if (!(kmem_flags & KMF_LITE) &&
4533 4534 (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4534 4535 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4535 4536 "enabled (kmem_flags = 0x%x). Performance degradation "
4536 4537 "and large memory overhead possible. See the Solaris "
4537 4538 "Tunable Parameters Reference Manual.", kmem_flags);
4538 4539 #endif /* not DEBUG */
4539 4540
4540 4541 kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4541 4542
4542 4543 kmem_ready = 1;
4543 4544
4544 4545 /*
4545 4546 * Initialize the platform-specific aligned/DMA memory allocator.
4546 4547 */
4547 4548 ka_init();
4548 4549
4549 4550 /*
4550 4551 * Initialize 32-bit ID cache.
4551 4552 */
4552 4553 id32_init();
4553 4554
4554 4555 /*
4555 4556 * Initialize the networking stack so modules loaded can
4556 4557 * register their callbacks.
4557 4558 */
4558 4559 netstack_init();
4559 4560 }
4560 4561
4561 4562 static void
4562 4563 kmem_move_init(void)
4563 4564 {
4564 4565 kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4565 4566 sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4566 4567 kmem_msb_arena, KMC_NOHASH);
4567 4568 kmem_move_cache = kmem_cache_create("kmem_move_cache",
4568 4569 sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4569 4570 kmem_msb_arena, KMC_NOHASH);
4570 4571
4571 4572 /*
4572 4573 * kmem guarantees that move callbacks are sequential and that even
4573 4574 * across multiple caches no two moves ever execute simultaneously.
4574 4575 * Move callbacks are processed on a separate taskq so that client code
4575 4576 * does not interfere with internal maintenance tasks.
4576 4577 */
4577 4578 kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4578 4579 minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4579 4580 }
4580 4581
4581 4582 void
4582 4583 kmem_thread_init(void)
4583 4584 {
4584 4585 kmem_move_init();
4585 4586 kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4586 4587 300, INT_MAX, TASKQ_PREPOPULATE);
4587 4588 }
4588 4589
4589 4590 void
4590 4591 kmem_mp_init(void)
4591 4592 {
4592 4593 mutex_enter(&cpu_lock);
4593 4594 register_cpu_setup_func(kmem_cpu_setup, NULL);
4594 4595 mutex_exit(&cpu_lock);
4595 4596
4596 4597 kmem_update_timeout(NULL);
4597 4598
4598 4599 taskq_mp_init();
4599 4600 }
4600 4601
4601 4602 /*
4602 4603 * Return the slab of the allocated buffer, or NULL if the buffer is not
4603 4604 * allocated. This function may be called with a known slab address to determine
4604 4605 * whether or not the buffer is allocated, or with a NULL slab address to obtain
4605 4606 * an allocated buffer's slab.
4606 4607 */
4607 4608 static kmem_slab_t *
4608 4609 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4609 4610 {
4610 4611 kmem_bufctl_t *bcp, *bufbcp;
4611 4612
4612 4613 ASSERT(MUTEX_HELD(&cp->cache_lock));
4613 4614 ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4614 4615
4615 4616 if (cp->cache_flags & KMF_HASH) {
4616 4617 for (bcp = *KMEM_HASH(cp, buf);
4617 4618 (bcp != NULL) && (bcp->bc_addr != buf);
4618 4619 bcp = bcp->bc_next) {
4619 4620 continue;
4620 4621 }
4621 4622 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4622 4623 return (bcp == NULL ? NULL : bcp->bc_slab);
4623 4624 }
4624 4625
4625 4626 if (sp == NULL) {
4626 4627 sp = KMEM_SLAB(cp, buf);
4627 4628 }
4628 4629 bufbcp = KMEM_BUFCTL(cp, buf);
4629 4630 for (bcp = sp->slab_head;
4630 4631 (bcp != NULL) && (bcp != bufbcp);
4631 4632 bcp = bcp->bc_next) {
4632 4633 continue;
4633 4634 }
4634 4635 return (bcp == NULL ? sp : NULL);
4635 4636 }
4636 4637
4637 4638 static boolean_t
4638 4639 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4639 4640 {
4640 4641 long refcnt = sp->slab_refcnt;
4641 4642
4642 4643 ASSERT(cp->cache_defrag != NULL);
4643 4644
4644 4645 /*
4645 4646 * For code coverage we want to be able to move an object within the
4646 4647 * same slab (the only partial slab) even if allocating the destination
4647 4648 * buffer resulted in a completely allocated slab.
4648 4649 */
4649 4650 if (flags & KMM_DEBUG) {
4650 4651 return ((flags & KMM_DESPERATE) ||
4651 4652 ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4652 4653 }
4653 4654
4654 4655 /* If we're desperate, we don't care if the client said NO. */
4655 4656 if (flags & KMM_DESPERATE) {
4656 4657 return (refcnt < sp->slab_chunks); /* any partial */
4657 4658 }
4658 4659
4659 4660 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4660 4661 return (B_FALSE);
4661 4662 }
4662 4663
4663 4664 if ((refcnt == 1) || kmem_move_any_partial) {
4664 4665 return (refcnt < sp->slab_chunks);
4665 4666 }
4666 4667
4667 4668 /*
4668 4669 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4669 4670 * slabs with a progressively higher percentage of used buffers can be
4670 4671 * reclaimed until the cache as a whole is no longer fragmented.
4671 4672 *
4672 4673 * sp->slab_refcnt kmd_reclaim_numer
4673 4674 * --------------- < ------------------
4674 4675 * sp->slab_chunks KMEM_VOID_FRACTION
4675 4676 */
4676 4677 return ((refcnt * KMEM_VOID_FRACTION) <
4677 4678 (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4678 4679 }
4679 4680
4680 4681 static void *
4681 4682 kmem_hunt_mag(kmem_cache_t *cp, kmem_magazine_t *m, int n, void *buf,
4682 4683 void *tbuf)
4683 4684 {
4684 4685 int i; /* magazine round index */
4685 4686
4686 4687 for (i = 0; i < n; i++) {
4687 4688 if (buf == m->mag_round[i]) {
4688 4689 if (cp->cache_flags & KMF_BUFTAG) {
4689 4690 (void) kmem_cache_free_debug(cp, tbuf,
4690 4691 caller());
4691 4692 }
4692 4693 m->mag_round[i] = tbuf;
4693 4694 return (buf);
4694 4695 }
4695 4696 }
4696 4697
4697 4698 return (NULL);
4698 4699 }
4699 4700
4700 4701 /*
4701 4702 * Hunt the magazine layer for the given buffer. If found, the buffer is
4702 4703 * removed from the magazine layer and returned, otherwise NULL is returned.
4703 4704 * The state of the returned buffer is freed and constructed.
4704 4705 */
4705 4706 static void *
4706 4707 kmem_hunt_mags(kmem_cache_t *cp, void *buf)
4707 4708 {
4708 4709 kmem_cpu_cache_t *ccp;
4709 4710 kmem_magazine_t *m;
4710 4711 int cpu_seqid;
4711 4712 int n; /* magazine rounds */
4712 4713 void *tbuf; /* temporary swap buffer */
4713 4714
4714 4715 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4715 4716
4716 4717 /*
4717 4718 * Allocated a buffer to swap with the one we hope to pull out of a
4718 4719 * magazine when found.
4719 4720 */
4720 4721 tbuf = kmem_cache_alloc(cp, KM_NOSLEEP);
4721 4722 if (tbuf == NULL) {
4722 4723 KMEM_STAT_ADD(kmem_move_stats.kms_hunt_alloc_fail);
4723 4724 return (NULL);
4724 4725 }
4725 4726 if (tbuf == buf) {
4726 4727 KMEM_STAT_ADD(kmem_move_stats.kms_hunt_lucky);
4727 4728 if (cp->cache_flags & KMF_BUFTAG) {
4728 4729 (void) kmem_cache_free_debug(cp, buf, caller());
4729 4730 }
4730 4731 return (buf);
4731 4732 }
4732 4733
4733 4734 /* Hunt the depot. */
4734 4735 mutex_enter(&cp->cache_depot_lock);
4735 4736 n = cp->cache_magtype->mt_magsize;
4736 4737 for (m = cp->cache_full.ml_list; m != NULL; m = m->mag_next) {
4737 4738 if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4738 4739 mutex_exit(&cp->cache_depot_lock);
4739 4740 return (buf);
4740 4741 }
4741 4742 }
4742 4743 mutex_exit(&cp->cache_depot_lock);
4743 4744
4744 4745 /* Hunt the per-CPU magazines. */
4745 4746 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4746 4747 ccp = &cp->cache_cpu[cpu_seqid];
4747 4748
4748 4749 mutex_enter(&ccp->cc_lock);
4749 4750 m = ccp->cc_loaded;
4750 4751 n = ccp->cc_rounds;
4751 4752 if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4752 4753 mutex_exit(&ccp->cc_lock);
4753 4754 return (buf);
4754 4755 }
4755 4756 m = ccp->cc_ploaded;
4756 4757 n = ccp->cc_prounds;
4757 4758 if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4758 4759 mutex_exit(&ccp->cc_lock);
4759 4760 return (buf);
4760 4761 }
4761 4762 mutex_exit(&ccp->cc_lock);
4762 4763 }
4763 4764
4764 4765 kmem_cache_free(cp, tbuf);
4765 4766 return (NULL);
4766 4767 }
4767 4768
4768 4769 /*
4769 4770 * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4770 4771 * or when the buffer is freed.
4771 4772 */
4772 4773 static void
4773 4774 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4774 4775 {
4775 4776 ASSERT(MUTEX_HELD(&cp->cache_lock));
4776 4777 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4777 4778
4778 4779 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4779 4780 return;
4780 4781 }
4781 4782
4782 4783 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4783 4784 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4784 4785 avl_remove(&cp->cache_partial_slabs, sp);
4785 4786 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4786 4787 sp->slab_stuck_offset = (uint32_t)-1;
4787 4788 avl_add(&cp->cache_partial_slabs, sp);
4788 4789 }
4789 4790 } else {
4790 4791 sp->slab_later_count = 0;
4791 4792 sp->slab_stuck_offset = (uint32_t)-1;
4792 4793 }
4793 4794 }
4794 4795
4795 4796 static void
4796 4797 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4797 4798 {
4798 4799 ASSERT(taskq_member(kmem_move_taskq, curthread));
4799 4800 ASSERT(MUTEX_HELD(&cp->cache_lock));
4800 4801 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4801 4802
4802 4803 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4803 4804 return;
4804 4805 }
4805 4806
4806 4807 avl_remove(&cp->cache_partial_slabs, sp);
4807 4808 sp->slab_later_count = 0;
4808 4809 sp->slab_flags |= KMEM_SLAB_NOMOVE;
4809 4810 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4810 4811 avl_add(&cp->cache_partial_slabs, sp);
4811 4812 }
4812 4813
4813 4814 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4814 4815
4815 4816 /*
4816 4817 * The move callback takes two buffer addresses, the buffer to be moved, and a
4817 4818 * newly allocated and constructed buffer selected by kmem as the destination.
4818 4819 * It also takes the size of the buffer and an optional user argument specified
4819 4820 * at cache creation time. kmem guarantees that the buffer to be moved has not
4820 4821 * been unmapped by the virtual memory subsystem. Beyond that, it cannot
4821 4822 * guarantee the present whereabouts of the buffer to be moved, so it is up to
4822 4823 * the client to safely determine whether or not it is still using the buffer.
4823 4824 * The client must not free either of the buffers passed to the move callback,
4824 4825 * since kmem wants to free them directly to the slab layer. The client response
4825 4826 * tells kmem which of the two buffers to free:
4826 4827 *
4827 4828 * YES kmem frees the old buffer (the move was successful)
4828 4829 * NO kmem frees the new buffer, marks the slab of the old buffer
4829 4830 * non-reclaimable to avoid bothering the client again
4830 4831 * LATER kmem frees the new buffer, increments slab_later_count
4831 4832 * DONT_KNOW kmem frees the new buffer, searches mags for the old buffer
4832 4833 * DONT_NEED kmem frees both the old buffer and the new buffer
4833 4834 *
4834 4835 * The pending callback argument now being processed contains both of the
4835 4836 * buffers (old and new) passed to the move callback function, the slab of the
4836 4837 * old buffer, and flags related to the move request, such as whether or not the
4837 4838 * system was desperate for memory.
4838 4839 *
4839 4840 * Slabs are not freed while there is a pending callback, but instead are kept
4840 4841 * on a deadlist, which is drained after the last callback completes. This means
4841 4842 * that slabs are safe to access until kmem_move_end(), no matter how many of
4842 4843 * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4843 4844 * zero for as long as the slab remains on the deadlist and until the slab is
4844 4845 * freed.
4845 4846 */
4846 4847 static void
4847 4848 kmem_move_buffer(kmem_move_t *callback)
4848 4849 {
4849 4850 kmem_cbrc_t response;
4850 4851 kmem_slab_t *sp = callback->kmm_from_slab;
4851 4852 kmem_cache_t *cp = sp->slab_cache;
4852 4853 boolean_t free_on_slab;
4853 4854
4854 4855 ASSERT(taskq_member(kmem_move_taskq, curthread));
4855 4856 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4856 4857 ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4857 4858
4858 4859 /*
4859 4860 * The number of allocated buffers on the slab may have changed since we
4860 4861 * last checked the slab's reclaimability (when the pending move was
4861 4862 * enqueued), or the client may have responded NO when asked to move
4862 4863 * another buffer on the same slab.
4863 4864 */
4864 4865 if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4865 4866 KMEM_STAT_ADD(kmem_move_stats.kms_no_longer_reclaimable);
4866 4867 KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4867 4868 kmem_move_stats.kms_notify_no_longer_reclaimable);
4868 4869 kmem_slab_free(cp, callback->kmm_to_buf);
4869 4870 kmem_move_end(cp, callback);
4870 4871 return;
4871 4872 }
4872 4873
4873 4874 /*
4874 4875 * Hunting magazines is expensive, so we'll wait to do that until the
4875 4876 * client responds KMEM_CBRC_DONT_KNOW. However, checking the slab layer
4876 4877 * is cheap, so we might as well do that here in case we can avoid
4877 4878 * bothering the client.
4878 4879 */
4879 4880 mutex_enter(&cp->cache_lock);
4880 4881 free_on_slab = (kmem_slab_allocated(cp, sp,
4881 4882 callback->kmm_from_buf) == NULL);
4882 4883 mutex_exit(&cp->cache_lock);
4883 4884
4884 4885 if (free_on_slab) {
4885 4886 KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_slab);
4886 4887 kmem_slab_free(cp, callback->kmm_to_buf);
4887 4888 kmem_move_end(cp, callback);
4888 4889 return;
4889 4890 }
4890 4891
4891 4892 if (cp->cache_flags & KMF_BUFTAG) {
4892 4893 /*
4893 4894 * Make kmem_cache_alloc_debug() apply the constructor for us.
4894 4895 */
4895 4896 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4896 4897 KM_NOSLEEP, 1, caller()) != 0) {
4897 4898 KMEM_STAT_ADD(kmem_move_stats.kms_alloc_fail);
4898 4899 kmem_move_end(cp, callback);
4899 4900 return;
4900 4901 }
4901 4902 } else if (cp->cache_constructor != NULL &&
4902 4903 cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4903 4904 KM_NOSLEEP) != 0) {
4904 4905 atomic_inc_64(&cp->cache_alloc_fail);
4905 4906 KMEM_STAT_ADD(kmem_move_stats.kms_constructor_fail);
4906 4907 kmem_slab_free(cp, callback->kmm_to_buf);
4907 4908 kmem_move_end(cp, callback);
4908 4909 return;
4909 4910 }
4910 4911
4911 4912 KMEM_STAT_ADD(kmem_move_stats.kms_callbacks);
4912 4913 KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4913 4914 kmem_move_stats.kms_notify_callbacks);
4914 4915 cp->cache_defrag->kmd_callbacks++;
4915 4916 cp->cache_defrag->kmd_thread = curthread;
4916 4917 cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4917 4918 cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4918 4919 DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4919 4920 callback);
4920 4921
4921 4922 response = cp->cache_move(callback->kmm_from_buf,
4922 4923 callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4923 4924
4924 4925 DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4925 4926 callback, kmem_cbrc_t, response);
4926 4927 cp->cache_defrag->kmd_thread = NULL;
4927 4928 cp->cache_defrag->kmd_from_buf = NULL;
4928 4929 cp->cache_defrag->kmd_to_buf = NULL;
4929 4930
4930 4931 if (response == KMEM_CBRC_YES) {
4931 4932 KMEM_STAT_ADD(kmem_move_stats.kms_yes);
4932 4933 cp->cache_defrag->kmd_yes++;
4933 4934 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4934 4935 /* slab safe to access until kmem_move_end() */
4935 4936 if (sp->slab_refcnt == 0)
4936 4937 cp->cache_defrag->kmd_slabs_freed++;
4937 4938 mutex_enter(&cp->cache_lock);
4938 4939 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4939 4940 mutex_exit(&cp->cache_lock);
4940 4941 kmem_move_end(cp, callback);
4941 4942 return;
4942 4943 }
4943 4944
4944 4945 switch (response) {
4945 4946 case KMEM_CBRC_NO:
4946 4947 KMEM_STAT_ADD(kmem_move_stats.kms_no);
4947 4948 cp->cache_defrag->kmd_no++;
4948 4949 mutex_enter(&cp->cache_lock);
4949 4950 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4950 4951 mutex_exit(&cp->cache_lock);
4951 4952 break;
4952 4953 case KMEM_CBRC_LATER:
4953 4954 KMEM_STAT_ADD(kmem_move_stats.kms_later);
4954 4955 cp->cache_defrag->kmd_later++;
4955 4956 mutex_enter(&cp->cache_lock);
4956 4957 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4957 4958 mutex_exit(&cp->cache_lock);
4958 4959 break;
4959 4960 }
4960 4961
4961 4962 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4962 4963 KMEM_STAT_ADD(kmem_move_stats.kms_disbelief);
4963 4964 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4964 4965 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4965 4966 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4966 4967 callback->kmm_from_buf);
4967 4968 }
4968 4969 mutex_exit(&cp->cache_lock);
4969 4970 break;
4970 4971 case KMEM_CBRC_DONT_NEED:
4971 4972 KMEM_STAT_ADD(kmem_move_stats.kms_dont_need);
4972 4973 cp->cache_defrag->kmd_dont_need++;
4973 4974 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4974 4975 if (sp->slab_refcnt == 0)
4975 4976 cp->cache_defrag->kmd_slabs_freed++;
4976 4977 mutex_enter(&cp->cache_lock);
4977 4978 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4978 4979 mutex_exit(&cp->cache_lock);
4979 4980 break;
4980 4981 case KMEM_CBRC_DONT_KNOW:
4981 4982 KMEM_STAT_ADD(kmem_move_stats.kms_dont_know);
4982 4983 cp->cache_defrag->kmd_dont_know++;
4983 4984 if (kmem_hunt_mags(cp, callback->kmm_from_buf) != NULL) {
4984 4985 KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_mag);
4985 4986 cp->cache_defrag->kmd_hunt_found++;
4986 4987 kmem_slab_free_constructed(cp, callback->kmm_from_buf,
4987 4988 B_TRUE);
4988 4989 if (sp->slab_refcnt == 0)
4989 4990 cp->cache_defrag->kmd_slabs_freed++;
4990 4991 mutex_enter(&cp->cache_lock);
4991 4992 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4992 4993 mutex_exit(&cp->cache_lock);
4993 4994 }
4994 4995 break;
4995 4996 default:
4996 4997 panic("'%s' (%p) unexpected move callback response %d\n",
4997 4998 cp->cache_name, (void *)cp, response);
4998 4999 }
4999 5000
5000 5001 kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
5001 5002 kmem_move_end(cp, callback);
5002 5003 }
5003 5004
5004 5005 /* Return B_FALSE if there is insufficient memory for the move request. */
5005 5006 static boolean_t
5006 5007 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
5007 5008 {
5008 5009 void *to_buf;
5009 5010 avl_index_t index;
5010 5011 kmem_move_t *callback, *pending;
5011 5012 ulong_t n;
5012 5013
5013 5014 ASSERT(taskq_member(kmem_taskq, curthread));
5014 5015 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
5015 5016 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5016 5017
5017 5018 callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
5018 5019 if (callback == NULL) {
5019 5020 KMEM_STAT_ADD(kmem_move_stats.kms_callback_alloc_fail);
5020 5021 return (B_FALSE);
5021 5022 }
5022 5023
5023 5024 callback->kmm_from_slab = sp;
5024 5025 callback->kmm_from_buf = buf;
5025 5026 callback->kmm_flags = flags;
5026 5027
5027 5028 mutex_enter(&cp->cache_lock);
5028 5029
5029 5030 n = avl_numnodes(&cp->cache_partial_slabs);
5030 5031 if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
5031 5032 mutex_exit(&cp->cache_lock);
5032 5033 kmem_cache_free(kmem_move_cache, callback);
5033 5034 return (B_TRUE); /* there is no need for the move request */
5034 5035 }
5035 5036
5036 5037 pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
5037 5038 if (pending != NULL) {
5038 5039 /*
5039 5040 * If the move is already pending and we're desperate now,
5040 5041 * update the move flags.
5041 5042 */
5042 5043 if (flags & KMM_DESPERATE) {
5043 5044 pending->kmm_flags |= KMM_DESPERATE;
5044 5045 }
5045 5046 mutex_exit(&cp->cache_lock);
5046 5047 KMEM_STAT_ADD(kmem_move_stats.kms_already_pending);
5047 5048 kmem_cache_free(kmem_move_cache, callback);
5048 5049 return (B_TRUE);
5049 5050 }
5050 5051
5051 5052 to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
5052 5053 B_FALSE);
5053 5054 callback->kmm_to_buf = to_buf;
5054 5055 avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
5055 5056
5056 5057 mutex_exit(&cp->cache_lock);
5057 5058
5058 5059 if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
5059 5060 callback, TQ_NOSLEEP)) {
5060 5061 KMEM_STAT_ADD(kmem_move_stats.kms_callback_taskq_fail);
5061 5062 mutex_enter(&cp->cache_lock);
5062 5063 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5063 5064 mutex_exit(&cp->cache_lock);
5064 5065 kmem_slab_free(cp, to_buf);
5065 5066 kmem_cache_free(kmem_move_cache, callback);
5066 5067 return (B_FALSE);
5067 5068 }
5068 5069
5069 5070 return (B_TRUE);
5070 5071 }
5071 5072
5072 5073 static void
5073 5074 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
5074 5075 {
5075 5076 avl_index_t index;
5076 5077
5077 5078 ASSERT(cp->cache_defrag != NULL);
5078 5079 ASSERT(taskq_member(kmem_move_taskq, curthread));
5079 5080 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
5080 5081
5081 5082 mutex_enter(&cp->cache_lock);
5082 5083 VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
5083 5084 callback->kmm_from_buf, &index) != NULL);
5084 5085 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
5085 5086 if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
5086 5087 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5087 5088 kmem_slab_t *sp;
5088 5089
5089 5090 /*
5090 5091 * The last pending move completed. Release all slabs from the
5091 5092 * front of the dead list except for any slab at the tail that
5092 5093 * needs to be released from the context of kmem_move_buffers().
5093 5094 * kmem deferred unmapping the buffers on these slabs in order
5094 5095 * to guarantee that buffers passed to the move callback have
5095 5096 * been touched only by kmem or by the client itself.
5096 5097 */
5097 5098 while ((sp = list_remove_head(deadlist)) != NULL) {
5098 5099 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
5099 5100 list_insert_tail(deadlist, sp);
5100 5101 break;
5101 5102 }
5102 5103 cp->cache_defrag->kmd_deadcount--;
5103 5104 cp->cache_slab_destroy++;
5104 5105 mutex_exit(&cp->cache_lock);
5105 5106 kmem_slab_destroy(cp, sp);
5106 5107 KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5107 5108 mutex_enter(&cp->cache_lock);
5108 5109 }
5109 5110 }
5110 5111 mutex_exit(&cp->cache_lock);
5111 5112 kmem_cache_free(kmem_move_cache, callback);
5112 5113 }
5113 5114
5114 5115 /*
5115 5116 * Move buffers from least used slabs first by scanning backwards from the end
5116 5117 * of the partial slab list. Scan at most max_scan candidate slabs and move
5117 5118 * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
5118 5119 * If desperate to reclaim memory, move buffers from any partial slab, otherwise
5119 5120 * skip slabs with a ratio of allocated buffers at or above the current
5120 5121 * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5121 5122 * scan is aborted) so that the caller can adjust the reclaimability threshold
5122 5123 * depending on how many reclaimable slabs it finds.
5123 5124 *
5124 5125 * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5125 5126 * move request, since it is not valid for kmem_move_begin() to call
5126 5127 * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5127 5128 */
5128 5129 static int
5129 5130 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5130 5131 int flags)
5131 5132 {
5132 5133 kmem_slab_t *sp;
5133 5134 void *buf;
5134 5135 int i, j; /* slab index, buffer index */
5135 5136 int s; /* reclaimable slabs */
5136 5137 int b; /* allocated (movable) buffers on reclaimable slab */
5137 5138 boolean_t success;
5138 5139 int refcnt;
5139 5140 int nomove;
5140 5141
5141 5142 ASSERT(taskq_member(kmem_taskq, curthread));
5142 5143 ASSERT(MUTEX_HELD(&cp->cache_lock));
5143 5144 ASSERT(kmem_move_cache != NULL);
5144 5145 ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5145 5146 ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5146 5147 avl_numnodes(&cp->cache_partial_slabs) > 1);
5147 5148
5148 5149 if (kmem_move_blocked) {
5149 5150 return (0);
5150 5151 }
5151 5152
5152 5153 if (kmem_move_fulltilt) {
5153 5154 flags |= KMM_DESPERATE;
5154 5155 }
5155 5156
5156 5157 if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5157 5158 /*
5158 5159 * Scan as many slabs as needed to find the desired number of
5159 5160 * candidate slabs.
5160 5161 */
5161 5162 max_scan = (size_t)-1;
5162 5163 }
5163 5164
5164 5165 if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5165 5166 /* Find as many candidate slabs as possible. */
5166 5167 max_slabs = (size_t)-1;
5167 5168 }
5168 5169
5169 5170 sp = avl_last(&cp->cache_partial_slabs);
5170 5171 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5171 5172 for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5172 5173 ((sp != avl_first(&cp->cache_partial_slabs)) ||
5173 5174 (flags & KMM_DEBUG));
5174 5175 sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5175 5176
5176 5177 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5177 5178 continue;
5178 5179 }
5179 5180 s++;
5180 5181
5181 5182 /* Look for allocated buffers to move. */
5182 5183 for (j = 0, b = 0, buf = sp->slab_base;
5183 5184 (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5184 5185 buf = (((char *)buf) + cp->cache_chunksize), j++) {
5185 5186
5186 5187 if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5187 5188 continue;
5188 5189 }
5189 5190
5190 5191 b++;
5191 5192
5192 5193 /*
5193 5194 * Prevent the slab from being destroyed while we drop
5194 5195 * cache_lock and while the pending move is not yet
5195 5196 * registered. Flag the pending move while
5196 5197 * kmd_moves_pending may still be empty, since we can't
5197 5198 * yet rely on a non-zero pending move count to prevent
5198 5199 * the slab from being destroyed.
5199 5200 */
5200 5201 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5201 5202 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5202 5203 /*
5203 5204 * Recheck refcnt and nomove after reacquiring the lock,
5204 5205 * since these control the order of partial slabs, and
5205 5206 * we want to know if we can pick up the scan where we
5206 5207 * left off.
5207 5208 */
5208 5209 refcnt = sp->slab_refcnt;
5209 5210 nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5210 5211 mutex_exit(&cp->cache_lock);
5211 5212
5212 5213 success = kmem_move_begin(cp, sp, buf, flags);
5213 5214
5214 5215 /*
5215 5216 * Now, before the lock is reacquired, kmem could
5216 5217 * process all pending move requests and purge the
5217 5218 * deadlist, so that upon reacquiring the lock, sp has
5218 5219 * been remapped. Or, the client may free all the
5219 5220 * objects on the slab while the pending moves are still
5220 5221 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5221 5222 * flag causes the slab to be put at the end of the
5222 5223 * deadlist and prevents it from being destroyed, since
5223 5224 * we plan to destroy it here after reacquiring the
5224 5225 * lock.
5225 5226 */
5226 5227 mutex_enter(&cp->cache_lock);
5227 5228 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5228 5229 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5229 5230
5230 5231 if (sp->slab_refcnt == 0) {
5231 5232 list_t *deadlist =
5232 5233 &cp->cache_defrag->kmd_deadlist;
5233 5234 list_remove(deadlist, sp);
5234 5235
5235 5236 if (!avl_is_empty(
5236 5237 &cp->cache_defrag->kmd_moves_pending)) {
5237 5238 /*
5238 5239 * A pending move makes it unsafe to
5239 5240 * destroy the slab, because even though
5240 5241 * the move is no longer needed, the
5241 5242 * context where that is determined
5242 5243 * requires the slab to exist.
5243 5244 * Fortunately, a pending move also
5244 5245 * means we don't need to destroy the
5245 5246 * slab here, since it will get
5246 5247 * destroyed along with any other slabs
5247 5248 * on the deadlist after the last
5248 5249 * pending move completes.
5249 5250 */
5250 5251 list_insert_head(deadlist, sp);
5251 5252 KMEM_STAT_ADD(kmem_move_stats.
5252 5253 kms_endscan_slab_dead);
5253 5254 return (-1);
5254 5255 }
5255 5256
5256 5257 /*
5257 5258 * Destroy the slab now if it was completely
5258 5259 * freed while we dropped cache_lock and there
5259 5260 * are no pending moves. Since slab_refcnt
5260 5261 * cannot change once it reaches zero, no new
5261 5262 * pending moves from that slab are possible.
5262 5263 */
5263 5264 cp->cache_defrag->kmd_deadcount--;
5264 5265 cp->cache_slab_destroy++;
5265 5266 mutex_exit(&cp->cache_lock);
5266 5267 kmem_slab_destroy(cp, sp);
5267 5268 KMEM_STAT_ADD(kmem_move_stats.
5268 5269 kms_dead_slabs_freed);
5269 5270 KMEM_STAT_ADD(kmem_move_stats.
5270 5271 kms_endscan_slab_destroyed);
5271 5272 mutex_enter(&cp->cache_lock);
5272 5273 /*
5273 5274 * Since we can't pick up the scan where we left
5274 5275 * off, abort the scan and say nothing about the
5275 5276 * number of reclaimable slabs.
5276 5277 */
5277 5278 return (-1);
5278 5279 }
5279 5280
5280 5281 if (!success) {
5281 5282 /*
5282 5283 * Abort the scan if there is not enough memory
5283 5284 * for the request and say nothing about the
5284 5285 * number of reclaimable slabs.
5285 5286 */
5286 5287 KMEM_STAT_COND_ADD(s < max_slabs,
5287 5288 kmem_move_stats.kms_endscan_nomem);
5288 5289 return (-1);
5289 5290 }
5290 5291
5291 5292 /*
5292 5293 * The slab's position changed while the lock was
5293 5294 * dropped, so we don't know where we are in the
5294 5295 * sequence any more.
5295 5296 */
5296 5297 if (sp->slab_refcnt != refcnt) {
5297 5298 /*
5298 5299 * If this is a KMM_DEBUG move, the slab_refcnt
5299 5300 * may have changed because we allocated a
5300 5301 * destination buffer on the same slab. In that
5301 5302 * case, we're not interested in counting it.
5302 5303 */
5303 5304 KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5304 5305 (s < max_slabs),
5305 5306 kmem_move_stats.kms_endscan_refcnt_changed);
5306 5307 return (-1);
5307 5308 }
5308 5309 if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove) {
5309 5310 KMEM_STAT_COND_ADD(s < max_slabs,
5310 5311 kmem_move_stats.kms_endscan_nomove_changed);
5311 5312 return (-1);
5312 5313 }
5313 5314
5314 5315 /*
5315 5316 * Generating a move request allocates a destination
5316 5317 * buffer from the slab layer, bumping the first partial
5317 5318 * slab if it is completely allocated. If the current
5318 5319 * slab becomes the first partial slab as a result, we
5319 5320 * can't continue to scan backwards.
5320 5321 *
5321 5322 * If this is a KMM_DEBUG move and we allocated the
5322 5323 * destination buffer from the last partial slab, then
5323 5324 * the buffer we're moving is on the same slab and our
5324 5325 * slab_refcnt has changed, causing us to return before
5325 5326 * reaching here if there are no partial slabs left.
5326 5327 */
5327 5328 ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5328 5329 if (sp == avl_first(&cp->cache_partial_slabs)) {
5329 5330 /*
5330 5331 * We're not interested in a second KMM_DEBUG
5331 5332 * move.
5332 5333 */
5333 5334 goto end_scan;
5334 5335 }
5335 5336 }
5336 5337 }
5337 5338 end_scan:
5338 5339
5339 5340 KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5340 5341 (s < max_slabs) &&
5341 5342 (sp == avl_first(&cp->cache_partial_slabs)),
5342 5343 kmem_move_stats.kms_endscan_freelist);
5343 5344
5344 5345 return (s);
5345 5346 }
5346 5347
5347 5348 typedef struct kmem_move_notify_args {
5348 5349 kmem_cache_t *kmna_cache;
5349 5350 void *kmna_buf;
5350 5351 } kmem_move_notify_args_t;
5351 5352
5352 5353 static void
5353 5354 kmem_cache_move_notify_task(void *arg)
5354 5355 {
5355 5356 kmem_move_notify_args_t *args = arg;
5356 5357 kmem_cache_t *cp = args->kmna_cache;
5357 5358 void *buf = args->kmna_buf;
5358 5359 kmem_slab_t *sp;
5359 5360
5360 5361 ASSERT(taskq_member(kmem_taskq, curthread));
5361 5362 ASSERT(list_link_active(&cp->cache_link));
5362 5363
5363 5364 kmem_free(args, sizeof (kmem_move_notify_args_t));
5364 5365 mutex_enter(&cp->cache_lock);
5365 5366 sp = kmem_slab_allocated(cp, NULL, buf);
5366 5367
5367 5368 /* Ignore the notification if the buffer is no longer allocated. */
5368 5369 if (sp == NULL) {
5369 5370 mutex_exit(&cp->cache_lock);
5370 5371 return;
5371 5372 }
5372 5373
5373 5374 /* Ignore the notification if there's no reason to move the buffer. */
5374 5375 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5375 5376 /*
5376 5377 * So far the notification is not ignored. Ignore the
5377 5378 * notification if the slab is not marked by an earlier refusal
5378 5379 * to move a buffer.
5379 5380 */
5380 5381 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5381 5382 (sp->slab_later_count == 0)) {
5382 5383 mutex_exit(&cp->cache_lock);
5383 5384 return;
5384 5385 }
5385 5386
5386 5387 kmem_slab_move_yes(cp, sp, buf);
5387 5388 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5388 5389 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5389 5390 mutex_exit(&cp->cache_lock);
5390 5391 /* see kmem_move_buffers() about dropping the lock */
5391 5392 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5392 5393 mutex_enter(&cp->cache_lock);
5393 5394 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5394 5395 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5395 5396 if (sp->slab_refcnt == 0) {
5396 5397 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5397 5398 list_remove(deadlist, sp);
5398 5399
5399 5400 if (!avl_is_empty(
5400 5401 &cp->cache_defrag->kmd_moves_pending)) {
5401 5402 list_insert_head(deadlist, sp);
5402 5403 mutex_exit(&cp->cache_lock);
5403 5404 KMEM_STAT_ADD(kmem_move_stats.
5404 5405 kms_notify_slab_dead);
5405 5406 return;
5406 5407 }
5407 5408
5408 5409 cp->cache_defrag->kmd_deadcount--;
5409 5410 cp->cache_slab_destroy++;
5410 5411 mutex_exit(&cp->cache_lock);
5411 5412 kmem_slab_destroy(cp, sp);
5412 5413 KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5413 5414 KMEM_STAT_ADD(kmem_move_stats.
5414 5415 kms_notify_slab_destroyed);
5415 5416 return;
5416 5417 }
5417 5418 } else {
5418 5419 kmem_slab_move_yes(cp, sp, buf);
5419 5420 }
5420 5421 mutex_exit(&cp->cache_lock);
5421 5422 }
5422 5423
5423 5424 void
5424 5425 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5425 5426 {
5426 5427 kmem_move_notify_args_t *args;
5427 5428
5428 5429 KMEM_STAT_ADD(kmem_move_stats.kms_notify);
5429 5430 args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5430 5431 if (args != NULL) {
5431 5432 args->kmna_cache = cp;
5432 5433 args->kmna_buf = buf;
5433 5434 if (!taskq_dispatch(kmem_taskq,
5434 5435 (task_func_t *)kmem_cache_move_notify_task, args,
5435 5436 TQ_NOSLEEP))
5436 5437 kmem_free(args, sizeof (kmem_move_notify_args_t));
5437 5438 }
5438 5439 }
5439 5440
5440 5441 static void
5441 5442 kmem_cache_defrag(kmem_cache_t *cp)
5442 5443 {
5443 5444 size_t n;
5444 5445
5445 5446 ASSERT(cp->cache_defrag != NULL);
5446 5447
5447 5448 mutex_enter(&cp->cache_lock);
5448 5449 n = avl_numnodes(&cp->cache_partial_slabs);
5449 5450 if (n > 1) {
5450 5451 /* kmem_move_buffers() drops and reacquires cache_lock */
5451 5452 KMEM_STAT_ADD(kmem_move_stats.kms_defrags);
5452 5453 cp->cache_defrag->kmd_defrags++;
5453 5454 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5454 5455 }
5455 5456 mutex_exit(&cp->cache_lock);
5456 5457 }
5457 5458
5458 5459 /* Is this cache above the fragmentation threshold? */
5459 5460 static boolean_t
5460 5461 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5461 5462 {
5462 5463 /*
5463 5464 * nfree kmem_frag_numer
5464 5465 * ------------------ > ---------------
5465 5466 * cp->cache_buftotal kmem_frag_denom
5466 5467 */
5467 5468 return ((nfree * kmem_frag_denom) >
5468 5469 (cp->cache_buftotal * kmem_frag_numer));
5469 5470 }
5470 5471
5471 5472 static boolean_t
5472 5473 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5473 5474 {
5474 5475 boolean_t fragmented;
5475 5476 uint64_t nfree;
5476 5477
5477 5478 ASSERT(MUTEX_HELD(&cp->cache_lock));
5478 5479 *doreap = B_FALSE;
5479 5480
5480 5481 if (kmem_move_fulltilt) {
5481 5482 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5482 5483 return (B_TRUE);
5483 5484 }
5484 5485 } else {
5485 5486 if ((cp->cache_complete_slab_count + avl_numnodes(
5486 5487 &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5487 5488 return (B_FALSE);
5488 5489 }
5489 5490 }
5490 5491
5491 5492 nfree = cp->cache_bufslab;
5492 5493 fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5493 5494 kmem_cache_frag_threshold(cp, nfree));
5494 5495
5495 5496 /*
5496 5497 * Free buffers in the magazine layer appear allocated from the point of
5497 5498 * view of the slab layer. We want to know if the slab layer would
5498 5499 * appear fragmented if we included free buffers from magazines that
5499 5500 * have fallen out of the working set.
5500 5501 */
5501 5502 if (!fragmented) {
5502 5503 long reap;
5503 5504
5504 5505 mutex_enter(&cp->cache_depot_lock);
5505 5506 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5506 5507 reap = MIN(reap, cp->cache_full.ml_total);
5507 5508 mutex_exit(&cp->cache_depot_lock);
5508 5509
5509 5510 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5510 5511 if (kmem_cache_frag_threshold(cp, nfree)) {
5511 5512 *doreap = B_TRUE;
5512 5513 }
5513 5514 }
5514 5515
5515 5516 return (fragmented);
5516 5517 }
5517 5518
5518 5519 /* Called periodically from kmem_taskq */
5519 5520 static void
5520 5521 kmem_cache_scan(kmem_cache_t *cp)
5521 5522 {
5522 5523 boolean_t reap = B_FALSE;
5523 5524 kmem_defrag_t *kmd;
5524 5525
5525 5526 ASSERT(taskq_member(kmem_taskq, curthread));
5526 5527
5527 5528 mutex_enter(&cp->cache_lock);
5528 5529
5529 5530 kmd = cp->cache_defrag;
5530 5531 if (kmd->kmd_consolidate > 0) {
5531 5532 kmd->kmd_consolidate--;
5532 5533 mutex_exit(&cp->cache_lock);
5533 5534 kmem_cache_reap(cp);
5534 5535 return;
5535 5536 }
5536 5537
5537 5538 if (kmem_cache_is_fragmented(cp, &reap)) {
5538 5539 size_t slabs_found;
5539 5540
5540 5541 /*
5541 5542 * Consolidate reclaimable slabs from the end of the partial
5542 5543 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5543 5544 * reclaimable slabs). Keep track of how many candidate slabs we
5544 5545 * looked for and how many we actually found so we can adjust
5545 5546 * the definition of a candidate slab if we're having trouble
5546 5547 * finding them.
5547 5548 *
5548 5549 * kmem_move_buffers() drops and reacquires cache_lock.
5549 5550 */
5550 5551 KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5551 5552 kmd->kmd_scans++;
5552 5553 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5553 5554 kmem_reclaim_max_slabs, 0);
5554 5555 if (slabs_found >= 0) {
5555 5556 kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5556 5557 kmd->kmd_slabs_found += slabs_found;
5557 5558 }
5558 5559
5559 5560 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5560 5561 kmd->kmd_tries = 0;
5561 5562
5562 5563 /*
5563 5564 * If we had difficulty finding candidate slabs in
5564 5565 * previous scans, adjust the threshold so that
5565 5566 * candidates are easier to find.
5566 5567 */
5567 5568 if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5568 5569 kmem_adjust_reclaim_threshold(kmd, -1);
5569 5570 } else if ((kmd->kmd_slabs_found * 2) <
5570 5571 kmd->kmd_slabs_sought) {
5571 5572 kmem_adjust_reclaim_threshold(kmd, 1);
5572 5573 }
5573 5574 kmd->kmd_slabs_sought = 0;
5574 5575 kmd->kmd_slabs_found = 0;
5575 5576 }
5576 5577 } else {
5577 5578 kmem_reset_reclaim_threshold(cp->cache_defrag);
5578 5579 #ifdef DEBUG
5579 5580 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5580 5581 /*
5581 5582 * In a debug kernel we want the consolidator to
5582 5583 * run occasionally even when there is plenty of
5583 5584 * memory.
5584 5585 */
5585 5586 uint16_t debug_rand;
5586 5587
5587 5588 (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5588 5589 if (!kmem_move_noreap &&
5589 5590 ((debug_rand % kmem_mtb_reap) == 0)) {
5590 5591 mutex_exit(&cp->cache_lock);
5591 5592 KMEM_STAT_ADD(kmem_move_stats.kms_debug_reaps);
5592 5593 kmem_cache_reap(cp);
5593 5594 return;
5594 5595 } else if ((debug_rand % kmem_mtb_move) == 0) {
5595 5596 KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5596 5597 KMEM_STAT_ADD(kmem_move_stats.kms_debug_scans);
5597 5598 kmd->kmd_scans++;
5598 5599 (void) kmem_move_buffers(cp,
5599 5600 kmem_reclaim_scan_range, 1, KMM_DEBUG);
5600 5601 }
5601 5602 }
5602 5603 #endif /* DEBUG */
5603 5604 }
5604 5605
5605 5606 mutex_exit(&cp->cache_lock);
5606 5607
5607 5608 if (reap) {
5608 5609 KMEM_STAT_ADD(kmem_move_stats.kms_scan_depot_ws_reaps);
5609 5610 kmem_depot_ws_reap(cp);
5610 5611 }
5611 5612 }
|
↓ open down ↓ |
1448 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX