Print this page
| 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 *
|
↓ open down ↓ |
12 lines elided |
↑ open up ↑ |
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 - * Copyright (c) 2015 Joyent, Inc. All rights reserved.
24 23 * Copyright (c) 2012, 2016 by Delphix. All rights reserved.
25 24 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
26 25 */
27 26
28 27 /*
29 28 * Kernel memory allocator, as described in the following two papers and a
30 29 * statement about the consolidator:
31 30 *
32 31 * Jeff Bonwick,
33 32 * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
34 33 * Proceedings of the Summer 1994 Usenix Conference.
35 34 * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
36 35 *
37 36 * Jeff Bonwick and Jonathan Adams,
38 37 * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
39 38 * Arbitrary Resources.
40 39 * Proceedings of the 2001 Usenix Conference.
41 40 * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
42 41 *
43 42 * kmem Slab Consolidator Big Theory Statement:
44 43 *
45 44 * 1. Motivation
46 45 *
47 46 * As stated in Bonwick94, slabs provide the following advantages over other
48 47 * allocation structures in terms of memory fragmentation:
49 48 *
50 49 * - Internal fragmentation (per-buffer wasted space) is minimal.
51 50 * - Severe external fragmentation (unused buffers on the free list) is
52 51 * unlikely.
53 52 *
54 53 * Segregating objects by size eliminates one source of external fragmentation,
55 54 * and according to Bonwick:
56 55 *
57 56 * The other reason that slabs reduce external fragmentation is that all
58 57 * objects in a slab are of the same type, so they have the same lifetime
59 58 * distribution. The resulting segregation of short-lived and long-lived
60 59 * objects at slab granularity reduces the likelihood of an entire page being
61 60 * held hostage due to a single long-lived allocation [Barrett93, Hanson90].
62 61 *
63 62 * While unlikely, severe external fragmentation remains possible. Clients that
64 63 * allocate both short- and long-lived objects from the same cache cannot
65 64 * anticipate the distribution of long-lived objects within the allocator's slab
66 65 * implementation. Even a small percentage of long-lived objects distributed
67 66 * randomly across many slabs can lead to a worst case scenario where the client
68 67 * frees the majority of its objects and the system gets back almost none of the
69 68 * slabs. Despite the client doing what it reasonably can to help the system
70 69 * reclaim memory, the allocator cannot shake free enough slabs because of
71 70 * lonely allocations stubbornly hanging on. Although the allocator is in a
72 71 * position to diagnose the fragmentation, there is nothing that the allocator
73 72 * by itself can do about it. It only takes a single allocated object to prevent
74 73 * an entire slab from being reclaimed, and any object handed out by
75 74 * kmem_cache_alloc() is by definition in the client's control. Conversely,
76 75 * although the client is in a position to move a long-lived object, it has no
77 76 * way of knowing if the object is causing fragmentation, and if so, where to
78 77 * move it. A solution necessarily requires further cooperation between the
79 78 * allocator and the client.
80 79 *
81 80 * 2. Move Callback
82 81 *
83 82 * The kmem slab consolidator therefore adds a move callback to the
84 83 * allocator/client interface, improving worst-case external fragmentation in
85 84 * kmem caches that supply a function to move objects from one memory location
86 85 * to another. In a situation of low memory kmem attempts to consolidate all of
87 86 * a cache's slabs at once; otherwise it works slowly to bring external
88 87 * fragmentation within the 1/8 limit guaranteed for internal fragmentation,
89 88 * thereby helping to avoid a low memory situation in the future.
90 89 *
91 90 * The callback has the following signature:
92 91 *
93 92 * kmem_cbrc_t move(void *old, void *new, size_t size, void *user_arg)
94 93 *
95 94 * It supplies the kmem client with two addresses: the allocated object that
96 95 * kmem wants to move and a buffer selected by kmem for the client to use as the
97 96 * copy destination. The callback is kmem's way of saying "Please get off of
98 97 * this buffer and use this one instead." kmem knows where it wants to move the
99 98 * object in order to best reduce fragmentation. All the client needs to know
100 99 * about the second argument (void *new) is that it is an allocated, constructed
101 100 * object ready to take the contents of the old object. When the move function
102 101 * is called, the system is likely to be low on memory, and the new object
103 102 * spares the client from having to worry about allocating memory for the
104 103 * requested move. The third argument supplies the size of the object, in case a
105 104 * single move function handles multiple caches whose objects differ only in
106 105 * size (such as zio_buf_512, zio_buf_1024, etc). Finally, the same optional
107 106 * user argument passed to the constructor, destructor, and reclaim functions is
108 107 * also passed to the move callback.
109 108 *
110 109 * 2.1 Setting the Move Callback
111 110 *
112 111 * The client sets the move callback after creating the cache and before
113 112 * allocating from it:
114 113 *
115 114 * object_cache = kmem_cache_create(...);
116 115 * kmem_cache_set_move(object_cache, object_move);
117 116 *
118 117 * 2.2 Move Callback Return Values
119 118 *
120 119 * Only the client knows about its own data and when is a good time to move it.
121 120 * The client is cooperating with kmem to return unused memory to the system,
122 121 * and kmem respectfully accepts this help at the client's convenience. When
123 122 * asked to move an object, the client can respond with any of the following:
124 123 *
125 124 * typedef enum kmem_cbrc {
126 125 * KMEM_CBRC_YES,
127 126 * KMEM_CBRC_NO,
128 127 * KMEM_CBRC_LATER,
129 128 * KMEM_CBRC_DONT_NEED,
130 129 * KMEM_CBRC_DONT_KNOW
131 130 * } kmem_cbrc_t;
132 131 *
133 132 * The client must not explicitly kmem_cache_free() either of the objects passed
134 133 * to the callback, since kmem wants to free them directly to the slab layer
135 134 * (bypassing the per-CPU magazine layer). The response tells kmem which of the
136 135 * objects to free:
137 136 *
138 137 * YES: (Did it) The client moved the object, so kmem frees the old one.
139 138 * NO: (Never) The client refused, so kmem frees the new object (the
140 139 * unused copy destination). kmem also marks the slab of the old
141 140 * object so as not to bother the client with further callbacks for
142 141 * that object as long as the slab remains on the partial slab list.
143 142 * (The system won't be getting the slab back as long as the
144 143 * immovable object holds it hostage, so there's no point in moving
145 144 * any of its objects.)
146 145 * LATER: The client is using the object and cannot move it now, so kmem
147 146 * frees the new object (the unused copy destination). kmem still
148 147 * attempts to move other objects off the slab, since it expects to
149 148 * succeed in clearing the slab in a later callback. The client
150 149 * should use LATER instead of NO if the object is likely to become
151 150 * movable very soon.
152 151 * DONT_NEED: The client no longer needs the object, so kmem frees the old along
|
↓ open down ↓ |
119 lines elided |
↑ open up ↑ |
153 152 * with the new object (the unused copy destination). This response
154 153 * is the client's opportunity to be a model citizen and give back as
155 154 * much as it can.
156 155 * DONT_KNOW: The client does not know about the object because
157 156 * a) the client has just allocated the object and not yet put it
158 157 * wherever it expects to find known objects
159 158 * b) the client has removed the object from wherever it expects to
160 159 * find known objects and is about to free it, or
161 160 * c) the client has freed the object.
162 161 * In all these cases (a, b, and c) kmem frees the new object (the
163 - * unused copy destination). In the first case, the object is in
164 - * use and the correct action is that for LATER; in the latter two
165 - * cases, we know that the object is either freed or about to be
166 - * freed, in which case it is either already in a magazine or about
167 - * to be in one. In these cases, we know that the object will either
168 - * be reallocated and reused, or it will end up in a full magazine
169 - * that will be reaped (thereby liberating the slab). Because it
170 - * is prohibitively expensive to differentiate these cases, and
171 - * because the defrag code is executed when we're low on memory
172 - * (thereby biasing the system to reclaim full magazines) we treat
173 - * all DONT_KNOW cases as LATER and rely on cache reaping to
174 - * generally clean up full magazines. While we take the same action
175 - * for these cases, we maintain their semantic distinction: if
176 - * defragmentation is not occurring, it is useful to know if this
177 - * is due to objects in use (LATER) or objects in an unknown state
178 - * of transition (DONT_KNOW).
162 + * unused copy destination) and searches for the old object in the
163 + * magazine layer. If found, the object is removed from the magazine
164 + * layer and freed to the slab layer so it will no longer hold the
165 + * slab hostage.
179 166 *
180 167 * 2.3 Object States
181 168 *
182 169 * Neither kmem nor the client can be assumed to know the object's whereabouts
183 170 * at the time of the callback. An object belonging to a kmem cache may be in
184 171 * any of the following states:
185 172 *
186 173 * 1. Uninitialized on the slab
187 174 * 2. Allocated from the slab but not constructed (still uninitialized)
188 175 * 3. Allocated from the slab, constructed, but not yet ready for business
189 176 * (not in a valid state for the move callback)
190 177 * 4. In use (valid and known to the client)
191 178 * 5. About to be freed (no longer in a valid state for the move callback)
192 179 * 6. Freed to a magazine (still constructed)
193 180 * 7. Allocated from a magazine, not yet ready for business (not in a valid
194 181 * state for the move callback), and about to return to state #4
195 182 * 8. Deconstructed on a magazine that is about to be freed
196 183 * 9. Freed to the slab
197 184 *
198 185 * Since the move callback may be called at any time while the object is in any
199 186 * of the above states (except state #1), the client needs a safe way to
200 187 * determine whether or not it knows about the object. Specifically, the client
201 188 * needs to know whether or not the object is in state #4, the only state in
202 189 * which a move is valid. If the object is in any other state, the client should
203 190 * immediately return KMEM_CBRC_DONT_KNOW, since it is unsafe to access any of
204 191 * the object's fields.
205 192 *
206 193 * Note that although an object may be in state #4 when kmem initiates the move
207 194 * request, the object may no longer be in that state by the time kmem actually
208 195 * calls the move function. Not only does the client free objects
209 196 * asynchronously, kmem itself puts move requests on a queue where thay are
210 197 * pending until kmem processes them from another context. Also, objects freed
211 198 * to a magazine appear allocated from the point of view of the slab layer, so
212 199 * kmem may even initiate requests for objects in a state other than state #4.
213 200 *
214 201 * 2.3.1 Magazine Layer
215 202 *
216 203 * An important insight revealed by the states listed above is that the magazine
217 204 * layer is populated only by kmem_cache_free(). Magazines of constructed
218 205 * objects are never populated directly from the slab layer (which contains raw,
219 206 * unconstructed objects). Whenever an allocation request cannot be satisfied
220 207 * from the magazine layer, the magazines are bypassed and the request is
221 208 * satisfied from the slab layer (creating a new slab if necessary). kmem calls
222 209 * the object constructor only when allocating from the slab layer, and only in
223 210 * response to kmem_cache_alloc() or to prepare the destination buffer passed in
224 211 * the move callback. kmem does not preconstruct objects in anticipation of
225 212 * kmem_cache_alloc().
226 213 *
227 214 * 2.3.2 Object Constructor and Destructor
228 215 *
229 216 * If the client supplies a destructor, it must be valid to call the destructor
230 217 * on a newly created object (immediately after the constructor).
231 218 *
232 219 * 2.4 Recognizing Known Objects
233 220 *
234 221 * There is a simple test to determine safely whether or not the client knows
235 222 * about a given object in the move callback. It relies on the fact that kmem
236 223 * guarantees that the object of the move callback has only been touched by the
237 224 * client itself or else by kmem. kmem does this by ensuring that none of the
238 225 * cache's slabs are freed to the virtual memory (VM) subsystem while a move
239 226 * callback is pending. When the last object on a slab is freed, if there is a
240 227 * pending move, kmem puts the slab on a per-cache dead list and defers freeing
241 228 * slabs on that list until all pending callbacks are completed. That way,
242 229 * clients can be certain that the object of a move callback is in one of the
243 230 * states listed above, making it possible to distinguish known objects (in
244 231 * state #4) using the two low order bits of any pointer member (with the
245 232 * exception of 'char *' or 'short *' which may not be 4-byte aligned on some
246 233 * platforms).
247 234 *
248 235 * The test works as long as the client always transitions objects from state #4
249 236 * (known, in use) to state #5 (about to be freed, invalid) by setting the low
250 237 * order bit of the client-designated pointer member. Since kmem only writes
251 238 * invalid memory patterns, such as 0xbaddcafe to uninitialized memory and
252 239 * 0xdeadbeef to freed memory, any scribbling on the object done by kmem is
253 240 * guaranteed to set at least one of the two low order bits. Therefore, given an
254 241 * object with a back pointer to a 'container_t *o_container', the client can
255 242 * test
256 243 *
257 244 * container_t *container = object->o_container;
258 245 * if ((uintptr_t)container & 0x3) {
259 246 * return (KMEM_CBRC_DONT_KNOW);
260 247 * }
261 248 *
262 249 * Typically, an object will have a pointer to some structure with a list or
263 250 * hash where objects from the cache are kept while in use. Assuming that the
264 251 * client has some way of knowing that the container structure is valid and will
265 252 * not go away during the move, and assuming that the structure includes a lock
266 253 * to protect whatever collection is used, then the client would continue as
267 254 * follows:
268 255 *
269 256 * // Ensure that the container structure does not go away.
270 257 * if (container_hold(container) == 0) {
271 258 * return (KMEM_CBRC_DONT_KNOW);
272 259 * }
273 260 * mutex_enter(&container->c_objects_lock);
274 261 * if (container != object->o_container) {
275 262 * mutex_exit(&container->c_objects_lock);
276 263 * container_rele(container);
277 264 * return (KMEM_CBRC_DONT_KNOW);
278 265 * }
279 266 *
280 267 * At this point the client knows that the object cannot be freed as long as
281 268 * c_objects_lock is held. Note that after acquiring the lock, the client must
282 269 * recheck the o_container pointer in case the object was removed just before
283 270 * acquiring the lock.
284 271 *
285 272 * When the client is about to free an object, it must first remove that object
286 273 * from the list, hash, or other structure where it is kept. At that time, to
287 274 * mark the object so it can be distinguished from the remaining, known objects,
288 275 * the client sets the designated low order bit:
289 276 *
290 277 * mutex_enter(&container->c_objects_lock);
|
↓ open down ↓ |
102 lines elided |
↑ open up ↑ |
291 278 * object->o_container = (void *)((uintptr_t)object->o_container | 0x1);
292 279 * list_remove(&container->c_objects, object);
293 280 * mutex_exit(&container->c_objects_lock);
294 281 *
295 282 * In the common case, the object is freed to the magazine layer, where it may
296 283 * be reused on a subsequent allocation without the overhead of calling the
297 284 * constructor. While in the magazine it appears allocated from the point of
298 285 * view of the slab layer, making it a candidate for the move callback. Most
299 286 * objects unrecognized by the client in the move callback fall into this
300 287 * category and are cheaply distinguished from known objects by the test
301 - * described earlier. Because searching magazines is prohibitively expensive
302 - * for kmem, clients that do not mark freed objects (and therefore return
303 - * KMEM_CBRC_DONT_KNOW for large numbers of objects) may find defragmentation
304 - * efficacy reduced.
288 + * described earlier. Since recognition is cheap for the client, and searching
289 + * magazines is expensive for kmem, kmem defers searching until the client first
290 + * returns KMEM_CBRC_DONT_KNOW. As long as the needed effort is reasonable, kmem
291 + * elsewhere does what it can to avoid bothering the client unnecessarily.
305 292 *
306 293 * Invalidating the designated pointer member before freeing the object marks
307 294 * the object to be avoided in the callback, and conversely, assigning a valid
308 295 * value to the designated pointer member after allocating the object makes the
309 296 * object fair game for the callback:
310 297 *
311 298 * ... allocate object ...
312 299 * ... set any initial state not set by the constructor ...
313 300 *
314 301 * mutex_enter(&container->c_objects_lock);
315 302 * list_insert_tail(&container->c_objects, object);
316 303 * membar_producer();
317 304 * object->o_container = container;
318 305 * mutex_exit(&container->c_objects_lock);
319 306 *
320 307 * Note that everything else must be valid before setting o_container makes the
321 308 * object fair game for the move callback. The membar_producer() call ensures
322 309 * that all the object's state is written to memory before setting the pointer
323 310 * that transitions the object from state #3 or #7 (allocated, constructed, not
324 311 * yet in use) to state #4 (in use, valid). That's important because the move
325 312 * function has to check the validity of the pointer before it can safely
326 313 * acquire the lock protecting the collection where it expects to find known
327 314 * objects.
328 315 *
329 316 * This method of distinguishing known objects observes the usual symmetry:
330 317 * invalidating the designated pointer is the first thing the client does before
331 318 * freeing the object, and setting the designated pointer is the last thing the
332 319 * client does after allocating the object. Of course, the client is not
333 320 * required to use this method. Fundamentally, how the client recognizes known
334 321 * objects is completely up to the client, but this method is recommended as an
335 322 * efficient and safe way to take advantage of the guarantees made by kmem. If
336 323 * the entire object is arbitrary data without any markable bits from a suitable
337 324 * pointer member, then the client must find some other method, such as
338 325 * searching a hash table of known objects.
339 326 *
340 327 * 2.5 Preventing Objects From Moving
341 328 *
342 329 * Besides a way to distinguish known objects, the other thing that the client
343 330 * needs is a strategy to ensure that an object will not move while the client
344 331 * is actively using it. The details of satisfying this requirement tend to be
345 332 * highly cache-specific. It might seem that the same rules that let a client
346 333 * remove an object safely should also decide when an object can be moved
347 334 * safely. However, any object state that makes a removal attempt invalid is
348 335 * likely to be long-lasting for objects that the client does not expect to
349 336 * remove. kmem knows nothing about the object state and is equally likely (from
350 337 * the client's point of view) to request a move for any object in the cache,
351 338 * whether prepared for removal or not. Even a low percentage of objects stuck
352 339 * in place by unremovability will defeat the consolidator if the stuck objects
353 340 * are the same long-lived allocations likely to hold slabs hostage.
354 341 * Fundamentally, the consolidator is not aimed at common cases. Severe external
355 342 * fragmentation is a worst case scenario manifested as sparsely allocated
356 343 * slabs, by definition a low percentage of the cache's objects. When deciding
357 344 * what makes an object movable, keep in mind the goal of the consolidator: to
358 345 * bring worst-case external fragmentation within the limits guaranteed for
359 346 * internal fragmentation. Removability is a poor criterion if it is likely to
360 347 * exclude more than an insignificant percentage of objects for long periods of
361 348 * time.
362 349 *
363 350 * A tricky general solution exists, and it has the advantage of letting you
364 351 * move any object at almost any moment, practically eliminating the likelihood
365 352 * that an object can hold a slab hostage. However, if there is a cache-specific
366 353 * way to ensure that an object is not actively in use in the vast majority of
367 354 * cases, a simpler solution that leverages this cache-specific knowledge is
368 355 * preferred.
369 356 *
370 357 * 2.5.1 Cache-Specific Solution
371 358 *
372 359 * As an example of a cache-specific solution, the ZFS znode cache takes
373 360 * advantage of the fact that the vast majority of znodes are only being
374 361 * referenced from the DNLC. (A typical case might be a few hundred in active
375 362 * use and a hundred thousand in the DNLC.) In the move callback, after the ZFS
376 363 * client has established that it recognizes the znode and can access its fields
377 364 * safely (using the method described earlier), it then tests whether the znode
378 365 * is referenced by anything other than the DNLC. If so, it assumes that the
379 366 * znode may be in active use and is unsafe to move, so it drops its locks and
380 367 * returns KMEM_CBRC_LATER. The advantage of this strategy is that everywhere
381 368 * else znodes are used, no change is needed to protect against the possibility
382 369 * of the znode moving. The disadvantage is that it remains possible for an
383 370 * application to hold a znode slab hostage with an open file descriptor.
384 371 * However, this case ought to be rare and the consolidator has a way to deal
385 372 * with it: If the client responds KMEM_CBRC_LATER repeatedly for the same
386 373 * object, kmem eventually stops believing it and treats the slab as if the
387 374 * client had responded KMEM_CBRC_NO. Having marked the hostage slab, kmem can
388 375 * then focus on getting it off of the partial slab list by allocating rather
389 376 * than freeing all of its objects. (Either way of getting a slab off the
390 377 * free list reduces fragmentation.)
391 378 *
392 379 * 2.5.2 General Solution
393 380 *
394 381 * The general solution, on the other hand, requires an explicit hold everywhere
395 382 * the object is used to prevent it from moving. To keep the client locking
396 383 * strategy as uncomplicated as possible, kmem guarantees the simplifying
397 384 * assumption that move callbacks are sequential, even across multiple caches.
398 385 * Internally, a global queue processed by a single thread supports all caches
399 386 * implementing the callback function. No matter how many caches supply a move
400 387 * function, the consolidator never moves more than one object at a time, so the
401 388 * client does not have to worry about tricky lock ordering involving several
402 389 * related objects from different kmem caches.
403 390 *
404 391 * The general solution implements the explicit hold as a read-write lock, which
405 392 * allows multiple readers to access an object from the cache simultaneously
406 393 * while a single writer is excluded from moving it. A single rwlock for the
407 394 * entire cache would lock out all threads from using any of the cache's objects
408 395 * even though only a single object is being moved, so to reduce contention,
409 396 * the client can fan out the single rwlock into an array of rwlocks hashed by
410 397 * the object address, making it probable that moving one object will not
411 398 * prevent other threads from using a different object. The rwlock cannot be a
412 399 * member of the object itself, because the possibility of the object moving
413 400 * makes it unsafe to access any of the object's fields until the lock is
414 401 * acquired.
415 402 *
416 403 * Assuming a small, fixed number of locks, it's possible that multiple objects
417 404 * will hash to the same lock. A thread that needs to use multiple objects in
418 405 * the same function may acquire the same lock multiple times. Since rwlocks are
419 406 * reentrant for readers, and since there is never more than a single writer at
420 407 * a time (assuming that the client acquires the lock as a writer only when
421 408 * moving an object inside the callback), there would seem to be no problem.
422 409 * However, a client locking multiple objects in the same function must handle
423 410 * one case of potential deadlock: Assume that thread A needs to prevent both
424 411 * object 1 and object 2 from moving, and thread B, the callback, meanwhile
425 412 * tries to move object 3. It's possible, if objects 1, 2, and 3 all hash to the
426 413 * same lock, that thread A will acquire the lock for object 1 as a reader
427 414 * before thread B sets the lock's write-wanted bit, preventing thread A from
428 415 * reacquiring the lock for object 2 as a reader. Unable to make forward
429 416 * progress, thread A will never release the lock for object 1, resulting in
430 417 * deadlock.
431 418 *
432 419 * There are two ways of avoiding the deadlock just described. The first is to
433 420 * use rw_tryenter() rather than rw_enter() in the callback function when
434 421 * attempting to acquire the lock as a writer. If tryenter discovers that the
435 422 * same object (or another object hashed to the same lock) is already in use, it
436 423 * aborts the callback and returns KMEM_CBRC_LATER. The second way is to use
437 424 * rprwlock_t (declared in common/fs/zfs/sys/rprwlock.h) instead of rwlock_t,
438 425 * since it allows a thread to acquire the lock as a reader in spite of a
439 426 * waiting writer. This second approach insists on moving the object now, no
440 427 * matter how many readers the move function must wait for in order to do so,
441 428 * and could delay the completion of the callback indefinitely (blocking
442 429 * callbacks to other clients). In practice, a less insistent callback using
443 430 * rw_tryenter() returns KMEM_CBRC_LATER infrequently enough that there seems
444 431 * little reason to use anything else.
445 432 *
446 433 * Avoiding deadlock is not the only problem that an implementation using an
447 434 * explicit hold needs to solve. Locking the object in the first place (to
448 435 * prevent it from moving) remains a problem, since the object could move
449 436 * between the time you obtain a pointer to the object and the time you acquire
450 437 * the rwlock hashed to that pointer value. Therefore the client needs to
451 438 * recheck the value of the pointer after acquiring the lock, drop the lock if
452 439 * the value has changed, and try again. This requires a level of indirection:
453 440 * something that points to the object rather than the object itself, that the
454 441 * client can access safely while attempting to acquire the lock. (The object
455 442 * itself cannot be referenced safely because it can move at any time.)
456 443 * The following lock-acquisition function takes whatever is safe to reference
457 444 * (arg), follows its pointer to the object (using function f), and tries as
458 445 * often as necessary to acquire the hashed lock and verify that the object
459 446 * still has not moved:
460 447 *
461 448 * object_t *
462 449 * object_hold(object_f f, void *arg)
463 450 * {
464 451 * object_t *op;
465 452 *
466 453 * op = f(arg);
467 454 * if (op == NULL) {
468 455 * return (NULL);
469 456 * }
470 457 *
471 458 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
472 459 * while (op != f(arg)) {
473 460 * rw_exit(OBJECT_RWLOCK(op));
474 461 * op = f(arg);
475 462 * if (op == NULL) {
476 463 * break;
477 464 * }
478 465 * rw_enter(OBJECT_RWLOCK(op), RW_READER);
479 466 * }
480 467 *
481 468 * return (op);
482 469 * }
483 470 *
484 471 * The OBJECT_RWLOCK macro hashes the object address to obtain the rwlock. The
485 472 * lock reacquisition loop, while necessary, almost never executes. The function
486 473 * pointer f (used to obtain the object pointer from arg) has the following type
487 474 * definition:
488 475 *
489 476 * typedef object_t *(*object_f)(void *arg);
490 477 *
491 478 * An object_f implementation is likely to be as simple as accessing a structure
492 479 * member:
493 480 *
494 481 * object_t *
495 482 * s_object(void *arg)
496 483 * {
497 484 * something_t *sp = arg;
498 485 * return (sp->s_object);
499 486 * }
500 487 *
501 488 * The flexibility of a function pointer allows the path to the object to be
502 489 * arbitrarily complex and also supports the notion that depending on where you
503 490 * are using the object, you may need to get it from someplace different.
504 491 *
505 492 * The function that releases the explicit hold is simpler because it does not
506 493 * have to worry about the object moving:
507 494 *
508 495 * void
509 496 * object_rele(object_t *op)
510 497 * {
511 498 * rw_exit(OBJECT_RWLOCK(op));
512 499 * }
513 500 *
514 501 * The caller is spared these details so that obtaining and releasing an
515 502 * explicit hold feels like a simple mutex_enter()/mutex_exit() pair. The caller
516 503 * of object_hold() only needs to know that the returned object pointer is valid
517 504 * if not NULL and that the object will not move until released.
518 505 *
519 506 * Although object_hold() prevents an object from moving, it does not prevent it
520 507 * from being freed. The caller must take measures before calling object_hold()
521 508 * (afterwards is too late) to ensure that the held object cannot be freed. The
522 509 * caller must do so without accessing the unsafe object reference, so any lock
523 510 * or reference count used to ensure the continued existence of the object must
524 511 * live outside the object itself.
525 512 *
526 513 * Obtaining a new object is a special case where an explicit hold is impossible
527 514 * for the caller. Any function that returns a newly allocated object (either as
528 515 * a return value, or as an in-out paramter) must return it already held; after
529 516 * the caller gets it is too late, since the object cannot be safely accessed
530 517 * without the level of indirection described earlier. The following
531 518 * object_alloc() example uses the same code shown earlier to transition a new
532 519 * object into the state of being recognized (by the client) as a known object.
533 520 * The function must acquire the hold (rw_enter) before that state transition
534 521 * makes the object movable:
535 522 *
536 523 * static object_t *
537 524 * object_alloc(container_t *container)
538 525 * {
539 526 * object_t *object = kmem_cache_alloc(object_cache, 0);
540 527 * ... set any initial state not set by the constructor ...
541 528 * rw_enter(OBJECT_RWLOCK(object), RW_READER);
542 529 * mutex_enter(&container->c_objects_lock);
543 530 * list_insert_tail(&container->c_objects, object);
544 531 * membar_producer();
545 532 * object->o_container = container;
546 533 * mutex_exit(&container->c_objects_lock);
547 534 * return (object);
548 535 * }
549 536 *
550 537 * Functions that implicitly acquire an object hold (any function that calls
551 538 * object_alloc() to supply an object for the caller) need to be carefully noted
552 539 * so that the matching object_rele() is not neglected. Otherwise, leaked holds
553 540 * prevent all objects hashed to the affected rwlocks from ever being moved.
554 541 *
555 542 * The pointer to a held object can be hashed to the holding rwlock even after
556 543 * the object has been freed. Although it is possible to release the hold
557 544 * after freeing the object, you may decide to release the hold implicitly in
558 545 * whatever function frees the object, so as to release the hold as soon as
559 546 * possible, and for the sake of symmetry with the function that implicitly
560 547 * acquires the hold when it allocates the object. Here, object_free() releases
561 548 * the hold acquired by object_alloc(). Its implicit object_rele() forms a
562 549 * matching pair with object_hold():
563 550 *
564 551 * void
565 552 * object_free(object_t *object)
566 553 * {
567 554 * container_t *container;
568 555 *
569 556 * ASSERT(object_held(object));
570 557 * container = object->o_container;
571 558 * mutex_enter(&container->c_objects_lock);
572 559 * object->o_container =
573 560 * (void *)((uintptr_t)object->o_container | 0x1);
574 561 * list_remove(&container->c_objects, object);
575 562 * mutex_exit(&container->c_objects_lock);
576 563 * object_rele(object);
577 564 * kmem_cache_free(object_cache, object);
578 565 * }
579 566 *
580 567 * Note that object_free() cannot safely accept an object pointer as an argument
581 568 * unless the object is already held. Any function that calls object_free()
582 569 * needs to be carefully noted since it similarly forms a matching pair with
583 570 * object_hold().
584 571 *
585 572 * To complete the picture, the following callback function implements the
586 573 * general solution by moving objects only if they are currently unheld:
587 574 *
588 575 * static kmem_cbrc_t
589 576 * object_move(void *buf, void *newbuf, size_t size, void *arg)
590 577 * {
591 578 * object_t *op = buf, *np = newbuf;
592 579 * container_t *container;
593 580 *
594 581 * container = op->o_container;
595 582 * if ((uintptr_t)container & 0x3) {
596 583 * return (KMEM_CBRC_DONT_KNOW);
597 584 * }
598 585 *
599 586 * // Ensure that the container structure does not go away.
600 587 * if (container_hold(container) == 0) {
601 588 * return (KMEM_CBRC_DONT_KNOW);
602 589 * }
603 590 *
604 591 * mutex_enter(&container->c_objects_lock);
605 592 * if (container != op->o_container) {
606 593 * mutex_exit(&container->c_objects_lock);
607 594 * container_rele(container);
608 595 * return (KMEM_CBRC_DONT_KNOW);
609 596 * }
610 597 *
611 598 * if (rw_tryenter(OBJECT_RWLOCK(op), RW_WRITER) == 0) {
612 599 * mutex_exit(&container->c_objects_lock);
613 600 * container_rele(container);
614 601 * return (KMEM_CBRC_LATER);
615 602 * }
616 603 *
617 604 * object_move_impl(op, np); // critical section
618 605 * rw_exit(OBJECT_RWLOCK(op));
619 606 *
620 607 * op->o_container = (void *)((uintptr_t)op->o_container | 0x1);
621 608 * list_link_replace(&op->o_link_node, &np->o_link_node);
622 609 * mutex_exit(&container->c_objects_lock);
623 610 * container_rele(container);
624 611 * return (KMEM_CBRC_YES);
625 612 * }
626 613 *
627 614 * Note that object_move() must invalidate the designated o_container pointer of
628 615 * the old object in the same way that object_free() does, since kmem will free
629 616 * the object in response to the KMEM_CBRC_YES return value.
630 617 *
631 618 * The lock order in object_move() differs from object_alloc(), which locks
632 619 * OBJECT_RWLOCK first and &container->c_objects_lock second, but as long as the
633 620 * callback uses rw_tryenter() (preventing the deadlock described earlier), it's
634 621 * not a problem. Holding the lock on the object list in the example above
635 622 * through the entire callback not only prevents the object from going away, it
636 623 * also allows you to lock the list elsewhere and know that none of its elements
637 624 * will move during iteration.
638 625 *
639 626 * Adding an explicit hold everywhere an object from the cache is used is tricky
640 627 * and involves much more change to client code than a cache-specific solution
641 628 * that leverages existing state to decide whether or not an object is
642 629 * movable. However, this approach has the advantage that no object remains
643 630 * immovable for any significant length of time, making it extremely unlikely
644 631 * that long-lived allocations can continue holding slabs hostage; and it works
645 632 * for any cache.
646 633 *
647 634 * 3. Consolidator Implementation
648 635 *
649 636 * Once the client supplies a move function that a) recognizes known objects and
650 637 * b) avoids moving objects that are actively in use, the remaining work is up
651 638 * to the consolidator to decide which objects to move and when to issue
652 639 * callbacks.
653 640 *
654 641 * The consolidator relies on the fact that a cache's slabs are ordered by
655 642 * usage. Each slab has a fixed number of objects. Depending on the slab's
656 643 * "color" (the offset of the first object from the beginning of the slab;
657 644 * offsets are staggered to mitigate false sharing of cache lines) it is either
658 645 * the maximum number of objects per slab determined at cache creation time or
659 646 * else the number closest to the maximum that fits within the space remaining
660 647 * after the initial offset. A completely allocated slab may contribute some
661 648 * internal fragmentation (per-slab overhead) but no external fragmentation, so
662 649 * it is of no interest to the consolidator. At the other extreme, slabs whose
663 650 * objects have all been freed to the slab are released to the virtual memory
664 651 * (VM) subsystem (objects freed to magazines are still allocated as far as the
665 652 * slab is concerned). External fragmentation exists when there are slabs
666 653 * somewhere between these extremes. A partial slab has at least one but not all
667 654 * of its objects allocated. The more partial slabs, and the fewer allocated
668 655 * objects on each of them, the higher the fragmentation. Hence the
669 656 * consolidator's overall strategy is to reduce the number of partial slabs by
670 657 * moving allocated objects from the least allocated slabs to the most allocated
671 658 * slabs.
672 659 *
673 660 * Partial slabs are kept in an AVL tree ordered by usage. Completely allocated
674 661 * slabs are kept separately in an unordered list. Since the majority of slabs
675 662 * tend to be completely allocated (a typical unfragmented cache may have
676 663 * thousands of complete slabs and only a single partial slab), separating
677 664 * complete slabs improves the efficiency of partial slab ordering, since the
678 665 * complete slabs do not affect the depth or balance of the AVL tree. This
679 666 * ordered sequence of partial slabs acts as a "free list" supplying objects for
680 667 * allocation requests.
681 668 *
682 669 * Objects are always allocated from the first partial slab in the free list,
683 670 * where the allocation is most likely to eliminate a partial slab (by
684 671 * completely allocating it). Conversely, when a single object from a completely
685 672 * allocated slab is freed to the slab, that slab is added to the front of the
686 673 * free list. Since most free list activity involves highly allocated slabs
687 674 * coming and going at the front of the list, slabs tend naturally toward the
688 675 * ideal order: highly allocated at the front, sparsely allocated at the back.
689 676 * Slabs with few allocated objects are likely to become completely free if they
690 677 * keep a safe distance away from the front of the free list. Slab misorders
691 678 * interfere with the natural tendency of slabs to become completely free or
692 679 * completely allocated. For example, a slab with a single allocated object
693 680 * needs only a single free to escape the cache; its natural desire is
694 681 * frustrated when it finds itself at the front of the list where a second
695 682 * allocation happens just before the free could have released it. Another slab
696 683 * with all but one object allocated might have supplied the buffer instead, so
697 684 * that both (as opposed to neither) of the slabs would have been taken off the
698 685 * free list.
699 686 *
700 687 * Although slabs tend naturally toward the ideal order, misorders allowed by a
701 688 * simple list implementation defeat the consolidator's strategy of merging
702 689 * least- and most-allocated slabs. Without an AVL tree to guarantee order, kmem
703 690 * needs another way to fix misorders to optimize its callback strategy. One
704 691 * approach is to periodically scan a limited number of slabs, advancing a
705 692 * marker to hold the current scan position, and to move extreme misorders to
706 693 * the front or back of the free list and to the front or back of the current
707 694 * scan range. By making consecutive scan ranges overlap by one slab, the least
708 695 * allocated slab in the current range can be carried along from the end of one
709 696 * scan to the start of the next.
710 697 *
711 698 * Maintaining partial slabs in an AVL tree relieves kmem of this additional
712 699 * task, however. Since most of the cache's activity is in the magazine layer,
713 700 * and allocations from the slab layer represent only a startup cost, the
714 701 * overhead of maintaining a balanced tree is not a significant concern compared
715 702 * to the opportunity of reducing complexity by eliminating the partial slab
716 703 * scanner just described. The overhead of an AVL tree is minimized by
717 704 * maintaining only partial slabs in the tree and keeping completely allocated
718 705 * slabs separately in a list. To avoid increasing the size of the slab
719 706 * structure the AVL linkage pointers are reused for the slab's list linkage,
720 707 * since the slab will always be either partial or complete, never stored both
721 708 * ways at the same time. To further minimize the overhead of the AVL tree the
722 709 * compare function that orders partial slabs by usage divides the range of
723 710 * allocated object counts into bins such that counts within the same bin are
724 711 * considered equal. Binning partial slabs makes it less likely that allocating
725 712 * or freeing a single object will change the slab's order, requiring a tree
726 713 * reinsertion (an avl_remove() followed by an avl_add(), both potentially
727 714 * requiring some rebalancing of the tree). Allocation counts closest to
728 715 * completely free and completely allocated are left unbinned (finely sorted) to
729 716 * better support the consolidator's strategy of merging slabs at either
730 717 * extreme.
731 718 *
732 719 * 3.1 Assessing Fragmentation and Selecting Candidate Slabs
733 720 *
734 721 * The consolidator piggybacks on the kmem maintenance thread and is called on
735 722 * the same interval as kmem_cache_update(), once per cache every fifteen
736 723 * seconds. kmem maintains a running count of unallocated objects in the slab
737 724 * layer (cache_bufslab). The consolidator checks whether that number exceeds
738 725 * 12.5% (1/8) of the total objects in the cache (cache_buftotal), and whether
739 726 * there is a significant number of slabs in the cache (arbitrarily a minimum
740 727 * 101 total slabs). Unused objects that have fallen out of the magazine layer's
741 728 * working set are included in the assessment, and magazines in the depot are
742 729 * reaped if those objects would lift cache_bufslab above the fragmentation
743 730 * threshold. Once the consolidator decides that a cache is fragmented, it looks
744 731 * for a candidate slab to reclaim, starting at the end of the partial slab free
745 732 * list and scanning backwards. At first the consolidator is choosy: only a slab
746 733 * with fewer than 12.5% (1/8) of its objects allocated qualifies (or else a
747 734 * single allocated object, regardless of percentage). If there is difficulty
748 735 * finding a candidate slab, kmem raises the allocation threshold incrementally,
749 736 * up to a maximum 87.5% (7/8), so that eventually the consolidator will reduce
750 737 * external fragmentation (unused objects on the free list) below 12.5% (1/8),
751 738 * even in the worst case of every slab in the cache being almost 7/8 allocated.
752 739 * The threshold can also be lowered incrementally when candidate slabs are easy
753 740 * to find, and the threshold is reset to the minimum 1/8 as soon as the cache
754 741 * is no longer fragmented.
755 742 *
756 743 * 3.2 Generating Callbacks
757 744 *
758 745 * Once an eligible slab is chosen, a callback is generated for every allocated
759 746 * object on the slab, in the hope that the client will move everything off the
760 747 * slab and make it reclaimable. Objects selected as move destinations are
761 748 * chosen from slabs at the front of the free list. Assuming slabs in the ideal
762 749 * order (most allocated at the front, least allocated at the back) and a
763 750 * cooperative client, the consolidator will succeed in removing slabs from both
764 751 * ends of the free list, completely allocating on the one hand and completely
765 752 * freeing on the other. Objects selected as move destinations are allocated in
766 753 * the kmem maintenance thread where move requests are enqueued. A separate
767 754 * callback thread removes pending callbacks from the queue and calls the
768 755 * client. The separate thread ensures that client code (the move function) does
769 756 * not interfere with internal kmem maintenance tasks. A map of pending
770 757 * callbacks keyed by object address (the object to be moved) is checked to
771 758 * ensure that duplicate callbacks are not generated for the same object.
772 759 * Allocating the move destination (the object to move to) prevents subsequent
773 760 * callbacks from selecting the same destination as an earlier pending callback.
774 761 *
775 762 * Move requests can also be generated by kmem_cache_reap() when the system is
776 763 * desperate for memory and by kmem_cache_move_notify(), called by the client to
777 764 * notify kmem that a move refused earlier with KMEM_CBRC_LATER is now possible.
778 765 * The map of pending callbacks is protected by the same lock that protects the
779 766 * slab layer.
780 767 *
781 768 * When the system is desperate for memory, kmem does not bother to determine
782 769 * whether or not the cache exceeds the fragmentation threshold, but tries to
783 770 * consolidate as many slabs as possible. Normally, the consolidator chews
784 771 * slowly, one sparsely allocated slab at a time during each maintenance
785 772 * interval that the cache is fragmented. When desperate, the consolidator
786 773 * starts at the last partial slab and enqueues callbacks for every allocated
787 774 * object on every partial slab, working backwards until it reaches the first
788 775 * partial slab. The first partial slab, meanwhile, advances in pace with the
789 776 * consolidator as allocations to supply move destinations for the enqueued
790 777 * callbacks use up the highly allocated slabs at the front of the free list.
791 778 * Ideally, the overgrown free list collapses like an accordion, starting at
792 779 * both ends and ending at the center with a single partial slab.
793 780 *
794 781 * 3.3 Client Responses
795 782 *
796 783 * When the client returns KMEM_CBRC_NO in response to the move callback, kmem
797 784 * marks the slab that supplied the stuck object non-reclaimable and moves it to
798 785 * front of the free list. The slab remains marked as long as it remains on the
799 786 * free list, and it appears more allocated to the partial slab compare function
800 787 * than any unmarked slab, no matter how many of its objects are allocated.
801 788 * Since even one immovable object ties up the entire slab, the goal is to
802 789 * completely allocate any slab that cannot be completely freed. kmem does not
803 790 * bother generating callbacks to move objects from a marked slab unless the
804 791 * system is desperate.
805 792 *
806 793 * When the client responds KMEM_CBRC_LATER, kmem increments a count for the
807 794 * slab. If the client responds LATER too many times, kmem disbelieves and
808 795 * treats the response as a NO. The count is cleared when the slab is taken off
809 796 * the partial slab list or when the client moves one of the slab's objects.
810 797 *
811 798 * 4. Observability
812 799 *
813 800 * A kmem cache's external fragmentation is best observed with 'mdb -k' using
814 801 * the ::kmem_slabs dcmd. For a complete description of the command, enter
815 802 * '::help kmem_slabs' at the mdb prompt.
816 803 */
817 804
818 805 #include <sys/kmem_impl.h>
819 806 #include <sys/vmem_impl.h>
820 807 #include <sys/param.h>
821 808 #include <sys/sysmacros.h>
822 809 #include <sys/vm.h>
823 810 #include <sys/proc.h>
824 811 #include <sys/tuneable.h>
825 812 #include <sys/systm.h>
826 813 #include <sys/cmn_err.h>
827 814 #include <sys/debug.h>
828 815 #include <sys/sdt.h>
829 816 #include <sys/mutex.h>
830 817 #include <sys/bitmap.h>
831 818 #include <sys/atomic.h>
832 819 #include <sys/kobj.h>
833 820 #include <sys/disp.h>
834 821 #include <vm/seg_kmem.h>
835 822 #include <sys/log.h>
836 823 #include <sys/callb.h>
837 824 #include <sys/taskq.h>
838 825 #include <sys/modctl.h>
839 826 #include <sys/reboot.h>
840 827 #include <sys/id32.h>
841 828 #include <sys/zone.h>
842 829 #include <sys/netstack.h>
843 830 #ifdef DEBUG
844 831 #include <sys/random.h>
845 832 #endif
846 833
847 834 extern void streams_msg_init(void);
848 835 extern int segkp_fromheap;
849 836 extern void segkp_cache_free(void);
850 837 extern int callout_init_done;
851 838
852 839 struct kmem_cache_kstat {
853 840 kstat_named_t kmc_buf_size;
854 841 kstat_named_t kmc_align;
855 842 kstat_named_t kmc_chunk_size;
856 843 kstat_named_t kmc_slab_size;
857 844 kstat_named_t kmc_alloc;
858 845 kstat_named_t kmc_alloc_fail;
859 846 kstat_named_t kmc_free;
860 847 kstat_named_t kmc_depot_alloc;
861 848 kstat_named_t kmc_depot_free;
862 849 kstat_named_t kmc_depot_contention;
863 850 kstat_named_t kmc_slab_alloc;
864 851 kstat_named_t kmc_slab_free;
865 852 kstat_named_t kmc_buf_constructed;
866 853 kstat_named_t kmc_buf_avail;
867 854 kstat_named_t kmc_buf_inuse;
868 855 kstat_named_t kmc_buf_total;
869 856 kstat_named_t kmc_buf_max;
870 857 kstat_named_t kmc_slab_create;
871 858 kstat_named_t kmc_slab_destroy;
872 859 kstat_named_t kmc_vmem_source;
873 860 kstat_named_t kmc_hash_size;
874 861 kstat_named_t kmc_hash_lookup_depth;
875 862 kstat_named_t kmc_hash_rescale;
876 863 kstat_named_t kmc_full_magazines;
877 864 kstat_named_t kmc_empty_magazines;
878 865 kstat_named_t kmc_magazine_size;
879 866 kstat_named_t kmc_reap; /* number of kmem_cache_reap() calls */
880 867 kstat_named_t kmc_defrag; /* attempts to defrag all partial slabs */
881 868 kstat_named_t kmc_scan; /* attempts to defrag one partial slab */
882 869 kstat_named_t kmc_move_callbacks; /* sum of yes, no, later, dn, dk */
883 870 kstat_named_t kmc_move_yes;
884 871 kstat_named_t kmc_move_no;
885 872 kstat_named_t kmc_move_later;
886 873 kstat_named_t kmc_move_dont_need;
887 874 kstat_named_t kmc_move_dont_know; /* obj unrecognized by client ... */
888 875 kstat_named_t kmc_move_hunt_found; /* ... but found in mag layer */
889 876 kstat_named_t kmc_move_slabs_freed; /* slabs freed by consolidator */
890 877 kstat_named_t kmc_move_reclaimable; /* buffers, if consolidator ran */
891 878 } kmem_cache_kstat = {
892 879 { "buf_size", KSTAT_DATA_UINT64 },
893 880 { "align", KSTAT_DATA_UINT64 },
894 881 { "chunk_size", KSTAT_DATA_UINT64 },
895 882 { "slab_size", KSTAT_DATA_UINT64 },
896 883 { "alloc", KSTAT_DATA_UINT64 },
897 884 { "alloc_fail", KSTAT_DATA_UINT64 },
898 885 { "free", KSTAT_DATA_UINT64 },
899 886 { "depot_alloc", KSTAT_DATA_UINT64 },
900 887 { "depot_free", KSTAT_DATA_UINT64 },
901 888 { "depot_contention", KSTAT_DATA_UINT64 },
902 889 { "slab_alloc", KSTAT_DATA_UINT64 },
903 890 { "slab_free", KSTAT_DATA_UINT64 },
904 891 { "buf_constructed", KSTAT_DATA_UINT64 },
905 892 { "buf_avail", KSTAT_DATA_UINT64 },
906 893 { "buf_inuse", KSTAT_DATA_UINT64 },
907 894 { "buf_total", KSTAT_DATA_UINT64 },
908 895 { "buf_max", KSTAT_DATA_UINT64 },
909 896 { "slab_create", KSTAT_DATA_UINT64 },
910 897 { "slab_destroy", KSTAT_DATA_UINT64 },
911 898 { "vmem_source", KSTAT_DATA_UINT64 },
912 899 { "hash_size", KSTAT_DATA_UINT64 },
913 900 { "hash_lookup_depth", KSTAT_DATA_UINT64 },
914 901 { "hash_rescale", KSTAT_DATA_UINT64 },
915 902 { "full_magazines", KSTAT_DATA_UINT64 },
916 903 { "empty_magazines", KSTAT_DATA_UINT64 },
917 904 { "magazine_size", KSTAT_DATA_UINT64 },
918 905 { "reap", KSTAT_DATA_UINT64 },
919 906 { "defrag", KSTAT_DATA_UINT64 },
920 907 { "scan", KSTAT_DATA_UINT64 },
921 908 { "move_callbacks", KSTAT_DATA_UINT64 },
922 909 { "move_yes", KSTAT_DATA_UINT64 },
923 910 { "move_no", KSTAT_DATA_UINT64 },
924 911 { "move_later", KSTAT_DATA_UINT64 },
925 912 { "move_dont_need", KSTAT_DATA_UINT64 },
926 913 { "move_dont_know", KSTAT_DATA_UINT64 },
927 914 { "move_hunt_found", KSTAT_DATA_UINT64 },
928 915 { "move_slabs_freed", KSTAT_DATA_UINT64 },
929 916 { "move_reclaimable", KSTAT_DATA_UINT64 },
930 917 };
931 918
932 919 static kmutex_t kmem_cache_kstat_lock;
933 920
934 921 /*
935 922 * The default set of caches to back kmem_alloc().
936 923 * These sizes should be reevaluated periodically.
937 924 *
938 925 * We want allocations that are multiples of the coherency granularity
939 926 * (64 bytes) to be satisfied from a cache which is a multiple of 64
940 927 * bytes, so that it will be 64-byte aligned. For all multiples of 64,
941 928 * the next kmem_cache_size greater than or equal to it must be a
942 929 * multiple of 64.
943 930 *
944 931 * We split the table into two sections: size <= 4k and size > 4k. This
945 932 * saves a lot of space and cache footprint in our cache tables.
946 933 */
947 934 static const int kmem_alloc_sizes[] = {
948 935 1 * 8,
949 936 2 * 8,
950 937 3 * 8,
951 938 4 * 8, 5 * 8, 6 * 8, 7 * 8,
952 939 4 * 16, 5 * 16, 6 * 16, 7 * 16,
953 940 4 * 32, 5 * 32, 6 * 32, 7 * 32,
954 941 4 * 64, 5 * 64, 6 * 64, 7 * 64,
955 942 4 * 128, 5 * 128, 6 * 128, 7 * 128,
956 943 P2ALIGN(8192 / 7, 64),
957 944 P2ALIGN(8192 / 6, 64),
958 945 P2ALIGN(8192 / 5, 64),
959 946 P2ALIGN(8192 / 4, 64),
960 947 P2ALIGN(8192 / 3, 64),
961 948 P2ALIGN(8192 / 2, 64),
962 949 };
963 950
964 951 static const int kmem_big_alloc_sizes[] = {
965 952 2 * 4096, 3 * 4096,
966 953 2 * 8192, 3 * 8192,
967 954 4 * 8192, 5 * 8192, 6 * 8192, 7 * 8192,
968 955 8 * 8192, 9 * 8192, 10 * 8192, 11 * 8192,
969 956 12 * 8192, 13 * 8192, 14 * 8192, 15 * 8192,
970 957 16 * 8192
971 958 };
972 959
973 960 #define KMEM_MAXBUF 4096
974 961 #define KMEM_BIG_MAXBUF_32BIT 32768
975 962 #define KMEM_BIG_MAXBUF 131072
976 963
977 964 #define KMEM_BIG_MULTIPLE 4096 /* big_alloc_sizes must be a multiple */
978 965 #define KMEM_BIG_SHIFT 12 /* lg(KMEM_BIG_MULTIPLE) */
979 966
980 967 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
981 968 static kmem_cache_t *kmem_big_alloc_table[KMEM_BIG_MAXBUF >> KMEM_BIG_SHIFT];
982 969
983 970 #define KMEM_ALLOC_TABLE_MAX (KMEM_MAXBUF >> KMEM_ALIGN_SHIFT)
984 971 static size_t kmem_big_alloc_table_max = 0; /* # of filled elements */
985 972
986 973 static kmem_magtype_t kmem_magtype[] = {
987 974 { 1, 8, 3200, 65536 },
988 975 { 3, 16, 256, 32768 },
989 976 { 7, 32, 64, 16384 },
990 977 { 15, 64, 0, 8192 },
991 978 { 31, 64, 0, 4096 },
992 979 { 47, 64, 0, 2048 },
993 980 { 63, 64, 0, 1024 },
994 981 { 95, 64, 0, 512 },
995 982 { 143, 64, 0, 0 },
996 983 };
997 984
998 985 static uint32_t kmem_reaping;
999 986 static uint32_t kmem_reaping_idspace;
1000 987
1001 988 /*
1002 989 * kmem tunables
1003 990 */
|
↓ open down ↓ |
689 lines elided |
↑ open up ↑ |
1004 991 clock_t kmem_reap_interval; /* cache reaping rate [15 * HZ ticks] */
1005 992 int kmem_depot_contention = 3; /* max failed tryenters per real interval */
1006 993 pgcnt_t kmem_reapahead = 0; /* start reaping N pages before pageout */
1007 994 int kmem_panic = 1; /* whether to panic on error */
1008 995 int kmem_logging = 1; /* kmem_log_enter() override */
1009 996 uint32_t kmem_mtbf = 0; /* mean time between failures [default: off] */
1010 997 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
1011 998 size_t kmem_content_log_size; /* content log size [2% of memory] */
1012 999 size_t kmem_failure_log_size; /* failure log [4 pages per CPU] */
1013 1000 size_t kmem_slab_log_size; /* slab create log [4 pages per CPU] */
1014 -size_t kmem_zerosized_log_size; /* zero-sized log [4 pages per CPU] */
1015 1001 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
1016 1002 size_t kmem_lite_minsize = 0; /* minimum buffer size for KMF_LITE */
1017 1003 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
1018 1004 int kmem_lite_pcs = 4; /* number of PCs to store in KMF_LITE mode */
1019 1005 size_t kmem_maxverify; /* maximum bytes to inspect in debug routines */
1020 1006 size_t kmem_minfirewall; /* hardware-enforced redzone threshold */
1021 1007
1022 -#ifdef DEBUG
1023 -int kmem_warn_zerosized = 1; /* whether to warn on zero-sized KM_SLEEP */
1024 -#else
1025 -int kmem_warn_zerosized = 0; /* whether to warn on zero-sized KM_SLEEP */
1026 -#endif
1027 -
1028 -int kmem_panic_zerosized = 0; /* whether to panic on zero-sized KM_SLEEP */
1029 -
1030 1008 #ifdef _LP64
1031 1009 size_t kmem_max_cached = KMEM_BIG_MAXBUF; /* maximum kmem_alloc cache */
1032 1010 #else
1033 1011 size_t kmem_max_cached = KMEM_BIG_MAXBUF_32BIT; /* maximum kmem_alloc cache */
1034 1012 #endif
1035 1013
1036 1014 #ifdef DEBUG
1037 1015 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
1038 1016 #else
1039 1017 int kmem_flags = 0;
1040 1018 #endif
1041 1019 int kmem_ready;
1042 1020
1043 1021 static kmem_cache_t *kmem_slab_cache;
1044 1022 static kmem_cache_t *kmem_bufctl_cache;
1045 1023 static kmem_cache_t *kmem_bufctl_audit_cache;
1046 1024
1047 1025 static kmutex_t kmem_cache_lock; /* inter-cache linkage only */
1048 1026 static list_t kmem_caches;
1049 1027
1050 1028 static taskq_t *kmem_taskq;
1051 1029 static kmutex_t kmem_flags_lock;
1052 1030 static vmem_t *kmem_metadata_arena;
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
1053 1031 static vmem_t *kmem_msb_arena; /* arena for metadata caches */
1054 1032 static vmem_t *kmem_cache_arena;
1055 1033 static vmem_t *kmem_hash_arena;
1056 1034 static vmem_t *kmem_log_arena;
1057 1035 static vmem_t *kmem_oversize_arena;
1058 1036 static vmem_t *kmem_va_arena;
1059 1037 static vmem_t *kmem_default_arena;
1060 1038 static vmem_t *kmem_firewall_va_arena;
1061 1039 static vmem_t *kmem_firewall_arena;
1062 1040
1063 -static int kmem_zerosized; /* # of zero-sized allocs */
1041 +/*
1042 + * Define KMEM_STATS to turn on statistic gathering. By default, it is only
1043 + * turned on when DEBUG is also defined.
1044 + */
1045 +#ifdef DEBUG
1046 +#define KMEM_STATS
1047 +#endif /* DEBUG */
1064 1048
1049 +#ifdef KMEM_STATS
1050 +#define KMEM_STAT_ADD(stat) ((stat)++)
1051 +#define KMEM_STAT_COND_ADD(cond, stat) ((void) (!(cond) || (stat)++))
1052 +#else
1053 +#define KMEM_STAT_ADD(stat) /* nothing */
1054 +#define KMEM_STAT_COND_ADD(cond, stat) /* nothing */
1055 +#endif /* KMEM_STATS */
1056 +
1065 1057 /*
1066 1058 * kmem slab consolidator thresholds (tunables)
1067 1059 */
1068 1060 size_t kmem_frag_minslabs = 101; /* minimum total slabs */
1069 1061 size_t kmem_frag_numer = 1; /* free buffers (numerator) */
1070 1062 size_t kmem_frag_denom = KMEM_VOID_FRACTION; /* buffers (denominator) */
1071 1063 /*
1072 1064 * Maximum number of slabs from which to move buffers during a single
1073 1065 * maintenance interval while the system is not low on memory.
1074 1066 */
1075 1067 size_t kmem_reclaim_max_slabs = 1;
1076 1068 /*
1077 1069 * Number of slabs to scan backwards from the end of the partial slab list
1078 1070 * when searching for buffers to relocate.
1079 1071 */
1080 1072 size_t kmem_reclaim_scan_range = 12;
1081 1073
1074 +#ifdef KMEM_STATS
1075 +static struct {
1076 + uint64_t kms_callbacks;
1077 + uint64_t kms_yes;
1078 + uint64_t kms_no;
1079 + uint64_t kms_later;
1080 + uint64_t kms_dont_need;
1081 + uint64_t kms_dont_know;
1082 + uint64_t kms_hunt_found_mag;
1083 + uint64_t kms_hunt_found_slab;
1084 + uint64_t kms_hunt_alloc_fail;
1085 + uint64_t kms_hunt_lucky;
1086 + uint64_t kms_notify;
1087 + uint64_t kms_notify_callbacks;
1088 + uint64_t kms_disbelief;
1089 + uint64_t kms_already_pending;
1090 + uint64_t kms_callback_alloc_fail;
1091 + uint64_t kms_callback_taskq_fail;
1092 + uint64_t kms_endscan_slab_dead;
1093 + uint64_t kms_endscan_slab_destroyed;
1094 + uint64_t kms_endscan_nomem;
1095 + uint64_t kms_endscan_refcnt_changed;
1096 + uint64_t kms_endscan_nomove_changed;
1097 + uint64_t kms_endscan_freelist;
1098 + uint64_t kms_avl_update;
1099 + uint64_t kms_avl_noupdate;
1100 + uint64_t kms_no_longer_reclaimable;
1101 + uint64_t kms_notify_no_longer_reclaimable;
1102 + uint64_t kms_notify_slab_dead;
1103 + uint64_t kms_notify_slab_destroyed;
1104 + uint64_t kms_alloc_fail;
1105 + uint64_t kms_constructor_fail;
1106 + uint64_t kms_dead_slabs_freed;
1107 + uint64_t kms_defrags;
1108 + uint64_t kms_scans;
1109 + uint64_t kms_scan_depot_ws_reaps;
1110 + uint64_t kms_debug_reaps;
1111 + uint64_t kms_debug_scans;
1112 +} kmem_move_stats;
1113 +#endif /* KMEM_STATS */
1114 +
1082 1115 /* consolidator knobs */
1083 1116 static boolean_t kmem_move_noreap;
1084 1117 static boolean_t kmem_move_blocked;
1085 1118 static boolean_t kmem_move_fulltilt;
1086 1119 static boolean_t kmem_move_any_partial;
1087 1120
1088 1121 #ifdef DEBUG
1089 1122 /*
1090 1123 * kmem consolidator debug tunables:
1091 1124 * Ensure code coverage by occasionally running the consolidator even when the
1092 1125 * caches are not fragmented (they may never be). These intervals are mean time
1093 1126 * in cache maintenance intervals (kmem_cache_update).
1094 1127 */
1095 1128 uint32_t kmem_mtb_move = 60; /* defrag 1 slab (~15min) */
1096 1129 uint32_t kmem_mtb_reap = 1800; /* defrag all slabs (~7.5hrs) */
1097 1130 #endif /* DEBUG */
1098 1131
1099 1132 static kmem_cache_t *kmem_defrag_cache;
1100 1133 static kmem_cache_t *kmem_move_cache;
1101 1134 static taskq_t *kmem_move_taskq;
|
↓ open down ↓ |
10 lines elided |
↑ open up ↑ |
1102 1135
1103 1136 static void kmem_cache_scan(kmem_cache_t *);
1104 1137 static void kmem_cache_defrag(kmem_cache_t *);
1105 1138 static void kmem_slab_prefill(kmem_cache_t *, kmem_slab_t *);
1106 1139
1107 1140
1108 1141 kmem_log_header_t *kmem_transaction_log;
1109 1142 kmem_log_header_t *kmem_content_log;
1110 1143 kmem_log_header_t *kmem_failure_log;
1111 1144 kmem_log_header_t *kmem_slab_log;
1112 -kmem_log_header_t *kmem_zerosized_log;
1113 1145
1114 1146 static int kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
1115 1147
1116 1148 #define KMEM_BUFTAG_LITE_ENTER(bt, count, caller) \
1117 1149 if ((count) > 0) { \
1118 1150 pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history; \
1119 1151 pc_t *_e; \
1120 1152 /* memmove() the old entries down one notch */ \
1121 1153 for (_e = &_s[(count) - 1]; _e > _s; _e--) \
1122 1154 *_e = *(_e - 1); \
1123 1155 *_s = (uintptr_t)(caller); \
1124 1156 }
1125 1157
1126 1158 #define KMERR_MODIFIED 0 /* buffer modified while on freelist */
1127 1159 #define KMERR_REDZONE 1 /* redzone violation (write past end of buf) */
1128 1160 #define KMERR_DUPFREE 2 /* freed a buffer twice */
1129 1161 #define KMERR_BADADDR 3 /* freed a bad (unallocated) address */
1130 1162 #define KMERR_BADBUFTAG 4 /* buftag corrupted */
1131 1163 #define KMERR_BADBUFCTL 5 /* bufctl corrupted */
1132 1164 #define KMERR_BADCACHE 6 /* freed a buffer to the wrong cache */
1133 1165 #define KMERR_BADSIZE 7 /* alloc size != free size */
1134 1166 #define KMERR_BADBASE 8 /* buffer base address wrong */
1135 1167
1136 1168 struct {
1137 1169 hrtime_t kmp_timestamp; /* timestamp of panic */
1138 1170 int kmp_error; /* type of kmem error */
1139 1171 void *kmp_buffer; /* buffer that induced panic */
1140 1172 void *kmp_realbuf; /* real start address for buffer */
1141 1173 kmem_cache_t *kmp_cache; /* buffer's cache according to client */
1142 1174 kmem_cache_t *kmp_realcache; /* actual cache containing buffer */
1143 1175 kmem_slab_t *kmp_slab; /* slab accoring to kmem_findslab() */
1144 1176 kmem_bufctl_t *kmp_bufctl; /* bufctl */
1145 1177 } kmem_panic_info;
1146 1178
1147 1179
1148 1180 static void
1149 1181 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
1150 1182 {
1151 1183 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1152 1184 uint64_t *buf = buf_arg;
1153 1185
1154 1186 while (buf < bufend)
1155 1187 *buf++ = pattern;
1156 1188 }
1157 1189
1158 1190 static void *
1159 1191 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
1160 1192 {
1161 1193 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1162 1194 uint64_t *buf;
1163 1195
1164 1196 for (buf = buf_arg; buf < bufend; buf++)
1165 1197 if (*buf != pattern)
1166 1198 return (buf);
1167 1199 return (NULL);
1168 1200 }
1169 1201
1170 1202 static void *
1171 1203 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
1172 1204 {
1173 1205 uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
1174 1206 uint64_t *buf;
1175 1207
1176 1208 for (buf = buf_arg; buf < bufend; buf++) {
1177 1209 if (*buf != old) {
1178 1210 copy_pattern(old, buf_arg,
1179 1211 (char *)buf - (char *)buf_arg);
1180 1212 return (buf);
1181 1213 }
1182 1214 *buf = new;
1183 1215 }
1184 1216
1185 1217 return (NULL);
1186 1218 }
1187 1219
1188 1220 static void
1189 1221 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1190 1222 {
1191 1223 kmem_cache_t *cp;
1192 1224
1193 1225 mutex_enter(&kmem_cache_lock);
1194 1226 for (cp = list_head(&kmem_caches); cp != NULL;
1195 1227 cp = list_next(&kmem_caches, cp))
1196 1228 if (tq != NULL)
1197 1229 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1198 1230 tqflag);
1199 1231 else
1200 1232 func(cp);
1201 1233 mutex_exit(&kmem_cache_lock);
1202 1234 }
1203 1235
1204 1236 static void
1205 1237 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
1206 1238 {
1207 1239 kmem_cache_t *cp;
1208 1240
1209 1241 mutex_enter(&kmem_cache_lock);
1210 1242 for (cp = list_head(&kmem_caches); cp != NULL;
1211 1243 cp = list_next(&kmem_caches, cp)) {
1212 1244 if (!(cp->cache_cflags & KMC_IDENTIFIER))
1213 1245 continue;
1214 1246 if (tq != NULL)
1215 1247 (void) taskq_dispatch(tq, (task_func_t *)func, cp,
1216 1248 tqflag);
1217 1249 else
1218 1250 func(cp);
1219 1251 }
1220 1252 mutex_exit(&kmem_cache_lock);
1221 1253 }
1222 1254
1223 1255 /*
1224 1256 * Debugging support. Given a buffer address, find its slab.
1225 1257 */
1226 1258 static kmem_slab_t *
1227 1259 kmem_findslab(kmem_cache_t *cp, void *buf)
1228 1260 {
1229 1261 kmem_slab_t *sp;
1230 1262
1231 1263 mutex_enter(&cp->cache_lock);
1232 1264 for (sp = list_head(&cp->cache_complete_slabs); sp != NULL;
1233 1265 sp = list_next(&cp->cache_complete_slabs, sp)) {
1234 1266 if (KMEM_SLAB_MEMBER(sp, buf)) {
1235 1267 mutex_exit(&cp->cache_lock);
1236 1268 return (sp);
1237 1269 }
1238 1270 }
1239 1271 for (sp = avl_first(&cp->cache_partial_slabs); sp != NULL;
1240 1272 sp = AVL_NEXT(&cp->cache_partial_slabs, sp)) {
1241 1273 if (KMEM_SLAB_MEMBER(sp, buf)) {
1242 1274 mutex_exit(&cp->cache_lock);
1243 1275 return (sp);
1244 1276 }
1245 1277 }
1246 1278 mutex_exit(&cp->cache_lock);
1247 1279
1248 1280 return (NULL);
1249 1281 }
1250 1282
1251 1283 static void
1252 1284 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
1253 1285 {
1254 1286 kmem_buftag_t *btp = NULL;
1255 1287 kmem_bufctl_t *bcp = NULL;
1256 1288 kmem_cache_t *cp = cparg;
1257 1289 kmem_slab_t *sp;
1258 1290 uint64_t *off;
1259 1291 void *buf = bufarg;
1260 1292
1261 1293 kmem_logging = 0; /* stop logging when a bad thing happens */
1262 1294
1263 1295 kmem_panic_info.kmp_timestamp = gethrtime();
1264 1296
1265 1297 sp = kmem_findslab(cp, buf);
1266 1298 if (sp == NULL) {
1267 1299 for (cp = list_tail(&kmem_caches); cp != NULL;
1268 1300 cp = list_prev(&kmem_caches, cp)) {
1269 1301 if ((sp = kmem_findslab(cp, buf)) != NULL)
1270 1302 break;
1271 1303 }
1272 1304 }
1273 1305
1274 1306 if (sp == NULL) {
1275 1307 cp = NULL;
1276 1308 error = KMERR_BADADDR;
1277 1309 } else {
1278 1310 if (cp != cparg)
1279 1311 error = KMERR_BADCACHE;
1280 1312 else
1281 1313 buf = (char *)bufarg - ((uintptr_t)bufarg -
1282 1314 (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1283 1315 if (buf != bufarg)
1284 1316 error = KMERR_BADBASE;
1285 1317 if (cp->cache_flags & KMF_BUFTAG)
1286 1318 btp = KMEM_BUFTAG(cp, buf);
1287 1319 if (cp->cache_flags & KMF_HASH) {
1288 1320 mutex_enter(&cp->cache_lock);
1289 1321 for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1290 1322 if (bcp->bc_addr == buf)
1291 1323 break;
1292 1324 mutex_exit(&cp->cache_lock);
1293 1325 if (bcp == NULL && btp != NULL)
1294 1326 bcp = btp->bt_bufctl;
1295 1327 if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
1296 1328 NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
1297 1329 bcp->bc_addr != buf) {
1298 1330 error = KMERR_BADBUFCTL;
1299 1331 bcp = NULL;
1300 1332 }
1301 1333 }
1302 1334 }
1303 1335
1304 1336 kmem_panic_info.kmp_error = error;
1305 1337 kmem_panic_info.kmp_buffer = bufarg;
1306 1338 kmem_panic_info.kmp_realbuf = buf;
1307 1339 kmem_panic_info.kmp_cache = cparg;
1308 1340 kmem_panic_info.kmp_realcache = cp;
1309 1341 kmem_panic_info.kmp_slab = sp;
1310 1342 kmem_panic_info.kmp_bufctl = bcp;
1311 1343
1312 1344 printf("kernel memory allocator: ");
1313 1345
1314 1346 switch (error) {
1315 1347
1316 1348 case KMERR_MODIFIED:
1317 1349 printf("buffer modified after being freed\n");
1318 1350 off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1319 1351 if (off == NULL) /* shouldn't happen */
1320 1352 off = buf;
1321 1353 printf("modification occurred at offset 0x%lx "
1322 1354 "(0x%llx replaced by 0x%llx)\n",
1323 1355 (uintptr_t)off - (uintptr_t)buf,
1324 1356 (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
1325 1357 break;
1326 1358
1327 1359 case KMERR_REDZONE:
1328 1360 printf("redzone violation: write past end of buffer\n");
1329 1361 break;
1330 1362
1331 1363 case KMERR_BADADDR:
1332 1364 printf("invalid free: buffer not in cache\n");
1333 1365 break;
1334 1366
1335 1367 case KMERR_DUPFREE:
1336 1368 printf("duplicate free: buffer freed twice\n");
1337 1369 break;
1338 1370
1339 1371 case KMERR_BADBUFTAG:
1340 1372 printf("boundary tag corrupted\n");
1341 1373 printf("bcp ^ bxstat = %lx, should be %lx\n",
1342 1374 (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1343 1375 KMEM_BUFTAG_FREE);
1344 1376 break;
1345 1377
1346 1378 case KMERR_BADBUFCTL:
1347 1379 printf("bufctl corrupted\n");
1348 1380 break;
1349 1381
1350 1382 case KMERR_BADCACHE:
1351 1383 printf("buffer freed to wrong cache\n");
1352 1384 printf("buffer was allocated from %s,\n", cp->cache_name);
1353 1385 printf("caller attempting free to %s.\n", cparg->cache_name);
1354 1386 break;
1355 1387
1356 1388 case KMERR_BADSIZE:
1357 1389 printf("bad free: free size (%u) != alloc size (%u)\n",
1358 1390 KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1359 1391 KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1360 1392 break;
1361 1393
1362 1394 case KMERR_BADBASE:
1363 1395 printf("bad free: free address (%p) != alloc address (%p)\n",
1364 1396 bufarg, buf);
1365 1397 break;
1366 1398 }
1367 1399
1368 1400 printf("buffer=%p bufctl=%p cache: %s\n",
1369 1401 bufarg, (void *)bcp, cparg->cache_name);
1370 1402
1371 1403 if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
1372 1404 error != KMERR_BADBUFCTL) {
1373 1405 int d;
1374 1406 timestruc_t ts;
1375 1407 kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
1376 1408
1377 1409 hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
1378 1410 printf("previous transaction on buffer %p:\n", buf);
1379 1411 printf("thread=%p time=T-%ld.%09ld slab=%p cache: %s\n",
1380 1412 (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1381 1413 (void *)sp, cp->cache_name);
1382 1414 for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
1383 1415 ulong_t off;
1384 1416 char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
1385 1417 printf("%s+%lx\n", sym ? sym : "?", off);
1386 1418 }
1387 1419 }
1388 1420 if (kmem_panic > 0)
1389 1421 panic("kernel heap corruption detected");
1390 1422 if (kmem_panic == 0)
1391 1423 debug_enter(NULL);
1392 1424 kmem_logging = 1; /* resume logging */
1393 1425 }
1394 1426
1395 1427 static kmem_log_header_t *
1396 1428 kmem_log_init(size_t logsize)
1397 1429 {
1398 1430 kmem_log_header_t *lhp;
1399 1431 int nchunks = 4 * max_ncpus;
1400 1432 size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
1401 1433 int i;
1402 1434
1403 1435 /*
1404 1436 * Make sure that lhp->lh_cpu[] is nicely aligned
1405 1437 * to prevent false sharing of cache lines.
1406 1438 */
1407 1439 lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
1408 1440 lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1409 1441 NULL, NULL, VM_SLEEP);
1410 1442 bzero(lhp, lhsize);
1411 1443
1412 1444 mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
1413 1445 lhp->lh_nchunks = nchunks;
1414 1446 lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
1415 1447 lhp->lh_base = vmem_alloc(kmem_log_arena,
1416 1448 lhp->lh_chunksize * nchunks, VM_SLEEP);
1417 1449 lhp->lh_free = vmem_alloc(kmem_log_arena,
1418 1450 nchunks * sizeof (int), VM_SLEEP);
1419 1451 bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1420 1452
1421 1453 for (i = 0; i < max_ncpus; i++) {
1422 1454 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1423 1455 mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
1424 1456 clhp->clh_chunk = i;
1425 1457 }
1426 1458
1427 1459 for (i = max_ncpus; i < nchunks; i++)
1428 1460 lhp->lh_free[i] = i;
1429 1461
1430 1462 lhp->lh_head = max_ncpus;
1431 1463 lhp->lh_tail = 0;
1432 1464
1433 1465 return (lhp);
1434 1466 }
1435 1467
1436 1468 static void *
1437 1469 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
1438 1470 {
1439 1471 void *logspace;
1440 1472 kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
1441 1473
1442 1474 if (lhp == NULL || kmem_logging == 0 || panicstr)
1443 1475 return (NULL);
1444 1476
1445 1477 mutex_enter(&clhp->clh_lock);
1446 1478 clhp->clh_hits++;
1447 1479 if (size > clhp->clh_avail) {
1448 1480 mutex_enter(&lhp->lh_lock);
1449 1481 lhp->lh_hits++;
1450 1482 lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1451 1483 lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1452 1484 clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1453 1485 lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1454 1486 clhp->clh_current = lhp->lh_base +
1455 1487 clhp->clh_chunk * lhp->lh_chunksize;
1456 1488 clhp->clh_avail = lhp->lh_chunksize;
1457 1489 if (size > lhp->lh_chunksize)
1458 1490 size = lhp->lh_chunksize;
1459 1491 mutex_exit(&lhp->lh_lock);
1460 1492 }
1461 1493 logspace = clhp->clh_current;
1462 1494 clhp->clh_current += size;
1463 1495 clhp->clh_avail -= size;
1464 1496 bcopy(data, logspace, size);
1465 1497 mutex_exit(&clhp->clh_lock);
1466 1498 return (logspace);
1467 1499 }
1468 1500
1469 1501 #define KMEM_AUDIT(lp, cp, bcp) \
1470 1502 { \
1471 1503 kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp); \
1472 1504 _bcp->bc_timestamp = gethrtime(); \
1473 1505 _bcp->bc_thread = curthread; \
1474 1506 _bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH); \
1475 1507 _bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp)); \
1476 1508 }
1477 1509
1478 1510 static void
1479 1511 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
1480 1512 kmem_slab_t *sp, void *addr)
1481 1513 {
1482 1514 kmem_bufctl_audit_t bca;
1483 1515
1484 1516 bzero(&bca, sizeof (kmem_bufctl_audit_t));
1485 1517 bca.bc_addr = addr;
1486 1518 bca.bc_slab = sp;
1487 1519 bca.bc_cache = cp;
1488 1520 KMEM_AUDIT(lp, cp, &bca);
1489 1521 }
1490 1522
1491 1523 /*
1492 1524 * Create a new slab for cache cp.
1493 1525 */
1494 1526 static kmem_slab_t *
1495 1527 kmem_slab_create(kmem_cache_t *cp, int kmflag)
1496 1528 {
1497 1529 size_t slabsize = cp->cache_slabsize;
1498 1530 size_t chunksize = cp->cache_chunksize;
1499 1531 int cache_flags = cp->cache_flags;
1500 1532 size_t color, chunks;
1501 1533 char *buf, *slab;
1502 1534 kmem_slab_t *sp;
1503 1535 kmem_bufctl_t *bcp;
1504 1536 vmem_t *vmp = cp->cache_arena;
1505 1537
1506 1538 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1507 1539
1508 1540 color = cp->cache_color + cp->cache_align;
1509 1541 if (color > cp->cache_maxcolor)
1510 1542 color = cp->cache_mincolor;
1511 1543 cp->cache_color = color;
1512 1544
1513 1545 slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
1514 1546
1515 1547 if (slab == NULL)
1516 1548 goto vmem_alloc_failure;
1517 1549
1518 1550 ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1519 1551
1520 1552 /*
1521 1553 * Reverify what was already checked in kmem_cache_set_move(), since the
1522 1554 * consolidator depends (for correctness) on slabs being initialized
1523 1555 * with the 0xbaddcafe memory pattern (setting a low order bit usable by
1524 1556 * clients to distinguish uninitialized memory from known objects).
1525 1557 */
1526 1558 ASSERT((cp->cache_move == NULL) || !(cp->cache_cflags & KMC_NOTOUCH));
1527 1559 if (!(cp->cache_cflags & KMC_NOTOUCH))
1528 1560 copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1529 1561
1530 1562 if (cache_flags & KMF_HASH) {
1531 1563 if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
1532 1564 goto slab_alloc_failure;
1533 1565 chunks = (slabsize - color) / chunksize;
1534 1566 } else {
1535 1567 sp = KMEM_SLAB(cp, slab);
1536 1568 chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
1537 1569 }
1538 1570
1539 1571 sp->slab_cache = cp;
1540 1572 sp->slab_head = NULL;
1541 1573 sp->slab_refcnt = 0;
1542 1574 sp->slab_base = buf = slab + color;
1543 1575 sp->slab_chunks = chunks;
1544 1576 sp->slab_stuck_offset = (uint32_t)-1;
1545 1577 sp->slab_later_count = 0;
1546 1578 sp->slab_flags = 0;
1547 1579
1548 1580 ASSERT(chunks > 0);
1549 1581 while (chunks-- != 0) {
1550 1582 if (cache_flags & KMF_HASH) {
1551 1583 bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
1552 1584 if (bcp == NULL)
1553 1585 goto bufctl_alloc_failure;
1554 1586 if (cache_flags & KMF_AUDIT) {
1555 1587 kmem_bufctl_audit_t *bcap =
1556 1588 (kmem_bufctl_audit_t *)bcp;
1557 1589 bzero(bcap, sizeof (kmem_bufctl_audit_t));
1558 1590 bcap->bc_cache = cp;
1559 1591 }
1560 1592 bcp->bc_addr = buf;
1561 1593 bcp->bc_slab = sp;
1562 1594 } else {
1563 1595 bcp = KMEM_BUFCTL(cp, buf);
1564 1596 }
1565 1597 if (cache_flags & KMF_BUFTAG) {
1566 1598 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1567 1599 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1568 1600 btp->bt_bufctl = bcp;
1569 1601 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1570 1602 if (cache_flags & KMF_DEADBEEF) {
1571 1603 copy_pattern(KMEM_FREE_PATTERN, buf,
1572 1604 cp->cache_verify);
1573 1605 }
1574 1606 }
1575 1607 bcp->bc_next = sp->slab_head;
1576 1608 sp->slab_head = bcp;
1577 1609 buf += chunksize;
1578 1610 }
1579 1611
1580 1612 kmem_log_event(kmem_slab_log, cp, sp, slab);
1581 1613
1582 1614 return (sp);
1583 1615
1584 1616 bufctl_alloc_failure:
1585 1617
1586 1618 while ((bcp = sp->slab_head) != NULL) {
1587 1619 sp->slab_head = bcp->bc_next;
1588 1620 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1589 1621 }
1590 1622 kmem_cache_free(kmem_slab_cache, sp);
1591 1623
1592 1624 slab_alloc_failure:
1593 1625
1594 1626 vmem_free(vmp, slab, slabsize);
1595 1627
1596 1628 vmem_alloc_failure:
1597 1629
1598 1630 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1599 1631 atomic_inc_64(&cp->cache_alloc_fail);
1600 1632
1601 1633 return (NULL);
1602 1634 }
1603 1635
1604 1636 /*
1605 1637 * Destroy a slab.
1606 1638 */
1607 1639 static void
1608 1640 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
1609 1641 {
1610 1642 vmem_t *vmp = cp->cache_arena;
1611 1643 void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1612 1644
1613 1645 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1614 1646 ASSERT(sp->slab_refcnt == 0);
1615 1647
1616 1648 if (cp->cache_flags & KMF_HASH) {
1617 1649 kmem_bufctl_t *bcp;
1618 1650 while ((bcp = sp->slab_head) != NULL) {
1619 1651 sp->slab_head = bcp->bc_next;
1620 1652 kmem_cache_free(cp->cache_bufctl_cache, bcp);
1621 1653 }
1622 1654 kmem_cache_free(kmem_slab_cache, sp);
1623 1655 }
1624 1656 vmem_free(vmp, slab, cp->cache_slabsize);
1625 1657 }
1626 1658
1627 1659 static void *
1628 1660 kmem_slab_alloc_impl(kmem_cache_t *cp, kmem_slab_t *sp, boolean_t prefill)
1629 1661 {
1630 1662 kmem_bufctl_t *bcp, **hash_bucket;
1631 1663 void *buf;
1632 1664 boolean_t new_slab = (sp->slab_refcnt == 0);
1633 1665
1634 1666 ASSERT(MUTEX_HELD(&cp->cache_lock));
1635 1667 /*
1636 1668 * kmem_slab_alloc() drops cache_lock when it creates a new slab, so we
1637 1669 * can't ASSERT(avl_is_empty(&cp->cache_partial_slabs)) here when the
1638 1670 * slab is newly created.
1639 1671 */
1640 1672 ASSERT(new_slab || (KMEM_SLAB_IS_PARTIAL(sp) &&
1641 1673 (sp == avl_first(&cp->cache_partial_slabs))));
1642 1674 ASSERT(sp->slab_cache == cp);
1643 1675
1644 1676 cp->cache_slab_alloc++;
1645 1677 cp->cache_bufslab--;
1646 1678 sp->slab_refcnt++;
1647 1679
1648 1680 bcp = sp->slab_head;
1649 1681 sp->slab_head = bcp->bc_next;
1650 1682
1651 1683 if (cp->cache_flags & KMF_HASH) {
1652 1684 /*
1653 1685 * Add buffer to allocated-address hash table.
1654 1686 */
1655 1687 buf = bcp->bc_addr;
1656 1688 hash_bucket = KMEM_HASH(cp, buf);
1657 1689 bcp->bc_next = *hash_bucket;
1658 1690 *hash_bucket = bcp;
1659 1691 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1660 1692 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1661 1693 }
1662 1694 } else {
1663 1695 buf = KMEM_BUF(cp, bcp);
1664 1696 }
1665 1697
1666 1698 ASSERT(KMEM_SLAB_MEMBER(sp, buf));
1667 1699
1668 1700 if (sp->slab_head == NULL) {
1669 1701 ASSERT(KMEM_SLAB_IS_ALL_USED(sp));
1670 1702 if (new_slab) {
1671 1703 ASSERT(sp->slab_chunks == 1);
1672 1704 } else {
1673 1705 ASSERT(sp->slab_chunks > 1); /* the slab was partial */
1674 1706 avl_remove(&cp->cache_partial_slabs, sp);
1675 1707 sp->slab_later_count = 0; /* clear history */
1676 1708 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
1677 1709 sp->slab_stuck_offset = (uint32_t)-1;
1678 1710 }
1679 1711 list_insert_head(&cp->cache_complete_slabs, sp);
1680 1712 cp->cache_complete_slab_count++;
1681 1713 return (buf);
1682 1714 }
1683 1715
1684 1716 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
1685 1717 /*
1686 1718 * Peek to see if the magazine layer is enabled before
1687 1719 * we prefill. We're not holding the cpu cache lock,
1688 1720 * so the peek could be wrong, but there's no harm in it.
1689 1721 */
1690 1722 if (new_slab && prefill && (cp->cache_flags & KMF_PREFILL) &&
1691 1723 (KMEM_CPU_CACHE(cp)->cc_magsize != 0)) {
1692 1724 kmem_slab_prefill(cp, sp);
1693 1725 return (buf);
1694 1726 }
1695 1727
1696 1728 if (new_slab) {
1697 1729 avl_add(&cp->cache_partial_slabs, sp);
1698 1730 return (buf);
1699 1731 }
1700 1732
1701 1733 /*
1702 1734 * The slab is now more allocated than it was, so the
1703 1735 * order remains unchanged.
1704 1736 */
1705 1737 ASSERT(!avl_update(&cp->cache_partial_slabs, sp));
1706 1738 return (buf);
1707 1739 }
1708 1740
1709 1741 /*
1710 1742 * Allocate a raw (unconstructed) buffer from cp's slab layer.
1711 1743 */
1712 1744 static void *
1713 1745 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
1714 1746 {
1715 1747 kmem_slab_t *sp;
1716 1748 void *buf;
1717 1749 boolean_t test_destructor;
1718 1750
1719 1751 mutex_enter(&cp->cache_lock);
1720 1752 test_destructor = (cp->cache_slab_alloc == 0);
1721 1753 sp = avl_first(&cp->cache_partial_slabs);
1722 1754 if (sp == NULL) {
1723 1755 ASSERT(cp->cache_bufslab == 0);
1724 1756
1725 1757 /*
1726 1758 * The freelist is empty. Create a new slab.
1727 1759 */
1728 1760 mutex_exit(&cp->cache_lock);
1729 1761 if ((sp = kmem_slab_create(cp, kmflag)) == NULL) {
1730 1762 return (NULL);
1731 1763 }
1732 1764 mutex_enter(&cp->cache_lock);
1733 1765 cp->cache_slab_create++;
1734 1766 if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1735 1767 cp->cache_bufmax = cp->cache_buftotal;
1736 1768 cp->cache_bufslab += sp->slab_chunks;
1737 1769 }
1738 1770
1739 1771 buf = kmem_slab_alloc_impl(cp, sp, B_TRUE);
1740 1772 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1741 1773 (cp->cache_complete_slab_count +
1742 1774 avl_numnodes(&cp->cache_partial_slabs) +
1743 1775 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1744 1776 mutex_exit(&cp->cache_lock);
1745 1777
1746 1778 if (test_destructor && cp->cache_destructor != NULL) {
1747 1779 /*
1748 1780 * On the first kmem_slab_alloc(), assert that it is valid to
1749 1781 * call the destructor on a newly constructed object without any
1750 1782 * client involvement.
1751 1783 */
1752 1784 if ((cp->cache_constructor == NULL) ||
1753 1785 cp->cache_constructor(buf, cp->cache_private,
1754 1786 kmflag) == 0) {
1755 1787 cp->cache_destructor(buf, cp->cache_private);
1756 1788 }
1757 1789 copy_pattern(KMEM_UNINITIALIZED_PATTERN, buf,
1758 1790 cp->cache_bufsize);
1759 1791 if (cp->cache_flags & KMF_DEADBEEF) {
1760 1792 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1761 1793 }
1762 1794 }
1763 1795
1764 1796 return (buf);
1765 1797 }
1766 1798
1767 1799 static void kmem_slab_move_yes(kmem_cache_t *, kmem_slab_t *, void *);
1768 1800
1769 1801 /*
1770 1802 * Free a raw (unconstructed) buffer to cp's slab layer.
1771 1803 */
1772 1804 static void
1773 1805 kmem_slab_free(kmem_cache_t *cp, void *buf)
1774 1806 {
1775 1807 kmem_slab_t *sp;
1776 1808 kmem_bufctl_t *bcp, **prev_bcpp;
1777 1809
1778 1810 ASSERT(buf != NULL);
1779 1811
1780 1812 mutex_enter(&cp->cache_lock);
1781 1813 cp->cache_slab_free++;
1782 1814
1783 1815 if (cp->cache_flags & KMF_HASH) {
1784 1816 /*
1785 1817 * Look up buffer in allocated-address hash table.
1786 1818 */
1787 1819 prev_bcpp = KMEM_HASH(cp, buf);
1788 1820 while ((bcp = *prev_bcpp) != NULL) {
1789 1821 if (bcp->bc_addr == buf) {
1790 1822 *prev_bcpp = bcp->bc_next;
1791 1823 sp = bcp->bc_slab;
1792 1824 break;
1793 1825 }
1794 1826 cp->cache_lookup_depth++;
1795 1827 prev_bcpp = &bcp->bc_next;
1796 1828 }
1797 1829 } else {
1798 1830 bcp = KMEM_BUFCTL(cp, buf);
1799 1831 sp = KMEM_SLAB(cp, buf);
1800 1832 }
1801 1833
1802 1834 if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
1803 1835 mutex_exit(&cp->cache_lock);
1804 1836 kmem_error(KMERR_BADADDR, cp, buf);
1805 1837 return;
1806 1838 }
1807 1839
1808 1840 if (KMEM_SLAB_OFFSET(sp, buf) == sp->slab_stuck_offset) {
1809 1841 /*
1810 1842 * If this is the buffer that prevented the consolidator from
1811 1843 * clearing the slab, we can reset the slab flags now that the
1812 1844 * buffer is freed. (It makes sense to do this in
1813 1845 * kmem_cache_free(), where the client gives up ownership of the
1814 1846 * buffer, but on the hot path the test is too expensive.)
1815 1847 */
1816 1848 kmem_slab_move_yes(cp, sp, buf);
1817 1849 }
1818 1850
1819 1851 if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
1820 1852 if (cp->cache_flags & KMF_CONTENTS)
1821 1853 ((kmem_bufctl_audit_t *)bcp)->bc_contents =
1822 1854 kmem_log_enter(kmem_content_log, buf,
1823 1855 cp->cache_contents);
1824 1856 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1825 1857 }
1826 1858
1827 1859 bcp->bc_next = sp->slab_head;
1828 1860 sp->slab_head = bcp;
1829 1861
1830 1862 cp->cache_bufslab++;
1831 1863 ASSERT(sp->slab_refcnt >= 1);
1832 1864
1833 1865 if (--sp->slab_refcnt == 0) {
1834 1866 /*
1835 1867 * There are no outstanding allocations from this slab,
1836 1868 * so we can reclaim the memory.
1837 1869 */
1838 1870 if (sp->slab_chunks == 1) {
1839 1871 list_remove(&cp->cache_complete_slabs, sp);
1840 1872 cp->cache_complete_slab_count--;
1841 1873 } else {
1842 1874 avl_remove(&cp->cache_partial_slabs, sp);
1843 1875 }
1844 1876
1845 1877 cp->cache_buftotal -= sp->slab_chunks;
1846 1878 cp->cache_bufslab -= sp->slab_chunks;
1847 1879 /*
1848 1880 * Defer releasing the slab to the virtual memory subsystem
1849 1881 * while there is a pending move callback, since we guarantee
1850 1882 * that buffers passed to the move callback have only been
1851 1883 * touched by kmem or by the client itself. Since the memory
1852 1884 * patterns baddcafe (uninitialized) and deadbeef (freed) both
1853 1885 * set at least one of the two lowest order bits, the client can
1854 1886 * test those bits in the move callback to determine whether or
1855 1887 * not it knows about the buffer (assuming that the client also
1856 1888 * sets one of those low order bits whenever it frees a buffer).
1857 1889 */
1858 1890 if (cp->cache_defrag == NULL ||
1859 1891 (avl_is_empty(&cp->cache_defrag->kmd_moves_pending) &&
1860 1892 !(sp->slab_flags & KMEM_SLAB_MOVE_PENDING))) {
1861 1893 cp->cache_slab_destroy++;
1862 1894 mutex_exit(&cp->cache_lock);
1863 1895 kmem_slab_destroy(cp, sp);
1864 1896 } else {
1865 1897 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
1866 1898 /*
1867 1899 * Slabs are inserted at both ends of the deadlist to
1868 1900 * distinguish between slabs freed while move callbacks
1869 1901 * are pending (list head) and a slab freed while the
1870 1902 * lock is dropped in kmem_move_buffers() (list tail) so
1871 1903 * that in both cases slab_destroy() is called from the
1872 1904 * right context.
1873 1905 */
1874 1906 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
1875 1907 list_insert_tail(deadlist, sp);
1876 1908 } else {
1877 1909 list_insert_head(deadlist, sp);
1878 1910 }
1879 1911 cp->cache_defrag->kmd_deadcount++;
1880 1912 mutex_exit(&cp->cache_lock);
1881 1913 }
1882 1914 return;
|
↓ open down ↓ |
760 lines elided |
↑ open up ↑ |
1883 1915 }
1884 1916
1885 1917 if (bcp->bc_next == NULL) {
1886 1918 /* Transition the slab from completely allocated to partial. */
1887 1919 ASSERT(sp->slab_refcnt == (sp->slab_chunks - 1));
1888 1920 ASSERT(sp->slab_chunks > 1);
1889 1921 list_remove(&cp->cache_complete_slabs, sp);
1890 1922 cp->cache_complete_slab_count--;
1891 1923 avl_add(&cp->cache_partial_slabs, sp);
1892 1924 } else {
1925 +#ifdef DEBUG
1926 + if (avl_update_gt(&cp->cache_partial_slabs, sp)) {
1927 + KMEM_STAT_ADD(kmem_move_stats.kms_avl_update);
1928 + } else {
1929 + KMEM_STAT_ADD(kmem_move_stats.kms_avl_noupdate);
1930 + }
1931 +#else
1893 1932 (void) avl_update_gt(&cp->cache_partial_slabs, sp);
1933 +#endif
1894 1934 }
1895 1935
1896 1936 ASSERT((cp->cache_slab_create - cp->cache_slab_destroy) ==
1897 1937 (cp->cache_complete_slab_count +
1898 1938 avl_numnodes(&cp->cache_partial_slabs) +
1899 1939 (cp->cache_defrag == NULL ? 0 : cp->cache_defrag->kmd_deadcount)));
1900 1940 mutex_exit(&cp->cache_lock);
1901 1941 }
1902 1942
1903 1943 /*
1904 1944 * Return -1 if kmem_error, 1 if constructor fails, 0 if successful.
1905 1945 */
1906 1946 static int
1907 1947 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
1908 1948 caddr_t caller)
1909 1949 {
1910 1950 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1911 1951 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1912 1952 uint32_t mtbf;
1913 1953
1914 1954 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1915 1955 kmem_error(KMERR_BADBUFTAG, cp, buf);
1916 1956 return (-1);
1917 1957 }
1918 1958
1919 1959 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
1920 1960
1921 1961 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
1922 1962 kmem_error(KMERR_BADBUFCTL, cp, buf);
1923 1963 return (-1);
1924 1964 }
1925 1965
1926 1966 if (cp->cache_flags & KMF_DEADBEEF) {
1927 1967 if (!construct && (cp->cache_flags & KMF_LITE)) {
1928 1968 if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
1929 1969 kmem_error(KMERR_MODIFIED, cp, buf);
1930 1970 return (-1);
1931 1971 }
1932 1972 if (cp->cache_constructor != NULL)
1933 1973 *(uint64_t *)buf = btp->bt_redzone;
1934 1974 else
1935 1975 *(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
1936 1976 } else {
1937 1977 construct = 1;
1938 1978 if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
1939 1979 KMEM_UNINITIALIZED_PATTERN, buf,
1940 1980 cp->cache_verify)) {
1941 1981 kmem_error(KMERR_MODIFIED, cp, buf);
1942 1982 return (-1);
1943 1983 }
1944 1984 }
1945 1985 }
1946 1986 btp->bt_redzone = KMEM_REDZONE_PATTERN;
1947 1987
1948 1988 if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
1949 1989 gethrtime() % mtbf == 0 &&
1950 1990 (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
1951 1991 kmem_log_event(kmem_failure_log, cp, NULL, NULL);
1952 1992 if (!construct && cp->cache_destructor != NULL)
1953 1993 cp->cache_destructor(buf, cp->cache_private);
1954 1994 } else {
1955 1995 mtbf = 0;
1956 1996 }
1957 1997
1958 1998 if (mtbf || (construct && cp->cache_constructor != NULL &&
1959 1999 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
1960 2000 atomic_inc_64(&cp->cache_alloc_fail);
1961 2001 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
1962 2002 if (cp->cache_flags & KMF_DEADBEEF)
1963 2003 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1964 2004 kmem_slab_free(cp, buf);
1965 2005 return (1);
1966 2006 }
1967 2007
1968 2008 if (cp->cache_flags & KMF_AUDIT) {
1969 2009 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
1970 2010 }
1971 2011
1972 2012 if ((cp->cache_flags & KMF_LITE) &&
1973 2013 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
1974 2014 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
1975 2015 }
1976 2016
1977 2017 return (0);
1978 2018 }
1979 2019
1980 2020 static int
1981 2021 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
1982 2022 {
1983 2023 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1984 2024 kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
1985 2025 kmem_slab_t *sp;
1986 2026
1987 2027 if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
1988 2028 if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
1989 2029 kmem_error(KMERR_DUPFREE, cp, buf);
1990 2030 return (-1);
1991 2031 }
1992 2032 sp = kmem_findslab(cp, buf);
1993 2033 if (sp == NULL || sp->slab_cache != cp)
1994 2034 kmem_error(KMERR_BADADDR, cp, buf);
1995 2035 else
1996 2036 kmem_error(KMERR_REDZONE, cp, buf);
1997 2037 return (-1);
1998 2038 }
1999 2039
2000 2040 btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
2001 2041
2002 2042 if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
2003 2043 kmem_error(KMERR_BADBUFCTL, cp, buf);
2004 2044 return (-1);
2005 2045 }
2006 2046
2007 2047 if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
2008 2048 kmem_error(KMERR_REDZONE, cp, buf);
2009 2049 return (-1);
2010 2050 }
2011 2051
2012 2052 if (cp->cache_flags & KMF_AUDIT) {
2013 2053 if (cp->cache_flags & KMF_CONTENTS)
2014 2054 bcp->bc_contents = kmem_log_enter(kmem_content_log,
2015 2055 buf, cp->cache_contents);
2016 2056 KMEM_AUDIT(kmem_transaction_log, cp, bcp);
2017 2057 }
2018 2058
2019 2059 if ((cp->cache_flags & KMF_LITE) &&
2020 2060 !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
2021 2061 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
2022 2062 }
2023 2063
2024 2064 if (cp->cache_flags & KMF_DEADBEEF) {
2025 2065 if (cp->cache_flags & KMF_LITE)
2026 2066 btp->bt_redzone = *(uint64_t *)buf;
2027 2067 else if (cp->cache_destructor != NULL)
2028 2068 cp->cache_destructor(buf, cp->cache_private);
2029 2069
2030 2070 copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
2031 2071 }
2032 2072
2033 2073 return (0);
2034 2074 }
2035 2075
2036 2076 /*
2037 2077 * Free each object in magazine mp to cp's slab layer, and free mp itself.
2038 2078 */
2039 2079 static void
2040 2080 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
2041 2081 {
2042 2082 int round;
2043 2083
2044 2084 ASSERT(!list_link_active(&cp->cache_link) ||
2045 2085 taskq_member(kmem_taskq, curthread));
2046 2086
2047 2087 for (round = 0; round < nrounds; round++) {
2048 2088 void *buf = mp->mag_round[round];
2049 2089
2050 2090 if (cp->cache_flags & KMF_DEADBEEF) {
2051 2091 if (verify_pattern(KMEM_FREE_PATTERN, buf,
2052 2092 cp->cache_verify) != NULL) {
2053 2093 kmem_error(KMERR_MODIFIED, cp, buf);
2054 2094 continue;
2055 2095 }
2056 2096 if ((cp->cache_flags & KMF_LITE) &&
2057 2097 cp->cache_destructor != NULL) {
2058 2098 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2059 2099 *(uint64_t *)buf = btp->bt_redzone;
2060 2100 cp->cache_destructor(buf, cp->cache_private);
2061 2101 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2062 2102 }
2063 2103 } else if (cp->cache_destructor != NULL) {
2064 2104 cp->cache_destructor(buf, cp->cache_private);
2065 2105 }
2066 2106
2067 2107 kmem_slab_free(cp, buf);
2068 2108 }
2069 2109 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2070 2110 kmem_cache_free(cp->cache_magtype->mt_cache, mp);
2071 2111 }
2072 2112
2073 2113 /*
2074 2114 * Allocate a magazine from the depot.
2075 2115 */
2076 2116 static kmem_magazine_t *
2077 2117 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
2078 2118 {
2079 2119 kmem_magazine_t *mp;
2080 2120
2081 2121 /*
2082 2122 * If we can't get the depot lock without contention,
2083 2123 * update our contention count. We use the depot
2084 2124 * contention rate to determine whether we need to
2085 2125 * increase the magazine size for better scalability.
2086 2126 */
2087 2127 if (!mutex_tryenter(&cp->cache_depot_lock)) {
2088 2128 mutex_enter(&cp->cache_depot_lock);
2089 2129 cp->cache_depot_contention++;
2090 2130 }
2091 2131
2092 2132 if ((mp = mlp->ml_list) != NULL) {
2093 2133 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2094 2134 mlp->ml_list = mp->mag_next;
2095 2135 if (--mlp->ml_total < mlp->ml_min)
2096 2136 mlp->ml_min = mlp->ml_total;
2097 2137 mlp->ml_alloc++;
2098 2138 }
2099 2139
2100 2140 mutex_exit(&cp->cache_depot_lock);
2101 2141
2102 2142 return (mp);
2103 2143 }
2104 2144
2105 2145 /*
2106 2146 * Free a magazine to the depot.
2107 2147 */
2108 2148 static void
2109 2149 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
2110 2150 {
2111 2151 mutex_enter(&cp->cache_depot_lock);
2112 2152 ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
2113 2153 mp->mag_next = mlp->ml_list;
2114 2154 mlp->ml_list = mp;
2115 2155 mlp->ml_total++;
2116 2156 mutex_exit(&cp->cache_depot_lock);
2117 2157 }
2118 2158
2119 2159 /*
2120 2160 * Update the working set statistics for cp's depot.
2121 2161 */
2122 2162 static void
2123 2163 kmem_depot_ws_update(kmem_cache_t *cp)
2124 2164 {
2125 2165 mutex_enter(&cp->cache_depot_lock);
2126 2166 cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
2127 2167 cp->cache_full.ml_min = cp->cache_full.ml_total;
2128 2168 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
2129 2169 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2130 2170 mutex_exit(&cp->cache_depot_lock);
2131 2171 }
2132 2172
2133 2173 /*
2134 2174 * Set the working set statistics for cp's depot to zero. (Everything is
2135 2175 * eligible for reaping.)
2136 2176 */
2137 2177 static void
2138 2178 kmem_depot_ws_zero(kmem_cache_t *cp)
2139 2179 {
2140 2180 mutex_enter(&cp->cache_depot_lock);
2141 2181 cp->cache_full.ml_reaplimit = cp->cache_full.ml_total;
2142 2182 cp->cache_full.ml_min = cp->cache_full.ml_total;
2143 2183 cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_total;
2144 2184 cp->cache_empty.ml_min = cp->cache_empty.ml_total;
2145 2185 mutex_exit(&cp->cache_depot_lock);
2146 2186 }
2147 2187
2148 2188 /*
2149 2189 * The number of bytes to reap before we call kpreempt(). The default (1MB)
2150 2190 * causes us to preempt reaping up to hundreds of times per second. Using a
2151 2191 * larger value (1GB) causes this to have virtually no effect.
2152 2192 */
2153 2193 size_t kmem_reap_preempt_bytes = 1024 * 1024;
2154 2194
2155 2195 /*
2156 2196 * Reap all magazines that have fallen out of the depot's working set.
2157 2197 */
2158 2198 static void
2159 2199 kmem_depot_ws_reap(kmem_cache_t *cp)
2160 2200 {
2161 2201 size_t bytes = 0;
2162 2202 long reap;
2163 2203 kmem_magazine_t *mp;
2164 2204
2165 2205 ASSERT(!list_link_active(&cp->cache_link) ||
2166 2206 taskq_member(kmem_taskq, curthread));
2167 2207
2168 2208 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
2169 2209 while (reap-- &&
2170 2210 (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL) {
2171 2211 kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
2172 2212 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2173 2213 if (bytes > kmem_reap_preempt_bytes) {
2174 2214 kpreempt(KPREEMPT_SYNC);
2175 2215 bytes = 0;
2176 2216 }
2177 2217 }
2178 2218
2179 2219 reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
2180 2220 while (reap-- &&
2181 2221 (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL) {
2182 2222 kmem_magazine_destroy(cp, mp, 0);
2183 2223 bytes += cp->cache_magtype->mt_magsize * cp->cache_bufsize;
2184 2224 if (bytes > kmem_reap_preempt_bytes) {
2185 2225 kpreempt(KPREEMPT_SYNC);
2186 2226 bytes = 0;
2187 2227 }
2188 2228 }
2189 2229 }
2190 2230
2191 2231 static void
2192 2232 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
2193 2233 {
2194 2234 ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
2195 2235 (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
2196 2236 ASSERT(ccp->cc_magsize > 0);
2197 2237
2198 2238 ccp->cc_ploaded = ccp->cc_loaded;
2199 2239 ccp->cc_prounds = ccp->cc_rounds;
2200 2240 ccp->cc_loaded = mp;
2201 2241 ccp->cc_rounds = rounds;
2202 2242 }
2203 2243
2204 2244 /*
2205 2245 * Intercept kmem alloc/free calls during crash dump in order to avoid
2206 2246 * changing kmem state while memory is being saved to the dump device.
2207 2247 * Otherwise, ::kmem_verify will report "corrupt buffers". Note that
2208 2248 * there are no locks because only one CPU calls kmem during a crash
2209 2249 * dump. To enable this feature, first create the associated vmem
2210 2250 * arena with VMC_DUMPSAFE.
2211 2251 */
2212 2252 static void *kmem_dump_start; /* start of pre-reserved heap */
2213 2253 static void *kmem_dump_end; /* end of heap area */
2214 2254 static void *kmem_dump_curr; /* current free heap pointer */
2215 2255 static size_t kmem_dump_size; /* size of heap area */
2216 2256
2217 2257 /* append to each buf created in the pre-reserved heap */
2218 2258 typedef struct kmem_dumpctl {
2219 2259 void *kdc_next; /* cache dump free list linkage */
2220 2260 } kmem_dumpctl_t;
2221 2261
2222 2262 #define KMEM_DUMPCTL(cp, buf) \
2223 2263 ((kmem_dumpctl_t *)P2ROUNDUP((uintptr_t)(buf) + (cp)->cache_bufsize, \
2224 2264 sizeof (void *)))
2225 2265
2226 2266 /* Keep some simple stats. */
2227 2267 #define KMEM_DUMP_LOGS (100)
2228 2268
2229 2269 typedef struct kmem_dump_log {
2230 2270 kmem_cache_t *kdl_cache;
2231 2271 uint_t kdl_allocs; /* # of dump allocations */
2232 2272 uint_t kdl_frees; /* # of dump frees */
2233 2273 uint_t kdl_alloc_fails; /* # of allocation failures */
2234 2274 uint_t kdl_free_nondump; /* # of non-dump frees */
2235 2275 uint_t kdl_unsafe; /* cache was used, but unsafe */
2236 2276 } kmem_dump_log_t;
2237 2277
2238 2278 static kmem_dump_log_t *kmem_dump_log;
2239 2279 static int kmem_dump_log_idx;
2240 2280
2241 2281 #define KDI_LOG(cp, stat) { \
2242 2282 kmem_dump_log_t *kdl; \
2243 2283 if ((kdl = (kmem_dump_log_t *)((cp)->cache_dumplog)) != NULL) { \
2244 2284 kdl->stat++; \
2245 2285 } else if (kmem_dump_log_idx < KMEM_DUMP_LOGS) { \
2246 2286 kdl = &kmem_dump_log[kmem_dump_log_idx++]; \
2247 2287 kdl->stat++; \
2248 2288 kdl->kdl_cache = (cp); \
2249 2289 (cp)->cache_dumplog = kdl; \
2250 2290 } \
2251 2291 }
2252 2292
2253 2293 /* set non zero for full report */
2254 2294 uint_t kmem_dump_verbose = 0;
2255 2295
2256 2296 /* stats for overize heap */
2257 2297 uint_t kmem_dump_oversize_allocs = 0;
2258 2298 uint_t kmem_dump_oversize_max = 0;
2259 2299
2260 2300 static void
2261 2301 kmem_dumppr(char **pp, char *e, const char *format, ...)
2262 2302 {
2263 2303 char *p = *pp;
2264 2304
2265 2305 if (p < e) {
2266 2306 int n;
2267 2307 va_list ap;
2268 2308
2269 2309 va_start(ap, format);
2270 2310 n = vsnprintf(p, e - p, format, ap);
2271 2311 va_end(ap);
2272 2312 *pp = p + n;
2273 2313 }
2274 2314 }
2275 2315
2276 2316 /*
2277 2317 * Called when dumpadm(1M) configures dump parameters.
2278 2318 */
2279 2319 void
2280 2320 kmem_dump_init(size_t size)
2281 2321 {
2282 2322 if (kmem_dump_start != NULL)
2283 2323 kmem_free(kmem_dump_start, kmem_dump_size);
2284 2324
2285 2325 if (kmem_dump_log == NULL)
2286 2326 kmem_dump_log = (kmem_dump_log_t *)kmem_zalloc(KMEM_DUMP_LOGS *
2287 2327 sizeof (kmem_dump_log_t), KM_SLEEP);
2288 2328
2289 2329 kmem_dump_start = kmem_alloc(size, KM_SLEEP);
2290 2330
2291 2331 if (kmem_dump_start != NULL) {
2292 2332 kmem_dump_size = size;
2293 2333 kmem_dump_curr = kmem_dump_start;
2294 2334 kmem_dump_end = (void *)((char *)kmem_dump_start + size);
2295 2335 copy_pattern(KMEM_UNINITIALIZED_PATTERN, kmem_dump_start, size);
2296 2336 } else {
2297 2337 kmem_dump_size = 0;
2298 2338 kmem_dump_curr = NULL;
2299 2339 kmem_dump_end = NULL;
2300 2340 }
2301 2341 }
2302 2342
2303 2343 /*
2304 2344 * Set flag for each kmem_cache_t if is safe to use alternate dump
2305 2345 * memory. Called just before panic crash dump starts. Set the flag
2306 2346 * for the calling CPU.
2307 2347 */
2308 2348 void
2309 2349 kmem_dump_begin(void)
2310 2350 {
2311 2351 ASSERT(panicstr != NULL);
2312 2352 if (kmem_dump_start != NULL) {
2313 2353 kmem_cache_t *cp;
2314 2354
2315 2355 for (cp = list_head(&kmem_caches); cp != NULL;
2316 2356 cp = list_next(&kmem_caches, cp)) {
2317 2357 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2318 2358
2319 2359 if (cp->cache_arena->vm_cflags & VMC_DUMPSAFE) {
2320 2360 cp->cache_flags |= KMF_DUMPDIVERT;
2321 2361 ccp->cc_flags |= KMF_DUMPDIVERT;
2322 2362 ccp->cc_dump_rounds = ccp->cc_rounds;
2323 2363 ccp->cc_dump_prounds = ccp->cc_prounds;
2324 2364 ccp->cc_rounds = ccp->cc_prounds = -1;
2325 2365 } else {
2326 2366 cp->cache_flags |= KMF_DUMPUNSAFE;
2327 2367 ccp->cc_flags |= KMF_DUMPUNSAFE;
2328 2368 }
2329 2369 }
2330 2370 }
2331 2371 }
2332 2372
2333 2373 /*
2334 2374 * finished dump intercept
2335 2375 * print any warnings on the console
2336 2376 * return verbose information to dumpsys() in the given buffer
2337 2377 */
2338 2378 size_t
2339 2379 kmem_dump_finish(char *buf, size_t size)
2340 2380 {
2341 2381 int kdi_idx;
2342 2382 int kdi_end = kmem_dump_log_idx;
2343 2383 int percent = 0;
2344 2384 int header = 0;
2345 2385 int warn = 0;
2346 2386 size_t used;
2347 2387 kmem_cache_t *cp;
2348 2388 kmem_dump_log_t *kdl;
2349 2389 char *e = buf + size;
2350 2390 char *p = buf;
2351 2391
2352 2392 if (kmem_dump_size == 0 || kmem_dump_verbose == 0)
2353 2393 return (0);
2354 2394
2355 2395 used = (char *)kmem_dump_curr - (char *)kmem_dump_start;
2356 2396 percent = (used * 100) / kmem_dump_size;
2357 2397
2358 2398 kmem_dumppr(&p, e, "%% heap used,%d\n", percent);
2359 2399 kmem_dumppr(&p, e, "used bytes,%ld\n", used);
2360 2400 kmem_dumppr(&p, e, "heap size,%ld\n", kmem_dump_size);
2361 2401 kmem_dumppr(&p, e, "Oversize allocs,%d\n",
2362 2402 kmem_dump_oversize_allocs);
2363 2403 kmem_dumppr(&p, e, "Oversize max size,%ld\n",
2364 2404 kmem_dump_oversize_max);
2365 2405
2366 2406 for (kdi_idx = 0; kdi_idx < kdi_end; kdi_idx++) {
2367 2407 kdl = &kmem_dump_log[kdi_idx];
2368 2408 cp = kdl->kdl_cache;
2369 2409 if (cp == NULL)
2370 2410 break;
2371 2411 if (kdl->kdl_alloc_fails)
2372 2412 ++warn;
2373 2413 if (header == 0) {
2374 2414 kmem_dumppr(&p, e,
2375 2415 "Cache Name,Allocs,Frees,Alloc Fails,"
2376 2416 "Nondump Frees,Unsafe Allocs/Frees\n");
2377 2417 header = 1;
2378 2418 }
2379 2419 kmem_dumppr(&p, e, "%s,%d,%d,%d,%d,%d\n",
2380 2420 cp->cache_name, kdl->kdl_allocs, kdl->kdl_frees,
2381 2421 kdl->kdl_alloc_fails, kdl->kdl_free_nondump,
2382 2422 kdl->kdl_unsafe);
2383 2423 }
2384 2424
2385 2425 /* return buffer size used */
2386 2426 if (p < e)
2387 2427 bzero(p, e - p);
2388 2428 return (p - buf);
2389 2429 }
2390 2430
2391 2431 /*
2392 2432 * Allocate a constructed object from alternate dump memory.
2393 2433 */
2394 2434 void *
2395 2435 kmem_cache_alloc_dump(kmem_cache_t *cp, int kmflag)
2396 2436 {
2397 2437 void *buf;
2398 2438 void *curr;
2399 2439 char *bufend;
2400 2440
2401 2441 /* return a constructed object */
2402 2442 if ((buf = cp->cache_dumpfreelist) != NULL) {
2403 2443 cp->cache_dumpfreelist = KMEM_DUMPCTL(cp, buf)->kdc_next;
2404 2444 KDI_LOG(cp, kdl_allocs);
2405 2445 return (buf);
2406 2446 }
2407 2447
2408 2448 /* create a new constructed object */
2409 2449 curr = kmem_dump_curr;
2410 2450 buf = (void *)P2ROUNDUP((uintptr_t)curr, cp->cache_align);
2411 2451 bufend = (char *)KMEM_DUMPCTL(cp, buf) + sizeof (kmem_dumpctl_t);
2412 2452
2413 2453 /* hat layer objects cannot cross a page boundary */
2414 2454 if (cp->cache_align < PAGESIZE) {
2415 2455 char *page = (char *)P2ROUNDUP((uintptr_t)buf, PAGESIZE);
2416 2456 if (bufend > page) {
2417 2457 bufend += page - (char *)buf;
2418 2458 buf = (void *)page;
2419 2459 }
2420 2460 }
2421 2461
2422 2462 /* fall back to normal alloc if reserved area is used up */
2423 2463 if (bufend > (char *)kmem_dump_end) {
2424 2464 kmem_dump_curr = kmem_dump_end;
2425 2465 KDI_LOG(cp, kdl_alloc_fails);
2426 2466 return (NULL);
2427 2467 }
2428 2468
2429 2469 /*
2430 2470 * Must advance curr pointer before calling a constructor that
2431 2471 * may also allocate memory.
2432 2472 */
2433 2473 kmem_dump_curr = bufend;
2434 2474
2435 2475 /* run constructor */
2436 2476 if (cp->cache_constructor != NULL &&
2437 2477 cp->cache_constructor(buf, cp->cache_private, kmflag)
2438 2478 != 0) {
2439 2479 #ifdef DEBUG
2440 2480 printf("name='%s' cache=0x%p: kmem cache constructor failed\n",
2441 2481 cp->cache_name, (void *)cp);
2442 2482 #endif
2443 2483 /* reset curr pointer iff no allocs were done */
2444 2484 if (kmem_dump_curr == bufend)
2445 2485 kmem_dump_curr = curr;
2446 2486
2447 2487 /* fall back to normal alloc if the constructor fails */
2448 2488 KDI_LOG(cp, kdl_alloc_fails);
2449 2489 return (NULL);
2450 2490 }
2451 2491
2452 2492 KDI_LOG(cp, kdl_allocs);
2453 2493 return (buf);
2454 2494 }
2455 2495
2456 2496 /*
2457 2497 * Free a constructed object in alternate dump memory.
2458 2498 */
2459 2499 int
2460 2500 kmem_cache_free_dump(kmem_cache_t *cp, void *buf)
2461 2501 {
2462 2502 /* save constructed buffers for next time */
2463 2503 if ((char *)buf >= (char *)kmem_dump_start &&
2464 2504 (char *)buf < (char *)kmem_dump_end) {
2465 2505 KMEM_DUMPCTL(cp, buf)->kdc_next = cp->cache_dumpfreelist;
2466 2506 cp->cache_dumpfreelist = buf;
2467 2507 KDI_LOG(cp, kdl_frees);
2468 2508 return (0);
2469 2509 }
2470 2510
2471 2511 /* count all non-dump buf frees */
2472 2512 KDI_LOG(cp, kdl_free_nondump);
2473 2513
2474 2514 /* just drop buffers that were allocated before dump started */
2475 2515 if (kmem_dump_curr < kmem_dump_end)
2476 2516 return (0);
2477 2517
2478 2518 /* fall back to normal free if reserved area is used up */
2479 2519 return (1);
2480 2520 }
2481 2521
2482 2522 /*
2483 2523 * Allocate a constructed object from cache cp.
2484 2524 */
2485 2525 void *
2486 2526 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
2487 2527 {
2488 2528 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2489 2529 kmem_magazine_t *fmp;
2490 2530 void *buf;
2491 2531
2492 2532 mutex_enter(&ccp->cc_lock);
2493 2533 for (;;) {
2494 2534 /*
2495 2535 * If there's an object available in the current CPU's
2496 2536 * loaded magazine, just take it and return.
2497 2537 */
2498 2538 if (ccp->cc_rounds > 0) {
2499 2539 buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
2500 2540 ccp->cc_alloc++;
2501 2541 mutex_exit(&ccp->cc_lock);
2502 2542 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPUNSAFE)) {
2503 2543 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2504 2544 ASSERT(!(ccp->cc_flags &
2505 2545 KMF_DUMPDIVERT));
2506 2546 KDI_LOG(cp, kdl_unsafe);
2507 2547 }
2508 2548 if ((ccp->cc_flags & KMF_BUFTAG) &&
2509 2549 kmem_cache_alloc_debug(cp, buf, kmflag, 0,
2510 2550 caller()) != 0) {
2511 2551 if (kmflag & KM_NOSLEEP)
2512 2552 return (NULL);
2513 2553 mutex_enter(&ccp->cc_lock);
2514 2554 continue;
2515 2555 }
2516 2556 }
2517 2557 return (buf);
2518 2558 }
2519 2559
2520 2560 /*
2521 2561 * The loaded magazine is empty. If the previously loaded
2522 2562 * magazine was full, exchange them and try again.
2523 2563 */
2524 2564 if (ccp->cc_prounds > 0) {
2525 2565 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2526 2566 continue;
2527 2567 }
2528 2568
2529 2569 /*
2530 2570 * Return an alternate buffer at dump time to preserve
2531 2571 * the heap.
2532 2572 */
2533 2573 if (ccp->cc_flags & (KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2534 2574 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2535 2575 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2536 2576 /* log it so that we can warn about it */
2537 2577 KDI_LOG(cp, kdl_unsafe);
2538 2578 } else {
2539 2579 if ((buf = kmem_cache_alloc_dump(cp, kmflag)) !=
2540 2580 NULL) {
2541 2581 mutex_exit(&ccp->cc_lock);
2542 2582 return (buf);
2543 2583 }
2544 2584 break; /* fall back to slab layer */
2545 2585 }
2546 2586 }
2547 2587
2548 2588 /*
2549 2589 * If the magazine layer is disabled, break out now.
2550 2590 */
2551 2591 if (ccp->cc_magsize == 0)
2552 2592 break;
2553 2593
2554 2594 /*
2555 2595 * Try to get a full magazine from the depot.
2556 2596 */
2557 2597 fmp = kmem_depot_alloc(cp, &cp->cache_full);
2558 2598 if (fmp != NULL) {
2559 2599 if (ccp->cc_ploaded != NULL)
2560 2600 kmem_depot_free(cp, &cp->cache_empty,
2561 2601 ccp->cc_ploaded);
2562 2602 kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
2563 2603 continue;
2564 2604 }
2565 2605
2566 2606 /*
2567 2607 * There are no full magazines in the depot,
2568 2608 * so fall through to the slab layer.
2569 2609 */
2570 2610 break;
2571 2611 }
2572 2612 mutex_exit(&ccp->cc_lock);
2573 2613
2574 2614 /*
2575 2615 * We couldn't allocate a constructed object from the magazine layer,
2576 2616 * so get a raw buffer from the slab layer and apply its constructor.
2577 2617 */
2578 2618 buf = kmem_slab_alloc(cp, kmflag);
2579 2619
2580 2620 if (buf == NULL)
2581 2621 return (NULL);
2582 2622
2583 2623 if (cp->cache_flags & KMF_BUFTAG) {
2584 2624 /*
2585 2625 * Make kmem_cache_alloc_debug() apply the constructor for us.
2586 2626 */
2587 2627 int rc = kmem_cache_alloc_debug(cp, buf, kmflag, 1, caller());
2588 2628 if (rc != 0) {
2589 2629 if (kmflag & KM_NOSLEEP)
2590 2630 return (NULL);
2591 2631 /*
2592 2632 * kmem_cache_alloc_debug() detected corruption
2593 2633 * but didn't panic (kmem_panic <= 0). We should not be
2594 2634 * here because the constructor failed (indicated by a
2595 2635 * return code of 1). Try again.
2596 2636 */
2597 2637 ASSERT(rc == -1);
2598 2638 return (kmem_cache_alloc(cp, kmflag));
2599 2639 }
2600 2640 return (buf);
2601 2641 }
2602 2642
2603 2643 if (cp->cache_constructor != NULL &&
2604 2644 cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
2605 2645 atomic_inc_64(&cp->cache_alloc_fail);
2606 2646 kmem_slab_free(cp, buf);
2607 2647 return (NULL);
2608 2648 }
2609 2649
2610 2650 return (buf);
2611 2651 }
2612 2652
2613 2653 /*
2614 2654 * The freed argument tells whether or not kmem_cache_free_debug() has already
2615 2655 * been called so that we can avoid the duplicate free error. For example, a
2616 2656 * buffer on a magazine has already been freed by the client but is still
2617 2657 * constructed.
2618 2658 */
2619 2659 static void
2620 2660 kmem_slab_free_constructed(kmem_cache_t *cp, void *buf, boolean_t freed)
2621 2661 {
2622 2662 if (!freed && (cp->cache_flags & KMF_BUFTAG))
2623 2663 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2624 2664 return;
2625 2665
2626 2666 /*
2627 2667 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
2628 2668 * kmem_cache_free_debug() will have already applied the destructor.
2629 2669 */
2630 2670 if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
2631 2671 cp->cache_destructor != NULL) {
2632 2672 if (cp->cache_flags & KMF_DEADBEEF) { /* KMF_LITE implied */
2633 2673 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2634 2674 *(uint64_t *)buf = btp->bt_redzone;
2635 2675 cp->cache_destructor(buf, cp->cache_private);
2636 2676 *(uint64_t *)buf = KMEM_FREE_PATTERN;
2637 2677 } else {
2638 2678 cp->cache_destructor(buf, cp->cache_private);
2639 2679 }
2640 2680 }
2641 2681
2642 2682 kmem_slab_free(cp, buf);
2643 2683 }
2644 2684
2645 2685 /*
2646 2686 * Used when there's no room to free a buffer to the per-CPU cache.
2647 2687 * Drops and re-acquires &ccp->cc_lock, and returns non-zero if the
2648 2688 * caller should try freeing to the per-CPU cache again.
2649 2689 * Note that we don't directly install the magazine in the cpu cache,
2650 2690 * since its state may have changed wildly while the lock was dropped.
2651 2691 */
2652 2692 static int
2653 2693 kmem_cpucache_magazine_alloc(kmem_cpu_cache_t *ccp, kmem_cache_t *cp)
2654 2694 {
2655 2695 kmem_magazine_t *emp;
2656 2696 kmem_magtype_t *mtp;
2657 2697
2658 2698 ASSERT(MUTEX_HELD(&ccp->cc_lock));
2659 2699 ASSERT(((uint_t)ccp->cc_rounds == ccp->cc_magsize ||
2660 2700 ((uint_t)ccp->cc_rounds == -1)) &&
2661 2701 ((uint_t)ccp->cc_prounds == ccp->cc_magsize ||
2662 2702 ((uint_t)ccp->cc_prounds == -1)));
2663 2703
2664 2704 emp = kmem_depot_alloc(cp, &cp->cache_empty);
2665 2705 if (emp != NULL) {
2666 2706 if (ccp->cc_ploaded != NULL)
2667 2707 kmem_depot_free(cp, &cp->cache_full,
2668 2708 ccp->cc_ploaded);
2669 2709 kmem_cpu_reload(ccp, emp, 0);
2670 2710 return (1);
2671 2711 }
2672 2712 /*
2673 2713 * There are no empty magazines in the depot,
2674 2714 * so try to allocate a new one. We must drop all locks
2675 2715 * across kmem_cache_alloc() because lower layers may
2676 2716 * attempt to allocate from this cache.
2677 2717 */
2678 2718 mtp = cp->cache_magtype;
2679 2719 mutex_exit(&ccp->cc_lock);
2680 2720 emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
2681 2721 mutex_enter(&ccp->cc_lock);
2682 2722
2683 2723 if (emp != NULL) {
2684 2724 /*
2685 2725 * We successfully allocated an empty magazine.
2686 2726 * However, we had to drop ccp->cc_lock to do it,
2687 2727 * so the cache's magazine size may have changed.
2688 2728 * If so, free the magazine and try again.
2689 2729 */
2690 2730 if (ccp->cc_magsize != mtp->mt_magsize) {
2691 2731 mutex_exit(&ccp->cc_lock);
2692 2732 kmem_cache_free(mtp->mt_cache, emp);
2693 2733 mutex_enter(&ccp->cc_lock);
2694 2734 return (1);
2695 2735 }
2696 2736
2697 2737 /*
2698 2738 * We got a magazine of the right size. Add it to
2699 2739 * the depot and try the whole dance again.
2700 2740 */
2701 2741 kmem_depot_free(cp, &cp->cache_empty, emp);
2702 2742 return (1);
2703 2743 }
2704 2744
2705 2745 /*
2706 2746 * We couldn't allocate an empty magazine,
2707 2747 * so fall through to the slab layer.
2708 2748 */
2709 2749 return (0);
2710 2750 }
2711 2751
2712 2752 /*
2713 2753 * Free a constructed object to cache cp.
2714 2754 */
2715 2755 void
2716 2756 kmem_cache_free(kmem_cache_t *cp, void *buf)
2717 2757 {
2718 2758 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2719 2759
2720 2760 /*
2721 2761 * The client must not free either of the buffers passed to the move
2722 2762 * callback function.
2723 2763 */
2724 2764 ASSERT(cp->cache_defrag == NULL ||
2725 2765 cp->cache_defrag->kmd_thread != curthread ||
2726 2766 (buf != cp->cache_defrag->kmd_from_buf &&
2727 2767 buf != cp->cache_defrag->kmd_to_buf));
2728 2768
2729 2769 if (ccp->cc_flags & (KMF_BUFTAG | KMF_DUMPDIVERT | KMF_DUMPUNSAFE)) {
2730 2770 if (ccp->cc_flags & KMF_DUMPUNSAFE) {
2731 2771 ASSERT(!(ccp->cc_flags & KMF_DUMPDIVERT));
2732 2772 /* log it so that we can warn about it */
2733 2773 KDI_LOG(cp, kdl_unsafe);
2734 2774 } else if (KMEM_DUMPCC(ccp) && !kmem_cache_free_dump(cp, buf)) {
2735 2775 return;
2736 2776 }
2737 2777 if (ccp->cc_flags & KMF_BUFTAG) {
2738 2778 if (kmem_cache_free_debug(cp, buf, caller()) == -1)
2739 2779 return;
2740 2780 }
2741 2781 }
2742 2782
2743 2783 mutex_enter(&ccp->cc_lock);
2744 2784 /*
2745 2785 * Any changes to this logic should be reflected in kmem_slab_prefill()
2746 2786 */
2747 2787 for (;;) {
2748 2788 /*
2749 2789 * If there's a slot available in the current CPU's
2750 2790 * loaded magazine, just put the object there and return.
2751 2791 */
2752 2792 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2753 2793 ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2754 2794 ccp->cc_free++;
2755 2795 mutex_exit(&ccp->cc_lock);
2756 2796 return;
2757 2797 }
2758 2798
2759 2799 /*
2760 2800 * The loaded magazine is full. If the previously loaded
2761 2801 * magazine was empty, exchange them and try again.
2762 2802 */
2763 2803 if (ccp->cc_prounds == 0) {
2764 2804 kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2765 2805 continue;
2766 2806 }
2767 2807
2768 2808 /*
2769 2809 * If the magazine layer is disabled, break out now.
2770 2810 */
2771 2811 if (ccp->cc_magsize == 0)
2772 2812 break;
2773 2813
2774 2814 if (!kmem_cpucache_magazine_alloc(ccp, cp)) {
2775 2815 /*
2776 2816 * We couldn't free our constructed object to the
2777 2817 * magazine layer, so apply its destructor and free it
2778 2818 * to the slab layer.
2779 2819 */
2780 2820 break;
2781 2821 }
2782 2822 }
2783 2823 mutex_exit(&ccp->cc_lock);
2784 2824 kmem_slab_free_constructed(cp, buf, B_TRUE);
2785 2825 }
2786 2826
2787 2827 static void
2788 2828 kmem_slab_prefill(kmem_cache_t *cp, kmem_slab_t *sp)
2789 2829 {
2790 2830 kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
2791 2831 int cache_flags = cp->cache_flags;
2792 2832
2793 2833 kmem_bufctl_t *next, *head;
2794 2834 size_t nbufs;
2795 2835
2796 2836 /*
2797 2837 * Completely allocate the newly created slab and put the pre-allocated
2798 2838 * buffers in magazines. Any of the buffers that cannot be put in
2799 2839 * magazines must be returned to the slab.
2800 2840 */
2801 2841 ASSERT(MUTEX_HELD(&cp->cache_lock));
2802 2842 ASSERT((cache_flags & (KMF_PREFILL|KMF_BUFTAG)) == KMF_PREFILL);
2803 2843 ASSERT(cp->cache_constructor == NULL);
2804 2844 ASSERT(sp->slab_cache == cp);
2805 2845 ASSERT(sp->slab_refcnt == 1);
2806 2846 ASSERT(sp->slab_head != NULL && sp->slab_chunks > sp->slab_refcnt);
2807 2847 ASSERT(avl_find(&cp->cache_partial_slabs, sp, NULL) == NULL);
2808 2848
2809 2849 head = sp->slab_head;
2810 2850 nbufs = (sp->slab_chunks - sp->slab_refcnt);
2811 2851 sp->slab_head = NULL;
2812 2852 sp->slab_refcnt += nbufs;
2813 2853 cp->cache_bufslab -= nbufs;
2814 2854 cp->cache_slab_alloc += nbufs;
2815 2855 list_insert_head(&cp->cache_complete_slabs, sp);
2816 2856 cp->cache_complete_slab_count++;
2817 2857 mutex_exit(&cp->cache_lock);
2818 2858 mutex_enter(&ccp->cc_lock);
2819 2859
2820 2860 while (head != NULL) {
2821 2861 void *buf = KMEM_BUF(cp, head);
2822 2862 /*
2823 2863 * If there's a slot available in the current CPU's
2824 2864 * loaded magazine, just put the object there and
2825 2865 * continue.
2826 2866 */
2827 2867 if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2828 2868 ccp->cc_loaded->mag_round[ccp->cc_rounds++] =
2829 2869 buf;
2830 2870 ccp->cc_free++;
2831 2871 nbufs--;
2832 2872 head = head->bc_next;
2833 2873 continue;
2834 2874 }
2835 2875
2836 2876 /*
2837 2877 * The loaded magazine is full. If the previously
2838 2878 * loaded magazine was empty, exchange them and try
2839 2879 * again.
2840 2880 */
2841 2881 if (ccp->cc_prounds == 0) {
2842 2882 kmem_cpu_reload(ccp, ccp->cc_ploaded,
2843 2883 ccp->cc_prounds);
2844 2884 continue;
2845 2885 }
2846 2886
2847 2887 /*
2848 2888 * If the magazine layer is disabled, break out now.
2849 2889 */
2850 2890
2851 2891 if (ccp->cc_magsize == 0) {
2852 2892 break;
2853 2893 }
2854 2894
2855 2895 if (!kmem_cpucache_magazine_alloc(ccp, cp))
2856 2896 break;
2857 2897 }
2858 2898 mutex_exit(&ccp->cc_lock);
2859 2899 if (nbufs != 0) {
2860 2900 ASSERT(head != NULL);
2861 2901
2862 2902 /*
2863 2903 * If there was a failure, return remaining objects to
2864 2904 * the slab
2865 2905 */
2866 2906 while (head != NULL) {
2867 2907 ASSERT(nbufs != 0);
2868 2908 next = head->bc_next;
2869 2909 head->bc_next = NULL;
2870 2910 kmem_slab_free(cp, KMEM_BUF(cp, head));
2871 2911 head = next;
2872 2912 nbufs--;
2873 2913 }
2874 2914 }
2875 2915 ASSERT(head == NULL);
2876 2916 ASSERT(nbufs == 0);
2877 2917 mutex_enter(&cp->cache_lock);
2878 2918 }
2879 2919
2880 2920 void *
2881 2921 kmem_zalloc(size_t size, int kmflag)
2882 2922 {
2883 2923 size_t index;
2884 2924 void *buf;
2885 2925
2886 2926 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2887 2927 kmem_cache_t *cp = kmem_alloc_table[index];
2888 2928 buf = kmem_cache_alloc(cp, kmflag);
2889 2929 if (buf != NULL) {
2890 2930 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
2891 2931 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2892 2932 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2893 2933 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2894 2934
2895 2935 if (cp->cache_flags & KMF_LITE) {
2896 2936 KMEM_BUFTAG_LITE_ENTER(btp,
2897 2937 kmem_lite_count, caller());
2898 2938 }
2899 2939 }
2900 2940 bzero(buf, size);
2901 2941 }
2902 2942 } else {
2903 2943 buf = kmem_alloc(size, kmflag);
2904 2944 if (buf != NULL)
2905 2945 bzero(buf, size);
2906 2946 }
2907 2947 return (buf);
2908 2948 }
2909 2949
2910 2950 void *
2911 2951 kmem_alloc(size_t size, int kmflag)
2912 2952 {
2913 2953 size_t index;
2914 2954 kmem_cache_t *cp;
2915 2955 void *buf;
2916 2956
|
↓ open down ↓ |
1013 lines elided |
↑ open up ↑ |
2917 2957 if ((index = ((size - 1) >> KMEM_ALIGN_SHIFT)) < KMEM_ALLOC_TABLE_MAX) {
2918 2958 cp = kmem_alloc_table[index];
2919 2959 /* fall through to kmem_cache_alloc() */
2920 2960
2921 2961 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2922 2962 kmem_big_alloc_table_max) {
2923 2963 cp = kmem_big_alloc_table[index];
2924 2964 /* fall through to kmem_cache_alloc() */
2925 2965
2926 2966 } else {
2927 - if (size == 0) {
2928 - if (kmflag != KM_SLEEP && !(kmflag & KM_PANIC))
2929 - return (NULL);
2930 -
2931 - /*
2932 - * If this is a sleeping allocation or one that has
2933 - * been specified to panic on allocation failure, we
2934 - * consider it to be deprecated behavior to allocate
2935 - * 0 bytes. If we have been configured to panic under
2936 - * this condition, we panic; if to warn, we warn -- and
2937 - * regardless, we log to the kmem_zerosized_log that
2938 - * that this condition has occurred (which gives us
2939 - * enough information to be able to debug it).
2940 - */
2941 - if (kmem_panic && kmem_panic_zerosized)
2942 - panic("attempted to kmem_alloc() size of 0");
2943 -
2944 - if (kmem_warn_zerosized) {
2945 - cmn_err(CE_WARN, "kmem_alloc(): sleeping "
2946 - "allocation with size of 0; "
2947 - "see kmem_zerosized_log for details");
2948 - }
2949 -
2950 - kmem_log_event(kmem_zerosized_log, NULL, NULL, NULL);
2951 -
2967 + if (size == 0)
2952 2968 return (NULL);
2953 - }
2954 2969
2955 2970 buf = vmem_alloc(kmem_oversize_arena, size,
2956 2971 kmflag & KM_VMFLAGS);
2957 2972 if (buf == NULL)
2958 2973 kmem_log_event(kmem_failure_log, NULL, NULL,
2959 2974 (void *)size);
2960 2975 else if (KMEM_DUMP(kmem_slab_cache)) {
2961 2976 /* stats for dump intercept */
2962 2977 kmem_dump_oversize_allocs++;
2963 2978 if (size > kmem_dump_oversize_max)
2964 2979 kmem_dump_oversize_max = size;
2965 2980 }
2966 2981 return (buf);
2967 2982 }
2968 2983
2969 2984 buf = kmem_cache_alloc(cp, kmflag);
2970 2985 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp) && buf != NULL) {
2971 2986 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
2972 2987 ((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
2973 2988 ((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
2974 2989
2975 2990 if (cp->cache_flags & KMF_LITE) {
2976 2991 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller());
2977 2992 }
2978 2993 }
2979 2994 return (buf);
2980 2995 }
2981 2996
2982 2997 void
2983 2998 kmem_free(void *buf, size_t size)
2984 2999 {
2985 3000 size_t index;
2986 3001 kmem_cache_t *cp;
2987 3002
2988 3003 if ((index = (size - 1) >> KMEM_ALIGN_SHIFT) < KMEM_ALLOC_TABLE_MAX) {
2989 3004 cp = kmem_alloc_table[index];
2990 3005 /* fall through to kmem_cache_free() */
2991 3006
2992 3007 } else if ((index = ((size - 1) >> KMEM_BIG_SHIFT)) <
2993 3008 kmem_big_alloc_table_max) {
2994 3009 cp = kmem_big_alloc_table[index];
2995 3010 /* fall through to kmem_cache_free() */
2996 3011
2997 3012 } else {
2998 3013 EQUIV(buf == NULL, size == 0);
2999 3014 if (buf == NULL && size == 0)
3000 3015 return;
3001 3016 vmem_free(kmem_oversize_arena, buf, size);
3002 3017 return;
3003 3018 }
3004 3019
3005 3020 if ((cp->cache_flags & KMF_BUFTAG) && !KMEM_DUMP(cp)) {
3006 3021 kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
3007 3022 uint32_t *ip = (uint32_t *)btp;
3008 3023 if (ip[1] != KMEM_SIZE_ENCODE(size)) {
3009 3024 if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
3010 3025 kmem_error(KMERR_DUPFREE, cp, buf);
3011 3026 return;
3012 3027 }
3013 3028 if (KMEM_SIZE_VALID(ip[1])) {
3014 3029 ip[0] = KMEM_SIZE_ENCODE(size);
3015 3030 kmem_error(KMERR_BADSIZE, cp, buf);
3016 3031 } else {
3017 3032 kmem_error(KMERR_REDZONE, cp, buf);
3018 3033 }
3019 3034 return;
3020 3035 }
3021 3036 if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
3022 3037 kmem_error(KMERR_REDZONE, cp, buf);
3023 3038 return;
3024 3039 }
3025 3040 btp->bt_redzone = KMEM_REDZONE_PATTERN;
3026 3041 if (cp->cache_flags & KMF_LITE) {
3027 3042 KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
3028 3043 caller());
3029 3044 }
3030 3045 }
3031 3046 kmem_cache_free(cp, buf);
3032 3047 }
3033 3048
3034 3049 void *
3035 3050 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
3036 3051 {
3037 3052 size_t realsize = size + vmp->vm_quantum;
3038 3053 void *addr;
3039 3054
3040 3055 /*
3041 3056 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
3042 3057 * vm_quantum will cause integer wraparound. Check for this, and
3043 3058 * blow off the firewall page in this case. Note that such a
3044 3059 * giant allocation (the entire kernel address space) can never
3045 3060 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
3046 3061 * or sleep forever (VM_SLEEP). Thus, there is no need for a
3047 3062 * corresponding check in kmem_firewall_va_free().
3048 3063 */
3049 3064 if (realsize < size)
3050 3065 realsize = size;
3051 3066
3052 3067 /*
3053 3068 * While boot still owns resource management, make sure that this
3054 3069 * redzone virtual address allocation is properly accounted for in
3055 3070 * OBPs "virtual-memory" "available" lists because we're
3056 3071 * effectively claiming them for a red zone. If we don't do this,
3057 3072 * the available lists become too fragmented and too large for the
3058 3073 * current boot/kernel memory list interface.
3059 3074 */
3060 3075 addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
3061 3076
3062 3077 if (addr != NULL && kvseg.s_base == NULL && realsize != size)
3063 3078 (void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
3064 3079
3065 3080 return (addr);
3066 3081 }
3067 3082
3068 3083 void
3069 3084 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
3070 3085 {
3071 3086 ASSERT((kvseg.s_base == NULL ?
3072 3087 va_to_pfn((char *)addr + size) :
3073 3088 hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
3074 3089
3075 3090 vmem_free(vmp, addr, size + vmp->vm_quantum);
3076 3091 }
3077 3092
3078 3093 /*
3079 3094 * Try to allocate at least `size' bytes of memory without sleeping or
3080 3095 * panicking. Return actual allocated size in `asize'. If allocation failed,
3081 3096 * try final allocation with sleep or panic allowed.
3082 3097 */
3083 3098 void *
3084 3099 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
3085 3100 {
3086 3101 void *p;
3087 3102
3088 3103 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3089 3104 do {
3090 3105 p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
3091 3106 if (p != NULL)
3092 3107 return (p);
3093 3108 *asize += KMEM_ALIGN;
3094 3109 } while (*asize <= PAGESIZE);
3095 3110
3096 3111 *asize = P2ROUNDUP(size, KMEM_ALIGN);
3097 3112 return (kmem_alloc(*asize, kmflag));
3098 3113 }
3099 3114
3100 3115 /*
3101 3116 * Reclaim all unused memory from a cache.
3102 3117 */
3103 3118 static void
3104 3119 kmem_cache_reap(kmem_cache_t *cp)
3105 3120 {
3106 3121 ASSERT(taskq_member(kmem_taskq, curthread));
3107 3122 cp->cache_reap++;
3108 3123
3109 3124 /*
3110 3125 * Ask the cache's owner to free some memory if possible.
3111 3126 * The idea is to handle things like the inode cache, which
3112 3127 * typically sits on a bunch of memory that it doesn't truly
3113 3128 * *need*. Reclaim policy is entirely up to the owner; this
3114 3129 * callback is just an advisory plea for help.
3115 3130 */
3116 3131 if (cp->cache_reclaim != NULL) {
3117 3132 long delta;
3118 3133
3119 3134 /*
3120 3135 * Reclaimed memory should be reapable (not included in the
3121 3136 * depot's working set).
3122 3137 */
3123 3138 delta = cp->cache_full.ml_total;
3124 3139 cp->cache_reclaim(cp->cache_private);
3125 3140 delta = cp->cache_full.ml_total - delta;
3126 3141 if (delta > 0) {
3127 3142 mutex_enter(&cp->cache_depot_lock);
3128 3143 cp->cache_full.ml_reaplimit += delta;
3129 3144 cp->cache_full.ml_min += delta;
3130 3145 mutex_exit(&cp->cache_depot_lock);
3131 3146 }
3132 3147 }
3133 3148
3134 3149 kmem_depot_ws_reap(cp);
3135 3150
3136 3151 if (cp->cache_defrag != NULL && !kmem_move_noreap) {
3137 3152 kmem_cache_defrag(cp);
3138 3153 }
3139 3154 }
3140 3155
3141 3156 static void
3142 3157 kmem_reap_timeout(void *flag_arg)
3143 3158 {
3144 3159 uint32_t *flag = (uint32_t *)flag_arg;
3145 3160
3146 3161 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3147 3162 *flag = 0;
3148 3163 }
3149 3164
3150 3165 static void
3151 3166 kmem_reap_done(void *flag)
3152 3167 {
3153 3168 if (!callout_init_done) {
3154 3169 /* can't schedule a timeout at this point */
3155 3170 kmem_reap_timeout(flag);
3156 3171 } else {
3157 3172 (void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
3158 3173 }
3159 3174 }
3160 3175
3161 3176 static void
3162 3177 kmem_reap_start(void *flag)
3163 3178 {
3164 3179 ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
3165 3180
3166 3181 if (flag == &kmem_reaping) {
3167 3182 kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3168 3183 /*
3169 3184 * if we have segkp under heap, reap segkp cache.
3170 3185 */
3171 3186 if (segkp_fromheap)
3172 3187 segkp_cache_free();
3173 3188 }
3174 3189 else
3175 3190 kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
3176 3191
3177 3192 /*
3178 3193 * We use taskq_dispatch() to schedule a timeout to clear
3179 3194 * the flag so that kmem_reap() becomes self-throttling:
3180 3195 * we won't reap again until the current reap completes *and*
3181 3196 * at least kmem_reap_interval ticks have elapsed.
3182 3197 */
3183 3198 if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
3184 3199 kmem_reap_done(flag);
3185 3200 }
3186 3201
3187 3202 static void
3188 3203 kmem_reap_common(void *flag_arg)
3189 3204 {
3190 3205 uint32_t *flag = (uint32_t *)flag_arg;
3191 3206
3192 3207 if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
3193 3208 atomic_cas_32(flag, 0, 1) != 0)
3194 3209 return;
3195 3210
3196 3211 /*
3197 3212 * It may not be kosher to do memory allocation when a reap is called
3198 3213 * (for example, if vmem_populate() is in the call chain). So we
3199 3214 * start the reap going with a TQ_NOALLOC dispatch. If the dispatch
3200 3215 * fails, we reset the flag, and the next reap will try again.
3201 3216 */
3202 3217 if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
3203 3218 *flag = 0;
3204 3219 }
3205 3220
3206 3221 /*
3207 3222 * Reclaim all unused memory from all caches. Called from the VM system
3208 3223 * when memory gets tight.
3209 3224 */
3210 3225 void
3211 3226 kmem_reap(void)
3212 3227 {
3213 3228 kmem_reap_common(&kmem_reaping);
3214 3229 }
3215 3230
3216 3231 /*
3217 3232 * Reclaim all unused memory from identifier arenas, called when a vmem
3218 3233 * arena not back by memory is exhausted. Since reaping memory-backed caches
3219 3234 * cannot help with identifier exhaustion, we avoid both a large amount of
3220 3235 * work and unwanted side-effects from reclaim callbacks.
3221 3236 */
3222 3237 void
3223 3238 kmem_reap_idspace(void)
3224 3239 {
3225 3240 kmem_reap_common(&kmem_reaping_idspace);
3226 3241 }
3227 3242
3228 3243 /*
3229 3244 * Purge all magazines from a cache and set its magazine limit to zero.
3230 3245 * All calls are serialized by the kmem_taskq lock, except for the final
3231 3246 * call from kmem_cache_destroy().
3232 3247 */
3233 3248 static void
3234 3249 kmem_cache_magazine_purge(kmem_cache_t *cp)
3235 3250 {
3236 3251 kmem_cpu_cache_t *ccp;
3237 3252 kmem_magazine_t *mp, *pmp;
3238 3253 int rounds, prounds, cpu_seqid;
3239 3254
3240 3255 ASSERT(!list_link_active(&cp->cache_link) ||
3241 3256 taskq_member(kmem_taskq, curthread));
3242 3257 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
3243 3258
3244 3259 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3245 3260 ccp = &cp->cache_cpu[cpu_seqid];
3246 3261
3247 3262 mutex_enter(&ccp->cc_lock);
3248 3263 mp = ccp->cc_loaded;
3249 3264 pmp = ccp->cc_ploaded;
3250 3265 rounds = ccp->cc_rounds;
3251 3266 prounds = ccp->cc_prounds;
3252 3267 ccp->cc_loaded = NULL;
3253 3268 ccp->cc_ploaded = NULL;
3254 3269 ccp->cc_rounds = -1;
3255 3270 ccp->cc_prounds = -1;
3256 3271 ccp->cc_magsize = 0;
3257 3272 mutex_exit(&ccp->cc_lock);
3258 3273
3259 3274 if (mp)
3260 3275 kmem_magazine_destroy(cp, mp, rounds);
3261 3276 if (pmp)
3262 3277 kmem_magazine_destroy(cp, pmp, prounds);
3263 3278 }
3264 3279
3265 3280 kmem_depot_ws_zero(cp);
3266 3281 kmem_depot_ws_reap(cp);
3267 3282 }
3268 3283
3269 3284 /*
3270 3285 * Enable per-cpu magazines on a cache.
3271 3286 */
3272 3287 static void
3273 3288 kmem_cache_magazine_enable(kmem_cache_t *cp)
3274 3289 {
3275 3290 int cpu_seqid;
3276 3291
3277 3292 if (cp->cache_flags & KMF_NOMAGAZINE)
3278 3293 return;
3279 3294
3280 3295 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3281 3296 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3282 3297 mutex_enter(&ccp->cc_lock);
3283 3298 ccp->cc_magsize = cp->cache_magtype->mt_magsize;
3284 3299 mutex_exit(&ccp->cc_lock);
3285 3300 }
3286 3301
3287 3302 }
3288 3303
3289 3304 /*
3290 3305 * Reap (almost) everything right now.
3291 3306 */
3292 3307 void
3293 3308 kmem_cache_reap_now(kmem_cache_t *cp)
3294 3309 {
3295 3310 ASSERT(list_link_active(&cp->cache_link));
3296 3311
3297 3312 kmem_depot_ws_zero(cp);
3298 3313
3299 3314 (void) taskq_dispatch(kmem_taskq,
3300 3315 (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
3301 3316 taskq_wait(kmem_taskq);
3302 3317 }
3303 3318
3304 3319 /*
3305 3320 * Recompute a cache's magazine size. The trade-off is that larger magazines
3306 3321 * provide a higher transfer rate with the depot, while smaller magazines
3307 3322 * reduce memory consumption. Magazine resizing is an expensive operation;
3308 3323 * it should not be done frequently.
3309 3324 *
3310 3325 * Changes to the magazine size are serialized by the kmem_taskq lock.
3311 3326 *
3312 3327 * Note: at present this only grows the magazine size. It might be useful
3313 3328 * to allow shrinkage too.
3314 3329 */
3315 3330 static void
3316 3331 kmem_cache_magazine_resize(kmem_cache_t *cp)
3317 3332 {
3318 3333 kmem_magtype_t *mtp = cp->cache_magtype;
3319 3334
3320 3335 ASSERT(taskq_member(kmem_taskq, curthread));
3321 3336
3322 3337 if (cp->cache_chunksize < mtp->mt_maxbuf) {
3323 3338 kmem_cache_magazine_purge(cp);
3324 3339 mutex_enter(&cp->cache_depot_lock);
3325 3340 cp->cache_magtype = ++mtp;
3326 3341 cp->cache_depot_contention_prev =
3327 3342 cp->cache_depot_contention + INT_MAX;
3328 3343 mutex_exit(&cp->cache_depot_lock);
3329 3344 kmem_cache_magazine_enable(cp);
3330 3345 }
3331 3346 }
3332 3347
3333 3348 /*
3334 3349 * Rescale a cache's hash table, so that the table size is roughly the
3335 3350 * cache size. We want the average lookup time to be extremely small.
3336 3351 */
3337 3352 static void
3338 3353 kmem_hash_rescale(kmem_cache_t *cp)
3339 3354 {
3340 3355 kmem_bufctl_t **old_table, **new_table, *bcp;
3341 3356 size_t old_size, new_size, h;
3342 3357
3343 3358 ASSERT(taskq_member(kmem_taskq, curthread));
3344 3359
3345 3360 new_size = MAX(KMEM_HASH_INITIAL,
3346 3361 1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
3347 3362 old_size = cp->cache_hash_mask + 1;
3348 3363
3349 3364 if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
3350 3365 return;
3351 3366
3352 3367 new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
3353 3368 VM_NOSLEEP);
3354 3369 if (new_table == NULL)
3355 3370 return;
3356 3371 bzero(new_table, new_size * sizeof (void *));
3357 3372
3358 3373 mutex_enter(&cp->cache_lock);
3359 3374
3360 3375 old_size = cp->cache_hash_mask + 1;
3361 3376 old_table = cp->cache_hash_table;
3362 3377
3363 3378 cp->cache_hash_mask = new_size - 1;
3364 3379 cp->cache_hash_table = new_table;
3365 3380 cp->cache_rescale++;
3366 3381
3367 3382 for (h = 0; h < old_size; h++) {
3368 3383 bcp = old_table[h];
3369 3384 while (bcp != NULL) {
3370 3385 void *addr = bcp->bc_addr;
3371 3386 kmem_bufctl_t *next_bcp = bcp->bc_next;
3372 3387 kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
3373 3388 bcp->bc_next = *hash_bucket;
3374 3389 *hash_bucket = bcp;
3375 3390 bcp = next_bcp;
3376 3391 }
3377 3392 }
3378 3393
3379 3394 mutex_exit(&cp->cache_lock);
3380 3395
3381 3396 vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
3382 3397 }
3383 3398
3384 3399 /*
3385 3400 * Perform periodic maintenance on a cache: hash rescaling, depot working-set
3386 3401 * update, magazine resizing, and slab consolidation.
3387 3402 */
3388 3403 static void
3389 3404 kmem_cache_update(kmem_cache_t *cp)
3390 3405 {
3391 3406 int need_hash_rescale = 0;
3392 3407 int need_magazine_resize = 0;
3393 3408
3394 3409 ASSERT(MUTEX_HELD(&kmem_cache_lock));
3395 3410
3396 3411 /*
3397 3412 * If the cache has become much larger or smaller than its hash table,
3398 3413 * fire off a request to rescale the hash table.
3399 3414 */
3400 3415 mutex_enter(&cp->cache_lock);
3401 3416
3402 3417 if ((cp->cache_flags & KMF_HASH) &&
3403 3418 (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
3404 3419 (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
3405 3420 cp->cache_hash_mask > KMEM_HASH_INITIAL)))
3406 3421 need_hash_rescale = 1;
3407 3422
3408 3423 mutex_exit(&cp->cache_lock);
3409 3424
3410 3425 /*
3411 3426 * Update the depot working set statistics.
3412 3427 */
3413 3428 kmem_depot_ws_update(cp);
3414 3429
3415 3430 /*
3416 3431 * If there's a lot of contention in the depot,
3417 3432 * increase the magazine size.
3418 3433 */
3419 3434 mutex_enter(&cp->cache_depot_lock);
3420 3435
3421 3436 if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
3422 3437 (int)(cp->cache_depot_contention -
3423 3438 cp->cache_depot_contention_prev) > kmem_depot_contention)
3424 3439 need_magazine_resize = 1;
3425 3440
3426 3441 cp->cache_depot_contention_prev = cp->cache_depot_contention;
3427 3442
3428 3443 mutex_exit(&cp->cache_depot_lock);
3429 3444
3430 3445 if (need_hash_rescale)
3431 3446 (void) taskq_dispatch(kmem_taskq,
3432 3447 (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
3433 3448
3434 3449 if (need_magazine_resize)
3435 3450 (void) taskq_dispatch(kmem_taskq,
3436 3451 (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
3437 3452
3438 3453 if (cp->cache_defrag != NULL)
3439 3454 (void) taskq_dispatch(kmem_taskq,
3440 3455 (task_func_t *)kmem_cache_scan, cp, TQ_NOSLEEP);
3441 3456 }
3442 3457
3443 3458 static void kmem_update(void *);
3444 3459
3445 3460 static void
3446 3461 kmem_update_timeout(void *dummy)
3447 3462 {
3448 3463 (void) timeout(kmem_update, dummy, kmem_reap_interval);
3449 3464 }
3450 3465
3451 3466 static void
3452 3467 kmem_update(void *dummy)
3453 3468 {
3454 3469 kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
3455 3470
3456 3471 /*
3457 3472 * We use taskq_dispatch() to reschedule the timeout so that
3458 3473 * kmem_update() becomes self-throttling: it won't schedule
3459 3474 * new tasks until all previous tasks have completed.
3460 3475 */
3461 3476 if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
3462 3477 kmem_update_timeout(NULL);
3463 3478 }
3464 3479
3465 3480 static int
3466 3481 kmem_cache_kstat_update(kstat_t *ksp, int rw)
3467 3482 {
3468 3483 struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
3469 3484 kmem_cache_t *cp = ksp->ks_private;
3470 3485 uint64_t cpu_buf_avail;
3471 3486 uint64_t buf_avail = 0;
3472 3487 int cpu_seqid;
3473 3488 long reap;
3474 3489
3475 3490 ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
3476 3491
3477 3492 if (rw == KSTAT_WRITE)
3478 3493 return (EACCES);
3479 3494
3480 3495 mutex_enter(&cp->cache_lock);
3481 3496
3482 3497 kmcp->kmc_alloc_fail.value.ui64 = cp->cache_alloc_fail;
3483 3498 kmcp->kmc_alloc.value.ui64 = cp->cache_slab_alloc;
3484 3499 kmcp->kmc_free.value.ui64 = cp->cache_slab_free;
3485 3500 kmcp->kmc_slab_alloc.value.ui64 = cp->cache_slab_alloc;
3486 3501 kmcp->kmc_slab_free.value.ui64 = cp->cache_slab_free;
3487 3502
3488 3503 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3489 3504 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3490 3505
3491 3506 mutex_enter(&ccp->cc_lock);
3492 3507
3493 3508 cpu_buf_avail = 0;
3494 3509 if (ccp->cc_rounds > 0)
3495 3510 cpu_buf_avail += ccp->cc_rounds;
3496 3511 if (ccp->cc_prounds > 0)
3497 3512 cpu_buf_avail += ccp->cc_prounds;
3498 3513
3499 3514 kmcp->kmc_alloc.value.ui64 += ccp->cc_alloc;
3500 3515 kmcp->kmc_free.value.ui64 += ccp->cc_free;
3501 3516 buf_avail += cpu_buf_avail;
3502 3517
3503 3518 mutex_exit(&ccp->cc_lock);
3504 3519 }
3505 3520
3506 3521 mutex_enter(&cp->cache_depot_lock);
3507 3522
3508 3523 kmcp->kmc_depot_alloc.value.ui64 = cp->cache_full.ml_alloc;
3509 3524 kmcp->kmc_depot_free.value.ui64 = cp->cache_empty.ml_alloc;
3510 3525 kmcp->kmc_depot_contention.value.ui64 = cp->cache_depot_contention;
3511 3526 kmcp->kmc_full_magazines.value.ui64 = cp->cache_full.ml_total;
3512 3527 kmcp->kmc_empty_magazines.value.ui64 = cp->cache_empty.ml_total;
3513 3528 kmcp->kmc_magazine_size.value.ui64 =
3514 3529 (cp->cache_flags & KMF_NOMAGAZINE) ?
3515 3530 0 : cp->cache_magtype->mt_magsize;
3516 3531
3517 3532 kmcp->kmc_alloc.value.ui64 += cp->cache_full.ml_alloc;
3518 3533 kmcp->kmc_free.value.ui64 += cp->cache_empty.ml_alloc;
3519 3534 buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
3520 3535
3521 3536 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
3522 3537 reap = MIN(reap, cp->cache_full.ml_total);
3523 3538
3524 3539 mutex_exit(&cp->cache_depot_lock);
3525 3540
3526 3541 kmcp->kmc_buf_size.value.ui64 = cp->cache_bufsize;
3527 3542 kmcp->kmc_align.value.ui64 = cp->cache_align;
3528 3543 kmcp->kmc_chunk_size.value.ui64 = cp->cache_chunksize;
3529 3544 kmcp->kmc_slab_size.value.ui64 = cp->cache_slabsize;
3530 3545 kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
3531 3546 buf_avail += cp->cache_bufslab;
3532 3547 kmcp->kmc_buf_avail.value.ui64 = buf_avail;
3533 3548 kmcp->kmc_buf_inuse.value.ui64 = cp->cache_buftotal - buf_avail;
3534 3549 kmcp->kmc_buf_total.value.ui64 = cp->cache_buftotal;
3535 3550 kmcp->kmc_buf_max.value.ui64 = cp->cache_bufmax;
3536 3551 kmcp->kmc_slab_create.value.ui64 = cp->cache_slab_create;
3537 3552 kmcp->kmc_slab_destroy.value.ui64 = cp->cache_slab_destroy;
3538 3553 kmcp->kmc_hash_size.value.ui64 = (cp->cache_flags & KMF_HASH) ?
3539 3554 cp->cache_hash_mask + 1 : 0;
3540 3555 kmcp->kmc_hash_lookup_depth.value.ui64 = cp->cache_lookup_depth;
3541 3556 kmcp->kmc_hash_rescale.value.ui64 = cp->cache_rescale;
3542 3557 kmcp->kmc_vmem_source.value.ui64 = cp->cache_arena->vm_id;
3543 3558 kmcp->kmc_reap.value.ui64 = cp->cache_reap;
3544 3559
3545 3560 if (cp->cache_defrag == NULL) {
3546 3561 kmcp->kmc_move_callbacks.value.ui64 = 0;
3547 3562 kmcp->kmc_move_yes.value.ui64 = 0;
3548 3563 kmcp->kmc_move_no.value.ui64 = 0;
3549 3564 kmcp->kmc_move_later.value.ui64 = 0;
3550 3565 kmcp->kmc_move_dont_need.value.ui64 = 0;
3551 3566 kmcp->kmc_move_dont_know.value.ui64 = 0;
3552 3567 kmcp->kmc_move_hunt_found.value.ui64 = 0;
3553 3568 kmcp->kmc_move_slabs_freed.value.ui64 = 0;
3554 3569 kmcp->kmc_defrag.value.ui64 = 0;
3555 3570 kmcp->kmc_scan.value.ui64 = 0;
3556 3571 kmcp->kmc_move_reclaimable.value.ui64 = 0;
|
↓ open down ↓ |
593 lines elided |
↑ open up ↑ |
3557 3572 } else {
3558 3573 int64_t reclaimable;
3559 3574
3560 3575 kmem_defrag_t *kd = cp->cache_defrag;
3561 3576 kmcp->kmc_move_callbacks.value.ui64 = kd->kmd_callbacks;
3562 3577 kmcp->kmc_move_yes.value.ui64 = kd->kmd_yes;
3563 3578 kmcp->kmc_move_no.value.ui64 = kd->kmd_no;
3564 3579 kmcp->kmc_move_later.value.ui64 = kd->kmd_later;
3565 3580 kmcp->kmc_move_dont_need.value.ui64 = kd->kmd_dont_need;
3566 3581 kmcp->kmc_move_dont_know.value.ui64 = kd->kmd_dont_know;
3567 - kmcp->kmc_move_hunt_found.value.ui64 = 0;
3582 + kmcp->kmc_move_hunt_found.value.ui64 = kd->kmd_hunt_found;
3568 3583 kmcp->kmc_move_slabs_freed.value.ui64 = kd->kmd_slabs_freed;
3569 3584 kmcp->kmc_defrag.value.ui64 = kd->kmd_defrags;
3570 3585 kmcp->kmc_scan.value.ui64 = kd->kmd_scans;
3571 3586
3572 3587 reclaimable = cp->cache_bufslab - (cp->cache_maxchunks - 1);
3573 3588 reclaimable = MAX(reclaimable, 0);
3574 3589 reclaimable += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
3575 3590 kmcp->kmc_move_reclaimable.value.ui64 = reclaimable;
3576 3591 }
3577 3592
3578 3593 mutex_exit(&cp->cache_lock);
3579 3594 return (0);
3580 3595 }
3581 3596
3582 3597 /*
3583 3598 * Return a named statistic about a particular cache.
3584 3599 * This shouldn't be called very often, so it's currently designed for
3585 3600 * simplicity (leverages existing kstat support) rather than efficiency.
3586 3601 */
3587 3602 uint64_t
3588 3603 kmem_cache_stat(kmem_cache_t *cp, char *name)
3589 3604 {
3590 3605 int i;
3591 3606 kstat_t *ksp = cp->cache_kstat;
3592 3607 kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
3593 3608 uint64_t value = 0;
3594 3609
3595 3610 if (ksp != NULL) {
3596 3611 mutex_enter(&kmem_cache_kstat_lock);
3597 3612 (void) kmem_cache_kstat_update(ksp, KSTAT_READ);
3598 3613 for (i = 0; i < ksp->ks_ndata; i++) {
3599 3614 if (strcmp(knp[i].name, name) == 0) {
3600 3615 value = knp[i].value.ui64;
3601 3616 break;
3602 3617 }
3603 3618 }
3604 3619 mutex_exit(&kmem_cache_kstat_lock);
3605 3620 }
3606 3621 return (value);
3607 3622 }
3608 3623
3609 3624 /*
3610 3625 * Return an estimate of currently available kernel heap memory.
3611 3626 * On 32-bit systems, physical memory may exceed virtual memory,
3612 3627 * we just truncate the result at 1GB.
3613 3628 */
3614 3629 size_t
3615 3630 kmem_avail(void)
3616 3631 {
3617 3632 spgcnt_t rmem = availrmem - tune.t_minarmem;
3618 3633 spgcnt_t fmem = freemem - minfree;
3619 3634
3620 3635 return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
3621 3636 1 << (30 - PAGESHIFT))));
3622 3637 }
3623 3638
3624 3639 /*
3625 3640 * Return the maximum amount of memory that is (in theory) allocatable
3626 3641 * from the heap. This may be used as an estimate only since there
3627 3642 * is no guarentee this space will still be available when an allocation
3628 3643 * request is made, nor that the space may be allocated in one big request
3629 3644 * due to kernel heap fragmentation.
3630 3645 */
3631 3646 size_t
3632 3647 kmem_maxavail(void)
3633 3648 {
3634 3649 spgcnt_t pmem = availrmem - tune.t_minarmem;
3635 3650 spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
3636 3651
3637 3652 return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
3638 3653 }
3639 3654
3640 3655 /*
3641 3656 * Indicate whether memory-intensive kmem debugging is enabled.
3642 3657 */
3643 3658 int
3644 3659 kmem_debugging(void)
3645 3660 {
3646 3661 return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
3647 3662 }
3648 3663
3649 3664 /* binning function, sorts finely at the two extremes */
3650 3665 #define KMEM_PARTIAL_SLAB_WEIGHT(sp, binshift) \
3651 3666 ((((sp)->slab_refcnt <= (binshift)) || \
3652 3667 (((sp)->slab_chunks - (sp)->slab_refcnt) <= (binshift))) \
3653 3668 ? -(sp)->slab_refcnt \
3654 3669 : -((binshift) + ((sp)->slab_refcnt >> (binshift))))
3655 3670
3656 3671 /*
3657 3672 * Minimizing the number of partial slabs on the freelist minimizes
3658 3673 * fragmentation (the ratio of unused buffers held by the slab layer). There are
3659 3674 * two ways to get a slab off of the freelist: 1) free all the buffers on the
3660 3675 * slab, and 2) allocate all the buffers on the slab. It follows that we want
3661 3676 * the most-used slabs at the front of the list where they have the best chance
3662 3677 * of being completely allocated, and the least-used slabs at a safe distance
3663 3678 * from the front to improve the odds that the few remaining buffers will all be
3664 3679 * freed before another allocation can tie up the slab. For that reason a slab
3665 3680 * with a higher slab_refcnt sorts less than than a slab with a lower
3666 3681 * slab_refcnt.
3667 3682 *
3668 3683 * However, if a slab has at least one buffer that is deemed unfreeable, we
3669 3684 * would rather have that slab at the front of the list regardless of
3670 3685 * slab_refcnt, since even one unfreeable buffer makes the entire slab
3671 3686 * unfreeable. If the client returns KMEM_CBRC_NO in response to a cache_move()
3672 3687 * callback, the slab is marked unfreeable for as long as it remains on the
3673 3688 * freelist.
3674 3689 */
3675 3690 static int
3676 3691 kmem_partial_slab_cmp(const void *p0, const void *p1)
3677 3692 {
3678 3693 const kmem_cache_t *cp;
3679 3694 const kmem_slab_t *s0 = p0;
3680 3695 const kmem_slab_t *s1 = p1;
3681 3696 int w0, w1;
3682 3697 size_t binshift;
3683 3698
3684 3699 ASSERT(KMEM_SLAB_IS_PARTIAL(s0));
3685 3700 ASSERT(KMEM_SLAB_IS_PARTIAL(s1));
3686 3701 ASSERT(s0->slab_cache == s1->slab_cache);
3687 3702 cp = s1->slab_cache;
3688 3703 ASSERT(MUTEX_HELD(&cp->cache_lock));
3689 3704 binshift = cp->cache_partial_binshift;
3690 3705
3691 3706 /* weight of first slab */
3692 3707 w0 = KMEM_PARTIAL_SLAB_WEIGHT(s0, binshift);
3693 3708 if (s0->slab_flags & KMEM_SLAB_NOMOVE) {
3694 3709 w0 -= cp->cache_maxchunks;
3695 3710 }
3696 3711
3697 3712 /* weight of second slab */
3698 3713 w1 = KMEM_PARTIAL_SLAB_WEIGHT(s1, binshift);
3699 3714 if (s1->slab_flags & KMEM_SLAB_NOMOVE) {
3700 3715 w1 -= cp->cache_maxchunks;
3701 3716 }
3702 3717
3703 3718 if (w0 < w1)
3704 3719 return (-1);
3705 3720 if (w0 > w1)
3706 3721 return (1);
3707 3722
3708 3723 /* compare pointer values */
3709 3724 if ((uintptr_t)s0 < (uintptr_t)s1)
3710 3725 return (-1);
3711 3726 if ((uintptr_t)s0 > (uintptr_t)s1)
3712 3727 return (1);
3713 3728
3714 3729 return (0);
3715 3730 }
3716 3731
3717 3732 /*
3718 3733 * It must be valid to call the destructor (if any) on a newly created object.
3719 3734 * That is, the constructor (if any) must leave the object in a valid state for
3720 3735 * the destructor.
3721 3736 */
3722 3737 kmem_cache_t *
3723 3738 kmem_cache_create(
3724 3739 char *name, /* descriptive name for this cache */
3725 3740 size_t bufsize, /* size of the objects it manages */
3726 3741 size_t align, /* required object alignment */
3727 3742 int (*constructor)(void *, void *, int), /* object constructor */
3728 3743 void (*destructor)(void *, void *), /* object destructor */
3729 3744 void (*reclaim)(void *), /* memory reclaim callback */
3730 3745 void *private, /* pass-thru arg for constr/destr/reclaim */
3731 3746 vmem_t *vmp, /* vmem source for slab allocation */
3732 3747 int cflags) /* cache creation flags */
3733 3748 {
3734 3749 int cpu_seqid;
3735 3750 size_t chunksize;
3736 3751 kmem_cache_t *cp;
3737 3752 kmem_magtype_t *mtp;
3738 3753 size_t csize = KMEM_CACHE_SIZE(max_ncpus);
3739 3754
3740 3755 #ifdef DEBUG
3741 3756 /*
3742 3757 * Cache names should conform to the rules for valid C identifiers
3743 3758 */
3744 3759 if (!strident_valid(name)) {
3745 3760 cmn_err(CE_CONT,
3746 3761 "kmem_cache_create: '%s' is an invalid cache name\n"
3747 3762 "cache names must conform to the rules for "
3748 3763 "C identifiers\n", name);
3749 3764 }
3750 3765 #endif /* DEBUG */
3751 3766
3752 3767 if (vmp == NULL)
3753 3768 vmp = kmem_default_arena;
3754 3769
3755 3770 /*
3756 3771 * If this kmem cache has an identifier vmem arena as its source, mark
3757 3772 * it such to allow kmem_reap_idspace().
3758 3773 */
3759 3774 ASSERT(!(cflags & KMC_IDENTIFIER)); /* consumer should not set this */
3760 3775 if (vmp->vm_cflags & VMC_IDENTIFIER)
3761 3776 cflags |= KMC_IDENTIFIER;
3762 3777
3763 3778 /*
3764 3779 * Get a kmem_cache structure. We arrange that cp->cache_cpu[]
3765 3780 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
3766 3781 * false sharing of per-CPU data.
3767 3782 */
3768 3783 cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
3769 3784 P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
3770 3785 bzero(cp, csize);
3771 3786 list_link_init(&cp->cache_link);
3772 3787
3773 3788 if (align == 0)
3774 3789 align = KMEM_ALIGN;
3775 3790
3776 3791 /*
3777 3792 * If we're not at least KMEM_ALIGN aligned, we can't use free
3778 3793 * memory to hold bufctl information (because we can't safely
3779 3794 * perform word loads and stores on it).
3780 3795 */
3781 3796 if (align < KMEM_ALIGN)
3782 3797 cflags |= KMC_NOTOUCH;
3783 3798
3784 3799 if (!ISP2(align) || align > vmp->vm_quantum)
3785 3800 panic("kmem_cache_create: bad alignment %lu", align);
3786 3801
3787 3802 mutex_enter(&kmem_flags_lock);
3788 3803 if (kmem_flags & KMF_RANDOMIZE)
3789 3804 kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
3790 3805 KMF_RANDOMIZE;
3791 3806 cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
3792 3807 mutex_exit(&kmem_flags_lock);
3793 3808
3794 3809 /*
3795 3810 * Make sure all the various flags are reasonable.
3796 3811 */
3797 3812 ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
3798 3813
3799 3814 if (cp->cache_flags & KMF_LITE) {
3800 3815 if (bufsize >= kmem_lite_minsize &&
3801 3816 align <= kmem_lite_maxalign &&
3802 3817 P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
3803 3818 cp->cache_flags |= KMF_BUFTAG;
3804 3819 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3805 3820 } else {
3806 3821 cp->cache_flags &= ~KMF_DEBUG;
3807 3822 }
3808 3823 }
3809 3824
3810 3825 if (cp->cache_flags & KMF_DEADBEEF)
3811 3826 cp->cache_flags |= KMF_REDZONE;
3812 3827
3813 3828 if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
3814 3829 cp->cache_flags |= KMF_NOMAGAZINE;
3815 3830
3816 3831 if (cflags & KMC_NODEBUG)
3817 3832 cp->cache_flags &= ~KMF_DEBUG;
3818 3833
3819 3834 if (cflags & KMC_NOTOUCH)
3820 3835 cp->cache_flags &= ~KMF_TOUCH;
3821 3836
3822 3837 if (cflags & KMC_PREFILL)
3823 3838 cp->cache_flags |= KMF_PREFILL;
3824 3839
3825 3840 if (cflags & KMC_NOHASH)
3826 3841 cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
3827 3842
3828 3843 if (cflags & KMC_NOMAGAZINE)
3829 3844 cp->cache_flags |= KMF_NOMAGAZINE;
3830 3845
3831 3846 if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
3832 3847 cp->cache_flags |= KMF_REDZONE;
3833 3848
3834 3849 if (!(cp->cache_flags & KMF_AUDIT))
3835 3850 cp->cache_flags &= ~KMF_CONTENTS;
3836 3851
3837 3852 if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
3838 3853 !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
3839 3854 cp->cache_flags |= KMF_FIREWALL;
3840 3855
3841 3856 if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
3842 3857 cp->cache_flags &= ~KMF_FIREWALL;
3843 3858
3844 3859 if (cp->cache_flags & KMF_FIREWALL) {
3845 3860 cp->cache_flags &= ~KMF_BUFTAG;
3846 3861 cp->cache_flags |= KMF_NOMAGAZINE;
3847 3862 ASSERT(vmp == kmem_default_arena);
3848 3863 vmp = kmem_firewall_arena;
3849 3864 }
3850 3865
3851 3866 /*
3852 3867 * Set cache properties.
3853 3868 */
3854 3869 (void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
3855 3870 strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN + 1);
3856 3871 cp->cache_bufsize = bufsize;
3857 3872 cp->cache_align = align;
3858 3873 cp->cache_constructor = constructor;
3859 3874 cp->cache_destructor = destructor;
3860 3875 cp->cache_reclaim = reclaim;
3861 3876 cp->cache_private = private;
3862 3877 cp->cache_arena = vmp;
3863 3878 cp->cache_cflags = cflags;
3864 3879
3865 3880 /*
3866 3881 * Determine the chunk size.
3867 3882 */
3868 3883 chunksize = bufsize;
3869 3884
3870 3885 if (align >= KMEM_ALIGN) {
3871 3886 chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
3872 3887 cp->cache_bufctl = chunksize - KMEM_ALIGN;
3873 3888 }
3874 3889
3875 3890 if (cp->cache_flags & KMF_BUFTAG) {
3876 3891 cp->cache_bufctl = chunksize;
3877 3892 cp->cache_buftag = chunksize;
3878 3893 if (cp->cache_flags & KMF_LITE)
3879 3894 chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
3880 3895 else
3881 3896 chunksize += sizeof (kmem_buftag_t);
3882 3897 }
3883 3898
3884 3899 if (cp->cache_flags & KMF_DEADBEEF) {
3885 3900 cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
3886 3901 if (cp->cache_flags & KMF_LITE)
3887 3902 cp->cache_verify = sizeof (uint64_t);
3888 3903 }
3889 3904
3890 3905 cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
3891 3906
3892 3907 cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
3893 3908
3894 3909 /*
3895 3910 * Now that we know the chunk size, determine the optimal slab size.
3896 3911 */
3897 3912 if (vmp == kmem_firewall_arena) {
3898 3913 cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
3899 3914 cp->cache_mincolor = cp->cache_slabsize - chunksize;
3900 3915 cp->cache_maxcolor = cp->cache_mincolor;
3901 3916 cp->cache_flags |= KMF_HASH;
3902 3917 ASSERT(!(cp->cache_flags & KMF_BUFTAG));
3903 3918 } else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
3904 3919 !(cp->cache_flags & KMF_AUDIT) &&
3905 3920 chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
3906 3921 cp->cache_slabsize = vmp->vm_quantum;
3907 3922 cp->cache_mincolor = 0;
3908 3923 cp->cache_maxcolor =
3909 3924 (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
3910 3925 ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
3911 3926 ASSERT(!(cp->cache_flags & KMF_AUDIT));
3912 3927 } else {
3913 3928 size_t chunks, bestfit, waste, slabsize;
3914 3929 size_t minwaste = LONG_MAX;
3915 3930
3916 3931 for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
3917 3932 slabsize = P2ROUNDUP(chunksize * chunks,
3918 3933 vmp->vm_quantum);
3919 3934 chunks = slabsize / chunksize;
3920 3935 waste = (slabsize % chunksize) / chunks;
3921 3936 if (waste < minwaste) {
3922 3937 minwaste = waste;
3923 3938 bestfit = slabsize;
3924 3939 }
3925 3940 }
3926 3941 if (cflags & KMC_QCACHE)
3927 3942 bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
3928 3943 cp->cache_slabsize = bestfit;
3929 3944 cp->cache_mincolor = 0;
3930 3945 cp->cache_maxcolor = bestfit % chunksize;
3931 3946 cp->cache_flags |= KMF_HASH;
3932 3947 }
3933 3948
3934 3949 cp->cache_maxchunks = (cp->cache_slabsize / cp->cache_chunksize);
3935 3950 cp->cache_partial_binshift = highbit(cp->cache_maxchunks / 16) + 1;
3936 3951
3937 3952 /*
3938 3953 * Disallowing prefill when either the DEBUG or HASH flag is set or when
3939 3954 * there is a constructor avoids some tricky issues with debug setup
3940 3955 * that may be revisited later. We cannot allow prefill in a
3941 3956 * metadata cache because of potential recursion.
3942 3957 */
3943 3958 if (vmp == kmem_msb_arena ||
3944 3959 cp->cache_flags & (KMF_HASH | KMF_BUFTAG) ||
3945 3960 cp->cache_constructor != NULL)
3946 3961 cp->cache_flags &= ~KMF_PREFILL;
3947 3962
3948 3963 if (cp->cache_flags & KMF_HASH) {
3949 3964 ASSERT(!(cflags & KMC_NOHASH));
3950 3965 cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
3951 3966 kmem_bufctl_audit_cache : kmem_bufctl_cache;
3952 3967 }
3953 3968
3954 3969 if (cp->cache_maxcolor >= vmp->vm_quantum)
3955 3970 cp->cache_maxcolor = vmp->vm_quantum - 1;
3956 3971
3957 3972 cp->cache_color = cp->cache_mincolor;
3958 3973
3959 3974 /*
3960 3975 * Initialize the rest of the slab layer.
3961 3976 */
3962 3977 mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
3963 3978
3964 3979 avl_create(&cp->cache_partial_slabs, kmem_partial_slab_cmp,
3965 3980 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3966 3981 /* LINTED: E_TRUE_LOGICAL_EXPR */
3967 3982 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
3968 3983 /* reuse partial slab AVL linkage for complete slab list linkage */
3969 3984 list_create(&cp->cache_complete_slabs,
3970 3985 sizeof (kmem_slab_t), offsetof(kmem_slab_t, slab_link));
3971 3986
3972 3987 if (cp->cache_flags & KMF_HASH) {
3973 3988 cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
3974 3989 KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
3975 3990 bzero(cp->cache_hash_table,
3976 3991 KMEM_HASH_INITIAL * sizeof (void *));
3977 3992 cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
3978 3993 cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
3979 3994 }
3980 3995
3981 3996 /*
3982 3997 * Initialize the depot.
3983 3998 */
3984 3999 mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
3985 4000
3986 4001 for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
3987 4002 continue;
3988 4003
3989 4004 cp->cache_magtype = mtp;
3990 4005
3991 4006 /*
3992 4007 * Initialize the CPU layer.
3993 4008 */
3994 4009 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
3995 4010 kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
3996 4011 mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
3997 4012 ccp->cc_flags = cp->cache_flags;
3998 4013 ccp->cc_rounds = -1;
3999 4014 ccp->cc_prounds = -1;
4000 4015 }
4001 4016
4002 4017 /*
4003 4018 * Create the cache's kstats.
4004 4019 */
4005 4020 if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
4006 4021 "kmem_cache", KSTAT_TYPE_NAMED,
4007 4022 sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
4008 4023 KSTAT_FLAG_VIRTUAL)) != NULL) {
4009 4024 cp->cache_kstat->ks_data = &kmem_cache_kstat;
4010 4025 cp->cache_kstat->ks_update = kmem_cache_kstat_update;
4011 4026 cp->cache_kstat->ks_private = cp;
4012 4027 cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
4013 4028 kstat_install(cp->cache_kstat);
4014 4029 }
4015 4030
4016 4031 /*
4017 4032 * Add the cache to the global list. This makes it visible
4018 4033 * to kmem_update(), so the cache must be ready for business.
4019 4034 */
4020 4035 mutex_enter(&kmem_cache_lock);
4021 4036 list_insert_tail(&kmem_caches, cp);
4022 4037 mutex_exit(&kmem_cache_lock);
4023 4038
4024 4039 if (kmem_ready)
4025 4040 kmem_cache_magazine_enable(cp);
4026 4041
4027 4042 return (cp);
4028 4043 }
4029 4044
4030 4045 static int
4031 4046 kmem_move_cmp(const void *buf, const void *p)
4032 4047 {
4033 4048 const kmem_move_t *kmm = p;
4034 4049 uintptr_t v1 = (uintptr_t)buf;
4035 4050 uintptr_t v2 = (uintptr_t)kmm->kmm_from_buf;
4036 4051 return (v1 < v2 ? -1 : (v1 > v2 ? 1 : 0));
4037 4052 }
4038 4053
4039 4054 static void
4040 4055 kmem_reset_reclaim_threshold(kmem_defrag_t *kmd)
4041 4056 {
4042 4057 kmd->kmd_reclaim_numer = 1;
4043 4058 }
4044 4059
4045 4060 /*
4046 4061 * Initially, when choosing candidate slabs for buffers to move, we want to be
4047 4062 * very selective and take only slabs that are less than
4048 4063 * (1 / KMEM_VOID_FRACTION) allocated. If we have difficulty finding candidate
4049 4064 * slabs, then we raise the allocation ceiling incrementally. The reclaim
4050 4065 * threshold is reset to (1 / KMEM_VOID_FRACTION) as soon as the cache is no
4051 4066 * longer fragmented.
4052 4067 */
4053 4068 static void
4054 4069 kmem_adjust_reclaim_threshold(kmem_defrag_t *kmd, int direction)
4055 4070 {
4056 4071 if (direction > 0) {
4057 4072 /* make it easier to find a candidate slab */
4058 4073 if (kmd->kmd_reclaim_numer < (KMEM_VOID_FRACTION - 1)) {
4059 4074 kmd->kmd_reclaim_numer++;
4060 4075 }
4061 4076 } else {
4062 4077 /* be more selective */
4063 4078 if (kmd->kmd_reclaim_numer > 1) {
4064 4079 kmd->kmd_reclaim_numer--;
4065 4080 }
4066 4081 }
4067 4082 }
4068 4083
4069 4084 void
4070 4085 kmem_cache_set_move(kmem_cache_t *cp,
4071 4086 kmem_cbrc_t (*move)(void *, void *, size_t, void *))
4072 4087 {
4073 4088 kmem_defrag_t *defrag;
4074 4089
4075 4090 ASSERT(move != NULL);
4076 4091 /*
4077 4092 * The consolidator does not support NOTOUCH caches because kmem cannot
4078 4093 * initialize their slabs with the 0xbaddcafe memory pattern, which sets
4079 4094 * a low order bit usable by clients to distinguish uninitialized memory
4080 4095 * from known objects (see kmem_slab_create).
4081 4096 */
4082 4097 ASSERT(!(cp->cache_cflags & KMC_NOTOUCH));
4083 4098 ASSERT(!(cp->cache_cflags & KMC_IDENTIFIER));
4084 4099
4085 4100 /*
4086 4101 * We should not be holding anyone's cache lock when calling
4087 4102 * kmem_cache_alloc(), so allocate in all cases before acquiring the
4088 4103 * lock.
4089 4104 */
4090 4105 defrag = kmem_cache_alloc(kmem_defrag_cache, KM_SLEEP);
4091 4106
4092 4107 mutex_enter(&cp->cache_lock);
4093 4108
4094 4109 if (KMEM_IS_MOVABLE(cp)) {
4095 4110 if (cp->cache_move == NULL) {
4096 4111 ASSERT(cp->cache_slab_alloc == 0);
4097 4112
4098 4113 cp->cache_defrag = defrag;
4099 4114 defrag = NULL; /* nothing to free */
4100 4115 bzero(cp->cache_defrag, sizeof (kmem_defrag_t));
4101 4116 avl_create(&cp->cache_defrag->kmd_moves_pending,
4102 4117 kmem_move_cmp, sizeof (kmem_move_t),
4103 4118 offsetof(kmem_move_t, kmm_entry));
4104 4119 /* LINTED: E_TRUE_LOGICAL_EXPR */
4105 4120 ASSERT(sizeof (list_node_t) <= sizeof (avl_node_t));
4106 4121 /* reuse the slab's AVL linkage for deadlist linkage */
4107 4122 list_create(&cp->cache_defrag->kmd_deadlist,
4108 4123 sizeof (kmem_slab_t),
4109 4124 offsetof(kmem_slab_t, slab_link));
4110 4125 kmem_reset_reclaim_threshold(cp->cache_defrag);
4111 4126 }
4112 4127 cp->cache_move = move;
4113 4128 }
4114 4129
4115 4130 mutex_exit(&cp->cache_lock);
4116 4131
4117 4132 if (defrag != NULL) {
4118 4133 kmem_cache_free(kmem_defrag_cache, defrag); /* unused */
4119 4134 }
4120 4135 }
4121 4136
4122 4137 void
4123 4138 kmem_cache_destroy(kmem_cache_t *cp)
4124 4139 {
4125 4140 int cpu_seqid;
4126 4141
4127 4142 /*
4128 4143 * Remove the cache from the global cache list so that no one else
4129 4144 * can schedule tasks on its behalf, wait for any pending tasks to
4130 4145 * complete, purge the cache, and then destroy it.
4131 4146 */
4132 4147 mutex_enter(&kmem_cache_lock);
4133 4148 list_remove(&kmem_caches, cp);
4134 4149 mutex_exit(&kmem_cache_lock);
4135 4150
4136 4151 if (kmem_taskq != NULL)
4137 4152 taskq_wait(kmem_taskq);
4138 4153
4139 4154 if (kmem_move_taskq != NULL && cp->cache_defrag != NULL)
4140 4155 taskq_wait(kmem_move_taskq);
4141 4156
4142 4157 kmem_cache_magazine_purge(cp);
4143 4158
4144 4159 mutex_enter(&cp->cache_lock);
4145 4160 if (cp->cache_buftotal != 0)
4146 4161 cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
4147 4162 cp->cache_name, (void *)cp);
4148 4163 if (cp->cache_defrag != NULL) {
4149 4164 avl_destroy(&cp->cache_defrag->kmd_moves_pending);
4150 4165 list_destroy(&cp->cache_defrag->kmd_deadlist);
4151 4166 kmem_cache_free(kmem_defrag_cache, cp->cache_defrag);
4152 4167 cp->cache_defrag = NULL;
4153 4168 }
4154 4169 /*
4155 4170 * The cache is now dead. There should be no further activity. We
4156 4171 * enforce this by setting land mines in the constructor, destructor,
4157 4172 * reclaim, and move routines that induce a kernel text fault if
4158 4173 * invoked.
4159 4174 */
4160 4175 cp->cache_constructor = (int (*)(void *, void *, int))1;
4161 4176 cp->cache_destructor = (void (*)(void *, void *))2;
4162 4177 cp->cache_reclaim = (void (*)(void *))3;
4163 4178 cp->cache_move = (kmem_cbrc_t (*)(void *, void *, size_t, void *))4;
4164 4179 mutex_exit(&cp->cache_lock);
4165 4180
4166 4181 kstat_delete(cp->cache_kstat);
4167 4182
4168 4183 if (cp->cache_hash_table != NULL)
4169 4184 vmem_free(kmem_hash_arena, cp->cache_hash_table,
4170 4185 (cp->cache_hash_mask + 1) * sizeof (void *));
4171 4186
4172 4187 for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
4173 4188 mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
4174 4189
4175 4190 mutex_destroy(&cp->cache_depot_lock);
4176 4191 mutex_destroy(&cp->cache_lock);
4177 4192
4178 4193 vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
4179 4194 }
4180 4195
4181 4196 /*ARGSUSED*/
4182 4197 static int
4183 4198 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
4184 4199 {
4185 4200 ASSERT(MUTEX_HELD(&cpu_lock));
4186 4201 if (what == CPU_UNCONFIG) {
4187 4202 kmem_cache_applyall(kmem_cache_magazine_purge,
4188 4203 kmem_taskq, TQ_SLEEP);
4189 4204 kmem_cache_applyall(kmem_cache_magazine_enable,
4190 4205 kmem_taskq, TQ_SLEEP);
4191 4206 }
4192 4207 return (0);
4193 4208 }
4194 4209
4195 4210 static void
4196 4211 kmem_alloc_caches_create(const int *array, size_t count,
4197 4212 kmem_cache_t **alloc_table, size_t maxbuf, uint_t shift)
4198 4213 {
4199 4214 char name[KMEM_CACHE_NAMELEN + 1];
4200 4215 size_t table_unit = (1 << shift); /* range of one alloc_table entry */
4201 4216 size_t size = table_unit;
4202 4217 int i;
4203 4218
4204 4219 for (i = 0; i < count; i++) {
4205 4220 size_t cache_size = array[i];
4206 4221 size_t align = KMEM_ALIGN;
4207 4222 kmem_cache_t *cp;
4208 4223
4209 4224 /* if the table has an entry for maxbuf, we're done */
4210 4225 if (size > maxbuf)
4211 4226 break;
4212 4227
4213 4228 /* cache size must be a multiple of the table unit */
4214 4229 ASSERT(P2PHASE(cache_size, table_unit) == 0);
4215 4230
4216 4231 /*
4217 4232 * If they allocate a multiple of the coherency granularity,
4218 4233 * they get a coherency-granularity-aligned address.
4219 4234 */
4220 4235 if (IS_P2ALIGNED(cache_size, 64))
4221 4236 align = 64;
4222 4237 if (IS_P2ALIGNED(cache_size, PAGESIZE))
4223 4238 align = PAGESIZE;
4224 4239 (void) snprintf(name, sizeof (name),
4225 4240 "kmem_alloc_%lu", cache_size);
4226 4241 cp = kmem_cache_create(name, cache_size, align,
4227 4242 NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
4228 4243
4229 4244 while (size <= cache_size) {
4230 4245 alloc_table[(size - 1) >> shift] = cp;
4231 4246 size += table_unit;
4232 4247 }
4233 4248 }
4234 4249
4235 4250 ASSERT(size > maxbuf); /* i.e. maxbuf <= max(cache_size) */
4236 4251 }
4237 4252
4238 4253 static void
4239 4254 kmem_cache_init(int pass, int use_large_pages)
4240 4255 {
4241 4256 int i;
4242 4257 size_t maxbuf;
4243 4258 kmem_magtype_t *mtp;
4244 4259
4245 4260 for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
4246 4261 char name[KMEM_CACHE_NAMELEN + 1];
4247 4262
4248 4263 mtp = &kmem_magtype[i];
4249 4264 (void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
4250 4265 mtp->mt_cache = kmem_cache_create(name,
4251 4266 (mtp->mt_magsize + 1) * sizeof (void *),
4252 4267 mtp->mt_align, NULL, NULL, NULL, NULL,
4253 4268 kmem_msb_arena, KMC_NOHASH);
4254 4269 }
4255 4270
4256 4271 kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
4257 4272 sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
4258 4273 kmem_msb_arena, KMC_NOHASH);
4259 4274
4260 4275 kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
4261 4276 sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
4262 4277 kmem_msb_arena, KMC_NOHASH);
4263 4278
4264 4279 kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
4265 4280 sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
4266 4281 kmem_msb_arena, KMC_NOHASH);
4267 4282
4268 4283 if (pass == 2) {
4269 4284 kmem_va_arena = vmem_create("kmem_va",
4270 4285 NULL, 0, PAGESIZE,
4271 4286 vmem_alloc, vmem_free, heap_arena,
4272 4287 8 * PAGESIZE, VM_SLEEP);
4273 4288
4274 4289 if (use_large_pages) {
4275 4290 kmem_default_arena = vmem_xcreate("kmem_default",
4276 4291 NULL, 0, PAGESIZE,
4277 4292 segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
4278 4293 0, VMC_DUMPSAFE | VM_SLEEP);
4279 4294 } else {
4280 4295 kmem_default_arena = vmem_create("kmem_default",
4281 4296 NULL, 0, PAGESIZE,
4282 4297 segkmem_alloc, segkmem_free, kmem_va_arena,
4283 4298 0, VMC_DUMPSAFE | VM_SLEEP);
4284 4299 }
4285 4300
4286 4301 /* Figure out what our maximum cache size is */
4287 4302 maxbuf = kmem_max_cached;
4288 4303 if (maxbuf <= KMEM_MAXBUF) {
4289 4304 maxbuf = 0;
4290 4305 kmem_max_cached = KMEM_MAXBUF;
4291 4306 } else {
4292 4307 size_t size = 0;
4293 4308 size_t max =
4294 4309 sizeof (kmem_big_alloc_sizes) / sizeof (int);
4295 4310 /*
4296 4311 * Round maxbuf up to an existing cache size. If maxbuf
4297 4312 * is larger than the largest cache, we truncate it to
4298 4313 * the largest cache's size.
4299 4314 */
4300 4315 for (i = 0; i < max; i++) {
4301 4316 size = kmem_big_alloc_sizes[i];
4302 4317 if (maxbuf <= size)
4303 4318 break;
4304 4319 }
4305 4320 kmem_max_cached = maxbuf = size;
4306 4321 }
4307 4322
4308 4323 /*
4309 4324 * The big alloc table may not be completely overwritten, so
4310 4325 * we clear out any stale cache pointers from the first pass.
4311 4326 */
4312 4327 bzero(kmem_big_alloc_table, sizeof (kmem_big_alloc_table));
4313 4328 } else {
4314 4329 /*
4315 4330 * During the first pass, the kmem_alloc_* caches
4316 4331 * are treated as metadata.
4317 4332 */
4318 4333 kmem_default_arena = kmem_msb_arena;
4319 4334 maxbuf = KMEM_BIG_MAXBUF_32BIT;
4320 4335 }
4321 4336
4322 4337 /*
4323 4338 * Set up the default caches to back kmem_alloc()
4324 4339 */
4325 4340 kmem_alloc_caches_create(
4326 4341 kmem_alloc_sizes, sizeof (kmem_alloc_sizes) / sizeof (int),
4327 4342 kmem_alloc_table, KMEM_MAXBUF, KMEM_ALIGN_SHIFT);
4328 4343
4329 4344 kmem_alloc_caches_create(
4330 4345 kmem_big_alloc_sizes, sizeof (kmem_big_alloc_sizes) / sizeof (int),
4331 4346 kmem_big_alloc_table, maxbuf, KMEM_BIG_SHIFT);
4332 4347
4333 4348 kmem_big_alloc_table_max = maxbuf >> KMEM_BIG_SHIFT;
4334 4349 }
4335 4350
4336 4351 void
4337 4352 kmem_init(void)
4338 4353 {
4339 4354 kmem_cache_t *cp;
4340 4355 int old_kmem_flags = kmem_flags;
4341 4356 int use_large_pages = 0;
4342 4357 size_t maxverify, minfirewall;
4343 4358
4344 4359 kstat_init();
4345 4360
4346 4361 /*
4347 4362 * Don't do firewalled allocations if the heap is less than 1TB
4348 4363 * (i.e. on a 32-bit kernel)
4349 4364 * The resulting VM_NEXTFIT allocations would create too much
4350 4365 * fragmentation in a small heap.
4351 4366 */
4352 4367 #if defined(_LP64)
4353 4368 maxverify = minfirewall = PAGESIZE / 2;
4354 4369 #else
4355 4370 maxverify = minfirewall = ULONG_MAX;
4356 4371 #endif
4357 4372
4358 4373 /* LINTED */
4359 4374 ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
4360 4375
4361 4376 list_create(&kmem_caches, sizeof (kmem_cache_t),
4362 4377 offsetof(kmem_cache_t, cache_link));
4363 4378
4364 4379 kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
4365 4380 vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
4366 4381 VM_SLEEP | VMC_NO_QCACHE);
4367 4382
4368 4383 kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
4369 4384 PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
4370 4385 VMC_DUMPSAFE | VM_SLEEP);
4371 4386
4372 4387 kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
4373 4388 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4374 4389
4375 4390 kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
4376 4391 segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
4377 4392
4378 4393 kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
4379 4394 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4380 4395
4381 4396 kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
4382 4397 NULL, 0, PAGESIZE,
4383 4398 kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
4384 4399 0, VM_SLEEP);
4385 4400
4386 4401 kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
4387 4402 segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0,
4388 4403 VMC_DUMPSAFE | VM_SLEEP);
4389 4404
4390 4405 /* temporary oversize arena for mod_read_system_file */
4391 4406 kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
4392 4407 segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
4393 4408
4394 4409 kmem_reap_interval = 15 * hz;
4395 4410
4396 4411 /*
4397 4412 * Read /etc/system. This is a chicken-and-egg problem because
4398 4413 * kmem_flags may be set in /etc/system, but mod_read_system_file()
4399 4414 * needs to use the allocator. The simplest solution is to create
4400 4415 * all the standard kmem caches, read /etc/system, destroy all the
4401 4416 * caches we just created, and then create them all again in light
4402 4417 * of the (possibly) new kmem_flags and other kmem tunables.
4403 4418 */
4404 4419 kmem_cache_init(1, 0);
4405 4420
4406 4421 mod_read_system_file(boothowto & RB_ASKNAME);
4407 4422
4408 4423 while ((cp = list_tail(&kmem_caches)) != NULL)
4409 4424 kmem_cache_destroy(cp);
4410 4425
4411 4426 vmem_destroy(kmem_oversize_arena);
4412 4427
4413 4428 if (old_kmem_flags & KMF_STICKY)
4414 4429 kmem_flags = old_kmem_flags;
4415 4430
4416 4431 if (!(kmem_flags & KMF_AUDIT))
4417 4432 vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
4418 4433
4419 4434 if (kmem_maxverify == 0)
4420 4435 kmem_maxverify = maxverify;
4421 4436
4422 4437 if (kmem_minfirewall == 0)
4423 4438 kmem_minfirewall = minfirewall;
4424 4439
4425 4440 /*
4426 4441 * give segkmem a chance to figure out if we are using large pages
4427 4442 * for the kernel heap
4428 4443 */
4429 4444 use_large_pages = segkmem_lpsetup();
4430 4445
4431 4446 /*
4432 4447 * To protect against corruption, we keep the actual number of callers
4433 4448 * KMF_LITE records seperate from the tunable. We arbitrarily clamp
4434 4449 * to 16, since the overhead for small buffers quickly gets out of
4435 4450 * hand.
4436 4451 *
4437 4452 * The real limit would depend on the needs of the largest KMC_NOHASH
4438 4453 * cache.
4439 4454 */
4440 4455 kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
4441 4456 kmem_lite_pcs = kmem_lite_count;
4442 4457
4443 4458 /*
4444 4459 * Normally, we firewall oversized allocations when possible, but
4445 4460 * if we are using large pages for kernel memory, and we don't have
4446 4461 * any non-LITE debugging flags set, we want to allocate oversized
4447 4462 * buffers from large pages, and so skip the firewalling.
4448 4463 */
4449 4464 if (use_large_pages &&
4450 4465 ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
4451 4466 kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
4452 4467 PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
4453 4468 0, VMC_DUMPSAFE | VM_SLEEP);
4454 4469 } else {
4455 4470 kmem_oversize_arena = vmem_create("kmem_oversize",
4456 4471 NULL, 0, PAGESIZE,
4457 4472 segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
4458 4473 kmem_firewall_va_arena : heap_arena, 0, VMC_DUMPSAFE |
4459 4474 VM_SLEEP);
4460 4475 }
4461 4476
4462 4477 kmem_cache_init(2, use_large_pages);
4463 4478
4464 4479 if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
4465 4480 if (kmem_transaction_log_size == 0)
4466 4481 kmem_transaction_log_size = kmem_maxavail() / 50;
|
↓ open down ↓ |
889 lines elided |
↑ open up ↑ |
4467 4482 kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
4468 4483 }
4469 4484
4470 4485 if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
4471 4486 if (kmem_content_log_size == 0)
4472 4487 kmem_content_log_size = kmem_maxavail() / 50;
4473 4488 kmem_content_log = kmem_log_init(kmem_content_log_size);
4474 4489 }
4475 4490
4476 4491 kmem_failure_log = kmem_log_init(kmem_failure_log_size);
4492 +
4477 4493 kmem_slab_log = kmem_log_init(kmem_slab_log_size);
4478 - kmem_zerosized_log = kmem_log_init(kmem_zerosized_log_size);
4479 4494
4480 4495 /*
4481 4496 * Initialize STREAMS message caches so allocb() is available.
4482 4497 * This allows us to initialize the logging framework (cmn_err(9F),
4483 4498 * strlog(9F), etc) so we can start recording messages.
4484 4499 */
4485 4500 streams_msg_init();
4486 4501
4487 4502 /*
4488 4503 * Initialize the ZSD framework in Zones so modules loaded henceforth
4489 4504 * can register their callbacks.
4490 4505 */
4491 4506 zone_zsd_init();
4492 4507
4493 4508 log_init();
4494 4509 taskq_init();
4495 4510
4496 4511 /*
4497 4512 * Warn about invalid or dangerous values of kmem_flags.
4498 4513 * Always warn about unsupported values.
4499 4514 */
4500 4515 if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
4501 4516 KMF_CONTENTS | KMF_LITE)) != 0) ||
4502 4517 ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
4503 4518 cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
4504 4519 "See the Solaris Tunable Parameters Reference Manual.",
4505 4520 kmem_flags);
4506 4521
4507 4522 #ifdef DEBUG
4508 4523 if ((kmem_flags & KMF_DEBUG) == 0)
4509 4524 cmn_err(CE_NOTE, "kmem debugging disabled.");
4510 4525 #else
4511 4526 /*
4512 4527 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
4513 4528 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
4514 4529 * if KMF_AUDIT is set). We should warn the user about the performance
4515 4530 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
4516 4531 * isn't set (since that disables AUDIT).
4517 4532 */
4518 4533 if (!(kmem_flags & KMF_LITE) &&
4519 4534 (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
4520 4535 cmn_err(CE_WARN, "High-overhead kmem debugging features "
4521 4536 "enabled (kmem_flags = 0x%x). Performance degradation "
4522 4537 "and large memory overhead possible. See the Solaris "
4523 4538 "Tunable Parameters Reference Manual.", kmem_flags);
4524 4539 #endif /* not DEBUG */
4525 4540
4526 4541 kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
4527 4542
4528 4543 kmem_ready = 1;
4529 4544
4530 4545 /*
4531 4546 * Initialize the platform-specific aligned/DMA memory allocator.
4532 4547 */
4533 4548 ka_init();
4534 4549
4535 4550 /*
4536 4551 * Initialize 32-bit ID cache.
4537 4552 */
4538 4553 id32_init();
4539 4554
4540 4555 /*
4541 4556 * Initialize the networking stack so modules loaded can
4542 4557 * register their callbacks.
4543 4558 */
4544 4559 netstack_init();
4545 4560 }
4546 4561
4547 4562 static void
4548 4563 kmem_move_init(void)
4549 4564 {
4550 4565 kmem_defrag_cache = kmem_cache_create("kmem_defrag_cache",
4551 4566 sizeof (kmem_defrag_t), 0, NULL, NULL, NULL, NULL,
4552 4567 kmem_msb_arena, KMC_NOHASH);
4553 4568 kmem_move_cache = kmem_cache_create("kmem_move_cache",
4554 4569 sizeof (kmem_move_t), 0, NULL, NULL, NULL, NULL,
4555 4570 kmem_msb_arena, KMC_NOHASH);
4556 4571
4557 4572 /*
4558 4573 * kmem guarantees that move callbacks are sequential and that even
4559 4574 * across multiple caches no two moves ever execute simultaneously.
4560 4575 * Move callbacks are processed on a separate taskq so that client code
4561 4576 * does not interfere with internal maintenance tasks.
4562 4577 */
4563 4578 kmem_move_taskq = taskq_create_instance("kmem_move_taskq", 0, 1,
4564 4579 minclsyspri, 100, INT_MAX, TASKQ_PREPOPULATE);
4565 4580 }
4566 4581
4567 4582 void
4568 4583 kmem_thread_init(void)
4569 4584 {
4570 4585 kmem_move_init();
4571 4586 kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
4572 4587 300, INT_MAX, TASKQ_PREPOPULATE);
4573 4588 }
4574 4589
4575 4590 void
4576 4591 kmem_mp_init(void)
4577 4592 {
4578 4593 mutex_enter(&cpu_lock);
4579 4594 register_cpu_setup_func(kmem_cpu_setup, NULL);
4580 4595 mutex_exit(&cpu_lock);
4581 4596
4582 4597 kmem_update_timeout(NULL);
4583 4598
4584 4599 taskq_mp_init();
4585 4600 }
4586 4601
4587 4602 /*
4588 4603 * Return the slab of the allocated buffer, or NULL if the buffer is not
4589 4604 * allocated. This function may be called with a known slab address to determine
4590 4605 * whether or not the buffer is allocated, or with a NULL slab address to obtain
4591 4606 * an allocated buffer's slab.
4592 4607 */
4593 4608 static kmem_slab_t *
4594 4609 kmem_slab_allocated(kmem_cache_t *cp, kmem_slab_t *sp, void *buf)
4595 4610 {
4596 4611 kmem_bufctl_t *bcp, *bufbcp;
4597 4612
4598 4613 ASSERT(MUTEX_HELD(&cp->cache_lock));
4599 4614 ASSERT(sp == NULL || KMEM_SLAB_MEMBER(sp, buf));
4600 4615
4601 4616 if (cp->cache_flags & KMF_HASH) {
4602 4617 for (bcp = *KMEM_HASH(cp, buf);
4603 4618 (bcp != NULL) && (bcp->bc_addr != buf);
4604 4619 bcp = bcp->bc_next) {
4605 4620 continue;
4606 4621 }
4607 4622 ASSERT(sp != NULL && bcp != NULL ? sp == bcp->bc_slab : 1);
4608 4623 return (bcp == NULL ? NULL : bcp->bc_slab);
4609 4624 }
4610 4625
4611 4626 if (sp == NULL) {
4612 4627 sp = KMEM_SLAB(cp, buf);
4613 4628 }
4614 4629 bufbcp = KMEM_BUFCTL(cp, buf);
4615 4630 for (bcp = sp->slab_head;
4616 4631 (bcp != NULL) && (bcp != bufbcp);
4617 4632 bcp = bcp->bc_next) {
4618 4633 continue;
4619 4634 }
4620 4635 return (bcp == NULL ? sp : NULL);
4621 4636 }
4622 4637
4623 4638 static boolean_t
4624 4639 kmem_slab_is_reclaimable(kmem_cache_t *cp, kmem_slab_t *sp, int flags)
4625 4640 {
4626 4641 long refcnt = sp->slab_refcnt;
4627 4642
4628 4643 ASSERT(cp->cache_defrag != NULL);
4629 4644
4630 4645 /*
4631 4646 * For code coverage we want to be able to move an object within the
4632 4647 * same slab (the only partial slab) even if allocating the destination
4633 4648 * buffer resulted in a completely allocated slab.
4634 4649 */
4635 4650 if (flags & KMM_DEBUG) {
4636 4651 return ((flags & KMM_DESPERATE) ||
4637 4652 ((sp->slab_flags & KMEM_SLAB_NOMOVE) == 0));
4638 4653 }
4639 4654
4640 4655 /* If we're desperate, we don't care if the client said NO. */
4641 4656 if (flags & KMM_DESPERATE) {
4642 4657 return (refcnt < sp->slab_chunks); /* any partial */
4643 4658 }
4644 4659
4645 4660 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4646 4661 return (B_FALSE);
4647 4662 }
4648 4663
4649 4664 if ((refcnt == 1) || kmem_move_any_partial) {
4650 4665 return (refcnt < sp->slab_chunks);
4651 4666 }
4652 4667
4653 4668 /*
4654 4669 * The reclaim threshold is adjusted at each kmem_cache_scan() so that
4655 4670 * slabs with a progressively higher percentage of used buffers can be
|
↓ open down ↓ |
167 lines elided |
↑ open up ↑ |
4656 4671 * reclaimed until the cache as a whole is no longer fragmented.
4657 4672 *
4658 4673 * sp->slab_refcnt kmd_reclaim_numer
4659 4674 * --------------- < ------------------
4660 4675 * sp->slab_chunks KMEM_VOID_FRACTION
4661 4676 */
4662 4677 return ((refcnt * KMEM_VOID_FRACTION) <
4663 4678 (sp->slab_chunks * cp->cache_defrag->kmd_reclaim_numer));
4664 4679 }
4665 4680
4681 +static void *
4682 +kmem_hunt_mag(kmem_cache_t *cp, kmem_magazine_t *m, int n, void *buf,
4683 + void *tbuf)
4684 +{
4685 + int i; /* magazine round index */
4686 +
4687 + for (i = 0; i < n; i++) {
4688 + if (buf == m->mag_round[i]) {
4689 + if (cp->cache_flags & KMF_BUFTAG) {
4690 + (void) kmem_cache_free_debug(cp, tbuf,
4691 + caller());
4692 + }
4693 + m->mag_round[i] = tbuf;
4694 + return (buf);
4695 + }
4696 + }
4697 +
4698 + return (NULL);
4699 +}
4700 +
4666 4701 /*
4702 + * Hunt the magazine layer for the given buffer. If found, the buffer is
4703 + * removed from the magazine layer and returned, otherwise NULL is returned.
4704 + * The state of the returned buffer is freed and constructed.
4705 + */
4706 +static void *
4707 +kmem_hunt_mags(kmem_cache_t *cp, void *buf)
4708 +{
4709 + kmem_cpu_cache_t *ccp;
4710 + kmem_magazine_t *m;
4711 + int cpu_seqid;
4712 + int n; /* magazine rounds */
4713 + void *tbuf; /* temporary swap buffer */
4714 +
4715 + ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4716 +
4717 + /*
4718 + * Allocated a buffer to swap with the one we hope to pull out of a
4719 + * magazine when found.
4720 + */
4721 + tbuf = kmem_cache_alloc(cp, KM_NOSLEEP);
4722 + if (tbuf == NULL) {
4723 + KMEM_STAT_ADD(kmem_move_stats.kms_hunt_alloc_fail);
4724 + return (NULL);
4725 + }
4726 + if (tbuf == buf) {
4727 + KMEM_STAT_ADD(kmem_move_stats.kms_hunt_lucky);
4728 + if (cp->cache_flags & KMF_BUFTAG) {
4729 + (void) kmem_cache_free_debug(cp, buf, caller());
4730 + }
4731 + return (buf);
4732 + }
4733 +
4734 + /* Hunt the depot. */
4735 + mutex_enter(&cp->cache_depot_lock);
4736 + n = cp->cache_magtype->mt_magsize;
4737 + for (m = cp->cache_full.ml_list; m != NULL; m = m->mag_next) {
4738 + if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4739 + mutex_exit(&cp->cache_depot_lock);
4740 + return (buf);
4741 + }
4742 + }
4743 + mutex_exit(&cp->cache_depot_lock);
4744 +
4745 + /* Hunt the per-CPU magazines. */
4746 + for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
4747 + ccp = &cp->cache_cpu[cpu_seqid];
4748 +
4749 + mutex_enter(&ccp->cc_lock);
4750 + m = ccp->cc_loaded;
4751 + n = ccp->cc_rounds;
4752 + if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4753 + mutex_exit(&ccp->cc_lock);
4754 + return (buf);
4755 + }
4756 + m = ccp->cc_ploaded;
4757 + n = ccp->cc_prounds;
4758 + if (kmem_hunt_mag(cp, m, n, buf, tbuf) != NULL) {
4759 + mutex_exit(&ccp->cc_lock);
4760 + return (buf);
4761 + }
4762 + mutex_exit(&ccp->cc_lock);
4763 + }
4764 +
4765 + kmem_cache_free(cp, tbuf);
4766 + return (NULL);
4767 +}
4768 +
4769 +/*
4667 4770 * May be called from the kmem_move_taskq, from kmem_cache_move_notify_task(),
4668 4771 * or when the buffer is freed.
4669 4772 */
4670 4773 static void
4671 4774 kmem_slab_move_yes(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4672 4775 {
4673 4776 ASSERT(MUTEX_HELD(&cp->cache_lock));
4674 4777 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4675 4778
4676 4779 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4677 4780 return;
4678 4781 }
4679 4782
4680 4783 if (sp->slab_flags & KMEM_SLAB_NOMOVE) {
4681 4784 if (KMEM_SLAB_OFFSET(sp, from_buf) == sp->slab_stuck_offset) {
4682 4785 avl_remove(&cp->cache_partial_slabs, sp);
4683 4786 sp->slab_flags &= ~KMEM_SLAB_NOMOVE;
4684 4787 sp->slab_stuck_offset = (uint32_t)-1;
4685 4788 avl_add(&cp->cache_partial_slabs, sp);
4686 4789 }
4687 4790 } else {
4688 4791 sp->slab_later_count = 0;
4689 4792 sp->slab_stuck_offset = (uint32_t)-1;
4690 4793 }
4691 4794 }
4692 4795
4693 4796 static void
4694 4797 kmem_slab_move_no(kmem_cache_t *cp, kmem_slab_t *sp, void *from_buf)
4695 4798 {
4696 4799 ASSERT(taskq_member(kmem_move_taskq, curthread));
4697 4800 ASSERT(MUTEX_HELD(&cp->cache_lock));
4698 4801 ASSERT(KMEM_SLAB_MEMBER(sp, from_buf));
4699 4802
4700 4803 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4701 4804 return;
4702 4805 }
4703 4806
4704 4807 avl_remove(&cp->cache_partial_slabs, sp);
4705 4808 sp->slab_later_count = 0;
4706 4809 sp->slab_flags |= KMEM_SLAB_NOMOVE;
4707 4810 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp, from_buf);
4708 4811 avl_add(&cp->cache_partial_slabs, sp);
4709 4812 }
4710 4813
4711 4814 static void kmem_move_end(kmem_cache_t *, kmem_move_t *);
4712 4815
4713 4816 /*
4714 4817 * The move callback takes two buffer addresses, the buffer to be moved, and a
4715 4818 * newly allocated and constructed buffer selected by kmem as the destination.
4716 4819 * It also takes the size of the buffer and an optional user argument specified
4717 4820 * at cache creation time. kmem guarantees that the buffer to be moved has not
4718 4821 * been unmapped by the virtual memory subsystem. Beyond that, it cannot
|
↓ open down ↓ |
42 lines elided |
↑ open up ↑ |
4719 4822 * guarantee the present whereabouts of the buffer to be moved, so it is up to
4720 4823 * the client to safely determine whether or not it is still using the buffer.
4721 4824 * The client must not free either of the buffers passed to the move callback,
4722 4825 * since kmem wants to free them directly to the slab layer. The client response
4723 4826 * tells kmem which of the two buffers to free:
4724 4827 *
4725 4828 * YES kmem frees the old buffer (the move was successful)
4726 4829 * NO kmem frees the new buffer, marks the slab of the old buffer
4727 4830 * non-reclaimable to avoid bothering the client again
4728 4831 * LATER kmem frees the new buffer, increments slab_later_count
4729 - * DONT_KNOW kmem frees the new buffer
4832 + * DONT_KNOW kmem frees the new buffer, searches mags for the old buffer
4730 4833 * DONT_NEED kmem frees both the old buffer and the new buffer
4731 4834 *
4732 4835 * The pending callback argument now being processed contains both of the
4733 4836 * buffers (old and new) passed to the move callback function, the slab of the
4734 4837 * old buffer, and flags related to the move request, such as whether or not the
4735 4838 * system was desperate for memory.
4736 4839 *
4737 4840 * Slabs are not freed while there is a pending callback, but instead are kept
4738 4841 * on a deadlist, which is drained after the last callback completes. This means
4739 4842 * that slabs are safe to access until kmem_move_end(), no matter how many of
4740 4843 * their buffers have been freed. Once slab_refcnt reaches zero, it stays at
4741 4844 * zero for as long as the slab remains on the deadlist and until the slab is
4742 4845 * freed.
4743 4846 */
4744 4847 static void
4745 4848 kmem_move_buffer(kmem_move_t *callback)
4746 4849 {
4747 4850 kmem_cbrc_t response;
4748 4851 kmem_slab_t *sp = callback->kmm_from_slab;
4749 4852 kmem_cache_t *cp = sp->slab_cache;
4750 4853 boolean_t free_on_slab;
4751 4854
4752 4855 ASSERT(taskq_member(kmem_move_taskq, curthread));
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
4753 4856 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4754 4857 ASSERT(KMEM_SLAB_MEMBER(sp, callback->kmm_from_buf));
4755 4858
4756 4859 /*
4757 4860 * The number of allocated buffers on the slab may have changed since we
4758 4861 * last checked the slab's reclaimability (when the pending move was
4759 4862 * enqueued), or the client may have responded NO when asked to move
4760 4863 * another buffer on the same slab.
4761 4864 */
4762 4865 if (!kmem_slab_is_reclaimable(cp, sp, callback->kmm_flags)) {
4866 + KMEM_STAT_ADD(kmem_move_stats.kms_no_longer_reclaimable);
4867 + KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4868 + kmem_move_stats.kms_notify_no_longer_reclaimable);
4763 4869 kmem_slab_free(cp, callback->kmm_to_buf);
4764 4870 kmem_move_end(cp, callback);
4765 4871 return;
4766 4872 }
4767 4873
4768 4874 /*
4769 - * Checking the slab layer is easy, so we might as well do that here
4770 - * in case we can avoid bothering the client.
4875 + * Hunting magazines is expensive, so we'll wait to do that until the
4876 + * client responds KMEM_CBRC_DONT_KNOW. However, checking the slab layer
4877 + * is cheap, so we might as well do that here in case we can avoid
4878 + * bothering the client.
4771 4879 */
4772 4880 mutex_enter(&cp->cache_lock);
4773 4881 free_on_slab = (kmem_slab_allocated(cp, sp,
4774 4882 callback->kmm_from_buf) == NULL);
4775 4883 mutex_exit(&cp->cache_lock);
4776 4884
4777 4885 if (free_on_slab) {
4886 + KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_slab);
4778 4887 kmem_slab_free(cp, callback->kmm_to_buf);
4779 4888 kmem_move_end(cp, callback);
4780 4889 return;
4781 4890 }
4782 4891
4783 4892 if (cp->cache_flags & KMF_BUFTAG) {
4784 4893 /*
4785 4894 * Make kmem_cache_alloc_debug() apply the constructor for us.
4786 4895 */
4787 4896 if (kmem_cache_alloc_debug(cp, callback->kmm_to_buf,
4788 4897 KM_NOSLEEP, 1, caller()) != 0) {
4898 + KMEM_STAT_ADD(kmem_move_stats.kms_alloc_fail);
4789 4899 kmem_move_end(cp, callback);
4790 4900 return;
4791 4901 }
4792 4902 } else if (cp->cache_constructor != NULL &&
4793 4903 cp->cache_constructor(callback->kmm_to_buf, cp->cache_private,
4794 4904 KM_NOSLEEP) != 0) {
4795 4905 atomic_inc_64(&cp->cache_alloc_fail);
4906 + KMEM_STAT_ADD(kmem_move_stats.kms_constructor_fail);
4796 4907 kmem_slab_free(cp, callback->kmm_to_buf);
4797 4908 kmem_move_end(cp, callback);
4798 4909 return;
4799 4910 }
4800 4911
4912 + KMEM_STAT_ADD(kmem_move_stats.kms_callbacks);
4913 + KMEM_STAT_COND_ADD((callback->kmm_flags & KMM_NOTIFY),
4914 + kmem_move_stats.kms_notify_callbacks);
4801 4915 cp->cache_defrag->kmd_callbacks++;
4802 4916 cp->cache_defrag->kmd_thread = curthread;
4803 4917 cp->cache_defrag->kmd_from_buf = callback->kmm_from_buf;
4804 4918 cp->cache_defrag->kmd_to_buf = callback->kmm_to_buf;
4805 4919 DTRACE_PROBE2(kmem__move__start, kmem_cache_t *, cp, kmem_move_t *,
4806 4920 callback);
4807 4921
4808 4922 response = cp->cache_move(callback->kmm_from_buf,
4809 4923 callback->kmm_to_buf, cp->cache_bufsize, cp->cache_private);
4810 4924
4811 4925 DTRACE_PROBE3(kmem__move__end, kmem_cache_t *, cp, kmem_move_t *,
4812 4926 callback, kmem_cbrc_t, response);
4813 4927 cp->cache_defrag->kmd_thread = NULL;
4814 4928 cp->cache_defrag->kmd_from_buf = NULL;
4815 4929 cp->cache_defrag->kmd_to_buf = NULL;
4816 4930
4817 4931 if (response == KMEM_CBRC_YES) {
4932 + KMEM_STAT_ADD(kmem_move_stats.kms_yes);
4818 4933 cp->cache_defrag->kmd_yes++;
4819 4934 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4820 4935 /* slab safe to access until kmem_move_end() */
4821 4936 if (sp->slab_refcnt == 0)
4822 4937 cp->cache_defrag->kmd_slabs_freed++;
4823 4938 mutex_enter(&cp->cache_lock);
4824 4939 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4825 4940 mutex_exit(&cp->cache_lock);
4826 4941 kmem_move_end(cp, callback);
4827 4942 return;
4828 4943 }
4829 4944
4830 4945 switch (response) {
4831 4946 case KMEM_CBRC_NO:
4947 + KMEM_STAT_ADD(kmem_move_stats.kms_no);
4832 4948 cp->cache_defrag->kmd_no++;
4833 4949 mutex_enter(&cp->cache_lock);
4834 4950 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4835 4951 mutex_exit(&cp->cache_lock);
4836 4952 break;
4837 4953 case KMEM_CBRC_LATER:
4954 + KMEM_STAT_ADD(kmem_move_stats.kms_later);
4838 4955 cp->cache_defrag->kmd_later++;
4839 4956 mutex_enter(&cp->cache_lock);
4840 4957 if (!KMEM_SLAB_IS_PARTIAL(sp)) {
4841 4958 mutex_exit(&cp->cache_lock);
4842 4959 break;
4843 4960 }
4844 4961
4845 4962 if (++sp->slab_later_count >= KMEM_DISBELIEF) {
4963 + KMEM_STAT_ADD(kmem_move_stats.kms_disbelief);
4846 4964 kmem_slab_move_no(cp, sp, callback->kmm_from_buf);
4847 4965 } else if (!(sp->slab_flags & KMEM_SLAB_NOMOVE)) {
4848 4966 sp->slab_stuck_offset = KMEM_SLAB_OFFSET(sp,
4849 4967 callback->kmm_from_buf);
4850 4968 }
4851 4969 mutex_exit(&cp->cache_lock);
4852 4970 break;
4853 4971 case KMEM_CBRC_DONT_NEED:
4972 + KMEM_STAT_ADD(kmem_move_stats.kms_dont_need);
4854 4973 cp->cache_defrag->kmd_dont_need++;
4855 4974 kmem_slab_free_constructed(cp, callback->kmm_from_buf, B_FALSE);
4856 4975 if (sp->slab_refcnt == 0)
4857 4976 cp->cache_defrag->kmd_slabs_freed++;
4858 4977 mutex_enter(&cp->cache_lock);
4859 4978 kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4860 4979 mutex_exit(&cp->cache_lock);
4861 4980 break;
4862 4981 case KMEM_CBRC_DONT_KNOW:
4863 - /*
4864 - * If we don't know if we can move this buffer or not, we'll
4865 - * just assume that we can't: if the buffer is in fact free,
4866 - * then it is sitting in one of the per-CPU magazines or in
4867 - * a full magazine in the depot layer. Either way, because
4868 - * defrag is induced in the same logic that reaps a cache,
4869 - * it's likely that full magazines will be returned to the
4870 - * system soon (thereby accomplishing what we're trying to
4871 - * accomplish here: return those magazines to their slabs).
4872 - * Given this, any work that we might do now to locate a buffer
4873 - * in a magazine is wasted (and expensive!) work; we bump
4874 - * a counter in this case and otherwise assume that we can't
4875 - * move it.
4876 - */
4982 + KMEM_STAT_ADD(kmem_move_stats.kms_dont_know);
4877 4983 cp->cache_defrag->kmd_dont_know++;
4984 + if (kmem_hunt_mags(cp, callback->kmm_from_buf) != NULL) {
4985 + KMEM_STAT_ADD(kmem_move_stats.kms_hunt_found_mag);
4986 + cp->cache_defrag->kmd_hunt_found++;
4987 + kmem_slab_free_constructed(cp, callback->kmm_from_buf,
4988 + B_TRUE);
4989 + if (sp->slab_refcnt == 0)
4990 + cp->cache_defrag->kmd_slabs_freed++;
4991 + mutex_enter(&cp->cache_lock);
4992 + kmem_slab_move_yes(cp, sp, callback->kmm_from_buf);
4993 + mutex_exit(&cp->cache_lock);
4994 + }
4878 4995 break;
4879 4996 default:
4880 4997 panic("'%s' (%p) unexpected move callback response %d\n",
4881 4998 cp->cache_name, (void *)cp, response);
4882 4999 }
4883 5000
4884 5001 kmem_slab_free_constructed(cp, callback->kmm_to_buf, B_FALSE);
4885 5002 kmem_move_end(cp, callback);
4886 5003 }
4887 5004
4888 5005 /* Return B_FALSE if there is insufficient memory for the move request. */
4889 5006 static boolean_t
4890 5007 kmem_move_begin(kmem_cache_t *cp, kmem_slab_t *sp, void *buf, int flags)
4891 5008 {
|
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
4892 5009 void *to_buf;
4893 5010 avl_index_t index;
4894 5011 kmem_move_t *callback, *pending;
4895 5012 ulong_t n;
4896 5013
4897 5014 ASSERT(taskq_member(kmem_taskq, curthread));
4898 5015 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4899 5016 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
4900 5017
4901 5018 callback = kmem_cache_alloc(kmem_move_cache, KM_NOSLEEP);
4902 -
4903 - if (callback == NULL)
5019 + if (callback == NULL) {
5020 + KMEM_STAT_ADD(kmem_move_stats.kms_callback_alloc_fail);
4904 5021 return (B_FALSE);
5022 + }
4905 5023
4906 5024 callback->kmm_from_slab = sp;
4907 5025 callback->kmm_from_buf = buf;
4908 5026 callback->kmm_flags = flags;
4909 5027
4910 5028 mutex_enter(&cp->cache_lock);
4911 5029
4912 5030 n = avl_numnodes(&cp->cache_partial_slabs);
4913 5031 if ((n == 0) || ((n == 1) && !(flags & KMM_DEBUG))) {
4914 5032 mutex_exit(&cp->cache_lock);
4915 5033 kmem_cache_free(kmem_move_cache, callback);
4916 5034 return (B_TRUE); /* there is no need for the move request */
4917 5035 }
4918 5036
|
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
4919 5037 pending = avl_find(&cp->cache_defrag->kmd_moves_pending, buf, &index);
4920 5038 if (pending != NULL) {
4921 5039 /*
4922 5040 * If the move is already pending and we're desperate now,
4923 5041 * update the move flags.
4924 5042 */
4925 5043 if (flags & KMM_DESPERATE) {
4926 5044 pending->kmm_flags |= KMM_DESPERATE;
4927 5045 }
4928 5046 mutex_exit(&cp->cache_lock);
5047 + KMEM_STAT_ADD(kmem_move_stats.kms_already_pending);
4929 5048 kmem_cache_free(kmem_move_cache, callback);
4930 5049 return (B_TRUE);
4931 5050 }
4932 5051
4933 5052 to_buf = kmem_slab_alloc_impl(cp, avl_first(&cp->cache_partial_slabs),
4934 5053 B_FALSE);
4935 5054 callback->kmm_to_buf = to_buf;
4936 5055 avl_insert(&cp->cache_defrag->kmd_moves_pending, callback, index);
4937 5056
4938 5057 mutex_exit(&cp->cache_lock);
4939 5058
4940 5059 if (!taskq_dispatch(kmem_move_taskq, (task_func_t *)kmem_move_buffer,
4941 5060 callback, TQ_NOSLEEP)) {
5061 + KMEM_STAT_ADD(kmem_move_stats.kms_callback_taskq_fail);
4942 5062 mutex_enter(&cp->cache_lock);
4943 5063 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4944 5064 mutex_exit(&cp->cache_lock);
4945 5065 kmem_slab_free(cp, to_buf);
4946 5066 kmem_cache_free(kmem_move_cache, callback);
4947 5067 return (B_FALSE);
4948 5068 }
4949 5069
4950 5070 return (B_TRUE);
4951 5071 }
4952 5072
4953 5073 static void
4954 5074 kmem_move_end(kmem_cache_t *cp, kmem_move_t *callback)
4955 5075 {
4956 5076 avl_index_t index;
4957 5077
4958 5078 ASSERT(cp->cache_defrag != NULL);
4959 5079 ASSERT(taskq_member(kmem_move_taskq, curthread));
4960 5080 ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
4961 5081
4962 5082 mutex_enter(&cp->cache_lock);
4963 5083 VERIFY(avl_find(&cp->cache_defrag->kmd_moves_pending,
4964 5084 callback->kmm_from_buf, &index) != NULL);
4965 5085 avl_remove(&cp->cache_defrag->kmd_moves_pending, callback);
4966 5086 if (avl_is_empty(&cp->cache_defrag->kmd_moves_pending)) {
4967 5087 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
4968 5088 kmem_slab_t *sp;
4969 5089
4970 5090 /*
4971 5091 * The last pending move completed. Release all slabs from the
4972 5092 * front of the dead list except for any slab at the tail that
4973 5093 * needs to be released from the context of kmem_move_buffers().
4974 5094 * kmem deferred unmapping the buffers on these slabs in order
4975 5095 * to guarantee that buffers passed to the move callback have
4976 5096 * been touched only by kmem or by the client itself.
|
↓ open down ↓ |
25 lines elided |
↑ open up ↑ |
4977 5097 */
4978 5098 while ((sp = list_remove_head(deadlist)) != NULL) {
4979 5099 if (sp->slab_flags & KMEM_SLAB_MOVE_PENDING) {
4980 5100 list_insert_tail(deadlist, sp);
4981 5101 break;
4982 5102 }
4983 5103 cp->cache_defrag->kmd_deadcount--;
4984 5104 cp->cache_slab_destroy++;
4985 5105 mutex_exit(&cp->cache_lock);
4986 5106 kmem_slab_destroy(cp, sp);
5107 + KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
4987 5108 mutex_enter(&cp->cache_lock);
4988 5109 }
4989 5110 }
4990 5111 mutex_exit(&cp->cache_lock);
4991 5112 kmem_cache_free(kmem_move_cache, callback);
4992 5113 }
4993 5114
4994 5115 /*
4995 5116 * Move buffers from least used slabs first by scanning backwards from the end
4996 5117 * of the partial slab list. Scan at most max_scan candidate slabs and move
4997 5118 * buffers from at most max_slabs slabs (0 for all partial slabs in both cases).
4998 5119 * If desperate to reclaim memory, move buffers from any partial slab, otherwise
4999 5120 * skip slabs with a ratio of allocated buffers at or above the current
5000 5121 * threshold. Return the number of unskipped slabs (at most max_slabs, -1 if the
5001 5122 * scan is aborted) so that the caller can adjust the reclaimability threshold
5002 5123 * depending on how many reclaimable slabs it finds.
5003 5124 *
5004 5125 * kmem_move_buffers() drops and reacquires cache_lock every time it issues a
5005 5126 * move request, since it is not valid for kmem_move_begin() to call
5006 5127 * kmem_cache_alloc() or taskq_dispatch() with cache_lock held.
5007 5128 */
5008 5129 static int
5009 5130 kmem_move_buffers(kmem_cache_t *cp, size_t max_scan, size_t max_slabs,
5010 5131 int flags)
5011 5132 {
5012 5133 kmem_slab_t *sp;
5013 5134 void *buf;
5014 5135 int i, j; /* slab index, buffer index */
5015 5136 int s; /* reclaimable slabs */
5016 5137 int b; /* allocated (movable) buffers on reclaimable slab */
5017 5138 boolean_t success;
5018 5139 int refcnt;
5019 5140 int nomove;
5020 5141
5021 5142 ASSERT(taskq_member(kmem_taskq, curthread));
5022 5143 ASSERT(MUTEX_HELD(&cp->cache_lock));
5023 5144 ASSERT(kmem_move_cache != NULL);
5024 5145 ASSERT(cp->cache_move != NULL && cp->cache_defrag != NULL);
5025 5146 ASSERT((flags & KMM_DEBUG) ? !avl_is_empty(&cp->cache_partial_slabs) :
5026 5147 avl_numnodes(&cp->cache_partial_slabs) > 1);
5027 5148
5028 5149 if (kmem_move_blocked) {
5029 5150 return (0);
5030 5151 }
5031 5152
5032 5153 if (kmem_move_fulltilt) {
5033 5154 flags |= KMM_DESPERATE;
5034 5155 }
5035 5156
5036 5157 if (max_scan == 0 || (flags & KMM_DESPERATE)) {
5037 5158 /*
5038 5159 * Scan as many slabs as needed to find the desired number of
5039 5160 * candidate slabs.
5040 5161 */
5041 5162 max_scan = (size_t)-1;
5042 5163 }
5043 5164
5044 5165 if (max_slabs == 0 || (flags & KMM_DESPERATE)) {
5045 5166 /* Find as many candidate slabs as possible. */
5046 5167 max_slabs = (size_t)-1;
5047 5168 }
5048 5169
5049 5170 sp = avl_last(&cp->cache_partial_slabs);
5050 5171 ASSERT(KMEM_SLAB_IS_PARTIAL(sp));
5051 5172 for (i = 0, s = 0; (i < max_scan) && (s < max_slabs) && (sp != NULL) &&
5052 5173 ((sp != avl_first(&cp->cache_partial_slabs)) ||
5053 5174 (flags & KMM_DEBUG));
5054 5175 sp = AVL_PREV(&cp->cache_partial_slabs, sp), i++) {
5055 5176
5056 5177 if (!kmem_slab_is_reclaimable(cp, sp, flags)) {
5057 5178 continue;
5058 5179 }
5059 5180 s++;
5060 5181
5061 5182 /* Look for allocated buffers to move. */
5062 5183 for (j = 0, b = 0, buf = sp->slab_base;
5063 5184 (j < sp->slab_chunks) && (b < sp->slab_refcnt);
5064 5185 buf = (((char *)buf) + cp->cache_chunksize), j++) {
5065 5186
5066 5187 if (kmem_slab_allocated(cp, sp, buf) == NULL) {
5067 5188 continue;
5068 5189 }
5069 5190
5070 5191 b++;
5071 5192
5072 5193 /*
5073 5194 * Prevent the slab from being destroyed while we drop
5074 5195 * cache_lock and while the pending move is not yet
5075 5196 * registered. Flag the pending move while
5076 5197 * kmd_moves_pending may still be empty, since we can't
5077 5198 * yet rely on a non-zero pending move count to prevent
5078 5199 * the slab from being destroyed.
5079 5200 */
5080 5201 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5081 5202 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5082 5203 /*
5083 5204 * Recheck refcnt and nomove after reacquiring the lock,
5084 5205 * since these control the order of partial slabs, and
5085 5206 * we want to know if we can pick up the scan where we
5086 5207 * left off.
5087 5208 */
5088 5209 refcnt = sp->slab_refcnt;
5089 5210 nomove = (sp->slab_flags & KMEM_SLAB_NOMOVE);
5090 5211 mutex_exit(&cp->cache_lock);
5091 5212
5092 5213 success = kmem_move_begin(cp, sp, buf, flags);
5093 5214
5094 5215 /*
5095 5216 * Now, before the lock is reacquired, kmem could
5096 5217 * process all pending move requests and purge the
5097 5218 * deadlist, so that upon reacquiring the lock, sp has
5098 5219 * been remapped. Or, the client may free all the
5099 5220 * objects on the slab while the pending moves are still
5100 5221 * on the taskq. Therefore, the KMEM_SLAB_MOVE_PENDING
5101 5222 * flag causes the slab to be put at the end of the
5102 5223 * deadlist and prevents it from being destroyed, since
5103 5224 * we plan to destroy it here after reacquiring the
5104 5225 * lock.
5105 5226 */
5106 5227 mutex_enter(&cp->cache_lock);
5107 5228 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5108 5229 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5109 5230
5110 5231 if (sp->slab_refcnt == 0) {
5111 5232 list_t *deadlist =
5112 5233 &cp->cache_defrag->kmd_deadlist;
5113 5234 list_remove(deadlist, sp);
5114 5235
5115 5236 if (!avl_is_empty(
5116 5237 &cp->cache_defrag->kmd_moves_pending)) {
5117 5238 /*
5118 5239 * A pending move makes it unsafe to
5119 5240 * destroy the slab, because even though
5120 5241 * the move is no longer needed, the
|
↓ open down ↓ |
124 lines elided |
↑ open up ↑ |
5121 5242 * context where that is determined
5122 5243 * requires the slab to exist.
5123 5244 * Fortunately, a pending move also
5124 5245 * means we don't need to destroy the
5125 5246 * slab here, since it will get
5126 5247 * destroyed along with any other slabs
5127 5248 * on the deadlist after the last
5128 5249 * pending move completes.
5129 5250 */
5130 5251 list_insert_head(deadlist, sp);
5252 + KMEM_STAT_ADD(kmem_move_stats.
5253 + kms_endscan_slab_dead);
5131 5254 return (-1);
5132 5255 }
5133 5256
5134 5257 /*
5135 5258 * Destroy the slab now if it was completely
5136 5259 * freed while we dropped cache_lock and there
5137 5260 * are no pending moves. Since slab_refcnt
5138 5261 * cannot change once it reaches zero, no new
5139 5262 * pending moves from that slab are possible.
5140 5263 */
5141 5264 cp->cache_defrag->kmd_deadcount--;
5142 5265 cp->cache_slab_destroy++;
5143 5266 mutex_exit(&cp->cache_lock);
5144 5267 kmem_slab_destroy(cp, sp);
5268 + KMEM_STAT_ADD(kmem_move_stats.
5269 + kms_dead_slabs_freed);
5270 + KMEM_STAT_ADD(kmem_move_stats.
5271 + kms_endscan_slab_destroyed);
5145 5272 mutex_enter(&cp->cache_lock);
5146 5273 /*
5147 5274 * Since we can't pick up the scan where we left
5148 5275 * off, abort the scan and say nothing about the
5149 5276 * number of reclaimable slabs.
5150 5277 */
5151 5278 return (-1);
5152 5279 }
5153 5280
5154 5281 if (!success) {
5155 5282 /*
5156 5283 * Abort the scan if there is not enough memory
5157 5284 * for the request and say nothing about the
5158 5285 * number of reclaimable slabs.
5159 5286 */
5287 + KMEM_STAT_COND_ADD(s < max_slabs,
5288 + kmem_move_stats.kms_endscan_nomem);
5160 5289 return (-1);
5161 5290 }
5162 5291
5163 5292 /*
5164 5293 * The slab's position changed while the lock was
5165 5294 * dropped, so we don't know where we are in the
5166 5295 * sequence any more.
5167 5296 */
5168 5297 if (sp->slab_refcnt != refcnt) {
5169 5298 /*
5170 5299 * If this is a KMM_DEBUG move, the slab_refcnt
5171 5300 * may have changed because we allocated a
5172 5301 * destination buffer on the same slab. In that
5173 5302 * case, we're not interested in counting it.
5174 5303 */
5304 + KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5305 + (s < max_slabs),
5306 + kmem_move_stats.kms_endscan_refcnt_changed);
5175 5307 return (-1);
5176 5308 }
5177 - if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove)
5309 + if ((sp->slab_flags & KMEM_SLAB_NOMOVE) != nomove) {
5310 + KMEM_STAT_COND_ADD(s < max_slabs,
5311 + kmem_move_stats.kms_endscan_nomove_changed);
5178 5312 return (-1);
5313 + }
5179 5314
5180 5315 /*
5181 5316 * Generating a move request allocates a destination
5182 5317 * buffer from the slab layer, bumping the first partial
5183 5318 * slab if it is completely allocated. If the current
5184 5319 * slab becomes the first partial slab as a result, we
5185 5320 * can't continue to scan backwards.
5186 5321 *
5187 5322 * If this is a KMM_DEBUG move and we allocated the
5188 5323 * destination buffer from the last partial slab, then
5189 5324 * the buffer we're moving is on the same slab and our
5190 5325 * slab_refcnt has changed, causing us to return before
5191 5326 * reaching here if there are no partial slabs left.
5192 5327 */
5193 5328 ASSERT(!avl_is_empty(&cp->cache_partial_slabs));
5194 5329 if (sp == avl_first(&cp->cache_partial_slabs)) {
|
↓ open down ↓ |
6 lines elided |
↑ open up ↑ |
5195 5330 /*
5196 5331 * We're not interested in a second KMM_DEBUG
5197 5332 * move.
5198 5333 */
5199 5334 goto end_scan;
5200 5335 }
5201 5336 }
5202 5337 }
5203 5338 end_scan:
5204 5339
5340 + KMEM_STAT_COND_ADD(!(flags & KMM_DEBUG) &&
5341 + (s < max_slabs) &&
5342 + (sp == avl_first(&cp->cache_partial_slabs)),
5343 + kmem_move_stats.kms_endscan_freelist);
5344 +
5205 5345 return (s);
5206 5346 }
5207 5347
5208 5348 typedef struct kmem_move_notify_args {
5209 5349 kmem_cache_t *kmna_cache;
5210 5350 void *kmna_buf;
5211 5351 } kmem_move_notify_args_t;
5212 5352
5213 5353 static void
5214 5354 kmem_cache_move_notify_task(void *arg)
5215 5355 {
5216 5356 kmem_move_notify_args_t *args = arg;
5217 5357 kmem_cache_t *cp = args->kmna_cache;
5218 5358 void *buf = args->kmna_buf;
5219 5359 kmem_slab_t *sp;
5220 5360
5221 5361 ASSERT(taskq_member(kmem_taskq, curthread));
5222 5362 ASSERT(list_link_active(&cp->cache_link));
5223 5363
5224 5364 kmem_free(args, sizeof (kmem_move_notify_args_t));
5225 5365 mutex_enter(&cp->cache_lock);
5226 5366 sp = kmem_slab_allocated(cp, NULL, buf);
5227 5367
5228 5368 /* Ignore the notification if the buffer is no longer allocated. */
5229 5369 if (sp == NULL) {
5230 5370 mutex_exit(&cp->cache_lock);
5231 5371 return;
5232 5372 }
5233 5373
5234 5374 /* Ignore the notification if there's no reason to move the buffer. */
5235 5375 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5236 5376 /*
5237 5377 * So far the notification is not ignored. Ignore the
5238 5378 * notification if the slab is not marked by an earlier refusal
5239 5379 * to move a buffer.
5240 5380 */
5241 5381 if (!(sp->slab_flags & KMEM_SLAB_NOMOVE) &&
5242 5382 (sp->slab_later_count == 0)) {
5243 5383 mutex_exit(&cp->cache_lock);
5244 5384 return;
5245 5385 }
5246 5386
5247 5387 kmem_slab_move_yes(cp, sp, buf);
5248 5388 ASSERT(!(sp->slab_flags & KMEM_SLAB_MOVE_PENDING));
5249 5389 sp->slab_flags |= KMEM_SLAB_MOVE_PENDING;
5250 5390 mutex_exit(&cp->cache_lock);
5251 5391 /* see kmem_move_buffers() about dropping the lock */
5252 5392 (void) kmem_move_begin(cp, sp, buf, KMM_NOTIFY);
5253 5393 mutex_enter(&cp->cache_lock);
|
↓ open down ↓ |
39 lines elided |
↑ open up ↑ |
5254 5394 ASSERT(sp->slab_flags & KMEM_SLAB_MOVE_PENDING);
5255 5395 sp->slab_flags &= ~KMEM_SLAB_MOVE_PENDING;
5256 5396 if (sp->slab_refcnt == 0) {
5257 5397 list_t *deadlist = &cp->cache_defrag->kmd_deadlist;
5258 5398 list_remove(deadlist, sp);
5259 5399
5260 5400 if (!avl_is_empty(
5261 5401 &cp->cache_defrag->kmd_moves_pending)) {
5262 5402 list_insert_head(deadlist, sp);
5263 5403 mutex_exit(&cp->cache_lock);
5404 + KMEM_STAT_ADD(kmem_move_stats.
5405 + kms_notify_slab_dead);
5264 5406 return;
5265 5407 }
5266 5408
5267 5409 cp->cache_defrag->kmd_deadcount--;
5268 5410 cp->cache_slab_destroy++;
5269 5411 mutex_exit(&cp->cache_lock);
5270 5412 kmem_slab_destroy(cp, sp);
5413 + KMEM_STAT_ADD(kmem_move_stats.kms_dead_slabs_freed);
5414 + KMEM_STAT_ADD(kmem_move_stats.
5415 + kms_notify_slab_destroyed);
5271 5416 return;
5272 5417 }
5273 5418 } else {
5274 5419 kmem_slab_move_yes(cp, sp, buf);
5275 5420 }
5276 5421 mutex_exit(&cp->cache_lock);
5277 5422 }
5278 5423
5279 5424 void
5280 5425 kmem_cache_move_notify(kmem_cache_t *cp, void *buf)
5281 5426 {
5282 5427 kmem_move_notify_args_t *args;
5283 5428
5429 + KMEM_STAT_ADD(kmem_move_stats.kms_notify);
5284 5430 args = kmem_alloc(sizeof (kmem_move_notify_args_t), KM_NOSLEEP);
5285 5431 if (args != NULL) {
5286 5432 args->kmna_cache = cp;
5287 5433 args->kmna_buf = buf;
5288 5434 if (!taskq_dispatch(kmem_taskq,
5289 5435 (task_func_t *)kmem_cache_move_notify_task, args,
5290 5436 TQ_NOSLEEP))
5291 5437 kmem_free(args, sizeof (kmem_move_notify_args_t));
5292 5438 }
5293 5439 }
5294 5440
5295 5441 static void
|
↓ open down ↓ |
2 lines elided |
↑ open up ↑ |
5296 5442 kmem_cache_defrag(kmem_cache_t *cp)
5297 5443 {
5298 5444 size_t n;
5299 5445
5300 5446 ASSERT(cp->cache_defrag != NULL);
5301 5447
5302 5448 mutex_enter(&cp->cache_lock);
5303 5449 n = avl_numnodes(&cp->cache_partial_slabs);
5304 5450 if (n > 1) {
5305 5451 /* kmem_move_buffers() drops and reacquires cache_lock */
5452 + KMEM_STAT_ADD(kmem_move_stats.kms_defrags);
5306 5453 cp->cache_defrag->kmd_defrags++;
5307 5454 (void) kmem_move_buffers(cp, n, 0, KMM_DESPERATE);
5308 5455 }
5309 5456 mutex_exit(&cp->cache_lock);
5310 5457 }
5311 5458
5312 5459 /* Is this cache above the fragmentation threshold? */
5313 5460 static boolean_t
5314 5461 kmem_cache_frag_threshold(kmem_cache_t *cp, uint64_t nfree)
5315 5462 {
5316 5463 /*
5317 5464 * nfree kmem_frag_numer
5318 5465 * ------------------ > ---------------
5319 5466 * cp->cache_buftotal kmem_frag_denom
5320 5467 */
5321 5468 return ((nfree * kmem_frag_denom) >
5322 5469 (cp->cache_buftotal * kmem_frag_numer));
5323 5470 }
5324 5471
5325 5472 static boolean_t
5326 5473 kmem_cache_is_fragmented(kmem_cache_t *cp, boolean_t *doreap)
5327 5474 {
5328 5475 boolean_t fragmented;
5329 5476 uint64_t nfree;
5330 5477
5331 5478 ASSERT(MUTEX_HELD(&cp->cache_lock));
5332 5479 *doreap = B_FALSE;
5333 5480
5334 5481 if (kmem_move_fulltilt) {
5335 5482 if (avl_numnodes(&cp->cache_partial_slabs) > 1) {
5336 5483 return (B_TRUE);
5337 5484 }
5338 5485 } else {
5339 5486 if ((cp->cache_complete_slab_count + avl_numnodes(
5340 5487 &cp->cache_partial_slabs)) < kmem_frag_minslabs) {
5341 5488 return (B_FALSE);
5342 5489 }
5343 5490 }
5344 5491
5345 5492 nfree = cp->cache_bufslab;
5346 5493 fragmented = ((avl_numnodes(&cp->cache_partial_slabs) > 1) &&
5347 5494 kmem_cache_frag_threshold(cp, nfree));
5348 5495
5349 5496 /*
5350 5497 * Free buffers in the magazine layer appear allocated from the point of
5351 5498 * view of the slab layer. We want to know if the slab layer would
5352 5499 * appear fragmented if we included free buffers from magazines that
5353 5500 * have fallen out of the working set.
5354 5501 */
5355 5502 if (!fragmented) {
5356 5503 long reap;
5357 5504
5358 5505 mutex_enter(&cp->cache_depot_lock);
5359 5506 reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
5360 5507 reap = MIN(reap, cp->cache_full.ml_total);
5361 5508 mutex_exit(&cp->cache_depot_lock);
5362 5509
5363 5510 nfree += ((uint64_t)reap * cp->cache_magtype->mt_magsize);
5364 5511 if (kmem_cache_frag_threshold(cp, nfree)) {
5365 5512 *doreap = B_TRUE;
5366 5513 }
5367 5514 }
5368 5515
5369 5516 return (fragmented);
5370 5517 }
5371 5518
5372 5519 /* Called periodically from kmem_taskq */
5373 5520 static void
5374 5521 kmem_cache_scan(kmem_cache_t *cp)
5375 5522 {
5376 5523 boolean_t reap = B_FALSE;
5377 5524 kmem_defrag_t *kmd;
5378 5525
5379 5526 ASSERT(taskq_member(kmem_taskq, curthread));
5380 5527
5381 5528 mutex_enter(&cp->cache_lock);
5382 5529
5383 5530 kmd = cp->cache_defrag;
5384 5531 if (kmd->kmd_consolidate > 0) {
5385 5532 kmd->kmd_consolidate--;
5386 5533 mutex_exit(&cp->cache_lock);
5387 5534 kmem_cache_reap(cp);
5388 5535 return;
5389 5536 }
5390 5537
5391 5538 if (kmem_cache_is_fragmented(cp, &reap)) {
5392 5539 size_t slabs_found;
5393 5540
|
↓ open down ↓ |
78 lines elided |
↑ open up ↑ |
5394 5541 /*
5395 5542 * Consolidate reclaimable slabs from the end of the partial
5396 5543 * slab list (scan at most kmem_reclaim_scan_range slabs to find
5397 5544 * reclaimable slabs). Keep track of how many candidate slabs we
5398 5545 * looked for and how many we actually found so we can adjust
5399 5546 * the definition of a candidate slab if we're having trouble
5400 5547 * finding them.
5401 5548 *
5402 5549 * kmem_move_buffers() drops and reacquires cache_lock.
5403 5550 */
5551 + KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5404 5552 kmd->kmd_scans++;
5405 5553 slabs_found = kmem_move_buffers(cp, kmem_reclaim_scan_range,
5406 5554 kmem_reclaim_max_slabs, 0);
5407 5555 if (slabs_found >= 0) {
5408 5556 kmd->kmd_slabs_sought += kmem_reclaim_max_slabs;
5409 5557 kmd->kmd_slabs_found += slabs_found;
5410 5558 }
5411 5559
5412 5560 if (++kmd->kmd_tries >= kmem_reclaim_scan_range) {
5413 5561 kmd->kmd_tries = 0;
5414 5562
5415 5563 /*
5416 5564 * If we had difficulty finding candidate slabs in
5417 5565 * previous scans, adjust the threshold so that
5418 5566 * candidates are easier to find.
5419 5567 */
5420 5568 if (kmd->kmd_slabs_found == kmd->kmd_slabs_sought) {
5421 5569 kmem_adjust_reclaim_threshold(kmd, -1);
5422 5570 } else if ((kmd->kmd_slabs_found * 2) <
5423 5571 kmd->kmd_slabs_sought) {
5424 5572 kmem_adjust_reclaim_threshold(kmd, 1);
5425 5573 }
5426 5574 kmd->kmd_slabs_sought = 0;
5427 5575 kmd->kmd_slabs_found = 0;
5428 5576 }
5429 5577 } else {
5430 5578 kmem_reset_reclaim_threshold(cp->cache_defrag);
5431 5579 #ifdef DEBUG
5432 5580 if (!avl_is_empty(&cp->cache_partial_slabs)) {
5433 5581 /*
|
↓ open down ↓ |
20 lines elided |
↑ open up ↑ |
5434 5582 * In a debug kernel we want the consolidator to
5435 5583 * run occasionally even when there is plenty of
5436 5584 * memory.
5437 5585 */
5438 5586 uint16_t debug_rand;
5439 5587
5440 5588 (void) random_get_bytes((uint8_t *)&debug_rand, 2);
5441 5589 if (!kmem_move_noreap &&
5442 5590 ((debug_rand % kmem_mtb_reap) == 0)) {
5443 5591 mutex_exit(&cp->cache_lock);
5592 + KMEM_STAT_ADD(kmem_move_stats.kms_debug_reaps);
5444 5593 kmem_cache_reap(cp);
5445 5594 return;
5446 5595 } else if ((debug_rand % kmem_mtb_move) == 0) {
5596 + KMEM_STAT_ADD(kmem_move_stats.kms_scans);
5597 + KMEM_STAT_ADD(kmem_move_stats.kms_debug_scans);
5447 5598 kmd->kmd_scans++;
5448 5599 (void) kmem_move_buffers(cp,
5449 5600 kmem_reclaim_scan_range, 1, KMM_DEBUG);
5450 5601 }
5451 5602 }
5452 5603 #endif /* DEBUG */
5453 5604 }
5454 5605
5455 5606 mutex_exit(&cp->cache_lock);
5456 5607
5457 - if (reap)
5608 + if (reap) {
5609 + KMEM_STAT_ADD(kmem_move_stats.kms_scan_depot_ws_reaps);
5458 5610 kmem_depot_ws_reap(cp);
5611 + }
5459 5612 }
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX