]>
Commit | Line | Data |
---|---|---|
716154c5 BB |
1 | /*****************************************************************************\ |
2 | * Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC. | |
3 | * Copyright (C) 2007 The Regents of the University of California. | |
4 | * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER). | |
5 | * Written by Brian Behlendorf <behlendorf1@llnl.gov>. | |
715f6251 | 6 | * UCRL-CODE-235197 |
7 | * | |
716154c5 BB |
8 | * This file is part of the SPL, Solaris Porting Layer. |
9 | * For details, see <http://github.com/behlendorf/spl/>. | |
715f6251 | 10 | * |
716154c5 BB |
11 | * The SPL is free software; you can redistribute it and/or modify it |
12 | * under the terms of the GNU General Public License as published by the | |
13 | * Free Software Foundation; either version 2 of the License, or (at your | |
14 | * option) any later version. | |
15 | * | |
16 | * The SPL is distributed in the hope that it will be useful, but WITHOUT | |
715f6251 | 17 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
18 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
19 | * for more details. | |
20 | * | |
21 | * You should have received a copy of the GNU General Public License along | |
716154c5 BB |
22 | * with the SPL. If not, see <http://www.gnu.org/licenses/>. |
23 | ***************************************************************************** | |
24 | * Solaris Porting Layer (SPL) Kmem Implementation. | |
25 | \*****************************************************************************/ | |
715f6251 | 26 | |
f4b37741 | 27 | #include <sys/kmem.h> |
55abb092 | 28 | #include <spl-debug.h> |
f1ca4da6 | 29 | |
b17edc10 BB |
30 | #ifdef SS_DEBUG_SUBSYS |
31 | #undef SS_DEBUG_SUBSYS | |
937879f1 | 32 | #endif |
33 | ||
b17edc10 | 34 | #define SS_DEBUG_SUBSYS SS_KMEM |
937879f1 | 35 | |
36b313da BB |
36 | /* |
37 | * The minimum amount of memory measured in pages to be free at all | |
38 | * times on the system. This is similar to Linux's zone->pages_min | |
39 | * multipled by the number of zones and is sized based on that. | |
40 | */ | |
41 | pgcnt_t minfree = 0; | |
42 | EXPORT_SYMBOL(minfree); | |
43 | ||
44 | /* | |
45 | * The desired amount of memory measured in pages to be free at all | |
46 | * times on the system. This is similar to Linux's zone->pages_low | |
47 | * multipled by the number of zones and is sized based on that. | |
48 | * Assuming all zones are being used roughly equally, when we drop | |
49 | * below this threshold async page reclamation is triggered. | |
50 | */ | |
51 | pgcnt_t desfree = 0; | |
52 | EXPORT_SYMBOL(desfree); | |
53 | ||
54 | /* | |
55 | * When above this amount of memory measures in pages the system is | |
56 | * determined to have enough free memory. This is similar to Linux's | |
57 | * zone->pages_high multipled by the number of zones and is sized based | |
58 | * on that. Assuming all zones are being used roughly equally, when | |
59 | * async page reclamation reaches this threshold it stops. | |
60 | */ | |
61 | pgcnt_t lotsfree = 0; | |
62 | EXPORT_SYMBOL(lotsfree); | |
63 | ||
64 | /* Unused always 0 in this implementation */ | |
65 | pgcnt_t needfree = 0; | |
66 | EXPORT_SYMBOL(needfree); | |
67 | ||
36b313da BB |
68 | pgcnt_t swapfs_minfree = 0; |
69 | EXPORT_SYMBOL(swapfs_minfree); | |
70 | ||
71 | pgcnt_t swapfs_reserve = 0; | |
72 | EXPORT_SYMBOL(swapfs_reserve); | |
73 | ||
36b313da BB |
74 | vmem_t *heap_arena = NULL; |
75 | EXPORT_SYMBOL(heap_arena); | |
76 | ||
77 | vmem_t *zio_alloc_arena = NULL; | |
78 | EXPORT_SYMBOL(zio_alloc_arena); | |
79 | ||
80 | vmem_t *zio_arena = NULL; | |
81 | EXPORT_SYMBOL(zio_arena); | |
82 | ||
d1ff2312 | 83 | #ifndef HAVE_GET_VMALLOC_INFO |
96dded38 | 84 | get_vmalloc_info_t get_vmalloc_info_fn = SYMBOL_POISON; |
d1ff2312 BB |
85 | EXPORT_SYMBOL(get_vmalloc_info_fn); |
86 | #endif /* HAVE_GET_VMALLOC_INFO */ | |
87 | ||
5232d256 BB |
88 | #ifdef HAVE_PGDAT_HELPERS |
89 | # ifndef HAVE_FIRST_ONLINE_PGDAT | |
96dded38 | 90 | first_online_pgdat_t first_online_pgdat_fn = SYMBOL_POISON; |
d1ff2312 | 91 | EXPORT_SYMBOL(first_online_pgdat_fn); |
5232d256 | 92 | # endif /* HAVE_FIRST_ONLINE_PGDAT */ |
36b313da | 93 | |
5232d256 | 94 | # ifndef HAVE_NEXT_ONLINE_PGDAT |
96dded38 | 95 | next_online_pgdat_t next_online_pgdat_fn = SYMBOL_POISON; |
d1ff2312 | 96 | EXPORT_SYMBOL(next_online_pgdat_fn); |
5232d256 | 97 | # endif /* HAVE_NEXT_ONLINE_PGDAT */ |
36b313da | 98 | |
5232d256 | 99 | # ifndef HAVE_NEXT_ZONE |
96dded38 | 100 | next_zone_t next_zone_fn = SYMBOL_POISON; |
d1ff2312 | 101 | EXPORT_SYMBOL(next_zone_fn); |
5232d256 BB |
102 | # endif /* HAVE_NEXT_ZONE */ |
103 | ||
104 | #else /* HAVE_PGDAT_HELPERS */ | |
105 | ||
106 | # ifndef HAVE_PGDAT_LIST | |
107 | struct pglist_data *pgdat_list_addr = SYMBOL_POISON; | |
108 | EXPORT_SYMBOL(pgdat_list_addr); | |
109 | # endif /* HAVE_PGDAT_LIST */ | |
110 | ||
111 | #endif /* HAVE_PGDAT_HELPERS */ | |
36b313da | 112 | |
6ae7fef5 | 113 | #ifdef NEED_GET_ZONE_COUNTS |
e11d6c5f | 114 | # ifndef HAVE_GET_ZONE_COUNTS |
96dded38 | 115 | get_zone_counts_t get_zone_counts_fn = SYMBOL_POISON; |
d1ff2312 | 116 | EXPORT_SYMBOL(get_zone_counts_fn); |
96dded38 | 117 | # endif /* HAVE_GET_ZONE_COUNTS */ |
4ab13d3b | 118 | |
e11d6c5f | 119 | unsigned long |
6ae7fef5 | 120 | spl_global_page_state(spl_zone_stat_item_t item) |
4ab13d3b BB |
121 | { |
122 | unsigned long active; | |
123 | unsigned long inactive; | |
124 | unsigned long free; | |
125 | ||
6ae7fef5 BB |
126 | get_zone_counts(&active, &inactive, &free); |
127 | switch (item) { | |
128 | case SPL_NR_FREE_PAGES: return free; | |
129 | case SPL_NR_INACTIVE: return inactive; | |
130 | case SPL_NR_ACTIVE: return active; | |
131 | default: ASSERT(0); /* Unsupported */ | |
e11d6c5f BB |
132 | } |
133 | ||
6ae7fef5 BB |
134 | return 0; |
135 | } | |
136 | #else | |
137 | # ifdef HAVE_GLOBAL_PAGE_STATE | |
138 | unsigned long | |
139 | spl_global_page_state(spl_zone_stat_item_t item) | |
140 | { | |
141 | unsigned long pages = 0; | |
142 | ||
143 | switch (item) { | |
144 | case SPL_NR_FREE_PAGES: | |
145 | # ifdef HAVE_ZONE_STAT_ITEM_NR_FREE_PAGES | |
146 | pages += global_page_state(NR_FREE_PAGES); | |
147 | # endif | |
148 | break; | |
149 | case SPL_NR_INACTIVE: | |
150 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE | |
151 | pages += global_page_state(NR_INACTIVE); | |
152 | # endif | |
153 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_ANON | |
154 | pages += global_page_state(NR_INACTIVE_ANON); | |
155 | # endif | |
156 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_FILE | |
157 | pages += global_page_state(NR_INACTIVE_FILE); | |
158 | # endif | |
159 | break; | |
160 | case SPL_NR_ACTIVE: | |
161 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE | |
162 | pages += global_page_state(NR_ACTIVE); | |
163 | # endif | |
164 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_ANON | |
165 | pages += global_page_state(NR_ACTIVE_ANON); | |
166 | # endif | |
167 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_FILE | |
168 | pages += global_page_state(NR_ACTIVE_FILE); | |
169 | # endif | |
170 | break; | |
171 | default: | |
172 | ASSERT(0); /* Unsupported */ | |
e11d6c5f BB |
173 | } |
174 | ||
6ae7fef5 BB |
175 | return pages; |
176 | } | |
96dded38 | 177 | # else |
6ae7fef5 | 178 | # error "Both global_page_state() and get_zone_counts() unavailable" |
96dded38 | 179 | # endif /* HAVE_GLOBAL_PAGE_STATE */ |
6ae7fef5 | 180 | #endif /* NEED_GET_ZONE_COUNTS */ |
e11d6c5f | 181 | EXPORT_SYMBOL(spl_global_page_state); |
4ab13d3b | 182 | |
e11d6c5f BB |
183 | pgcnt_t |
184 | spl_kmem_availrmem(void) | |
185 | { | |
4ab13d3b | 186 | /* The amount of easily available memory */ |
6ae7fef5 BB |
187 | return (spl_global_page_state(SPL_NR_FREE_PAGES) + |
188 | spl_global_page_state(SPL_NR_INACTIVE)); | |
4ab13d3b BB |
189 | } |
190 | EXPORT_SYMBOL(spl_kmem_availrmem); | |
191 | ||
192 | size_t | |
193 | vmem_size(vmem_t *vmp, int typemask) | |
194 | { | |
d1ff2312 BB |
195 | struct vmalloc_info vmi; |
196 | size_t size = 0; | |
197 | ||
4ab13d3b BB |
198 | ASSERT(vmp == NULL); |
199 | ASSERT(typemask & (VMEM_ALLOC | VMEM_FREE)); | |
200 | ||
d1ff2312 BB |
201 | get_vmalloc_info(&vmi); |
202 | if (typemask & VMEM_ALLOC) | |
203 | size += (size_t)vmi.used; | |
204 | ||
205 | if (typemask & VMEM_FREE) | |
206 | size += (size_t)(VMALLOC_TOTAL - vmi.used); | |
207 | ||
208 | return size; | |
4ab13d3b BB |
209 | } |
210 | EXPORT_SYMBOL(vmem_size); | |
4ab13d3b | 211 | |
b868e22f BB |
212 | int |
213 | kmem_debugging(void) | |
214 | { | |
215 | return 0; | |
216 | } | |
217 | EXPORT_SYMBOL(kmem_debugging); | |
218 | ||
219 | #ifndef HAVE_KVASPRINTF | |
220 | /* Simplified asprintf. */ | |
221 | char *kvasprintf(gfp_t gfp, const char *fmt, va_list ap) | |
222 | { | |
223 | unsigned int len; | |
224 | char *p; | |
225 | va_list aq; | |
226 | ||
227 | va_copy(aq, ap); | |
228 | len = vsnprintf(NULL, 0, fmt, aq); | |
229 | va_end(aq); | |
230 | ||
231 | p = kmalloc(len+1, gfp); | |
232 | if (!p) | |
233 | return NULL; | |
234 | ||
235 | vsnprintf(p, len+1, fmt, ap); | |
236 | ||
237 | return p; | |
238 | } | |
239 | EXPORT_SYMBOL(kvasprintf); | |
240 | #endif /* HAVE_KVASPRINTF */ | |
241 | ||
e6de04b7 BB |
242 | char * |
243 | kmem_vasprintf(const char *fmt, va_list ap) | |
244 | { | |
245 | va_list aq; | |
246 | char *ptr; | |
247 | ||
e6de04b7 | 248 | do { |
2c762de8 | 249 | va_copy(aq, ap); |
e6de04b7 | 250 | ptr = kvasprintf(GFP_KERNEL, fmt, aq); |
2c762de8 | 251 | va_end(aq); |
e6de04b7 | 252 | } while (ptr == NULL); |
e6de04b7 BB |
253 | |
254 | return ptr; | |
255 | } | |
256 | EXPORT_SYMBOL(kmem_vasprintf); | |
257 | ||
b868e22f BB |
258 | char * |
259 | kmem_asprintf(const char *fmt, ...) | |
260 | { | |
e6de04b7 | 261 | va_list ap; |
b868e22f BB |
262 | char *ptr; |
263 | ||
b868e22f | 264 | do { |
2c762de8 | 265 | va_start(ap, fmt); |
e6de04b7 | 266 | ptr = kvasprintf(GFP_KERNEL, fmt, ap); |
2c762de8 | 267 | va_end(ap); |
b868e22f | 268 | } while (ptr == NULL); |
b868e22f BB |
269 | |
270 | return ptr; | |
271 | } | |
272 | EXPORT_SYMBOL(kmem_asprintf); | |
273 | ||
10129680 BB |
274 | static char * |
275 | __strdup(const char *str, int flags) | |
276 | { | |
277 | char *ptr; | |
278 | int n; | |
279 | ||
280 | n = strlen(str); | |
281 | ptr = kmalloc_nofail(n + 1, flags); | |
282 | if (ptr) | |
283 | memcpy(ptr, str, n + 1); | |
284 | ||
285 | return ptr; | |
286 | } | |
287 | ||
288 | char * | |
289 | strdup(const char *str) | |
290 | { | |
291 | return __strdup(str, KM_SLEEP); | |
292 | } | |
293 | EXPORT_SYMBOL(strdup); | |
294 | ||
295 | void | |
296 | strfree(char *str) | |
297 | { | |
298 | kmem_free(str, strlen(str) + 1); | |
299 | } | |
300 | EXPORT_SYMBOL(strfree); | |
301 | ||
f1ca4da6 | 302 | /* |
2fb9b26a | 303 | * Memory allocation interfaces and debugging for basic kmem_* |
055ffd98 BB |
304 | * and vmem_* style memory allocation. When DEBUG_KMEM is enabled |
305 | * the SPL will keep track of the total memory allocated, and | |
306 | * report any memory leaked when the module is unloaded. | |
f1ca4da6 | 307 | */ |
308 | #ifdef DEBUG_KMEM | |
d04c8a56 | 309 | |
f1ca4da6 | 310 | /* Shim layer memory accounting */ |
d04c8a56 | 311 | # ifdef HAVE_ATOMIC64_T |
550f1705 | 312 | atomic64_t kmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 313 | unsigned long long kmem_alloc_max = 0; |
550f1705 | 314 | atomic64_t vmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 315 | unsigned long long vmem_alloc_max = 0; |
10129680 | 316 | # else /* HAVE_ATOMIC64_T */ |
d04c8a56 BB |
317 | atomic_t kmem_alloc_used = ATOMIC_INIT(0); |
318 | unsigned long long kmem_alloc_max = 0; | |
319 | atomic_t vmem_alloc_used = ATOMIC_INIT(0); | |
320 | unsigned long long vmem_alloc_max = 0; | |
10129680 | 321 | # endif /* HAVE_ATOMIC64_T */ |
79b31f36 | 322 | |
ff449ac4 | 323 | EXPORT_SYMBOL(kmem_alloc_used); |
324 | EXPORT_SYMBOL(kmem_alloc_max); | |
325 | EXPORT_SYMBOL(vmem_alloc_used); | |
326 | EXPORT_SYMBOL(vmem_alloc_max); | |
ff449ac4 | 327 | |
055ffd98 BB |
328 | /* When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked |
329 | * but also the location of every alloc and free. When the SPL module is | |
330 | * unloaded a list of all leaked addresses and where they were allocated | |
331 | * will be dumped to the console. Enabling this feature has a significant | |
332 | * impact on performance but it makes finding memory leaks straight forward. | |
333 | * | |
334 | * Not surprisingly with debugging enabled the xmem_locks are very highly | |
335 | * contended particularly on xfree(). If we want to run with this detailed | |
336 | * debugging enabled for anything other than debugging we need to minimize | |
337 | * the contention by moving to a lock per xmem_table entry model. | |
a0f6da3d | 338 | */ |
055ffd98 | 339 | # ifdef DEBUG_KMEM_TRACKING |
a0f6da3d | 340 | |
341 | # define KMEM_HASH_BITS 10 | |
342 | # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS) | |
343 | ||
344 | # define VMEM_HASH_BITS 10 | |
345 | # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS) | |
346 | ||
347 | typedef struct kmem_debug { | |
348 | struct hlist_node kd_hlist; /* Hash node linkage */ | |
349 | struct list_head kd_list; /* List of all allocations */ | |
350 | void *kd_addr; /* Allocation pointer */ | |
351 | size_t kd_size; /* Allocation size */ | |
352 | const char *kd_func; /* Allocation function */ | |
353 | int kd_line; /* Allocation line */ | |
354 | } kmem_debug_t; | |
355 | ||
d6a26c6a | 356 | spinlock_t kmem_lock; |
357 | struct hlist_head kmem_table[KMEM_TABLE_SIZE]; | |
358 | struct list_head kmem_list; | |
359 | ||
13cdca65 | 360 | spinlock_t vmem_lock; |
361 | struct hlist_head vmem_table[VMEM_TABLE_SIZE]; | |
362 | struct list_head vmem_list; | |
363 | ||
d6a26c6a | 364 | EXPORT_SYMBOL(kmem_lock); |
365 | EXPORT_SYMBOL(kmem_table); | |
366 | EXPORT_SYMBOL(kmem_list); | |
367 | ||
13cdca65 | 368 | EXPORT_SYMBOL(vmem_lock); |
369 | EXPORT_SYMBOL(vmem_table); | |
370 | EXPORT_SYMBOL(vmem_list); | |
a0f6da3d | 371 | |
372 | static kmem_debug_t * | |
10129680 | 373 | kmem_del_init(spinlock_t *lock, struct hlist_head *table, int bits, void *addr) |
a0f6da3d | 374 | { |
375 | struct hlist_head *head; | |
376 | struct hlist_node *node; | |
377 | struct kmem_debug *p; | |
378 | unsigned long flags; | |
b17edc10 | 379 | SENTRY; |
a0f6da3d | 380 | |
381 | spin_lock_irqsave(lock, flags); | |
382 | ||
383 | head = &table[hash_ptr(addr, bits)]; | |
384 | hlist_for_each_entry_rcu(p, node, head, kd_hlist) { | |
385 | if (p->kd_addr == addr) { | |
386 | hlist_del_init(&p->kd_hlist); | |
387 | list_del_init(&p->kd_list); | |
388 | spin_unlock_irqrestore(lock, flags); | |
389 | return p; | |
390 | } | |
391 | } | |
392 | ||
393 | spin_unlock_irqrestore(lock, flags); | |
394 | ||
b17edc10 | 395 | SRETURN(NULL); |
a0f6da3d | 396 | } |
397 | ||
398 | void * | |
399 | kmem_alloc_track(size_t size, int flags, const char *func, int line, | |
400 | int node_alloc, int node) | |
401 | { | |
402 | void *ptr = NULL; | |
403 | kmem_debug_t *dptr; | |
404 | unsigned long irq_flags; | |
b17edc10 | 405 | SENTRY; |
a0f6da3d | 406 | |
10129680 | 407 | /* Function may be called with KM_NOSLEEP so failure is possible */ |
c89fdee4 | 408 | dptr = (kmem_debug_t *) kmalloc_nofail(sizeof(kmem_debug_t), |
a0f6da3d | 409 | flags & ~__GFP_ZERO); |
410 | ||
10129680 | 411 | if (unlikely(dptr == NULL)) { |
b17edc10 | 412 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, "debug " |
3cb77549 BB |
413 | "kmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n", |
414 | sizeof(kmem_debug_t), flags, func, line, | |
415 | kmem_alloc_used_read(), kmem_alloc_max); | |
a0f6da3d | 416 | } else { |
10129680 BB |
417 | /* |
418 | * Marked unlikely because we should never be doing this, | |
419 | * we tolerate to up 2 pages but a single page is best. | |
420 | */ | |
23d91792 | 421 | if (unlikely((size > PAGE_SIZE*2) && !(flags & KM_NODEBUG))) { |
b17edc10 | 422 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, "large " |
3cb77549 BB |
423 | "kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n", |
424 | (unsigned long long) size, flags, func, line, | |
d04c8a56 | 425 | kmem_alloc_used_read(), kmem_alloc_max); |
5198ea0e BB |
426 | spl_debug_dumpstack(NULL); |
427 | } | |
a0f6da3d | 428 | |
10129680 BB |
429 | /* |
430 | * We use __strdup() below because the string pointed to by | |
c8e60837 | 431 | * __FUNCTION__ might not be available by the time we want |
10129680 BB |
432 | * to print it since the module might have been unloaded. |
433 | * This can only fail in the KM_NOSLEEP case. | |
434 | */ | |
435 | dptr->kd_func = __strdup(func, flags & ~__GFP_ZERO); | |
c8e60837 | 436 | if (unlikely(dptr->kd_func == NULL)) { |
437 | kfree(dptr); | |
b17edc10 | 438 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
10129680 | 439 | "debug __strdup() at %s:%d failed (%lld/%llu)\n", |
3cb77549 | 440 | func, line, kmem_alloc_used_read(), kmem_alloc_max); |
c8e60837 | 441 | goto out; |
442 | } | |
443 | ||
a0f6da3d | 444 | /* Use the correct allocator */ |
445 | if (node_alloc) { | |
446 | ASSERT(!(flags & __GFP_ZERO)); | |
c89fdee4 | 447 | ptr = kmalloc_node_nofail(size, flags, node); |
a0f6da3d | 448 | } else if (flags & __GFP_ZERO) { |
c89fdee4 | 449 | ptr = kzalloc_nofail(size, flags & ~__GFP_ZERO); |
a0f6da3d | 450 | } else { |
c89fdee4 | 451 | ptr = kmalloc_nofail(size, flags); |
a0f6da3d | 452 | } |
453 | ||
454 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 455 | kfree(dptr->kd_func); |
a0f6da3d | 456 | kfree(dptr); |
b17edc10 | 457 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, "kmem_alloc" |
3cb77549 BB |
458 | "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n", |
459 | (unsigned long long) size, flags, func, line, | |
d04c8a56 | 460 | kmem_alloc_used_read(), kmem_alloc_max); |
a0f6da3d | 461 | goto out; |
462 | } | |
463 | ||
d04c8a56 BB |
464 | kmem_alloc_used_add(size); |
465 | if (unlikely(kmem_alloc_used_read() > kmem_alloc_max)) | |
466 | kmem_alloc_max = kmem_alloc_used_read(); | |
a0f6da3d | 467 | |
468 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
469 | INIT_LIST_HEAD(&dptr->kd_list); | |
470 | ||
471 | dptr->kd_addr = ptr; | |
472 | dptr->kd_size = size; | |
a0f6da3d | 473 | dptr->kd_line = line; |
474 | ||
475 | spin_lock_irqsave(&kmem_lock, irq_flags); | |
476 | hlist_add_head_rcu(&dptr->kd_hlist, | |
477 | &kmem_table[hash_ptr(ptr, KMEM_HASH_BITS)]); | |
478 | list_add_tail(&dptr->kd_list, &kmem_list); | |
479 | spin_unlock_irqrestore(&kmem_lock, irq_flags); | |
480 | ||
b17edc10 | 481 | SDEBUG_LIMIT(SD_INFO, |
3cb77549 BB |
482 | "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n", |
483 | (unsigned long long) size, flags, func, line, ptr, | |
484 | kmem_alloc_used_read(), kmem_alloc_max); | |
a0f6da3d | 485 | } |
486 | out: | |
b17edc10 | 487 | SRETURN(ptr); |
a0f6da3d | 488 | } |
489 | EXPORT_SYMBOL(kmem_alloc_track); | |
490 | ||
491 | void | |
492 | kmem_free_track(void *ptr, size_t size) | |
493 | { | |
494 | kmem_debug_t *dptr; | |
b17edc10 | 495 | SENTRY; |
a0f6da3d | 496 | |
497 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
498 | (unsigned long long) size); | |
499 | ||
500 | dptr = kmem_del_init(&kmem_lock, kmem_table, KMEM_HASH_BITS, ptr); | |
501 | ||
10129680 BB |
502 | /* Must exist in hash due to kmem_alloc() */ |
503 | ASSERT(dptr); | |
a0f6da3d | 504 | |
505 | /* Size must match */ | |
506 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
507 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
508 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
509 | ||
d04c8a56 | 510 | kmem_alloc_used_sub(size); |
b17edc10 | 511 | SDEBUG_LIMIT(SD_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, |
d04c8a56 | 512 | (unsigned long long) size, kmem_alloc_used_read(), |
a0f6da3d | 513 | kmem_alloc_max); |
514 | ||
c8e60837 | 515 | kfree(dptr->kd_func); |
516 | ||
a0f6da3d | 517 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
518 | kfree(dptr); | |
519 | ||
520 | memset(ptr, 0x5a, size); | |
521 | kfree(ptr); | |
522 | ||
b17edc10 | 523 | SEXIT; |
a0f6da3d | 524 | } |
525 | EXPORT_SYMBOL(kmem_free_track); | |
526 | ||
527 | void * | |
528 | vmem_alloc_track(size_t size, int flags, const char *func, int line) | |
529 | { | |
530 | void *ptr = NULL; | |
531 | kmem_debug_t *dptr; | |
532 | unsigned long irq_flags; | |
b17edc10 | 533 | SENTRY; |
a0f6da3d | 534 | |
535 | ASSERT(flags & KM_SLEEP); | |
536 | ||
10129680 | 537 | /* Function may be called with KM_NOSLEEP so failure is possible */ |
ef1c7a06 BB |
538 | dptr = (kmem_debug_t *) kmalloc_nofail(sizeof(kmem_debug_t), |
539 | flags & ~__GFP_ZERO); | |
10129680 | 540 | if (unlikely(dptr == NULL)) { |
b17edc10 | 541 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, "debug " |
3cb77549 BB |
542 | "vmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n", |
543 | sizeof(kmem_debug_t), flags, func, line, | |
544 | vmem_alloc_used_read(), vmem_alloc_max); | |
a0f6da3d | 545 | } else { |
10129680 BB |
546 | /* |
547 | * We use __strdup() below because the string pointed to by | |
c8e60837 | 548 | * __FUNCTION__ might not be available by the time we want |
10129680 BB |
549 | * to print it, since the module might have been unloaded. |
550 | * This can never fail because we have already asserted | |
551 | * that flags is KM_SLEEP. | |
552 | */ | |
553 | dptr->kd_func = __strdup(func, flags & ~__GFP_ZERO); | |
c8e60837 | 554 | if (unlikely(dptr->kd_func == NULL)) { |
555 | kfree(dptr); | |
b17edc10 | 556 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
10129680 | 557 | "debug __strdup() at %s:%d failed (%lld/%llu)\n", |
3cb77549 | 558 | func, line, vmem_alloc_used_read(), vmem_alloc_max); |
c8e60837 | 559 | goto out; |
560 | } | |
561 | ||
10129680 BB |
562 | /* Use the correct allocator */ |
563 | if (flags & __GFP_ZERO) { | |
564 | ptr = vzalloc_nofail(size, flags & ~__GFP_ZERO); | |
565 | } else { | |
566 | ptr = vmalloc_nofail(size, flags); | |
567 | } | |
a0f6da3d | 568 | |
569 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 570 | kfree(dptr->kd_func); |
a0f6da3d | 571 | kfree(dptr); |
b17edc10 | 572 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, "vmem_alloc" |
3cb77549 BB |
573 | "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n", |
574 | (unsigned long long) size, flags, func, line, | |
d04c8a56 | 575 | vmem_alloc_used_read(), vmem_alloc_max); |
a0f6da3d | 576 | goto out; |
577 | } | |
578 | ||
d04c8a56 BB |
579 | vmem_alloc_used_add(size); |
580 | if (unlikely(vmem_alloc_used_read() > vmem_alloc_max)) | |
581 | vmem_alloc_max = vmem_alloc_used_read(); | |
a0f6da3d | 582 | |
583 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
584 | INIT_LIST_HEAD(&dptr->kd_list); | |
585 | ||
586 | dptr->kd_addr = ptr; | |
587 | dptr->kd_size = size; | |
a0f6da3d | 588 | dptr->kd_line = line; |
589 | ||
590 | spin_lock_irqsave(&vmem_lock, irq_flags); | |
591 | hlist_add_head_rcu(&dptr->kd_hlist, | |
592 | &vmem_table[hash_ptr(ptr, VMEM_HASH_BITS)]); | |
593 | list_add_tail(&dptr->kd_list, &vmem_list); | |
594 | spin_unlock_irqrestore(&vmem_lock, irq_flags); | |
595 | ||
b17edc10 | 596 | SDEBUG_LIMIT(SD_INFO, |
3cb77549 BB |
597 | "vmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n", |
598 | (unsigned long long) size, flags, func, line, | |
599 | ptr, vmem_alloc_used_read(), vmem_alloc_max); | |
a0f6da3d | 600 | } |
601 | out: | |
b17edc10 | 602 | SRETURN(ptr); |
a0f6da3d | 603 | } |
604 | EXPORT_SYMBOL(vmem_alloc_track); | |
605 | ||
606 | void | |
607 | vmem_free_track(void *ptr, size_t size) | |
608 | { | |
609 | kmem_debug_t *dptr; | |
b17edc10 | 610 | SENTRY; |
a0f6da3d | 611 | |
612 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
613 | (unsigned long long) size); | |
614 | ||
615 | dptr = kmem_del_init(&vmem_lock, vmem_table, VMEM_HASH_BITS, ptr); | |
10129680 BB |
616 | |
617 | /* Must exist in hash due to vmem_alloc() */ | |
618 | ASSERT(dptr); | |
a0f6da3d | 619 | |
620 | /* Size must match */ | |
621 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
622 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
623 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
624 | ||
d04c8a56 | 625 | vmem_alloc_used_sub(size); |
b17edc10 | 626 | SDEBUG_LIMIT(SD_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, |
d04c8a56 | 627 | (unsigned long long) size, vmem_alloc_used_read(), |
a0f6da3d | 628 | vmem_alloc_max); |
629 | ||
c8e60837 | 630 | kfree(dptr->kd_func); |
631 | ||
a0f6da3d | 632 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
633 | kfree(dptr); | |
634 | ||
635 | memset(ptr, 0x5a, size); | |
636 | vfree(ptr); | |
637 | ||
b17edc10 | 638 | SEXIT; |
a0f6da3d | 639 | } |
640 | EXPORT_SYMBOL(vmem_free_track); | |
641 | ||
642 | # else /* DEBUG_KMEM_TRACKING */ | |
643 | ||
644 | void * | |
645 | kmem_alloc_debug(size_t size, int flags, const char *func, int line, | |
646 | int node_alloc, int node) | |
647 | { | |
648 | void *ptr; | |
b17edc10 | 649 | SENTRY; |
a0f6da3d | 650 | |
10129680 BB |
651 | /* |
652 | * Marked unlikely because we should never be doing this, | |
653 | * we tolerate to up 2 pages but a single page is best. | |
654 | */ | |
23d91792 | 655 | if (unlikely((size > PAGE_SIZE * 2) && !(flags & KM_NODEBUG))) { |
b17edc10 | 656 | SDEBUG(SD_CONSOLE | SD_WARNING, |
10129680 | 657 | "large kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n", |
3cb77549 | 658 | (unsigned long long) size, flags, func, line, |
d04c8a56 | 659 | kmem_alloc_used_read(), kmem_alloc_max); |
5198ea0e BB |
660 | spl_debug_dumpstack(NULL); |
661 | } | |
a0f6da3d | 662 | |
663 | /* Use the correct allocator */ | |
664 | if (node_alloc) { | |
665 | ASSERT(!(flags & __GFP_ZERO)); | |
c89fdee4 | 666 | ptr = kmalloc_node_nofail(size, flags, node); |
a0f6da3d | 667 | } else if (flags & __GFP_ZERO) { |
c89fdee4 | 668 | ptr = kzalloc_nofail(size, flags & (~__GFP_ZERO)); |
a0f6da3d | 669 | } else { |
c89fdee4 | 670 | ptr = kmalloc_nofail(size, flags); |
a0f6da3d | 671 | } |
672 | ||
10129680 | 673 | if (unlikely(ptr == NULL)) { |
b17edc10 | 674 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
3cb77549 BB |
675 | "kmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n", |
676 | (unsigned long long) size, flags, func, line, | |
d04c8a56 | 677 | kmem_alloc_used_read(), kmem_alloc_max); |
a0f6da3d | 678 | } else { |
d04c8a56 BB |
679 | kmem_alloc_used_add(size); |
680 | if (unlikely(kmem_alloc_used_read() > kmem_alloc_max)) | |
681 | kmem_alloc_max = kmem_alloc_used_read(); | |
a0f6da3d | 682 | |
b17edc10 | 683 | SDEBUG_LIMIT(SD_INFO, |
3cb77549 BB |
684 | "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n", |
685 | (unsigned long long) size, flags, func, line, ptr, | |
10129680 | 686 | kmem_alloc_used_read(), kmem_alloc_max); |
a0f6da3d | 687 | } |
10129680 | 688 | |
b17edc10 | 689 | SRETURN(ptr); |
a0f6da3d | 690 | } |
691 | EXPORT_SYMBOL(kmem_alloc_debug); | |
692 | ||
693 | void | |
694 | kmem_free_debug(void *ptr, size_t size) | |
695 | { | |
b17edc10 | 696 | SENTRY; |
a0f6da3d | 697 | |
698 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
699 | (unsigned long long) size); | |
700 | ||
d04c8a56 | 701 | kmem_alloc_used_sub(size); |
b17edc10 | 702 | SDEBUG_LIMIT(SD_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, |
d04c8a56 | 703 | (unsigned long long) size, kmem_alloc_used_read(), |
a0f6da3d | 704 | kmem_alloc_max); |
a0f6da3d | 705 | kfree(ptr); |
706 | ||
b17edc10 | 707 | SEXIT; |
a0f6da3d | 708 | } |
709 | EXPORT_SYMBOL(kmem_free_debug); | |
710 | ||
711 | void * | |
712 | vmem_alloc_debug(size_t size, int flags, const char *func, int line) | |
713 | { | |
714 | void *ptr; | |
b17edc10 | 715 | SENTRY; |
a0f6da3d | 716 | |
717 | ASSERT(flags & KM_SLEEP); | |
718 | ||
10129680 BB |
719 | /* Use the correct allocator */ |
720 | if (flags & __GFP_ZERO) { | |
721 | ptr = vzalloc_nofail(size, flags & (~__GFP_ZERO)); | |
722 | } else { | |
723 | ptr = vmalloc_nofail(size, flags); | |
724 | } | |
725 | ||
726 | if (unlikely(ptr == NULL)) { | |
b17edc10 | 727 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
3cb77549 BB |
728 | "vmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n", |
729 | (unsigned long long) size, flags, func, line, | |
d04c8a56 | 730 | vmem_alloc_used_read(), vmem_alloc_max); |
a0f6da3d | 731 | } else { |
d04c8a56 BB |
732 | vmem_alloc_used_add(size); |
733 | if (unlikely(vmem_alloc_used_read() > vmem_alloc_max)) | |
734 | vmem_alloc_max = vmem_alloc_used_read(); | |
a0f6da3d | 735 | |
b17edc10 | 736 | SDEBUG_LIMIT(SD_INFO, "vmem_alloc(%llu, 0x%x) = %p " |
a0f6da3d | 737 | "(%lld/%llu)\n", (unsigned long long) size, flags, ptr, |
d04c8a56 | 738 | vmem_alloc_used_read(), vmem_alloc_max); |
a0f6da3d | 739 | } |
740 | ||
b17edc10 | 741 | SRETURN(ptr); |
a0f6da3d | 742 | } |
743 | EXPORT_SYMBOL(vmem_alloc_debug); | |
744 | ||
745 | void | |
746 | vmem_free_debug(void *ptr, size_t size) | |
747 | { | |
b17edc10 | 748 | SENTRY; |
a0f6da3d | 749 | |
750 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
751 | (unsigned long long) size); | |
752 | ||
d04c8a56 | 753 | vmem_alloc_used_sub(size); |
b17edc10 | 754 | SDEBUG_LIMIT(SD_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, |
d04c8a56 | 755 | (unsigned long long) size, vmem_alloc_used_read(), |
a0f6da3d | 756 | vmem_alloc_max); |
a0f6da3d | 757 | vfree(ptr); |
758 | ||
b17edc10 | 759 | SEXIT; |
a0f6da3d | 760 | } |
761 | EXPORT_SYMBOL(vmem_free_debug); | |
762 | ||
763 | # endif /* DEBUG_KMEM_TRACKING */ | |
764 | #endif /* DEBUG_KMEM */ | |
765 | ||
10129680 BB |
766 | /* |
767 | * Slab allocation interfaces | |
768 | * | |
769 | * While the Linux slab implementation was inspired by the Solaris | |
770 | * implemenation I cannot use it to emulate the Solaris APIs. I | |
771 | * require two features which are not provided by the Linux slab. | |
772 | * | |
773 | * 1) Constructors AND destructors. Recent versions of the Linux | |
774 | * kernel have removed support for destructors. This is a deal | |
775 | * breaker for the SPL which contains particularly expensive | |
776 | * initializers for mutex's, condition variables, etc. We also | |
777 | * require a minimal level of cleanup for these data types unlike | |
778 | * many Linux data type which do need to be explicitly destroyed. | |
779 | * | |
780 | * 2) Virtual address space backed slab. Callers of the Solaris slab | |
781 | * expect it to work well for both small are very large allocations. | |
782 | * Because of memory fragmentation the Linux slab which is backed | |
783 | * by kmalloc'ed memory performs very badly when confronted with | |
784 | * large numbers of large allocations. Basing the slab on the | |
785 | * virtual address space removes the need for contigeous pages | |
786 | * and greatly improve performance for large allocations. | |
787 | * | |
788 | * For these reasons, the SPL has its own slab implementation with | |
789 | * the needed features. It is not as highly optimized as either the | |
790 | * Solaris or Linux slabs, but it should get me most of what is | |
791 | * needed until it can be optimized or obsoleted by another approach. | |
792 | * | |
793 | * One serious concern I do have about this method is the relatively | |
794 | * small virtual address space on 32bit arches. This will seriously | |
795 | * constrain the size of the slab caches and their performance. | |
796 | * | |
797 | * XXX: Improve the partial slab list by carefully maintaining a | |
798 | * strict ordering of fullest to emptiest slabs based on | |
799 | * the slab reference count. This gaurentees the when freeing | |
800 | * slabs back to the system we need only linearly traverse the | |
801 | * last N slabs in the list to discover all the freeable slabs. | |
802 | * | |
803 | * XXX: NUMA awareness for optionally allocating memory close to a | |
804 | * particular core. This can be adventageous if you know the slab | |
805 | * object will be short lived and primarily accessed from one core. | |
806 | * | |
807 | * XXX: Slab coloring may also yield performance improvements and would | |
808 | * be desirable to implement. | |
809 | */ | |
810 | ||
811 | struct list_head spl_kmem_cache_list; /* List of caches */ | |
812 | struct rw_semaphore spl_kmem_cache_sem; /* Cache list lock */ | |
813 | ||
814 | static int spl_cache_flush(spl_kmem_cache_t *skc, | |
815 | spl_kmem_magazine_t *skm, int flush); | |
816 | ||
817 | #ifdef HAVE_SET_SHRINKER | |
818 | static struct shrinker *spl_kmem_cache_shrinker; | |
819 | #else | |
820 | static int spl_kmem_cache_generic_shrinker(int nr_to_scan, | |
821 | unsigned int gfp_mask); | |
822 | static struct shrinker spl_kmem_cache_shrinker = { | |
823 | .shrink = spl_kmem_cache_generic_shrinker, | |
824 | .seeks = KMC_DEFAULT_SEEKS, | |
825 | }; | |
826 | #endif | |
827 | ||
a1502d76 | 828 | static void * |
829 | kv_alloc(spl_kmem_cache_t *skc, int size, int flags) | |
fece7c99 | 830 | { |
a1502d76 | 831 | void *ptr; |
f1ca4da6 | 832 | |
8b45dda2 BB |
833 | ASSERT(ISP2(size)); |
834 | ||
835 | if (skc->skc_flags & KMC_KMEM) | |
836 | ptr = (void *)__get_free_pages(flags, get_order(size)); | |
837 | else | |
838 | ptr = __vmalloc(size, flags | __GFP_HIGHMEM, PAGE_KERNEL); | |
839 | ||
840 | /* Resulting allocated memory will be page aligned */ | |
841 | ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE)); | |
fece7c99 | 842 | |
a1502d76 | 843 | return ptr; |
844 | } | |
fece7c99 | 845 | |
a1502d76 | 846 | static void |
847 | kv_free(spl_kmem_cache_t *skc, void *ptr, int size) | |
848 | { | |
8b45dda2 BB |
849 | ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE)); |
850 | ASSERT(ISP2(size)); | |
851 | ||
852 | if (skc->skc_flags & KMC_KMEM) | |
853 | free_pages((unsigned long)ptr, get_order(size)); | |
854 | else | |
855 | vfree(ptr); | |
856 | } | |
857 | ||
858 | /* | |
859 | * Required space for each aligned sks. | |
860 | */ | |
861 | static inline uint32_t | |
862 | spl_sks_size(spl_kmem_cache_t *skc) | |
863 | { | |
864 | return P2ROUNDUP_TYPED(sizeof(spl_kmem_slab_t), | |
865 | skc->skc_obj_align, uint32_t); | |
866 | } | |
867 | ||
868 | /* | |
869 | * Required space for each aligned object. | |
870 | */ | |
871 | static inline uint32_t | |
872 | spl_obj_size(spl_kmem_cache_t *skc) | |
873 | { | |
874 | uint32_t align = skc->skc_obj_align; | |
875 | ||
876 | return P2ROUNDUP_TYPED(skc->skc_obj_size, align, uint32_t) + | |
877 | P2ROUNDUP_TYPED(sizeof(spl_kmem_obj_t), align, uint32_t); | |
878 | } | |
879 | ||
880 | /* | |
881 | * Lookup the spl_kmem_object_t for an object given that object. | |
882 | */ | |
883 | static inline spl_kmem_obj_t * | |
884 | spl_sko_from_obj(spl_kmem_cache_t *skc, void *obj) | |
885 | { | |
886 | return obj + P2ROUNDUP_TYPED(skc->skc_obj_size, | |
887 | skc->skc_obj_align, uint32_t); | |
888 | } | |
889 | ||
890 | /* | |
891 | * Required space for each offslab object taking in to account alignment | |
892 | * restrictions and the power-of-two requirement of kv_alloc(). | |
893 | */ | |
894 | static inline uint32_t | |
895 | spl_offslab_size(spl_kmem_cache_t *skc) | |
896 | { | |
897 | return 1UL << (highbit(spl_obj_size(skc)) + 1); | |
fece7c99 | 898 | } |
899 | ||
ea3e6ca9 BB |
900 | /* |
901 | * It's important that we pack the spl_kmem_obj_t structure and the | |
48e0606a BB |
902 | * actual objects in to one large address space to minimize the number |
903 | * of calls to the allocator. It is far better to do a few large | |
904 | * allocations and then subdivide it ourselves. Now which allocator | |
905 | * we use requires balancing a few trade offs. | |
906 | * | |
907 | * For small objects we use kmem_alloc() because as long as you are | |
908 | * only requesting a small number of pages (ideally just one) its cheap. | |
909 | * However, when you start requesting multiple pages with kmem_alloc() | |
910 | * it gets increasingly expensive since it requires contigeous pages. | |
911 | * For this reason we shift to vmem_alloc() for slabs of large objects | |
912 | * which removes the need for contigeous pages. We do not use | |
913 | * vmem_alloc() in all cases because there is significant locking | |
914 | * overhead in __get_vm_area_node(). This function takes a single | |
915 | * global lock when aquiring an available virtual address range which | |
916 | * serializes all vmem_alloc()'s for all slab caches. Using slightly | |
917 | * different allocation functions for small and large objects should | |
918 | * give us the best of both worlds. | |
919 | * | |
920 | * KMC_ONSLAB KMC_OFFSLAB | |
921 | * | |
922 | * +------------------------+ +-----------------+ | |
923 | * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+ | |
924 | * | skc_obj_size <-+ | | +-----------------+ | | | |
925 | * | spl_kmem_obj_t | | | | | |
926 | * | skc_obj_size <---+ | +-----------------+ | | | |
927 | * | spl_kmem_obj_t | | | skc_obj_size | <-+ | | |
928 | * | ... v | | spl_kmem_obj_t | | | |
929 | * +------------------------+ +-----------------+ v | |
930 | */ | |
fece7c99 | 931 | static spl_kmem_slab_t * |
a1502d76 | 932 | spl_slab_alloc(spl_kmem_cache_t *skc, int flags) |
fece7c99 | 933 | { |
934 | spl_kmem_slab_t *sks; | |
a1502d76 | 935 | spl_kmem_obj_t *sko, *n; |
936 | void *base, *obj; | |
8b45dda2 BB |
937 | uint32_t obj_size, offslab_size = 0; |
938 | int i, rc = 0; | |
48e0606a | 939 | |
a1502d76 | 940 | base = kv_alloc(skc, skc->skc_slab_size, flags); |
941 | if (base == NULL) | |
b17edc10 | 942 | SRETURN(NULL); |
fece7c99 | 943 | |
a1502d76 | 944 | sks = (spl_kmem_slab_t *)base; |
945 | sks->sks_magic = SKS_MAGIC; | |
946 | sks->sks_objs = skc->skc_slab_objs; | |
947 | sks->sks_age = jiffies; | |
948 | sks->sks_cache = skc; | |
949 | INIT_LIST_HEAD(&sks->sks_list); | |
950 | INIT_LIST_HEAD(&sks->sks_free_list); | |
951 | sks->sks_ref = 0; | |
8b45dda2 | 952 | obj_size = spl_obj_size(skc); |
48e0606a | 953 | |
8b45dda2 BB |
954 | if (skc->skc_flags * KMC_OFFSLAB) |
955 | offslab_size = spl_offslab_size(skc); | |
fece7c99 | 956 | |
957 | for (i = 0; i < sks->sks_objs; i++) { | |
a1502d76 | 958 | if (skc->skc_flags & KMC_OFFSLAB) { |
8b45dda2 | 959 | obj = kv_alloc(skc, offslab_size, flags); |
a1502d76 | 960 | if (!obj) |
b17edc10 | 961 | SGOTO(out, rc = -ENOMEM); |
a1502d76 | 962 | } else { |
8b45dda2 | 963 | obj = base + spl_sks_size(skc) + (i * obj_size); |
a1502d76 | 964 | } |
965 | ||
8b45dda2 BB |
966 | ASSERT(IS_P2ALIGNED(obj, skc->skc_obj_align)); |
967 | sko = spl_sko_from_obj(skc, obj); | |
fece7c99 | 968 | sko->sko_addr = obj; |
969 | sko->sko_magic = SKO_MAGIC; | |
970 | sko->sko_slab = sks; | |
971 | INIT_LIST_HEAD(&sko->sko_list); | |
fece7c99 | 972 | list_add_tail(&sko->sko_list, &sks->sks_free_list); |
973 | } | |
974 | ||
fece7c99 | 975 | list_for_each_entry(sko, &sks->sks_free_list, sko_list) |
976 | if (skc->skc_ctor) | |
977 | skc->skc_ctor(sko->sko_addr, skc->skc_private, flags); | |
2fb9b26a | 978 | out: |
a1502d76 | 979 | if (rc) { |
980 | if (skc->skc_flags & KMC_OFFSLAB) | |
48e0606a BB |
981 | list_for_each_entry_safe(sko, n, &sks->sks_free_list, |
982 | sko_list) | |
8b45dda2 | 983 | kv_free(skc, sko->sko_addr, offslab_size); |
fece7c99 | 984 | |
a1502d76 | 985 | kv_free(skc, base, skc->skc_slab_size); |
986 | sks = NULL; | |
fece7c99 | 987 | } |
988 | ||
b17edc10 | 989 | SRETURN(sks); |
fece7c99 | 990 | } |
991 | ||
ea3e6ca9 BB |
992 | /* |
993 | * Remove a slab from complete or partial list, it must be called with | |
994 | * the 'skc->skc_lock' held but the actual free must be performed | |
995 | * outside the lock to prevent deadlocking on vmem addresses. | |
fece7c99 | 996 | */ |
f1ca4da6 | 997 | static void |
ea3e6ca9 BB |
998 | spl_slab_free(spl_kmem_slab_t *sks, |
999 | struct list_head *sks_list, struct list_head *sko_list) | |
1000 | { | |
2fb9b26a | 1001 | spl_kmem_cache_t *skc; |
b17edc10 | 1002 | SENTRY; |
57d86234 | 1003 | |
2fb9b26a | 1004 | ASSERT(sks->sks_magic == SKS_MAGIC); |
4afaaefa | 1005 | ASSERT(sks->sks_ref == 0); |
d6a26c6a | 1006 | |
fece7c99 | 1007 | skc = sks->sks_cache; |
1008 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
d46630e0 | 1009 | ASSERT(spin_is_locked(&skc->skc_lock)); |
f1ca4da6 | 1010 | |
1a944a7d BB |
1011 | /* |
1012 | * Update slab/objects counters in the cache, then remove the | |
1013 | * slab from the skc->skc_partial_list. Finally add the slab | |
1014 | * and all its objects in to the private work lists where the | |
1015 | * destructors will be called and the memory freed to the system. | |
1016 | */ | |
fece7c99 | 1017 | skc->skc_obj_total -= sks->sks_objs; |
1018 | skc->skc_slab_total--; | |
1019 | list_del(&sks->sks_list); | |
ea3e6ca9 | 1020 | list_add(&sks->sks_list, sks_list); |
1a944a7d BB |
1021 | list_splice_init(&sks->sks_free_list, sko_list); |
1022 | ||
b17edc10 | 1023 | SEXIT; |
2fb9b26a | 1024 | } |
d6a26c6a | 1025 | |
ea3e6ca9 BB |
1026 | /* |
1027 | * Traverses all the partial slabs attached to a cache and free those | |
1028 | * which which are currently empty, and have not been touched for | |
37db7d8c BB |
1029 | * skc_delay seconds to avoid thrashing. The count argument is |
1030 | * passed to optionally cap the number of slabs reclaimed, a count | |
1031 | * of zero means try and reclaim everything. When flag is set we | |
1032 | * always free an available slab regardless of age. | |
ea3e6ca9 BB |
1033 | */ |
1034 | static void | |
37db7d8c | 1035 | spl_slab_reclaim(spl_kmem_cache_t *skc, int count, int flag) |
2fb9b26a | 1036 | { |
1037 | spl_kmem_slab_t *sks, *m; | |
ea3e6ca9 BB |
1038 | spl_kmem_obj_t *sko, *n; |
1039 | LIST_HEAD(sks_list); | |
1040 | LIST_HEAD(sko_list); | |
8b45dda2 BB |
1041 | uint32_t size = 0; |
1042 | int i = 0; | |
b17edc10 | 1043 | SENTRY; |
2fb9b26a | 1044 | |
2fb9b26a | 1045 | /* |
ea3e6ca9 BB |
1046 | * Move empty slabs and objects which have not been touched in |
1047 | * skc_delay seconds on to private lists to be freed outside | |
1a944a7d BB |
1048 | * the spin lock. This delay time is important to avoid thrashing |
1049 | * however when flag is set the delay will not be used. | |
2fb9b26a | 1050 | */ |
ea3e6ca9 | 1051 | spin_lock(&skc->skc_lock); |
1a944a7d BB |
1052 | list_for_each_entry_safe_reverse(sks,m,&skc->skc_partial_list,sks_list){ |
1053 | /* | |
1054 | * All empty slabs are at the end of skc->skc_partial_list, | |
1055 | * therefore once a non-empty slab is found we can stop | |
1056 | * scanning. Additionally, stop when reaching the target | |
1057 | * reclaim 'count' if a non-zero threshhold is given. | |
1058 | */ | |
1059 | if ((sks->sks_ref > 0) || (count && i > count)) | |
37db7d8c BB |
1060 | break; |
1061 | ||
37db7d8c | 1062 | if (time_after(jiffies,sks->sks_age+skc->skc_delay*HZ)||flag) { |
ea3e6ca9 | 1063 | spl_slab_free(sks, &sks_list, &sko_list); |
37db7d8c BB |
1064 | i++; |
1065 | } | |
ea3e6ca9 BB |
1066 | } |
1067 | spin_unlock(&skc->skc_lock); | |
1068 | ||
1069 | /* | |
1a944a7d BB |
1070 | * The following two loops ensure all the object destructors are |
1071 | * run, any offslab objects are freed, and the slabs themselves | |
1072 | * are freed. This is all done outside the skc->skc_lock since | |
1073 | * this allows the destructor to sleep, and allows us to perform | |
1074 | * a conditional reschedule when a freeing a large number of | |
1075 | * objects and slabs back to the system. | |
ea3e6ca9 | 1076 | */ |
1a944a7d | 1077 | if (skc->skc_flags & KMC_OFFSLAB) |
8b45dda2 | 1078 | size = spl_offslab_size(skc); |
ea3e6ca9 | 1079 | |
1a944a7d BB |
1080 | list_for_each_entry_safe(sko, n, &sko_list, sko_list) { |
1081 | ASSERT(sko->sko_magic == SKO_MAGIC); | |
1082 | ||
1083 | if (skc->skc_dtor) | |
1084 | skc->skc_dtor(sko->sko_addr, skc->skc_private); | |
1085 | ||
1086 | if (skc->skc_flags & KMC_OFFSLAB) | |
ea3e6ca9 | 1087 | kv_free(skc, sko->sko_addr, size); |
1a944a7d BB |
1088 | |
1089 | cond_resched(); | |
2fb9b26a | 1090 | } |
1091 | ||
37db7d8c | 1092 | list_for_each_entry_safe(sks, m, &sks_list, sks_list) { |
1a944a7d | 1093 | ASSERT(sks->sks_magic == SKS_MAGIC); |
ea3e6ca9 | 1094 | kv_free(skc, sks, skc->skc_slab_size); |
37db7d8c BB |
1095 | cond_resched(); |
1096 | } | |
ea3e6ca9 | 1097 | |
b17edc10 | 1098 | SEXIT; |
f1ca4da6 | 1099 | } |
1100 | ||
ea3e6ca9 BB |
1101 | /* |
1102 | * Called regularly on all caches to age objects out of the magazines | |
1103 | * which have not been access in skc->skc_delay seconds. This prevents | |
1104 | * idle magazines from holding memory which might be better used by | |
1105 | * other caches or parts of the system. The delay is present to | |
1106 | * prevent thrashing the magazine. | |
1107 | */ | |
1108 | static void | |
1109 | spl_magazine_age(void *data) | |
f1ca4da6 | 1110 | { |
9b1b8e4c BB |
1111 | spl_kmem_magazine_t *skm = |
1112 | spl_get_work_data(data, spl_kmem_magazine_t, skm_work.work); | |
1113 | spl_kmem_cache_t *skc = skm->skm_cache; | |
1114 | int i = smp_processor_id(); | |
1115 | ||
1116 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1117 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
1118 | ASSERT(skc->skc_mag[i] == skm); | |
f1ca4da6 | 1119 | |
ea3e6ca9 BB |
1120 | if (skm->skm_avail > 0 && |
1121 | time_after(jiffies, skm->skm_age + skc->skc_delay * HZ)) | |
1122 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
9b1b8e4c BB |
1123 | |
1124 | if (!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)) | |
1125 | schedule_delayed_work_on(i, &skm->skm_work, | |
1126 | skc->skc_delay / 3 * HZ); | |
ea3e6ca9 | 1127 | } |
4efd4118 | 1128 | |
ea3e6ca9 BB |
1129 | /* |
1130 | * Called regularly to keep a downward pressure on the size of idle | |
1131 | * magazines and to release free slabs from the cache. This function | |
1132 | * never calls the registered reclaim function, that only occures | |
1133 | * under memory pressure or with a direct call to spl_kmem_reap(). | |
1134 | */ | |
1135 | static void | |
1136 | spl_cache_age(void *data) | |
1137 | { | |
9b1b8e4c | 1138 | spl_kmem_cache_t *skc = |
ea3e6ca9 BB |
1139 | spl_get_work_data(data, spl_kmem_cache_t, skc_work.work); |
1140 | ||
1141 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
37db7d8c | 1142 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
ea3e6ca9 BB |
1143 | |
1144 | if (!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)) | |
37db7d8c | 1145 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
2fb9b26a | 1146 | } |
f1ca4da6 | 1147 | |
ea3e6ca9 | 1148 | /* |
8b45dda2 | 1149 | * Size a slab based on the size of each aligned object plus spl_kmem_obj_t. |
ea3e6ca9 BB |
1150 | * When on-slab we want to target SPL_KMEM_CACHE_OBJ_PER_SLAB. However, |
1151 | * for very small objects we may end up with more than this so as not | |
1152 | * to waste space in the minimal allocation of a single page. Also for | |
1153 | * very large objects we may use as few as SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN, | |
1154 | * lower than this and we will fail. | |
1155 | */ | |
48e0606a BB |
1156 | static int |
1157 | spl_slab_size(spl_kmem_cache_t *skc, uint32_t *objs, uint32_t *size) | |
1158 | { | |
8b45dda2 | 1159 | uint32_t sks_size, obj_size, max_size; |
48e0606a BB |
1160 | |
1161 | if (skc->skc_flags & KMC_OFFSLAB) { | |
ea3e6ca9 | 1162 | *objs = SPL_KMEM_CACHE_OBJ_PER_SLAB; |
48e0606a BB |
1163 | *size = sizeof(spl_kmem_slab_t); |
1164 | } else { | |
8b45dda2 BB |
1165 | sks_size = spl_sks_size(skc); |
1166 | obj_size = spl_obj_size(skc); | |
ea3e6ca9 BB |
1167 | |
1168 | if (skc->skc_flags & KMC_KMEM) | |
aa600d8a | 1169 | max_size = ((uint32_t)1 << (MAX_ORDER-3)) * PAGE_SIZE; |
ea3e6ca9 BB |
1170 | else |
1171 | max_size = (32 * 1024 * 1024); | |
48e0606a | 1172 | |
8b45dda2 BB |
1173 | /* Power of two sized slab */ |
1174 | for (*size = PAGE_SIZE; *size <= max_size; *size *= 2) { | |
ea3e6ca9 BB |
1175 | *objs = (*size - sks_size) / obj_size; |
1176 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB) | |
b17edc10 | 1177 | SRETURN(0); |
ea3e6ca9 | 1178 | } |
48e0606a | 1179 | |
ea3e6ca9 | 1180 | /* |
8b45dda2 | 1181 | * Unable to satisfy target objects per slab, fall back to |
ea3e6ca9 BB |
1182 | * allocating a maximally sized slab and assuming it can |
1183 | * contain the minimum objects count use it. If not fail. | |
1184 | */ | |
1185 | *size = max_size; | |
1186 | *objs = (*size - sks_size) / obj_size; | |
1187 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN) | |
b17edc10 | 1188 | SRETURN(0); |
48e0606a BB |
1189 | } |
1190 | ||
b17edc10 | 1191 | SRETURN(-ENOSPC); |
48e0606a BB |
1192 | } |
1193 | ||
ea3e6ca9 BB |
1194 | /* |
1195 | * Make a guess at reasonable per-cpu magazine size based on the size of | |
1196 | * each object and the cost of caching N of them in each magazine. Long | |
1197 | * term this should really adapt based on an observed usage heuristic. | |
1198 | */ | |
4afaaefa | 1199 | static int |
1200 | spl_magazine_size(spl_kmem_cache_t *skc) | |
1201 | { | |
8b45dda2 BB |
1202 | uint32_t obj_size = spl_obj_size(skc); |
1203 | int size; | |
b17edc10 | 1204 | SENTRY; |
4afaaefa | 1205 | |
ea3e6ca9 | 1206 | /* Per-magazine sizes below assume a 4Kib page size */ |
8b45dda2 | 1207 | if (obj_size > (PAGE_SIZE * 256)) |
ea3e6ca9 | 1208 | size = 4; /* Minimum 4Mib per-magazine */ |
8b45dda2 | 1209 | else if (obj_size > (PAGE_SIZE * 32)) |
ea3e6ca9 | 1210 | size = 16; /* Minimum 2Mib per-magazine */ |
8b45dda2 | 1211 | else if (obj_size > (PAGE_SIZE)) |
ea3e6ca9 | 1212 | size = 64; /* Minimum 256Kib per-magazine */ |
8b45dda2 | 1213 | else if (obj_size > (PAGE_SIZE / 4)) |
ea3e6ca9 | 1214 | size = 128; /* Minimum 128Kib per-magazine */ |
4afaaefa | 1215 | else |
ea3e6ca9 | 1216 | size = 256; |
4afaaefa | 1217 | |
b17edc10 | 1218 | SRETURN(size); |
4afaaefa | 1219 | } |
1220 | ||
ea3e6ca9 BB |
1221 | /* |
1222 | * Allocate a per-cpu magazine to assoicate with a specific core. | |
1223 | */ | |
4afaaefa | 1224 | static spl_kmem_magazine_t * |
1225 | spl_magazine_alloc(spl_kmem_cache_t *skc, int node) | |
1226 | { | |
1227 | spl_kmem_magazine_t *skm; | |
1228 | int size = sizeof(spl_kmem_magazine_t) + | |
1229 | sizeof(void *) * skc->skc_mag_size; | |
b17edc10 | 1230 | SENTRY; |
4afaaefa | 1231 | |
c89fdee4 | 1232 | skm = kmem_alloc_node(size, KM_SLEEP, node); |
4afaaefa | 1233 | if (skm) { |
1234 | skm->skm_magic = SKM_MAGIC; | |
1235 | skm->skm_avail = 0; | |
1236 | skm->skm_size = skc->skc_mag_size; | |
1237 | skm->skm_refill = skc->skc_mag_refill; | |
9b1b8e4c BB |
1238 | skm->skm_cache = skc; |
1239 | spl_init_delayed_work(&skm->skm_work, spl_magazine_age, skm); | |
ea3e6ca9 | 1240 | skm->skm_age = jiffies; |
4afaaefa | 1241 | } |
1242 | ||
b17edc10 | 1243 | SRETURN(skm); |
4afaaefa | 1244 | } |
1245 | ||
ea3e6ca9 BB |
1246 | /* |
1247 | * Free a per-cpu magazine assoicated with a specific core. | |
1248 | */ | |
4afaaefa | 1249 | static void |
1250 | spl_magazine_free(spl_kmem_magazine_t *skm) | |
1251 | { | |
a0f6da3d | 1252 | int size = sizeof(spl_kmem_magazine_t) + |
1253 | sizeof(void *) * skm->skm_size; | |
1254 | ||
b17edc10 | 1255 | SENTRY; |
4afaaefa | 1256 | ASSERT(skm->skm_magic == SKM_MAGIC); |
1257 | ASSERT(skm->skm_avail == 0); | |
a0f6da3d | 1258 | |
1259 | kmem_free(skm, size); | |
b17edc10 | 1260 | SEXIT; |
4afaaefa | 1261 | } |
1262 | ||
ea3e6ca9 BB |
1263 | /* |
1264 | * Create all pre-cpu magazines of reasonable sizes. | |
1265 | */ | |
4afaaefa | 1266 | static int |
1267 | spl_magazine_create(spl_kmem_cache_t *skc) | |
1268 | { | |
37db7d8c | 1269 | int i; |
b17edc10 | 1270 | SENTRY; |
4afaaefa | 1271 | |
1272 | skc->skc_mag_size = spl_magazine_size(skc); | |
ea3e6ca9 | 1273 | skc->skc_mag_refill = (skc->skc_mag_size + 1) / 2; |
4afaaefa | 1274 | |
37db7d8c BB |
1275 | for_each_online_cpu(i) { |
1276 | skc->skc_mag[i] = spl_magazine_alloc(skc, cpu_to_node(i)); | |
1277 | if (!skc->skc_mag[i]) { | |
1278 | for (i--; i >= 0; i--) | |
1279 | spl_magazine_free(skc->skc_mag[i]); | |
4afaaefa | 1280 | |
b17edc10 | 1281 | SRETURN(-ENOMEM); |
37db7d8c BB |
1282 | } |
1283 | } | |
4afaaefa | 1284 | |
9b1b8e4c BB |
1285 | /* Only after everything is allocated schedule magazine work */ |
1286 | for_each_online_cpu(i) | |
1287 | schedule_delayed_work_on(i, &skc->skc_mag[i]->skm_work, | |
1288 | skc->skc_delay / 3 * HZ); | |
1289 | ||
b17edc10 | 1290 | SRETURN(0); |
4afaaefa | 1291 | } |
1292 | ||
ea3e6ca9 BB |
1293 | /* |
1294 | * Destroy all pre-cpu magazines. | |
1295 | */ | |
4afaaefa | 1296 | static void |
1297 | spl_magazine_destroy(spl_kmem_cache_t *skc) | |
1298 | { | |
37db7d8c BB |
1299 | spl_kmem_magazine_t *skm; |
1300 | int i; | |
b17edc10 | 1301 | SENTRY; |
37db7d8c BB |
1302 | |
1303 | for_each_online_cpu(i) { | |
1304 | skm = skc->skc_mag[i]; | |
1305 | (void)spl_cache_flush(skc, skm, skm->skm_avail); | |
1306 | spl_magazine_free(skm); | |
1307 | } | |
1308 | ||
b17edc10 | 1309 | SEXIT; |
4afaaefa | 1310 | } |
1311 | ||
ea3e6ca9 BB |
1312 | /* |
1313 | * Create a object cache based on the following arguments: | |
1314 | * name cache name | |
1315 | * size cache object size | |
1316 | * align cache object alignment | |
1317 | * ctor cache object constructor | |
1318 | * dtor cache object destructor | |
1319 | * reclaim cache object reclaim | |
1320 | * priv cache private data for ctor/dtor/reclaim | |
1321 | * vmp unused must be NULL | |
1322 | * flags | |
1323 | * KMC_NOTOUCH Disable cache object aging (unsupported) | |
1324 | * KMC_NODEBUG Disable debugging (unsupported) | |
1325 | * KMC_NOMAGAZINE Disable magazine (unsupported) | |
1326 | * KMC_NOHASH Disable hashing (unsupported) | |
1327 | * KMC_QCACHE Disable qcache (unsupported) | |
1328 | * KMC_KMEM Force kmem backed cache | |
1329 | * KMC_VMEM Force vmem backed cache | |
1330 | * KMC_OFFSLAB Locate objects off the slab | |
1331 | */ | |
2fb9b26a | 1332 | spl_kmem_cache_t * |
1333 | spl_kmem_cache_create(char *name, size_t size, size_t align, | |
1334 | spl_kmem_ctor_t ctor, | |
1335 | spl_kmem_dtor_t dtor, | |
1336 | spl_kmem_reclaim_t reclaim, | |
1337 | void *priv, void *vmp, int flags) | |
1338 | { | |
1339 | spl_kmem_cache_t *skc; | |
a1502d76 | 1340 | int rc, kmem_flags = KM_SLEEP; |
b17edc10 | 1341 | SENTRY; |
937879f1 | 1342 | |
a1502d76 | 1343 | ASSERTF(!(flags & KMC_NOMAGAZINE), "Bad KMC_NOMAGAZINE (%x)\n", flags); |
1344 | ASSERTF(!(flags & KMC_NOHASH), "Bad KMC_NOHASH (%x)\n", flags); | |
1345 | ASSERTF(!(flags & KMC_QCACHE), "Bad KMC_QCACHE (%x)\n", flags); | |
48e0606a | 1346 | ASSERT(vmp == NULL); |
a1502d76 | 1347 | |
2fb9b26a | 1348 | /* We may be called when there is a non-zero preempt_count or |
1349 | * interrupts are disabled is which case we must not sleep. | |
1350 | */ | |
e9d7a2be | 1351 | if (current_thread_info()->preempt_count || irqs_disabled()) |
2fb9b26a | 1352 | kmem_flags = KM_NOSLEEP; |
0a6fd143 | 1353 | |
5198ea0e BB |
1354 | /* Allocate memry for a new cache an initialize it. Unfortunately, |
1355 | * this usually ends up being a large allocation of ~32k because | |
1356 | * we need to allocate enough memory for the worst case number of | |
1357 | * cpus in the magazine, skc_mag[NR_CPUS]. Because of this we | |
23d91792 | 1358 | * explicitly pass KM_NODEBUG to suppress the kmem warning */ |
5198ea0e | 1359 | skc = (spl_kmem_cache_t *)kmem_zalloc(sizeof(*skc), |
23d91792 | 1360 | kmem_flags | KM_NODEBUG); |
e9d7a2be | 1361 | if (skc == NULL) |
b17edc10 | 1362 | SRETURN(NULL); |
d61e12af | 1363 | |
2fb9b26a | 1364 | skc->skc_magic = SKC_MAGIC; |
2fb9b26a | 1365 | skc->skc_name_size = strlen(name) + 1; |
1366 | skc->skc_name = (char *)kmem_alloc(skc->skc_name_size, kmem_flags); | |
1367 | if (skc->skc_name == NULL) { | |
1368 | kmem_free(skc, sizeof(*skc)); | |
b17edc10 | 1369 | SRETURN(NULL); |
2fb9b26a | 1370 | } |
1371 | strncpy(skc->skc_name, name, skc->skc_name_size); | |
1372 | ||
e9d7a2be | 1373 | skc->skc_ctor = ctor; |
1374 | skc->skc_dtor = dtor; | |
1375 | skc->skc_reclaim = reclaim; | |
2fb9b26a | 1376 | skc->skc_private = priv; |
1377 | skc->skc_vmp = vmp; | |
1378 | skc->skc_flags = flags; | |
1379 | skc->skc_obj_size = size; | |
48e0606a | 1380 | skc->skc_obj_align = SPL_KMEM_CACHE_ALIGN; |
2fb9b26a | 1381 | skc->skc_delay = SPL_KMEM_CACHE_DELAY; |
37db7d8c | 1382 | skc->skc_reap = SPL_KMEM_CACHE_REAP; |
ea3e6ca9 | 1383 | atomic_set(&skc->skc_ref, 0); |
2fb9b26a | 1384 | |
2fb9b26a | 1385 | INIT_LIST_HEAD(&skc->skc_list); |
1386 | INIT_LIST_HEAD(&skc->skc_complete_list); | |
1387 | INIT_LIST_HEAD(&skc->skc_partial_list); | |
d46630e0 | 1388 | spin_lock_init(&skc->skc_lock); |
e9d7a2be | 1389 | skc->skc_slab_fail = 0; |
1390 | skc->skc_slab_create = 0; | |
1391 | skc->skc_slab_destroy = 0; | |
2fb9b26a | 1392 | skc->skc_slab_total = 0; |
1393 | skc->skc_slab_alloc = 0; | |
1394 | skc->skc_slab_max = 0; | |
1395 | skc->skc_obj_total = 0; | |
1396 | skc->skc_obj_alloc = 0; | |
1397 | skc->skc_obj_max = 0; | |
a1502d76 | 1398 | |
48e0606a | 1399 | if (align) { |
8b45dda2 BB |
1400 | VERIFY(ISP2(align)); |
1401 | VERIFY3U(align, >=, SPL_KMEM_CACHE_ALIGN); /* Min alignment */ | |
1402 | VERIFY3U(align, <=, PAGE_SIZE); /* Max alignment */ | |
48e0606a BB |
1403 | skc->skc_obj_align = align; |
1404 | } | |
1405 | ||
a1502d76 | 1406 | /* If none passed select a cache type based on object size */ |
1407 | if (!(skc->skc_flags & (KMC_KMEM | KMC_VMEM))) { | |
8b45dda2 | 1408 | if (spl_obj_size(skc) < (PAGE_SIZE / 8)) |
a1502d76 | 1409 | skc->skc_flags |= KMC_KMEM; |
8b45dda2 | 1410 | else |
a1502d76 | 1411 | skc->skc_flags |= KMC_VMEM; |
a1502d76 | 1412 | } |
1413 | ||
48e0606a BB |
1414 | rc = spl_slab_size(skc, &skc->skc_slab_objs, &skc->skc_slab_size); |
1415 | if (rc) | |
b17edc10 | 1416 | SGOTO(out, rc); |
4afaaefa | 1417 | |
1418 | rc = spl_magazine_create(skc); | |
48e0606a | 1419 | if (rc) |
b17edc10 | 1420 | SGOTO(out, rc); |
2fb9b26a | 1421 | |
ea3e6ca9 | 1422 | spl_init_delayed_work(&skc->skc_work, spl_cache_age, skc); |
37db7d8c | 1423 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
ea3e6ca9 | 1424 | |
2fb9b26a | 1425 | down_write(&spl_kmem_cache_sem); |
e9d7a2be | 1426 | list_add_tail(&skc->skc_list, &spl_kmem_cache_list); |
2fb9b26a | 1427 | up_write(&spl_kmem_cache_sem); |
1428 | ||
b17edc10 | 1429 | SRETURN(skc); |
48e0606a BB |
1430 | out: |
1431 | kmem_free(skc->skc_name, skc->skc_name_size); | |
1432 | kmem_free(skc, sizeof(*skc)); | |
b17edc10 | 1433 | SRETURN(NULL); |
f1ca4da6 | 1434 | } |
2fb9b26a | 1435 | EXPORT_SYMBOL(spl_kmem_cache_create); |
f1ca4da6 | 1436 | |
ea3e6ca9 BB |
1437 | /* |
1438 | * Destroy a cache and all objects assoicated with the cache. | |
1439 | */ | |
2fb9b26a | 1440 | void |
1441 | spl_kmem_cache_destroy(spl_kmem_cache_t *skc) | |
f1ca4da6 | 1442 | { |
ea3e6ca9 | 1443 | DECLARE_WAIT_QUEUE_HEAD(wq); |
9b1b8e4c | 1444 | int i; |
b17edc10 | 1445 | SENTRY; |
f1ca4da6 | 1446 | |
e9d7a2be | 1447 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1448 | ||
1449 | down_write(&spl_kmem_cache_sem); | |
1450 | list_del_init(&skc->skc_list); | |
1451 | up_write(&spl_kmem_cache_sem); | |
2fb9b26a | 1452 | |
ea3e6ca9 BB |
1453 | /* Cancel any and wait for any pending delayed work */ |
1454 | ASSERT(!test_and_set_bit(KMC_BIT_DESTROY, &skc->skc_flags)); | |
1455 | cancel_delayed_work(&skc->skc_work); | |
9b1b8e4c BB |
1456 | for_each_online_cpu(i) |
1457 | cancel_delayed_work(&skc->skc_mag[i]->skm_work); | |
1458 | ||
ea3e6ca9 BB |
1459 | flush_scheduled_work(); |
1460 | ||
1461 | /* Wait until all current callers complete, this is mainly | |
1462 | * to catch the case where a low memory situation triggers a | |
1463 | * cache reaping action which races with this destroy. */ | |
1464 | wait_event(wq, atomic_read(&skc->skc_ref) == 0); | |
1465 | ||
4afaaefa | 1466 | spl_magazine_destroy(skc); |
37db7d8c | 1467 | spl_slab_reclaim(skc, 0, 1); |
d46630e0 | 1468 | spin_lock(&skc->skc_lock); |
d6a26c6a | 1469 | |
2fb9b26a | 1470 | /* Validate there are no objects in use and free all the |
4afaaefa | 1471 | * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */ |
ea3e6ca9 BB |
1472 | ASSERT3U(skc->skc_slab_alloc, ==, 0); |
1473 | ASSERT3U(skc->skc_obj_alloc, ==, 0); | |
1474 | ASSERT3U(skc->skc_slab_total, ==, 0); | |
1475 | ASSERT3U(skc->skc_obj_total, ==, 0); | |
2fb9b26a | 1476 | ASSERT(list_empty(&skc->skc_complete_list)); |
a1502d76 | 1477 | |
2fb9b26a | 1478 | kmem_free(skc->skc_name, skc->skc_name_size); |
d46630e0 | 1479 | spin_unlock(&skc->skc_lock); |
ff449ac4 | 1480 | |
4afaaefa | 1481 | kmem_free(skc, sizeof(*skc)); |
2fb9b26a | 1482 | |
b17edc10 | 1483 | SEXIT; |
f1ca4da6 | 1484 | } |
2fb9b26a | 1485 | EXPORT_SYMBOL(spl_kmem_cache_destroy); |
f1ca4da6 | 1486 | |
ea3e6ca9 BB |
1487 | /* |
1488 | * Allocate an object from a slab attached to the cache. This is used to | |
1489 | * repopulate the per-cpu magazine caches in batches when they run low. | |
1490 | */ | |
4afaaefa | 1491 | static void * |
1492 | spl_cache_obj(spl_kmem_cache_t *skc, spl_kmem_slab_t *sks) | |
f1ca4da6 | 1493 | { |
2fb9b26a | 1494 | spl_kmem_obj_t *sko; |
f1ca4da6 | 1495 | |
e9d7a2be | 1496 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1497 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
4afaaefa | 1498 | ASSERT(spin_is_locked(&skc->skc_lock)); |
2fb9b26a | 1499 | |
a1502d76 | 1500 | sko = list_entry(sks->sks_free_list.next, spl_kmem_obj_t, sko_list); |
4afaaefa | 1501 | ASSERT(sko->sko_magic == SKO_MAGIC); |
1502 | ASSERT(sko->sko_addr != NULL); | |
2fb9b26a | 1503 | |
a1502d76 | 1504 | /* Remove from sks_free_list */ |
4afaaefa | 1505 | list_del_init(&sko->sko_list); |
2fb9b26a | 1506 | |
4afaaefa | 1507 | sks->sks_age = jiffies; |
1508 | sks->sks_ref++; | |
1509 | skc->skc_obj_alloc++; | |
2fb9b26a | 1510 | |
4afaaefa | 1511 | /* Track max obj usage statistics */ |
1512 | if (skc->skc_obj_alloc > skc->skc_obj_max) | |
1513 | skc->skc_obj_max = skc->skc_obj_alloc; | |
2fb9b26a | 1514 | |
4afaaefa | 1515 | /* Track max slab usage statistics */ |
1516 | if (sks->sks_ref == 1) { | |
1517 | skc->skc_slab_alloc++; | |
f1ca4da6 | 1518 | |
4afaaefa | 1519 | if (skc->skc_slab_alloc > skc->skc_slab_max) |
1520 | skc->skc_slab_max = skc->skc_slab_alloc; | |
2fb9b26a | 1521 | } |
1522 | ||
4afaaefa | 1523 | return sko->sko_addr; |
1524 | } | |
c30df9c8 | 1525 | |
ea3e6ca9 BB |
1526 | /* |
1527 | * No available objects on any slabsi, create a new slab. Since this | |
1528 | * is an expensive operation we do it without holding the spinlock and | |
1529 | * only briefly aquire it when we link in the fully allocated and | |
1530 | * constructed slab. | |
4afaaefa | 1531 | */ |
1532 | static spl_kmem_slab_t * | |
1533 | spl_cache_grow(spl_kmem_cache_t *skc, int flags) | |
1534 | { | |
e9d7a2be | 1535 | spl_kmem_slab_t *sks; |
b17edc10 | 1536 | SENTRY; |
f1ca4da6 | 1537 | |
e9d7a2be | 1538 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1539 | local_irq_enable(); |
1540 | might_sleep(); | |
e9d7a2be | 1541 | |
ea3e6ca9 BB |
1542 | /* |
1543 | * Before allocating a new slab check if the slab is being reaped. | |
1544 | * If it is there is a good chance we can wait until it finishes | |
1545 | * and then use one of the newly freed but not aged-out slabs. | |
1546 | */ | |
1547 | if (test_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
1548 | schedule(); | |
b17edc10 | 1549 | SGOTO(out, sks= NULL); |
4afaaefa | 1550 | } |
2fb9b26a | 1551 | |
ea3e6ca9 | 1552 | /* Allocate a new slab for the cache */ |
23d91792 | 1553 | sks = spl_slab_alloc(skc, flags | __GFP_NORETRY | KM_NODEBUG); |
ea3e6ca9 | 1554 | if (sks == NULL) |
b17edc10 | 1555 | SGOTO(out, sks = NULL); |
4afaaefa | 1556 | |
ea3e6ca9 | 1557 | /* Link the new empty slab in to the end of skc_partial_list. */ |
d46630e0 | 1558 | spin_lock(&skc->skc_lock); |
2fb9b26a | 1559 | skc->skc_slab_total++; |
1560 | skc->skc_obj_total += sks->sks_objs; | |
1561 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
d46630e0 | 1562 | spin_unlock(&skc->skc_lock); |
ea3e6ca9 BB |
1563 | out: |
1564 | local_irq_disable(); | |
4afaaefa | 1565 | |
b17edc10 | 1566 | SRETURN(sks); |
f1ca4da6 | 1567 | } |
1568 | ||
ea3e6ca9 BB |
1569 | /* |
1570 | * Refill a per-cpu magazine with objects from the slabs for this | |
1571 | * cache. Ideally the magazine can be repopulated using existing | |
1572 | * objects which have been released, however if we are unable to | |
1573 | * locate enough free objects new slabs of objects will be created. | |
1574 | */ | |
4afaaefa | 1575 | static int |
1576 | spl_cache_refill(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flags) | |
f1ca4da6 | 1577 | { |
e9d7a2be | 1578 | spl_kmem_slab_t *sks; |
1579 | int rc = 0, refill; | |
b17edc10 | 1580 | SENTRY; |
f1ca4da6 | 1581 | |
e9d7a2be | 1582 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1583 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1584 | ||
e9d7a2be | 1585 | refill = MIN(skm->skm_refill, skm->skm_size - skm->skm_avail); |
d46630e0 | 1586 | spin_lock(&skc->skc_lock); |
ff449ac4 | 1587 | |
4afaaefa | 1588 | while (refill > 0) { |
ea3e6ca9 | 1589 | /* No slabs available we may need to grow the cache */ |
4afaaefa | 1590 | if (list_empty(&skc->skc_partial_list)) { |
1591 | spin_unlock(&skc->skc_lock); | |
ff449ac4 | 1592 | |
4afaaefa | 1593 | sks = spl_cache_grow(skc, flags); |
1594 | if (!sks) | |
b17edc10 | 1595 | SGOTO(out, rc); |
4afaaefa | 1596 | |
1597 | /* Rescheduled to different CPU skm is not local */ | |
1598 | if (skm != skc->skc_mag[smp_processor_id()]) | |
b17edc10 | 1599 | SGOTO(out, rc); |
e9d7a2be | 1600 | |
1601 | /* Potentially rescheduled to the same CPU but | |
1602 | * allocations may have occured from this CPU while | |
1603 | * we were sleeping so recalculate max refill. */ | |
1604 | refill = MIN(refill, skm->skm_size - skm->skm_avail); | |
4afaaefa | 1605 | |
1606 | spin_lock(&skc->skc_lock); | |
1607 | continue; | |
1608 | } | |
d46630e0 | 1609 | |
4afaaefa | 1610 | /* Grab the next available slab */ |
1611 | sks = list_entry((&skc->skc_partial_list)->next, | |
1612 | spl_kmem_slab_t, sks_list); | |
1613 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
1614 | ASSERT(sks->sks_ref < sks->sks_objs); | |
1615 | ASSERT(!list_empty(&sks->sks_free_list)); | |
d46630e0 | 1616 | |
4afaaefa | 1617 | /* Consume as many objects as needed to refill the requested |
e9d7a2be | 1618 | * cache. We must also be careful not to overfill it. */ |
1619 | while (sks->sks_ref < sks->sks_objs && refill-- > 0 && ++rc) { | |
1620 | ASSERT(skm->skm_avail < skm->skm_size); | |
1621 | ASSERT(rc < skm->skm_size); | |
4afaaefa | 1622 | skm->skm_objs[skm->skm_avail++]=spl_cache_obj(skc,sks); |
e9d7a2be | 1623 | } |
f1ca4da6 | 1624 | |
4afaaefa | 1625 | /* Move slab to skc_complete_list when full */ |
1626 | if (sks->sks_ref == sks->sks_objs) { | |
1627 | list_del(&sks->sks_list); | |
1628 | list_add(&sks->sks_list, &skc->skc_complete_list); | |
2fb9b26a | 1629 | } |
1630 | } | |
57d86234 | 1631 | |
4afaaefa | 1632 | spin_unlock(&skc->skc_lock); |
1633 | out: | |
1634 | /* Returns the number of entries added to cache */ | |
b17edc10 | 1635 | SRETURN(rc); |
4afaaefa | 1636 | } |
1637 | ||
ea3e6ca9 BB |
1638 | /* |
1639 | * Release an object back to the slab from which it came. | |
1640 | */ | |
4afaaefa | 1641 | static void |
1642 | spl_cache_shrink(spl_kmem_cache_t *skc, void *obj) | |
1643 | { | |
e9d7a2be | 1644 | spl_kmem_slab_t *sks = NULL; |
4afaaefa | 1645 | spl_kmem_obj_t *sko = NULL; |
b17edc10 | 1646 | SENTRY; |
4afaaefa | 1647 | |
e9d7a2be | 1648 | ASSERT(skc->skc_magic == SKC_MAGIC); |
4afaaefa | 1649 | ASSERT(spin_is_locked(&skc->skc_lock)); |
1650 | ||
8b45dda2 | 1651 | sko = spl_sko_from_obj(skc, obj); |
a1502d76 | 1652 | ASSERT(sko->sko_magic == SKO_MAGIC); |
4afaaefa | 1653 | sks = sko->sko_slab; |
a1502d76 | 1654 | ASSERT(sks->sks_magic == SKS_MAGIC); |
2fb9b26a | 1655 | ASSERT(sks->sks_cache == skc); |
2fb9b26a | 1656 | list_add(&sko->sko_list, &sks->sks_free_list); |
d6a26c6a | 1657 | |
2fb9b26a | 1658 | sks->sks_age = jiffies; |
4afaaefa | 1659 | sks->sks_ref--; |
2fb9b26a | 1660 | skc->skc_obj_alloc--; |
f1ca4da6 | 1661 | |
2fb9b26a | 1662 | /* Move slab to skc_partial_list when no longer full. Slabs |
4afaaefa | 1663 | * are added to the head to keep the partial list is quasi-full |
1664 | * sorted order. Fuller at the head, emptier at the tail. */ | |
1665 | if (sks->sks_ref == (sks->sks_objs - 1)) { | |
2fb9b26a | 1666 | list_del(&sks->sks_list); |
1667 | list_add(&sks->sks_list, &skc->skc_partial_list); | |
1668 | } | |
f1ca4da6 | 1669 | |
2fb9b26a | 1670 | /* Move emply slabs to the end of the partial list so |
4afaaefa | 1671 | * they can be easily found and freed during reclamation. */ |
1672 | if (sks->sks_ref == 0) { | |
2fb9b26a | 1673 | list_del(&sks->sks_list); |
1674 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
1675 | skc->skc_slab_alloc--; | |
1676 | } | |
1677 | ||
b17edc10 | 1678 | SEXIT; |
4afaaefa | 1679 | } |
1680 | ||
ea3e6ca9 BB |
1681 | /* |
1682 | * Release a batch of objects from a per-cpu magazine back to their | |
1683 | * respective slabs. This occurs when we exceed the magazine size, | |
1684 | * are under memory pressure, when the cache is idle, or during | |
1685 | * cache cleanup. The flush argument contains the number of entries | |
1686 | * to remove from the magazine. | |
1687 | */ | |
4afaaefa | 1688 | static int |
1689 | spl_cache_flush(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flush) | |
1690 | { | |
1691 | int i, count = MIN(flush, skm->skm_avail); | |
b17edc10 | 1692 | SENTRY; |
4afaaefa | 1693 | |
e9d7a2be | 1694 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1695 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
4afaaefa | 1696 | |
ea3e6ca9 BB |
1697 | /* |
1698 | * XXX: Currently we simply return objects from the magazine to | |
1699 | * the slabs in fifo order. The ideal thing to do from a memory | |
1700 | * fragmentation standpoint is to cheaply determine the set of | |
1701 | * objects in the magazine which will result in the largest | |
1702 | * number of free slabs if released from the magazine. | |
1703 | */ | |
4afaaefa | 1704 | spin_lock(&skc->skc_lock); |
1705 | for (i = 0; i < count; i++) | |
1706 | spl_cache_shrink(skc, skm->skm_objs[i]); | |
1707 | ||
e9d7a2be | 1708 | skm->skm_avail -= count; |
1709 | memmove(skm->skm_objs, &(skm->skm_objs[count]), | |
4afaaefa | 1710 | sizeof(void *) * skm->skm_avail); |
1711 | ||
d46630e0 | 1712 | spin_unlock(&skc->skc_lock); |
4afaaefa | 1713 | |
b17edc10 | 1714 | SRETURN(count); |
4afaaefa | 1715 | } |
1716 | ||
ea3e6ca9 BB |
1717 | /* |
1718 | * Allocate an object from the per-cpu magazine, or if the magazine | |
1719 | * is empty directly allocate from a slab and repopulate the magazine. | |
1720 | */ | |
4afaaefa | 1721 | void * |
1722 | spl_kmem_cache_alloc(spl_kmem_cache_t *skc, int flags) | |
1723 | { | |
1724 | spl_kmem_magazine_t *skm; | |
1725 | unsigned long irq_flags; | |
1726 | void *obj = NULL; | |
b17edc10 | 1727 | SENTRY; |
4afaaefa | 1728 | |
e9d7a2be | 1729 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1730 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1731 | ASSERT(flags & KM_SLEEP); | |
1732 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1733 | local_irq_save(irq_flags); |
1734 | ||
1735 | restart: | |
1736 | /* Safe to update per-cpu structure without lock, but | |
1737 | * in the restart case we must be careful to reaquire | |
1738 | * the local magazine since this may have changed | |
1739 | * when we need to grow the cache. */ | |
1740 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1741 | ASSERTF(skm->skm_magic == SKM_MAGIC, "%x != %x: %s/%p/%p %x/%x/%x\n", |
1742 | skm->skm_magic, SKM_MAGIC, skc->skc_name, skc, skm, | |
1743 | skm->skm_size, skm->skm_refill, skm->skm_avail); | |
4afaaefa | 1744 | |
1745 | if (likely(skm->skm_avail)) { | |
1746 | /* Object available in CPU cache, use it */ | |
1747 | obj = skm->skm_objs[--skm->skm_avail]; | |
ea3e6ca9 | 1748 | skm->skm_age = jiffies; |
4afaaefa | 1749 | } else { |
1750 | /* Per-CPU cache empty, directly allocate from | |
1751 | * the slab and refill the per-CPU cache. */ | |
1752 | (void)spl_cache_refill(skc, skm, flags); | |
b17edc10 | 1753 | SGOTO(restart, obj = NULL); |
4afaaefa | 1754 | } |
1755 | ||
1756 | local_irq_restore(irq_flags); | |
fece7c99 | 1757 | ASSERT(obj); |
8b45dda2 | 1758 | ASSERT(IS_P2ALIGNED(obj, skc->skc_obj_align)); |
4afaaefa | 1759 | |
1760 | /* Pre-emptively migrate object to CPU L1 cache */ | |
1761 | prefetchw(obj); | |
ea3e6ca9 | 1762 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1763 | |
b17edc10 | 1764 | SRETURN(obj); |
4afaaefa | 1765 | } |
1766 | EXPORT_SYMBOL(spl_kmem_cache_alloc); | |
1767 | ||
ea3e6ca9 BB |
1768 | /* |
1769 | * Free an object back to the local per-cpu magazine, there is no | |
1770 | * guarantee that this is the same magazine the object was originally | |
1771 | * allocated from. We may need to flush entire from the magazine | |
1772 | * back to the slabs to make space. | |
1773 | */ | |
4afaaefa | 1774 | void |
1775 | spl_kmem_cache_free(spl_kmem_cache_t *skc, void *obj) | |
1776 | { | |
1777 | spl_kmem_magazine_t *skm; | |
1778 | unsigned long flags; | |
b17edc10 | 1779 | SENTRY; |
4afaaefa | 1780 | |
e9d7a2be | 1781 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1782 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1783 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1784 | local_irq_save(flags); |
1785 | ||
1786 | /* Safe to update per-cpu structure without lock, but | |
1787 | * no remote memory allocation tracking is being performed | |
1788 | * it is entirely possible to allocate an object from one | |
1789 | * CPU cache and return it to another. */ | |
1790 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1791 | ASSERT(skm->skm_magic == SKM_MAGIC); |
4afaaefa | 1792 | |
1793 | /* Per-CPU cache full, flush it to make space */ | |
1794 | if (unlikely(skm->skm_avail >= skm->skm_size)) | |
1795 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
1796 | ||
1797 | /* Available space in cache, use it */ | |
1798 | skm->skm_objs[skm->skm_avail++] = obj; | |
1799 | ||
1800 | local_irq_restore(flags); | |
ea3e6ca9 | 1801 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1802 | |
b17edc10 | 1803 | SEXIT; |
f1ca4da6 | 1804 | } |
2fb9b26a | 1805 | EXPORT_SYMBOL(spl_kmem_cache_free); |
5c2bb9b2 | 1806 | |
ea3e6ca9 BB |
1807 | /* |
1808 | * The generic shrinker function for all caches. Under linux a shrinker | |
1809 | * may not be tightly coupled with a slab cache. In fact linux always | |
1810 | * systematically trys calling all registered shrinker callbacks which | |
1811 | * report that they contain unused objects. Because of this we only | |
1812 | * register one shrinker function in the shim layer for all slab caches. | |
1813 | * We always attempt to shrink all caches when this generic shrinker | |
1814 | * is called. The shrinker should return the number of free objects | |
1815 | * in the cache when called with nr_to_scan == 0 but not attempt to | |
1816 | * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan | |
1817 | * objects should be freed, because Solaris semantics are to free | |
1818 | * all available objects we may free more objects than requested. | |
1819 | */ | |
2fb9b26a | 1820 | static int |
4afaaefa | 1821 | spl_kmem_cache_generic_shrinker(int nr_to_scan, unsigned int gfp_mask) |
2fb9b26a | 1822 | { |
e9d7a2be | 1823 | spl_kmem_cache_t *skc; |
ea3e6ca9 | 1824 | int unused = 0; |
5c2bb9b2 | 1825 | |
e9d7a2be | 1826 | down_read(&spl_kmem_cache_sem); |
ea3e6ca9 BB |
1827 | list_for_each_entry(skc, &spl_kmem_cache_list, skc_list) { |
1828 | if (nr_to_scan) | |
1829 | spl_kmem_cache_reap_now(skc); | |
1830 | ||
1831 | /* | |
1832 | * Presume everything alloc'ed in reclaimable, this ensures | |
1833 | * we are called again with nr_to_scan > 0 so can try and | |
1834 | * reclaim. The exact number is not important either so | |
1835 | * we forgo taking this already highly contented lock. | |
1836 | */ | |
1837 | unused += skc->skc_obj_alloc; | |
1838 | } | |
e9d7a2be | 1839 | up_read(&spl_kmem_cache_sem); |
2fb9b26a | 1840 | |
ea3e6ca9 | 1841 | return (unused * sysctl_vfs_cache_pressure) / 100; |
5c2bb9b2 | 1842 | } |
5c2bb9b2 | 1843 | |
ea3e6ca9 BB |
1844 | /* |
1845 | * Call the registered reclaim function for a cache. Depending on how | |
1846 | * many and which objects are released it may simply repopulate the | |
1847 | * local magazine which will then need to age-out. Objects which cannot | |
1848 | * fit in the magazine we will be released back to their slabs which will | |
1849 | * also need to age out before being release. This is all just best | |
1850 | * effort and we do not want to thrash creating and destroying slabs. | |
1851 | */ | |
57d86234 | 1852 | void |
2fb9b26a | 1853 | spl_kmem_cache_reap_now(spl_kmem_cache_t *skc) |
57d86234 | 1854 | { |
b17edc10 | 1855 | SENTRY; |
e9d7a2be | 1856 | |
1857 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
ea3e6ca9 | 1858 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
2fb9b26a | 1859 | |
ea3e6ca9 BB |
1860 | /* Prevent concurrent cache reaping when contended */ |
1861 | if (test_and_set_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
b17edc10 | 1862 | SEXIT; |
ea3e6ca9 BB |
1863 | return; |
1864 | } | |
2fb9b26a | 1865 | |
ea3e6ca9 | 1866 | atomic_inc(&skc->skc_ref); |
4afaaefa | 1867 | |
ea3e6ca9 BB |
1868 | if (skc->skc_reclaim) |
1869 | skc->skc_reclaim(skc->skc_private); | |
4afaaefa | 1870 | |
37db7d8c | 1871 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
ea3e6ca9 BB |
1872 | clear_bit(KMC_BIT_REAPING, &skc->skc_flags); |
1873 | atomic_dec(&skc->skc_ref); | |
4afaaefa | 1874 | |
b17edc10 | 1875 | SEXIT; |
57d86234 | 1876 | } |
2fb9b26a | 1877 | EXPORT_SYMBOL(spl_kmem_cache_reap_now); |
57d86234 | 1878 | |
ea3e6ca9 BB |
1879 | /* |
1880 | * Reap all free slabs from all registered caches. | |
1881 | */ | |
f1b59d26 | 1882 | void |
2fb9b26a | 1883 | spl_kmem_reap(void) |
937879f1 | 1884 | { |
4afaaefa | 1885 | spl_kmem_cache_generic_shrinker(KMC_REAP_CHUNK, GFP_KERNEL); |
f1ca4da6 | 1886 | } |
2fb9b26a | 1887 | EXPORT_SYMBOL(spl_kmem_reap); |
5d86345d | 1888 | |
ff449ac4 | 1889 | #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING) |
c6dc93d6 | 1890 | static char * |
4afaaefa | 1891 | spl_sprintf_addr(kmem_debug_t *kd, char *str, int len, int min) |
d6a26c6a | 1892 | { |
e9d7a2be | 1893 | int size = ((len - 1) < kd->kd_size) ? (len - 1) : kd->kd_size; |
d6a26c6a | 1894 | int i, flag = 1; |
1895 | ||
1896 | ASSERT(str != NULL && len >= 17); | |
e9d7a2be | 1897 | memset(str, 0, len); |
d6a26c6a | 1898 | |
1899 | /* Check for a fully printable string, and while we are at | |
1900 | * it place the printable characters in the passed buffer. */ | |
1901 | for (i = 0; i < size; i++) { | |
e9d7a2be | 1902 | str[i] = ((char *)(kd->kd_addr))[i]; |
1903 | if (isprint(str[i])) { | |
1904 | continue; | |
1905 | } else { | |
1906 | /* Minimum number of printable characters found | |
1907 | * to make it worthwhile to print this as ascii. */ | |
1908 | if (i > min) | |
1909 | break; | |
1910 | ||
1911 | flag = 0; | |
1912 | break; | |
1913 | } | |
d6a26c6a | 1914 | } |
1915 | ||
1916 | if (!flag) { | |
1917 | sprintf(str, "%02x%02x%02x%02x%02x%02x%02x%02x", | |
1918 | *((uint8_t *)kd->kd_addr), | |
1919 | *((uint8_t *)kd->kd_addr + 2), | |
1920 | *((uint8_t *)kd->kd_addr + 4), | |
1921 | *((uint8_t *)kd->kd_addr + 6), | |
1922 | *((uint8_t *)kd->kd_addr + 8), | |
1923 | *((uint8_t *)kd->kd_addr + 10), | |
1924 | *((uint8_t *)kd->kd_addr + 12), | |
1925 | *((uint8_t *)kd->kd_addr + 14)); | |
1926 | } | |
1927 | ||
1928 | return str; | |
1929 | } | |
1930 | ||
a1502d76 | 1931 | static int |
1932 | spl_kmem_init_tracking(struct list_head *list, spinlock_t *lock, int size) | |
1933 | { | |
1934 | int i; | |
b17edc10 | 1935 | SENTRY; |
a1502d76 | 1936 | |
1937 | spin_lock_init(lock); | |
1938 | INIT_LIST_HEAD(list); | |
1939 | ||
1940 | for (i = 0; i < size; i++) | |
1941 | INIT_HLIST_HEAD(&kmem_table[i]); | |
1942 | ||
b17edc10 | 1943 | SRETURN(0); |
a1502d76 | 1944 | } |
1945 | ||
ff449ac4 | 1946 | static void |
1947 | spl_kmem_fini_tracking(struct list_head *list, spinlock_t *lock) | |
5d86345d | 1948 | { |
2fb9b26a | 1949 | unsigned long flags; |
1950 | kmem_debug_t *kd; | |
1951 | char str[17]; | |
b17edc10 | 1952 | SENTRY; |
2fb9b26a | 1953 | |
ff449ac4 | 1954 | spin_lock_irqsave(lock, flags); |
1955 | if (!list_empty(list)) | |
a0f6da3d | 1956 | printk(KERN_WARNING "%-16s %-5s %-16s %s:%s\n", "address", |
1957 | "size", "data", "func", "line"); | |
2fb9b26a | 1958 | |
ff449ac4 | 1959 | list_for_each_entry(kd, list, kd_list) |
a0f6da3d | 1960 | printk(KERN_WARNING "%p %-5d %-16s %s:%d\n", kd->kd_addr, |
b6b2acc6 | 1961 | (int)kd->kd_size, spl_sprintf_addr(kd, str, 17, 8), |
2fb9b26a | 1962 | kd->kd_func, kd->kd_line); |
1963 | ||
ff449ac4 | 1964 | spin_unlock_irqrestore(lock, flags); |
b17edc10 | 1965 | SEXIT; |
ff449ac4 | 1966 | } |
1967 | #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
a1502d76 | 1968 | #define spl_kmem_init_tracking(list, lock, size) |
ff449ac4 | 1969 | #define spl_kmem_fini_tracking(list, lock) |
1970 | #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
1971 | ||
36b313da BB |
1972 | static void |
1973 | spl_kmem_init_globals(void) | |
1974 | { | |
1975 | struct zone *zone; | |
1976 | ||
1977 | /* For now all zones are includes, it may be wise to restrict | |
1978 | * this to normal and highmem zones if we see problems. */ | |
1979 | for_each_zone(zone) { | |
1980 | ||
1981 | if (!populated_zone(zone)) | |
1982 | continue; | |
1983 | ||
baf2979e BB |
1984 | minfree += min_wmark_pages(zone); |
1985 | desfree += low_wmark_pages(zone); | |
1986 | lotsfree += high_wmark_pages(zone); | |
36b313da | 1987 | } |
4ab13d3b BB |
1988 | |
1989 | /* Solaris default values */ | |
96dded38 BB |
1990 | swapfs_minfree = MAX(2*1024*1024 >> PAGE_SHIFT, physmem >> 3); |
1991 | swapfs_reserve = MIN(4*1024*1024 >> PAGE_SHIFT, physmem >> 4); | |
36b313da BB |
1992 | } |
1993 | ||
d1ff2312 BB |
1994 | /* |
1995 | * Called at module init when it is safe to use spl_kallsyms_lookup_name() | |
1996 | */ | |
1997 | int | |
1998 | spl_kmem_init_kallsyms_lookup(void) | |
1999 | { | |
2000 | #ifndef HAVE_GET_VMALLOC_INFO | |
2001 | get_vmalloc_info_fn = (get_vmalloc_info_t) | |
2002 | spl_kallsyms_lookup_name("get_vmalloc_info"); | |
e11d6c5f BB |
2003 | if (!get_vmalloc_info_fn) { |
2004 | printk(KERN_ERR "Error: Unknown symbol get_vmalloc_info\n"); | |
d1ff2312 | 2005 | return -EFAULT; |
e11d6c5f | 2006 | } |
d1ff2312 BB |
2007 | #endif /* HAVE_GET_VMALLOC_INFO */ |
2008 | ||
5232d256 BB |
2009 | #ifdef HAVE_PGDAT_HELPERS |
2010 | # ifndef HAVE_FIRST_ONLINE_PGDAT | |
d1ff2312 BB |
2011 | first_online_pgdat_fn = (first_online_pgdat_t) |
2012 | spl_kallsyms_lookup_name("first_online_pgdat"); | |
e11d6c5f BB |
2013 | if (!first_online_pgdat_fn) { |
2014 | printk(KERN_ERR "Error: Unknown symbol first_online_pgdat\n"); | |
d1ff2312 | 2015 | return -EFAULT; |
e11d6c5f | 2016 | } |
5232d256 | 2017 | # endif /* HAVE_FIRST_ONLINE_PGDAT */ |
d1ff2312 | 2018 | |
5232d256 | 2019 | # ifndef HAVE_NEXT_ONLINE_PGDAT |
d1ff2312 BB |
2020 | next_online_pgdat_fn = (next_online_pgdat_t) |
2021 | spl_kallsyms_lookup_name("next_online_pgdat"); | |
e11d6c5f BB |
2022 | if (!next_online_pgdat_fn) { |
2023 | printk(KERN_ERR "Error: Unknown symbol next_online_pgdat\n"); | |
d1ff2312 | 2024 | return -EFAULT; |
e11d6c5f | 2025 | } |
5232d256 | 2026 | # endif /* HAVE_NEXT_ONLINE_PGDAT */ |
d1ff2312 | 2027 | |
5232d256 | 2028 | # ifndef HAVE_NEXT_ZONE |
d1ff2312 BB |
2029 | next_zone_fn = (next_zone_t) |
2030 | spl_kallsyms_lookup_name("next_zone"); | |
e11d6c5f BB |
2031 | if (!next_zone_fn) { |
2032 | printk(KERN_ERR "Error: Unknown symbol next_zone\n"); | |
d1ff2312 | 2033 | return -EFAULT; |
e11d6c5f | 2034 | } |
5232d256 BB |
2035 | # endif /* HAVE_NEXT_ZONE */ |
2036 | ||
2037 | #else /* HAVE_PGDAT_HELPERS */ | |
2038 | ||
2039 | # ifndef HAVE_PGDAT_LIST | |
124ca8a5 | 2040 | pgdat_list_addr = *(struct pglist_data **) |
5232d256 BB |
2041 | spl_kallsyms_lookup_name("pgdat_list"); |
2042 | if (!pgdat_list_addr) { | |
2043 | printk(KERN_ERR "Error: Unknown symbol pgdat_list\n"); | |
2044 | return -EFAULT; | |
2045 | } | |
2046 | # endif /* HAVE_PGDAT_LIST */ | |
2047 | #endif /* HAVE_PGDAT_HELPERS */ | |
d1ff2312 | 2048 | |
6ae7fef5 | 2049 | #if defined(NEED_GET_ZONE_COUNTS) && !defined(HAVE_GET_ZONE_COUNTS) |
d1ff2312 BB |
2050 | get_zone_counts_fn = (get_zone_counts_t) |
2051 | spl_kallsyms_lookup_name("get_zone_counts"); | |
e11d6c5f BB |
2052 | if (!get_zone_counts_fn) { |
2053 | printk(KERN_ERR "Error: Unknown symbol get_zone_counts\n"); | |
d1ff2312 | 2054 | return -EFAULT; |
e11d6c5f | 2055 | } |
6ae7fef5 | 2056 | #endif /* NEED_GET_ZONE_COUNTS && !HAVE_GET_ZONE_COUNTS */ |
d1ff2312 BB |
2057 | |
2058 | /* | |
2059 | * It is now safe to initialize the global tunings which rely on | |
2060 | * the use of the for_each_zone() macro. This macro in turns | |
2061 | * depends on the *_pgdat symbols which are now available. | |
2062 | */ | |
2063 | spl_kmem_init_globals(); | |
2064 | ||
2065 | return 0; | |
2066 | } | |
2067 | ||
a1502d76 | 2068 | int |
2069 | spl_kmem_init(void) | |
2070 | { | |
2071 | int rc = 0; | |
b17edc10 | 2072 | SENTRY; |
a1502d76 | 2073 | |
2074 | init_rwsem(&spl_kmem_cache_sem); | |
2075 | INIT_LIST_HEAD(&spl_kmem_cache_list); | |
2076 | ||
2077 | #ifdef HAVE_SET_SHRINKER | |
2078 | spl_kmem_cache_shrinker = set_shrinker(KMC_DEFAULT_SEEKS, | |
2079 | spl_kmem_cache_generic_shrinker); | |
2080 | if (spl_kmem_cache_shrinker == NULL) | |
b17edc10 | 2081 | SRETURN(rc = -ENOMEM); |
a1502d76 | 2082 | #else |
2083 | register_shrinker(&spl_kmem_cache_shrinker); | |
2084 | #endif | |
2085 | ||
2086 | #ifdef DEBUG_KMEM | |
d04c8a56 BB |
2087 | kmem_alloc_used_set(0); |
2088 | vmem_alloc_used_set(0); | |
a1502d76 | 2089 | |
2090 | spl_kmem_init_tracking(&kmem_list, &kmem_lock, KMEM_TABLE_SIZE); | |
2091 | spl_kmem_init_tracking(&vmem_list, &vmem_lock, VMEM_TABLE_SIZE); | |
2092 | #endif | |
b17edc10 | 2093 | SRETURN(rc); |
a1502d76 | 2094 | } |
2095 | ||
ff449ac4 | 2096 | void |
2097 | spl_kmem_fini(void) | |
2098 | { | |
2099 | #ifdef DEBUG_KMEM | |
2100 | /* Display all unreclaimed memory addresses, including the | |
2101 | * allocation size and the first few bytes of what's located | |
2102 | * at that address to aid in debugging. Performance is not | |
2103 | * a serious concern here since it is module unload time. */ | |
d04c8a56 | 2104 | if (kmem_alloc_used_read() != 0) |
b17edc10 | 2105 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
3cb77549 BB |
2106 | "kmem leaked %ld/%ld bytes\n", |
2107 | kmem_alloc_used_read(), kmem_alloc_max); | |
ff449ac4 | 2108 | |
2fb9b26a | 2109 | |
d04c8a56 | 2110 | if (vmem_alloc_used_read() != 0) |
b17edc10 | 2111 | SDEBUG_LIMIT(SD_CONSOLE | SD_WARNING, |
3cb77549 BB |
2112 | "vmem leaked %ld/%ld bytes\n", |
2113 | vmem_alloc_used_read(), vmem_alloc_max); | |
2fb9b26a | 2114 | |
ff449ac4 | 2115 | spl_kmem_fini_tracking(&kmem_list, &kmem_lock); |
2116 | spl_kmem_fini_tracking(&vmem_list, &vmem_lock); | |
2117 | #endif /* DEBUG_KMEM */ | |
b17edc10 | 2118 | SENTRY; |
2fb9b26a | 2119 | |
2120 | #ifdef HAVE_SET_SHRINKER | |
2121 | remove_shrinker(spl_kmem_cache_shrinker); | |
2122 | #else | |
2123 | unregister_shrinker(&spl_kmem_cache_shrinker); | |
5d86345d | 2124 | #endif |
2fb9b26a | 2125 | |
b17edc10 | 2126 | SEXIT; |
5d86345d | 2127 | } |