]>
Commit | Line | Data |
---|---|---|
715f6251 | 1 | /* |
2 | * This file is part of the SPL: Solaris Porting Layer. | |
3 | * | |
4 | * Copyright (c) 2008 Lawrence Livermore National Security, LLC. | |
5 | * Produced at Lawrence Livermore National Laboratory | |
6 | * Written by: | |
7 | * Brian Behlendorf <behlendorf1@llnl.gov>, | |
8 | * Herb Wartens <wartens2@llnl.gov>, | |
9 | * Jim Garlick <garlick@llnl.gov> | |
10 | * UCRL-CODE-235197 | |
11 | * | |
12 | * This is free software; you can redistribute it and/or modify it | |
13 | * under the terms of the GNU General Public License as published by | |
14 | * the Free Software Foundation; either version 2 of the License, or | |
15 | * (at your option) any later version. | |
16 | * | |
17 | * This is distributed in the hope that it will be useful, but WITHOUT | |
18 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
19 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
20 | * for more details. | |
21 | * | |
22 | * You should have received a copy of the GNU General Public License along | |
23 | * with this program; if not, write to the Free Software Foundation, Inc., | |
24 | * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. | |
25 | */ | |
26 | ||
f4b37741 | 27 | #include <sys/kmem.h> |
f1ca4da6 | 28 | |
937879f1 | 29 | #ifdef DEBUG_SUBSYSTEM |
a0f6da3d | 30 | # undef DEBUG_SUBSYSTEM |
937879f1 | 31 | #endif |
32 | ||
33 | #define DEBUG_SUBSYSTEM S_KMEM | |
34 | ||
36b313da BB |
35 | /* |
36 | * The minimum amount of memory measured in pages to be free at all | |
37 | * times on the system. This is similar to Linux's zone->pages_min | |
38 | * multipled by the number of zones and is sized based on that. | |
39 | */ | |
40 | pgcnt_t minfree = 0; | |
41 | EXPORT_SYMBOL(minfree); | |
42 | ||
43 | /* | |
44 | * The desired amount of memory measured in pages to be free at all | |
45 | * times on the system. This is similar to Linux's zone->pages_low | |
46 | * multipled by the number of zones and is sized based on that. | |
47 | * Assuming all zones are being used roughly equally, when we drop | |
48 | * below this threshold async page reclamation is triggered. | |
49 | */ | |
50 | pgcnt_t desfree = 0; | |
51 | EXPORT_SYMBOL(desfree); | |
52 | ||
53 | /* | |
54 | * When above this amount of memory measures in pages the system is | |
55 | * determined to have enough free memory. This is similar to Linux's | |
56 | * zone->pages_high multipled by the number of zones and is sized based | |
57 | * on that. Assuming all zones are being used roughly equally, when | |
58 | * async page reclamation reaches this threshold it stops. | |
59 | */ | |
60 | pgcnt_t lotsfree = 0; | |
61 | EXPORT_SYMBOL(lotsfree); | |
62 | ||
63 | /* Unused always 0 in this implementation */ | |
64 | pgcnt_t needfree = 0; | |
65 | EXPORT_SYMBOL(needfree); | |
66 | ||
36b313da BB |
67 | pgcnt_t swapfs_minfree = 0; |
68 | EXPORT_SYMBOL(swapfs_minfree); | |
69 | ||
70 | pgcnt_t swapfs_reserve = 0; | |
71 | EXPORT_SYMBOL(swapfs_reserve); | |
72 | ||
36b313da BB |
73 | vmem_t *heap_arena = NULL; |
74 | EXPORT_SYMBOL(heap_arena); | |
75 | ||
76 | vmem_t *zio_alloc_arena = NULL; | |
77 | EXPORT_SYMBOL(zio_alloc_arena); | |
78 | ||
79 | vmem_t *zio_arena = NULL; | |
80 | EXPORT_SYMBOL(zio_arena); | |
81 | ||
82 | #ifndef HAVE_FIRST_ONLINE_PGDAT | |
4ab13d3b BB |
83 | struct pglist_data * |
84 | first_online_pgdat(void) | |
36b313da BB |
85 | { |
86 | return NODE_DATA(first_online_node); | |
87 | } | |
4ab13d3b | 88 | EXPORT_SYMBOL(first_online_pgdat); |
36b313da BB |
89 | #endif /* HAVE_FIRST_ONLINE_PGDAT */ |
90 | ||
91 | #ifndef HAVE_NEXT_ONLINE_PGDAT | |
4ab13d3b BB |
92 | struct pglist_data * |
93 | next_online_pgdat(struct pglist_data *pgdat) | |
36b313da BB |
94 | { |
95 | int nid = next_online_node(pgdat->node_id); | |
96 | ||
97 | if (nid == MAX_NUMNODES) | |
98 | return NULL; | |
99 | ||
100 | return NODE_DATA(nid); | |
101 | } | |
4ab13d3b | 102 | EXPORT_SYMBOL(next_online_pgdat); |
36b313da BB |
103 | #endif /* HAVE_NEXT_ONLINE_PGDAT */ |
104 | ||
105 | #ifndef HAVE_NEXT_ZONE | |
4ab13d3b BB |
106 | struct zone * |
107 | next_zone(struct zone *zone) | |
36b313da BB |
108 | { |
109 | pg_data_t *pgdat = zone->zone_pgdat; | |
110 | ||
111 | if (zone < pgdat->node_zones + MAX_NR_ZONES - 1) | |
112 | zone++; | |
113 | else { | |
114 | pgdat = next_online_pgdat(pgdat); | |
115 | if (pgdat) | |
116 | zone = pgdat->node_zones; | |
117 | else | |
118 | zone = NULL; | |
119 | } | |
120 | return zone; | |
121 | } | |
4ab13d3b | 122 | EXPORT_SYMBOL(next_zone); |
36b313da BB |
123 | #endif /* HAVE_NEXT_ZONE */ |
124 | ||
4ab13d3b BB |
125 | #ifndef HAVE_GET_ZONE_COUNTS |
126 | void | |
127 | __get_zone_counts(unsigned long *active, unsigned long *inactive, | |
128 | unsigned long *free, struct pglist_data *pgdat) | |
129 | { | |
130 | struct zone *zones = pgdat->node_zones; | |
131 | int i; | |
132 | ||
133 | *active = 0; | |
134 | *inactive = 0; | |
135 | *free = 0; | |
136 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
137 | *active += zones[i].nr_active; | |
138 | *inactive += zones[i].nr_inactive; | |
139 | *free += zones[i].free_pages; | |
140 | } | |
141 | } | |
142 | ||
143 | void | |
144 | get_zone_counts(unsigned long *active, unsigned long *inactive, | |
145 | unsigned long *free) | |
146 | { | |
147 | struct pglist_data *pgdat; | |
148 | ||
149 | *active = 0; | |
150 | *inactive = 0; | |
151 | *free = 0; | |
152 | for_each_online_pgdat(pgdat) { | |
153 | unsigned long l, m, n; | |
154 | __get_zone_counts(&l, &m, &n, pgdat); | |
155 | *active += l; | |
156 | *inactive += m; | |
157 | *free += n; | |
158 | } | |
159 | } | |
160 | EXPORT_SYMBOL(get_zone_counts); | |
161 | #endif /* HAVE_GET_ZONE_COUNTS */ | |
162 | ||
163 | pgcnt_t | |
164 | spl_kmem_availrmem(void) | |
165 | { | |
166 | unsigned long active; | |
167 | unsigned long inactive; | |
168 | unsigned long free; | |
169 | ||
170 | get_zone_counts(&active, &inactive, &free); | |
171 | ||
172 | /* The amount of easily available memory */ | |
173 | return free + inactive; | |
174 | } | |
175 | EXPORT_SYMBOL(spl_kmem_availrmem); | |
176 | ||
177 | size_t | |
178 | vmem_size(vmem_t *vmp, int typemask) | |
179 | { | |
180 | /* Arena's unsupported */ | |
181 | ASSERT(vmp == NULL); | |
182 | ASSERT(typemask & (VMEM_ALLOC | VMEM_FREE)); | |
183 | ||
184 | return 0; | |
185 | } | |
186 | EXPORT_SYMBOL(vmem_size); | |
187 | ||
188 | ||
f1ca4da6 | 189 | /* |
2fb9b26a | 190 | * Memory allocation interfaces and debugging for basic kmem_* |
191 | * and vmem_* style memory allocation. When DEBUG_KMEM is enable | |
192 | * all allocations will be tracked when they are allocated and | |
193 | * freed. When the SPL module is unload a list of all leaked | |
194 | * addresses and where they were allocated will be dumped to the | |
195 | * console. Enabling this feature has a significant impant on | |
196 | * performance but it makes finding memory leaks staight forward. | |
f1ca4da6 | 197 | */ |
198 | #ifdef DEBUG_KMEM | |
199 | /* Shim layer memory accounting */ | |
550f1705 | 200 | atomic64_t kmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 201 | unsigned long long kmem_alloc_max = 0; |
550f1705 | 202 | atomic64_t vmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 203 | unsigned long long vmem_alloc_max = 0; |
c19c06f3 | 204 | int kmem_warning_flag = 1; |
79b31f36 | 205 | |
ff449ac4 | 206 | EXPORT_SYMBOL(kmem_alloc_used); |
207 | EXPORT_SYMBOL(kmem_alloc_max); | |
208 | EXPORT_SYMBOL(vmem_alloc_used); | |
209 | EXPORT_SYMBOL(vmem_alloc_max); | |
210 | EXPORT_SYMBOL(kmem_warning_flag); | |
211 | ||
a0f6da3d | 212 | # ifdef DEBUG_KMEM_TRACKING |
213 | ||
214 | /* XXX - Not to surprisingly with debugging enabled the xmem_locks are very | |
215 | * highly contended particularly on xfree(). If we want to run with this | |
216 | * detailed debugging enabled for anything other than debugging we need to | |
217 | * minimize the contention by moving to a lock per xmem_table entry model. | |
218 | */ | |
219 | ||
220 | # define KMEM_HASH_BITS 10 | |
221 | # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS) | |
222 | ||
223 | # define VMEM_HASH_BITS 10 | |
224 | # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS) | |
225 | ||
226 | typedef struct kmem_debug { | |
227 | struct hlist_node kd_hlist; /* Hash node linkage */ | |
228 | struct list_head kd_list; /* List of all allocations */ | |
229 | void *kd_addr; /* Allocation pointer */ | |
230 | size_t kd_size; /* Allocation size */ | |
231 | const char *kd_func; /* Allocation function */ | |
232 | int kd_line; /* Allocation line */ | |
233 | } kmem_debug_t; | |
234 | ||
d6a26c6a | 235 | spinlock_t kmem_lock; |
236 | struct hlist_head kmem_table[KMEM_TABLE_SIZE]; | |
237 | struct list_head kmem_list; | |
238 | ||
13cdca65 | 239 | spinlock_t vmem_lock; |
240 | struct hlist_head vmem_table[VMEM_TABLE_SIZE]; | |
241 | struct list_head vmem_list; | |
242 | ||
d6a26c6a | 243 | EXPORT_SYMBOL(kmem_lock); |
244 | EXPORT_SYMBOL(kmem_table); | |
245 | EXPORT_SYMBOL(kmem_list); | |
246 | ||
13cdca65 | 247 | EXPORT_SYMBOL(vmem_lock); |
248 | EXPORT_SYMBOL(vmem_table); | |
249 | EXPORT_SYMBOL(vmem_list); | |
a0f6da3d | 250 | # endif |
13cdca65 | 251 | |
c19c06f3 | 252 | int kmem_set_warning(int flag) { return (kmem_warning_flag = !!flag); } |
253 | #else | |
254 | int kmem_set_warning(int flag) { return 0; } | |
f1ca4da6 | 255 | #endif |
c19c06f3 | 256 | EXPORT_SYMBOL(kmem_set_warning); |
f1ca4da6 | 257 | |
258 | /* | |
259 | * Slab allocation interfaces | |
260 | * | |
2fb9b26a | 261 | * While the Linux slab implementation was inspired by the Solaris |
262 | * implemenation I cannot use it to emulate the Solaris APIs. I | |
263 | * require two features which are not provided by the Linux slab. | |
264 | * | |
265 | * 1) Constructors AND destructors. Recent versions of the Linux | |
266 | * kernel have removed support for destructors. This is a deal | |
267 | * breaker for the SPL which contains particularly expensive | |
268 | * initializers for mutex's, condition variables, etc. We also | |
a0f6da3d | 269 | * require a minimal level of cleanup for these data types unlike |
270 | * many Linux data type which do need to be explicitly destroyed. | |
2fb9b26a | 271 | * |
a0f6da3d | 272 | * 2) Virtual address space backed slab. Callers of the Solaris slab |
2fb9b26a | 273 | * expect it to work well for both small are very large allocations. |
274 | * Because of memory fragmentation the Linux slab which is backed | |
275 | * by kmalloc'ed memory performs very badly when confronted with | |
276 | * large numbers of large allocations. Basing the slab on the | |
277 | * virtual address space removes the need for contigeous pages | |
278 | * and greatly improve performance for large allocations. | |
279 | * | |
280 | * For these reasons, the SPL has its own slab implementation with | |
281 | * the needed features. It is not as highly optimized as either the | |
282 | * Solaris or Linux slabs, but it should get me most of what is | |
283 | * needed until it can be optimized or obsoleted by another approach. | |
284 | * | |
285 | * One serious concern I do have about this method is the relatively | |
286 | * small virtual address space on 32bit arches. This will seriously | |
287 | * constrain the size of the slab caches and their performance. | |
288 | * | |
2fb9b26a | 289 | * XXX: Improve the partial slab list by carefully maintaining a |
290 | * strict ordering of fullest to emptiest slabs based on | |
291 | * the slab reference count. This gaurentees the when freeing | |
292 | * slabs back to the system we need only linearly traverse the | |
293 | * last N slabs in the list to discover all the freeable slabs. | |
294 | * | |
295 | * XXX: NUMA awareness for optionally allocating memory close to a | |
296 | * particular core. This can be adventageous if you know the slab | |
297 | * object will be short lived and primarily accessed from one core. | |
298 | * | |
299 | * XXX: Slab coloring may also yield performance improvements and would | |
300 | * be desirable to implement. | |
f1ca4da6 | 301 | */ |
2fb9b26a | 302 | |
a0f6da3d | 303 | struct list_head spl_kmem_cache_list; /* List of caches */ |
304 | struct rw_semaphore spl_kmem_cache_sem; /* Cache list lock */ | |
c30df9c8 | 305 | |
4afaaefa | 306 | static int spl_cache_flush(spl_kmem_cache_t *skc, |
a0f6da3d | 307 | spl_kmem_magazine_t *skm, int flush); |
4afaaefa | 308 | |
57d86234 | 309 | #ifdef HAVE_SET_SHRINKER |
2fb9b26a | 310 | static struct shrinker *spl_kmem_cache_shrinker; |
57d86234 | 311 | #else |
4afaaefa | 312 | static int spl_kmem_cache_generic_shrinker(int nr_to_scan, |
a0f6da3d | 313 | unsigned int gfp_mask); |
2fb9b26a | 314 | static struct shrinker spl_kmem_cache_shrinker = { |
4afaaefa | 315 | .shrink = spl_kmem_cache_generic_shrinker, |
57d86234 | 316 | .seeks = KMC_DEFAULT_SEEKS, |
317 | }; | |
318 | #endif | |
f1ca4da6 | 319 | |
a0f6da3d | 320 | #ifdef DEBUG_KMEM |
321 | # ifdef DEBUG_KMEM_TRACKING | |
322 | ||
323 | static kmem_debug_t * | |
324 | kmem_del_init(spinlock_t *lock, struct hlist_head *table, int bits, | |
325 | void *addr) | |
326 | { | |
327 | struct hlist_head *head; | |
328 | struct hlist_node *node; | |
329 | struct kmem_debug *p; | |
330 | unsigned long flags; | |
331 | ENTRY; | |
332 | ||
333 | spin_lock_irqsave(lock, flags); | |
334 | ||
335 | head = &table[hash_ptr(addr, bits)]; | |
336 | hlist_for_each_entry_rcu(p, node, head, kd_hlist) { | |
337 | if (p->kd_addr == addr) { | |
338 | hlist_del_init(&p->kd_hlist); | |
339 | list_del_init(&p->kd_list); | |
340 | spin_unlock_irqrestore(lock, flags); | |
341 | return p; | |
342 | } | |
343 | } | |
344 | ||
345 | spin_unlock_irqrestore(lock, flags); | |
346 | ||
347 | RETURN(NULL); | |
348 | } | |
349 | ||
350 | void * | |
351 | kmem_alloc_track(size_t size, int flags, const char *func, int line, | |
352 | int node_alloc, int node) | |
353 | { | |
354 | void *ptr = NULL; | |
355 | kmem_debug_t *dptr; | |
356 | unsigned long irq_flags; | |
357 | ENTRY; | |
358 | ||
359 | dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t), | |
360 | flags & ~__GFP_ZERO); | |
361 | ||
362 | if (dptr == NULL) { | |
363 | CWARN("kmem_alloc(%ld, 0x%x) debug failed\n", | |
364 | sizeof(kmem_debug_t), flags); | |
365 | } else { | |
366 | /* Marked unlikely because we should never be doing this, | |
367 | * we tolerate to up 2 pages but a single page is best. */ | |
368 | if (unlikely((size) > (PAGE_SIZE * 2)) && kmem_warning_flag) | |
369 | CWARN("Large kmem_alloc(%llu, 0x%x) (%lld/%llu)\n", | |
370 | (unsigned long long) size, flags, | |
371 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
372 | ||
c8e60837 | 373 | /* We use kstrdup() below because the string pointed to by |
374 | * __FUNCTION__ might not be available by the time we want | |
375 | * to print it since the module might have been unloaded. */ | |
376 | dptr->kd_func = kstrdup(func, flags & ~__GFP_ZERO); | |
377 | if (unlikely(dptr->kd_func == NULL)) { | |
378 | kfree(dptr); | |
379 | CWARN("kstrdup() failed in kmem_alloc(%llu, 0x%x) " | |
380 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
381 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
382 | goto out; | |
383 | } | |
384 | ||
a0f6da3d | 385 | /* Use the correct allocator */ |
386 | if (node_alloc) { | |
387 | ASSERT(!(flags & __GFP_ZERO)); | |
388 | ptr = kmalloc_node(size, flags, node); | |
389 | } else if (flags & __GFP_ZERO) { | |
390 | ptr = kzalloc(size, flags & ~__GFP_ZERO); | |
391 | } else { | |
392 | ptr = kmalloc(size, flags); | |
393 | } | |
394 | ||
395 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 396 | kfree(dptr->kd_func); |
a0f6da3d | 397 | kfree(dptr); |
398 | CWARN("kmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
399 | (unsigned long long) size, flags, | |
400 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
401 | goto out; | |
402 | } | |
403 | ||
404 | atomic64_add(size, &kmem_alloc_used); | |
405 | if (unlikely(atomic64_read(&kmem_alloc_used) > | |
406 | kmem_alloc_max)) | |
407 | kmem_alloc_max = | |
408 | atomic64_read(&kmem_alloc_used); | |
409 | ||
410 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
411 | INIT_LIST_HEAD(&dptr->kd_list); | |
412 | ||
413 | dptr->kd_addr = ptr; | |
414 | dptr->kd_size = size; | |
a0f6da3d | 415 | dptr->kd_line = line; |
416 | ||
417 | spin_lock_irqsave(&kmem_lock, irq_flags); | |
418 | hlist_add_head_rcu(&dptr->kd_hlist, | |
419 | &kmem_table[hash_ptr(ptr, KMEM_HASH_BITS)]); | |
420 | list_add_tail(&dptr->kd_list, &kmem_list); | |
421 | spin_unlock_irqrestore(&kmem_lock, irq_flags); | |
422 | ||
423 | CDEBUG_LIMIT(D_INFO, "kmem_alloc(%llu, 0x%x) = %p " | |
424 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
425 | ptr, atomic64_read(&kmem_alloc_used), | |
426 | kmem_alloc_max); | |
427 | } | |
428 | out: | |
429 | RETURN(ptr); | |
430 | } | |
431 | EXPORT_SYMBOL(kmem_alloc_track); | |
432 | ||
433 | void | |
434 | kmem_free_track(void *ptr, size_t size) | |
435 | { | |
436 | kmem_debug_t *dptr; | |
437 | ENTRY; | |
438 | ||
439 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
440 | (unsigned long long) size); | |
441 | ||
442 | dptr = kmem_del_init(&kmem_lock, kmem_table, KMEM_HASH_BITS, ptr); | |
443 | ||
444 | ASSERT(dptr); /* Must exist in hash due to kmem_alloc() */ | |
445 | ||
446 | /* Size must match */ | |
447 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
448 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
449 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
450 | ||
451 | atomic64_sub(size, &kmem_alloc_used); | |
452 | ||
453 | CDEBUG_LIMIT(D_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
454 | (unsigned long long) size, atomic64_read(&kmem_alloc_used), | |
455 | kmem_alloc_max); | |
456 | ||
c8e60837 | 457 | kfree(dptr->kd_func); |
458 | ||
a0f6da3d | 459 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
460 | kfree(dptr); | |
461 | ||
462 | memset(ptr, 0x5a, size); | |
463 | kfree(ptr); | |
464 | ||
465 | EXIT; | |
466 | } | |
467 | EXPORT_SYMBOL(kmem_free_track); | |
468 | ||
469 | void * | |
470 | vmem_alloc_track(size_t size, int flags, const char *func, int line) | |
471 | { | |
472 | void *ptr = NULL; | |
473 | kmem_debug_t *dptr; | |
474 | unsigned long irq_flags; | |
475 | ENTRY; | |
476 | ||
477 | ASSERT(flags & KM_SLEEP); | |
478 | ||
479 | dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t), flags); | |
480 | if (dptr == NULL) { | |
481 | CWARN("vmem_alloc(%ld, 0x%x) debug failed\n", | |
482 | sizeof(kmem_debug_t), flags); | |
483 | } else { | |
c8e60837 | 484 | /* We use kstrdup() below because the string pointed to by |
485 | * __FUNCTION__ might not be available by the time we want | |
486 | * to print it, since the module might have been unloaded. */ | |
487 | dptr->kd_func = kstrdup(func, flags & ~__GFP_ZERO); | |
488 | if (unlikely(dptr->kd_func == NULL)) { | |
489 | kfree(dptr); | |
490 | CWARN("kstrdup() failed in vmem_alloc(%llu, 0x%x) " | |
491 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
492 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
493 | goto out; | |
494 | } | |
495 | ||
a0f6da3d | 496 | ptr = __vmalloc(size, (flags | __GFP_HIGHMEM) & ~__GFP_ZERO, |
497 | PAGE_KERNEL); | |
498 | ||
499 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 500 | kfree(dptr->kd_func); |
a0f6da3d | 501 | kfree(dptr); |
502 | CWARN("vmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
503 | (unsigned long long) size, flags, | |
504 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
505 | goto out; | |
506 | } | |
507 | ||
508 | if (flags & __GFP_ZERO) | |
509 | memset(ptr, 0, size); | |
510 | ||
511 | atomic64_add(size, &vmem_alloc_used); | |
512 | if (unlikely(atomic64_read(&vmem_alloc_used) > | |
513 | vmem_alloc_max)) | |
514 | vmem_alloc_max = | |
515 | atomic64_read(&vmem_alloc_used); | |
516 | ||
517 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
518 | INIT_LIST_HEAD(&dptr->kd_list); | |
519 | ||
520 | dptr->kd_addr = ptr; | |
521 | dptr->kd_size = size; | |
a0f6da3d | 522 | dptr->kd_line = line; |
523 | ||
524 | spin_lock_irqsave(&vmem_lock, irq_flags); | |
525 | hlist_add_head_rcu(&dptr->kd_hlist, | |
526 | &vmem_table[hash_ptr(ptr, VMEM_HASH_BITS)]); | |
527 | list_add_tail(&dptr->kd_list, &vmem_list); | |
528 | spin_unlock_irqrestore(&vmem_lock, irq_flags); | |
529 | ||
530 | CDEBUG_LIMIT(D_INFO, "vmem_alloc(%llu, 0x%x) = %p " | |
531 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
532 | ptr, atomic64_read(&vmem_alloc_used), | |
533 | vmem_alloc_max); | |
534 | } | |
535 | out: | |
536 | RETURN(ptr); | |
537 | } | |
538 | EXPORT_SYMBOL(vmem_alloc_track); | |
539 | ||
540 | void | |
541 | vmem_free_track(void *ptr, size_t size) | |
542 | { | |
543 | kmem_debug_t *dptr; | |
544 | ENTRY; | |
545 | ||
546 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
547 | (unsigned long long) size); | |
548 | ||
549 | dptr = kmem_del_init(&vmem_lock, vmem_table, VMEM_HASH_BITS, ptr); | |
550 | ASSERT(dptr); /* Must exist in hash due to vmem_alloc() */ | |
551 | ||
552 | /* Size must match */ | |
553 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
554 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
555 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
556 | ||
557 | atomic64_sub(size, &vmem_alloc_used); | |
558 | CDEBUG_LIMIT(D_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
559 | (unsigned long long) size, atomic64_read(&vmem_alloc_used), | |
560 | vmem_alloc_max); | |
561 | ||
c8e60837 | 562 | kfree(dptr->kd_func); |
563 | ||
a0f6da3d | 564 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
565 | kfree(dptr); | |
566 | ||
567 | memset(ptr, 0x5a, size); | |
568 | vfree(ptr); | |
569 | ||
570 | EXIT; | |
571 | } | |
572 | EXPORT_SYMBOL(vmem_free_track); | |
573 | ||
574 | # else /* DEBUG_KMEM_TRACKING */ | |
575 | ||
576 | void * | |
577 | kmem_alloc_debug(size_t size, int flags, const char *func, int line, | |
578 | int node_alloc, int node) | |
579 | { | |
580 | void *ptr; | |
581 | ENTRY; | |
582 | ||
583 | /* Marked unlikely because we should never be doing this, | |
584 | * we tolerate to up 2 pages but a single page is best. */ | |
585 | if (unlikely(size > (PAGE_SIZE * 2)) && kmem_warning_flag) | |
586 | CWARN("Large kmem_alloc(%llu, 0x%x) (%lld/%llu)\n", | |
587 | (unsigned long long) size, flags, | |
588 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
589 | ||
590 | /* Use the correct allocator */ | |
591 | if (node_alloc) { | |
592 | ASSERT(!(flags & __GFP_ZERO)); | |
593 | ptr = kmalloc_node(size, flags, node); | |
594 | } else if (flags & __GFP_ZERO) { | |
595 | ptr = kzalloc(size, flags & (~__GFP_ZERO)); | |
596 | } else { | |
597 | ptr = kmalloc(size, flags); | |
598 | } | |
599 | ||
600 | if (ptr == NULL) { | |
601 | CWARN("kmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
602 | (unsigned long long) size, flags, | |
603 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
604 | } else { | |
605 | atomic64_add(size, &kmem_alloc_used); | |
606 | if (unlikely(atomic64_read(&kmem_alloc_used) > kmem_alloc_max)) | |
607 | kmem_alloc_max = atomic64_read(&kmem_alloc_used); | |
608 | ||
609 | CDEBUG_LIMIT(D_INFO, "kmem_alloc(%llu, 0x%x) = %p " | |
610 | "(%lld/%llu)\n", (unsigned long long) size, flags, ptr, | |
611 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
612 | } | |
613 | RETURN(ptr); | |
614 | } | |
615 | EXPORT_SYMBOL(kmem_alloc_debug); | |
616 | ||
617 | void | |
618 | kmem_free_debug(void *ptr, size_t size) | |
619 | { | |
620 | ENTRY; | |
621 | ||
622 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
623 | (unsigned long long) size); | |
624 | ||
625 | atomic64_sub(size, &kmem_alloc_used); | |
626 | ||
627 | CDEBUG_LIMIT(D_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
628 | (unsigned long long) size, atomic64_read(&kmem_alloc_used), | |
629 | kmem_alloc_max); | |
630 | ||
631 | memset(ptr, 0x5a, size); | |
632 | kfree(ptr); | |
633 | ||
634 | EXIT; | |
635 | } | |
636 | EXPORT_SYMBOL(kmem_free_debug); | |
637 | ||
638 | void * | |
639 | vmem_alloc_debug(size_t size, int flags, const char *func, int line) | |
640 | { | |
641 | void *ptr; | |
642 | ENTRY; | |
643 | ||
644 | ASSERT(flags & KM_SLEEP); | |
645 | ||
646 | ptr = __vmalloc(size, (flags | __GFP_HIGHMEM) & ~__GFP_ZERO, | |
647 | PAGE_KERNEL); | |
648 | if (ptr == NULL) { | |
649 | CWARN("vmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
650 | (unsigned long long) size, flags, | |
651 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
652 | } else { | |
653 | if (flags & __GFP_ZERO) | |
654 | memset(ptr, 0, size); | |
655 | ||
656 | atomic64_add(size, &vmem_alloc_used); | |
657 | ||
658 | if (unlikely(atomic64_read(&vmem_alloc_used) > vmem_alloc_max)) | |
659 | vmem_alloc_max = atomic64_read(&vmem_alloc_used); | |
660 | ||
661 | CDEBUG_LIMIT(D_INFO, "vmem_alloc(%llu, 0x%x) = %p " | |
662 | "(%lld/%llu)\n", (unsigned long long) size, flags, ptr, | |
663 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
664 | } | |
665 | ||
666 | RETURN(ptr); | |
667 | } | |
668 | EXPORT_SYMBOL(vmem_alloc_debug); | |
669 | ||
670 | void | |
671 | vmem_free_debug(void *ptr, size_t size) | |
672 | { | |
673 | ENTRY; | |
674 | ||
675 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
676 | (unsigned long long) size); | |
677 | ||
678 | atomic64_sub(size, &vmem_alloc_used); | |
679 | ||
680 | CDEBUG_LIMIT(D_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
681 | (unsigned long long) size, atomic64_read(&vmem_alloc_used), | |
682 | vmem_alloc_max); | |
683 | ||
684 | memset(ptr, 0x5a, size); | |
685 | vfree(ptr); | |
686 | ||
687 | EXIT; | |
688 | } | |
689 | EXPORT_SYMBOL(vmem_free_debug); | |
690 | ||
691 | # endif /* DEBUG_KMEM_TRACKING */ | |
692 | #endif /* DEBUG_KMEM */ | |
693 | ||
a1502d76 | 694 | static void * |
695 | kv_alloc(spl_kmem_cache_t *skc, int size, int flags) | |
fece7c99 | 696 | { |
a1502d76 | 697 | void *ptr; |
f1ca4da6 | 698 | |
a1502d76 | 699 | if (skc->skc_flags & KMC_KMEM) { |
700 | if (size > (2 * PAGE_SIZE)) { | |
701 | ptr = (void *)__get_free_pages(flags, get_order(size)); | |
702 | } else | |
703 | ptr = kmem_alloc(size, flags); | |
704 | } else { | |
705 | ptr = vmem_alloc(size, flags); | |
d6a26c6a | 706 | } |
fece7c99 | 707 | |
a1502d76 | 708 | return ptr; |
709 | } | |
fece7c99 | 710 | |
a1502d76 | 711 | static void |
712 | kv_free(spl_kmem_cache_t *skc, void *ptr, int size) | |
713 | { | |
714 | if (skc->skc_flags & KMC_KMEM) { | |
715 | if (size > (2 * PAGE_SIZE)) | |
716 | free_pages((unsigned long)ptr, get_order(size)); | |
717 | else | |
718 | kmem_free(ptr, size); | |
719 | } else { | |
720 | vmem_free(ptr, size); | |
721 | } | |
fece7c99 | 722 | } |
723 | ||
ea3e6ca9 BB |
724 | /* |
725 | * It's important that we pack the spl_kmem_obj_t structure and the | |
48e0606a BB |
726 | * actual objects in to one large address space to minimize the number |
727 | * of calls to the allocator. It is far better to do a few large | |
728 | * allocations and then subdivide it ourselves. Now which allocator | |
729 | * we use requires balancing a few trade offs. | |
730 | * | |
731 | * For small objects we use kmem_alloc() because as long as you are | |
732 | * only requesting a small number of pages (ideally just one) its cheap. | |
733 | * However, when you start requesting multiple pages with kmem_alloc() | |
734 | * it gets increasingly expensive since it requires contigeous pages. | |
735 | * For this reason we shift to vmem_alloc() for slabs of large objects | |
736 | * which removes the need for contigeous pages. We do not use | |
737 | * vmem_alloc() in all cases because there is significant locking | |
738 | * overhead in __get_vm_area_node(). This function takes a single | |
739 | * global lock when aquiring an available virtual address range which | |
740 | * serializes all vmem_alloc()'s for all slab caches. Using slightly | |
741 | * different allocation functions for small and large objects should | |
742 | * give us the best of both worlds. | |
743 | * | |
744 | * KMC_ONSLAB KMC_OFFSLAB | |
745 | * | |
746 | * +------------------------+ +-----------------+ | |
747 | * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+ | |
748 | * | skc_obj_size <-+ | | +-----------------+ | | | |
749 | * | spl_kmem_obj_t | | | | | |
750 | * | skc_obj_size <---+ | +-----------------+ | | | |
751 | * | spl_kmem_obj_t | | | skc_obj_size | <-+ | | |
752 | * | ... v | | spl_kmem_obj_t | | | |
753 | * +------------------------+ +-----------------+ v | |
754 | */ | |
fece7c99 | 755 | static spl_kmem_slab_t * |
a1502d76 | 756 | spl_slab_alloc(spl_kmem_cache_t *skc, int flags) |
fece7c99 | 757 | { |
758 | spl_kmem_slab_t *sks; | |
a1502d76 | 759 | spl_kmem_obj_t *sko, *n; |
760 | void *base, *obj; | |
48e0606a BB |
761 | int i, align, size, rc = 0; |
762 | ||
a1502d76 | 763 | base = kv_alloc(skc, skc->skc_slab_size, flags); |
764 | if (base == NULL) | |
fece7c99 | 765 | RETURN(NULL); |
766 | ||
a1502d76 | 767 | sks = (spl_kmem_slab_t *)base; |
768 | sks->sks_magic = SKS_MAGIC; | |
769 | sks->sks_objs = skc->skc_slab_objs; | |
770 | sks->sks_age = jiffies; | |
771 | sks->sks_cache = skc; | |
772 | INIT_LIST_HEAD(&sks->sks_list); | |
773 | INIT_LIST_HEAD(&sks->sks_free_list); | |
774 | sks->sks_ref = 0; | |
48e0606a BB |
775 | |
776 | align = skc->skc_obj_align; | |
777 | size = P2ROUNDUP(skc->skc_obj_size, align) + | |
778 | P2ROUNDUP(sizeof(spl_kmem_obj_t), align); | |
fece7c99 | 779 | |
780 | for (i = 0; i < sks->sks_objs; i++) { | |
a1502d76 | 781 | if (skc->skc_flags & KMC_OFFSLAB) { |
782 | obj = kv_alloc(skc, size, flags); | |
783 | if (!obj) | |
784 | GOTO(out, rc = -ENOMEM); | |
785 | } else { | |
48e0606a BB |
786 | obj = base + |
787 | P2ROUNDUP(sizeof(spl_kmem_slab_t), align) + | |
788 | (i * size); | |
a1502d76 | 789 | } |
790 | ||
48e0606a | 791 | sko = obj + P2ROUNDUP(skc->skc_obj_size, align); |
fece7c99 | 792 | sko->sko_addr = obj; |
793 | sko->sko_magic = SKO_MAGIC; | |
794 | sko->sko_slab = sks; | |
795 | INIT_LIST_HEAD(&sko->sko_list); | |
fece7c99 | 796 | list_add_tail(&sko->sko_list, &sks->sks_free_list); |
797 | } | |
798 | ||
fece7c99 | 799 | list_for_each_entry(sko, &sks->sks_free_list, sko_list) |
800 | if (skc->skc_ctor) | |
801 | skc->skc_ctor(sko->sko_addr, skc->skc_private, flags); | |
2fb9b26a | 802 | out: |
a1502d76 | 803 | if (rc) { |
804 | if (skc->skc_flags & KMC_OFFSLAB) | |
48e0606a BB |
805 | list_for_each_entry_safe(sko, n, &sks->sks_free_list, |
806 | sko_list) | |
a1502d76 | 807 | kv_free(skc, sko->sko_addr, size); |
fece7c99 | 808 | |
a1502d76 | 809 | kv_free(skc, base, skc->skc_slab_size); |
810 | sks = NULL; | |
fece7c99 | 811 | } |
812 | ||
a1502d76 | 813 | RETURN(sks); |
fece7c99 | 814 | } |
815 | ||
ea3e6ca9 BB |
816 | /* |
817 | * Remove a slab from complete or partial list, it must be called with | |
818 | * the 'skc->skc_lock' held but the actual free must be performed | |
819 | * outside the lock to prevent deadlocking on vmem addresses. | |
fece7c99 | 820 | */ |
f1ca4da6 | 821 | static void |
ea3e6ca9 BB |
822 | spl_slab_free(spl_kmem_slab_t *sks, |
823 | struct list_head *sks_list, struct list_head *sko_list) | |
824 | { | |
2fb9b26a | 825 | spl_kmem_cache_t *skc; |
2fb9b26a | 826 | ENTRY; |
57d86234 | 827 | |
2fb9b26a | 828 | ASSERT(sks->sks_magic == SKS_MAGIC); |
4afaaefa | 829 | ASSERT(sks->sks_ref == 0); |
d6a26c6a | 830 | |
fece7c99 | 831 | skc = sks->sks_cache; |
832 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
d46630e0 | 833 | ASSERT(spin_is_locked(&skc->skc_lock)); |
f1ca4da6 | 834 | |
1a944a7d BB |
835 | /* |
836 | * Update slab/objects counters in the cache, then remove the | |
837 | * slab from the skc->skc_partial_list. Finally add the slab | |
838 | * and all its objects in to the private work lists where the | |
839 | * destructors will be called and the memory freed to the system. | |
840 | */ | |
fece7c99 | 841 | skc->skc_obj_total -= sks->sks_objs; |
842 | skc->skc_slab_total--; | |
843 | list_del(&sks->sks_list); | |
ea3e6ca9 | 844 | list_add(&sks->sks_list, sks_list); |
1a944a7d BB |
845 | list_splice_init(&sks->sks_free_list, sko_list); |
846 | ||
2fb9b26a | 847 | EXIT; |
848 | } | |
d6a26c6a | 849 | |
ea3e6ca9 BB |
850 | /* |
851 | * Traverses all the partial slabs attached to a cache and free those | |
852 | * which which are currently empty, and have not been touched for | |
37db7d8c BB |
853 | * skc_delay seconds to avoid thrashing. The count argument is |
854 | * passed to optionally cap the number of slabs reclaimed, a count | |
855 | * of zero means try and reclaim everything. When flag is set we | |
856 | * always free an available slab regardless of age. | |
ea3e6ca9 BB |
857 | */ |
858 | static void | |
37db7d8c | 859 | spl_slab_reclaim(spl_kmem_cache_t *skc, int count, int flag) |
2fb9b26a | 860 | { |
861 | spl_kmem_slab_t *sks, *m; | |
ea3e6ca9 BB |
862 | spl_kmem_obj_t *sko, *n; |
863 | LIST_HEAD(sks_list); | |
864 | LIST_HEAD(sko_list); | |
1a944a7d | 865 | int size = 0, i = 0; |
2fb9b26a | 866 | ENTRY; |
867 | ||
2fb9b26a | 868 | /* |
ea3e6ca9 BB |
869 | * Move empty slabs and objects which have not been touched in |
870 | * skc_delay seconds on to private lists to be freed outside | |
1a944a7d BB |
871 | * the spin lock. This delay time is important to avoid thrashing |
872 | * however when flag is set the delay will not be used. | |
2fb9b26a | 873 | */ |
ea3e6ca9 | 874 | spin_lock(&skc->skc_lock); |
1a944a7d BB |
875 | list_for_each_entry_safe_reverse(sks,m,&skc->skc_partial_list,sks_list){ |
876 | /* | |
877 | * All empty slabs are at the end of skc->skc_partial_list, | |
878 | * therefore once a non-empty slab is found we can stop | |
879 | * scanning. Additionally, stop when reaching the target | |
880 | * reclaim 'count' if a non-zero threshhold is given. | |
881 | */ | |
882 | if ((sks->sks_ref > 0) || (count && i > count)) | |
37db7d8c BB |
883 | break; |
884 | ||
37db7d8c | 885 | if (time_after(jiffies,sks->sks_age+skc->skc_delay*HZ)||flag) { |
ea3e6ca9 | 886 | spl_slab_free(sks, &sks_list, &sko_list); |
37db7d8c BB |
887 | i++; |
888 | } | |
ea3e6ca9 BB |
889 | } |
890 | spin_unlock(&skc->skc_lock); | |
891 | ||
892 | /* | |
1a944a7d BB |
893 | * The following two loops ensure all the object destructors are |
894 | * run, any offslab objects are freed, and the slabs themselves | |
895 | * are freed. This is all done outside the skc->skc_lock since | |
896 | * this allows the destructor to sleep, and allows us to perform | |
897 | * a conditional reschedule when a freeing a large number of | |
898 | * objects and slabs back to the system. | |
ea3e6ca9 | 899 | */ |
1a944a7d | 900 | if (skc->skc_flags & KMC_OFFSLAB) |
ea3e6ca9 BB |
901 | size = P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align) + |
902 | P2ROUNDUP(sizeof(spl_kmem_obj_t), skc->skc_obj_align); | |
903 | ||
1a944a7d BB |
904 | list_for_each_entry_safe(sko, n, &sko_list, sko_list) { |
905 | ASSERT(sko->sko_magic == SKO_MAGIC); | |
906 | ||
907 | if (skc->skc_dtor) | |
908 | skc->skc_dtor(sko->sko_addr, skc->skc_private); | |
909 | ||
910 | if (skc->skc_flags & KMC_OFFSLAB) | |
ea3e6ca9 | 911 | kv_free(skc, sko->sko_addr, size); |
1a944a7d BB |
912 | |
913 | cond_resched(); | |
2fb9b26a | 914 | } |
915 | ||
37db7d8c | 916 | list_for_each_entry_safe(sks, m, &sks_list, sks_list) { |
1a944a7d | 917 | ASSERT(sks->sks_magic == SKS_MAGIC); |
ea3e6ca9 | 918 | kv_free(skc, sks, skc->skc_slab_size); |
37db7d8c BB |
919 | cond_resched(); |
920 | } | |
ea3e6ca9 BB |
921 | |
922 | EXIT; | |
f1ca4da6 | 923 | } |
924 | ||
ea3e6ca9 BB |
925 | /* |
926 | * Called regularly on all caches to age objects out of the magazines | |
927 | * which have not been access in skc->skc_delay seconds. This prevents | |
928 | * idle magazines from holding memory which might be better used by | |
929 | * other caches or parts of the system. The delay is present to | |
930 | * prevent thrashing the magazine. | |
931 | */ | |
932 | static void | |
933 | spl_magazine_age(void *data) | |
f1ca4da6 | 934 | { |
ea3e6ca9 BB |
935 | spl_kmem_cache_t *skc = data; |
936 | spl_kmem_magazine_t *skm = skc->skc_mag[smp_processor_id()]; | |
f1ca4da6 | 937 | |
ea3e6ca9 BB |
938 | if (skm->skm_avail > 0 && |
939 | time_after(jiffies, skm->skm_age + skc->skc_delay * HZ)) | |
940 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
941 | } | |
4efd4118 | 942 | |
ea3e6ca9 BB |
943 | /* |
944 | * Called regularly to keep a downward pressure on the size of idle | |
945 | * magazines and to release free slabs from the cache. This function | |
946 | * never calls the registered reclaim function, that only occures | |
947 | * under memory pressure or with a direct call to spl_kmem_reap(). | |
948 | */ | |
949 | static void | |
950 | spl_cache_age(void *data) | |
951 | { | |
952 | spl_kmem_cache_t *skc = | |
953 | spl_get_work_data(data, spl_kmem_cache_t, skc_work.work); | |
954 | ||
955 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
37db7d8c BB |
956 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
957 | spl_on_each_cpu(spl_magazine_age, skc, 0); | |
ea3e6ca9 BB |
958 | |
959 | if (!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)) | |
37db7d8c | 960 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
2fb9b26a | 961 | } |
f1ca4da6 | 962 | |
ea3e6ca9 BB |
963 | /* |
964 | * Size a slab based on the size of each aliged object plus spl_kmem_obj_t. | |
965 | * When on-slab we want to target SPL_KMEM_CACHE_OBJ_PER_SLAB. However, | |
966 | * for very small objects we may end up with more than this so as not | |
967 | * to waste space in the minimal allocation of a single page. Also for | |
968 | * very large objects we may use as few as SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN, | |
969 | * lower than this and we will fail. | |
970 | */ | |
48e0606a BB |
971 | static int |
972 | spl_slab_size(spl_kmem_cache_t *skc, uint32_t *objs, uint32_t *size) | |
973 | { | |
ea3e6ca9 | 974 | int sks_size, obj_size, max_size, align; |
48e0606a BB |
975 | |
976 | if (skc->skc_flags & KMC_OFFSLAB) { | |
ea3e6ca9 | 977 | *objs = SPL_KMEM_CACHE_OBJ_PER_SLAB; |
48e0606a BB |
978 | *size = sizeof(spl_kmem_slab_t); |
979 | } else { | |
ea3e6ca9 BB |
980 | align = skc->skc_obj_align; |
981 | sks_size = P2ROUNDUP(sizeof(spl_kmem_slab_t), align); | |
982 | obj_size = P2ROUNDUP(skc->skc_obj_size, align) + | |
983 | P2ROUNDUP(sizeof(spl_kmem_obj_t), align); | |
984 | ||
985 | if (skc->skc_flags & KMC_KMEM) | |
986 | max_size = ((uint64_t)1 << (MAX_ORDER-1)) * PAGE_SIZE; | |
987 | else | |
988 | max_size = (32 * 1024 * 1024); | |
48e0606a | 989 | |
ea3e6ca9 BB |
990 | for (*size = PAGE_SIZE; *size <= max_size; *size += PAGE_SIZE) { |
991 | *objs = (*size - sks_size) / obj_size; | |
992 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB) | |
993 | RETURN(0); | |
994 | } | |
48e0606a | 995 | |
ea3e6ca9 BB |
996 | /* |
997 | * Unable to satisfy target objets per slab, fallback to | |
998 | * allocating a maximally sized slab and assuming it can | |
999 | * contain the minimum objects count use it. If not fail. | |
1000 | */ | |
1001 | *size = max_size; | |
1002 | *objs = (*size - sks_size) / obj_size; | |
1003 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN) | |
1004 | RETURN(0); | |
48e0606a BB |
1005 | } |
1006 | ||
ea3e6ca9 | 1007 | RETURN(-ENOSPC); |
48e0606a BB |
1008 | } |
1009 | ||
ea3e6ca9 BB |
1010 | /* |
1011 | * Make a guess at reasonable per-cpu magazine size based on the size of | |
1012 | * each object and the cost of caching N of them in each magazine. Long | |
1013 | * term this should really adapt based on an observed usage heuristic. | |
1014 | */ | |
4afaaefa | 1015 | static int |
1016 | spl_magazine_size(spl_kmem_cache_t *skc) | |
1017 | { | |
48e0606a | 1018 | int size, align = skc->skc_obj_align; |
4afaaefa | 1019 | ENTRY; |
1020 | ||
ea3e6ca9 | 1021 | /* Per-magazine sizes below assume a 4Kib page size */ |
48e0606a | 1022 | if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE * 256)) |
ea3e6ca9 | 1023 | size = 4; /* Minimum 4Mib per-magazine */ |
48e0606a | 1024 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE * 32)) |
ea3e6ca9 | 1025 | size = 16; /* Minimum 2Mib per-magazine */ |
48e0606a | 1026 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE)) |
ea3e6ca9 | 1027 | size = 64; /* Minimum 256Kib per-magazine */ |
48e0606a | 1028 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE / 4)) |
ea3e6ca9 | 1029 | size = 128; /* Minimum 128Kib per-magazine */ |
4afaaefa | 1030 | else |
ea3e6ca9 | 1031 | size = 256; |
4afaaefa | 1032 | |
1033 | RETURN(size); | |
1034 | } | |
1035 | ||
ea3e6ca9 BB |
1036 | /* |
1037 | * Allocate a per-cpu magazine to assoicate with a specific core. | |
1038 | */ | |
4afaaefa | 1039 | static spl_kmem_magazine_t * |
1040 | spl_magazine_alloc(spl_kmem_cache_t *skc, int node) | |
1041 | { | |
1042 | spl_kmem_magazine_t *skm; | |
1043 | int size = sizeof(spl_kmem_magazine_t) + | |
1044 | sizeof(void *) * skc->skc_mag_size; | |
1045 | ENTRY; | |
1046 | ||
ea3e6ca9 | 1047 | skm = kmem_alloc_node(size, GFP_KERNEL | __GFP_NOFAIL, node); |
4afaaefa | 1048 | if (skm) { |
1049 | skm->skm_magic = SKM_MAGIC; | |
1050 | skm->skm_avail = 0; | |
1051 | skm->skm_size = skc->skc_mag_size; | |
1052 | skm->skm_refill = skc->skc_mag_refill; | |
ea3e6ca9 | 1053 | skm->skm_age = jiffies; |
4afaaefa | 1054 | } |
1055 | ||
1056 | RETURN(skm); | |
1057 | } | |
1058 | ||
ea3e6ca9 BB |
1059 | /* |
1060 | * Free a per-cpu magazine assoicated with a specific core. | |
1061 | */ | |
4afaaefa | 1062 | static void |
1063 | spl_magazine_free(spl_kmem_magazine_t *skm) | |
1064 | { | |
a0f6da3d | 1065 | int size = sizeof(spl_kmem_magazine_t) + |
1066 | sizeof(void *) * skm->skm_size; | |
1067 | ||
4afaaefa | 1068 | ENTRY; |
1069 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1070 | ASSERT(skm->skm_avail == 0); | |
a0f6da3d | 1071 | |
1072 | kmem_free(skm, size); | |
4afaaefa | 1073 | EXIT; |
1074 | } | |
1075 | ||
ea3e6ca9 BB |
1076 | /* |
1077 | * Create all pre-cpu magazines of reasonable sizes. | |
1078 | */ | |
4afaaefa | 1079 | static int |
1080 | spl_magazine_create(spl_kmem_cache_t *skc) | |
1081 | { | |
37db7d8c | 1082 | int i; |
4afaaefa | 1083 | ENTRY; |
1084 | ||
1085 | skc->skc_mag_size = spl_magazine_size(skc); | |
ea3e6ca9 | 1086 | skc->skc_mag_refill = (skc->skc_mag_size + 1) / 2; |
4afaaefa | 1087 | |
37db7d8c BB |
1088 | for_each_online_cpu(i) { |
1089 | skc->skc_mag[i] = spl_magazine_alloc(skc, cpu_to_node(i)); | |
1090 | if (!skc->skc_mag[i]) { | |
1091 | for (i--; i >= 0; i--) | |
1092 | spl_magazine_free(skc->skc_mag[i]); | |
4afaaefa | 1093 | |
37db7d8c BB |
1094 | RETURN(-ENOMEM); |
1095 | } | |
1096 | } | |
4afaaefa | 1097 | |
37db7d8c | 1098 | RETURN(0); |
4afaaefa | 1099 | } |
1100 | ||
ea3e6ca9 BB |
1101 | /* |
1102 | * Destroy all pre-cpu magazines. | |
1103 | */ | |
4afaaefa | 1104 | static void |
1105 | spl_magazine_destroy(spl_kmem_cache_t *skc) | |
1106 | { | |
37db7d8c BB |
1107 | spl_kmem_magazine_t *skm; |
1108 | int i; | |
4afaaefa | 1109 | ENTRY; |
37db7d8c BB |
1110 | |
1111 | for_each_online_cpu(i) { | |
1112 | skm = skc->skc_mag[i]; | |
1113 | (void)spl_cache_flush(skc, skm, skm->skm_avail); | |
1114 | spl_magazine_free(skm); | |
1115 | } | |
1116 | ||
4afaaefa | 1117 | EXIT; |
1118 | } | |
1119 | ||
ea3e6ca9 BB |
1120 | /* |
1121 | * Create a object cache based on the following arguments: | |
1122 | * name cache name | |
1123 | * size cache object size | |
1124 | * align cache object alignment | |
1125 | * ctor cache object constructor | |
1126 | * dtor cache object destructor | |
1127 | * reclaim cache object reclaim | |
1128 | * priv cache private data for ctor/dtor/reclaim | |
1129 | * vmp unused must be NULL | |
1130 | * flags | |
1131 | * KMC_NOTOUCH Disable cache object aging (unsupported) | |
1132 | * KMC_NODEBUG Disable debugging (unsupported) | |
1133 | * KMC_NOMAGAZINE Disable magazine (unsupported) | |
1134 | * KMC_NOHASH Disable hashing (unsupported) | |
1135 | * KMC_QCACHE Disable qcache (unsupported) | |
1136 | * KMC_KMEM Force kmem backed cache | |
1137 | * KMC_VMEM Force vmem backed cache | |
1138 | * KMC_OFFSLAB Locate objects off the slab | |
1139 | */ | |
2fb9b26a | 1140 | spl_kmem_cache_t * |
1141 | spl_kmem_cache_create(char *name, size_t size, size_t align, | |
1142 | spl_kmem_ctor_t ctor, | |
1143 | spl_kmem_dtor_t dtor, | |
1144 | spl_kmem_reclaim_t reclaim, | |
1145 | void *priv, void *vmp, int flags) | |
1146 | { | |
1147 | spl_kmem_cache_t *skc; | |
a1502d76 | 1148 | int rc, kmem_flags = KM_SLEEP; |
2fb9b26a | 1149 | ENTRY; |
937879f1 | 1150 | |
a1502d76 | 1151 | ASSERTF(!(flags & KMC_NOMAGAZINE), "Bad KMC_NOMAGAZINE (%x)\n", flags); |
1152 | ASSERTF(!(flags & KMC_NOHASH), "Bad KMC_NOHASH (%x)\n", flags); | |
1153 | ASSERTF(!(flags & KMC_QCACHE), "Bad KMC_QCACHE (%x)\n", flags); | |
48e0606a | 1154 | ASSERT(vmp == NULL); |
a1502d76 | 1155 | |
2fb9b26a | 1156 | /* We may be called when there is a non-zero preempt_count or |
1157 | * interrupts are disabled is which case we must not sleep. | |
1158 | */ | |
e9d7a2be | 1159 | if (current_thread_info()->preempt_count || irqs_disabled()) |
2fb9b26a | 1160 | kmem_flags = KM_NOSLEEP; |
0a6fd143 | 1161 | |
2fb9b26a | 1162 | /* Allocate new cache memory and initialize. */ |
ff449ac4 | 1163 | skc = (spl_kmem_cache_t *)kmem_zalloc(sizeof(*skc), kmem_flags); |
e9d7a2be | 1164 | if (skc == NULL) |
2fb9b26a | 1165 | RETURN(NULL); |
d61e12af | 1166 | |
2fb9b26a | 1167 | skc->skc_magic = SKC_MAGIC; |
2fb9b26a | 1168 | skc->skc_name_size = strlen(name) + 1; |
1169 | skc->skc_name = (char *)kmem_alloc(skc->skc_name_size, kmem_flags); | |
1170 | if (skc->skc_name == NULL) { | |
1171 | kmem_free(skc, sizeof(*skc)); | |
1172 | RETURN(NULL); | |
1173 | } | |
1174 | strncpy(skc->skc_name, name, skc->skc_name_size); | |
1175 | ||
e9d7a2be | 1176 | skc->skc_ctor = ctor; |
1177 | skc->skc_dtor = dtor; | |
1178 | skc->skc_reclaim = reclaim; | |
2fb9b26a | 1179 | skc->skc_private = priv; |
1180 | skc->skc_vmp = vmp; | |
1181 | skc->skc_flags = flags; | |
1182 | skc->skc_obj_size = size; | |
48e0606a | 1183 | skc->skc_obj_align = SPL_KMEM_CACHE_ALIGN; |
2fb9b26a | 1184 | skc->skc_delay = SPL_KMEM_CACHE_DELAY; |
37db7d8c | 1185 | skc->skc_reap = SPL_KMEM_CACHE_REAP; |
ea3e6ca9 | 1186 | atomic_set(&skc->skc_ref, 0); |
2fb9b26a | 1187 | |
2fb9b26a | 1188 | INIT_LIST_HEAD(&skc->skc_list); |
1189 | INIT_LIST_HEAD(&skc->skc_complete_list); | |
1190 | INIT_LIST_HEAD(&skc->skc_partial_list); | |
d46630e0 | 1191 | spin_lock_init(&skc->skc_lock); |
e9d7a2be | 1192 | skc->skc_slab_fail = 0; |
1193 | skc->skc_slab_create = 0; | |
1194 | skc->skc_slab_destroy = 0; | |
2fb9b26a | 1195 | skc->skc_slab_total = 0; |
1196 | skc->skc_slab_alloc = 0; | |
1197 | skc->skc_slab_max = 0; | |
1198 | skc->skc_obj_total = 0; | |
1199 | skc->skc_obj_alloc = 0; | |
1200 | skc->skc_obj_max = 0; | |
a1502d76 | 1201 | |
48e0606a BB |
1202 | if (align) { |
1203 | ASSERT((align & (align - 1)) == 0); /* Power of two */ | |
1204 | ASSERT(align >= SPL_KMEM_CACHE_ALIGN); /* Minimum size */ | |
1205 | skc->skc_obj_align = align; | |
1206 | } | |
1207 | ||
a1502d76 | 1208 | /* If none passed select a cache type based on object size */ |
1209 | if (!(skc->skc_flags & (KMC_KMEM | KMC_VMEM))) { | |
48e0606a BB |
1210 | if (P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align) < |
1211 | (PAGE_SIZE / 8)) { | |
a1502d76 | 1212 | skc->skc_flags |= KMC_KMEM; |
1213 | } else { | |
1214 | skc->skc_flags |= KMC_VMEM; | |
1215 | } | |
1216 | } | |
1217 | ||
48e0606a BB |
1218 | rc = spl_slab_size(skc, &skc->skc_slab_objs, &skc->skc_slab_size); |
1219 | if (rc) | |
1220 | GOTO(out, rc); | |
4afaaefa | 1221 | |
1222 | rc = spl_magazine_create(skc); | |
48e0606a BB |
1223 | if (rc) |
1224 | GOTO(out, rc); | |
2fb9b26a | 1225 | |
ea3e6ca9 | 1226 | spl_init_delayed_work(&skc->skc_work, spl_cache_age, skc); |
37db7d8c | 1227 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
ea3e6ca9 | 1228 | |
2fb9b26a | 1229 | down_write(&spl_kmem_cache_sem); |
e9d7a2be | 1230 | list_add_tail(&skc->skc_list, &spl_kmem_cache_list); |
2fb9b26a | 1231 | up_write(&spl_kmem_cache_sem); |
1232 | ||
e9d7a2be | 1233 | RETURN(skc); |
48e0606a BB |
1234 | out: |
1235 | kmem_free(skc->skc_name, skc->skc_name_size); | |
1236 | kmem_free(skc, sizeof(*skc)); | |
1237 | RETURN(NULL); | |
f1ca4da6 | 1238 | } |
2fb9b26a | 1239 | EXPORT_SYMBOL(spl_kmem_cache_create); |
f1ca4da6 | 1240 | |
ea3e6ca9 BB |
1241 | /* |
1242 | * Destroy a cache and all objects assoicated with the cache. | |
1243 | */ | |
2fb9b26a | 1244 | void |
1245 | spl_kmem_cache_destroy(spl_kmem_cache_t *skc) | |
f1ca4da6 | 1246 | { |
ea3e6ca9 | 1247 | DECLARE_WAIT_QUEUE_HEAD(wq); |
2fb9b26a | 1248 | ENTRY; |
f1ca4da6 | 1249 | |
e9d7a2be | 1250 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1251 | ||
1252 | down_write(&spl_kmem_cache_sem); | |
1253 | list_del_init(&skc->skc_list); | |
1254 | up_write(&spl_kmem_cache_sem); | |
2fb9b26a | 1255 | |
ea3e6ca9 BB |
1256 | /* Cancel any and wait for any pending delayed work */ |
1257 | ASSERT(!test_and_set_bit(KMC_BIT_DESTROY, &skc->skc_flags)); | |
1258 | cancel_delayed_work(&skc->skc_work); | |
1259 | flush_scheduled_work(); | |
1260 | ||
1261 | /* Wait until all current callers complete, this is mainly | |
1262 | * to catch the case where a low memory situation triggers a | |
1263 | * cache reaping action which races with this destroy. */ | |
1264 | wait_event(wq, atomic_read(&skc->skc_ref) == 0); | |
1265 | ||
4afaaefa | 1266 | spl_magazine_destroy(skc); |
37db7d8c | 1267 | spl_slab_reclaim(skc, 0, 1); |
d46630e0 | 1268 | spin_lock(&skc->skc_lock); |
d6a26c6a | 1269 | |
2fb9b26a | 1270 | /* Validate there are no objects in use and free all the |
4afaaefa | 1271 | * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */ |
ea3e6ca9 BB |
1272 | ASSERT3U(skc->skc_slab_alloc, ==, 0); |
1273 | ASSERT3U(skc->skc_obj_alloc, ==, 0); | |
1274 | ASSERT3U(skc->skc_slab_total, ==, 0); | |
1275 | ASSERT3U(skc->skc_obj_total, ==, 0); | |
2fb9b26a | 1276 | ASSERT(list_empty(&skc->skc_complete_list)); |
a1502d76 | 1277 | |
2fb9b26a | 1278 | kmem_free(skc->skc_name, skc->skc_name_size); |
d46630e0 | 1279 | spin_unlock(&skc->skc_lock); |
ff449ac4 | 1280 | |
4afaaefa | 1281 | kmem_free(skc, sizeof(*skc)); |
2fb9b26a | 1282 | |
1283 | EXIT; | |
f1ca4da6 | 1284 | } |
2fb9b26a | 1285 | EXPORT_SYMBOL(spl_kmem_cache_destroy); |
f1ca4da6 | 1286 | |
ea3e6ca9 BB |
1287 | /* |
1288 | * Allocate an object from a slab attached to the cache. This is used to | |
1289 | * repopulate the per-cpu magazine caches in batches when they run low. | |
1290 | */ | |
4afaaefa | 1291 | static void * |
1292 | spl_cache_obj(spl_kmem_cache_t *skc, spl_kmem_slab_t *sks) | |
f1ca4da6 | 1293 | { |
2fb9b26a | 1294 | spl_kmem_obj_t *sko; |
f1ca4da6 | 1295 | |
e9d7a2be | 1296 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1297 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
4afaaefa | 1298 | ASSERT(spin_is_locked(&skc->skc_lock)); |
2fb9b26a | 1299 | |
a1502d76 | 1300 | sko = list_entry(sks->sks_free_list.next, spl_kmem_obj_t, sko_list); |
4afaaefa | 1301 | ASSERT(sko->sko_magic == SKO_MAGIC); |
1302 | ASSERT(sko->sko_addr != NULL); | |
2fb9b26a | 1303 | |
a1502d76 | 1304 | /* Remove from sks_free_list */ |
4afaaefa | 1305 | list_del_init(&sko->sko_list); |
2fb9b26a | 1306 | |
4afaaefa | 1307 | sks->sks_age = jiffies; |
1308 | sks->sks_ref++; | |
1309 | skc->skc_obj_alloc++; | |
2fb9b26a | 1310 | |
4afaaefa | 1311 | /* Track max obj usage statistics */ |
1312 | if (skc->skc_obj_alloc > skc->skc_obj_max) | |
1313 | skc->skc_obj_max = skc->skc_obj_alloc; | |
2fb9b26a | 1314 | |
4afaaefa | 1315 | /* Track max slab usage statistics */ |
1316 | if (sks->sks_ref == 1) { | |
1317 | skc->skc_slab_alloc++; | |
f1ca4da6 | 1318 | |
4afaaefa | 1319 | if (skc->skc_slab_alloc > skc->skc_slab_max) |
1320 | skc->skc_slab_max = skc->skc_slab_alloc; | |
2fb9b26a | 1321 | } |
1322 | ||
4afaaefa | 1323 | return sko->sko_addr; |
1324 | } | |
c30df9c8 | 1325 | |
ea3e6ca9 BB |
1326 | /* |
1327 | * No available objects on any slabsi, create a new slab. Since this | |
1328 | * is an expensive operation we do it without holding the spinlock and | |
1329 | * only briefly aquire it when we link in the fully allocated and | |
1330 | * constructed slab. | |
4afaaefa | 1331 | */ |
1332 | static spl_kmem_slab_t * | |
1333 | spl_cache_grow(spl_kmem_cache_t *skc, int flags) | |
1334 | { | |
e9d7a2be | 1335 | spl_kmem_slab_t *sks; |
4afaaefa | 1336 | ENTRY; |
f1ca4da6 | 1337 | |
e9d7a2be | 1338 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1339 | local_irq_enable(); |
1340 | might_sleep(); | |
e9d7a2be | 1341 | |
ea3e6ca9 BB |
1342 | /* |
1343 | * Before allocating a new slab check if the slab is being reaped. | |
1344 | * If it is there is a good chance we can wait until it finishes | |
1345 | * and then use one of the newly freed but not aged-out slabs. | |
1346 | */ | |
1347 | if (test_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
1348 | schedule(); | |
1349 | GOTO(out, sks= NULL); | |
4afaaefa | 1350 | } |
2fb9b26a | 1351 | |
ea3e6ca9 BB |
1352 | /* Allocate a new slab for the cache */ |
1353 | sks = spl_slab_alloc(skc, flags | __GFP_NORETRY | __GFP_NOWARN); | |
1354 | if (sks == NULL) | |
1355 | GOTO(out, sks = NULL); | |
4afaaefa | 1356 | |
ea3e6ca9 | 1357 | /* Link the new empty slab in to the end of skc_partial_list. */ |
d46630e0 | 1358 | spin_lock(&skc->skc_lock); |
2fb9b26a | 1359 | skc->skc_slab_total++; |
1360 | skc->skc_obj_total += sks->sks_objs; | |
1361 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
d46630e0 | 1362 | spin_unlock(&skc->skc_lock); |
ea3e6ca9 BB |
1363 | out: |
1364 | local_irq_disable(); | |
4afaaefa | 1365 | |
1366 | RETURN(sks); | |
f1ca4da6 | 1367 | } |
1368 | ||
ea3e6ca9 BB |
1369 | /* |
1370 | * Refill a per-cpu magazine with objects from the slabs for this | |
1371 | * cache. Ideally the magazine can be repopulated using existing | |
1372 | * objects which have been released, however if we are unable to | |
1373 | * locate enough free objects new slabs of objects will be created. | |
1374 | */ | |
4afaaefa | 1375 | static int |
1376 | spl_cache_refill(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flags) | |
f1ca4da6 | 1377 | { |
e9d7a2be | 1378 | spl_kmem_slab_t *sks; |
1379 | int rc = 0, refill; | |
937879f1 | 1380 | ENTRY; |
f1ca4da6 | 1381 | |
e9d7a2be | 1382 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1383 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1384 | ||
e9d7a2be | 1385 | refill = MIN(skm->skm_refill, skm->skm_size - skm->skm_avail); |
d46630e0 | 1386 | spin_lock(&skc->skc_lock); |
ff449ac4 | 1387 | |
4afaaefa | 1388 | while (refill > 0) { |
ea3e6ca9 | 1389 | /* No slabs available we may need to grow the cache */ |
4afaaefa | 1390 | if (list_empty(&skc->skc_partial_list)) { |
1391 | spin_unlock(&skc->skc_lock); | |
ff449ac4 | 1392 | |
4afaaefa | 1393 | sks = spl_cache_grow(skc, flags); |
1394 | if (!sks) | |
e9d7a2be | 1395 | GOTO(out, rc); |
4afaaefa | 1396 | |
1397 | /* Rescheduled to different CPU skm is not local */ | |
1398 | if (skm != skc->skc_mag[smp_processor_id()]) | |
e9d7a2be | 1399 | GOTO(out, rc); |
1400 | ||
1401 | /* Potentially rescheduled to the same CPU but | |
1402 | * allocations may have occured from this CPU while | |
1403 | * we were sleeping so recalculate max refill. */ | |
1404 | refill = MIN(refill, skm->skm_size - skm->skm_avail); | |
4afaaefa | 1405 | |
1406 | spin_lock(&skc->skc_lock); | |
1407 | continue; | |
1408 | } | |
d46630e0 | 1409 | |
4afaaefa | 1410 | /* Grab the next available slab */ |
1411 | sks = list_entry((&skc->skc_partial_list)->next, | |
1412 | spl_kmem_slab_t, sks_list); | |
1413 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
1414 | ASSERT(sks->sks_ref < sks->sks_objs); | |
1415 | ASSERT(!list_empty(&sks->sks_free_list)); | |
d46630e0 | 1416 | |
4afaaefa | 1417 | /* Consume as many objects as needed to refill the requested |
e9d7a2be | 1418 | * cache. We must also be careful not to overfill it. */ |
1419 | while (sks->sks_ref < sks->sks_objs && refill-- > 0 && ++rc) { | |
1420 | ASSERT(skm->skm_avail < skm->skm_size); | |
1421 | ASSERT(rc < skm->skm_size); | |
4afaaefa | 1422 | skm->skm_objs[skm->skm_avail++]=spl_cache_obj(skc,sks); |
e9d7a2be | 1423 | } |
f1ca4da6 | 1424 | |
4afaaefa | 1425 | /* Move slab to skc_complete_list when full */ |
1426 | if (sks->sks_ref == sks->sks_objs) { | |
1427 | list_del(&sks->sks_list); | |
1428 | list_add(&sks->sks_list, &skc->skc_complete_list); | |
2fb9b26a | 1429 | } |
1430 | } | |
57d86234 | 1431 | |
4afaaefa | 1432 | spin_unlock(&skc->skc_lock); |
1433 | out: | |
1434 | /* Returns the number of entries added to cache */ | |
e9d7a2be | 1435 | RETURN(rc); |
4afaaefa | 1436 | } |
1437 | ||
ea3e6ca9 BB |
1438 | /* |
1439 | * Release an object back to the slab from which it came. | |
1440 | */ | |
4afaaefa | 1441 | static void |
1442 | spl_cache_shrink(spl_kmem_cache_t *skc, void *obj) | |
1443 | { | |
e9d7a2be | 1444 | spl_kmem_slab_t *sks = NULL; |
4afaaefa | 1445 | spl_kmem_obj_t *sko = NULL; |
1446 | ENTRY; | |
1447 | ||
e9d7a2be | 1448 | ASSERT(skc->skc_magic == SKC_MAGIC); |
4afaaefa | 1449 | ASSERT(spin_is_locked(&skc->skc_lock)); |
1450 | ||
48e0606a | 1451 | sko = obj + P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align); |
a1502d76 | 1452 | ASSERT(sko->sko_magic == SKO_MAGIC); |
4afaaefa | 1453 | |
1454 | sks = sko->sko_slab; | |
a1502d76 | 1455 | ASSERT(sks->sks_magic == SKS_MAGIC); |
2fb9b26a | 1456 | ASSERT(sks->sks_cache == skc); |
2fb9b26a | 1457 | list_add(&sko->sko_list, &sks->sks_free_list); |
d6a26c6a | 1458 | |
2fb9b26a | 1459 | sks->sks_age = jiffies; |
4afaaefa | 1460 | sks->sks_ref--; |
2fb9b26a | 1461 | skc->skc_obj_alloc--; |
f1ca4da6 | 1462 | |
2fb9b26a | 1463 | /* Move slab to skc_partial_list when no longer full. Slabs |
4afaaefa | 1464 | * are added to the head to keep the partial list is quasi-full |
1465 | * sorted order. Fuller at the head, emptier at the tail. */ | |
1466 | if (sks->sks_ref == (sks->sks_objs - 1)) { | |
2fb9b26a | 1467 | list_del(&sks->sks_list); |
1468 | list_add(&sks->sks_list, &skc->skc_partial_list); | |
1469 | } | |
f1ca4da6 | 1470 | |
2fb9b26a | 1471 | /* Move emply slabs to the end of the partial list so |
4afaaefa | 1472 | * they can be easily found and freed during reclamation. */ |
1473 | if (sks->sks_ref == 0) { | |
2fb9b26a | 1474 | list_del(&sks->sks_list); |
1475 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
1476 | skc->skc_slab_alloc--; | |
1477 | } | |
1478 | ||
4afaaefa | 1479 | EXIT; |
1480 | } | |
1481 | ||
ea3e6ca9 BB |
1482 | /* |
1483 | * Release a batch of objects from a per-cpu magazine back to their | |
1484 | * respective slabs. This occurs when we exceed the magazine size, | |
1485 | * are under memory pressure, when the cache is idle, or during | |
1486 | * cache cleanup. The flush argument contains the number of entries | |
1487 | * to remove from the magazine. | |
1488 | */ | |
4afaaefa | 1489 | static int |
1490 | spl_cache_flush(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flush) | |
1491 | { | |
1492 | int i, count = MIN(flush, skm->skm_avail); | |
1493 | ENTRY; | |
1494 | ||
e9d7a2be | 1495 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1496 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
4afaaefa | 1497 | |
ea3e6ca9 BB |
1498 | /* |
1499 | * XXX: Currently we simply return objects from the magazine to | |
1500 | * the slabs in fifo order. The ideal thing to do from a memory | |
1501 | * fragmentation standpoint is to cheaply determine the set of | |
1502 | * objects in the magazine which will result in the largest | |
1503 | * number of free slabs if released from the magazine. | |
1504 | */ | |
4afaaefa | 1505 | spin_lock(&skc->skc_lock); |
1506 | for (i = 0; i < count; i++) | |
1507 | spl_cache_shrink(skc, skm->skm_objs[i]); | |
1508 | ||
e9d7a2be | 1509 | skm->skm_avail -= count; |
1510 | memmove(skm->skm_objs, &(skm->skm_objs[count]), | |
4afaaefa | 1511 | sizeof(void *) * skm->skm_avail); |
1512 | ||
d46630e0 | 1513 | spin_unlock(&skc->skc_lock); |
4afaaefa | 1514 | |
1515 | RETURN(count); | |
1516 | } | |
1517 | ||
ea3e6ca9 BB |
1518 | /* |
1519 | * Allocate an object from the per-cpu magazine, or if the magazine | |
1520 | * is empty directly allocate from a slab and repopulate the magazine. | |
1521 | */ | |
4afaaefa | 1522 | void * |
1523 | spl_kmem_cache_alloc(spl_kmem_cache_t *skc, int flags) | |
1524 | { | |
1525 | spl_kmem_magazine_t *skm; | |
1526 | unsigned long irq_flags; | |
1527 | void *obj = NULL; | |
1528 | ENTRY; | |
1529 | ||
e9d7a2be | 1530 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1531 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1532 | ASSERT(flags & KM_SLEEP); | |
1533 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1534 | local_irq_save(irq_flags); |
1535 | ||
1536 | restart: | |
1537 | /* Safe to update per-cpu structure without lock, but | |
1538 | * in the restart case we must be careful to reaquire | |
1539 | * the local magazine since this may have changed | |
1540 | * when we need to grow the cache. */ | |
1541 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1542 | ASSERTF(skm->skm_magic == SKM_MAGIC, "%x != %x: %s/%p/%p %x/%x/%x\n", |
1543 | skm->skm_magic, SKM_MAGIC, skc->skc_name, skc, skm, | |
1544 | skm->skm_size, skm->skm_refill, skm->skm_avail); | |
4afaaefa | 1545 | |
1546 | if (likely(skm->skm_avail)) { | |
1547 | /* Object available in CPU cache, use it */ | |
1548 | obj = skm->skm_objs[--skm->skm_avail]; | |
ea3e6ca9 | 1549 | skm->skm_age = jiffies; |
4afaaefa | 1550 | } else { |
1551 | /* Per-CPU cache empty, directly allocate from | |
1552 | * the slab and refill the per-CPU cache. */ | |
1553 | (void)spl_cache_refill(skc, skm, flags); | |
1554 | GOTO(restart, obj = NULL); | |
1555 | } | |
1556 | ||
1557 | local_irq_restore(irq_flags); | |
fece7c99 | 1558 | ASSERT(obj); |
48e0606a | 1559 | ASSERT(((unsigned long)(obj) % skc->skc_obj_align) == 0); |
4afaaefa | 1560 | |
1561 | /* Pre-emptively migrate object to CPU L1 cache */ | |
1562 | prefetchw(obj); | |
ea3e6ca9 | 1563 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1564 | |
1565 | RETURN(obj); | |
1566 | } | |
1567 | EXPORT_SYMBOL(spl_kmem_cache_alloc); | |
1568 | ||
ea3e6ca9 BB |
1569 | /* |
1570 | * Free an object back to the local per-cpu magazine, there is no | |
1571 | * guarantee that this is the same magazine the object was originally | |
1572 | * allocated from. We may need to flush entire from the magazine | |
1573 | * back to the slabs to make space. | |
1574 | */ | |
4afaaefa | 1575 | void |
1576 | spl_kmem_cache_free(spl_kmem_cache_t *skc, void *obj) | |
1577 | { | |
1578 | spl_kmem_magazine_t *skm; | |
1579 | unsigned long flags; | |
1580 | ENTRY; | |
1581 | ||
e9d7a2be | 1582 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1583 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1584 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1585 | local_irq_save(flags); |
1586 | ||
1587 | /* Safe to update per-cpu structure without lock, but | |
1588 | * no remote memory allocation tracking is being performed | |
1589 | * it is entirely possible to allocate an object from one | |
1590 | * CPU cache and return it to another. */ | |
1591 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1592 | ASSERT(skm->skm_magic == SKM_MAGIC); |
4afaaefa | 1593 | |
1594 | /* Per-CPU cache full, flush it to make space */ | |
1595 | if (unlikely(skm->skm_avail >= skm->skm_size)) | |
1596 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
1597 | ||
1598 | /* Available space in cache, use it */ | |
1599 | skm->skm_objs[skm->skm_avail++] = obj; | |
1600 | ||
1601 | local_irq_restore(flags); | |
ea3e6ca9 | 1602 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1603 | |
1604 | EXIT; | |
f1ca4da6 | 1605 | } |
2fb9b26a | 1606 | EXPORT_SYMBOL(spl_kmem_cache_free); |
5c2bb9b2 | 1607 | |
ea3e6ca9 BB |
1608 | /* |
1609 | * The generic shrinker function for all caches. Under linux a shrinker | |
1610 | * may not be tightly coupled with a slab cache. In fact linux always | |
1611 | * systematically trys calling all registered shrinker callbacks which | |
1612 | * report that they contain unused objects. Because of this we only | |
1613 | * register one shrinker function in the shim layer for all slab caches. | |
1614 | * We always attempt to shrink all caches when this generic shrinker | |
1615 | * is called. The shrinker should return the number of free objects | |
1616 | * in the cache when called with nr_to_scan == 0 but not attempt to | |
1617 | * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan | |
1618 | * objects should be freed, because Solaris semantics are to free | |
1619 | * all available objects we may free more objects than requested. | |
1620 | */ | |
2fb9b26a | 1621 | static int |
4afaaefa | 1622 | spl_kmem_cache_generic_shrinker(int nr_to_scan, unsigned int gfp_mask) |
2fb9b26a | 1623 | { |
e9d7a2be | 1624 | spl_kmem_cache_t *skc; |
ea3e6ca9 | 1625 | int unused = 0; |
5c2bb9b2 | 1626 | |
e9d7a2be | 1627 | down_read(&spl_kmem_cache_sem); |
ea3e6ca9 BB |
1628 | list_for_each_entry(skc, &spl_kmem_cache_list, skc_list) { |
1629 | if (nr_to_scan) | |
1630 | spl_kmem_cache_reap_now(skc); | |
1631 | ||
1632 | /* | |
1633 | * Presume everything alloc'ed in reclaimable, this ensures | |
1634 | * we are called again with nr_to_scan > 0 so can try and | |
1635 | * reclaim. The exact number is not important either so | |
1636 | * we forgo taking this already highly contented lock. | |
1637 | */ | |
1638 | unused += skc->skc_obj_alloc; | |
1639 | } | |
e9d7a2be | 1640 | up_read(&spl_kmem_cache_sem); |
2fb9b26a | 1641 | |
ea3e6ca9 | 1642 | return (unused * sysctl_vfs_cache_pressure) / 100; |
5c2bb9b2 | 1643 | } |
5c2bb9b2 | 1644 | |
ea3e6ca9 BB |
1645 | /* |
1646 | * Call the registered reclaim function for a cache. Depending on how | |
1647 | * many and which objects are released it may simply repopulate the | |
1648 | * local magazine which will then need to age-out. Objects which cannot | |
1649 | * fit in the magazine we will be released back to their slabs which will | |
1650 | * also need to age out before being release. This is all just best | |
1651 | * effort and we do not want to thrash creating and destroying slabs. | |
1652 | */ | |
57d86234 | 1653 | void |
2fb9b26a | 1654 | spl_kmem_cache_reap_now(spl_kmem_cache_t *skc) |
57d86234 | 1655 | { |
2fb9b26a | 1656 | ENTRY; |
e9d7a2be | 1657 | |
1658 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
ea3e6ca9 | 1659 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
2fb9b26a | 1660 | |
ea3e6ca9 BB |
1661 | /* Prevent concurrent cache reaping when contended */ |
1662 | if (test_and_set_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
1663 | EXIT; | |
1664 | return; | |
1665 | } | |
2fb9b26a | 1666 | |
ea3e6ca9 | 1667 | atomic_inc(&skc->skc_ref); |
4afaaefa | 1668 | |
ea3e6ca9 BB |
1669 | if (skc->skc_reclaim) |
1670 | skc->skc_reclaim(skc->skc_private); | |
4afaaefa | 1671 | |
37db7d8c | 1672 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
ea3e6ca9 BB |
1673 | clear_bit(KMC_BIT_REAPING, &skc->skc_flags); |
1674 | atomic_dec(&skc->skc_ref); | |
4afaaefa | 1675 | |
2fb9b26a | 1676 | EXIT; |
57d86234 | 1677 | } |
2fb9b26a | 1678 | EXPORT_SYMBOL(spl_kmem_cache_reap_now); |
57d86234 | 1679 | |
ea3e6ca9 BB |
1680 | /* |
1681 | * Reap all free slabs from all registered caches. | |
1682 | */ | |
f1b59d26 | 1683 | void |
2fb9b26a | 1684 | spl_kmem_reap(void) |
937879f1 | 1685 | { |
4afaaefa | 1686 | spl_kmem_cache_generic_shrinker(KMC_REAP_CHUNK, GFP_KERNEL); |
f1ca4da6 | 1687 | } |
2fb9b26a | 1688 | EXPORT_SYMBOL(spl_kmem_reap); |
5d86345d | 1689 | |
ff449ac4 | 1690 | #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING) |
c6dc93d6 | 1691 | static char * |
4afaaefa | 1692 | spl_sprintf_addr(kmem_debug_t *kd, char *str, int len, int min) |
d6a26c6a | 1693 | { |
e9d7a2be | 1694 | int size = ((len - 1) < kd->kd_size) ? (len - 1) : kd->kd_size; |
d6a26c6a | 1695 | int i, flag = 1; |
1696 | ||
1697 | ASSERT(str != NULL && len >= 17); | |
e9d7a2be | 1698 | memset(str, 0, len); |
d6a26c6a | 1699 | |
1700 | /* Check for a fully printable string, and while we are at | |
1701 | * it place the printable characters in the passed buffer. */ | |
1702 | for (i = 0; i < size; i++) { | |
e9d7a2be | 1703 | str[i] = ((char *)(kd->kd_addr))[i]; |
1704 | if (isprint(str[i])) { | |
1705 | continue; | |
1706 | } else { | |
1707 | /* Minimum number of printable characters found | |
1708 | * to make it worthwhile to print this as ascii. */ | |
1709 | if (i > min) | |
1710 | break; | |
1711 | ||
1712 | flag = 0; | |
1713 | break; | |
1714 | } | |
d6a26c6a | 1715 | } |
1716 | ||
1717 | if (!flag) { | |
1718 | sprintf(str, "%02x%02x%02x%02x%02x%02x%02x%02x", | |
1719 | *((uint8_t *)kd->kd_addr), | |
1720 | *((uint8_t *)kd->kd_addr + 2), | |
1721 | *((uint8_t *)kd->kd_addr + 4), | |
1722 | *((uint8_t *)kd->kd_addr + 6), | |
1723 | *((uint8_t *)kd->kd_addr + 8), | |
1724 | *((uint8_t *)kd->kd_addr + 10), | |
1725 | *((uint8_t *)kd->kd_addr + 12), | |
1726 | *((uint8_t *)kd->kd_addr + 14)); | |
1727 | } | |
1728 | ||
1729 | return str; | |
1730 | } | |
1731 | ||
a1502d76 | 1732 | static int |
1733 | spl_kmem_init_tracking(struct list_head *list, spinlock_t *lock, int size) | |
1734 | { | |
1735 | int i; | |
1736 | ENTRY; | |
1737 | ||
1738 | spin_lock_init(lock); | |
1739 | INIT_LIST_HEAD(list); | |
1740 | ||
1741 | for (i = 0; i < size; i++) | |
1742 | INIT_HLIST_HEAD(&kmem_table[i]); | |
1743 | ||
1744 | RETURN(0); | |
1745 | } | |
1746 | ||
ff449ac4 | 1747 | static void |
1748 | spl_kmem_fini_tracking(struct list_head *list, spinlock_t *lock) | |
5d86345d | 1749 | { |
2fb9b26a | 1750 | unsigned long flags; |
1751 | kmem_debug_t *kd; | |
1752 | char str[17]; | |
a1502d76 | 1753 | ENTRY; |
2fb9b26a | 1754 | |
ff449ac4 | 1755 | spin_lock_irqsave(lock, flags); |
1756 | if (!list_empty(list)) | |
a0f6da3d | 1757 | printk(KERN_WARNING "%-16s %-5s %-16s %s:%s\n", "address", |
1758 | "size", "data", "func", "line"); | |
2fb9b26a | 1759 | |
ff449ac4 | 1760 | list_for_each_entry(kd, list, kd_list) |
a0f6da3d | 1761 | printk(KERN_WARNING "%p %-5d %-16s %s:%d\n", kd->kd_addr, |
b6b2acc6 | 1762 | (int)kd->kd_size, spl_sprintf_addr(kd, str, 17, 8), |
2fb9b26a | 1763 | kd->kd_func, kd->kd_line); |
1764 | ||
ff449ac4 | 1765 | spin_unlock_irqrestore(lock, flags); |
a1502d76 | 1766 | EXIT; |
ff449ac4 | 1767 | } |
1768 | #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
a1502d76 | 1769 | #define spl_kmem_init_tracking(list, lock, size) |
ff449ac4 | 1770 | #define spl_kmem_fini_tracking(list, lock) |
1771 | #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
1772 | ||
36b313da BB |
1773 | static void |
1774 | spl_kmem_init_globals(void) | |
1775 | { | |
1776 | struct zone *zone; | |
1777 | ||
1778 | /* For now all zones are includes, it may be wise to restrict | |
1779 | * this to normal and highmem zones if we see problems. */ | |
1780 | for_each_zone(zone) { | |
1781 | ||
1782 | if (!populated_zone(zone)) | |
1783 | continue; | |
1784 | ||
1785 | minfree += zone->pages_min; | |
1786 | desfree += zone->pages_low; | |
1787 | lotsfree += zone->pages_high; | |
1788 | } | |
4ab13d3b BB |
1789 | |
1790 | /* Solaris default values */ | |
1791 | swapfs_minfree = MAX(2*1024*1024 / PAGE_SIZE, physmem / 8); | |
1792 | swapfs_reserve = MIN(4*1024*1024 / PAGE_SIZE, physmem / 16); | |
36b313da BB |
1793 | } |
1794 | ||
a1502d76 | 1795 | int |
1796 | spl_kmem_init(void) | |
1797 | { | |
1798 | int rc = 0; | |
1799 | ENTRY; | |
1800 | ||
1801 | init_rwsem(&spl_kmem_cache_sem); | |
1802 | INIT_LIST_HEAD(&spl_kmem_cache_list); | |
36b313da | 1803 | spl_kmem_init_globals(); |
a1502d76 | 1804 | |
1805 | #ifdef HAVE_SET_SHRINKER | |
1806 | spl_kmem_cache_shrinker = set_shrinker(KMC_DEFAULT_SEEKS, | |
1807 | spl_kmem_cache_generic_shrinker); | |
1808 | if (spl_kmem_cache_shrinker == NULL) | |
f78a933f | 1809 | RETURN(rc = -ENOMEM); |
a1502d76 | 1810 | #else |
1811 | register_shrinker(&spl_kmem_cache_shrinker); | |
1812 | #endif | |
1813 | ||
1814 | #ifdef DEBUG_KMEM | |
1815 | atomic64_set(&kmem_alloc_used, 0); | |
1816 | atomic64_set(&vmem_alloc_used, 0); | |
1817 | ||
1818 | spl_kmem_init_tracking(&kmem_list, &kmem_lock, KMEM_TABLE_SIZE); | |
1819 | spl_kmem_init_tracking(&vmem_list, &vmem_lock, VMEM_TABLE_SIZE); | |
1820 | #endif | |
a1502d76 | 1821 | RETURN(rc); |
1822 | } | |
1823 | ||
ff449ac4 | 1824 | void |
1825 | spl_kmem_fini(void) | |
1826 | { | |
1827 | #ifdef DEBUG_KMEM | |
1828 | /* Display all unreclaimed memory addresses, including the | |
1829 | * allocation size and the first few bytes of what's located | |
1830 | * at that address to aid in debugging. Performance is not | |
1831 | * a serious concern here since it is module unload time. */ | |
1832 | if (atomic64_read(&kmem_alloc_used) != 0) | |
1833 | CWARN("kmem leaked %ld/%ld bytes\n", | |
550f1705 | 1834 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); |
ff449ac4 | 1835 | |
2fb9b26a | 1836 | |
1837 | if (atomic64_read(&vmem_alloc_used) != 0) | |
1838 | CWARN("vmem leaked %ld/%ld bytes\n", | |
550f1705 | 1839 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); |
2fb9b26a | 1840 | |
ff449ac4 | 1841 | spl_kmem_fini_tracking(&kmem_list, &kmem_lock); |
1842 | spl_kmem_fini_tracking(&vmem_list, &vmem_lock); | |
1843 | #endif /* DEBUG_KMEM */ | |
2fb9b26a | 1844 | ENTRY; |
1845 | ||
1846 | #ifdef HAVE_SET_SHRINKER | |
1847 | remove_shrinker(spl_kmem_cache_shrinker); | |
1848 | #else | |
1849 | unregister_shrinker(&spl_kmem_cache_shrinker); | |
5d86345d | 1850 | #endif |
2fb9b26a | 1851 | |
937879f1 | 1852 | EXIT; |
5d86345d | 1853 | } |