]>
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 | ||
d1ff2312 | 82 | #ifndef HAVE_GET_VMALLOC_INFO |
96dded38 | 83 | get_vmalloc_info_t get_vmalloc_info_fn = SYMBOL_POISON; |
d1ff2312 BB |
84 | EXPORT_SYMBOL(get_vmalloc_info_fn); |
85 | #endif /* HAVE_GET_VMALLOC_INFO */ | |
86 | ||
5232d256 BB |
87 | #ifdef HAVE_PGDAT_HELPERS |
88 | # ifndef HAVE_FIRST_ONLINE_PGDAT | |
96dded38 | 89 | first_online_pgdat_t first_online_pgdat_fn = SYMBOL_POISON; |
d1ff2312 | 90 | EXPORT_SYMBOL(first_online_pgdat_fn); |
5232d256 | 91 | # endif /* HAVE_FIRST_ONLINE_PGDAT */ |
36b313da | 92 | |
5232d256 | 93 | # ifndef HAVE_NEXT_ONLINE_PGDAT |
96dded38 | 94 | next_online_pgdat_t next_online_pgdat_fn = SYMBOL_POISON; |
d1ff2312 | 95 | EXPORT_SYMBOL(next_online_pgdat_fn); |
5232d256 | 96 | # endif /* HAVE_NEXT_ONLINE_PGDAT */ |
36b313da | 97 | |
5232d256 | 98 | # ifndef HAVE_NEXT_ZONE |
96dded38 | 99 | next_zone_t next_zone_fn = SYMBOL_POISON; |
d1ff2312 | 100 | EXPORT_SYMBOL(next_zone_fn); |
5232d256 BB |
101 | # endif /* HAVE_NEXT_ZONE */ |
102 | ||
103 | #else /* HAVE_PGDAT_HELPERS */ | |
104 | ||
105 | # ifndef HAVE_PGDAT_LIST | |
106 | struct pglist_data *pgdat_list_addr = SYMBOL_POISON; | |
107 | EXPORT_SYMBOL(pgdat_list_addr); | |
108 | # endif /* HAVE_PGDAT_LIST */ | |
109 | ||
110 | #endif /* HAVE_PGDAT_HELPERS */ | |
36b313da | 111 | |
6ae7fef5 | 112 | #ifdef NEED_GET_ZONE_COUNTS |
e11d6c5f | 113 | # ifndef HAVE_GET_ZONE_COUNTS |
96dded38 | 114 | get_zone_counts_t get_zone_counts_fn = SYMBOL_POISON; |
d1ff2312 | 115 | EXPORT_SYMBOL(get_zone_counts_fn); |
96dded38 | 116 | # endif /* HAVE_GET_ZONE_COUNTS */ |
4ab13d3b | 117 | |
e11d6c5f | 118 | unsigned long |
6ae7fef5 | 119 | spl_global_page_state(spl_zone_stat_item_t item) |
4ab13d3b BB |
120 | { |
121 | unsigned long active; | |
122 | unsigned long inactive; | |
123 | unsigned long free; | |
124 | ||
6ae7fef5 BB |
125 | get_zone_counts(&active, &inactive, &free); |
126 | switch (item) { | |
127 | case SPL_NR_FREE_PAGES: return free; | |
128 | case SPL_NR_INACTIVE: return inactive; | |
129 | case SPL_NR_ACTIVE: return active; | |
130 | default: ASSERT(0); /* Unsupported */ | |
e11d6c5f BB |
131 | } |
132 | ||
6ae7fef5 BB |
133 | return 0; |
134 | } | |
135 | #else | |
136 | # ifdef HAVE_GLOBAL_PAGE_STATE | |
137 | unsigned long | |
138 | spl_global_page_state(spl_zone_stat_item_t item) | |
139 | { | |
140 | unsigned long pages = 0; | |
141 | ||
142 | switch (item) { | |
143 | case SPL_NR_FREE_PAGES: | |
144 | # ifdef HAVE_ZONE_STAT_ITEM_NR_FREE_PAGES | |
145 | pages += global_page_state(NR_FREE_PAGES); | |
146 | # endif | |
147 | break; | |
148 | case SPL_NR_INACTIVE: | |
149 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE | |
150 | pages += global_page_state(NR_INACTIVE); | |
151 | # endif | |
152 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_ANON | |
153 | pages += global_page_state(NR_INACTIVE_ANON); | |
154 | # endif | |
155 | # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_FILE | |
156 | pages += global_page_state(NR_INACTIVE_FILE); | |
157 | # endif | |
158 | break; | |
159 | case SPL_NR_ACTIVE: | |
160 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE | |
161 | pages += global_page_state(NR_ACTIVE); | |
162 | # endif | |
163 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_ANON | |
164 | pages += global_page_state(NR_ACTIVE_ANON); | |
165 | # endif | |
166 | # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_FILE | |
167 | pages += global_page_state(NR_ACTIVE_FILE); | |
168 | # endif | |
169 | break; | |
170 | default: | |
171 | ASSERT(0); /* Unsupported */ | |
e11d6c5f BB |
172 | } |
173 | ||
6ae7fef5 BB |
174 | return pages; |
175 | } | |
96dded38 | 176 | # else |
6ae7fef5 | 177 | # error "Both global_page_state() and get_zone_counts() unavailable" |
96dded38 | 178 | # endif /* HAVE_GLOBAL_PAGE_STATE */ |
6ae7fef5 | 179 | #endif /* NEED_GET_ZONE_COUNTS */ |
e11d6c5f | 180 | EXPORT_SYMBOL(spl_global_page_state); |
4ab13d3b | 181 | |
e11d6c5f BB |
182 | pgcnt_t |
183 | spl_kmem_availrmem(void) | |
184 | { | |
4ab13d3b | 185 | /* The amount of easily available memory */ |
6ae7fef5 BB |
186 | return (spl_global_page_state(SPL_NR_FREE_PAGES) + |
187 | spl_global_page_state(SPL_NR_INACTIVE)); | |
4ab13d3b BB |
188 | } |
189 | EXPORT_SYMBOL(spl_kmem_availrmem); | |
190 | ||
191 | size_t | |
192 | vmem_size(vmem_t *vmp, int typemask) | |
193 | { | |
d1ff2312 BB |
194 | struct vmalloc_info vmi; |
195 | size_t size = 0; | |
196 | ||
4ab13d3b BB |
197 | ASSERT(vmp == NULL); |
198 | ASSERT(typemask & (VMEM_ALLOC | VMEM_FREE)); | |
199 | ||
d1ff2312 BB |
200 | get_vmalloc_info(&vmi); |
201 | if (typemask & VMEM_ALLOC) | |
202 | size += (size_t)vmi.used; | |
203 | ||
204 | if (typemask & VMEM_FREE) | |
205 | size += (size_t)(VMALLOC_TOTAL - vmi.used); | |
206 | ||
207 | return size; | |
4ab13d3b BB |
208 | } |
209 | EXPORT_SYMBOL(vmem_size); | |
4ab13d3b | 210 | |
f1ca4da6 | 211 | /* |
2fb9b26a | 212 | * Memory allocation interfaces and debugging for basic kmem_* |
055ffd98 BB |
213 | * and vmem_* style memory allocation. When DEBUG_KMEM is enabled |
214 | * the SPL will keep track of the total memory allocated, and | |
215 | * report any memory leaked when the module is unloaded. | |
f1ca4da6 | 216 | */ |
217 | #ifdef DEBUG_KMEM | |
218 | /* Shim layer memory accounting */ | |
550f1705 | 219 | atomic64_t kmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 220 | unsigned long long kmem_alloc_max = 0; |
550f1705 | 221 | atomic64_t vmem_alloc_used = ATOMIC64_INIT(0); |
a0f6da3d | 222 | unsigned long long vmem_alloc_max = 0; |
c19c06f3 | 223 | int kmem_warning_flag = 1; |
79b31f36 | 224 | |
ff449ac4 | 225 | EXPORT_SYMBOL(kmem_alloc_used); |
226 | EXPORT_SYMBOL(kmem_alloc_max); | |
227 | EXPORT_SYMBOL(vmem_alloc_used); | |
228 | EXPORT_SYMBOL(vmem_alloc_max); | |
229 | EXPORT_SYMBOL(kmem_warning_flag); | |
230 | ||
055ffd98 BB |
231 | /* When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked |
232 | * but also the location of every alloc and free. When the SPL module is | |
233 | * unloaded a list of all leaked addresses and where they were allocated | |
234 | * will be dumped to the console. Enabling this feature has a significant | |
235 | * impact on performance but it makes finding memory leaks straight forward. | |
236 | * | |
237 | * Not surprisingly with debugging enabled the xmem_locks are very highly | |
238 | * contended particularly on xfree(). If we want to run with this detailed | |
239 | * debugging enabled for anything other than debugging we need to minimize | |
240 | * the contention by moving to a lock per xmem_table entry model. | |
a0f6da3d | 241 | */ |
055ffd98 | 242 | # ifdef DEBUG_KMEM_TRACKING |
a0f6da3d | 243 | |
244 | # define KMEM_HASH_BITS 10 | |
245 | # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS) | |
246 | ||
247 | # define VMEM_HASH_BITS 10 | |
248 | # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS) | |
249 | ||
250 | typedef struct kmem_debug { | |
251 | struct hlist_node kd_hlist; /* Hash node linkage */ | |
252 | struct list_head kd_list; /* List of all allocations */ | |
253 | void *kd_addr; /* Allocation pointer */ | |
254 | size_t kd_size; /* Allocation size */ | |
255 | const char *kd_func; /* Allocation function */ | |
256 | int kd_line; /* Allocation line */ | |
257 | } kmem_debug_t; | |
258 | ||
d6a26c6a | 259 | spinlock_t kmem_lock; |
260 | struct hlist_head kmem_table[KMEM_TABLE_SIZE]; | |
261 | struct list_head kmem_list; | |
262 | ||
13cdca65 | 263 | spinlock_t vmem_lock; |
264 | struct hlist_head vmem_table[VMEM_TABLE_SIZE]; | |
265 | struct list_head vmem_list; | |
266 | ||
d6a26c6a | 267 | EXPORT_SYMBOL(kmem_lock); |
268 | EXPORT_SYMBOL(kmem_table); | |
269 | EXPORT_SYMBOL(kmem_list); | |
270 | ||
13cdca65 | 271 | EXPORT_SYMBOL(vmem_lock); |
272 | EXPORT_SYMBOL(vmem_table); | |
273 | EXPORT_SYMBOL(vmem_list); | |
a0f6da3d | 274 | # endif |
13cdca65 | 275 | |
c19c06f3 | 276 | int kmem_set_warning(int flag) { return (kmem_warning_flag = !!flag); } |
277 | #else | |
278 | int kmem_set_warning(int flag) { return 0; } | |
f1ca4da6 | 279 | #endif |
c19c06f3 | 280 | EXPORT_SYMBOL(kmem_set_warning); |
f1ca4da6 | 281 | |
282 | /* | |
283 | * Slab allocation interfaces | |
284 | * | |
2fb9b26a | 285 | * While the Linux slab implementation was inspired by the Solaris |
286 | * implemenation I cannot use it to emulate the Solaris APIs. I | |
287 | * require two features which are not provided by the Linux slab. | |
288 | * | |
289 | * 1) Constructors AND destructors. Recent versions of the Linux | |
290 | * kernel have removed support for destructors. This is a deal | |
291 | * breaker for the SPL which contains particularly expensive | |
292 | * initializers for mutex's, condition variables, etc. We also | |
a0f6da3d | 293 | * require a minimal level of cleanup for these data types unlike |
294 | * many Linux data type which do need to be explicitly destroyed. | |
2fb9b26a | 295 | * |
a0f6da3d | 296 | * 2) Virtual address space backed slab. Callers of the Solaris slab |
2fb9b26a | 297 | * expect it to work well for both small are very large allocations. |
298 | * Because of memory fragmentation the Linux slab which is backed | |
299 | * by kmalloc'ed memory performs very badly when confronted with | |
300 | * large numbers of large allocations. Basing the slab on the | |
301 | * virtual address space removes the need for contigeous pages | |
302 | * and greatly improve performance for large allocations. | |
303 | * | |
304 | * For these reasons, the SPL has its own slab implementation with | |
305 | * the needed features. It is not as highly optimized as either the | |
306 | * Solaris or Linux slabs, but it should get me most of what is | |
307 | * needed until it can be optimized or obsoleted by another approach. | |
308 | * | |
309 | * One serious concern I do have about this method is the relatively | |
310 | * small virtual address space on 32bit arches. This will seriously | |
311 | * constrain the size of the slab caches and their performance. | |
312 | * | |
2fb9b26a | 313 | * XXX: Improve the partial slab list by carefully maintaining a |
314 | * strict ordering of fullest to emptiest slabs based on | |
315 | * the slab reference count. This gaurentees the when freeing | |
316 | * slabs back to the system we need only linearly traverse the | |
317 | * last N slabs in the list to discover all the freeable slabs. | |
318 | * | |
319 | * XXX: NUMA awareness for optionally allocating memory close to a | |
320 | * particular core. This can be adventageous if you know the slab | |
321 | * object will be short lived and primarily accessed from one core. | |
322 | * | |
323 | * XXX: Slab coloring may also yield performance improvements and would | |
324 | * be desirable to implement. | |
f1ca4da6 | 325 | */ |
2fb9b26a | 326 | |
a0f6da3d | 327 | struct list_head spl_kmem_cache_list; /* List of caches */ |
328 | struct rw_semaphore spl_kmem_cache_sem; /* Cache list lock */ | |
c30df9c8 | 329 | |
4afaaefa | 330 | static int spl_cache_flush(spl_kmem_cache_t *skc, |
a0f6da3d | 331 | spl_kmem_magazine_t *skm, int flush); |
4afaaefa | 332 | |
57d86234 | 333 | #ifdef HAVE_SET_SHRINKER |
2fb9b26a | 334 | static struct shrinker *spl_kmem_cache_shrinker; |
57d86234 | 335 | #else |
4afaaefa | 336 | static int spl_kmem_cache_generic_shrinker(int nr_to_scan, |
a0f6da3d | 337 | unsigned int gfp_mask); |
2fb9b26a | 338 | static struct shrinker spl_kmem_cache_shrinker = { |
4afaaefa | 339 | .shrink = spl_kmem_cache_generic_shrinker, |
57d86234 | 340 | .seeks = KMC_DEFAULT_SEEKS, |
341 | }; | |
342 | #endif | |
f1ca4da6 | 343 | |
a0f6da3d | 344 | #ifdef DEBUG_KMEM |
345 | # ifdef DEBUG_KMEM_TRACKING | |
346 | ||
347 | static kmem_debug_t * | |
348 | kmem_del_init(spinlock_t *lock, struct hlist_head *table, int bits, | |
349 | void *addr) | |
350 | { | |
351 | struct hlist_head *head; | |
352 | struct hlist_node *node; | |
353 | struct kmem_debug *p; | |
354 | unsigned long flags; | |
355 | ENTRY; | |
356 | ||
357 | spin_lock_irqsave(lock, flags); | |
358 | ||
359 | head = &table[hash_ptr(addr, bits)]; | |
360 | hlist_for_each_entry_rcu(p, node, head, kd_hlist) { | |
361 | if (p->kd_addr == addr) { | |
362 | hlist_del_init(&p->kd_hlist); | |
363 | list_del_init(&p->kd_list); | |
364 | spin_unlock_irqrestore(lock, flags); | |
365 | return p; | |
366 | } | |
367 | } | |
368 | ||
369 | spin_unlock_irqrestore(lock, flags); | |
370 | ||
371 | RETURN(NULL); | |
372 | } | |
373 | ||
374 | void * | |
375 | kmem_alloc_track(size_t size, int flags, const char *func, int line, | |
376 | int node_alloc, int node) | |
377 | { | |
378 | void *ptr = NULL; | |
379 | kmem_debug_t *dptr; | |
380 | unsigned long irq_flags; | |
381 | ENTRY; | |
382 | ||
383 | dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t), | |
384 | flags & ~__GFP_ZERO); | |
385 | ||
386 | if (dptr == NULL) { | |
387 | CWARN("kmem_alloc(%ld, 0x%x) debug failed\n", | |
388 | sizeof(kmem_debug_t), flags); | |
389 | } else { | |
390 | /* Marked unlikely because we should never be doing this, | |
391 | * we tolerate to up 2 pages but a single page is best. */ | |
392 | if (unlikely((size) > (PAGE_SIZE * 2)) && kmem_warning_flag) | |
393 | CWARN("Large kmem_alloc(%llu, 0x%x) (%lld/%llu)\n", | |
394 | (unsigned long long) size, flags, | |
395 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
396 | ||
c8e60837 | 397 | /* We use kstrdup() below because the string pointed to by |
398 | * __FUNCTION__ might not be available by the time we want | |
399 | * to print it since the module might have been unloaded. */ | |
400 | dptr->kd_func = kstrdup(func, flags & ~__GFP_ZERO); | |
401 | if (unlikely(dptr->kd_func == NULL)) { | |
402 | kfree(dptr); | |
403 | CWARN("kstrdup() failed in kmem_alloc(%llu, 0x%x) " | |
404 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
405 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
406 | goto out; | |
407 | } | |
408 | ||
a0f6da3d | 409 | /* Use the correct allocator */ |
410 | if (node_alloc) { | |
411 | ASSERT(!(flags & __GFP_ZERO)); | |
412 | ptr = kmalloc_node(size, flags, node); | |
413 | } else if (flags & __GFP_ZERO) { | |
414 | ptr = kzalloc(size, flags & ~__GFP_ZERO); | |
415 | } else { | |
416 | ptr = kmalloc(size, flags); | |
417 | } | |
418 | ||
419 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 420 | kfree(dptr->kd_func); |
a0f6da3d | 421 | kfree(dptr); |
422 | CWARN("kmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
423 | (unsigned long long) size, flags, | |
424 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
425 | goto out; | |
426 | } | |
427 | ||
428 | atomic64_add(size, &kmem_alloc_used); | |
429 | if (unlikely(atomic64_read(&kmem_alloc_used) > | |
430 | kmem_alloc_max)) | |
431 | kmem_alloc_max = | |
432 | atomic64_read(&kmem_alloc_used); | |
433 | ||
434 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
435 | INIT_LIST_HEAD(&dptr->kd_list); | |
436 | ||
437 | dptr->kd_addr = ptr; | |
438 | dptr->kd_size = size; | |
a0f6da3d | 439 | dptr->kd_line = line; |
440 | ||
441 | spin_lock_irqsave(&kmem_lock, irq_flags); | |
442 | hlist_add_head_rcu(&dptr->kd_hlist, | |
443 | &kmem_table[hash_ptr(ptr, KMEM_HASH_BITS)]); | |
444 | list_add_tail(&dptr->kd_list, &kmem_list); | |
445 | spin_unlock_irqrestore(&kmem_lock, irq_flags); | |
446 | ||
447 | CDEBUG_LIMIT(D_INFO, "kmem_alloc(%llu, 0x%x) = %p " | |
448 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
449 | ptr, atomic64_read(&kmem_alloc_used), | |
450 | kmem_alloc_max); | |
451 | } | |
452 | out: | |
453 | RETURN(ptr); | |
454 | } | |
455 | EXPORT_SYMBOL(kmem_alloc_track); | |
456 | ||
457 | void | |
458 | kmem_free_track(void *ptr, size_t size) | |
459 | { | |
460 | kmem_debug_t *dptr; | |
461 | ENTRY; | |
462 | ||
463 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
464 | (unsigned long long) size); | |
465 | ||
466 | dptr = kmem_del_init(&kmem_lock, kmem_table, KMEM_HASH_BITS, ptr); | |
467 | ||
468 | ASSERT(dptr); /* Must exist in hash due to kmem_alloc() */ | |
469 | ||
470 | /* Size must match */ | |
471 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
472 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
473 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
474 | ||
475 | atomic64_sub(size, &kmem_alloc_used); | |
476 | ||
477 | CDEBUG_LIMIT(D_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
478 | (unsigned long long) size, atomic64_read(&kmem_alloc_used), | |
479 | kmem_alloc_max); | |
480 | ||
c8e60837 | 481 | kfree(dptr->kd_func); |
482 | ||
a0f6da3d | 483 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
484 | kfree(dptr); | |
485 | ||
486 | memset(ptr, 0x5a, size); | |
487 | kfree(ptr); | |
488 | ||
489 | EXIT; | |
490 | } | |
491 | EXPORT_SYMBOL(kmem_free_track); | |
492 | ||
493 | void * | |
494 | vmem_alloc_track(size_t size, int flags, const char *func, int line) | |
495 | { | |
496 | void *ptr = NULL; | |
497 | kmem_debug_t *dptr; | |
498 | unsigned long irq_flags; | |
499 | ENTRY; | |
500 | ||
501 | ASSERT(flags & KM_SLEEP); | |
502 | ||
503 | dptr = (kmem_debug_t *) kmalloc(sizeof(kmem_debug_t), flags); | |
504 | if (dptr == NULL) { | |
505 | CWARN("vmem_alloc(%ld, 0x%x) debug failed\n", | |
506 | sizeof(kmem_debug_t), flags); | |
507 | } else { | |
c8e60837 | 508 | /* We use kstrdup() below because the string pointed to by |
509 | * __FUNCTION__ might not be available by the time we want | |
510 | * to print it, since the module might have been unloaded. */ | |
511 | dptr->kd_func = kstrdup(func, flags & ~__GFP_ZERO); | |
512 | if (unlikely(dptr->kd_func == NULL)) { | |
513 | kfree(dptr); | |
514 | CWARN("kstrdup() failed in vmem_alloc(%llu, 0x%x) " | |
515 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
516 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
517 | goto out; | |
518 | } | |
519 | ||
a0f6da3d | 520 | ptr = __vmalloc(size, (flags | __GFP_HIGHMEM) & ~__GFP_ZERO, |
521 | PAGE_KERNEL); | |
522 | ||
523 | if (unlikely(ptr == NULL)) { | |
c8e60837 | 524 | kfree(dptr->kd_func); |
a0f6da3d | 525 | kfree(dptr); |
526 | CWARN("vmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
527 | (unsigned long long) size, flags, | |
528 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
529 | goto out; | |
530 | } | |
531 | ||
532 | if (flags & __GFP_ZERO) | |
533 | memset(ptr, 0, size); | |
534 | ||
535 | atomic64_add(size, &vmem_alloc_used); | |
536 | if (unlikely(atomic64_read(&vmem_alloc_used) > | |
537 | vmem_alloc_max)) | |
538 | vmem_alloc_max = | |
539 | atomic64_read(&vmem_alloc_used); | |
540 | ||
541 | INIT_HLIST_NODE(&dptr->kd_hlist); | |
542 | INIT_LIST_HEAD(&dptr->kd_list); | |
543 | ||
544 | dptr->kd_addr = ptr; | |
545 | dptr->kd_size = size; | |
a0f6da3d | 546 | dptr->kd_line = line; |
547 | ||
548 | spin_lock_irqsave(&vmem_lock, irq_flags); | |
549 | hlist_add_head_rcu(&dptr->kd_hlist, | |
550 | &vmem_table[hash_ptr(ptr, VMEM_HASH_BITS)]); | |
551 | list_add_tail(&dptr->kd_list, &vmem_list); | |
552 | spin_unlock_irqrestore(&vmem_lock, irq_flags); | |
553 | ||
554 | CDEBUG_LIMIT(D_INFO, "vmem_alloc(%llu, 0x%x) = %p " | |
555 | "(%lld/%llu)\n", (unsigned long long) size, flags, | |
556 | ptr, atomic64_read(&vmem_alloc_used), | |
557 | vmem_alloc_max); | |
558 | } | |
559 | out: | |
560 | RETURN(ptr); | |
561 | } | |
562 | EXPORT_SYMBOL(vmem_alloc_track); | |
563 | ||
564 | void | |
565 | vmem_free_track(void *ptr, size_t size) | |
566 | { | |
567 | kmem_debug_t *dptr; | |
568 | ENTRY; | |
569 | ||
570 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
571 | (unsigned long long) size); | |
572 | ||
573 | dptr = kmem_del_init(&vmem_lock, vmem_table, VMEM_HASH_BITS, ptr); | |
574 | ASSERT(dptr); /* Must exist in hash due to vmem_alloc() */ | |
575 | ||
576 | /* Size must match */ | |
577 | ASSERTF(dptr->kd_size == size, "kd_size (%llu) != size (%llu), " | |
578 | "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr->kd_size, | |
579 | (unsigned long long) size, dptr->kd_func, dptr->kd_line); | |
580 | ||
581 | atomic64_sub(size, &vmem_alloc_used); | |
582 | CDEBUG_LIMIT(D_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
583 | (unsigned long long) size, atomic64_read(&vmem_alloc_used), | |
584 | vmem_alloc_max); | |
585 | ||
c8e60837 | 586 | kfree(dptr->kd_func); |
587 | ||
a0f6da3d | 588 | memset(dptr, 0x5a, sizeof(kmem_debug_t)); |
589 | kfree(dptr); | |
590 | ||
591 | memset(ptr, 0x5a, size); | |
592 | vfree(ptr); | |
593 | ||
594 | EXIT; | |
595 | } | |
596 | EXPORT_SYMBOL(vmem_free_track); | |
597 | ||
598 | # else /* DEBUG_KMEM_TRACKING */ | |
599 | ||
600 | void * | |
601 | kmem_alloc_debug(size_t size, int flags, const char *func, int line, | |
602 | int node_alloc, int node) | |
603 | { | |
604 | void *ptr; | |
605 | ENTRY; | |
606 | ||
607 | /* Marked unlikely because we should never be doing this, | |
608 | * we tolerate to up 2 pages but a single page is best. */ | |
609 | if (unlikely(size > (PAGE_SIZE * 2)) && kmem_warning_flag) | |
610 | CWARN("Large kmem_alloc(%llu, 0x%x) (%lld/%llu)\n", | |
611 | (unsigned long long) size, flags, | |
612 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
613 | ||
614 | /* Use the correct allocator */ | |
615 | if (node_alloc) { | |
616 | ASSERT(!(flags & __GFP_ZERO)); | |
617 | ptr = kmalloc_node(size, flags, node); | |
618 | } else if (flags & __GFP_ZERO) { | |
619 | ptr = kzalloc(size, flags & (~__GFP_ZERO)); | |
620 | } else { | |
621 | ptr = kmalloc(size, flags); | |
622 | } | |
623 | ||
624 | if (ptr == NULL) { | |
625 | CWARN("kmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
626 | (unsigned long long) size, flags, | |
627 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
628 | } else { | |
629 | atomic64_add(size, &kmem_alloc_used); | |
630 | if (unlikely(atomic64_read(&kmem_alloc_used) > kmem_alloc_max)) | |
631 | kmem_alloc_max = atomic64_read(&kmem_alloc_used); | |
632 | ||
633 | CDEBUG_LIMIT(D_INFO, "kmem_alloc(%llu, 0x%x) = %p " | |
634 | "(%lld/%llu)\n", (unsigned long long) size, flags, ptr, | |
635 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); | |
636 | } | |
637 | RETURN(ptr); | |
638 | } | |
639 | EXPORT_SYMBOL(kmem_alloc_debug); | |
640 | ||
641 | void | |
642 | kmem_free_debug(void *ptr, size_t size) | |
643 | { | |
644 | ENTRY; | |
645 | ||
646 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
647 | (unsigned long long) size); | |
648 | ||
649 | atomic64_sub(size, &kmem_alloc_used); | |
650 | ||
651 | CDEBUG_LIMIT(D_INFO, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
652 | (unsigned long long) size, atomic64_read(&kmem_alloc_used), | |
653 | kmem_alloc_max); | |
654 | ||
655 | memset(ptr, 0x5a, size); | |
656 | kfree(ptr); | |
657 | ||
658 | EXIT; | |
659 | } | |
660 | EXPORT_SYMBOL(kmem_free_debug); | |
661 | ||
662 | void * | |
663 | vmem_alloc_debug(size_t size, int flags, const char *func, int line) | |
664 | { | |
665 | void *ptr; | |
666 | ENTRY; | |
667 | ||
668 | ASSERT(flags & KM_SLEEP); | |
669 | ||
670 | ptr = __vmalloc(size, (flags | __GFP_HIGHMEM) & ~__GFP_ZERO, | |
671 | PAGE_KERNEL); | |
672 | if (ptr == NULL) { | |
673 | CWARN("vmem_alloc(%llu, 0x%x) failed (%lld/%llu)\n", | |
674 | (unsigned long long) size, flags, | |
675 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
676 | } else { | |
677 | if (flags & __GFP_ZERO) | |
678 | memset(ptr, 0, size); | |
679 | ||
680 | atomic64_add(size, &vmem_alloc_used); | |
681 | ||
682 | if (unlikely(atomic64_read(&vmem_alloc_used) > vmem_alloc_max)) | |
683 | vmem_alloc_max = atomic64_read(&vmem_alloc_used); | |
684 | ||
685 | CDEBUG_LIMIT(D_INFO, "vmem_alloc(%llu, 0x%x) = %p " | |
686 | "(%lld/%llu)\n", (unsigned long long) size, flags, ptr, | |
687 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); | |
688 | } | |
689 | ||
690 | RETURN(ptr); | |
691 | } | |
692 | EXPORT_SYMBOL(vmem_alloc_debug); | |
693 | ||
694 | void | |
695 | vmem_free_debug(void *ptr, size_t size) | |
696 | { | |
697 | ENTRY; | |
698 | ||
699 | ASSERTF(ptr || size > 0, "ptr: %p, size: %llu", ptr, | |
700 | (unsigned long long) size); | |
701 | ||
702 | atomic64_sub(size, &vmem_alloc_used); | |
703 | ||
704 | CDEBUG_LIMIT(D_INFO, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr, | |
705 | (unsigned long long) size, atomic64_read(&vmem_alloc_used), | |
706 | vmem_alloc_max); | |
707 | ||
708 | memset(ptr, 0x5a, size); | |
709 | vfree(ptr); | |
710 | ||
711 | EXIT; | |
712 | } | |
713 | EXPORT_SYMBOL(vmem_free_debug); | |
714 | ||
715 | # endif /* DEBUG_KMEM_TRACKING */ | |
716 | #endif /* DEBUG_KMEM */ | |
717 | ||
a1502d76 | 718 | static void * |
719 | kv_alloc(spl_kmem_cache_t *skc, int size, int flags) | |
fece7c99 | 720 | { |
a1502d76 | 721 | void *ptr; |
f1ca4da6 | 722 | |
a1502d76 | 723 | if (skc->skc_flags & KMC_KMEM) { |
724 | if (size > (2 * PAGE_SIZE)) { | |
725 | ptr = (void *)__get_free_pages(flags, get_order(size)); | |
726 | } else | |
727 | ptr = kmem_alloc(size, flags); | |
728 | } else { | |
729 | ptr = vmem_alloc(size, flags); | |
d6a26c6a | 730 | } |
fece7c99 | 731 | |
a1502d76 | 732 | return ptr; |
733 | } | |
fece7c99 | 734 | |
a1502d76 | 735 | static void |
736 | kv_free(spl_kmem_cache_t *skc, void *ptr, int size) | |
737 | { | |
738 | if (skc->skc_flags & KMC_KMEM) { | |
739 | if (size > (2 * PAGE_SIZE)) | |
740 | free_pages((unsigned long)ptr, get_order(size)); | |
741 | else | |
742 | kmem_free(ptr, size); | |
743 | } else { | |
744 | vmem_free(ptr, size); | |
745 | } | |
fece7c99 | 746 | } |
747 | ||
ea3e6ca9 BB |
748 | /* |
749 | * It's important that we pack the spl_kmem_obj_t structure and the | |
48e0606a BB |
750 | * actual objects in to one large address space to minimize the number |
751 | * of calls to the allocator. It is far better to do a few large | |
752 | * allocations and then subdivide it ourselves. Now which allocator | |
753 | * we use requires balancing a few trade offs. | |
754 | * | |
755 | * For small objects we use kmem_alloc() because as long as you are | |
756 | * only requesting a small number of pages (ideally just one) its cheap. | |
757 | * However, when you start requesting multiple pages with kmem_alloc() | |
758 | * it gets increasingly expensive since it requires contigeous pages. | |
759 | * For this reason we shift to vmem_alloc() for slabs of large objects | |
760 | * which removes the need for contigeous pages. We do not use | |
761 | * vmem_alloc() in all cases because there is significant locking | |
762 | * overhead in __get_vm_area_node(). This function takes a single | |
763 | * global lock when aquiring an available virtual address range which | |
764 | * serializes all vmem_alloc()'s for all slab caches. Using slightly | |
765 | * different allocation functions for small and large objects should | |
766 | * give us the best of both worlds. | |
767 | * | |
768 | * KMC_ONSLAB KMC_OFFSLAB | |
769 | * | |
770 | * +------------------------+ +-----------------+ | |
771 | * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+ | |
772 | * | skc_obj_size <-+ | | +-----------------+ | | | |
773 | * | spl_kmem_obj_t | | | | | |
774 | * | skc_obj_size <---+ | +-----------------+ | | | |
775 | * | spl_kmem_obj_t | | | skc_obj_size | <-+ | | |
776 | * | ... v | | spl_kmem_obj_t | | | |
777 | * +------------------------+ +-----------------+ v | |
778 | */ | |
fece7c99 | 779 | static spl_kmem_slab_t * |
a1502d76 | 780 | spl_slab_alloc(spl_kmem_cache_t *skc, int flags) |
fece7c99 | 781 | { |
782 | spl_kmem_slab_t *sks; | |
a1502d76 | 783 | spl_kmem_obj_t *sko, *n; |
784 | void *base, *obj; | |
48e0606a BB |
785 | int i, align, size, rc = 0; |
786 | ||
a1502d76 | 787 | base = kv_alloc(skc, skc->skc_slab_size, flags); |
788 | if (base == NULL) | |
fece7c99 | 789 | RETURN(NULL); |
790 | ||
a1502d76 | 791 | sks = (spl_kmem_slab_t *)base; |
792 | sks->sks_magic = SKS_MAGIC; | |
793 | sks->sks_objs = skc->skc_slab_objs; | |
794 | sks->sks_age = jiffies; | |
795 | sks->sks_cache = skc; | |
796 | INIT_LIST_HEAD(&sks->sks_list); | |
797 | INIT_LIST_HEAD(&sks->sks_free_list); | |
798 | sks->sks_ref = 0; | |
48e0606a BB |
799 | |
800 | align = skc->skc_obj_align; | |
801 | size = P2ROUNDUP(skc->skc_obj_size, align) + | |
802 | P2ROUNDUP(sizeof(spl_kmem_obj_t), align); | |
fece7c99 | 803 | |
804 | for (i = 0; i < sks->sks_objs; i++) { | |
a1502d76 | 805 | if (skc->skc_flags & KMC_OFFSLAB) { |
806 | obj = kv_alloc(skc, size, flags); | |
807 | if (!obj) | |
808 | GOTO(out, rc = -ENOMEM); | |
809 | } else { | |
48e0606a BB |
810 | obj = base + |
811 | P2ROUNDUP(sizeof(spl_kmem_slab_t), align) + | |
812 | (i * size); | |
a1502d76 | 813 | } |
814 | ||
48e0606a | 815 | sko = obj + P2ROUNDUP(skc->skc_obj_size, align); |
fece7c99 | 816 | sko->sko_addr = obj; |
817 | sko->sko_magic = SKO_MAGIC; | |
818 | sko->sko_slab = sks; | |
819 | INIT_LIST_HEAD(&sko->sko_list); | |
fece7c99 | 820 | list_add_tail(&sko->sko_list, &sks->sks_free_list); |
821 | } | |
822 | ||
fece7c99 | 823 | list_for_each_entry(sko, &sks->sks_free_list, sko_list) |
824 | if (skc->skc_ctor) | |
825 | skc->skc_ctor(sko->sko_addr, skc->skc_private, flags); | |
2fb9b26a | 826 | out: |
a1502d76 | 827 | if (rc) { |
828 | if (skc->skc_flags & KMC_OFFSLAB) | |
48e0606a BB |
829 | list_for_each_entry_safe(sko, n, &sks->sks_free_list, |
830 | sko_list) | |
a1502d76 | 831 | kv_free(skc, sko->sko_addr, size); |
fece7c99 | 832 | |
a1502d76 | 833 | kv_free(skc, base, skc->skc_slab_size); |
834 | sks = NULL; | |
fece7c99 | 835 | } |
836 | ||
a1502d76 | 837 | RETURN(sks); |
fece7c99 | 838 | } |
839 | ||
ea3e6ca9 BB |
840 | /* |
841 | * Remove a slab from complete or partial list, it must be called with | |
842 | * the 'skc->skc_lock' held but the actual free must be performed | |
843 | * outside the lock to prevent deadlocking on vmem addresses. | |
fece7c99 | 844 | */ |
f1ca4da6 | 845 | static void |
ea3e6ca9 BB |
846 | spl_slab_free(spl_kmem_slab_t *sks, |
847 | struct list_head *sks_list, struct list_head *sko_list) | |
848 | { | |
2fb9b26a | 849 | spl_kmem_cache_t *skc; |
2fb9b26a | 850 | ENTRY; |
57d86234 | 851 | |
2fb9b26a | 852 | ASSERT(sks->sks_magic == SKS_MAGIC); |
4afaaefa | 853 | ASSERT(sks->sks_ref == 0); |
d6a26c6a | 854 | |
fece7c99 | 855 | skc = sks->sks_cache; |
856 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
d46630e0 | 857 | ASSERT(spin_is_locked(&skc->skc_lock)); |
f1ca4da6 | 858 | |
1a944a7d BB |
859 | /* |
860 | * Update slab/objects counters in the cache, then remove the | |
861 | * slab from the skc->skc_partial_list. Finally add the slab | |
862 | * and all its objects in to the private work lists where the | |
863 | * destructors will be called and the memory freed to the system. | |
864 | */ | |
fece7c99 | 865 | skc->skc_obj_total -= sks->sks_objs; |
866 | skc->skc_slab_total--; | |
867 | list_del(&sks->sks_list); | |
ea3e6ca9 | 868 | list_add(&sks->sks_list, sks_list); |
1a944a7d BB |
869 | list_splice_init(&sks->sks_free_list, sko_list); |
870 | ||
2fb9b26a | 871 | EXIT; |
872 | } | |
d6a26c6a | 873 | |
ea3e6ca9 BB |
874 | /* |
875 | * Traverses all the partial slabs attached to a cache and free those | |
876 | * which which are currently empty, and have not been touched for | |
37db7d8c BB |
877 | * skc_delay seconds to avoid thrashing. The count argument is |
878 | * passed to optionally cap the number of slabs reclaimed, a count | |
879 | * of zero means try and reclaim everything. When flag is set we | |
880 | * always free an available slab regardless of age. | |
ea3e6ca9 BB |
881 | */ |
882 | static void | |
37db7d8c | 883 | spl_slab_reclaim(spl_kmem_cache_t *skc, int count, int flag) |
2fb9b26a | 884 | { |
885 | spl_kmem_slab_t *sks, *m; | |
ea3e6ca9 BB |
886 | spl_kmem_obj_t *sko, *n; |
887 | LIST_HEAD(sks_list); | |
888 | LIST_HEAD(sko_list); | |
1a944a7d | 889 | int size = 0, i = 0; |
2fb9b26a | 890 | ENTRY; |
891 | ||
2fb9b26a | 892 | /* |
ea3e6ca9 BB |
893 | * Move empty slabs and objects which have not been touched in |
894 | * skc_delay seconds on to private lists to be freed outside | |
1a944a7d BB |
895 | * the spin lock. This delay time is important to avoid thrashing |
896 | * however when flag is set the delay will not be used. | |
2fb9b26a | 897 | */ |
ea3e6ca9 | 898 | spin_lock(&skc->skc_lock); |
1a944a7d BB |
899 | list_for_each_entry_safe_reverse(sks,m,&skc->skc_partial_list,sks_list){ |
900 | /* | |
901 | * All empty slabs are at the end of skc->skc_partial_list, | |
902 | * therefore once a non-empty slab is found we can stop | |
903 | * scanning. Additionally, stop when reaching the target | |
904 | * reclaim 'count' if a non-zero threshhold is given. | |
905 | */ | |
906 | if ((sks->sks_ref > 0) || (count && i > count)) | |
37db7d8c BB |
907 | break; |
908 | ||
37db7d8c | 909 | if (time_after(jiffies,sks->sks_age+skc->skc_delay*HZ)||flag) { |
ea3e6ca9 | 910 | spl_slab_free(sks, &sks_list, &sko_list); |
37db7d8c BB |
911 | i++; |
912 | } | |
ea3e6ca9 BB |
913 | } |
914 | spin_unlock(&skc->skc_lock); | |
915 | ||
916 | /* | |
1a944a7d BB |
917 | * The following two loops ensure all the object destructors are |
918 | * run, any offslab objects are freed, and the slabs themselves | |
919 | * are freed. This is all done outside the skc->skc_lock since | |
920 | * this allows the destructor to sleep, and allows us to perform | |
921 | * a conditional reschedule when a freeing a large number of | |
922 | * objects and slabs back to the system. | |
ea3e6ca9 | 923 | */ |
1a944a7d | 924 | if (skc->skc_flags & KMC_OFFSLAB) |
ea3e6ca9 BB |
925 | size = P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align) + |
926 | P2ROUNDUP(sizeof(spl_kmem_obj_t), skc->skc_obj_align); | |
927 | ||
1a944a7d BB |
928 | list_for_each_entry_safe(sko, n, &sko_list, sko_list) { |
929 | ASSERT(sko->sko_magic == SKO_MAGIC); | |
930 | ||
931 | if (skc->skc_dtor) | |
932 | skc->skc_dtor(sko->sko_addr, skc->skc_private); | |
933 | ||
934 | if (skc->skc_flags & KMC_OFFSLAB) | |
ea3e6ca9 | 935 | kv_free(skc, sko->sko_addr, size); |
1a944a7d BB |
936 | |
937 | cond_resched(); | |
2fb9b26a | 938 | } |
939 | ||
37db7d8c | 940 | list_for_each_entry_safe(sks, m, &sks_list, sks_list) { |
1a944a7d | 941 | ASSERT(sks->sks_magic == SKS_MAGIC); |
ea3e6ca9 | 942 | kv_free(skc, sks, skc->skc_slab_size); |
37db7d8c BB |
943 | cond_resched(); |
944 | } | |
ea3e6ca9 BB |
945 | |
946 | EXIT; | |
f1ca4da6 | 947 | } |
948 | ||
ea3e6ca9 BB |
949 | /* |
950 | * Called regularly on all caches to age objects out of the magazines | |
951 | * which have not been access in skc->skc_delay seconds. This prevents | |
952 | * idle magazines from holding memory which might be better used by | |
953 | * other caches or parts of the system. The delay is present to | |
954 | * prevent thrashing the magazine. | |
955 | */ | |
956 | static void | |
957 | spl_magazine_age(void *data) | |
f1ca4da6 | 958 | { |
9b1b8e4c BB |
959 | spl_kmem_magazine_t *skm = |
960 | spl_get_work_data(data, spl_kmem_magazine_t, skm_work.work); | |
961 | spl_kmem_cache_t *skc = skm->skm_cache; | |
962 | int i = smp_processor_id(); | |
963 | ||
964 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
965 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
966 | ASSERT(skc->skc_mag[i] == skm); | |
f1ca4da6 | 967 | |
ea3e6ca9 BB |
968 | if (skm->skm_avail > 0 && |
969 | time_after(jiffies, skm->skm_age + skc->skc_delay * HZ)) | |
970 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
9b1b8e4c BB |
971 | |
972 | if (!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)) | |
973 | schedule_delayed_work_on(i, &skm->skm_work, | |
974 | skc->skc_delay / 3 * HZ); | |
ea3e6ca9 | 975 | } |
4efd4118 | 976 | |
ea3e6ca9 BB |
977 | /* |
978 | * Called regularly to keep a downward pressure on the size of idle | |
979 | * magazines and to release free slabs from the cache. This function | |
980 | * never calls the registered reclaim function, that only occures | |
981 | * under memory pressure or with a direct call to spl_kmem_reap(). | |
982 | */ | |
983 | static void | |
984 | spl_cache_age(void *data) | |
985 | { | |
9b1b8e4c | 986 | spl_kmem_cache_t *skc = |
ea3e6ca9 BB |
987 | spl_get_work_data(data, spl_kmem_cache_t, skc_work.work); |
988 | ||
989 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
37db7d8c | 990 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
ea3e6ca9 BB |
991 | |
992 | if (!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)) | |
37db7d8c | 993 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
2fb9b26a | 994 | } |
f1ca4da6 | 995 | |
ea3e6ca9 BB |
996 | /* |
997 | * Size a slab based on the size of each aliged object plus spl_kmem_obj_t. | |
998 | * When on-slab we want to target SPL_KMEM_CACHE_OBJ_PER_SLAB. However, | |
999 | * for very small objects we may end up with more than this so as not | |
1000 | * to waste space in the minimal allocation of a single page. Also for | |
1001 | * very large objects we may use as few as SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN, | |
1002 | * lower than this and we will fail. | |
1003 | */ | |
48e0606a BB |
1004 | static int |
1005 | spl_slab_size(spl_kmem_cache_t *skc, uint32_t *objs, uint32_t *size) | |
1006 | { | |
ea3e6ca9 | 1007 | int sks_size, obj_size, max_size, align; |
48e0606a BB |
1008 | |
1009 | if (skc->skc_flags & KMC_OFFSLAB) { | |
ea3e6ca9 | 1010 | *objs = SPL_KMEM_CACHE_OBJ_PER_SLAB; |
48e0606a BB |
1011 | *size = sizeof(spl_kmem_slab_t); |
1012 | } else { | |
ea3e6ca9 BB |
1013 | align = skc->skc_obj_align; |
1014 | sks_size = P2ROUNDUP(sizeof(spl_kmem_slab_t), align); | |
1015 | obj_size = P2ROUNDUP(skc->skc_obj_size, align) + | |
1016 | P2ROUNDUP(sizeof(spl_kmem_obj_t), align); | |
1017 | ||
1018 | if (skc->skc_flags & KMC_KMEM) | |
1019 | max_size = ((uint64_t)1 << (MAX_ORDER-1)) * PAGE_SIZE; | |
1020 | else | |
1021 | max_size = (32 * 1024 * 1024); | |
48e0606a | 1022 | |
ea3e6ca9 BB |
1023 | for (*size = PAGE_SIZE; *size <= max_size; *size += PAGE_SIZE) { |
1024 | *objs = (*size - sks_size) / obj_size; | |
1025 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB) | |
1026 | RETURN(0); | |
1027 | } | |
48e0606a | 1028 | |
ea3e6ca9 BB |
1029 | /* |
1030 | * Unable to satisfy target objets per slab, fallback to | |
1031 | * allocating a maximally sized slab and assuming it can | |
1032 | * contain the minimum objects count use it. If not fail. | |
1033 | */ | |
1034 | *size = max_size; | |
1035 | *objs = (*size - sks_size) / obj_size; | |
1036 | if (*objs >= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN) | |
1037 | RETURN(0); | |
48e0606a BB |
1038 | } |
1039 | ||
ea3e6ca9 | 1040 | RETURN(-ENOSPC); |
48e0606a BB |
1041 | } |
1042 | ||
ea3e6ca9 BB |
1043 | /* |
1044 | * Make a guess at reasonable per-cpu magazine size based on the size of | |
1045 | * each object and the cost of caching N of them in each magazine. Long | |
1046 | * term this should really adapt based on an observed usage heuristic. | |
1047 | */ | |
4afaaefa | 1048 | static int |
1049 | spl_magazine_size(spl_kmem_cache_t *skc) | |
1050 | { | |
48e0606a | 1051 | int size, align = skc->skc_obj_align; |
4afaaefa | 1052 | ENTRY; |
1053 | ||
ea3e6ca9 | 1054 | /* Per-magazine sizes below assume a 4Kib page size */ |
48e0606a | 1055 | if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE * 256)) |
ea3e6ca9 | 1056 | size = 4; /* Minimum 4Mib per-magazine */ |
48e0606a | 1057 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE * 32)) |
ea3e6ca9 | 1058 | size = 16; /* Minimum 2Mib per-magazine */ |
48e0606a | 1059 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE)) |
ea3e6ca9 | 1060 | size = 64; /* Minimum 256Kib per-magazine */ |
48e0606a | 1061 | else if (P2ROUNDUP(skc->skc_obj_size, align) > (PAGE_SIZE / 4)) |
ea3e6ca9 | 1062 | size = 128; /* Minimum 128Kib per-magazine */ |
4afaaefa | 1063 | else |
ea3e6ca9 | 1064 | size = 256; |
4afaaefa | 1065 | |
1066 | RETURN(size); | |
1067 | } | |
1068 | ||
ea3e6ca9 BB |
1069 | /* |
1070 | * Allocate a per-cpu magazine to assoicate with a specific core. | |
1071 | */ | |
4afaaefa | 1072 | static spl_kmem_magazine_t * |
1073 | spl_magazine_alloc(spl_kmem_cache_t *skc, int node) | |
1074 | { | |
1075 | spl_kmem_magazine_t *skm; | |
1076 | int size = sizeof(spl_kmem_magazine_t) + | |
1077 | sizeof(void *) * skc->skc_mag_size; | |
1078 | ENTRY; | |
1079 | ||
ea3e6ca9 | 1080 | skm = kmem_alloc_node(size, GFP_KERNEL | __GFP_NOFAIL, node); |
4afaaefa | 1081 | if (skm) { |
1082 | skm->skm_magic = SKM_MAGIC; | |
1083 | skm->skm_avail = 0; | |
1084 | skm->skm_size = skc->skc_mag_size; | |
1085 | skm->skm_refill = skc->skc_mag_refill; | |
9b1b8e4c BB |
1086 | skm->skm_cache = skc; |
1087 | spl_init_delayed_work(&skm->skm_work, spl_magazine_age, skm); | |
ea3e6ca9 | 1088 | skm->skm_age = jiffies; |
4afaaefa | 1089 | } |
1090 | ||
1091 | RETURN(skm); | |
1092 | } | |
1093 | ||
ea3e6ca9 BB |
1094 | /* |
1095 | * Free a per-cpu magazine assoicated with a specific core. | |
1096 | */ | |
4afaaefa | 1097 | static void |
1098 | spl_magazine_free(spl_kmem_magazine_t *skm) | |
1099 | { | |
a0f6da3d | 1100 | int size = sizeof(spl_kmem_magazine_t) + |
1101 | sizeof(void *) * skm->skm_size; | |
1102 | ||
4afaaefa | 1103 | ENTRY; |
1104 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1105 | ASSERT(skm->skm_avail == 0); | |
a0f6da3d | 1106 | |
1107 | kmem_free(skm, size); | |
4afaaefa | 1108 | EXIT; |
1109 | } | |
1110 | ||
ea3e6ca9 BB |
1111 | /* |
1112 | * Create all pre-cpu magazines of reasonable sizes. | |
1113 | */ | |
4afaaefa | 1114 | static int |
1115 | spl_magazine_create(spl_kmem_cache_t *skc) | |
1116 | { | |
37db7d8c | 1117 | int i; |
4afaaefa | 1118 | ENTRY; |
1119 | ||
1120 | skc->skc_mag_size = spl_magazine_size(skc); | |
ea3e6ca9 | 1121 | skc->skc_mag_refill = (skc->skc_mag_size + 1) / 2; |
4afaaefa | 1122 | |
37db7d8c BB |
1123 | for_each_online_cpu(i) { |
1124 | skc->skc_mag[i] = spl_magazine_alloc(skc, cpu_to_node(i)); | |
1125 | if (!skc->skc_mag[i]) { | |
1126 | for (i--; i >= 0; i--) | |
1127 | spl_magazine_free(skc->skc_mag[i]); | |
4afaaefa | 1128 | |
37db7d8c BB |
1129 | RETURN(-ENOMEM); |
1130 | } | |
1131 | } | |
4afaaefa | 1132 | |
9b1b8e4c BB |
1133 | /* Only after everything is allocated schedule magazine work */ |
1134 | for_each_online_cpu(i) | |
1135 | schedule_delayed_work_on(i, &skc->skc_mag[i]->skm_work, | |
1136 | skc->skc_delay / 3 * HZ); | |
1137 | ||
37db7d8c | 1138 | RETURN(0); |
4afaaefa | 1139 | } |
1140 | ||
ea3e6ca9 BB |
1141 | /* |
1142 | * Destroy all pre-cpu magazines. | |
1143 | */ | |
4afaaefa | 1144 | static void |
1145 | spl_magazine_destroy(spl_kmem_cache_t *skc) | |
1146 | { | |
37db7d8c BB |
1147 | spl_kmem_magazine_t *skm; |
1148 | int i; | |
4afaaefa | 1149 | ENTRY; |
37db7d8c BB |
1150 | |
1151 | for_each_online_cpu(i) { | |
1152 | skm = skc->skc_mag[i]; | |
1153 | (void)spl_cache_flush(skc, skm, skm->skm_avail); | |
1154 | spl_magazine_free(skm); | |
1155 | } | |
1156 | ||
4afaaefa | 1157 | EXIT; |
1158 | } | |
1159 | ||
ea3e6ca9 BB |
1160 | /* |
1161 | * Create a object cache based on the following arguments: | |
1162 | * name cache name | |
1163 | * size cache object size | |
1164 | * align cache object alignment | |
1165 | * ctor cache object constructor | |
1166 | * dtor cache object destructor | |
1167 | * reclaim cache object reclaim | |
1168 | * priv cache private data for ctor/dtor/reclaim | |
1169 | * vmp unused must be NULL | |
1170 | * flags | |
1171 | * KMC_NOTOUCH Disable cache object aging (unsupported) | |
1172 | * KMC_NODEBUG Disable debugging (unsupported) | |
1173 | * KMC_NOMAGAZINE Disable magazine (unsupported) | |
1174 | * KMC_NOHASH Disable hashing (unsupported) | |
1175 | * KMC_QCACHE Disable qcache (unsupported) | |
1176 | * KMC_KMEM Force kmem backed cache | |
1177 | * KMC_VMEM Force vmem backed cache | |
1178 | * KMC_OFFSLAB Locate objects off the slab | |
1179 | */ | |
2fb9b26a | 1180 | spl_kmem_cache_t * |
1181 | spl_kmem_cache_create(char *name, size_t size, size_t align, | |
1182 | spl_kmem_ctor_t ctor, | |
1183 | spl_kmem_dtor_t dtor, | |
1184 | spl_kmem_reclaim_t reclaim, | |
1185 | void *priv, void *vmp, int flags) | |
1186 | { | |
1187 | spl_kmem_cache_t *skc; | |
a1502d76 | 1188 | int rc, kmem_flags = KM_SLEEP; |
2fb9b26a | 1189 | ENTRY; |
937879f1 | 1190 | |
a1502d76 | 1191 | ASSERTF(!(flags & KMC_NOMAGAZINE), "Bad KMC_NOMAGAZINE (%x)\n", flags); |
1192 | ASSERTF(!(flags & KMC_NOHASH), "Bad KMC_NOHASH (%x)\n", flags); | |
1193 | ASSERTF(!(flags & KMC_QCACHE), "Bad KMC_QCACHE (%x)\n", flags); | |
48e0606a | 1194 | ASSERT(vmp == NULL); |
a1502d76 | 1195 | |
2fb9b26a | 1196 | /* We may be called when there is a non-zero preempt_count or |
1197 | * interrupts are disabled is which case we must not sleep. | |
1198 | */ | |
e9d7a2be | 1199 | if (current_thread_info()->preempt_count || irqs_disabled()) |
2fb9b26a | 1200 | kmem_flags = KM_NOSLEEP; |
0a6fd143 | 1201 | |
2fb9b26a | 1202 | /* Allocate new cache memory and initialize. */ |
ff449ac4 | 1203 | skc = (spl_kmem_cache_t *)kmem_zalloc(sizeof(*skc), kmem_flags); |
e9d7a2be | 1204 | if (skc == NULL) |
2fb9b26a | 1205 | RETURN(NULL); |
d61e12af | 1206 | |
2fb9b26a | 1207 | skc->skc_magic = SKC_MAGIC; |
2fb9b26a | 1208 | skc->skc_name_size = strlen(name) + 1; |
1209 | skc->skc_name = (char *)kmem_alloc(skc->skc_name_size, kmem_flags); | |
1210 | if (skc->skc_name == NULL) { | |
1211 | kmem_free(skc, sizeof(*skc)); | |
1212 | RETURN(NULL); | |
1213 | } | |
1214 | strncpy(skc->skc_name, name, skc->skc_name_size); | |
1215 | ||
e9d7a2be | 1216 | skc->skc_ctor = ctor; |
1217 | skc->skc_dtor = dtor; | |
1218 | skc->skc_reclaim = reclaim; | |
2fb9b26a | 1219 | skc->skc_private = priv; |
1220 | skc->skc_vmp = vmp; | |
1221 | skc->skc_flags = flags; | |
1222 | skc->skc_obj_size = size; | |
48e0606a | 1223 | skc->skc_obj_align = SPL_KMEM_CACHE_ALIGN; |
2fb9b26a | 1224 | skc->skc_delay = SPL_KMEM_CACHE_DELAY; |
37db7d8c | 1225 | skc->skc_reap = SPL_KMEM_CACHE_REAP; |
ea3e6ca9 | 1226 | atomic_set(&skc->skc_ref, 0); |
2fb9b26a | 1227 | |
2fb9b26a | 1228 | INIT_LIST_HEAD(&skc->skc_list); |
1229 | INIT_LIST_HEAD(&skc->skc_complete_list); | |
1230 | INIT_LIST_HEAD(&skc->skc_partial_list); | |
d46630e0 | 1231 | spin_lock_init(&skc->skc_lock); |
e9d7a2be | 1232 | skc->skc_slab_fail = 0; |
1233 | skc->skc_slab_create = 0; | |
1234 | skc->skc_slab_destroy = 0; | |
2fb9b26a | 1235 | skc->skc_slab_total = 0; |
1236 | skc->skc_slab_alloc = 0; | |
1237 | skc->skc_slab_max = 0; | |
1238 | skc->skc_obj_total = 0; | |
1239 | skc->skc_obj_alloc = 0; | |
1240 | skc->skc_obj_max = 0; | |
a1502d76 | 1241 | |
48e0606a BB |
1242 | if (align) { |
1243 | ASSERT((align & (align - 1)) == 0); /* Power of two */ | |
1244 | ASSERT(align >= SPL_KMEM_CACHE_ALIGN); /* Minimum size */ | |
1245 | skc->skc_obj_align = align; | |
1246 | } | |
1247 | ||
a1502d76 | 1248 | /* If none passed select a cache type based on object size */ |
1249 | if (!(skc->skc_flags & (KMC_KMEM | KMC_VMEM))) { | |
48e0606a BB |
1250 | if (P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align) < |
1251 | (PAGE_SIZE / 8)) { | |
a1502d76 | 1252 | skc->skc_flags |= KMC_KMEM; |
1253 | } else { | |
1254 | skc->skc_flags |= KMC_VMEM; | |
1255 | } | |
1256 | } | |
1257 | ||
48e0606a BB |
1258 | rc = spl_slab_size(skc, &skc->skc_slab_objs, &skc->skc_slab_size); |
1259 | if (rc) | |
1260 | GOTO(out, rc); | |
4afaaefa | 1261 | |
1262 | rc = spl_magazine_create(skc); | |
48e0606a BB |
1263 | if (rc) |
1264 | GOTO(out, rc); | |
2fb9b26a | 1265 | |
ea3e6ca9 | 1266 | spl_init_delayed_work(&skc->skc_work, spl_cache_age, skc); |
37db7d8c | 1267 | schedule_delayed_work(&skc->skc_work, skc->skc_delay / 3 * HZ); |
ea3e6ca9 | 1268 | |
2fb9b26a | 1269 | down_write(&spl_kmem_cache_sem); |
e9d7a2be | 1270 | list_add_tail(&skc->skc_list, &spl_kmem_cache_list); |
2fb9b26a | 1271 | up_write(&spl_kmem_cache_sem); |
1272 | ||
e9d7a2be | 1273 | RETURN(skc); |
48e0606a BB |
1274 | out: |
1275 | kmem_free(skc->skc_name, skc->skc_name_size); | |
1276 | kmem_free(skc, sizeof(*skc)); | |
1277 | RETURN(NULL); | |
f1ca4da6 | 1278 | } |
2fb9b26a | 1279 | EXPORT_SYMBOL(spl_kmem_cache_create); |
f1ca4da6 | 1280 | |
ea3e6ca9 BB |
1281 | /* |
1282 | * Destroy a cache and all objects assoicated with the cache. | |
1283 | */ | |
2fb9b26a | 1284 | void |
1285 | spl_kmem_cache_destroy(spl_kmem_cache_t *skc) | |
f1ca4da6 | 1286 | { |
ea3e6ca9 | 1287 | DECLARE_WAIT_QUEUE_HEAD(wq); |
9b1b8e4c | 1288 | int i; |
2fb9b26a | 1289 | ENTRY; |
f1ca4da6 | 1290 | |
e9d7a2be | 1291 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1292 | ||
1293 | down_write(&spl_kmem_cache_sem); | |
1294 | list_del_init(&skc->skc_list); | |
1295 | up_write(&spl_kmem_cache_sem); | |
2fb9b26a | 1296 | |
ea3e6ca9 BB |
1297 | /* Cancel any and wait for any pending delayed work */ |
1298 | ASSERT(!test_and_set_bit(KMC_BIT_DESTROY, &skc->skc_flags)); | |
1299 | cancel_delayed_work(&skc->skc_work); | |
9b1b8e4c BB |
1300 | for_each_online_cpu(i) |
1301 | cancel_delayed_work(&skc->skc_mag[i]->skm_work); | |
1302 | ||
ea3e6ca9 BB |
1303 | flush_scheduled_work(); |
1304 | ||
1305 | /* Wait until all current callers complete, this is mainly | |
1306 | * to catch the case where a low memory situation triggers a | |
1307 | * cache reaping action which races with this destroy. */ | |
1308 | wait_event(wq, atomic_read(&skc->skc_ref) == 0); | |
1309 | ||
4afaaefa | 1310 | spl_magazine_destroy(skc); |
37db7d8c | 1311 | spl_slab_reclaim(skc, 0, 1); |
d46630e0 | 1312 | spin_lock(&skc->skc_lock); |
d6a26c6a | 1313 | |
2fb9b26a | 1314 | /* Validate there are no objects in use and free all the |
4afaaefa | 1315 | * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */ |
ea3e6ca9 BB |
1316 | ASSERT3U(skc->skc_slab_alloc, ==, 0); |
1317 | ASSERT3U(skc->skc_obj_alloc, ==, 0); | |
1318 | ASSERT3U(skc->skc_slab_total, ==, 0); | |
1319 | ASSERT3U(skc->skc_obj_total, ==, 0); | |
2fb9b26a | 1320 | ASSERT(list_empty(&skc->skc_complete_list)); |
a1502d76 | 1321 | |
2fb9b26a | 1322 | kmem_free(skc->skc_name, skc->skc_name_size); |
d46630e0 | 1323 | spin_unlock(&skc->skc_lock); |
ff449ac4 | 1324 | |
4afaaefa | 1325 | kmem_free(skc, sizeof(*skc)); |
2fb9b26a | 1326 | |
1327 | EXIT; | |
f1ca4da6 | 1328 | } |
2fb9b26a | 1329 | EXPORT_SYMBOL(spl_kmem_cache_destroy); |
f1ca4da6 | 1330 | |
ea3e6ca9 BB |
1331 | /* |
1332 | * Allocate an object from a slab attached to the cache. This is used to | |
1333 | * repopulate the per-cpu magazine caches in batches when they run low. | |
1334 | */ | |
4afaaefa | 1335 | static void * |
1336 | spl_cache_obj(spl_kmem_cache_t *skc, spl_kmem_slab_t *sks) | |
f1ca4da6 | 1337 | { |
2fb9b26a | 1338 | spl_kmem_obj_t *sko; |
f1ca4da6 | 1339 | |
e9d7a2be | 1340 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1341 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
4afaaefa | 1342 | ASSERT(spin_is_locked(&skc->skc_lock)); |
2fb9b26a | 1343 | |
a1502d76 | 1344 | sko = list_entry(sks->sks_free_list.next, spl_kmem_obj_t, sko_list); |
4afaaefa | 1345 | ASSERT(sko->sko_magic == SKO_MAGIC); |
1346 | ASSERT(sko->sko_addr != NULL); | |
2fb9b26a | 1347 | |
a1502d76 | 1348 | /* Remove from sks_free_list */ |
4afaaefa | 1349 | list_del_init(&sko->sko_list); |
2fb9b26a | 1350 | |
4afaaefa | 1351 | sks->sks_age = jiffies; |
1352 | sks->sks_ref++; | |
1353 | skc->skc_obj_alloc++; | |
2fb9b26a | 1354 | |
4afaaefa | 1355 | /* Track max obj usage statistics */ |
1356 | if (skc->skc_obj_alloc > skc->skc_obj_max) | |
1357 | skc->skc_obj_max = skc->skc_obj_alloc; | |
2fb9b26a | 1358 | |
4afaaefa | 1359 | /* Track max slab usage statistics */ |
1360 | if (sks->sks_ref == 1) { | |
1361 | skc->skc_slab_alloc++; | |
f1ca4da6 | 1362 | |
4afaaefa | 1363 | if (skc->skc_slab_alloc > skc->skc_slab_max) |
1364 | skc->skc_slab_max = skc->skc_slab_alloc; | |
2fb9b26a | 1365 | } |
1366 | ||
4afaaefa | 1367 | return sko->sko_addr; |
1368 | } | |
c30df9c8 | 1369 | |
ea3e6ca9 BB |
1370 | /* |
1371 | * No available objects on any slabsi, create a new slab. Since this | |
1372 | * is an expensive operation we do it without holding the spinlock and | |
1373 | * only briefly aquire it when we link in the fully allocated and | |
1374 | * constructed slab. | |
4afaaefa | 1375 | */ |
1376 | static spl_kmem_slab_t * | |
1377 | spl_cache_grow(spl_kmem_cache_t *skc, int flags) | |
1378 | { | |
e9d7a2be | 1379 | spl_kmem_slab_t *sks; |
4afaaefa | 1380 | ENTRY; |
f1ca4da6 | 1381 | |
e9d7a2be | 1382 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1383 | local_irq_enable(); |
1384 | might_sleep(); | |
e9d7a2be | 1385 | |
ea3e6ca9 BB |
1386 | /* |
1387 | * Before allocating a new slab check if the slab is being reaped. | |
1388 | * If it is there is a good chance we can wait until it finishes | |
1389 | * and then use one of the newly freed but not aged-out slabs. | |
1390 | */ | |
1391 | if (test_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
1392 | schedule(); | |
1393 | GOTO(out, sks= NULL); | |
4afaaefa | 1394 | } |
2fb9b26a | 1395 | |
ea3e6ca9 BB |
1396 | /* Allocate a new slab for the cache */ |
1397 | sks = spl_slab_alloc(skc, flags | __GFP_NORETRY | __GFP_NOWARN); | |
1398 | if (sks == NULL) | |
1399 | GOTO(out, sks = NULL); | |
4afaaefa | 1400 | |
ea3e6ca9 | 1401 | /* Link the new empty slab in to the end of skc_partial_list. */ |
d46630e0 | 1402 | spin_lock(&skc->skc_lock); |
2fb9b26a | 1403 | skc->skc_slab_total++; |
1404 | skc->skc_obj_total += sks->sks_objs; | |
1405 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
d46630e0 | 1406 | spin_unlock(&skc->skc_lock); |
ea3e6ca9 BB |
1407 | out: |
1408 | local_irq_disable(); | |
4afaaefa | 1409 | |
1410 | RETURN(sks); | |
f1ca4da6 | 1411 | } |
1412 | ||
ea3e6ca9 BB |
1413 | /* |
1414 | * Refill a per-cpu magazine with objects from the slabs for this | |
1415 | * cache. Ideally the magazine can be repopulated using existing | |
1416 | * objects which have been released, however if we are unable to | |
1417 | * locate enough free objects new slabs of objects will be created. | |
1418 | */ | |
4afaaefa | 1419 | static int |
1420 | spl_cache_refill(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flags) | |
f1ca4da6 | 1421 | { |
e9d7a2be | 1422 | spl_kmem_slab_t *sks; |
1423 | int rc = 0, refill; | |
937879f1 | 1424 | ENTRY; |
f1ca4da6 | 1425 | |
e9d7a2be | 1426 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1427 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
1428 | ||
e9d7a2be | 1429 | refill = MIN(skm->skm_refill, skm->skm_size - skm->skm_avail); |
d46630e0 | 1430 | spin_lock(&skc->skc_lock); |
ff449ac4 | 1431 | |
4afaaefa | 1432 | while (refill > 0) { |
ea3e6ca9 | 1433 | /* No slabs available we may need to grow the cache */ |
4afaaefa | 1434 | if (list_empty(&skc->skc_partial_list)) { |
1435 | spin_unlock(&skc->skc_lock); | |
ff449ac4 | 1436 | |
4afaaefa | 1437 | sks = spl_cache_grow(skc, flags); |
1438 | if (!sks) | |
e9d7a2be | 1439 | GOTO(out, rc); |
4afaaefa | 1440 | |
1441 | /* Rescheduled to different CPU skm is not local */ | |
1442 | if (skm != skc->skc_mag[smp_processor_id()]) | |
e9d7a2be | 1443 | GOTO(out, rc); |
1444 | ||
1445 | /* Potentially rescheduled to the same CPU but | |
1446 | * allocations may have occured from this CPU while | |
1447 | * we were sleeping so recalculate max refill. */ | |
1448 | refill = MIN(refill, skm->skm_size - skm->skm_avail); | |
4afaaefa | 1449 | |
1450 | spin_lock(&skc->skc_lock); | |
1451 | continue; | |
1452 | } | |
d46630e0 | 1453 | |
4afaaefa | 1454 | /* Grab the next available slab */ |
1455 | sks = list_entry((&skc->skc_partial_list)->next, | |
1456 | spl_kmem_slab_t, sks_list); | |
1457 | ASSERT(sks->sks_magic == SKS_MAGIC); | |
1458 | ASSERT(sks->sks_ref < sks->sks_objs); | |
1459 | ASSERT(!list_empty(&sks->sks_free_list)); | |
d46630e0 | 1460 | |
4afaaefa | 1461 | /* Consume as many objects as needed to refill the requested |
e9d7a2be | 1462 | * cache. We must also be careful not to overfill it. */ |
1463 | while (sks->sks_ref < sks->sks_objs && refill-- > 0 && ++rc) { | |
1464 | ASSERT(skm->skm_avail < skm->skm_size); | |
1465 | ASSERT(rc < skm->skm_size); | |
4afaaefa | 1466 | skm->skm_objs[skm->skm_avail++]=spl_cache_obj(skc,sks); |
e9d7a2be | 1467 | } |
f1ca4da6 | 1468 | |
4afaaefa | 1469 | /* Move slab to skc_complete_list when full */ |
1470 | if (sks->sks_ref == sks->sks_objs) { | |
1471 | list_del(&sks->sks_list); | |
1472 | list_add(&sks->sks_list, &skc->skc_complete_list); | |
2fb9b26a | 1473 | } |
1474 | } | |
57d86234 | 1475 | |
4afaaefa | 1476 | spin_unlock(&skc->skc_lock); |
1477 | out: | |
1478 | /* Returns the number of entries added to cache */ | |
e9d7a2be | 1479 | RETURN(rc); |
4afaaefa | 1480 | } |
1481 | ||
ea3e6ca9 BB |
1482 | /* |
1483 | * Release an object back to the slab from which it came. | |
1484 | */ | |
4afaaefa | 1485 | static void |
1486 | spl_cache_shrink(spl_kmem_cache_t *skc, void *obj) | |
1487 | { | |
e9d7a2be | 1488 | spl_kmem_slab_t *sks = NULL; |
4afaaefa | 1489 | spl_kmem_obj_t *sko = NULL; |
1490 | ENTRY; | |
1491 | ||
e9d7a2be | 1492 | ASSERT(skc->skc_magic == SKC_MAGIC); |
4afaaefa | 1493 | ASSERT(spin_is_locked(&skc->skc_lock)); |
1494 | ||
48e0606a | 1495 | sko = obj + P2ROUNDUP(skc->skc_obj_size, skc->skc_obj_align); |
a1502d76 | 1496 | ASSERT(sko->sko_magic == SKO_MAGIC); |
4afaaefa | 1497 | |
1498 | sks = sko->sko_slab; | |
a1502d76 | 1499 | ASSERT(sks->sks_magic == SKS_MAGIC); |
2fb9b26a | 1500 | ASSERT(sks->sks_cache == skc); |
2fb9b26a | 1501 | list_add(&sko->sko_list, &sks->sks_free_list); |
d6a26c6a | 1502 | |
2fb9b26a | 1503 | sks->sks_age = jiffies; |
4afaaefa | 1504 | sks->sks_ref--; |
2fb9b26a | 1505 | skc->skc_obj_alloc--; |
f1ca4da6 | 1506 | |
2fb9b26a | 1507 | /* Move slab to skc_partial_list when no longer full. Slabs |
4afaaefa | 1508 | * are added to the head to keep the partial list is quasi-full |
1509 | * sorted order. Fuller at the head, emptier at the tail. */ | |
1510 | if (sks->sks_ref == (sks->sks_objs - 1)) { | |
2fb9b26a | 1511 | list_del(&sks->sks_list); |
1512 | list_add(&sks->sks_list, &skc->skc_partial_list); | |
1513 | } | |
f1ca4da6 | 1514 | |
2fb9b26a | 1515 | /* Move emply slabs to the end of the partial list so |
4afaaefa | 1516 | * they can be easily found and freed during reclamation. */ |
1517 | if (sks->sks_ref == 0) { | |
2fb9b26a | 1518 | list_del(&sks->sks_list); |
1519 | list_add_tail(&sks->sks_list, &skc->skc_partial_list); | |
1520 | skc->skc_slab_alloc--; | |
1521 | } | |
1522 | ||
4afaaefa | 1523 | EXIT; |
1524 | } | |
1525 | ||
ea3e6ca9 BB |
1526 | /* |
1527 | * Release a batch of objects from a per-cpu magazine back to their | |
1528 | * respective slabs. This occurs when we exceed the magazine size, | |
1529 | * are under memory pressure, when the cache is idle, or during | |
1530 | * cache cleanup. The flush argument contains the number of entries | |
1531 | * to remove from the magazine. | |
1532 | */ | |
4afaaefa | 1533 | static int |
1534 | spl_cache_flush(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flush) | |
1535 | { | |
1536 | int i, count = MIN(flush, skm->skm_avail); | |
1537 | ENTRY; | |
1538 | ||
e9d7a2be | 1539 | ASSERT(skc->skc_magic == SKC_MAGIC); |
1540 | ASSERT(skm->skm_magic == SKM_MAGIC); | |
4afaaefa | 1541 | |
ea3e6ca9 BB |
1542 | /* |
1543 | * XXX: Currently we simply return objects from the magazine to | |
1544 | * the slabs in fifo order. The ideal thing to do from a memory | |
1545 | * fragmentation standpoint is to cheaply determine the set of | |
1546 | * objects in the magazine which will result in the largest | |
1547 | * number of free slabs if released from the magazine. | |
1548 | */ | |
4afaaefa | 1549 | spin_lock(&skc->skc_lock); |
1550 | for (i = 0; i < count; i++) | |
1551 | spl_cache_shrink(skc, skm->skm_objs[i]); | |
1552 | ||
e9d7a2be | 1553 | skm->skm_avail -= count; |
1554 | memmove(skm->skm_objs, &(skm->skm_objs[count]), | |
4afaaefa | 1555 | sizeof(void *) * skm->skm_avail); |
1556 | ||
d46630e0 | 1557 | spin_unlock(&skc->skc_lock); |
4afaaefa | 1558 | |
1559 | RETURN(count); | |
1560 | } | |
1561 | ||
ea3e6ca9 BB |
1562 | /* |
1563 | * Allocate an object from the per-cpu magazine, or if the magazine | |
1564 | * is empty directly allocate from a slab and repopulate the magazine. | |
1565 | */ | |
4afaaefa | 1566 | void * |
1567 | spl_kmem_cache_alloc(spl_kmem_cache_t *skc, int flags) | |
1568 | { | |
1569 | spl_kmem_magazine_t *skm; | |
1570 | unsigned long irq_flags; | |
1571 | void *obj = NULL; | |
1572 | ENTRY; | |
1573 | ||
e9d7a2be | 1574 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1575 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1576 | ASSERT(flags & KM_SLEEP); | |
1577 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1578 | local_irq_save(irq_flags); |
1579 | ||
1580 | restart: | |
1581 | /* Safe to update per-cpu structure without lock, but | |
1582 | * in the restart case we must be careful to reaquire | |
1583 | * the local magazine since this may have changed | |
1584 | * when we need to grow the cache. */ | |
1585 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1586 | ASSERTF(skm->skm_magic == SKM_MAGIC, "%x != %x: %s/%p/%p %x/%x/%x\n", |
1587 | skm->skm_magic, SKM_MAGIC, skc->skc_name, skc, skm, | |
1588 | skm->skm_size, skm->skm_refill, skm->skm_avail); | |
4afaaefa | 1589 | |
1590 | if (likely(skm->skm_avail)) { | |
1591 | /* Object available in CPU cache, use it */ | |
1592 | obj = skm->skm_objs[--skm->skm_avail]; | |
ea3e6ca9 | 1593 | skm->skm_age = jiffies; |
4afaaefa | 1594 | } else { |
1595 | /* Per-CPU cache empty, directly allocate from | |
1596 | * the slab and refill the per-CPU cache. */ | |
1597 | (void)spl_cache_refill(skc, skm, flags); | |
1598 | GOTO(restart, obj = NULL); | |
1599 | } | |
1600 | ||
1601 | local_irq_restore(irq_flags); | |
fece7c99 | 1602 | ASSERT(obj); |
48e0606a | 1603 | ASSERT(((unsigned long)(obj) % skc->skc_obj_align) == 0); |
4afaaefa | 1604 | |
1605 | /* Pre-emptively migrate object to CPU L1 cache */ | |
1606 | prefetchw(obj); | |
ea3e6ca9 | 1607 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1608 | |
1609 | RETURN(obj); | |
1610 | } | |
1611 | EXPORT_SYMBOL(spl_kmem_cache_alloc); | |
1612 | ||
ea3e6ca9 BB |
1613 | /* |
1614 | * Free an object back to the local per-cpu magazine, there is no | |
1615 | * guarantee that this is the same magazine the object was originally | |
1616 | * allocated from. We may need to flush entire from the magazine | |
1617 | * back to the slabs to make space. | |
1618 | */ | |
4afaaefa | 1619 | void |
1620 | spl_kmem_cache_free(spl_kmem_cache_t *skc, void *obj) | |
1621 | { | |
1622 | spl_kmem_magazine_t *skm; | |
1623 | unsigned long flags; | |
1624 | ENTRY; | |
1625 | ||
e9d7a2be | 1626 | ASSERT(skc->skc_magic == SKC_MAGIC); |
ea3e6ca9 BB |
1627 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
1628 | atomic_inc(&skc->skc_ref); | |
4afaaefa | 1629 | local_irq_save(flags); |
1630 | ||
1631 | /* Safe to update per-cpu structure without lock, but | |
1632 | * no remote memory allocation tracking is being performed | |
1633 | * it is entirely possible to allocate an object from one | |
1634 | * CPU cache and return it to another. */ | |
1635 | skm = skc->skc_mag[smp_processor_id()]; | |
e9d7a2be | 1636 | ASSERT(skm->skm_magic == SKM_MAGIC); |
4afaaefa | 1637 | |
1638 | /* Per-CPU cache full, flush it to make space */ | |
1639 | if (unlikely(skm->skm_avail >= skm->skm_size)) | |
1640 | (void)spl_cache_flush(skc, skm, skm->skm_refill); | |
1641 | ||
1642 | /* Available space in cache, use it */ | |
1643 | skm->skm_objs[skm->skm_avail++] = obj; | |
1644 | ||
1645 | local_irq_restore(flags); | |
ea3e6ca9 | 1646 | atomic_dec(&skc->skc_ref); |
4afaaefa | 1647 | |
1648 | EXIT; | |
f1ca4da6 | 1649 | } |
2fb9b26a | 1650 | EXPORT_SYMBOL(spl_kmem_cache_free); |
5c2bb9b2 | 1651 | |
ea3e6ca9 BB |
1652 | /* |
1653 | * The generic shrinker function for all caches. Under linux a shrinker | |
1654 | * may not be tightly coupled with a slab cache. In fact linux always | |
1655 | * systematically trys calling all registered shrinker callbacks which | |
1656 | * report that they contain unused objects. Because of this we only | |
1657 | * register one shrinker function in the shim layer for all slab caches. | |
1658 | * We always attempt to shrink all caches when this generic shrinker | |
1659 | * is called. The shrinker should return the number of free objects | |
1660 | * in the cache when called with nr_to_scan == 0 but not attempt to | |
1661 | * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan | |
1662 | * objects should be freed, because Solaris semantics are to free | |
1663 | * all available objects we may free more objects than requested. | |
1664 | */ | |
2fb9b26a | 1665 | static int |
4afaaefa | 1666 | spl_kmem_cache_generic_shrinker(int nr_to_scan, unsigned int gfp_mask) |
2fb9b26a | 1667 | { |
e9d7a2be | 1668 | spl_kmem_cache_t *skc; |
ea3e6ca9 | 1669 | int unused = 0; |
5c2bb9b2 | 1670 | |
e9d7a2be | 1671 | down_read(&spl_kmem_cache_sem); |
ea3e6ca9 BB |
1672 | list_for_each_entry(skc, &spl_kmem_cache_list, skc_list) { |
1673 | if (nr_to_scan) | |
1674 | spl_kmem_cache_reap_now(skc); | |
1675 | ||
1676 | /* | |
1677 | * Presume everything alloc'ed in reclaimable, this ensures | |
1678 | * we are called again with nr_to_scan > 0 so can try and | |
1679 | * reclaim. The exact number is not important either so | |
1680 | * we forgo taking this already highly contented lock. | |
1681 | */ | |
1682 | unused += skc->skc_obj_alloc; | |
1683 | } | |
e9d7a2be | 1684 | up_read(&spl_kmem_cache_sem); |
2fb9b26a | 1685 | |
ea3e6ca9 | 1686 | return (unused * sysctl_vfs_cache_pressure) / 100; |
5c2bb9b2 | 1687 | } |
5c2bb9b2 | 1688 | |
ea3e6ca9 BB |
1689 | /* |
1690 | * Call the registered reclaim function for a cache. Depending on how | |
1691 | * many and which objects are released it may simply repopulate the | |
1692 | * local magazine which will then need to age-out. Objects which cannot | |
1693 | * fit in the magazine we will be released back to their slabs which will | |
1694 | * also need to age out before being release. This is all just best | |
1695 | * effort and we do not want to thrash creating and destroying slabs. | |
1696 | */ | |
57d86234 | 1697 | void |
2fb9b26a | 1698 | spl_kmem_cache_reap_now(spl_kmem_cache_t *skc) |
57d86234 | 1699 | { |
2fb9b26a | 1700 | ENTRY; |
e9d7a2be | 1701 | |
1702 | ASSERT(skc->skc_magic == SKC_MAGIC); | |
ea3e6ca9 | 1703 | ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags)); |
2fb9b26a | 1704 | |
ea3e6ca9 BB |
1705 | /* Prevent concurrent cache reaping when contended */ |
1706 | if (test_and_set_bit(KMC_BIT_REAPING, &skc->skc_flags)) { | |
1707 | EXIT; | |
1708 | return; | |
1709 | } | |
2fb9b26a | 1710 | |
ea3e6ca9 | 1711 | atomic_inc(&skc->skc_ref); |
4afaaefa | 1712 | |
ea3e6ca9 BB |
1713 | if (skc->skc_reclaim) |
1714 | skc->skc_reclaim(skc->skc_private); | |
4afaaefa | 1715 | |
37db7d8c | 1716 | spl_slab_reclaim(skc, skc->skc_reap, 0); |
ea3e6ca9 BB |
1717 | clear_bit(KMC_BIT_REAPING, &skc->skc_flags); |
1718 | atomic_dec(&skc->skc_ref); | |
4afaaefa | 1719 | |
2fb9b26a | 1720 | EXIT; |
57d86234 | 1721 | } |
2fb9b26a | 1722 | EXPORT_SYMBOL(spl_kmem_cache_reap_now); |
57d86234 | 1723 | |
ea3e6ca9 BB |
1724 | /* |
1725 | * Reap all free slabs from all registered caches. | |
1726 | */ | |
f1b59d26 | 1727 | void |
2fb9b26a | 1728 | spl_kmem_reap(void) |
937879f1 | 1729 | { |
4afaaefa | 1730 | spl_kmem_cache_generic_shrinker(KMC_REAP_CHUNK, GFP_KERNEL); |
f1ca4da6 | 1731 | } |
2fb9b26a | 1732 | EXPORT_SYMBOL(spl_kmem_reap); |
5d86345d | 1733 | |
ff449ac4 | 1734 | #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING) |
c6dc93d6 | 1735 | static char * |
4afaaefa | 1736 | spl_sprintf_addr(kmem_debug_t *kd, char *str, int len, int min) |
d6a26c6a | 1737 | { |
e9d7a2be | 1738 | int size = ((len - 1) < kd->kd_size) ? (len - 1) : kd->kd_size; |
d6a26c6a | 1739 | int i, flag = 1; |
1740 | ||
1741 | ASSERT(str != NULL && len >= 17); | |
e9d7a2be | 1742 | memset(str, 0, len); |
d6a26c6a | 1743 | |
1744 | /* Check for a fully printable string, and while we are at | |
1745 | * it place the printable characters in the passed buffer. */ | |
1746 | for (i = 0; i < size; i++) { | |
e9d7a2be | 1747 | str[i] = ((char *)(kd->kd_addr))[i]; |
1748 | if (isprint(str[i])) { | |
1749 | continue; | |
1750 | } else { | |
1751 | /* Minimum number of printable characters found | |
1752 | * to make it worthwhile to print this as ascii. */ | |
1753 | if (i > min) | |
1754 | break; | |
1755 | ||
1756 | flag = 0; | |
1757 | break; | |
1758 | } | |
d6a26c6a | 1759 | } |
1760 | ||
1761 | if (!flag) { | |
1762 | sprintf(str, "%02x%02x%02x%02x%02x%02x%02x%02x", | |
1763 | *((uint8_t *)kd->kd_addr), | |
1764 | *((uint8_t *)kd->kd_addr + 2), | |
1765 | *((uint8_t *)kd->kd_addr + 4), | |
1766 | *((uint8_t *)kd->kd_addr + 6), | |
1767 | *((uint8_t *)kd->kd_addr + 8), | |
1768 | *((uint8_t *)kd->kd_addr + 10), | |
1769 | *((uint8_t *)kd->kd_addr + 12), | |
1770 | *((uint8_t *)kd->kd_addr + 14)); | |
1771 | } | |
1772 | ||
1773 | return str; | |
1774 | } | |
1775 | ||
a1502d76 | 1776 | static int |
1777 | spl_kmem_init_tracking(struct list_head *list, spinlock_t *lock, int size) | |
1778 | { | |
1779 | int i; | |
1780 | ENTRY; | |
1781 | ||
1782 | spin_lock_init(lock); | |
1783 | INIT_LIST_HEAD(list); | |
1784 | ||
1785 | for (i = 0; i < size; i++) | |
1786 | INIT_HLIST_HEAD(&kmem_table[i]); | |
1787 | ||
1788 | RETURN(0); | |
1789 | } | |
1790 | ||
ff449ac4 | 1791 | static void |
1792 | spl_kmem_fini_tracking(struct list_head *list, spinlock_t *lock) | |
5d86345d | 1793 | { |
2fb9b26a | 1794 | unsigned long flags; |
1795 | kmem_debug_t *kd; | |
1796 | char str[17]; | |
a1502d76 | 1797 | ENTRY; |
2fb9b26a | 1798 | |
ff449ac4 | 1799 | spin_lock_irqsave(lock, flags); |
1800 | if (!list_empty(list)) | |
a0f6da3d | 1801 | printk(KERN_WARNING "%-16s %-5s %-16s %s:%s\n", "address", |
1802 | "size", "data", "func", "line"); | |
2fb9b26a | 1803 | |
ff449ac4 | 1804 | list_for_each_entry(kd, list, kd_list) |
a0f6da3d | 1805 | printk(KERN_WARNING "%p %-5d %-16s %s:%d\n", kd->kd_addr, |
b6b2acc6 | 1806 | (int)kd->kd_size, spl_sprintf_addr(kd, str, 17, 8), |
2fb9b26a | 1807 | kd->kd_func, kd->kd_line); |
1808 | ||
ff449ac4 | 1809 | spin_unlock_irqrestore(lock, flags); |
a1502d76 | 1810 | EXIT; |
ff449ac4 | 1811 | } |
1812 | #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
a1502d76 | 1813 | #define spl_kmem_init_tracking(list, lock, size) |
ff449ac4 | 1814 | #define spl_kmem_fini_tracking(list, lock) |
1815 | #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */ | |
1816 | ||
36b313da BB |
1817 | static void |
1818 | spl_kmem_init_globals(void) | |
1819 | { | |
1820 | struct zone *zone; | |
1821 | ||
1822 | /* For now all zones are includes, it may be wise to restrict | |
1823 | * this to normal and highmem zones if we see problems. */ | |
1824 | for_each_zone(zone) { | |
1825 | ||
1826 | if (!populated_zone(zone)) | |
1827 | continue; | |
1828 | ||
1829 | minfree += zone->pages_min; | |
1830 | desfree += zone->pages_low; | |
1831 | lotsfree += zone->pages_high; | |
1832 | } | |
4ab13d3b BB |
1833 | |
1834 | /* Solaris default values */ | |
96dded38 BB |
1835 | swapfs_minfree = MAX(2*1024*1024 >> PAGE_SHIFT, physmem >> 3); |
1836 | swapfs_reserve = MIN(4*1024*1024 >> PAGE_SHIFT, physmem >> 4); | |
36b313da BB |
1837 | } |
1838 | ||
d1ff2312 BB |
1839 | /* |
1840 | * Called at module init when it is safe to use spl_kallsyms_lookup_name() | |
1841 | */ | |
1842 | int | |
1843 | spl_kmem_init_kallsyms_lookup(void) | |
1844 | { | |
1845 | #ifndef HAVE_GET_VMALLOC_INFO | |
1846 | get_vmalloc_info_fn = (get_vmalloc_info_t) | |
1847 | spl_kallsyms_lookup_name("get_vmalloc_info"); | |
e11d6c5f BB |
1848 | if (!get_vmalloc_info_fn) { |
1849 | printk(KERN_ERR "Error: Unknown symbol get_vmalloc_info\n"); | |
d1ff2312 | 1850 | return -EFAULT; |
e11d6c5f | 1851 | } |
d1ff2312 BB |
1852 | #endif /* HAVE_GET_VMALLOC_INFO */ |
1853 | ||
5232d256 BB |
1854 | #ifdef HAVE_PGDAT_HELPERS |
1855 | # ifndef HAVE_FIRST_ONLINE_PGDAT | |
d1ff2312 BB |
1856 | first_online_pgdat_fn = (first_online_pgdat_t) |
1857 | spl_kallsyms_lookup_name("first_online_pgdat"); | |
e11d6c5f BB |
1858 | if (!first_online_pgdat_fn) { |
1859 | printk(KERN_ERR "Error: Unknown symbol first_online_pgdat\n"); | |
d1ff2312 | 1860 | return -EFAULT; |
e11d6c5f | 1861 | } |
5232d256 | 1862 | # endif /* HAVE_FIRST_ONLINE_PGDAT */ |
d1ff2312 | 1863 | |
5232d256 | 1864 | # ifndef HAVE_NEXT_ONLINE_PGDAT |
d1ff2312 BB |
1865 | next_online_pgdat_fn = (next_online_pgdat_t) |
1866 | spl_kallsyms_lookup_name("next_online_pgdat"); | |
e11d6c5f BB |
1867 | if (!next_online_pgdat_fn) { |
1868 | printk(KERN_ERR "Error: Unknown symbol next_online_pgdat\n"); | |
d1ff2312 | 1869 | return -EFAULT; |
e11d6c5f | 1870 | } |
5232d256 | 1871 | # endif /* HAVE_NEXT_ONLINE_PGDAT */ |
d1ff2312 | 1872 | |
5232d256 | 1873 | # ifndef HAVE_NEXT_ZONE |
d1ff2312 BB |
1874 | next_zone_fn = (next_zone_t) |
1875 | spl_kallsyms_lookup_name("next_zone"); | |
e11d6c5f BB |
1876 | if (!next_zone_fn) { |
1877 | printk(KERN_ERR "Error: Unknown symbol next_zone\n"); | |
d1ff2312 | 1878 | return -EFAULT; |
e11d6c5f | 1879 | } |
5232d256 BB |
1880 | # endif /* HAVE_NEXT_ZONE */ |
1881 | ||
1882 | #else /* HAVE_PGDAT_HELPERS */ | |
1883 | ||
1884 | # ifndef HAVE_PGDAT_LIST | |
124ca8a5 | 1885 | pgdat_list_addr = *(struct pglist_data **) |
5232d256 BB |
1886 | spl_kallsyms_lookup_name("pgdat_list"); |
1887 | if (!pgdat_list_addr) { | |
1888 | printk(KERN_ERR "Error: Unknown symbol pgdat_list\n"); | |
1889 | return -EFAULT; | |
1890 | } | |
1891 | # endif /* HAVE_PGDAT_LIST */ | |
1892 | #endif /* HAVE_PGDAT_HELPERS */ | |
d1ff2312 | 1893 | |
6ae7fef5 | 1894 | #if defined(NEED_GET_ZONE_COUNTS) && !defined(HAVE_GET_ZONE_COUNTS) |
d1ff2312 BB |
1895 | get_zone_counts_fn = (get_zone_counts_t) |
1896 | spl_kallsyms_lookup_name("get_zone_counts"); | |
e11d6c5f BB |
1897 | if (!get_zone_counts_fn) { |
1898 | printk(KERN_ERR "Error: Unknown symbol get_zone_counts\n"); | |
d1ff2312 | 1899 | return -EFAULT; |
e11d6c5f | 1900 | } |
6ae7fef5 | 1901 | #endif /* NEED_GET_ZONE_COUNTS && !HAVE_GET_ZONE_COUNTS */ |
d1ff2312 BB |
1902 | |
1903 | /* | |
1904 | * It is now safe to initialize the global tunings which rely on | |
1905 | * the use of the for_each_zone() macro. This macro in turns | |
1906 | * depends on the *_pgdat symbols which are now available. | |
1907 | */ | |
1908 | spl_kmem_init_globals(); | |
1909 | ||
1910 | return 0; | |
1911 | } | |
1912 | ||
a1502d76 | 1913 | int |
1914 | spl_kmem_init(void) | |
1915 | { | |
1916 | int rc = 0; | |
1917 | ENTRY; | |
1918 | ||
1919 | init_rwsem(&spl_kmem_cache_sem); | |
1920 | INIT_LIST_HEAD(&spl_kmem_cache_list); | |
1921 | ||
1922 | #ifdef HAVE_SET_SHRINKER | |
1923 | spl_kmem_cache_shrinker = set_shrinker(KMC_DEFAULT_SEEKS, | |
1924 | spl_kmem_cache_generic_shrinker); | |
1925 | if (spl_kmem_cache_shrinker == NULL) | |
f78a933f | 1926 | RETURN(rc = -ENOMEM); |
a1502d76 | 1927 | #else |
1928 | register_shrinker(&spl_kmem_cache_shrinker); | |
1929 | #endif | |
1930 | ||
1931 | #ifdef DEBUG_KMEM | |
1932 | atomic64_set(&kmem_alloc_used, 0); | |
1933 | atomic64_set(&vmem_alloc_used, 0); | |
1934 | ||
1935 | spl_kmem_init_tracking(&kmem_list, &kmem_lock, KMEM_TABLE_SIZE); | |
1936 | spl_kmem_init_tracking(&vmem_list, &vmem_lock, VMEM_TABLE_SIZE); | |
1937 | #endif | |
a1502d76 | 1938 | RETURN(rc); |
1939 | } | |
1940 | ||
ff449ac4 | 1941 | void |
1942 | spl_kmem_fini(void) | |
1943 | { | |
1944 | #ifdef DEBUG_KMEM | |
1945 | /* Display all unreclaimed memory addresses, including the | |
1946 | * allocation size and the first few bytes of what's located | |
1947 | * at that address to aid in debugging. Performance is not | |
1948 | * a serious concern here since it is module unload time. */ | |
1949 | if (atomic64_read(&kmem_alloc_used) != 0) | |
1950 | CWARN("kmem leaked %ld/%ld bytes\n", | |
550f1705 | 1951 | atomic64_read(&kmem_alloc_used), kmem_alloc_max); |
ff449ac4 | 1952 | |
2fb9b26a | 1953 | |
1954 | if (atomic64_read(&vmem_alloc_used) != 0) | |
1955 | CWARN("vmem leaked %ld/%ld bytes\n", | |
550f1705 | 1956 | atomic64_read(&vmem_alloc_used), vmem_alloc_max); |
2fb9b26a | 1957 | |
ff449ac4 | 1958 | spl_kmem_fini_tracking(&kmem_list, &kmem_lock); |
1959 | spl_kmem_fini_tracking(&vmem_list, &vmem_lock); | |
1960 | #endif /* DEBUG_KMEM */ | |
2fb9b26a | 1961 | ENTRY; |
1962 | ||
1963 | #ifdef HAVE_SET_SHRINKER | |
1964 | remove_shrinker(spl_kmem_cache_shrinker); | |
1965 | #else | |
1966 | unregister_shrinker(&spl_kmem_cache_shrinker); | |
5d86345d | 1967 | #endif |
2fb9b26a | 1968 | |
937879f1 | 1969 | EXIT; |
5d86345d | 1970 | } |