1 /*****************************************************************************\
2 * Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3 * Copyright (C) 2007 The Regents of the University of California.
4 * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5 * Written by Brian Behlendorf <behlendorf1@llnl.gov>.
8 * This file is part of the SPL, Solaris Porting Layer.
9 * For details, see <http://zfsonlinux.org/>.
11 * The SPL is free software; you can redistribute it and/or modify it
12 * under the terms of the GNU General Public License as published by the
13 * Free Software Foundation; either version 2 of the License, or (at your
14 * option) any later version.
16 * The SPL is distributed in the hope that it will be useful, but WITHOUT
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
21 * You should have received a copy of the GNU General Public License along
22 * with the SPL. If not, see <http://www.gnu.org/licenses/>.
23 *****************************************************************************
24 * Solaris Porting Layer (SPL) Kmem Implementation.
25 \*****************************************************************************/
28 #include <spl-debug.h>
30 #ifdef SS_DEBUG_SUBSYS
31 #undef SS_DEBUG_SUBSYS
34 #define SS_DEBUG_SUBSYS SS_KMEM
37 * Within the scope of spl-kmem.c file the kmem_cache_* definitions
38 * are removed to allow access to the real Linux slab allocator.
40 #undef kmem_cache_destroy
41 #undef kmem_cache_create
42 #undef kmem_cache_alloc
43 #undef kmem_cache_free
47 * Cache expiration was implemented because it was part of the default Solaris
48 * kmem_cache behavior. The idea is that per-cpu objects which haven't been
49 * accessed in several seconds should be returned to the cache. On the other
50 * hand Linux slabs never move objects back to the slabs unless there is
51 * memory pressure on the system. By default the Linux method is enabled
52 * because it has been shown to improve responsiveness on low memory systems.
53 * This policy may be changed by setting KMC_EXPIRE_AGE or KMC_EXPIRE_MEM.
55 unsigned int spl_kmem_cache_expire
= KMC_EXPIRE_MEM
;
56 EXPORT_SYMBOL(spl_kmem_cache_expire
);
57 module_param(spl_kmem_cache_expire
, uint
, 0644);
58 MODULE_PARM_DESC(spl_kmem_cache_expire
, "By age (0x1) or low memory (0x2)");
60 unsigned int spl_kmem_cache_obj_per_slab
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
61 module_param(spl_kmem_cache_obj_per_slab
, uint
, 0644);
62 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab
, "Number of objects per slab");
64 unsigned int spl_kmem_cache_obj_per_slab_min
= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN
;
65 module_param(spl_kmem_cache_obj_per_slab_min
, uint
, 0644);
66 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab_min
,
67 "Minimal number of objects per slab");
69 unsigned int spl_kmem_cache_max_size
= 32;
70 module_param(spl_kmem_cache_max_size
, uint
, 0644);
71 MODULE_PARM_DESC(spl_kmem_cache_max_size
, "Maximum size of slab in MB");
73 unsigned int spl_kmem_cache_slab_limit
= 0;
74 module_param(spl_kmem_cache_slab_limit
, uint
, 0644);
75 MODULE_PARM_DESC(spl_kmem_cache_slab_limit
,
76 "Objects less than N bytes use the Linux slab");
78 unsigned int spl_kmem_cache_kmem_limit
= (PAGE_SIZE
/ 4);
79 module_param(spl_kmem_cache_kmem_limit
, uint
, 0644);
80 MODULE_PARM_DESC(spl_kmem_cache_kmem_limit
,
81 "Objects less than N bytes use the kmalloc");
84 * The minimum amount of memory measured in pages to be free at all
85 * times on the system. This is similar to Linux's zone->pages_min
86 * multiplied by the number of zones and is sized based on that.
89 EXPORT_SYMBOL(minfree
);
92 * The desired amount of memory measured in pages to be free at all
93 * times on the system. This is similar to Linux's zone->pages_low
94 * multiplied by the number of zones and is sized based on that.
95 * Assuming all zones are being used roughly equally, when we drop
96 * below this threshold asynchronous page reclamation is triggered.
99 EXPORT_SYMBOL(desfree
);
102 * When above this amount of memory measures in pages the system is
103 * determined to have enough free memory. This is similar to Linux's
104 * zone->pages_high multiplied by the number of zones and is sized based
105 * on that. Assuming all zones are being used roughly equally, when
106 * asynchronous page reclamation reaches this threshold it stops.
108 pgcnt_t lotsfree
= 0;
109 EXPORT_SYMBOL(lotsfree
);
111 /* Unused always 0 in this implementation */
112 pgcnt_t needfree
= 0;
113 EXPORT_SYMBOL(needfree
);
115 pgcnt_t swapfs_minfree
= 0;
116 EXPORT_SYMBOL(swapfs_minfree
);
118 pgcnt_t swapfs_reserve
= 0;
119 EXPORT_SYMBOL(swapfs_reserve
);
121 vmem_t
*heap_arena
= NULL
;
122 EXPORT_SYMBOL(heap_arena
);
124 vmem_t
*zio_alloc_arena
= NULL
;
125 EXPORT_SYMBOL(zio_alloc_arena
);
127 vmem_t
*zio_arena
= NULL
;
128 EXPORT_SYMBOL(zio_arena
);
130 #ifndef HAVE_GET_VMALLOC_INFO
131 get_vmalloc_info_t get_vmalloc_info_fn
= SYMBOL_POISON
;
132 EXPORT_SYMBOL(get_vmalloc_info_fn
);
133 #endif /* HAVE_GET_VMALLOC_INFO */
135 #ifdef HAVE_PGDAT_HELPERS
136 # ifndef HAVE_FIRST_ONLINE_PGDAT
137 first_online_pgdat_t first_online_pgdat_fn
= SYMBOL_POISON
;
138 EXPORT_SYMBOL(first_online_pgdat_fn
);
139 # endif /* HAVE_FIRST_ONLINE_PGDAT */
141 # ifndef HAVE_NEXT_ONLINE_PGDAT
142 next_online_pgdat_t next_online_pgdat_fn
= SYMBOL_POISON
;
143 EXPORT_SYMBOL(next_online_pgdat_fn
);
144 # endif /* HAVE_NEXT_ONLINE_PGDAT */
146 # ifndef HAVE_NEXT_ZONE
147 next_zone_t next_zone_fn
= SYMBOL_POISON
;
148 EXPORT_SYMBOL(next_zone_fn
);
149 # endif /* HAVE_NEXT_ZONE */
151 #else /* HAVE_PGDAT_HELPERS */
153 # ifndef HAVE_PGDAT_LIST
154 struct pglist_data
*pgdat_list_addr
= SYMBOL_POISON
;
155 EXPORT_SYMBOL(pgdat_list_addr
);
156 # endif /* HAVE_PGDAT_LIST */
158 #endif /* HAVE_PGDAT_HELPERS */
160 #ifdef NEED_GET_ZONE_COUNTS
161 # ifndef HAVE_GET_ZONE_COUNTS
162 get_zone_counts_t get_zone_counts_fn
= SYMBOL_POISON
;
163 EXPORT_SYMBOL(get_zone_counts_fn
);
164 # endif /* HAVE_GET_ZONE_COUNTS */
167 spl_global_page_state(spl_zone_stat_item_t item
)
169 unsigned long active
;
170 unsigned long inactive
;
173 get_zone_counts(&active
, &inactive
, &free
);
175 case SPL_NR_FREE_PAGES
: return free
;
176 case SPL_NR_INACTIVE
: return inactive
;
177 case SPL_NR_ACTIVE
: return active
;
178 default: ASSERT(0); /* Unsupported */
184 # ifdef HAVE_GLOBAL_PAGE_STATE
186 spl_global_page_state(spl_zone_stat_item_t item
)
188 unsigned long pages
= 0;
191 case SPL_NR_FREE_PAGES
:
192 # ifdef HAVE_ZONE_STAT_ITEM_NR_FREE_PAGES
193 pages
+= global_page_state(NR_FREE_PAGES
);
196 case SPL_NR_INACTIVE
:
197 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE
198 pages
+= global_page_state(NR_INACTIVE
);
200 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_ANON
201 pages
+= global_page_state(NR_INACTIVE_ANON
);
203 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_FILE
204 pages
+= global_page_state(NR_INACTIVE_FILE
);
208 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE
209 pages
+= global_page_state(NR_ACTIVE
);
211 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_ANON
212 pages
+= global_page_state(NR_ACTIVE_ANON
);
214 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_FILE
215 pages
+= global_page_state(NR_ACTIVE_FILE
);
219 ASSERT(0); /* Unsupported */
225 # error "Both global_page_state() and get_zone_counts() unavailable"
226 # endif /* HAVE_GLOBAL_PAGE_STATE */
227 #endif /* NEED_GET_ZONE_COUNTS */
228 EXPORT_SYMBOL(spl_global_page_state
);
230 #ifndef HAVE_SHRINK_DCACHE_MEMORY
231 shrink_dcache_memory_t shrink_dcache_memory_fn
= SYMBOL_POISON
;
232 EXPORT_SYMBOL(shrink_dcache_memory_fn
);
233 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
235 #ifndef HAVE_SHRINK_ICACHE_MEMORY
236 shrink_icache_memory_t shrink_icache_memory_fn
= SYMBOL_POISON
;
237 EXPORT_SYMBOL(shrink_icache_memory_fn
);
238 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
241 spl_kmem_availrmem(void)
243 /* The amount of easily available memory */
244 return (spl_global_page_state(SPL_NR_FREE_PAGES
) +
245 spl_global_page_state(SPL_NR_INACTIVE
));
247 EXPORT_SYMBOL(spl_kmem_availrmem
);
250 vmem_size(vmem_t
*vmp
, int typemask
)
252 struct vmalloc_info vmi
;
256 ASSERT(typemask
& (VMEM_ALLOC
| VMEM_FREE
));
258 get_vmalloc_info(&vmi
);
259 if (typemask
& VMEM_ALLOC
)
260 size
+= (size_t)vmi
.used
;
262 if (typemask
& VMEM_FREE
)
263 size
+= (size_t)(VMALLOC_TOTAL
- vmi
.used
);
267 EXPORT_SYMBOL(vmem_size
);
274 EXPORT_SYMBOL(kmem_debugging
);
276 #ifndef HAVE_KVASPRINTF
277 /* Simplified asprintf. */
278 char *kvasprintf(gfp_t gfp
, const char *fmt
, va_list ap
)
285 len
= vsnprintf(NULL
, 0, fmt
, aq
);
288 p
= kmalloc(len
+1, gfp
);
292 vsnprintf(p
, len
+1, fmt
, ap
);
296 EXPORT_SYMBOL(kvasprintf
);
297 #endif /* HAVE_KVASPRINTF */
300 kmem_vasprintf(const char *fmt
, va_list ap
)
307 ptr
= kvasprintf(GFP_KERNEL
, fmt
, aq
);
309 } while (ptr
== NULL
);
313 EXPORT_SYMBOL(kmem_vasprintf
);
316 kmem_asprintf(const char *fmt
, ...)
323 ptr
= kvasprintf(GFP_KERNEL
, fmt
, ap
);
325 } while (ptr
== NULL
);
329 EXPORT_SYMBOL(kmem_asprintf
);
332 __strdup(const char *str
, int flags
)
338 ptr
= kmalloc_nofail(n
+ 1, flags
);
340 memcpy(ptr
, str
, n
+ 1);
346 strdup(const char *str
)
348 return __strdup(str
, KM_SLEEP
);
350 EXPORT_SYMBOL(strdup
);
357 EXPORT_SYMBOL(strfree
);
360 * Memory allocation interfaces and debugging for basic kmem_*
361 * and vmem_* style memory allocation. When DEBUG_KMEM is enabled
362 * the SPL will keep track of the total memory allocated, and
363 * report any memory leaked when the module is unloaded.
367 /* Shim layer memory accounting */
368 # ifdef HAVE_ATOMIC64_T
369 atomic64_t kmem_alloc_used
= ATOMIC64_INIT(0);
370 unsigned long long kmem_alloc_max
= 0;
371 atomic64_t vmem_alloc_used
= ATOMIC64_INIT(0);
372 unsigned long long vmem_alloc_max
= 0;
373 # else /* HAVE_ATOMIC64_T */
374 atomic_t kmem_alloc_used
= ATOMIC_INIT(0);
375 unsigned long long kmem_alloc_max
= 0;
376 atomic_t vmem_alloc_used
= ATOMIC_INIT(0);
377 unsigned long long vmem_alloc_max
= 0;
378 # endif /* HAVE_ATOMIC64_T */
380 EXPORT_SYMBOL(kmem_alloc_used
);
381 EXPORT_SYMBOL(kmem_alloc_max
);
382 EXPORT_SYMBOL(vmem_alloc_used
);
383 EXPORT_SYMBOL(vmem_alloc_max
);
385 /* When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked
386 * but also the location of every alloc and free. When the SPL module is
387 * unloaded a list of all leaked addresses and where they were allocated
388 * will be dumped to the console. Enabling this feature has a significant
389 * impact on performance but it makes finding memory leaks straight forward.
391 * Not surprisingly with debugging enabled the xmem_locks are very highly
392 * contended particularly on xfree(). If we want to run with this detailed
393 * debugging enabled for anything other than debugging we need to minimize
394 * the contention by moving to a lock per xmem_table entry model.
396 # ifdef DEBUG_KMEM_TRACKING
398 # define KMEM_HASH_BITS 10
399 # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS)
401 # define VMEM_HASH_BITS 10
402 # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS)
404 typedef struct kmem_debug
{
405 struct hlist_node kd_hlist
; /* Hash node linkage */
406 struct list_head kd_list
; /* List of all allocations */
407 void *kd_addr
; /* Allocation pointer */
408 size_t kd_size
; /* Allocation size */
409 const char *kd_func
; /* Allocation function */
410 int kd_line
; /* Allocation line */
413 spinlock_t kmem_lock
;
414 struct hlist_head kmem_table
[KMEM_TABLE_SIZE
];
415 struct list_head kmem_list
;
417 spinlock_t vmem_lock
;
418 struct hlist_head vmem_table
[VMEM_TABLE_SIZE
];
419 struct list_head vmem_list
;
421 EXPORT_SYMBOL(kmem_lock
);
422 EXPORT_SYMBOL(kmem_table
);
423 EXPORT_SYMBOL(kmem_list
);
425 EXPORT_SYMBOL(vmem_lock
);
426 EXPORT_SYMBOL(vmem_table
);
427 EXPORT_SYMBOL(vmem_list
);
429 static kmem_debug_t
*
430 kmem_del_init(spinlock_t
*lock
, struct hlist_head
*table
, int bits
, const void *addr
)
432 struct hlist_head
*head
;
433 struct hlist_node
*node
;
434 struct kmem_debug
*p
;
438 spin_lock_irqsave(lock
, flags
);
440 head
= &table
[hash_ptr((void *)addr
, bits
)];
441 hlist_for_each(node
, head
) {
442 p
= list_entry(node
, struct kmem_debug
, kd_hlist
);
443 if (p
->kd_addr
== addr
) {
444 hlist_del_init(&p
->kd_hlist
);
445 list_del_init(&p
->kd_list
);
446 spin_unlock_irqrestore(lock
, flags
);
451 spin_unlock_irqrestore(lock
, flags
);
457 kmem_alloc_track(size_t size
, int flags
, const char *func
, int line
,
458 int node_alloc
, int node
)
462 unsigned long irq_flags
;
465 /* Function may be called with KM_NOSLEEP so failure is possible */
466 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
467 flags
& ~__GFP_ZERO
);
469 if (unlikely(dptr
== NULL
)) {
470 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
471 "kmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
472 sizeof(kmem_debug_t
), flags
, func
, line
,
473 kmem_alloc_used_read(), kmem_alloc_max
);
476 * Marked unlikely because we should never be doing this,
477 * we tolerate to up 2 pages but a single page is best.
479 if (unlikely((size
> PAGE_SIZE
*2) && !(flags
& KM_NODEBUG
))) {
480 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "large "
481 "kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
482 (unsigned long long) size
, flags
, func
, line
,
483 kmem_alloc_used_read(), kmem_alloc_max
);
484 spl_debug_dumpstack(NULL
);
488 * We use __strdup() below because the string pointed to by
489 * __FUNCTION__ might not be available by the time we want
490 * to print it since the module might have been unloaded.
491 * This can only fail in the KM_NOSLEEP case.
493 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
494 if (unlikely(dptr
->kd_func
== NULL
)) {
496 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
497 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
498 func
, line
, kmem_alloc_used_read(), kmem_alloc_max
);
502 /* Use the correct allocator */
504 ASSERT(!(flags
& __GFP_ZERO
));
505 ptr
= kmalloc_node_nofail(size
, flags
, node
);
506 } else if (flags
& __GFP_ZERO
) {
507 ptr
= kzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
509 ptr
= kmalloc_nofail(size
, flags
);
512 if (unlikely(ptr
== NULL
)) {
513 kfree(dptr
->kd_func
);
515 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "kmem_alloc"
516 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
517 (unsigned long long) size
, flags
, func
, line
,
518 kmem_alloc_used_read(), kmem_alloc_max
);
522 kmem_alloc_used_add(size
);
523 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
524 kmem_alloc_max
= kmem_alloc_used_read();
526 INIT_HLIST_NODE(&dptr
->kd_hlist
);
527 INIT_LIST_HEAD(&dptr
->kd_list
);
530 dptr
->kd_size
= size
;
531 dptr
->kd_line
= line
;
533 spin_lock_irqsave(&kmem_lock
, irq_flags
);
534 hlist_add_head(&dptr
->kd_hlist
,
535 &kmem_table
[hash_ptr(ptr
, KMEM_HASH_BITS
)]);
536 list_add_tail(&dptr
->kd_list
, &kmem_list
);
537 spin_unlock_irqrestore(&kmem_lock
, irq_flags
);
539 SDEBUG_LIMIT(SD_INFO
,
540 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
541 (unsigned long long) size
, flags
, func
, line
, ptr
,
542 kmem_alloc_used_read(), kmem_alloc_max
);
547 EXPORT_SYMBOL(kmem_alloc_track
);
550 kmem_free_track(const void *ptr
, size_t size
)
555 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
556 (unsigned long long) size
);
558 dptr
= kmem_del_init(&kmem_lock
, kmem_table
, KMEM_HASH_BITS
, ptr
);
560 /* Must exist in hash due to kmem_alloc() */
563 /* Size must match */
564 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
565 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
566 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
568 kmem_alloc_used_sub(size
);
569 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
570 (unsigned long long) size
, kmem_alloc_used_read(),
573 kfree(dptr
->kd_func
);
575 memset((void *)dptr
, 0x5a, sizeof(kmem_debug_t
));
578 memset((void *)ptr
, 0x5a, size
);
583 EXPORT_SYMBOL(kmem_free_track
);
586 vmem_alloc_track(size_t size
, int flags
, const char *func
, int line
)
590 unsigned long irq_flags
;
593 ASSERT(flags
& KM_SLEEP
);
595 /* Function may be called with KM_NOSLEEP so failure is possible */
596 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
597 flags
& ~__GFP_ZERO
);
598 if (unlikely(dptr
== NULL
)) {
599 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
600 "vmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
601 sizeof(kmem_debug_t
), flags
, func
, line
,
602 vmem_alloc_used_read(), vmem_alloc_max
);
605 * We use __strdup() below because the string pointed to by
606 * __FUNCTION__ might not be available by the time we want
607 * to print it, since the module might have been unloaded.
608 * This can never fail because we have already asserted
609 * that flags is KM_SLEEP.
611 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
612 if (unlikely(dptr
->kd_func
== NULL
)) {
614 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
615 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
616 func
, line
, vmem_alloc_used_read(), vmem_alloc_max
);
620 /* Use the correct allocator */
621 if (flags
& __GFP_ZERO
) {
622 ptr
= vzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
624 ptr
= vmalloc_nofail(size
, flags
);
627 if (unlikely(ptr
== NULL
)) {
628 kfree(dptr
->kd_func
);
630 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "vmem_alloc"
631 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
632 (unsigned long long) size
, flags
, func
, line
,
633 vmem_alloc_used_read(), vmem_alloc_max
);
637 vmem_alloc_used_add(size
);
638 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
639 vmem_alloc_max
= vmem_alloc_used_read();
641 INIT_HLIST_NODE(&dptr
->kd_hlist
);
642 INIT_LIST_HEAD(&dptr
->kd_list
);
645 dptr
->kd_size
= size
;
646 dptr
->kd_line
= line
;
648 spin_lock_irqsave(&vmem_lock
, irq_flags
);
649 hlist_add_head(&dptr
->kd_hlist
,
650 &vmem_table
[hash_ptr(ptr
, VMEM_HASH_BITS
)]);
651 list_add_tail(&dptr
->kd_list
, &vmem_list
);
652 spin_unlock_irqrestore(&vmem_lock
, irq_flags
);
654 SDEBUG_LIMIT(SD_INFO
,
655 "vmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
656 (unsigned long long) size
, flags
, func
, line
,
657 ptr
, vmem_alloc_used_read(), vmem_alloc_max
);
662 EXPORT_SYMBOL(vmem_alloc_track
);
665 vmem_free_track(const void *ptr
, size_t size
)
670 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
671 (unsigned long long) size
);
673 dptr
= kmem_del_init(&vmem_lock
, vmem_table
, VMEM_HASH_BITS
, ptr
);
675 /* Must exist in hash due to vmem_alloc() */
678 /* Size must match */
679 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
680 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
681 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
683 vmem_alloc_used_sub(size
);
684 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
685 (unsigned long long) size
, vmem_alloc_used_read(),
688 kfree(dptr
->kd_func
);
690 memset((void *)dptr
, 0x5a, sizeof(kmem_debug_t
));
693 memset((void *)ptr
, 0x5a, size
);
698 EXPORT_SYMBOL(vmem_free_track
);
700 # else /* DEBUG_KMEM_TRACKING */
703 kmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
,
704 int node_alloc
, int node
)
710 * Marked unlikely because we should never be doing this,
711 * we tolerate to up 2 pages but a single page is best.
713 if (unlikely((size
> PAGE_SIZE
* 2) && !(flags
& KM_NODEBUG
))) {
714 SDEBUG(SD_CONSOLE
| SD_WARNING
,
715 "large kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
716 (unsigned long long) size
, flags
, func
, line
,
717 kmem_alloc_used_read(), kmem_alloc_max
);
721 /* Use the correct allocator */
723 ASSERT(!(flags
& __GFP_ZERO
));
724 ptr
= kmalloc_node_nofail(size
, flags
, node
);
725 } else if (flags
& __GFP_ZERO
) {
726 ptr
= kzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
728 ptr
= kmalloc_nofail(size
, flags
);
731 if (unlikely(ptr
== NULL
)) {
732 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
733 "kmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
734 (unsigned long long) size
, flags
, func
, line
,
735 kmem_alloc_used_read(), kmem_alloc_max
);
737 kmem_alloc_used_add(size
);
738 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
739 kmem_alloc_max
= kmem_alloc_used_read();
741 SDEBUG_LIMIT(SD_INFO
,
742 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
743 (unsigned long long) size
, flags
, func
, line
, ptr
,
744 kmem_alloc_used_read(), kmem_alloc_max
);
749 EXPORT_SYMBOL(kmem_alloc_debug
);
752 kmem_free_debug(const void *ptr
, size_t size
)
756 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
757 (unsigned long long) size
);
759 kmem_alloc_used_sub(size
);
760 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
761 (unsigned long long) size
, kmem_alloc_used_read(),
767 EXPORT_SYMBOL(kmem_free_debug
);
770 vmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
)
775 ASSERT(flags
& KM_SLEEP
);
777 /* Use the correct allocator */
778 if (flags
& __GFP_ZERO
) {
779 ptr
= vzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
781 ptr
= vmalloc_nofail(size
, flags
);
784 if (unlikely(ptr
== NULL
)) {
785 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
786 "vmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
787 (unsigned long long) size
, flags
, func
, line
,
788 vmem_alloc_used_read(), vmem_alloc_max
);
790 vmem_alloc_used_add(size
);
791 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
792 vmem_alloc_max
= vmem_alloc_used_read();
794 SDEBUG_LIMIT(SD_INFO
, "vmem_alloc(%llu, 0x%x) = %p "
795 "(%lld/%llu)\n", (unsigned long long) size
, flags
, ptr
,
796 vmem_alloc_used_read(), vmem_alloc_max
);
801 EXPORT_SYMBOL(vmem_alloc_debug
);
804 vmem_free_debug(const void *ptr
, size_t size
)
808 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
809 (unsigned long long) size
);
811 vmem_alloc_used_sub(size
);
812 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
813 (unsigned long long) size
, vmem_alloc_used_read(),
819 EXPORT_SYMBOL(vmem_free_debug
);
821 # endif /* DEBUG_KMEM_TRACKING */
822 #endif /* DEBUG_KMEM */
825 * Slab allocation interfaces
827 * While the Linux slab implementation was inspired by the Solaris
828 * implementation I cannot use it to emulate the Solaris APIs. I
829 * require two features which are not provided by the Linux slab.
831 * 1) Constructors AND destructors. Recent versions of the Linux
832 * kernel have removed support for destructors. This is a deal
833 * breaker for the SPL which contains particularly expensive
834 * initializers for mutex's, condition variables, etc. We also
835 * require a minimal level of cleanup for these data types unlike
836 * many Linux data type which do need to be explicitly destroyed.
838 * 2) Virtual address space backed slab. Callers of the Solaris slab
839 * expect it to work well for both small are very large allocations.
840 * Because of memory fragmentation the Linux slab which is backed
841 * by kmalloc'ed memory performs very badly when confronted with
842 * large numbers of large allocations. Basing the slab on the
843 * virtual address space removes the need for contiguous pages
844 * and greatly improve performance for large allocations.
846 * For these reasons, the SPL has its own slab implementation with
847 * the needed features. It is not as highly optimized as either the
848 * Solaris or Linux slabs, but it should get me most of what is
849 * needed until it can be optimized or obsoleted by another approach.
851 * One serious concern I do have about this method is the relatively
852 * small virtual address space on 32bit arches. This will seriously
853 * constrain the size of the slab caches and their performance.
855 * XXX: Improve the partial slab list by carefully maintaining a
856 * strict ordering of fullest to emptiest slabs based on
857 * the slab reference count. This guarantees the when freeing
858 * slabs back to the system we need only linearly traverse the
859 * last N slabs in the list to discover all the freeable slabs.
861 * XXX: NUMA awareness for optionally allocating memory close to a
862 * particular core. This can be advantageous if you know the slab
863 * object will be short lived and primarily accessed from one core.
865 * XXX: Slab coloring may also yield performance improvements and would
866 * be desirable to implement.
869 struct list_head spl_kmem_cache_list
; /* List of caches */
870 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
871 taskq_t
*spl_kmem_cache_taskq
; /* Task queue for ageing / reclaim */
873 static void spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
);
875 SPL_SHRINKER_CALLBACK_FWD_DECLARE(spl_kmem_cache_generic_shrinker
);
876 SPL_SHRINKER_DECLARE(spl_kmem_cache_shrinker
,
877 spl_kmem_cache_generic_shrinker
, KMC_DEFAULT_SEEKS
);
880 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
886 if (skc
->skc_flags
& KMC_KMEM
)
887 ptr
= (void *)__get_free_pages(flags
| __GFP_COMP
,
890 ptr
= __vmalloc(size
, flags
| __GFP_HIGHMEM
, PAGE_KERNEL
);
892 /* Resulting allocated memory will be page aligned */
893 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
899 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
901 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
905 * The Linux direct reclaim path uses this out of band value to
906 * determine if forward progress is being made. Normally this is
907 * incremented by kmem_freepages() which is part of the various
908 * Linux slab implementations. However, since we are using none
909 * of that infrastructure we are responsible for incrementing it.
911 if (current
->reclaim_state
)
912 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
914 if (skc
->skc_flags
& KMC_KMEM
)
915 free_pages((unsigned long)ptr
, get_order(size
));
921 * Required space for each aligned sks.
923 static inline uint32_t
924 spl_sks_size(spl_kmem_cache_t
*skc
)
926 return P2ROUNDUP_TYPED(sizeof(spl_kmem_slab_t
),
927 skc
->skc_obj_align
, uint32_t);
931 * Required space for each aligned object.
933 static inline uint32_t
934 spl_obj_size(spl_kmem_cache_t
*skc
)
936 uint32_t align
= skc
->skc_obj_align
;
938 return P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
939 P2ROUNDUP_TYPED(sizeof(spl_kmem_obj_t
), align
, uint32_t);
943 * Lookup the spl_kmem_object_t for an object given that object.
945 static inline spl_kmem_obj_t
*
946 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
948 return obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
949 skc
->skc_obj_align
, uint32_t);
953 * Required space for each offslab object taking in to account alignment
954 * restrictions and the power-of-two requirement of kv_alloc().
956 static inline uint32_t
957 spl_offslab_size(spl_kmem_cache_t
*skc
)
959 return 1UL << (highbit(spl_obj_size(skc
)) + 1);
963 * It's important that we pack the spl_kmem_obj_t structure and the
964 * actual objects in to one large address space to minimize the number
965 * of calls to the allocator. It is far better to do a few large
966 * allocations and then subdivide it ourselves. Now which allocator
967 * we use requires balancing a few trade offs.
969 * For small objects we use kmem_alloc() because as long as you are
970 * only requesting a small number of pages (ideally just one) its cheap.
971 * However, when you start requesting multiple pages with kmem_alloc()
972 * it gets increasingly expensive since it requires contiguous pages.
973 * For this reason we shift to vmem_alloc() for slabs of large objects
974 * which removes the need for contiguous pages. We do not use
975 * vmem_alloc() in all cases because there is significant locking
976 * overhead in __get_vm_area_node(). This function takes a single
977 * global lock when acquiring an available virtual address range which
978 * serializes all vmem_alloc()'s for all slab caches. Using slightly
979 * different allocation functions for small and large objects should
980 * give us the best of both worlds.
982 * KMC_ONSLAB KMC_OFFSLAB
984 * +------------------------+ +-----------------+
985 * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+
986 * | skc_obj_size <-+ | | +-----------------+ | |
987 * | spl_kmem_obj_t | | | |
988 * | skc_obj_size <---+ | +-----------------+ | |
989 * | spl_kmem_obj_t | | | skc_obj_size | <-+ |
990 * | ... v | | spl_kmem_obj_t | |
991 * +------------------------+ +-----------------+ v
993 static spl_kmem_slab_t
*
994 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
996 spl_kmem_slab_t
*sks
;
997 spl_kmem_obj_t
*sko
, *n
;
999 uint32_t obj_size
, offslab_size
= 0;
1002 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
1006 sks
= (spl_kmem_slab_t
*)base
;
1007 sks
->sks_magic
= SKS_MAGIC
;
1008 sks
->sks_objs
= skc
->skc_slab_objs
;
1009 sks
->sks_age
= jiffies
;
1010 sks
->sks_cache
= skc
;
1011 INIT_LIST_HEAD(&sks
->sks_list
);
1012 INIT_LIST_HEAD(&sks
->sks_free_list
);
1014 obj_size
= spl_obj_size(skc
);
1016 if (skc
->skc_flags
& KMC_OFFSLAB
)
1017 offslab_size
= spl_offslab_size(skc
);
1019 for (i
= 0; i
< sks
->sks_objs
; i
++) {
1020 if (skc
->skc_flags
& KMC_OFFSLAB
) {
1021 obj
= kv_alloc(skc
, offslab_size
, flags
);
1023 SGOTO(out
, rc
= -ENOMEM
);
1025 obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
1028 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
1029 sko
= spl_sko_from_obj(skc
, obj
);
1030 sko
->sko_addr
= obj
;
1031 sko
->sko_magic
= SKO_MAGIC
;
1032 sko
->sko_slab
= sks
;
1033 INIT_LIST_HEAD(&sko
->sko_list
);
1034 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
1037 list_for_each_entry(sko
, &sks
->sks_free_list
, sko_list
)
1039 skc
->skc_ctor(sko
->sko_addr
, skc
->skc_private
, flags
);
1042 if (skc
->skc_flags
& KMC_OFFSLAB
)
1043 list_for_each_entry_safe(sko
, n
, &sks
->sks_free_list
,
1045 kv_free(skc
, sko
->sko_addr
, offslab_size
);
1047 kv_free(skc
, base
, skc
->skc_slab_size
);
1055 * Remove a slab from complete or partial list, it must be called with
1056 * the 'skc->skc_lock' held but the actual free must be performed
1057 * outside the lock to prevent deadlocking on vmem addresses.
1060 spl_slab_free(spl_kmem_slab_t
*sks
,
1061 struct list_head
*sks_list
, struct list_head
*sko_list
)
1063 spl_kmem_cache_t
*skc
;
1066 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1067 ASSERT(sks
->sks_ref
== 0);
1069 skc
= sks
->sks_cache
;
1070 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1071 ASSERT(spin_is_locked(&skc
->skc_lock
));
1074 * Update slab/objects counters in the cache, then remove the
1075 * slab from the skc->skc_partial_list. Finally add the slab
1076 * and all its objects in to the private work lists where the
1077 * destructors will be called and the memory freed to the system.
1079 skc
->skc_obj_total
-= sks
->sks_objs
;
1080 skc
->skc_slab_total
--;
1081 list_del(&sks
->sks_list
);
1082 list_add(&sks
->sks_list
, sks_list
);
1083 list_splice_init(&sks
->sks_free_list
, sko_list
);
1089 * Traverses all the partial slabs attached to a cache and free those
1090 * which which are currently empty, and have not been touched for
1091 * skc_delay seconds to avoid thrashing. The count argument is
1092 * passed to optionally cap the number of slabs reclaimed, a count
1093 * of zero means try and reclaim everything. When flag is set we
1094 * always free an available slab regardless of age.
1097 spl_slab_reclaim(spl_kmem_cache_t
*skc
, int count
, int flag
)
1099 spl_kmem_slab_t
*sks
, *m
;
1100 spl_kmem_obj_t
*sko
, *n
;
1101 LIST_HEAD(sks_list
);
1102 LIST_HEAD(sko_list
);
1108 * Move empty slabs and objects which have not been touched in
1109 * skc_delay seconds on to private lists to be freed outside
1110 * the spin lock. This delay time is important to avoid thrashing
1111 * however when flag is set the delay will not be used.
1113 spin_lock(&skc
->skc_lock
);
1114 list_for_each_entry_safe_reverse(sks
,m
,&skc
->skc_partial_list
,sks_list
){
1116 * All empty slabs are at the end of skc->skc_partial_list,
1117 * therefore once a non-empty slab is found we can stop
1118 * scanning. Additionally, stop when reaching the target
1119 * reclaim 'count' if a non-zero threshold is given.
1121 if ((sks
->sks_ref
> 0) || (count
&& i
>= count
))
1124 if (time_after(jiffies
,sks
->sks_age
+skc
->skc_delay
*HZ
)||flag
) {
1125 spl_slab_free(sks
, &sks_list
, &sko_list
);
1129 spin_unlock(&skc
->skc_lock
);
1132 * The following two loops ensure all the object destructors are
1133 * run, any offslab objects are freed, and the slabs themselves
1134 * are freed. This is all done outside the skc->skc_lock since
1135 * this allows the destructor to sleep, and allows us to perform
1136 * a conditional reschedule when a freeing a large number of
1137 * objects and slabs back to the system.
1139 if (skc
->skc_flags
& KMC_OFFSLAB
)
1140 size
= spl_offslab_size(skc
);
1142 list_for_each_entry_safe(sko
, n
, &sko_list
, sko_list
) {
1143 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1146 skc
->skc_dtor(sko
->sko_addr
, skc
->skc_private
);
1148 if (skc
->skc_flags
& KMC_OFFSLAB
)
1149 kv_free(skc
, sko
->sko_addr
, size
);
1152 list_for_each_entry_safe(sks
, m
, &sks_list
, sks_list
) {
1153 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1154 kv_free(skc
, sks
, skc
->skc_slab_size
);
1160 static spl_kmem_emergency_t
*
1161 spl_emergency_search(struct rb_root
*root
, void *obj
)
1163 struct rb_node
*node
= root
->rb_node
;
1164 spl_kmem_emergency_t
*ske
;
1165 unsigned long address
= (unsigned long)obj
;
1168 ske
= container_of(node
, spl_kmem_emergency_t
, ske_node
);
1170 if (address
< (unsigned long)ske
->ske_obj
)
1171 node
= node
->rb_left
;
1172 else if (address
> (unsigned long)ske
->ske_obj
)
1173 node
= node
->rb_right
;
1182 spl_emergency_insert(struct rb_root
*root
, spl_kmem_emergency_t
*ske
)
1184 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
1185 spl_kmem_emergency_t
*ske_tmp
;
1186 unsigned long address
= (unsigned long)ske
->ske_obj
;
1189 ske_tmp
= container_of(*new, spl_kmem_emergency_t
, ske_node
);
1192 if (address
< (unsigned long)ske_tmp
->ske_obj
)
1193 new = &((*new)->rb_left
);
1194 else if (address
> (unsigned long)ske_tmp
->ske_obj
)
1195 new = &((*new)->rb_right
);
1200 rb_link_node(&ske
->ske_node
, parent
, new);
1201 rb_insert_color(&ske
->ske_node
, root
);
1207 * Allocate a single emergency object and track it in a red black tree.
1210 spl_emergency_alloc(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1212 spl_kmem_emergency_t
*ske
;
1216 /* Last chance use a partial slab if one now exists */
1217 spin_lock(&skc
->skc_lock
);
1218 empty
= list_empty(&skc
->skc_partial_list
);
1219 spin_unlock(&skc
->skc_lock
);
1223 ske
= kmalloc(sizeof(*ske
), flags
);
1227 ske
->ske_obj
= kmalloc(skc
->skc_obj_size
, flags
);
1228 if (ske
->ske_obj
== NULL
) {
1233 spin_lock(&skc
->skc_lock
);
1234 empty
= spl_emergency_insert(&skc
->skc_emergency_tree
, ske
);
1235 if (likely(empty
)) {
1236 skc
->skc_obj_total
++;
1237 skc
->skc_obj_emergency
++;
1238 if (skc
->skc_obj_emergency
> skc
->skc_obj_emergency_max
)
1239 skc
->skc_obj_emergency_max
= skc
->skc_obj_emergency
;
1241 spin_unlock(&skc
->skc_lock
);
1243 if (unlikely(!empty
)) {
1244 kfree(ske
->ske_obj
);
1250 skc
->skc_ctor(ske
->ske_obj
, skc
->skc_private
, flags
);
1252 *obj
= ske
->ske_obj
;
1258 * Locate the passed object in the red black tree and free it.
1261 spl_emergency_free(spl_kmem_cache_t
*skc
, void *obj
)
1263 spl_kmem_emergency_t
*ske
;
1266 spin_lock(&skc
->skc_lock
);
1267 ske
= spl_emergency_search(&skc
->skc_emergency_tree
, obj
);
1269 rb_erase(&ske
->ske_node
, &skc
->skc_emergency_tree
);
1270 skc
->skc_obj_emergency
--;
1271 skc
->skc_obj_total
--;
1273 spin_unlock(&skc
->skc_lock
);
1275 if (unlikely(ske
== NULL
))
1279 skc
->skc_dtor(ske
->ske_obj
, skc
->skc_private
);
1281 kfree(ske
->ske_obj
);
1288 * Release objects from the per-cpu magazine back to their slab. The flush
1289 * argument contains the max number of entries to remove from the magazine.
1292 __spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
1294 int i
, count
= MIN(flush
, skm
->skm_avail
);
1297 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1298 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1299 ASSERT(spin_is_locked(&skc
->skc_lock
));
1301 for (i
= 0; i
< count
; i
++)
1302 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
1304 skm
->skm_avail
-= count
;
1305 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
1306 sizeof(void *) * skm
->skm_avail
);
1312 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
1314 spin_lock(&skc
->skc_lock
);
1315 __spl_cache_flush(skc
, skm
, flush
);
1316 spin_unlock(&skc
->skc_lock
);
1320 spl_magazine_age(void *data
)
1322 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1323 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
1325 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1326 ASSERT(skm
->skm_cpu
== smp_processor_id());
1327 ASSERT(irqs_disabled());
1329 /* There are no available objects or they are too young to age out */
1330 if ((skm
->skm_avail
== 0) ||
1331 time_before(jiffies
, skm
->skm_age
+ skc
->skc_delay
* HZ
))
1335 * Because we're executing in interrupt context we may have
1336 * interrupted the holder of this lock. To avoid a potential
1337 * deadlock return if the lock is contended.
1339 if (!spin_trylock(&skc
->skc_lock
))
1342 __spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1343 spin_unlock(&skc
->skc_lock
);
1347 * Called regularly to keep a downward pressure on the cache.
1349 * Objects older than skc->skc_delay seconds in the per-cpu magazines will
1350 * be returned to the caches. This is done to prevent idle magazines from
1351 * holding memory which could be better used elsewhere. The delay is
1352 * present to prevent thrashing the magazine.
1354 * The newly released objects may result in empty partial slabs. Those
1355 * slabs should be released to the system. Otherwise moving the objects
1356 * out of the magazines is just wasted work.
1359 spl_cache_age(void *data
)
1361 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1364 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1366 /* Dynamically disabled at run time */
1367 if (!(spl_kmem_cache_expire
& KMC_EXPIRE_AGE
))
1370 atomic_inc(&skc
->skc_ref
);
1372 if (!(skc
->skc_flags
& KMC_NOMAGAZINE
))
1373 spl_on_each_cpu(spl_magazine_age
, skc
, 1);
1375 spl_slab_reclaim(skc
, skc
->skc_reap
, 0);
1377 while (!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && !id
) {
1378 id
= taskq_dispatch_delay(
1379 spl_kmem_cache_taskq
, spl_cache_age
, skc
, TQ_SLEEP
,
1380 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1382 /* Destroy issued after dispatch immediately cancel it */
1383 if (test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && id
)
1384 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1387 spin_lock(&skc
->skc_lock
);
1388 skc
->skc_taskqid
= id
;
1389 spin_unlock(&skc
->skc_lock
);
1391 atomic_dec(&skc
->skc_ref
);
1395 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
1396 * When on-slab we want to target spl_kmem_cache_obj_per_slab. However,
1397 * for very small objects we may end up with more than this so as not
1398 * to waste space in the minimal allocation of a single page. Also for
1399 * very large objects we may use as few as spl_kmem_cache_obj_per_slab_min,
1400 * lower than this and we will fail.
1403 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
1405 uint32_t sks_size
, obj_size
, max_size
;
1407 if (skc
->skc_flags
& KMC_OFFSLAB
) {
1408 *objs
= spl_kmem_cache_obj_per_slab
;
1409 *size
= P2ROUNDUP(sizeof(spl_kmem_slab_t
), PAGE_SIZE
);
1412 sks_size
= spl_sks_size(skc
);
1413 obj_size
= spl_obj_size(skc
);
1415 if (skc
->skc_flags
& KMC_KMEM
)
1416 max_size
= ((uint32_t)1 << (MAX_ORDER
-3)) * PAGE_SIZE
;
1418 max_size
= (spl_kmem_cache_max_size
* 1024 * 1024);
1420 /* Power of two sized slab */
1421 for (*size
= PAGE_SIZE
; *size
<= max_size
; *size
*= 2) {
1422 *objs
= (*size
- sks_size
) / obj_size
;
1423 if (*objs
>= spl_kmem_cache_obj_per_slab
)
1428 * Unable to satisfy target objects per slab, fall back to
1429 * allocating a maximally sized slab and assuming it can
1430 * contain the minimum objects count use it. If not fail.
1433 *objs
= (*size
- sks_size
) / obj_size
;
1434 if (*objs
>= (spl_kmem_cache_obj_per_slab_min
))
1442 * Make a guess at reasonable per-cpu magazine size based on the size of
1443 * each object and the cost of caching N of them in each magazine. Long
1444 * term this should really adapt based on an observed usage heuristic.
1447 spl_magazine_size(spl_kmem_cache_t
*skc
)
1449 uint32_t obj_size
= spl_obj_size(skc
);
1453 /* Per-magazine sizes below assume a 4Kib page size */
1454 if (obj_size
> (PAGE_SIZE
* 256))
1455 size
= 4; /* Minimum 4Mib per-magazine */
1456 else if (obj_size
> (PAGE_SIZE
* 32))
1457 size
= 16; /* Minimum 2Mib per-magazine */
1458 else if (obj_size
> (PAGE_SIZE
))
1459 size
= 64; /* Minimum 256Kib per-magazine */
1460 else if (obj_size
> (PAGE_SIZE
/ 4))
1461 size
= 128; /* Minimum 128Kib per-magazine */
1469 * Allocate a per-cpu magazine to associate with a specific core.
1471 static spl_kmem_magazine_t
*
1472 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
1474 spl_kmem_magazine_t
*skm
;
1475 int size
= sizeof(spl_kmem_magazine_t
) +
1476 sizeof(void *) * skc
->skc_mag_size
;
1479 skm
= kmem_alloc_node(size
, KM_SLEEP
, cpu_to_node(cpu
));
1481 skm
->skm_magic
= SKM_MAGIC
;
1483 skm
->skm_size
= skc
->skc_mag_size
;
1484 skm
->skm_refill
= skc
->skc_mag_refill
;
1485 skm
->skm_cache
= skc
;
1486 skm
->skm_age
= jiffies
;
1494 * Free a per-cpu magazine associated with a specific core.
1497 spl_magazine_free(spl_kmem_magazine_t
*skm
)
1499 int size
= sizeof(spl_kmem_magazine_t
) +
1500 sizeof(void *) * skm
->skm_size
;
1503 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1504 ASSERT(skm
->skm_avail
== 0);
1506 kmem_free(skm
, size
);
1511 * Create all pre-cpu magazines of reasonable sizes.
1514 spl_magazine_create(spl_kmem_cache_t
*skc
)
1519 if (skc
->skc_flags
& KMC_NOMAGAZINE
)
1522 skc
->skc_mag_size
= spl_magazine_size(skc
);
1523 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
1525 for_each_online_cpu(i
) {
1526 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
1527 if (!skc
->skc_mag
[i
]) {
1528 for (i
--; i
>= 0; i
--)
1529 spl_magazine_free(skc
->skc_mag
[i
]);
1539 * Destroy all pre-cpu magazines.
1542 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
1544 spl_kmem_magazine_t
*skm
;
1548 if (skc
->skc_flags
& KMC_NOMAGAZINE
) {
1553 for_each_online_cpu(i
) {
1554 skm
= skc
->skc_mag
[i
];
1555 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1556 spl_magazine_free(skm
);
1563 * Create a object cache based on the following arguments:
1565 * size cache object size
1566 * align cache object alignment
1567 * ctor cache object constructor
1568 * dtor cache object destructor
1569 * reclaim cache object reclaim
1570 * priv cache private data for ctor/dtor/reclaim
1571 * vmp unused must be NULL
1573 * KMC_NOTOUCH Disable cache object aging (unsupported)
1574 * KMC_NODEBUG Disable debugging (unsupported)
1575 * KMC_NOHASH Disable hashing (unsupported)
1576 * KMC_QCACHE Disable qcache (unsupported)
1577 * KMC_NOMAGAZINE Enabled for kmem/vmem, Disabled for Linux slab
1578 * KMC_KMEM Force kmem backed cache
1579 * KMC_VMEM Force vmem backed cache
1580 * KMC_SLAB Force Linux slab backed cache
1581 * KMC_OFFSLAB Locate objects off the slab
1584 spl_kmem_cache_create(char *name
, size_t size
, size_t align
,
1585 spl_kmem_ctor_t ctor
,
1586 spl_kmem_dtor_t dtor
,
1587 spl_kmem_reclaim_t reclaim
,
1588 void *priv
, void *vmp
, int flags
)
1590 spl_kmem_cache_t
*skc
;
1594 ASSERTF(!(flags
& KMC_NOMAGAZINE
), "Bad KMC_NOMAGAZINE (%x)\n", flags
);
1595 ASSERTF(!(flags
& KMC_NOHASH
), "Bad KMC_NOHASH (%x)\n", flags
);
1596 ASSERTF(!(flags
& KMC_QCACHE
), "Bad KMC_QCACHE (%x)\n", flags
);
1597 ASSERT(vmp
== NULL
);
1602 * Allocate memory for a new cache an initialize it. Unfortunately,
1603 * this usually ends up being a large allocation of ~32k because
1604 * we need to allocate enough memory for the worst case number of
1605 * cpus in the magazine, skc_mag[NR_CPUS]. Because of this we
1606 * explicitly pass KM_NODEBUG to suppress the kmem warning
1608 skc
= kmem_zalloc(sizeof(*skc
), KM_SLEEP
| KM_NODEBUG
);
1612 skc
->skc_magic
= SKC_MAGIC
;
1613 skc
->skc_name_size
= strlen(name
) + 1;
1614 skc
->skc_name
= (char *)kmem_alloc(skc
->skc_name_size
, KM_SLEEP
);
1615 if (skc
->skc_name
== NULL
) {
1616 kmem_free(skc
, sizeof(*skc
));
1619 strncpy(skc
->skc_name
, name
, skc
->skc_name_size
);
1621 skc
->skc_ctor
= ctor
;
1622 skc
->skc_dtor
= dtor
;
1623 skc
->skc_reclaim
= reclaim
;
1624 skc
->skc_private
= priv
;
1626 skc
->skc_linux_cache
= NULL
;
1627 skc
->skc_flags
= flags
;
1628 skc
->skc_obj_size
= size
;
1629 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
1630 skc
->skc_delay
= SPL_KMEM_CACHE_DELAY
;
1631 skc
->skc_reap
= SPL_KMEM_CACHE_REAP
;
1632 atomic_set(&skc
->skc_ref
, 0);
1634 INIT_LIST_HEAD(&skc
->skc_list
);
1635 INIT_LIST_HEAD(&skc
->skc_complete_list
);
1636 INIT_LIST_HEAD(&skc
->skc_partial_list
);
1637 skc
->skc_emergency_tree
= RB_ROOT
;
1638 spin_lock_init(&skc
->skc_lock
);
1639 init_waitqueue_head(&skc
->skc_waitq
);
1640 skc
->skc_slab_fail
= 0;
1641 skc
->skc_slab_create
= 0;
1642 skc
->skc_slab_destroy
= 0;
1643 skc
->skc_slab_total
= 0;
1644 skc
->skc_slab_alloc
= 0;
1645 skc
->skc_slab_max
= 0;
1646 skc
->skc_obj_total
= 0;
1647 skc
->skc_obj_alloc
= 0;
1648 skc
->skc_obj_max
= 0;
1649 skc
->skc_obj_deadlock
= 0;
1650 skc
->skc_obj_emergency
= 0;
1651 skc
->skc_obj_emergency_max
= 0;
1654 * Verify the requested alignment restriction is sane.
1657 VERIFY(ISP2(align
));
1658 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
);
1659 VERIFY3U(align
, <=, PAGE_SIZE
);
1660 skc
->skc_obj_align
= align
;
1664 * When no specific type of slab is requested (kmem, vmem, or
1665 * linuxslab) then select a cache type based on the object size
1666 * and default tunables.
1668 if (!(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
))) {
1671 * Objects smaller than spl_kmem_cache_slab_limit can
1672 * use the Linux slab for better space-efficiency. By
1673 * default this functionality is disabled until its
1674 * performance characters are fully understood.
1676 if (spl_kmem_cache_slab_limit
&&
1677 size
<= (size_t)spl_kmem_cache_slab_limit
)
1678 skc
->skc_flags
|= KMC_SLAB
;
1681 * Small objects, less than spl_kmem_cache_kmem_limit per
1682 * object should use kmem because their slabs are small.
1684 else if (spl_obj_size(skc
) <= spl_kmem_cache_kmem_limit
)
1685 skc
->skc_flags
|= KMC_KMEM
;
1688 * All other objects are considered large and are placed
1689 * on vmem backed slabs.
1692 skc
->skc_flags
|= KMC_VMEM
;
1696 * Given the type of slab allocate the required resources.
1698 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
1699 rc
= spl_slab_size(skc
,
1700 &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
1704 rc
= spl_magazine_create(skc
);
1708 skc
->skc_linux_cache
= kmem_cache_create(
1709 skc
->skc_name
, size
, align
, 0, NULL
);
1710 if (skc
->skc_linux_cache
== NULL
)
1711 SGOTO(out
, rc
= ENOMEM
);
1713 kmem_cache_set_allocflags(skc
, __GFP_COMP
);
1714 skc
->skc_flags
|= KMC_NOMAGAZINE
;
1717 if (spl_kmem_cache_expire
& KMC_EXPIRE_AGE
)
1718 skc
->skc_taskqid
= taskq_dispatch_delay(spl_kmem_cache_taskq
,
1719 spl_cache_age
, skc
, TQ_SLEEP
,
1720 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1722 down_write(&spl_kmem_cache_sem
);
1723 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
1724 up_write(&spl_kmem_cache_sem
);
1728 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1729 kmem_free(skc
, sizeof(*skc
));
1732 EXPORT_SYMBOL(spl_kmem_cache_create
);
1735 * Register a move callback to for cache defragmentation.
1736 * XXX: Unimplemented but harmless to stub out for now.
1739 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
1740 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
1742 ASSERT(move
!= NULL
);
1744 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
1747 * Destroy a cache and all objects associated with the cache.
1750 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
1752 DECLARE_WAIT_QUEUE_HEAD(wq
);
1756 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1757 ASSERT(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
));
1759 down_write(&spl_kmem_cache_sem
);
1760 list_del_init(&skc
->skc_list
);
1761 up_write(&spl_kmem_cache_sem
);
1763 /* Cancel any and wait for any pending delayed tasks */
1764 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1766 spin_lock(&skc
->skc_lock
);
1767 id
= skc
->skc_taskqid
;
1768 spin_unlock(&skc
->skc_lock
);
1770 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1772 /* Wait until all current callers complete, this is mainly
1773 * to catch the case where a low memory situation triggers a
1774 * cache reaping action which races with this destroy. */
1775 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
1777 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
1778 spl_magazine_destroy(skc
);
1779 spl_slab_reclaim(skc
, 0, 1);
1781 ASSERT(skc
->skc_flags
& KMC_SLAB
);
1782 kmem_cache_destroy(skc
->skc_linux_cache
);
1785 spin_lock(&skc
->skc_lock
);
1787 /* Validate there are no objects in use and free all the
1788 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */
1789 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
1790 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
1791 ASSERT3U(skc
->skc_slab_total
, ==, 0);
1792 ASSERT3U(skc
->skc_obj_total
, ==, 0);
1793 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
1794 ASSERT(list_empty(&skc
->skc_complete_list
));
1796 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1797 spin_unlock(&skc
->skc_lock
);
1799 kmem_free(skc
, sizeof(*skc
));
1803 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
1806 * Allocate an object from a slab attached to the cache. This is used to
1807 * repopulate the per-cpu magazine caches in batches when they run low.
1810 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
1812 spl_kmem_obj_t
*sko
;
1814 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1815 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1816 ASSERT(spin_is_locked(&skc
->skc_lock
));
1818 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
1819 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1820 ASSERT(sko
->sko_addr
!= NULL
);
1822 /* Remove from sks_free_list */
1823 list_del_init(&sko
->sko_list
);
1825 sks
->sks_age
= jiffies
;
1827 skc
->skc_obj_alloc
++;
1829 /* Track max obj usage statistics */
1830 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
1831 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
1833 /* Track max slab usage statistics */
1834 if (sks
->sks_ref
== 1) {
1835 skc
->skc_slab_alloc
++;
1837 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
1838 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
1841 return sko
->sko_addr
;
1845 * Generic slab allocation function to run by the global work queues.
1846 * It is responsible for allocating a new slab, linking it in to the list
1847 * of partial slabs, and then waking any waiters.
1850 spl_cache_grow_work(void *data
)
1852 spl_kmem_alloc_t
*ska
= (spl_kmem_alloc_t
*)data
;
1853 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
1854 spl_kmem_slab_t
*sks
;
1856 sks
= spl_slab_alloc(skc
, ska
->ska_flags
| __GFP_NORETRY
| KM_NODEBUG
);
1857 spin_lock(&skc
->skc_lock
);
1859 skc
->skc_slab_total
++;
1860 skc
->skc_obj_total
+= sks
->sks_objs
;
1861 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1864 atomic_dec(&skc
->skc_ref
);
1865 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1866 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1867 wake_up_all(&skc
->skc_waitq
);
1868 spin_unlock(&skc
->skc_lock
);
1874 * Returns non-zero when a new slab should be available.
1877 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
1879 return !test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1883 spl_cache_reclaim_wait(void *word
)
1890 * No available objects on any slabs, create a new slab. Note that this
1891 * functionality is disabled for KMC_SLAB caches which are backed by the
1895 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1900 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1901 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
1906 * Before allocating a new slab wait for any reaping to complete and
1907 * then return so the local magazine can be rechecked for new objects.
1909 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1910 rc
= wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1911 spl_cache_reclaim_wait
, TASK_UNINTERRUPTIBLE
);
1912 SRETURN(rc
? rc
: -EAGAIN
);
1916 * This is handled by dispatching a work request to the global work
1917 * queue. This allows us to asynchronously allocate a new slab while
1918 * retaining the ability to safely fall back to a smaller synchronous
1919 * allocations to ensure forward progress is always maintained.
1921 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1922 spl_kmem_alloc_t
*ska
;
1924 ska
= kmalloc(sizeof(*ska
), flags
);
1926 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1927 wake_up_all(&skc
->skc_waitq
);
1931 atomic_inc(&skc
->skc_ref
);
1932 ska
->ska_cache
= skc
;
1933 ska
->ska_flags
= flags
& ~__GFP_FS
;
1934 taskq_init_ent(&ska
->ska_tqe
);
1935 taskq_dispatch_ent(spl_kmem_cache_taskq
,
1936 spl_cache_grow_work
, ska
, 0, &ska
->ska_tqe
);
1940 * The goal here is to only detect the rare case where a virtual slab
1941 * allocation has deadlocked. We must be careful to minimize the use
1942 * of emergency objects which are more expensive to track. Therefore,
1943 * we set a very long timeout for the asynchronous allocation and if
1944 * the timeout is reached the cache is flagged as deadlocked. From
1945 * this point only new emergency objects will be allocated until the
1946 * asynchronous allocation completes and clears the deadlocked flag.
1948 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1949 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1951 remaining
= wait_event_timeout(skc
->skc_waitq
,
1952 spl_cache_grow_wait(skc
), HZ
);
1954 if (!remaining
&& test_bit(KMC_BIT_VMEM
, &skc
->skc_flags
)) {
1955 spin_lock(&skc
->skc_lock
);
1956 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1957 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1958 skc
->skc_obj_deadlock
++;
1960 spin_unlock(&skc
->skc_lock
);
1970 * Refill a per-cpu magazine with objects from the slabs for this cache.
1971 * Ideally the magazine can be repopulated using existing objects which have
1972 * been released, however if we are unable to locate enough free objects new
1973 * slabs of objects will be created. On success NULL is returned, otherwise
1974 * the address of a single emergency object is returned for use by the caller.
1977 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1979 spl_kmem_slab_t
*sks
;
1980 int count
= 0, rc
, refill
;
1984 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1985 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1987 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1988 spin_lock(&skc
->skc_lock
);
1990 while (refill
> 0) {
1991 /* No slabs available we may need to grow the cache */
1992 if (list_empty(&skc
->skc_partial_list
)) {
1993 spin_unlock(&skc
->skc_lock
);
1996 rc
= spl_cache_grow(skc
, flags
, &obj
);
1997 local_irq_disable();
1999 /* Emergency object for immediate use by caller */
2000 if (rc
== 0 && obj
!= NULL
)
2006 /* Rescheduled to different CPU skm is not local */
2007 if (skm
!= skc
->skc_mag
[smp_processor_id()])
2010 /* Potentially rescheduled to the same CPU but
2011 * allocations may have occurred from this CPU while
2012 * we were sleeping so recalculate max refill. */
2013 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
2015 spin_lock(&skc
->skc_lock
);
2019 /* Grab the next available slab */
2020 sks
= list_entry((&skc
->skc_partial_list
)->next
,
2021 spl_kmem_slab_t
, sks_list
);
2022 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
2023 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
2024 ASSERT(!list_empty(&sks
->sks_free_list
));
2026 /* Consume as many objects as needed to refill the requested
2027 * cache. We must also be careful not to overfill it. */
2028 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 && ++count
) {
2029 ASSERT(skm
->skm_avail
< skm
->skm_size
);
2030 ASSERT(count
< skm
->skm_size
);
2031 skm
->skm_objs
[skm
->skm_avail
++]=spl_cache_obj(skc
,sks
);
2034 /* Move slab to skc_complete_list when full */
2035 if (sks
->sks_ref
== sks
->sks_objs
) {
2036 list_del(&sks
->sks_list
);
2037 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
2041 spin_unlock(&skc
->skc_lock
);
2047 * Release an object back to the slab from which it came.
2050 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
2052 spl_kmem_slab_t
*sks
= NULL
;
2053 spl_kmem_obj_t
*sko
= NULL
;
2056 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2057 ASSERT(spin_is_locked(&skc
->skc_lock
));
2059 sko
= spl_sko_from_obj(skc
, obj
);
2060 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
2061 sks
= sko
->sko_slab
;
2062 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
2063 ASSERT(sks
->sks_cache
== skc
);
2064 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
2066 sks
->sks_age
= jiffies
;
2068 skc
->skc_obj_alloc
--;
2070 /* Move slab to skc_partial_list when no longer full. Slabs
2071 * are added to the head to keep the partial list is quasi-full
2072 * sorted order. Fuller at the head, emptier at the tail. */
2073 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
2074 list_del(&sks
->sks_list
);
2075 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
2078 /* Move empty slabs to the end of the partial list so
2079 * they can be easily found and freed during reclamation. */
2080 if (sks
->sks_ref
== 0) {
2081 list_del(&sks
->sks_list
);
2082 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
2083 skc
->skc_slab_alloc
--;
2090 * Allocate an object from the per-cpu magazine, or if the magazine
2091 * is empty directly allocate from a slab and repopulate the magazine.
2094 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
2096 spl_kmem_magazine_t
*skm
;
2100 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2101 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2102 ASSERT(flags
& KM_SLEEP
);
2104 atomic_inc(&skc
->skc_ref
);
2107 * Allocate directly from a Linux slab. All optimizations are left
2108 * to the underlying cache we only need to guarantee that KM_SLEEP
2109 * callers will never fail.
2111 if (skc
->skc_flags
& KMC_SLAB
) {
2112 struct kmem_cache
*slc
= skc
->skc_linux_cache
;
2115 obj
= kmem_cache_alloc(slc
, flags
| __GFP_COMP
);
2116 if (obj
&& skc
->skc_ctor
)
2117 skc
->skc_ctor(obj
, skc
->skc_private
, flags
);
2119 } while ((obj
== NULL
) && !(flags
& KM_NOSLEEP
));
2121 atomic_dec(&skc
->skc_ref
);
2125 local_irq_disable();
2128 /* Safe to update per-cpu structure without lock, but
2129 * in the restart case we must be careful to reacquire
2130 * the local magazine since this may have changed
2131 * when we need to grow the cache. */
2132 skm
= skc
->skc_mag
[smp_processor_id()];
2133 ASSERTF(skm
->skm_magic
== SKM_MAGIC
, "%x != %x: %s/%p/%p %x/%x/%x\n",
2134 skm
->skm_magic
, SKM_MAGIC
, skc
->skc_name
, skc
, skm
,
2135 skm
->skm_size
, skm
->skm_refill
, skm
->skm_avail
);
2137 if (likely(skm
->skm_avail
)) {
2138 /* Object available in CPU cache, use it */
2139 obj
= skm
->skm_objs
[--skm
->skm_avail
];
2140 skm
->skm_age
= jiffies
;
2142 obj
= spl_cache_refill(skc
, skm
, flags
);
2144 SGOTO(restart
, obj
= NULL
);
2149 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
2151 /* Pre-emptively migrate object to CPU L1 cache */
2153 atomic_dec(&skc
->skc_ref
);
2157 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
2160 * Free an object back to the local per-cpu magazine, there is no
2161 * guarantee that this is the same magazine the object was originally
2162 * allocated from. We may need to flush entire from the magazine
2163 * back to the slabs to make space.
2166 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
2168 spl_kmem_magazine_t
*skm
;
2169 unsigned long flags
;
2172 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2173 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2174 atomic_inc(&skc
->skc_ref
);
2177 * Free the object from the Linux underlying Linux slab.
2179 if (skc
->skc_flags
& KMC_SLAB
) {
2181 skc
->skc_dtor(obj
, skc
->skc_private
);
2183 kmem_cache_free(skc
->skc_linux_cache
, obj
);
2188 * Only virtual slabs may have emergency objects and these objects
2189 * are guaranteed to have physical addresses. They must be removed
2190 * from the tree of emergency objects and the freed.
2192 if ((skc
->skc_flags
& KMC_VMEM
) && !kmem_virt(obj
))
2193 SGOTO(out
, spl_emergency_free(skc
, obj
));
2195 local_irq_save(flags
);
2197 /* Safe to update per-cpu structure without lock, but
2198 * no remote memory allocation tracking is being performed
2199 * it is entirely possible to allocate an object from one
2200 * CPU cache and return it to another. */
2201 skm
= skc
->skc_mag
[smp_processor_id()];
2202 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
2204 /* Per-CPU cache full, flush it to make space */
2205 if (unlikely(skm
->skm_avail
>= skm
->skm_size
))
2206 spl_cache_flush(skc
, skm
, skm
->skm_refill
);
2208 /* Available space in cache, use it */
2209 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
2211 local_irq_restore(flags
);
2213 atomic_dec(&skc
->skc_ref
);
2217 EXPORT_SYMBOL(spl_kmem_cache_free
);
2220 * The generic shrinker function for all caches. Under Linux a shrinker
2221 * may not be tightly coupled with a slab cache. In fact Linux always
2222 * systematically tries calling all registered shrinker callbacks which
2223 * report that they contain unused objects. Because of this we only
2224 * register one shrinker function in the shim layer for all slab caches.
2225 * We always attempt to shrink all caches when this generic shrinker
2226 * is called. The shrinker should return the number of free objects
2227 * in the cache when called with nr_to_scan == 0 but not attempt to
2228 * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan
2229 * objects should be freed, which differs from Solaris semantics.
2230 * Solaris semantics are to free all available objects which may (and
2231 * probably will) be more objects than the requested nr_to_scan.
2234 __spl_kmem_cache_generic_shrinker(struct shrinker
*shrink
,
2235 struct shrink_control
*sc
)
2237 spl_kmem_cache_t
*skc
;
2240 down_read(&spl_kmem_cache_sem
);
2241 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
2243 spl_kmem_cache_reap_now(skc
,
2244 MAX(sc
->nr_to_scan
>> fls64(skc
->skc_slab_objs
), 1));
2247 * Presume everything alloc'ed in reclaimable, this ensures
2248 * we are called again with nr_to_scan > 0 so can try and
2249 * reclaim. The exact number is not important either so
2250 * we forgo taking this already highly contented lock.
2252 unused
+= skc
->skc_obj_alloc
;
2254 up_read(&spl_kmem_cache_sem
);
2257 * After performing reclaim always return -1 to indicate we cannot
2258 * perform additional reclaim. This prevents shrink_slabs() from
2259 * repeatedly invoking this generic shrinker and potentially spinning.
2267 SPL_SHRINKER_CALLBACK_WRAPPER(spl_kmem_cache_generic_shrinker
);
2270 * Call the registered reclaim function for a cache. Depending on how
2271 * many and which objects are released it may simply repopulate the
2272 * local magazine which will then need to age-out. Objects which cannot
2273 * fit in the magazine we will be released back to their slabs which will
2274 * also need to age out before being release. This is all just best
2275 * effort and we do not want to thrash creating and destroying slabs.
2278 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
, int count
)
2282 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2283 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2285 atomic_inc(&skc
->skc_ref
);
2288 * Execute the registered reclaim callback if it exists. The
2289 * per-cpu caches will be drained when is set KMC_EXPIRE_MEM.
2291 if (skc
->skc_flags
& KMC_SLAB
) {
2292 if (skc
->skc_reclaim
)
2293 skc
->skc_reclaim(skc
->skc_private
);
2295 if (spl_kmem_cache_expire
& KMC_EXPIRE_MEM
)
2296 kmem_cache_shrink(skc
->skc_linux_cache
);
2302 * Prevent concurrent cache reaping when contended.
2304 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
))
2308 * When a reclaim function is available it may be invoked repeatedly
2309 * until at least a single slab can be freed. This ensures that we
2310 * do free memory back to the system. This helps minimize the chance
2311 * of an OOM event when the bulk of memory is used by the slab.
2313 * When free slabs are already available the reclaim callback will be
2314 * skipped. Additionally, if no forward progress is detected despite
2315 * a reclaim function the cache will be skipped to avoid deadlock.
2317 * Longer term this would be the correct place to add the code which
2318 * repacks the slabs in order minimize fragmentation.
2320 if (skc
->skc_reclaim
) {
2321 uint64_t objects
= UINT64_MAX
;
2325 spin_lock(&skc
->skc_lock
);
2327 (skc
->skc_slab_total
> 0) &&
2328 ((skc
->skc_slab_total
- skc
->skc_slab_alloc
) == 0) &&
2329 (skc
->skc_obj_alloc
< objects
);
2331 objects
= skc
->skc_obj_alloc
;
2332 spin_unlock(&skc
->skc_lock
);
2335 skc
->skc_reclaim(skc
->skc_private
);
2337 } while (do_reclaim
);
2340 /* Reclaim from the magazine then the slabs ignoring age and delay. */
2341 if (spl_kmem_cache_expire
& KMC_EXPIRE_MEM
) {
2342 spl_kmem_magazine_t
*skm
;
2343 unsigned long irq_flags
;
2345 local_irq_save(irq_flags
);
2346 skm
= skc
->skc_mag
[smp_processor_id()];
2347 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
2348 local_irq_restore(irq_flags
);
2351 spl_slab_reclaim(skc
, count
, 1);
2352 clear_bit(KMC_BIT_REAPING
, &skc
->skc_flags
);
2353 smp_mb__after_clear_bit();
2354 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
2356 atomic_dec(&skc
->skc_ref
);
2360 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
2363 * Reap all free slabs from all registered caches.
2368 struct shrink_control sc
;
2370 sc
.nr_to_scan
= KMC_REAP_CHUNK
;
2371 sc
.gfp_mask
= GFP_KERNEL
;
2373 __spl_kmem_cache_generic_shrinker(NULL
, &sc
);
2375 EXPORT_SYMBOL(spl_kmem_reap
);
2377 #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING)
2379 spl_sprintf_addr(kmem_debug_t
*kd
, char *str
, int len
, int min
)
2381 int size
= ((len
- 1) < kd
->kd_size
) ? (len
- 1) : kd
->kd_size
;
2384 ASSERT(str
!= NULL
&& len
>= 17);
2385 memset(str
, 0, len
);
2387 /* Check for a fully printable string, and while we are at
2388 * it place the printable characters in the passed buffer. */
2389 for (i
= 0; i
< size
; i
++) {
2390 str
[i
] = ((char *)(kd
->kd_addr
))[i
];
2391 if (isprint(str
[i
])) {
2394 /* Minimum number of printable characters found
2395 * to make it worthwhile to print this as ascii. */
2405 sprintf(str
, "%02x%02x%02x%02x%02x%02x%02x%02x",
2406 *((uint8_t *)kd
->kd_addr
),
2407 *((uint8_t *)kd
->kd_addr
+ 2),
2408 *((uint8_t *)kd
->kd_addr
+ 4),
2409 *((uint8_t *)kd
->kd_addr
+ 6),
2410 *((uint8_t *)kd
->kd_addr
+ 8),
2411 *((uint8_t *)kd
->kd_addr
+ 10),
2412 *((uint8_t *)kd
->kd_addr
+ 12),
2413 *((uint8_t *)kd
->kd_addr
+ 14));
2420 spl_kmem_init_tracking(struct list_head
*list
, spinlock_t
*lock
, int size
)
2425 spin_lock_init(lock
);
2426 INIT_LIST_HEAD(list
);
2428 for (i
= 0; i
< size
; i
++)
2429 INIT_HLIST_HEAD(&kmem_table
[i
]);
2435 spl_kmem_fini_tracking(struct list_head
*list
, spinlock_t
*lock
)
2437 unsigned long flags
;
2442 spin_lock_irqsave(lock
, flags
);
2443 if (!list_empty(list
))
2444 printk(KERN_WARNING
"%-16s %-5s %-16s %s:%s\n", "address",
2445 "size", "data", "func", "line");
2447 list_for_each_entry(kd
, list
, kd_list
)
2448 printk(KERN_WARNING
"%p %-5d %-16s %s:%d\n", kd
->kd_addr
,
2449 (int)kd
->kd_size
, spl_sprintf_addr(kd
, str
, 17, 8),
2450 kd
->kd_func
, kd
->kd_line
);
2452 spin_unlock_irqrestore(lock
, flags
);
2455 #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2456 #define spl_kmem_init_tracking(list, lock, size)
2457 #define spl_kmem_fini_tracking(list, lock)
2458 #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2461 spl_kmem_init_globals(void)
2465 /* For now all zones are includes, it may be wise to restrict
2466 * this to normal and highmem zones if we see problems. */
2467 for_each_zone(zone
) {
2469 if (!populated_zone(zone
))
2472 minfree
+= min_wmark_pages(zone
);
2473 desfree
+= low_wmark_pages(zone
);
2474 lotsfree
+= high_wmark_pages(zone
);
2477 /* Solaris default values */
2478 swapfs_minfree
= MAX(2*1024*1024 >> PAGE_SHIFT
, physmem
>> 3);
2479 swapfs_reserve
= MIN(4*1024*1024 >> PAGE_SHIFT
, physmem
>> 4);
2483 * Called at module init when it is safe to use spl_kallsyms_lookup_name()
2486 spl_kmem_init_kallsyms_lookup(void)
2488 #ifndef HAVE_GET_VMALLOC_INFO
2489 get_vmalloc_info_fn
= (get_vmalloc_info_t
)
2490 spl_kallsyms_lookup_name("get_vmalloc_info");
2491 if (!get_vmalloc_info_fn
) {
2492 printk(KERN_ERR
"Error: Unknown symbol get_vmalloc_info\n");
2495 #endif /* HAVE_GET_VMALLOC_INFO */
2497 #ifdef HAVE_PGDAT_HELPERS
2498 # ifndef HAVE_FIRST_ONLINE_PGDAT
2499 first_online_pgdat_fn
= (first_online_pgdat_t
)
2500 spl_kallsyms_lookup_name("first_online_pgdat");
2501 if (!first_online_pgdat_fn
) {
2502 printk(KERN_ERR
"Error: Unknown symbol first_online_pgdat\n");
2505 # endif /* HAVE_FIRST_ONLINE_PGDAT */
2507 # ifndef HAVE_NEXT_ONLINE_PGDAT
2508 next_online_pgdat_fn
= (next_online_pgdat_t
)
2509 spl_kallsyms_lookup_name("next_online_pgdat");
2510 if (!next_online_pgdat_fn
) {
2511 printk(KERN_ERR
"Error: Unknown symbol next_online_pgdat\n");
2514 # endif /* HAVE_NEXT_ONLINE_PGDAT */
2516 # ifndef HAVE_NEXT_ZONE
2517 next_zone_fn
= (next_zone_t
)
2518 spl_kallsyms_lookup_name("next_zone");
2519 if (!next_zone_fn
) {
2520 printk(KERN_ERR
"Error: Unknown symbol next_zone\n");
2523 # endif /* HAVE_NEXT_ZONE */
2525 #else /* HAVE_PGDAT_HELPERS */
2527 # ifndef HAVE_PGDAT_LIST
2528 pgdat_list_addr
= *(struct pglist_data
**)
2529 spl_kallsyms_lookup_name("pgdat_list");
2530 if (!pgdat_list_addr
) {
2531 printk(KERN_ERR
"Error: Unknown symbol pgdat_list\n");
2534 # endif /* HAVE_PGDAT_LIST */
2535 #endif /* HAVE_PGDAT_HELPERS */
2537 #if defined(NEED_GET_ZONE_COUNTS) && !defined(HAVE_GET_ZONE_COUNTS)
2538 get_zone_counts_fn
= (get_zone_counts_t
)
2539 spl_kallsyms_lookup_name("get_zone_counts");
2540 if (!get_zone_counts_fn
) {
2541 printk(KERN_ERR
"Error: Unknown symbol get_zone_counts\n");
2544 #endif /* NEED_GET_ZONE_COUNTS && !HAVE_GET_ZONE_COUNTS */
2547 * It is now safe to initialize the global tunings which rely on
2548 * the use of the for_each_zone() macro. This macro in turns
2549 * depends on the *_pgdat symbols which are now available.
2551 spl_kmem_init_globals();
2553 #ifndef HAVE_SHRINK_DCACHE_MEMORY
2554 /* When shrink_dcache_memory_fn == NULL support is disabled */
2555 shrink_dcache_memory_fn
= (shrink_dcache_memory_t
)
2556 spl_kallsyms_lookup_name("shrink_dcache_memory");
2557 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
2559 #ifndef HAVE_SHRINK_ICACHE_MEMORY
2560 /* When shrink_icache_memory_fn == NULL support is disabled */
2561 shrink_icache_memory_fn
= (shrink_icache_memory_t
)
2562 spl_kallsyms_lookup_name("shrink_icache_memory");
2563 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
2575 kmem_alloc_used_set(0);
2576 vmem_alloc_used_set(0);
2578 spl_kmem_init_tracking(&kmem_list
, &kmem_lock
, KMEM_TABLE_SIZE
);
2579 spl_kmem_init_tracking(&vmem_list
, &vmem_lock
, VMEM_TABLE_SIZE
);
2582 init_rwsem(&spl_kmem_cache_sem
);
2583 INIT_LIST_HEAD(&spl_kmem_cache_list
);
2584 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
2585 1, maxclsyspri
, 1, 32, TASKQ_PREPOPULATE
);
2587 spl_register_shrinker(&spl_kmem_cache_shrinker
);
2597 spl_unregister_shrinker(&spl_kmem_cache_shrinker
);
2598 taskq_destroy(spl_kmem_cache_taskq
);
2601 /* Display all unreclaimed memory addresses, including the
2602 * allocation size and the first few bytes of what's located
2603 * at that address to aid in debugging. Performance is not
2604 * a serious concern here since it is module unload time. */
2605 if (kmem_alloc_used_read() != 0)
2606 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2607 "kmem leaked %ld/%ld bytes\n",
2608 kmem_alloc_used_read(), kmem_alloc_max
);
2611 if (vmem_alloc_used_read() != 0)
2612 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2613 "vmem leaked %ld/%ld bytes\n",
2614 vmem_alloc_used_read(), vmem_alloc_max
);
2616 spl_kmem_fini_tracking(&kmem_list
, &kmem_lock
);
2617 spl_kmem_fini_tracking(&vmem_list
, &vmem_lock
);
2618 #endif /* DEBUG_KMEM */