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://github.com/behlendorf/spl/>.
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 * The minimum amount of memory measured in pages to be free at all
38 * times on the system. This is similar to Linux's zone->pages_min
39 * multiplied by the number of zones and is sized based on that.
42 EXPORT_SYMBOL(minfree
);
45 * The desired amount of memory measured in pages to be free at all
46 * times on the system. This is similar to Linux's zone->pages_low
47 * multiplied by the number of zones and is sized based on that.
48 * Assuming all zones are being used roughly equally, when we drop
49 * below this threshold asynchronous page reclamation is triggered.
52 EXPORT_SYMBOL(desfree
);
55 * When above this amount of memory measures in pages the system is
56 * determined to have enough free memory. This is similar to Linux's
57 * zone->pages_high multiplied by the number of zones and is sized based
58 * on that. Assuming all zones are being used roughly equally, when
59 * asynchronous page reclamation reaches this threshold it stops.
62 EXPORT_SYMBOL(lotsfree
);
64 /* Unused always 0 in this implementation */
66 EXPORT_SYMBOL(needfree
);
68 pgcnt_t swapfs_minfree
= 0;
69 EXPORT_SYMBOL(swapfs_minfree
);
71 pgcnt_t swapfs_reserve
= 0;
72 EXPORT_SYMBOL(swapfs_reserve
);
74 vmem_t
*heap_arena
= NULL
;
75 EXPORT_SYMBOL(heap_arena
);
77 vmem_t
*zio_alloc_arena
= NULL
;
78 EXPORT_SYMBOL(zio_alloc_arena
);
80 vmem_t
*zio_arena
= NULL
;
81 EXPORT_SYMBOL(zio_arena
);
83 #ifndef HAVE_GET_VMALLOC_INFO
84 get_vmalloc_info_t get_vmalloc_info_fn
= SYMBOL_POISON
;
85 EXPORT_SYMBOL(get_vmalloc_info_fn
);
86 #endif /* HAVE_GET_VMALLOC_INFO */
88 #ifdef HAVE_PGDAT_HELPERS
89 # ifndef HAVE_FIRST_ONLINE_PGDAT
90 first_online_pgdat_t first_online_pgdat_fn
= SYMBOL_POISON
;
91 EXPORT_SYMBOL(first_online_pgdat_fn
);
92 # endif /* HAVE_FIRST_ONLINE_PGDAT */
94 # ifndef HAVE_NEXT_ONLINE_PGDAT
95 next_online_pgdat_t next_online_pgdat_fn
= SYMBOL_POISON
;
96 EXPORT_SYMBOL(next_online_pgdat_fn
);
97 # endif /* HAVE_NEXT_ONLINE_PGDAT */
99 # ifndef HAVE_NEXT_ZONE
100 next_zone_t next_zone_fn
= SYMBOL_POISON
;
101 EXPORT_SYMBOL(next_zone_fn
);
102 # endif /* HAVE_NEXT_ZONE */
104 #else /* HAVE_PGDAT_HELPERS */
106 # ifndef HAVE_PGDAT_LIST
107 struct pglist_data
*pgdat_list_addr
= SYMBOL_POISON
;
108 EXPORT_SYMBOL(pgdat_list_addr
);
109 # endif /* HAVE_PGDAT_LIST */
111 #endif /* HAVE_PGDAT_HELPERS */
113 #ifdef NEED_GET_ZONE_COUNTS
114 # ifndef HAVE_GET_ZONE_COUNTS
115 get_zone_counts_t get_zone_counts_fn
= SYMBOL_POISON
;
116 EXPORT_SYMBOL(get_zone_counts_fn
);
117 # endif /* HAVE_GET_ZONE_COUNTS */
120 spl_global_page_state(spl_zone_stat_item_t item
)
122 unsigned long active
;
123 unsigned long inactive
;
126 get_zone_counts(&active
, &inactive
, &free
);
128 case SPL_NR_FREE_PAGES
: return free
;
129 case SPL_NR_INACTIVE
: return inactive
;
130 case SPL_NR_ACTIVE
: return active
;
131 default: ASSERT(0); /* Unsupported */
137 # ifdef HAVE_GLOBAL_PAGE_STATE
139 spl_global_page_state(spl_zone_stat_item_t item
)
141 unsigned long pages
= 0;
144 case SPL_NR_FREE_PAGES
:
145 # ifdef HAVE_ZONE_STAT_ITEM_NR_FREE_PAGES
146 pages
+= global_page_state(NR_FREE_PAGES
);
149 case SPL_NR_INACTIVE
:
150 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE
151 pages
+= global_page_state(NR_INACTIVE
);
153 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_ANON
154 pages
+= global_page_state(NR_INACTIVE_ANON
);
156 # ifdef HAVE_ZONE_STAT_ITEM_NR_INACTIVE_FILE
157 pages
+= global_page_state(NR_INACTIVE_FILE
);
161 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE
162 pages
+= global_page_state(NR_ACTIVE
);
164 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_ANON
165 pages
+= global_page_state(NR_ACTIVE_ANON
);
167 # ifdef HAVE_ZONE_STAT_ITEM_NR_ACTIVE_FILE
168 pages
+= global_page_state(NR_ACTIVE_FILE
);
172 ASSERT(0); /* Unsupported */
178 # error "Both global_page_state() and get_zone_counts() unavailable"
179 # endif /* HAVE_GLOBAL_PAGE_STATE */
180 #endif /* NEED_GET_ZONE_COUNTS */
181 EXPORT_SYMBOL(spl_global_page_state
);
183 #ifndef HAVE_SHRINK_DCACHE_MEMORY
184 shrink_dcache_memory_t shrink_dcache_memory_fn
= SYMBOL_POISON
;
185 EXPORT_SYMBOL(shrink_dcache_memory_fn
);
186 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
188 #ifndef HAVE_SHRINK_ICACHE_MEMORY
189 shrink_icache_memory_t shrink_icache_memory_fn
= SYMBOL_POISON
;
190 EXPORT_SYMBOL(shrink_icache_memory_fn
);
191 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
194 spl_kmem_availrmem(void)
196 /* The amount of easily available memory */
197 return (spl_global_page_state(SPL_NR_FREE_PAGES
) +
198 spl_global_page_state(SPL_NR_INACTIVE
));
200 EXPORT_SYMBOL(spl_kmem_availrmem
);
203 vmem_size(vmem_t
*vmp
, int typemask
)
205 struct vmalloc_info vmi
;
209 ASSERT(typemask
& (VMEM_ALLOC
| VMEM_FREE
));
211 get_vmalloc_info(&vmi
);
212 if (typemask
& VMEM_ALLOC
)
213 size
+= (size_t)vmi
.used
;
215 if (typemask
& VMEM_FREE
)
216 size
+= (size_t)(VMALLOC_TOTAL
- vmi
.used
);
220 EXPORT_SYMBOL(vmem_size
);
227 EXPORT_SYMBOL(kmem_debugging
);
229 #ifndef HAVE_KVASPRINTF
230 /* Simplified asprintf. */
231 char *kvasprintf(gfp_t gfp
, const char *fmt
, va_list ap
)
238 len
= vsnprintf(NULL
, 0, fmt
, aq
);
241 p
= kmalloc(len
+1, gfp
);
245 vsnprintf(p
, len
+1, fmt
, ap
);
249 EXPORT_SYMBOL(kvasprintf
);
250 #endif /* HAVE_KVASPRINTF */
253 kmem_vasprintf(const char *fmt
, va_list ap
)
260 ptr
= kvasprintf(GFP_KERNEL
, fmt
, aq
);
262 } while (ptr
== NULL
);
266 EXPORT_SYMBOL(kmem_vasprintf
);
269 kmem_asprintf(const char *fmt
, ...)
276 ptr
= kvasprintf(GFP_KERNEL
, fmt
, ap
);
278 } while (ptr
== NULL
);
282 EXPORT_SYMBOL(kmem_asprintf
);
285 __strdup(const char *str
, int flags
)
291 ptr
= kmalloc_nofail(n
+ 1, flags
);
293 memcpy(ptr
, str
, n
+ 1);
299 strdup(const char *str
)
301 return __strdup(str
, KM_SLEEP
);
303 EXPORT_SYMBOL(strdup
);
310 EXPORT_SYMBOL(strfree
);
313 * Memory allocation interfaces and debugging for basic kmem_*
314 * and vmem_* style memory allocation. When DEBUG_KMEM is enabled
315 * the SPL will keep track of the total memory allocated, and
316 * report any memory leaked when the module is unloaded.
320 /* Shim layer memory accounting */
321 # ifdef HAVE_ATOMIC64_T
322 atomic64_t kmem_alloc_used
= ATOMIC64_INIT(0);
323 unsigned long long kmem_alloc_max
= 0;
324 atomic64_t vmem_alloc_used
= ATOMIC64_INIT(0);
325 unsigned long long vmem_alloc_max
= 0;
326 # else /* HAVE_ATOMIC64_T */
327 atomic_t kmem_alloc_used
= ATOMIC_INIT(0);
328 unsigned long long kmem_alloc_max
= 0;
329 atomic_t vmem_alloc_used
= ATOMIC_INIT(0);
330 unsigned long long vmem_alloc_max
= 0;
331 # endif /* HAVE_ATOMIC64_T */
333 EXPORT_SYMBOL(kmem_alloc_used
);
334 EXPORT_SYMBOL(kmem_alloc_max
);
335 EXPORT_SYMBOL(vmem_alloc_used
);
336 EXPORT_SYMBOL(vmem_alloc_max
);
338 /* When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked
339 * but also the location of every alloc and free. When the SPL module is
340 * unloaded a list of all leaked addresses and where they were allocated
341 * will be dumped to the console. Enabling this feature has a significant
342 * impact on performance but it makes finding memory leaks straight forward.
344 * Not surprisingly with debugging enabled the xmem_locks are very highly
345 * contended particularly on xfree(). If we want to run with this detailed
346 * debugging enabled for anything other than debugging we need to minimize
347 * the contention by moving to a lock per xmem_table entry model.
349 # ifdef DEBUG_KMEM_TRACKING
351 # define KMEM_HASH_BITS 10
352 # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS)
354 # define VMEM_HASH_BITS 10
355 # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS)
357 typedef struct kmem_debug
{
358 struct hlist_node kd_hlist
; /* Hash node linkage */
359 struct list_head kd_list
; /* List of all allocations */
360 void *kd_addr
; /* Allocation pointer */
361 size_t kd_size
; /* Allocation size */
362 const char *kd_func
; /* Allocation function */
363 int kd_line
; /* Allocation line */
366 spinlock_t kmem_lock
;
367 struct hlist_head kmem_table
[KMEM_TABLE_SIZE
];
368 struct list_head kmem_list
;
370 spinlock_t vmem_lock
;
371 struct hlist_head vmem_table
[VMEM_TABLE_SIZE
];
372 struct list_head vmem_list
;
374 EXPORT_SYMBOL(kmem_lock
);
375 EXPORT_SYMBOL(kmem_table
);
376 EXPORT_SYMBOL(kmem_list
);
378 EXPORT_SYMBOL(vmem_lock
);
379 EXPORT_SYMBOL(vmem_table
);
380 EXPORT_SYMBOL(vmem_list
);
382 static kmem_debug_t
*
383 kmem_del_init(spinlock_t
*lock
, struct hlist_head
*table
, int bits
, const void *addr
)
385 struct hlist_head
*head
;
386 struct hlist_node
*node
;
387 struct kmem_debug
*p
;
391 spin_lock_irqsave(lock
, flags
);
393 head
= &table
[hash_ptr(addr
, bits
)];
394 hlist_for_each_entry_rcu(p
, node
, head
, kd_hlist
) {
395 if (p
->kd_addr
== addr
) {
396 hlist_del_init(&p
->kd_hlist
);
397 list_del_init(&p
->kd_list
);
398 spin_unlock_irqrestore(lock
, flags
);
403 spin_unlock_irqrestore(lock
, flags
);
409 kmem_alloc_track(size_t size
, int flags
, const char *func
, int line
,
410 int node_alloc
, int node
)
414 unsigned long irq_flags
;
417 /* Function may be called with KM_NOSLEEP so failure is possible */
418 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
419 flags
& ~__GFP_ZERO
);
421 if (unlikely(dptr
== NULL
)) {
422 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
423 "kmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
424 sizeof(kmem_debug_t
), flags
, func
, line
,
425 kmem_alloc_used_read(), kmem_alloc_max
);
428 * Marked unlikely because we should never be doing this,
429 * we tolerate to up 2 pages but a single page is best.
431 if (unlikely((size
> PAGE_SIZE
*2) && !(flags
& KM_NODEBUG
))) {
432 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "large "
433 "kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
434 (unsigned long long) size
, flags
, func
, line
,
435 kmem_alloc_used_read(), kmem_alloc_max
);
436 spl_debug_dumpstack(NULL
);
440 * We use __strdup() below because the string pointed to by
441 * __FUNCTION__ might not be available by the time we want
442 * to print it since the module might have been unloaded.
443 * This can only fail in the KM_NOSLEEP case.
445 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
446 if (unlikely(dptr
->kd_func
== NULL
)) {
448 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
449 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
450 func
, line
, kmem_alloc_used_read(), kmem_alloc_max
);
454 /* Use the correct allocator */
456 ASSERT(!(flags
& __GFP_ZERO
));
457 ptr
= kmalloc_node_nofail(size
, flags
, node
);
458 } else if (flags
& __GFP_ZERO
) {
459 ptr
= kzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
461 ptr
= kmalloc_nofail(size
, flags
);
464 if (unlikely(ptr
== NULL
)) {
465 kfree(dptr
->kd_func
);
467 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "kmem_alloc"
468 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
469 (unsigned long long) size
, flags
, func
, line
,
470 kmem_alloc_used_read(), kmem_alloc_max
);
474 kmem_alloc_used_add(size
);
475 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
476 kmem_alloc_max
= kmem_alloc_used_read();
478 INIT_HLIST_NODE(&dptr
->kd_hlist
);
479 INIT_LIST_HEAD(&dptr
->kd_list
);
482 dptr
->kd_size
= size
;
483 dptr
->kd_line
= line
;
485 spin_lock_irqsave(&kmem_lock
, irq_flags
);
486 hlist_add_head_rcu(&dptr
->kd_hlist
,
487 &kmem_table
[hash_ptr(ptr
, KMEM_HASH_BITS
)]);
488 list_add_tail(&dptr
->kd_list
, &kmem_list
);
489 spin_unlock_irqrestore(&kmem_lock
, irq_flags
);
491 SDEBUG_LIMIT(SD_INFO
,
492 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
493 (unsigned long long) size
, flags
, func
, line
, ptr
,
494 kmem_alloc_used_read(), kmem_alloc_max
);
499 EXPORT_SYMBOL(kmem_alloc_track
);
502 kmem_free_track(const void *ptr
, size_t size
)
507 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
508 (unsigned long long) size
);
510 dptr
= kmem_del_init(&kmem_lock
, kmem_table
, KMEM_HASH_BITS
, ptr
);
512 /* Must exist in hash due to kmem_alloc() */
515 /* Size must match */
516 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
517 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
518 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
520 kmem_alloc_used_sub(size
);
521 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
522 (unsigned long long) size
, kmem_alloc_used_read(),
525 kfree(dptr
->kd_func
);
527 memset(dptr
, 0x5a, sizeof(kmem_debug_t
));
530 memset(ptr
, 0x5a, size
);
535 EXPORT_SYMBOL(kmem_free_track
);
538 vmem_alloc_track(size_t size
, int flags
, const char *func
, int line
)
542 unsigned long irq_flags
;
545 ASSERT(flags
& KM_SLEEP
);
547 /* Function may be called with KM_NOSLEEP so failure is possible */
548 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
549 flags
& ~__GFP_ZERO
);
550 if (unlikely(dptr
== NULL
)) {
551 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
552 "vmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
553 sizeof(kmem_debug_t
), flags
, func
, line
,
554 vmem_alloc_used_read(), vmem_alloc_max
);
557 * We use __strdup() below because the string pointed to by
558 * __FUNCTION__ might not be available by the time we want
559 * to print it, since the module might have been unloaded.
560 * This can never fail because we have already asserted
561 * that flags is KM_SLEEP.
563 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
564 if (unlikely(dptr
->kd_func
== NULL
)) {
566 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
567 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
568 func
, line
, vmem_alloc_used_read(), vmem_alloc_max
);
572 /* Use the correct allocator */
573 if (flags
& __GFP_ZERO
) {
574 ptr
= vzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
576 ptr
= vmalloc_nofail(size
, flags
);
579 if (unlikely(ptr
== NULL
)) {
580 kfree(dptr
->kd_func
);
582 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "vmem_alloc"
583 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
584 (unsigned long long) size
, flags
, func
, line
,
585 vmem_alloc_used_read(), vmem_alloc_max
);
589 vmem_alloc_used_add(size
);
590 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
591 vmem_alloc_max
= vmem_alloc_used_read();
593 INIT_HLIST_NODE(&dptr
->kd_hlist
);
594 INIT_LIST_HEAD(&dptr
->kd_list
);
597 dptr
->kd_size
= size
;
598 dptr
->kd_line
= line
;
600 spin_lock_irqsave(&vmem_lock
, irq_flags
);
601 hlist_add_head_rcu(&dptr
->kd_hlist
,
602 &vmem_table
[hash_ptr(ptr
, VMEM_HASH_BITS
)]);
603 list_add_tail(&dptr
->kd_list
, &vmem_list
);
604 spin_unlock_irqrestore(&vmem_lock
, irq_flags
);
606 SDEBUG_LIMIT(SD_INFO
,
607 "vmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
608 (unsigned long long) size
, flags
, func
, line
,
609 ptr
, vmem_alloc_used_read(), vmem_alloc_max
);
614 EXPORT_SYMBOL(vmem_alloc_track
);
617 vmem_free_track(const void *ptr
, size_t size
)
622 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
623 (unsigned long long) size
);
625 dptr
= kmem_del_init(&vmem_lock
, vmem_table
, VMEM_HASH_BITS
, ptr
);
627 /* Must exist in hash due to vmem_alloc() */
630 /* Size must match */
631 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
632 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
633 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
635 vmem_alloc_used_sub(size
);
636 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
637 (unsigned long long) size
, vmem_alloc_used_read(),
640 kfree(dptr
->kd_func
);
642 memset(dptr
, 0x5a, sizeof(kmem_debug_t
));
645 memset(ptr
, 0x5a, size
);
650 EXPORT_SYMBOL(vmem_free_track
);
652 # else /* DEBUG_KMEM_TRACKING */
655 kmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
,
656 int node_alloc
, int node
)
662 * Marked unlikely because we should never be doing this,
663 * we tolerate to up 2 pages but a single page is best.
665 if (unlikely((size
> PAGE_SIZE
* 2) && !(flags
& KM_NODEBUG
))) {
666 SDEBUG(SD_CONSOLE
| SD_WARNING
,
667 "large kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
668 (unsigned long long) size
, flags
, func
, line
,
669 kmem_alloc_used_read(), kmem_alloc_max
);
673 /* Use the correct allocator */
675 ASSERT(!(flags
& __GFP_ZERO
));
676 ptr
= kmalloc_node_nofail(size
, flags
, node
);
677 } else if (flags
& __GFP_ZERO
) {
678 ptr
= kzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
680 ptr
= kmalloc_nofail(size
, flags
);
683 if (unlikely(ptr
== NULL
)) {
684 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
685 "kmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
686 (unsigned long long) size
, flags
, func
, line
,
687 kmem_alloc_used_read(), kmem_alloc_max
);
689 kmem_alloc_used_add(size
);
690 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
691 kmem_alloc_max
= kmem_alloc_used_read();
693 SDEBUG_LIMIT(SD_INFO
,
694 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
695 (unsigned long long) size
, flags
, func
, line
, ptr
,
696 kmem_alloc_used_read(), kmem_alloc_max
);
701 EXPORT_SYMBOL(kmem_alloc_debug
);
704 kmem_free_debug(const void *ptr
, size_t size
)
708 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
709 (unsigned long long) size
);
711 kmem_alloc_used_sub(size
);
712 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
713 (unsigned long long) size
, kmem_alloc_used_read(),
719 EXPORT_SYMBOL(kmem_free_debug
);
722 vmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
)
727 ASSERT(flags
& KM_SLEEP
);
729 /* Use the correct allocator */
730 if (flags
& __GFP_ZERO
) {
731 ptr
= vzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
733 ptr
= vmalloc_nofail(size
, flags
);
736 if (unlikely(ptr
== NULL
)) {
737 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
738 "vmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
739 (unsigned long long) size
, flags
, func
, line
,
740 vmem_alloc_used_read(), vmem_alloc_max
);
742 vmem_alloc_used_add(size
);
743 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
744 vmem_alloc_max
= vmem_alloc_used_read();
746 SDEBUG_LIMIT(SD_INFO
, "vmem_alloc(%llu, 0x%x) = %p "
747 "(%lld/%llu)\n", (unsigned long long) size
, flags
, ptr
,
748 vmem_alloc_used_read(), vmem_alloc_max
);
753 EXPORT_SYMBOL(vmem_alloc_debug
);
756 vmem_free_debug(const void *ptr
, size_t size
)
760 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
761 (unsigned long long) size
);
763 vmem_alloc_used_sub(size
);
764 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
765 (unsigned long long) size
, vmem_alloc_used_read(),
771 EXPORT_SYMBOL(vmem_free_debug
);
773 # endif /* DEBUG_KMEM_TRACKING */
774 #endif /* DEBUG_KMEM */
777 * Slab allocation interfaces
779 * While the Linux slab implementation was inspired by the Solaris
780 * implementation I cannot use it to emulate the Solaris APIs. I
781 * require two features which are not provided by the Linux slab.
783 * 1) Constructors AND destructors. Recent versions of the Linux
784 * kernel have removed support for destructors. This is a deal
785 * breaker for the SPL which contains particularly expensive
786 * initializers for mutex's, condition variables, etc. We also
787 * require a minimal level of cleanup for these data types unlike
788 * many Linux data type which do need to be explicitly destroyed.
790 * 2) Virtual address space backed slab. Callers of the Solaris slab
791 * expect it to work well for both small are very large allocations.
792 * Because of memory fragmentation the Linux slab which is backed
793 * by kmalloc'ed memory performs very badly when confronted with
794 * large numbers of large allocations. Basing the slab on the
795 * virtual address space removes the need for contiguous pages
796 * and greatly improve performance for large allocations.
798 * For these reasons, the SPL has its own slab implementation with
799 * the needed features. It is not as highly optimized as either the
800 * Solaris or Linux slabs, but it should get me most of what is
801 * needed until it can be optimized or obsoleted by another approach.
803 * One serious concern I do have about this method is the relatively
804 * small virtual address space on 32bit arches. This will seriously
805 * constrain the size of the slab caches and their performance.
807 * XXX: Improve the partial slab list by carefully maintaining a
808 * strict ordering of fullest to emptiest slabs based on
809 * the slab reference count. This guarantees the when freeing
810 * slabs back to the system we need only linearly traverse the
811 * last N slabs in the list to discover all the freeable slabs.
813 * XXX: NUMA awareness for optionally allocating memory close to a
814 * particular core. This can be advantageous if you know the slab
815 * object will be short lived and primarily accessed from one core.
817 * XXX: Slab coloring may also yield performance improvements and would
818 * be desirable to implement.
821 struct list_head spl_kmem_cache_list
; /* List of caches */
822 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
823 taskq_t
*spl_kmem_cache_taskq
; /* Task queue for ageing / reclaim */
825 static void spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
);
827 SPL_SHRINKER_CALLBACK_FWD_DECLARE(spl_kmem_cache_generic_shrinker
);
828 SPL_SHRINKER_DECLARE(spl_kmem_cache_shrinker
,
829 spl_kmem_cache_generic_shrinker
, KMC_DEFAULT_SEEKS
);
832 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
838 if (skc
->skc_flags
& KMC_KMEM
)
839 ptr
= (void *)__get_free_pages(flags
, get_order(size
));
841 ptr
= __vmalloc(size
, flags
| __GFP_HIGHMEM
, PAGE_KERNEL
);
843 /* Resulting allocated memory will be page aligned */
844 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
850 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
852 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
856 * The Linux direct reclaim path uses this out of band value to
857 * determine if forward progress is being made. Normally this is
858 * incremented by kmem_freepages() which is part of the various
859 * Linux slab implementations. However, since we are using none
860 * of that infrastructure we are responsible for incrementing it.
862 if (current
->reclaim_state
)
863 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
865 if (skc
->skc_flags
& KMC_KMEM
)
866 free_pages((unsigned long)ptr
, get_order(size
));
872 * Required space for each aligned sks.
874 static inline uint32_t
875 spl_sks_size(spl_kmem_cache_t
*skc
)
877 return P2ROUNDUP_TYPED(sizeof(spl_kmem_slab_t
),
878 skc
->skc_obj_align
, uint32_t);
882 * Required space for each aligned object.
884 static inline uint32_t
885 spl_obj_size(spl_kmem_cache_t
*skc
)
887 uint32_t align
= skc
->skc_obj_align
;
889 return P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
890 P2ROUNDUP_TYPED(sizeof(spl_kmem_obj_t
), align
, uint32_t);
894 * Lookup the spl_kmem_object_t for an object given that object.
896 static inline spl_kmem_obj_t
*
897 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
899 return obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
900 skc
->skc_obj_align
, uint32_t);
904 * Required space for each offslab object taking in to account alignment
905 * restrictions and the power-of-two requirement of kv_alloc().
907 static inline uint32_t
908 spl_offslab_size(spl_kmem_cache_t
*skc
)
910 return 1UL << (highbit(spl_obj_size(skc
)) + 1);
914 * It's important that we pack the spl_kmem_obj_t structure and the
915 * actual objects in to one large address space to minimize the number
916 * of calls to the allocator. It is far better to do a few large
917 * allocations and then subdivide it ourselves. Now which allocator
918 * we use requires balancing a few trade offs.
920 * For small objects we use kmem_alloc() because as long as you are
921 * only requesting a small number of pages (ideally just one) its cheap.
922 * However, when you start requesting multiple pages with kmem_alloc()
923 * it gets increasingly expensive since it requires contiguous pages.
924 * For this reason we shift to vmem_alloc() for slabs of large objects
925 * which removes the need for contiguous pages. We do not use
926 * vmem_alloc() in all cases because there is significant locking
927 * overhead in __get_vm_area_node(). This function takes a single
928 * global lock when acquiring an available virtual address range which
929 * serializes all vmem_alloc()'s for all slab caches. Using slightly
930 * different allocation functions for small and large objects should
931 * give us the best of both worlds.
933 * KMC_ONSLAB KMC_OFFSLAB
935 * +------------------------+ +-----------------+
936 * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+
937 * | skc_obj_size <-+ | | +-----------------+ | |
938 * | spl_kmem_obj_t | | | |
939 * | skc_obj_size <---+ | +-----------------+ | |
940 * | spl_kmem_obj_t | | | skc_obj_size | <-+ |
941 * | ... v | | spl_kmem_obj_t | |
942 * +------------------------+ +-----------------+ v
944 static spl_kmem_slab_t
*
945 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
947 spl_kmem_slab_t
*sks
;
948 spl_kmem_obj_t
*sko
, *n
;
950 uint32_t obj_size
, offslab_size
= 0;
953 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
957 sks
= (spl_kmem_slab_t
*)base
;
958 sks
->sks_magic
= SKS_MAGIC
;
959 sks
->sks_objs
= skc
->skc_slab_objs
;
960 sks
->sks_age
= jiffies
;
961 sks
->sks_cache
= skc
;
962 INIT_LIST_HEAD(&sks
->sks_list
);
963 INIT_LIST_HEAD(&sks
->sks_free_list
);
965 obj_size
= spl_obj_size(skc
);
967 if (skc
->skc_flags
& KMC_OFFSLAB
)
968 offslab_size
= spl_offslab_size(skc
);
970 for (i
= 0; i
< sks
->sks_objs
; i
++) {
971 if (skc
->skc_flags
& KMC_OFFSLAB
) {
972 obj
= kv_alloc(skc
, offslab_size
, flags
);
974 SGOTO(out
, rc
= -ENOMEM
);
976 obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
979 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
980 sko
= spl_sko_from_obj(skc
, obj
);
982 sko
->sko_magic
= SKO_MAGIC
;
984 INIT_LIST_HEAD(&sko
->sko_list
);
985 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
988 list_for_each_entry(sko
, &sks
->sks_free_list
, sko_list
)
990 skc
->skc_ctor(sko
->sko_addr
, skc
->skc_private
, flags
);
993 if (skc
->skc_flags
& KMC_OFFSLAB
)
994 list_for_each_entry_safe(sko
, n
, &sks
->sks_free_list
,
996 kv_free(skc
, sko
->sko_addr
, offslab_size
);
998 kv_free(skc
, base
, skc
->skc_slab_size
);
1006 * Remove a slab from complete or partial list, it must be called with
1007 * the 'skc->skc_lock' held but the actual free must be performed
1008 * outside the lock to prevent deadlocking on vmem addresses.
1011 spl_slab_free(spl_kmem_slab_t
*sks
,
1012 struct list_head
*sks_list
, struct list_head
*sko_list
)
1014 spl_kmem_cache_t
*skc
;
1017 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1018 ASSERT(sks
->sks_ref
== 0);
1020 skc
= sks
->sks_cache
;
1021 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1022 ASSERT(spin_is_locked(&skc
->skc_lock
));
1025 * Update slab/objects counters in the cache, then remove the
1026 * slab from the skc->skc_partial_list. Finally add the slab
1027 * and all its objects in to the private work lists where the
1028 * destructors will be called and the memory freed to the system.
1030 skc
->skc_obj_total
-= sks
->sks_objs
;
1031 skc
->skc_slab_total
--;
1032 list_del(&sks
->sks_list
);
1033 list_add(&sks
->sks_list
, sks_list
);
1034 list_splice_init(&sks
->sks_free_list
, sko_list
);
1040 * Traverses all the partial slabs attached to a cache and free those
1041 * which which are currently empty, and have not been touched for
1042 * skc_delay seconds to avoid thrashing. The count argument is
1043 * passed to optionally cap the number of slabs reclaimed, a count
1044 * of zero means try and reclaim everything. When flag is set we
1045 * always free an available slab regardless of age.
1048 spl_slab_reclaim(spl_kmem_cache_t
*skc
, int count
, int flag
)
1050 spl_kmem_slab_t
*sks
, *m
;
1051 spl_kmem_obj_t
*sko
, *n
;
1052 LIST_HEAD(sks_list
);
1053 LIST_HEAD(sko_list
);
1059 * Move empty slabs and objects which have not been touched in
1060 * skc_delay seconds on to private lists to be freed outside
1061 * the spin lock. This delay time is important to avoid thrashing
1062 * however when flag is set the delay will not be used.
1064 spin_lock(&skc
->skc_lock
);
1065 list_for_each_entry_safe_reverse(sks
,m
,&skc
->skc_partial_list
,sks_list
){
1067 * All empty slabs are at the end of skc->skc_partial_list,
1068 * therefore once a non-empty slab is found we can stop
1069 * scanning. Additionally, stop when reaching the target
1070 * reclaim 'count' if a non-zero threshold is given.
1072 if ((sks
->sks_ref
> 0) || (count
&& i
>= count
))
1075 if (time_after(jiffies
,sks
->sks_age
+skc
->skc_delay
*HZ
)||flag
) {
1076 spl_slab_free(sks
, &sks_list
, &sko_list
);
1080 spin_unlock(&skc
->skc_lock
);
1083 * The following two loops ensure all the object destructors are
1084 * run, any offslab objects are freed, and the slabs themselves
1085 * are freed. This is all done outside the skc->skc_lock since
1086 * this allows the destructor to sleep, and allows us to perform
1087 * a conditional reschedule when a freeing a large number of
1088 * objects and slabs back to the system.
1090 if (skc
->skc_flags
& KMC_OFFSLAB
)
1091 size
= spl_offslab_size(skc
);
1093 list_for_each_entry_safe(sko
, n
, &sko_list
, sko_list
) {
1094 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1097 skc
->skc_dtor(sko
->sko_addr
, skc
->skc_private
);
1099 if (skc
->skc_flags
& KMC_OFFSLAB
)
1100 kv_free(skc
, sko
->sko_addr
, size
);
1105 list_for_each_entry_safe(sks
, m
, &sks_list
, sks_list
) {
1106 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1107 kv_free(skc
, sks
, skc
->skc_slab_size
);
1114 static spl_kmem_emergency_t
*
1115 spl_emergency_search(struct rb_root
*root
, void *obj
)
1117 struct rb_node
*node
= root
->rb_node
;
1118 spl_kmem_emergency_t
*ske
;
1119 unsigned long address
= (unsigned long)obj
;
1122 ske
= container_of(node
, spl_kmem_emergency_t
, ske_node
);
1124 if (address
< (unsigned long)ske
->ske_obj
)
1125 node
= node
->rb_left
;
1126 else if (address
> (unsigned long)ske
->ske_obj
)
1127 node
= node
->rb_right
;
1136 spl_emergency_insert(struct rb_root
*root
, spl_kmem_emergency_t
*ske
)
1138 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
1139 spl_kmem_emergency_t
*ske_tmp
;
1140 unsigned long address
= (unsigned long)ske
->ske_obj
;
1143 ske_tmp
= container_of(*new, spl_kmem_emergency_t
, ske_node
);
1146 if (address
< (unsigned long)ske_tmp
->ske_obj
)
1147 new = &((*new)->rb_left
);
1148 else if (address
> (unsigned long)ske_tmp
->ske_obj
)
1149 new = &((*new)->rb_right
);
1154 rb_link_node(&ske
->ske_node
, parent
, new);
1155 rb_insert_color(&ske
->ske_node
, root
);
1161 * Allocate a single emergency object and track it in a red black tree.
1164 spl_emergency_alloc(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1166 spl_kmem_emergency_t
*ske
;
1170 /* Last chance use a partial slab if one now exists */
1171 spin_lock(&skc
->skc_lock
);
1172 empty
= list_empty(&skc
->skc_partial_list
);
1173 spin_unlock(&skc
->skc_lock
);
1177 ske
= kmalloc(sizeof(*ske
), flags
);
1181 ske
->ske_obj
= kmalloc(skc
->skc_obj_size
, flags
);
1182 if (ske
->ske_obj
== NULL
) {
1187 spin_lock(&skc
->skc_lock
);
1188 empty
= spl_emergency_insert(&skc
->skc_emergency_tree
, ske
);
1189 if (likely(empty
)) {
1190 skc
->skc_obj_total
++;
1191 skc
->skc_obj_emergency
++;
1192 if (skc
->skc_obj_emergency
> skc
->skc_obj_emergency_max
)
1193 skc
->skc_obj_emergency_max
= skc
->skc_obj_emergency
;
1195 spin_unlock(&skc
->skc_lock
);
1197 if (unlikely(!empty
)) {
1198 kfree(ske
->ske_obj
);
1204 skc
->skc_ctor(ske
->ske_obj
, skc
->skc_private
, flags
);
1206 *obj
= ske
->ske_obj
;
1212 * Locate the passed object in the red black tree and free it.
1215 spl_emergency_free(spl_kmem_cache_t
*skc
, void *obj
)
1217 spl_kmem_emergency_t
*ske
;
1220 spin_lock(&skc
->skc_lock
);
1221 ske
= spl_emergency_search(&skc
->skc_emergency_tree
, obj
);
1223 rb_erase(&ske
->ske_node
, &skc
->skc_emergency_tree
);
1224 skc
->skc_obj_emergency
--;
1225 skc
->skc_obj_total
--;
1227 spin_unlock(&skc
->skc_lock
);
1229 if (unlikely(ske
== NULL
))
1233 skc
->skc_dtor(ske
->ske_obj
, skc
->skc_private
);
1235 kfree(ske
->ske_obj
);
1242 * Release objects from the per-cpu magazine back to their slab. The flush
1243 * argument contains the max number of entries to remove from the magazine.
1246 __spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
1248 int i
, count
= MIN(flush
, skm
->skm_avail
);
1251 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1252 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1253 ASSERT(spin_is_locked(&skc
->skc_lock
));
1255 for (i
= 0; i
< count
; i
++)
1256 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
1258 skm
->skm_avail
-= count
;
1259 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
1260 sizeof(void *) * skm
->skm_avail
);
1266 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
1268 spin_lock(&skc
->skc_lock
);
1269 __spl_cache_flush(skc
, skm
, flush
);
1270 spin_unlock(&skc
->skc_lock
);
1274 spl_magazine_age(void *data
)
1276 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1277 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
1279 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1280 ASSERT(skm
->skm_cpu
== smp_processor_id());
1281 ASSERT(irqs_disabled());
1283 /* There are no available objects or they are too young to age out */
1284 if ((skm
->skm_avail
== 0) ||
1285 time_before(jiffies
, skm
->skm_age
+ skc
->skc_delay
* HZ
))
1289 * Because we're executing in interrupt context we may have
1290 * interrupted the holder of this lock. To avoid a potential
1291 * deadlock return if the lock is contended.
1293 if (!spin_trylock(&skc
->skc_lock
))
1296 __spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1297 spin_unlock(&skc
->skc_lock
);
1301 * Called regularly to keep a downward pressure on the cache.
1303 * Objects older than skc->skc_delay seconds in the per-cpu magazines will
1304 * be returned to the caches. This is done to prevent idle magazines from
1305 * holding memory which could be better used elsewhere. The delay is
1306 * present to prevent thrashing the magazine.
1308 * The newly released objects may result in empty partial slabs. Those
1309 * slabs should be released to the system. Otherwise moving the objects
1310 * out of the magazines is just wasted work.
1313 spl_cache_age(void *data
)
1315 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1318 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1320 atomic_inc(&skc
->skc_ref
);
1321 spl_on_each_cpu(spl_magazine_age
, skc
, 1);
1322 spl_slab_reclaim(skc
, skc
->skc_reap
, 0);
1324 while (!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && !id
) {
1325 id
= taskq_dispatch_delay(
1326 spl_kmem_cache_taskq
, spl_cache_age
, skc
, TQ_SLEEP
,
1327 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1329 /* Destroy issued after dispatch immediately cancel it */
1330 if (test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && id
)
1331 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1334 spin_lock(&skc
->skc_lock
);
1335 skc
->skc_taskqid
= id
;
1336 spin_unlock(&skc
->skc_lock
);
1338 atomic_dec(&skc
->skc_ref
);
1342 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
1343 * When on-slab we want to target SPL_KMEM_CACHE_OBJ_PER_SLAB. However,
1344 * for very small objects we may end up with more than this so as not
1345 * to waste space in the minimal allocation of a single page. Also for
1346 * very large objects we may use as few as SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN,
1347 * lower than this and we will fail.
1350 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
1352 uint32_t sks_size
, obj_size
, max_size
;
1354 if (skc
->skc_flags
& KMC_OFFSLAB
) {
1355 *objs
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
1356 *size
= sizeof(spl_kmem_slab_t
);
1358 sks_size
= spl_sks_size(skc
);
1359 obj_size
= spl_obj_size(skc
);
1361 if (skc
->skc_flags
& KMC_KMEM
)
1362 max_size
= ((uint32_t)1 << (MAX_ORDER
-3)) * PAGE_SIZE
;
1364 max_size
= (32 * 1024 * 1024);
1366 /* Power of two sized slab */
1367 for (*size
= PAGE_SIZE
; *size
<= max_size
; *size
*= 2) {
1368 *objs
= (*size
- sks_size
) / obj_size
;
1369 if (*objs
>= SPL_KMEM_CACHE_OBJ_PER_SLAB
)
1374 * Unable to satisfy target objects per slab, fall back to
1375 * allocating a maximally sized slab and assuming it can
1376 * contain the minimum objects count use it. If not fail.
1379 *objs
= (*size
- sks_size
) / obj_size
;
1380 if (*objs
>= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN
)
1388 * Make a guess at reasonable per-cpu magazine size based on the size of
1389 * each object and the cost of caching N of them in each magazine. Long
1390 * term this should really adapt based on an observed usage heuristic.
1393 spl_magazine_size(spl_kmem_cache_t
*skc
)
1395 uint32_t obj_size
= spl_obj_size(skc
);
1399 /* Per-magazine sizes below assume a 4Kib page size */
1400 if (obj_size
> (PAGE_SIZE
* 256))
1401 size
= 4; /* Minimum 4Mib per-magazine */
1402 else if (obj_size
> (PAGE_SIZE
* 32))
1403 size
= 16; /* Minimum 2Mib per-magazine */
1404 else if (obj_size
> (PAGE_SIZE
))
1405 size
= 64; /* Minimum 256Kib per-magazine */
1406 else if (obj_size
> (PAGE_SIZE
/ 4))
1407 size
= 128; /* Minimum 128Kib per-magazine */
1415 * Allocate a per-cpu magazine to associate with a specific core.
1417 static spl_kmem_magazine_t
*
1418 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
1420 spl_kmem_magazine_t
*skm
;
1421 int size
= sizeof(spl_kmem_magazine_t
) +
1422 sizeof(void *) * skc
->skc_mag_size
;
1425 skm
= kmem_alloc_node(size
, KM_SLEEP
, cpu_to_node(cpu
));
1427 skm
->skm_magic
= SKM_MAGIC
;
1429 skm
->skm_size
= skc
->skc_mag_size
;
1430 skm
->skm_refill
= skc
->skc_mag_refill
;
1431 skm
->skm_cache
= skc
;
1432 skm
->skm_age
= jiffies
;
1440 * Free a per-cpu magazine associated with a specific core.
1443 spl_magazine_free(spl_kmem_magazine_t
*skm
)
1445 int size
= sizeof(spl_kmem_magazine_t
) +
1446 sizeof(void *) * skm
->skm_size
;
1449 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1450 ASSERT(skm
->skm_avail
== 0);
1452 kmem_free(skm
, size
);
1457 * Create all pre-cpu magazines of reasonable sizes.
1460 spl_magazine_create(spl_kmem_cache_t
*skc
)
1465 skc
->skc_mag_size
= spl_magazine_size(skc
);
1466 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
1468 for_each_online_cpu(i
) {
1469 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
1470 if (!skc
->skc_mag
[i
]) {
1471 for (i
--; i
>= 0; i
--)
1472 spl_magazine_free(skc
->skc_mag
[i
]);
1482 * Destroy all pre-cpu magazines.
1485 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
1487 spl_kmem_magazine_t
*skm
;
1491 for_each_online_cpu(i
) {
1492 skm
= skc
->skc_mag
[i
];
1493 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1494 spl_magazine_free(skm
);
1501 * Create a object cache based on the following arguments:
1503 * size cache object size
1504 * align cache object alignment
1505 * ctor cache object constructor
1506 * dtor cache object destructor
1507 * reclaim cache object reclaim
1508 * priv cache private data for ctor/dtor/reclaim
1509 * vmp unused must be NULL
1511 * KMC_NOTOUCH Disable cache object aging (unsupported)
1512 * KMC_NODEBUG Disable debugging (unsupported)
1513 * KMC_NOMAGAZINE Disable magazine (unsupported)
1514 * KMC_NOHASH Disable hashing (unsupported)
1515 * KMC_QCACHE Disable qcache (unsupported)
1516 * KMC_KMEM Force kmem backed cache
1517 * KMC_VMEM Force vmem backed cache
1518 * KMC_OFFSLAB Locate objects off the slab
1521 spl_kmem_cache_create(char *name
, size_t size
, size_t align
,
1522 spl_kmem_ctor_t ctor
,
1523 spl_kmem_dtor_t dtor
,
1524 spl_kmem_reclaim_t reclaim
,
1525 void *priv
, void *vmp
, int flags
)
1527 spl_kmem_cache_t
*skc
;
1531 ASSERTF(!(flags
& KMC_NOMAGAZINE
), "Bad KMC_NOMAGAZINE (%x)\n", flags
);
1532 ASSERTF(!(flags
& KMC_NOHASH
), "Bad KMC_NOHASH (%x)\n", flags
);
1533 ASSERTF(!(flags
& KMC_QCACHE
), "Bad KMC_QCACHE (%x)\n", flags
);
1534 ASSERT(vmp
== NULL
);
1539 * Allocate memory for a new cache an initialize it. Unfortunately,
1540 * this usually ends up being a large allocation of ~32k because
1541 * we need to allocate enough memory for the worst case number of
1542 * cpus in the magazine, skc_mag[NR_CPUS]. Because of this we
1543 * explicitly pass KM_NODEBUG to suppress the kmem warning
1545 skc
= kmem_zalloc(sizeof(*skc
), KM_SLEEP
| KM_NODEBUG
);
1549 skc
->skc_magic
= SKC_MAGIC
;
1550 skc
->skc_name_size
= strlen(name
) + 1;
1551 skc
->skc_name
= (char *)kmem_alloc(skc
->skc_name_size
, KM_SLEEP
);
1552 if (skc
->skc_name
== NULL
) {
1553 kmem_free(skc
, sizeof(*skc
));
1556 strncpy(skc
->skc_name
, name
, skc
->skc_name_size
);
1558 skc
->skc_ctor
= ctor
;
1559 skc
->skc_dtor
= dtor
;
1560 skc
->skc_reclaim
= reclaim
;
1561 skc
->skc_private
= priv
;
1563 skc
->skc_flags
= flags
;
1564 skc
->skc_obj_size
= size
;
1565 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
1566 skc
->skc_delay
= SPL_KMEM_CACHE_DELAY
;
1567 skc
->skc_reap
= SPL_KMEM_CACHE_REAP
;
1568 atomic_set(&skc
->skc_ref
, 0);
1570 INIT_LIST_HEAD(&skc
->skc_list
);
1571 INIT_LIST_HEAD(&skc
->skc_complete_list
);
1572 INIT_LIST_HEAD(&skc
->skc_partial_list
);
1573 skc
->skc_emergency_tree
= RB_ROOT
;
1574 spin_lock_init(&skc
->skc_lock
);
1575 init_waitqueue_head(&skc
->skc_waitq
);
1576 skc
->skc_slab_fail
= 0;
1577 skc
->skc_slab_create
= 0;
1578 skc
->skc_slab_destroy
= 0;
1579 skc
->skc_slab_total
= 0;
1580 skc
->skc_slab_alloc
= 0;
1581 skc
->skc_slab_max
= 0;
1582 skc
->skc_obj_total
= 0;
1583 skc
->skc_obj_alloc
= 0;
1584 skc
->skc_obj_max
= 0;
1585 skc
->skc_obj_deadlock
= 0;
1586 skc
->skc_obj_emergency
= 0;
1587 skc
->skc_obj_emergency_max
= 0;
1590 VERIFY(ISP2(align
));
1591 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
); /* Min alignment */
1592 VERIFY3U(align
, <=, PAGE_SIZE
); /* Max alignment */
1593 skc
->skc_obj_align
= align
;
1596 /* If none passed select a cache type based on object size */
1597 if (!(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
))) {
1598 if (spl_obj_size(skc
) < (PAGE_SIZE
/ 8))
1599 skc
->skc_flags
|= KMC_KMEM
;
1601 skc
->skc_flags
|= KMC_VMEM
;
1604 rc
= spl_slab_size(skc
, &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
1608 rc
= spl_magazine_create(skc
);
1612 skc
->skc_taskqid
= taskq_dispatch_delay(spl_kmem_cache_taskq
,
1613 spl_cache_age
, skc
, TQ_SLEEP
,
1614 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1616 down_write(&spl_kmem_cache_sem
);
1617 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
1618 up_write(&spl_kmem_cache_sem
);
1622 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1623 kmem_free(skc
, sizeof(*skc
));
1626 EXPORT_SYMBOL(spl_kmem_cache_create
);
1629 * Register a move callback to for cache defragmentation.
1630 * XXX: Unimplemented but harmless to stub out for now.
1633 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
1634 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
1636 ASSERT(move
!= NULL
);
1638 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
1641 * Destroy a cache and all objects associated with the cache.
1644 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
1646 DECLARE_WAIT_QUEUE_HEAD(wq
);
1650 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1652 down_write(&spl_kmem_cache_sem
);
1653 list_del_init(&skc
->skc_list
);
1654 up_write(&spl_kmem_cache_sem
);
1656 /* Cancel any and wait for any pending delayed tasks */
1657 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1659 spin_lock(&skc
->skc_lock
);
1660 id
= skc
->skc_taskqid
;
1661 spin_unlock(&skc
->skc_lock
);
1663 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1665 /* Wait until all current callers complete, this is mainly
1666 * to catch the case where a low memory situation triggers a
1667 * cache reaping action which races with this destroy. */
1668 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
1670 spl_magazine_destroy(skc
);
1671 spl_slab_reclaim(skc
, 0, 1);
1672 spin_lock(&skc
->skc_lock
);
1674 /* Validate there are no objects in use and free all the
1675 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */
1676 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
1677 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
1678 ASSERT3U(skc
->skc_slab_total
, ==, 0);
1679 ASSERT3U(skc
->skc_obj_total
, ==, 0);
1680 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
1681 ASSERT(list_empty(&skc
->skc_complete_list
));
1683 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1684 spin_unlock(&skc
->skc_lock
);
1686 kmem_free(skc
, sizeof(*skc
));
1690 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
1693 * Allocate an object from a slab attached to the cache. This is used to
1694 * repopulate the per-cpu magazine caches in batches when they run low.
1697 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
1699 spl_kmem_obj_t
*sko
;
1701 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1702 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1703 ASSERT(spin_is_locked(&skc
->skc_lock
));
1705 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
1706 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1707 ASSERT(sko
->sko_addr
!= NULL
);
1709 /* Remove from sks_free_list */
1710 list_del_init(&sko
->sko_list
);
1712 sks
->sks_age
= jiffies
;
1714 skc
->skc_obj_alloc
++;
1716 /* Track max obj usage statistics */
1717 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
1718 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
1720 /* Track max slab usage statistics */
1721 if (sks
->sks_ref
== 1) {
1722 skc
->skc_slab_alloc
++;
1724 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
1725 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
1728 return sko
->sko_addr
;
1732 * Generic slab allocation function to run by the global work queues.
1733 * It is responsible for allocating a new slab, linking it in to the list
1734 * of partial slabs, and then waking any waiters.
1737 spl_cache_grow_work(void *data
)
1739 spl_kmem_alloc_t
*ska
= (spl_kmem_alloc_t
*)data
;
1740 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
1741 spl_kmem_slab_t
*sks
;
1743 sks
= spl_slab_alloc(skc
, ska
->ska_flags
| __GFP_NORETRY
| KM_NODEBUG
);
1744 spin_lock(&skc
->skc_lock
);
1746 skc
->skc_slab_total
++;
1747 skc
->skc_obj_total
+= sks
->sks_objs
;
1748 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1751 atomic_dec(&skc
->skc_ref
);
1752 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1753 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1754 wake_up_all(&skc
->skc_waitq
);
1755 spin_unlock(&skc
->skc_lock
);
1761 * Returns non-zero when a new slab should be available.
1764 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
1766 return !test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1770 spl_cache_reclaim_wait(void *word
)
1777 * No available objects on any slabs, create a new slab.
1780 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1785 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1790 * Before allocating a new slab wait for any reaping to complete and
1791 * then return so the local magazine can be rechecked for new objects.
1793 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1794 rc
= wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1795 spl_cache_reclaim_wait
, TASK_UNINTERRUPTIBLE
);
1796 SRETURN(rc
? rc
: -EAGAIN
);
1800 * This is handled by dispatching a work request to the global work
1801 * queue. This allows us to asynchronously allocate a new slab while
1802 * retaining the ability to safely fall back to a smaller synchronous
1803 * allocations to ensure forward progress is always maintained.
1805 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1806 spl_kmem_alloc_t
*ska
;
1808 ska
= kmalloc(sizeof(*ska
), flags
);
1810 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1811 wake_up_all(&skc
->skc_waitq
);
1815 atomic_inc(&skc
->skc_ref
);
1816 ska
->ska_cache
= skc
;
1817 ska
->ska_flags
= flags
& ~__GFP_FS
;
1818 taskq_init_ent(&ska
->ska_tqe
);
1819 taskq_dispatch_ent(spl_kmem_cache_taskq
,
1820 spl_cache_grow_work
, ska
, 0, &ska
->ska_tqe
);
1824 * The goal here is to only detect the rare case where a virtual slab
1825 * allocation has deadlocked. We must be careful to minimize the use
1826 * of emergency objects which are more expensive to track. Therefore,
1827 * we set a very long timeout for the asynchronous allocation and if
1828 * the timeout is reached the cache is flagged as deadlocked. From
1829 * this point only new emergency objects will be allocated until the
1830 * asynchronous allocation completes and clears the deadlocked flag.
1832 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1833 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1835 remaining
= wait_event_timeout(skc
->skc_waitq
,
1836 spl_cache_grow_wait(skc
), HZ
);
1838 if (!remaining
&& test_bit(KMC_BIT_VMEM
, &skc
->skc_flags
)) {
1839 spin_lock(&skc
->skc_lock
);
1840 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1841 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1842 skc
->skc_obj_deadlock
++;
1844 spin_unlock(&skc
->skc_lock
);
1854 * Refill a per-cpu magazine with objects from the slabs for this cache.
1855 * Ideally the magazine can be repopulated using existing objects which have
1856 * been released, however if we are unable to locate enough free objects new
1857 * slabs of objects will be created. On success NULL is returned, otherwise
1858 * the address of a single emergency object is returned for use by the caller.
1861 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1863 spl_kmem_slab_t
*sks
;
1864 int count
= 0, rc
, refill
;
1868 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1869 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1871 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1872 spin_lock(&skc
->skc_lock
);
1874 while (refill
> 0) {
1875 /* No slabs available we may need to grow the cache */
1876 if (list_empty(&skc
->skc_partial_list
)) {
1877 spin_unlock(&skc
->skc_lock
);
1880 rc
= spl_cache_grow(skc
, flags
, &obj
);
1881 local_irq_disable();
1883 /* Emergency object for immediate use by caller */
1884 if (rc
== 0 && obj
!= NULL
)
1890 /* Rescheduled to different CPU skm is not local */
1891 if (skm
!= skc
->skc_mag
[smp_processor_id()])
1894 /* Potentially rescheduled to the same CPU but
1895 * allocations may have occurred from this CPU while
1896 * we were sleeping so recalculate max refill. */
1897 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
1899 spin_lock(&skc
->skc_lock
);
1903 /* Grab the next available slab */
1904 sks
= list_entry((&skc
->skc_partial_list
)->next
,
1905 spl_kmem_slab_t
, sks_list
);
1906 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1907 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
1908 ASSERT(!list_empty(&sks
->sks_free_list
));
1910 /* Consume as many objects as needed to refill the requested
1911 * cache. We must also be careful not to overfill it. */
1912 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 && ++count
) {
1913 ASSERT(skm
->skm_avail
< skm
->skm_size
);
1914 ASSERT(count
< skm
->skm_size
);
1915 skm
->skm_objs
[skm
->skm_avail
++]=spl_cache_obj(skc
,sks
);
1918 /* Move slab to skc_complete_list when full */
1919 if (sks
->sks_ref
== sks
->sks_objs
) {
1920 list_del(&sks
->sks_list
);
1921 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
1925 spin_unlock(&skc
->skc_lock
);
1931 * Release an object back to the slab from which it came.
1934 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
1936 spl_kmem_slab_t
*sks
= NULL
;
1937 spl_kmem_obj_t
*sko
= NULL
;
1940 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1941 ASSERT(spin_is_locked(&skc
->skc_lock
));
1943 sko
= spl_sko_from_obj(skc
, obj
);
1944 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1945 sks
= sko
->sko_slab
;
1946 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1947 ASSERT(sks
->sks_cache
== skc
);
1948 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
1950 sks
->sks_age
= jiffies
;
1952 skc
->skc_obj_alloc
--;
1954 /* Move slab to skc_partial_list when no longer full. Slabs
1955 * are added to the head to keep the partial list is quasi-full
1956 * sorted order. Fuller at the head, emptier at the tail. */
1957 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
1958 list_del(&sks
->sks_list
);
1959 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
1962 /* Move empty slabs to the end of the partial list so
1963 * they can be easily found and freed during reclamation. */
1964 if (sks
->sks_ref
== 0) {
1965 list_del(&sks
->sks_list
);
1966 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1967 skc
->skc_slab_alloc
--;
1974 * Allocate an object from the per-cpu magazine, or if the magazine
1975 * is empty directly allocate from a slab and repopulate the magazine.
1978 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
1980 spl_kmem_magazine_t
*skm
;
1981 unsigned long irq_flags
;
1985 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1986 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1987 ASSERT(flags
& KM_SLEEP
);
1988 atomic_inc(&skc
->skc_ref
);
1989 local_irq_save(irq_flags
);
1992 /* Safe to update per-cpu structure without lock, but
1993 * in the restart case we must be careful to reacquire
1994 * the local magazine since this may have changed
1995 * when we need to grow the cache. */
1996 skm
= skc
->skc_mag
[smp_processor_id()];
1997 ASSERTF(skm
->skm_magic
== SKM_MAGIC
, "%x != %x: %s/%p/%p %x/%x/%x\n",
1998 skm
->skm_magic
, SKM_MAGIC
, skc
->skc_name
, skc
, skm
,
1999 skm
->skm_size
, skm
->skm_refill
, skm
->skm_avail
);
2001 if (likely(skm
->skm_avail
)) {
2002 /* Object available in CPU cache, use it */
2003 obj
= skm
->skm_objs
[--skm
->skm_avail
];
2004 skm
->skm_age
= jiffies
;
2006 obj
= spl_cache_refill(skc
, skm
, flags
);
2008 SGOTO(restart
, obj
= NULL
);
2011 local_irq_restore(irq_flags
);
2013 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
2015 /* Pre-emptively migrate object to CPU L1 cache */
2017 atomic_dec(&skc
->skc_ref
);
2021 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
2024 * Free an object back to the local per-cpu magazine, there is no
2025 * guarantee that this is the same magazine the object was originally
2026 * allocated from. We may need to flush entire from the magazine
2027 * back to the slabs to make space.
2030 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
2032 spl_kmem_magazine_t
*skm
;
2033 unsigned long flags
;
2036 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2037 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2038 atomic_inc(&skc
->skc_ref
);
2041 * Only virtual slabs may have emergency objects and these objects
2042 * are guaranteed to have physical addresses. They must be removed
2043 * from the tree of emergency objects and the freed.
2045 if ((skc
->skc_flags
& KMC_VMEM
) && !kmem_virt(obj
))
2046 SGOTO(out
, spl_emergency_free(skc
, obj
));
2048 local_irq_save(flags
);
2050 /* Safe to update per-cpu structure without lock, but
2051 * no remote memory allocation tracking is being performed
2052 * it is entirely possible to allocate an object from one
2053 * CPU cache and return it to another. */
2054 skm
= skc
->skc_mag
[smp_processor_id()];
2055 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
2057 /* Per-CPU cache full, flush it to make space */
2058 if (unlikely(skm
->skm_avail
>= skm
->skm_size
))
2059 spl_cache_flush(skc
, skm
, skm
->skm_refill
);
2061 /* Available space in cache, use it */
2062 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
2064 local_irq_restore(flags
);
2066 atomic_dec(&skc
->skc_ref
);
2070 EXPORT_SYMBOL(spl_kmem_cache_free
);
2073 * The generic shrinker function for all caches. Under Linux a shrinker
2074 * may not be tightly coupled with a slab cache. In fact Linux always
2075 * systematically tries calling all registered shrinker callbacks which
2076 * report that they contain unused objects. Because of this we only
2077 * register one shrinker function in the shim layer for all slab caches.
2078 * We always attempt to shrink all caches when this generic shrinker
2079 * is called. The shrinker should return the number of free objects
2080 * in the cache when called with nr_to_scan == 0 but not attempt to
2081 * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan
2082 * objects should be freed, which differs from Solaris semantics.
2083 * Solaris semantics are to free all available objects which may (and
2084 * probably will) be more objects than the requested nr_to_scan.
2087 __spl_kmem_cache_generic_shrinker(struct shrinker
*shrink
,
2088 struct shrink_control
*sc
)
2090 spl_kmem_cache_t
*skc
;
2093 down_read(&spl_kmem_cache_sem
);
2094 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
2096 spl_kmem_cache_reap_now(skc
,
2097 MAX(sc
->nr_to_scan
>> fls64(skc
->skc_slab_objs
), 1));
2100 * Presume everything alloc'ed in reclaimable, this ensures
2101 * we are called again with nr_to_scan > 0 so can try and
2102 * reclaim. The exact number is not important either so
2103 * we forgo taking this already highly contented lock.
2105 unused
+= skc
->skc_obj_alloc
;
2107 up_read(&spl_kmem_cache_sem
);
2109 return (unused
* sysctl_vfs_cache_pressure
) / 100;
2112 SPL_SHRINKER_CALLBACK_WRAPPER(spl_kmem_cache_generic_shrinker
);
2115 * Call the registered reclaim function for a cache. Depending on how
2116 * many and which objects are released it may simply repopulate the
2117 * local magazine which will then need to age-out. Objects which cannot
2118 * fit in the magazine we will be released back to their slabs which will
2119 * also need to age out before being release. This is all just best
2120 * effort and we do not want to thrash creating and destroying slabs.
2123 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
, int count
)
2127 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2128 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2130 /* Prevent concurrent cache reaping when contended */
2131 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
2136 atomic_inc(&skc
->skc_ref
);
2139 * When a reclaim function is available it may be invoked repeatedly
2140 * until at least a single slab can be freed. This ensures that we
2141 * do free memory back to the system. This helps minimize the chance
2142 * of an OOM event when the bulk of memory is used by the slab.
2144 * When free slabs are already available the reclaim callback will be
2145 * skipped. Additionally, if no forward progress is detected despite
2146 * a reclaim function the cache will be skipped to avoid deadlock.
2148 * Longer term this would be the correct place to add the code which
2149 * repacks the slabs in order minimize fragmentation.
2151 if (skc
->skc_reclaim
) {
2152 uint64_t objects
= UINT64_MAX
;
2156 spin_lock(&skc
->skc_lock
);
2158 (skc
->skc_slab_total
> 0) &&
2159 ((skc
->skc_slab_total
- skc
->skc_slab_alloc
) == 0) &&
2160 (skc
->skc_obj_alloc
< objects
);
2162 objects
= skc
->skc_obj_alloc
;
2163 spin_unlock(&skc
->skc_lock
);
2166 skc
->skc_reclaim(skc
->skc_private
);
2168 } while (do_reclaim
);
2171 /* Reclaim from the cache, ignoring it's age and delay. */
2172 spl_slab_reclaim(skc
, count
, 1);
2173 clear_bit(KMC_BIT_REAPING
, &skc
->skc_flags
);
2174 smp_mb__after_clear_bit();
2175 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
2177 atomic_dec(&skc
->skc_ref
);
2181 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
2184 * Reap all free slabs from all registered caches.
2189 struct shrink_control sc
;
2191 sc
.nr_to_scan
= KMC_REAP_CHUNK
;
2192 sc
.gfp_mask
= GFP_KERNEL
;
2194 __spl_kmem_cache_generic_shrinker(NULL
, &sc
);
2196 EXPORT_SYMBOL(spl_kmem_reap
);
2198 #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING)
2200 spl_sprintf_addr(kmem_debug_t
*kd
, char *str
, int len
, int min
)
2202 int size
= ((len
- 1) < kd
->kd_size
) ? (len
- 1) : kd
->kd_size
;
2205 ASSERT(str
!= NULL
&& len
>= 17);
2206 memset(str
, 0, len
);
2208 /* Check for a fully printable string, and while we are at
2209 * it place the printable characters in the passed buffer. */
2210 for (i
= 0; i
< size
; i
++) {
2211 str
[i
] = ((char *)(kd
->kd_addr
))[i
];
2212 if (isprint(str
[i
])) {
2215 /* Minimum number of printable characters found
2216 * to make it worthwhile to print this as ascii. */
2226 sprintf(str
, "%02x%02x%02x%02x%02x%02x%02x%02x",
2227 *((uint8_t *)kd
->kd_addr
),
2228 *((uint8_t *)kd
->kd_addr
+ 2),
2229 *((uint8_t *)kd
->kd_addr
+ 4),
2230 *((uint8_t *)kd
->kd_addr
+ 6),
2231 *((uint8_t *)kd
->kd_addr
+ 8),
2232 *((uint8_t *)kd
->kd_addr
+ 10),
2233 *((uint8_t *)kd
->kd_addr
+ 12),
2234 *((uint8_t *)kd
->kd_addr
+ 14));
2241 spl_kmem_init_tracking(struct list_head
*list
, spinlock_t
*lock
, int size
)
2246 spin_lock_init(lock
);
2247 INIT_LIST_HEAD(list
);
2249 for (i
= 0; i
< size
; i
++)
2250 INIT_HLIST_HEAD(&kmem_table
[i
]);
2256 spl_kmem_fini_tracking(struct list_head
*list
, spinlock_t
*lock
)
2258 unsigned long flags
;
2263 spin_lock_irqsave(lock
, flags
);
2264 if (!list_empty(list
))
2265 printk(KERN_WARNING
"%-16s %-5s %-16s %s:%s\n", "address",
2266 "size", "data", "func", "line");
2268 list_for_each_entry(kd
, list
, kd_list
)
2269 printk(KERN_WARNING
"%p %-5d %-16s %s:%d\n", kd
->kd_addr
,
2270 (int)kd
->kd_size
, spl_sprintf_addr(kd
, str
, 17, 8),
2271 kd
->kd_func
, kd
->kd_line
);
2273 spin_unlock_irqrestore(lock
, flags
);
2276 #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2277 #define spl_kmem_init_tracking(list, lock, size)
2278 #define spl_kmem_fini_tracking(list, lock)
2279 #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2282 spl_kmem_init_globals(void)
2286 /* For now all zones are includes, it may be wise to restrict
2287 * this to normal and highmem zones if we see problems. */
2288 for_each_zone(zone
) {
2290 if (!populated_zone(zone
))
2293 minfree
+= min_wmark_pages(zone
);
2294 desfree
+= low_wmark_pages(zone
);
2295 lotsfree
+= high_wmark_pages(zone
);
2298 /* Solaris default values */
2299 swapfs_minfree
= MAX(2*1024*1024 >> PAGE_SHIFT
, physmem
>> 3);
2300 swapfs_reserve
= MIN(4*1024*1024 >> PAGE_SHIFT
, physmem
>> 4);
2304 * Called at module init when it is safe to use spl_kallsyms_lookup_name()
2307 spl_kmem_init_kallsyms_lookup(void)
2309 #ifndef HAVE_GET_VMALLOC_INFO
2310 get_vmalloc_info_fn
= (get_vmalloc_info_t
)
2311 spl_kallsyms_lookup_name("get_vmalloc_info");
2312 if (!get_vmalloc_info_fn
) {
2313 printk(KERN_ERR
"Error: Unknown symbol get_vmalloc_info\n");
2316 #endif /* HAVE_GET_VMALLOC_INFO */
2318 #ifdef HAVE_PGDAT_HELPERS
2319 # ifndef HAVE_FIRST_ONLINE_PGDAT
2320 first_online_pgdat_fn
= (first_online_pgdat_t
)
2321 spl_kallsyms_lookup_name("first_online_pgdat");
2322 if (!first_online_pgdat_fn
) {
2323 printk(KERN_ERR
"Error: Unknown symbol first_online_pgdat\n");
2326 # endif /* HAVE_FIRST_ONLINE_PGDAT */
2328 # ifndef HAVE_NEXT_ONLINE_PGDAT
2329 next_online_pgdat_fn
= (next_online_pgdat_t
)
2330 spl_kallsyms_lookup_name("next_online_pgdat");
2331 if (!next_online_pgdat_fn
) {
2332 printk(KERN_ERR
"Error: Unknown symbol next_online_pgdat\n");
2335 # endif /* HAVE_NEXT_ONLINE_PGDAT */
2337 # ifndef HAVE_NEXT_ZONE
2338 next_zone_fn
= (next_zone_t
)
2339 spl_kallsyms_lookup_name("next_zone");
2340 if (!next_zone_fn
) {
2341 printk(KERN_ERR
"Error: Unknown symbol next_zone\n");
2344 # endif /* HAVE_NEXT_ZONE */
2346 #else /* HAVE_PGDAT_HELPERS */
2348 # ifndef HAVE_PGDAT_LIST
2349 pgdat_list_addr
= *(struct pglist_data
**)
2350 spl_kallsyms_lookup_name("pgdat_list");
2351 if (!pgdat_list_addr
) {
2352 printk(KERN_ERR
"Error: Unknown symbol pgdat_list\n");
2355 # endif /* HAVE_PGDAT_LIST */
2356 #endif /* HAVE_PGDAT_HELPERS */
2358 #if defined(NEED_GET_ZONE_COUNTS) && !defined(HAVE_GET_ZONE_COUNTS)
2359 get_zone_counts_fn
= (get_zone_counts_t
)
2360 spl_kallsyms_lookup_name("get_zone_counts");
2361 if (!get_zone_counts_fn
) {
2362 printk(KERN_ERR
"Error: Unknown symbol get_zone_counts\n");
2365 #endif /* NEED_GET_ZONE_COUNTS && !HAVE_GET_ZONE_COUNTS */
2368 * It is now safe to initialize the global tunings which rely on
2369 * the use of the for_each_zone() macro. This macro in turns
2370 * depends on the *_pgdat symbols which are now available.
2372 spl_kmem_init_globals();
2374 #ifndef HAVE_SHRINK_DCACHE_MEMORY
2375 /* When shrink_dcache_memory_fn == NULL support is disabled */
2376 shrink_dcache_memory_fn
= (shrink_dcache_memory_t
)
2377 spl_kallsyms_lookup_name("shrink_dcache_memory");
2378 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
2380 #ifndef HAVE_SHRINK_ICACHE_MEMORY
2381 /* When shrink_icache_memory_fn == NULL support is disabled */
2382 shrink_icache_memory_fn
= (shrink_icache_memory_t
)
2383 spl_kallsyms_lookup_name("shrink_icache_memory");
2384 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
2395 init_rwsem(&spl_kmem_cache_sem
);
2396 INIT_LIST_HEAD(&spl_kmem_cache_list
);
2397 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
2398 1, maxclsyspri
, 1, 32, TASKQ_PREPOPULATE
);
2400 spl_register_shrinker(&spl_kmem_cache_shrinker
);
2403 kmem_alloc_used_set(0);
2404 vmem_alloc_used_set(0);
2406 spl_kmem_init_tracking(&kmem_list
, &kmem_lock
, KMEM_TABLE_SIZE
);
2407 spl_kmem_init_tracking(&vmem_list
, &vmem_lock
, VMEM_TABLE_SIZE
);
2416 /* Display all unreclaimed memory addresses, including the
2417 * allocation size and the first few bytes of what's located
2418 * at that address to aid in debugging. Performance is not
2419 * a serious concern here since it is module unload time. */
2420 if (kmem_alloc_used_read() != 0)
2421 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2422 "kmem leaked %ld/%ld bytes\n",
2423 kmem_alloc_used_read(), kmem_alloc_max
);
2426 if (vmem_alloc_used_read() != 0)
2427 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2428 "vmem leaked %ld/%ld bytes\n",
2429 vmem_alloc_used_read(), vmem_alloc_max
);
2431 spl_kmem_fini_tracking(&kmem_list
, &kmem_lock
);
2432 spl_kmem_fini_tracking(&vmem_list
, &vmem_lock
);
2433 #endif /* DEBUG_KMEM */
2436 spl_unregister_shrinker(&spl_kmem_cache_shrinker
);
2437 taskq_destroy(spl_kmem_cache_taskq
);