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 #if !defined(HAVE_INVALIDATE_INODES) && !defined(HAVE_INVALIDATE_INODES_CHECK)
184 invalidate_inodes_t invalidate_inodes_fn
= SYMBOL_POISON
;
185 EXPORT_SYMBOL(invalidate_inodes_fn
);
186 #endif /* !HAVE_INVALIDATE_INODES && !HAVE_INVALIDATE_INODES_CHECK */
188 #ifndef HAVE_SHRINK_DCACHE_MEMORY
189 shrink_dcache_memory_t shrink_dcache_memory_fn
= SYMBOL_POISON
;
190 EXPORT_SYMBOL(shrink_dcache_memory_fn
);
191 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
193 #ifndef HAVE_SHRINK_ICACHE_MEMORY
194 shrink_icache_memory_t shrink_icache_memory_fn
= SYMBOL_POISON
;
195 EXPORT_SYMBOL(shrink_icache_memory_fn
);
196 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
199 spl_kmem_availrmem(void)
201 /* The amount of easily available memory */
202 return (spl_global_page_state(SPL_NR_FREE_PAGES
) +
203 spl_global_page_state(SPL_NR_INACTIVE
));
205 EXPORT_SYMBOL(spl_kmem_availrmem
);
208 vmem_size(vmem_t
*vmp
, int typemask
)
210 struct vmalloc_info vmi
;
214 ASSERT(typemask
& (VMEM_ALLOC
| VMEM_FREE
));
216 get_vmalloc_info(&vmi
);
217 if (typemask
& VMEM_ALLOC
)
218 size
+= (size_t)vmi
.used
;
220 if (typemask
& VMEM_FREE
)
221 size
+= (size_t)(VMALLOC_TOTAL
- vmi
.used
);
225 EXPORT_SYMBOL(vmem_size
);
232 EXPORT_SYMBOL(kmem_debugging
);
234 #ifndef HAVE_KVASPRINTF
235 /* Simplified asprintf. */
236 char *kvasprintf(gfp_t gfp
, const char *fmt
, va_list ap
)
243 len
= vsnprintf(NULL
, 0, fmt
, aq
);
246 p
= kmalloc(len
+1, gfp
);
250 vsnprintf(p
, len
+1, fmt
, ap
);
254 EXPORT_SYMBOL(kvasprintf
);
255 #endif /* HAVE_KVASPRINTF */
258 kmem_vasprintf(const char *fmt
, va_list ap
)
265 ptr
= kvasprintf(GFP_KERNEL
, fmt
, aq
);
267 } while (ptr
== NULL
);
271 EXPORT_SYMBOL(kmem_vasprintf
);
274 kmem_asprintf(const char *fmt
, ...)
281 ptr
= kvasprintf(GFP_KERNEL
, fmt
, ap
);
283 } while (ptr
== NULL
);
287 EXPORT_SYMBOL(kmem_asprintf
);
290 __strdup(const char *str
, int flags
)
296 ptr
= kmalloc_nofail(n
+ 1, flags
);
298 memcpy(ptr
, str
, n
+ 1);
304 strdup(const char *str
)
306 return __strdup(str
, KM_SLEEP
);
308 EXPORT_SYMBOL(strdup
);
315 EXPORT_SYMBOL(strfree
);
318 * Memory allocation interfaces and debugging for basic kmem_*
319 * and vmem_* style memory allocation. When DEBUG_KMEM is enabled
320 * the SPL will keep track of the total memory allocated, and
321 * report any memory leaked when the module is unloaded.
325 /* Shim layer memory accounting */
326 # ifdef HAVE_ATOMIC64_T
327 atomic64_t kmem_alloc_used
= ATOMIC64_INIT(0);
328 unsigned long long kmem_alloc_max
= 0;
329 atomic64_t vmem_alloc_used
= ATOMIC64_INIT(0);
330 unsigned long long vmem_alloc_max
= 0;
331 # else /* HAVE_ATOMIC64_T */
332 atomic_t kmem_alloc_used
= ATOMIC_INIT(0);
333 unsigned long long kmem_alloc_max
= 0;
334 atomic_t vmem_alloc_used
= ATOMIC_INIT(0);
335 unsigned long long vmem_alloc_max
= 0;
336 # endif /* HAVE_ATOMIC64_T */
338 EXPORT_SYMBOL(kmem_alloc_used
);
339 EXPORT_SYMBOL(kmem_alloc_max
);
340 EXPORT_SYMBOL(vmem_alloc_used
);
341 EXPORT_SYMBOL(vmem_alloc_max
);
343 /* When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked
344 * but also the location of every alloc and free. When the SPL module is
345 * unloaded a list of all leaked addresses and where they were allocated
346 * will be dumped to the console. Enabling this feature has a significant
347 * impact on performance but it makes finding memory leaks straight forward.
349 * Not surprisingly with debugging enabled the xmem_locks are very highly
350 * contended particularly on xfree(). If we want to run with this detailed
351 * debugging enabled for anything other than debugging we need to minimize
352 * the contention by moving to a lock per xmem_table entry model.
354 # ifdef DEBUG_KMEM_TRACKING
356 # define KMEM_HASH_BITS 10
357 # define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS)
359 # define VMEM_HASH_BITS 10
360 # define VMEM_TABLE_SIZE (1 << VMEM_HASH_BITS)
362 typedef struct kmem_debug
{
363 struct hlist_node kd_hlist
; /* Hash node linkage */
364 struct list_head kd_list
; /* List of all allocations */
365 void *kd_addr
; /* Allocation pointer */
366 size_t kd_size
; /* Allocation size */
367 const char *kd_func
; /* Allocation function */
368 int kd_line
; /* Allocation line */
371 spinlock_t kmem_lock
;
372 struct hlist_head kmem_table
[KMEM_TABLE_SIZE
];
373 struct list_head kmem_list
;
375 spinlock_t vmem_lock
;
376 struct hlist_head vmem_table
[VMEM_TABLE_SIZE
];
377 struct list_head vmem_list
;
379 EXPORT_SYMBOL(kmem_lock
);
380 EXPORT_SYMBOL(kmem_table
);
381 EXPORT_SYMBOL(kmem_list
);
383 EXPORT_SYMBOL(vmem_lock
);
384 EXPORT_SYMBOL(vmem_table
);
385 EXPORT_SYMBOL(vmem_list
);
387 static kmem_debug_t
*
388 kmem_del_init(spinlock_t
*lock
, struct hlist_head
*table
, int bits
, const void *addr
)
390 struct hlist_head
*head
;
391 struct hlist_node
*node
;
392 struct kmem_debug
*p
;
396 spin_lock_irqsave(lock
, flags
);
398 head
= &table
[hash_ptr(addr
, bits
)];
399 hlist_for_each_entry_rcu(p
, node
, head
, kd_hlist
) {
400 if (p
->kd_addr
== addr
) {
401 hlist_del_init(&p
->kd_hlist
);
402 list_del_init(&p
->kd_list
);
403 spin_unlock_irqrestore(lock
, flags
);
408 spin_unlock_irqrestore(lock
, flags
);
414 kmem_alloc_track(size_t size
, int flags
, const char *func
, int line
,
415 int node_alloc
, int node
)
419 unsigned long irq_flags
;
422 /* Function may be called with KM_NOSLEEP so failure is possible */
423 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
424 flags
& ~__GFP_ZERO
);
426 if (unlikely(dptr
== NULL
)) {
427 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
428 "kmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
429 sizeof(kmem_debug_t
), flags
, func
, line
,
430 kmem_alloc_used_read(), kmem_alloc_max
);
433 * Marked unlikely because we should never be doing this,
434 * we tolerate to up 2 pages but a single page is best.
436 if (unlikely((size
> PAGE_SIZE
*2) && !(flags
& KM_NODEBUG
))) {
437 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "large "
438 "kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
439 (unsigned long long) size
, flags
, func
, line
,
440 kmem_alloc_used_read(), kmem_alloc_max
);
441 spl_debug_dumpstack(NULL
);
445 * We use __strdup() below because the string pointed to by
446 * __FUNCTION__ might not be available by the time we want
447 * to print it since the module might have been unloaded.
448 * This can only fail in the KM_NOSLEEP case.
450 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
451 if (unlikely(dptr
->kd_func
== NULL
)) {
453 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
454 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
455 func
, line
, kmem_alloc_used_read(), kmem_alloc_max
);
459 /* Use the correct allocator */
461 ASSERT(!(flags
& __GFP_ZERO
));
462 ptr
= kmalloc_node_nofail(size
, flags
, node
);
463 } else if (flags
& __GFP_ZERO
) {
464 ptr
= kzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
466 ptr
= kmalloc_nofail(size
, flags
);
469 if (unlikely(ptr
== NULL
)) {
470 kfree(dptr
->kd_func
);
472 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "kmem_alloc"
473 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
474 (unsigned long long) size
, flags
, func
, line
,
475 kmem_alloc_used_read(), kmem_alloc_max
);
479 kmem_alloc_used_add(size
);
480 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
481 kmem_alloc_max
= kmem_alloc_used_read();
483 INIT_HLIST_NODE(&dptr
->kd_hlist
);
484 INIT_LIST_HEAD(&dptr
->kd_list
);
487 dptr
->kd_size
= size
;
488 dptr
->kd_line
= line
;
490 spin_lock_irqsave(&kmem_lock
, irq_flags
);
491 hlist_add_head_rcu(&dptr
->kd_hlist
,
492 &kmem_table
[hash_ptr(ptr
, KMEM_HASH_BITS
)]);
493 list_add_tail(&dptr
->kd_list
, &kmem_list
);
494 spin_unlock_irqrestore(&kmem_lock
, irq_flags
);
496 SDEBUG_LIMIT(SD_INFO
,
497 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
498 (unsigned long long) size
, flags
, func
, line
, ptr
,
499 kmem_alloc_used_read(), kmem_alloc_max
);
504 EXPORT_SYMBOL(kmem_alloc_track
);
507 kmem_free_track(const void *ptr
, size_t size
)
512 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
513 (unsigned long long) size
);
515 dptr
= kmem_del_init(&kmem_lock
, kmem_table
, KMEM_HASH_BITS
, ptr
);
517 /* Must exist in hash due to kmem_alloc() */
520 /* Size must match */
521 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
522 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
523 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
525 kmem_alloc_used_sub(size
);
526 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
527 (unsigned long long) size
, kmem_alloc_used_read(),
530 kfree(dptr
->kd_func
);
532 memset(dptr
, 0x5a, sizeof(kmem_debug_t
));
535 memset(ptr
, 0x5a, size
);
540 EXPORT_SYMBOL(kmem_free_track
);
543 vmem_alloc_track(size_t size
, int flags
, const char *func
, int line
)
547 unsigned long irq_flags
;
550 ASSERT(flags
& KM_SLEEP
);
552 /* Function may be called with KM_NOSLEEP so failure is possible */
553 dptr
= (kmem_debug_t
*) kmalloc_nofail(sizeof(kmem_debug_t
),
554 flags
& ~__GFP_ZERO
);
555 if (unlikely(dptr
== NULL
)) {
556 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "debug "
557 "vmem_alloc(%ld, 0x%x) at %s:%d failed (%lld/%llu)\n",
558 sizeof(kmem_debug_t
), flags
, func
, line
,
559 vmem_alloc_used_read(), vmem_alloc_max
);
562 * We use __strdup() below because the string pointed to by
563 * __FUNCTION__ might not be available by the time we want
564 * to print it, since the module might have been unloaded.
565 * This can never fail because we have already asserted
566 * that flags is KM_SLEEP.
568 dptr
->kd_func
= __strdup(func
, flags
& ~__GFP_ZERO
);
569 if (unlikely(dptr
->kd_func
== NULL
)) {
571 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
572 "debug __strdup() at %s:%d failed (%lld/%llu)\n",
573 func
, line
, vmem_alloc_used_read(), vmem_alloc_max
);
577 /* Use the correct allocator */
578 if (flags
& __GFP_ZERO
) {
579 ptr
= vzalloc_nofail(size
, flags
& ~__GFP_ZERO
);
581 ptr
= vmalloc_nofail(size
, flags
);
584 if (unlikely(ptr
== NULL
)) {
585 kfree(dptr
->kd_func
);
587 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
, "vmem_alloc"
588 "(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
589 (unsigned long long) size
, flags
, func
, line
,
590 vmem_alloc_used_read(), vmem_alloc_max
);
594 vmem_alloc_used_add(size
);
595 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
596 vmem_alloc_max
= vmem_alloc_used_read();
598 INIT_HLIST_NODE(&dptr
->kd_hlist
);
599 INIT_LIST_HEAD(&dptr
->kd_list
);
602 dptr
->kd_size
= size
;
603 dptr
->kd_line
= line
;
605 spin_lock_irqsave(&vmem_lock
, irq_flags
);
606 hlist_add_head_rcu(&dptr
->kd_hlist
,
607 &vmem_table
[hash_ptr(ptr
, VMEM_HASH_BITS
)]);
608 list_add_tail(&dptr
->kd_list
, &vmem_list
);
609 spin_unlock_irqrestore(&vmem_lock
, irq_flags
);
611 SDEBUG_LIMIT(SD_INFO
,
612 "vmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
613 (unsigned long long) size
, flags
, func
, line
,
614 ptr
, vmem_alloc_used_read(), vmem_alloc_max
);
619 EXPORT_SYMBOL(vmem_alloc_track
);
622 vmem_free_track(const void *ptr
, size_t size
)
627 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
628 (unsigned long long) size
);
630 dptr
= kmem_del_init(&vmem_lock
, vmem_table
, VMEM_HASH_BITS
, ptr
);
632 /* Must exist in hash due to vmem_alloc() */
635 /* Size must match */
636 ASSERTF(dptr
->kd_size
== size
, "kd_size (%llu) != size (%llu), "
637 "kd_func = %s, kd_line = %d\n", (unsigned long long) dptr
->kd_size
,
638 (unsigned long long) size
, dptr
->kd_func
, dptr
->kd_line
);
640 vmem_alloc_used_sub(size
);
641 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
642 (unsigned long long) size
, vmem_alloc_used_read(),
645 kfree(dptr
->kd_func
);
647 memset(dptr
, 0x5a, sizeof(kmem_debug_t
));
650 memset(ptr
, 0x5a, size
);
655 EXPORT_SYMBOL(vmem_free_track
);
657 # else /* DEBUG_KMEM_TRACKING */
660 kmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
,
661 int node_alloc
, int node
)
667 * Marked unlikely because we should never be doing this,
668 * we tolerate to up 2 pages but a single page is best.
670 if (unlikely((size
> PAGE_SIZE
* 2) && !(flags
& KM_NODEBUG
))) {
671 SDEBUG(SD_CONSOLE
| SD_WARNING
,
672 "large kmem_alloc(%llu, 0x%x) at %s:%d (%lld/%llu)\n",
673 (unsigned long long) size
, flags
, func
, line
,
674 kmem_alloc_used_read(), kmem_alloc_max
);
678 /* Use the correct allocator */
680 ASSERT(!(flags
& __GFP_ZERO
));
681 ptr
= kmalloc_node_nofail(size
, flags
, node
);
682 } else if (flags
& __GFP_ZERO
) {
683 ptr
= kzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
685 ptr
= kmalloc_nofail(size
, flags
);
688 if (unlikely(ptr
== NULL
)) {
689 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
690 "kmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
691 (unsigned long long) size
, flags
, func
, line
,
692 kmem_alloc_used_read(), kmem_alloc_max
);
694 kmem_alloc_used_add(size
);
695 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
696 kmem_alloc_max
= kmem_alloc_used_read();
698 SDEBUG_LIMIT(SD_INFO
,
699 "kmem_alloc(%llu, 0x%x) at %s:%d = %p (%lld/%llu)\n",
700 (unsigned long long) size
, flags
, func
, line
, ptr
,
701 kmem_alloc_used_read(), kmem_alloc_max
);
706 EXPORT_SYMBOL(kmem_alloc_debug
);
709 kmem_free_debug(const void *ptr
, size_t size
)
713 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
714 (unsigned long long) size
);
716 kmem_alloc_used_sub(size
);
717 SDEBUG_LIMIT(SD_INFO
, "kmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
718 (unsigned long long) size
, kmem_alloc_used_read(),
724 EXPORT_SYMBOL(kmem_free_debug
);
727 vmem_alloc_debug(size_t size
, int flags
, const char *func
, int line
)
732 ASSERT(flags
& KM_SLEEP
);
734 /* Use the correct allocator */
735 if (flags
& __GFP_ZERO
) {
736 ptr
= vzalloc_nofail(size
, flags
& (~__GFP_ZERO
));
738 ptr
= vmalloc_nofail(size
, flags
);
741 if (unlikely(ptr
== NULL
)) {
742 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
743 "vmem_alloc(%llu, 0x%x) at %s:%d failed (%lld/%llu)\n",
744 (unsigned long long) size
, flags
, func
, line
,
745 vmem_alloc_used_read(), vmem_alloc_max
);
747 vmem_alloc_used_add(size
);
748 if (unlikely(vmem_alloc_used_read() > vmem_alloc_max
))
749 vmem_alloc_max
= vmem_alloc_used_read();
751 SDEBUG_LIMIT(SD_INFO
, "vmem_alloc(%llu, 0x%x) = %p "
752 "(%lld/%llu)\n", (unsigned long long) size
, flags
, ptr
,
753 vmem_alloc_used_read(), vmem_alloc_max
);
758 EXPORT_SYMBOL(vmem_alloc_debug
);
761 vmem_free_debug(const void *ptr
, size_t size
)
765 ASSERTF(ptr
|| size
> 0, "ptr: %p, size: %llu", ptr
,
766 (unsigned long long) size
);
768 vmem_alloc_used_sub(size
);
769 SDEBUG_LIMIT(SD_INFO
, "vmem_free(%p, %llu) (%lld/%llu)\n", ptr
,
770 (unsigned long long) size
, vmem_alloc_used_read(),
776 EXPORT_SYMBOL(vmem_free_debug
);
778 # endif /* DEBUG_KMEM_TRACKING */
779 #endif /* DEBUG_KMEM */
782 * Slab allocation interfaces
784 * While the Linux slab implementation was inspired by the Solaris
785 * implementation I cannot use it to emulate the Solaris APIs. I
786 * require two features which are not provided by the Linux slab.
788 * 1) Constructors AND destructors. Recent versions of the Linux
789 * kernel have removed support for destructors. This is a deal
790 * breaker for the SPL which contains particularly expensive
791 * initializers for mutex's, condition variables, etc. We also
792 * require a minimal level of cleanup for these data types unlike
793 * many Linux data type which do need to be explicitly destroyed.
795 * 2) Virtual address space backed slab. Callers of the Solaris slab
796 * expect it to work well for both small are very large allocations.
797 * Because of memory fragmentation the Linux slab which is backed
798 * by kmalloc'ed memory performs very badly when confronted with
799 * large numbers of large allocations. Basing the slab on the
800 * virtual address space removes the need for contiguous pages
801 * and greatly improve performance for large allocations.
803 * For these reasons, the SPL has its own slab implementation with
804 * the needed features. It is not as highly optimized as either the
805 * Solaris or Linux slabs, but it should get me most of what is
806 * needed until it can be optimized or obsoleted by another approach.
808 * One serious concern I do have about this method is the relatively
809 * small virtual address space on 32bit arches. This will seriously
810 * constrain the size of the slab caches and their performance.
812 * XXX: Improve the partial slab list by carefully maintaining a
813 * strict ordering of fullest to emptiest slabs based on
814 * the slab reference count. This guarantees the when freeing
815 * slabs back to the system we need only linearly traverse the
816 * last N slabs in the list to discover all the freeable slabs.
818 * XXX: NUMA awareness for optionally allocating memory close to a
819 * particular core. This can be advantageous if you know the slab
820 * object will be short lived and primarily accessed from one core.
822 * XXX: Slab coloring may also yield performance improvements and would
823 * be desirable to implement.
826 struct list_head spl_kmem_cache_list
; /* List of caches */
827 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
828 taskq_t
*spl_kmem_cache_taskq
; /* Task queue for ageing / reclaim */
830 static int spl_cache_flush(spl_kmem_cache_t
*skc
,
831 spl_kmem_magazine_t
*skm
, int flush
);
833 SPL_SHRINKER_CALLBACK_FWD_DECLARE(spl_kmem_cache_generic_shrinker
);
834 SPL_SHRINKER_DECLARE(spl_kmem_cache_shrinker
,
835 spl_kmem_cache_generic_shrinker
, KMC_DEFAULT_SEEKS
);
838 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
844 if (skc
->skc_flags
& KMC_KMEM
)
845 ptr
= (void *)__get_free_pages(flags
, get_order(size
));
847 ptr
= __vmalloc(size
, flags
| __GFP_HIGHMEM
, PAGE_KERNEL
);
849 /* Resulting allocated memory will be page aligned */
850 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
856 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
858 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
862 * The Linux direct reclaim path uses this out of band value to
863 * determine if forward progress is being made. Normally this is
864 * incremented by kmem_freepages() which is part of the various
865 * Linux slab implementations. However, since we are using none
866 * of that infrastructure we are responsible for incrementing it.
868 if (current
->reclaim_state
)
869 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
871 if (skc
->skc_flags
& KMC_KMEM
)
872 free_pages((unsigned long)ptr
, get_order(size
));
878 * Required space for each aligned sks.
880 static inline uint32_t
881 spl_sks_size(spl_kmem_cache_t
*skc
)
883 return P2ROUNDUP_TYPED(sizeof(spl_kmem_slab_t
),
884 skc
->skc_obj_align
, uint32_t);
888 * Required space for each aligned object.
890 static inline uint32_t
891 spl_obj_size(spl_kmem_cache_t
*skc
)
893 uint32_t align
= skc
->skc_obj_align
;
895 return P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
896 P2ROUNDUP_TYPED(sizeof(spl_kmem_obj_t
), align
, uint32_t);
900 * Lookup the spl_kmem_object_t for an object given that object.
902 static inline spl_kmem_obj_t
*
903 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
905 return obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
906 skc
->skc_obj_align
, uint32_t);
910 * Required space for each offslab object taking in to account alignment
911 * restrictions and the power-of-two requirement of kv_alloc().
913 static inline uint32_t
914 spl_offslab_size(spl_kmem_cache_t
*skc
)
916 return 1UL << (highbit(spl_obj_size(skc
)) + 1);
920 * It's important that we pack the spl_kmem_obj_t structure and the
921 * actual objects in to one large address space to minimize the number
922 * of calls to the allocator. It is far better to do a few large
923 * allocations and then subdivide it ourselves. Now which allocator
924 * we use requires balancing a few trade offs.
926 * For small objects we use kmem_alloc() because as long as you are
927 * only requesting a small number of pages (ideally just one) its cheap.
928 * However, when you start requesting multiple pages with kmem_alloc()
929 * it gets increasingly expensive since it requires contiguous pages.
930 * For this reason we shift to vmem_alloc() for slabs of large objects
931 * which removes the need for contiguous pages. We do not use
932 * vmem_alloc() in all cases because there is significant locking
933 * overhead in __get_vm_area_node(). This function takes a single
934 * global lock when acquiring an available virtual address range which
935 * serializes all vmem_alloc()'s for all slab caches. Using slightly
936 * different allocation functions for small and large objects should
937 * give us the best of both worlds.
939 * KMC_ONSLAB KMC_OFFSLAB
941 * +------------------------+ +-----------------+
942 * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+
943 * | skc_obj_size <-+ | | +-----------------+ | |
944 * | spl_kmem_obj_t | | | |
945 * | skc_obj_size <---+ | +-----------------+ | |
946 * | spl_kmem_obj_t | | | skc_obj_size | <-+ |
947 * | ... v | | spl_kmem_obj_t | |
948 * +------------------------+ +-----------------+ v
950 static spl_kmem_slab_t
*
951 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
953 spl_kmem_slab_t
*sks
;
954 spl_kmem_obj_t
*sko
, *n
;
956 uint32_t obj_size
, offslab_size
= 0;
959 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
963 sks
= (spl_kmem_slab_t
*)base
;
964 sks
->sks_magic
= SKS_MAGIC
;
965 sks
->sks_objs
= skc
->skc_slab_objs
;
966 sks
->sks_age
= jiffies
;
967 sks
->sks_cache
= skc
;
968 INIT_LIST_HEAD(&sks
->sks_list
);
969 INIT_LIST_HEAD(&sks
->sks_free_list
);
971 obj_size
= spl_obj_size(skc
);
973 if (skc
->skc_flags
& KMC_OFFSLAB
)
974 offslab_size
= spl_offslab_size(skc
);
976 for (i
= 0; i
< sks
->sks_objs
; i
++) {
977 if (skc
->skc_flags
& KMC_OFFSLAB
) {
978 obj
= kv_alloc(skc
, offslab_size
, flags
);
980 SGOTO(out
, rc
= -ENOMEM
);
982 obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
985 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
986 sko
= spl_sko_from_obj(skc
, obj
);
988 sko
->sko_magic
= SKO_MAGIC
;
990 INIT_LIST_HEAD(&sko
->sko_list
);
991 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
994 list_for_each_entry(sko
, &sks
->sks_free_list
, sko_list
)
996 skc
->skc_ctor(sko
->sko_addr
, skc
->skc_private
, flags
);
999 if (skc
->skc_flags
& KMC_OFFSLAB
)
1000 list_for_each_entry_safe(sko
, n
, &sks
->sks_free_list
,
1002 kv_free(skc
, sko
->sko_addr
, offslab_size
);
1004 kv_free(skc
, base
, skc
->skc_slab_size
);
1012 * Remove a slab from complete or partial list, it must be called with
1013 * the 'skc->skc_lock' held but the actual free must be performed
1014 * outside the lock to prevent deadlocking on vmem addresses.
1017 spl_slab_free(spl_kmem_slab_t
*sks
,
1018 struct list_head
*sks_list
, struct list_head
*sko_list
)
1020 spl_kmem_cache_t
*skc
;
1023 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1024 ASSERT(sks
->sks_ref
== 0);
1026 skc
= sks
->sks_cache
;
1027 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1028 ASSERT(spin_is_locked(&skc
->skc_lock
));
1031 * Update slab/objects counters in the cache, then remove the
1032 * slab from the skc->skc_partial_list. Finally add the slab
1033 * and all its objects in to the private work lists where the
1034 * destructors will be called and the memory freed to the system.
1036 skc
->skc_obj_total
-= sks
->sks_objs
;
1037 skc
->skc_slab_total
--;
1038 list_del(&sks
->sks_list
);
1039 list_add(&sks
->sks_list
, sks_list
);
1040 list_splice_init(&sks
->sks_free_list
, sko_list
);
1046 * Traverses all the partial slabs attached to a cache and free those
1047 * which which are currently empty, and have not been touched for
1048 * skc_delay seconds to avoid thrashing. The count argument is
1049 * passed to optionally cap the number of slabs reclaimed, a count
1050 * of zero means try and reclaim everything. When flag is set we
1051 * always free an available slab regardless of age.
1054 spl_slab_reclaim(spl_kmem_cache_t
*skc
, int count
, int flag
)
1056 spl_kmem_slab_t
*sks
, *m
;
1057 spl_kmem_obj_t
*sko
, *n
;
1058 LIST_HEAD(sks_list
);
1059 LIST_HEAD(sko_list
);
1065 * Move empty slabs and objects which have not been touched in
1066 * skc_delay seconds on to private lists to be freed outside
1067 * the spin lock. This delay time is important to avoid thrashing
1068 * however when flag is set the delay will not be used.
1070 spin_lock(&skc
->skc_lock
);
1071 list_for_each_entry_safe_reverse(sks
,m
,&skc
->skc_partial_list
,sks_list
){
1073 * All empty slabs are at the end of skc->skc_partial_list,
1074 * therefore once a non-empty slab is found we can stop
1075 * scanning. Additionally, stop when reaching the target
1076 * reclaim 'count' if a non-zero threshold is given.
1078 if ((sks
->sks_ref
> 0) || (count
&& i
>= count
))
1081 if (time_after(jiffies
,sks
->sks_age
+skc
->skc_delay
*HZ
)||flag
) {
1082 spl_slab_free(sks
, &sks_list
, &sko_list
);
1086 spin_unlock(&skc
->skc_lock
);
1089 * The following two loops ensure all the object destructors are
1090 * run, any offslab objects are freed, and the slabs themselves
1091 * are freed. This is all done outside the skc->skc_lock since
1092 * this allows the destructor to sleep, and allows us to perform
1093 * a conditional reschedule when a freeing a large number of
1094 * objects and slabs back to the system.
1096 if (skc
->skc_flags
& KMC_OFFSLAB
)
1097 size
= spl_offslab_size(skc
);
1099 list_for_each_entry_safe(sko
, n
, &sko_list
, sko_list
) {
1100 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1103 skc
->skc_dtor(sko
->sko_addr
, skc
->skc_private
);
1105 if (skc
->skc_flags
& KMC_OFFSLAB
)
1106 kv_free(skc
, sko
->sko_addr
, size
);
1111 list_for_each_entry_safe(sks
, m
, &sks_list
, sks_list
) {
1112 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1113 kv_free(skc
, sks
, skc
->skc_slab_size
);
1120 static spl_kmem_emergency_t
*
1121 spl_emergency_search(struct rb_root
*root
, void *obj
)
1123 struct rb_node
*node
= root
->rb_node
;
1124 spl_kmem_emergency_t
*ske
;
1125 unsigned long address
= (unsigned long)obj
;
1128 ske
= container_of(node
, spl_kmem_emergency_t
, ske_node
);
1130 if (address
< (unsigned long)ske
->ske_obj
)
1131 node
= node
->rb_left
;
1132 else if (address
> (unsigned long)ske
->ske_obj
)
1133 node
= node
->rb_right
;
1142 spl_emergency_insert(struct rb_root
*root
, spl_kmem_emergency_t
*ske
)
1144 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
1145 spl_kmem_emergency_t
*ske_tmp
;
1146 unsigned long address
= (unsigned long)ske
->ske_obj
;
1149 ske_tmp
= container_of(*new, spl_kmem_emergency_t
, ske_node
);
1152 if (address
< (unsigned long)ske_tmp
->ske_obj
)
1153 new = &((*new)->rb_left
);
1154 else if (address
> (unsigned long)ske_tmp
->ske_obj
)
1155 new = &((*new)->rb_right
);
1160 rb_link_node(&ske
->ske_node
, parent
, new);
1161 rb_insert_color(&ske
->ske_node
, root
);
1167 * Allocate a single emergency object and track it in a red black tree.
1170 spl_emergency_alloc(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1172 spl_kmem_emergency_t
*ske
;
1176 /* Last chance use a partial slab if one now exists */
1177 spin_lock(&skc
->skc_lock
);
1178 empty
= list_empty(&skc
->skc_partial_list
);
1179 spin_unlock(&skc
->skc_lock
);
1183 ske
= kmalloc(sizeof(*ske
), flags
);
1187 ske
->ske_obj
= kmalloc(skc
->skc_obj_size
, flags
);
1188 if (ske
->ske_obj
== NULL
) {
1193 spin_lock(&skc
->skc_lock
);
1194 empty
= spl_emergency_insert(&skc
->skc_emergency_tree
, ske
);
1195 if (likely(empty
)) {
1196 skc
->skc_obj_total
++;
1197 skc
->skc_obj_emergency
++;
1198 if (skc
->skc_obj_emergency
> skc
->skc_obj_emergency_max
)
1199 skc
->skc_obj_emergency_max
= skc
->skc_obj_emergency
;
1201 spin_unlock(&skc
->skc_lock
);
1203 if (unlikely(!empty
)) {
1204 kfree(ske
->ske_obj
);
1210 skc
->skc_ctor(ske
->ske_obj
, skc
->skc_private
, flags
);
1212 *obj
= ske
->ske_obj
;
1218 * Locate the passed object in the red black tree and free it.
1221 spl_emergency_free(spl_kmem_cache_t
*skc
, void *obj
)
1223 spl_kmem_emergency_t
*ske
;
1226 spin_lock(&skc
->skc_lock
);
1227 ske
= spl_emergency_search(&skc
->skc_emergency_tree
, obj
);
1229 rb_erase(&ske
->ske_node
, &skc
->skc_emergency_tree
);
1230 skc
->skc_obj_emergency
--;
1231 skc
->skc_obj_total
--;
1233 spin_unlock(&skc
->skc_lock
);
1235 if (unlikely(ske
== NULL
))
1239 skc
->skc_dtor(ske
->ske_obj
, skc
->skc_private
);
1241 kfree(ske
->ske_obj
);
1248 spl_magazine_age(void *data
)
1250 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1251 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
1253 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1254 ASSERT(skm
->skm_cpu
== smp_processor_id());
1256 if (skm
->skm_avail
> 0)
1257 if (time_after(jiffies
, skm
->skm_age
+ skc
->skc_delay
* HZ
))
1258 (void) spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1262 * Called regularly to keep a downward pressure on the cache.
1264 * Objects older than skc->skc_delay seconds in the per-cpu magazines will
1265 * be returned to the caches. This is done to prevent idle magazines from
1266 * holding memory which could be better used elsewhere. The delay is
1267 * present to prevent thrashing the magazine.
1269 * The newly released objects may result in empty partial slabs. Those
1270 * slabs should be released to the system. Otherwise moving the objects
1271 * out of the magazines is just wasted work.
1274 spl_cache_age(void *data
)
1276 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
1279 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1281 atomic_inc(&skc
->skc_ref
);
1282 spl_on_each_cpu(spl_magazine_age
, skc
, 1);
1283 spl_slab_reclaim(skc
, skc
->skc_reap
, 0);
1285 while (!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && !id
) {
1286 id
= taskq_dispatch_delay(
1287 spl_kmem_cache_taskq
, spl_cache_age
, skc
, TQ_SLEEP
,
1288 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1290 /* Destroy issued after dispatch immediately cancel it */
1291 if (test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && id
)
1292 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1295 spin_lock(&skc
->skc_lock
);
1296 skc
->skc_taskqid
= id
;
1297 spin_unlock(&skc
->skc_lock
);
1299 atomic_dec(&skc
->skc_ref
);
1303 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
1304 * When on-slab we want to target SPL_KMEM_CACHE_OBJ_PER_SLAB. However,
1305 * for very small objects we may end up with more than this so as not
1306 * to waste space in the minimal allocation of a single page. Also for
1307 * very large objects we may use as few as SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN,
1308 * lower than this and we will fail.
1311 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
1313 uint32_t sks_size
, obj_size
, max_size
;
1315 if (skc
->skc_flags
& KMC_OFFSLAB
) {
1316 *objs
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
1317 *size
= sizeof(spl_kmem_slab_t
);
1319 sks_size
= spl_sks_size(skc
);
1320 obj_size
= spl_obj_size(skc
);
1322 if (skc
->skc_flags
& KMC_KMEM
)
1323 max_size
= ((uint32_t)1 << (MAX_ORDER
-3)) * PAGE_SIZE
;
1325 max_size
= (32 * 1024 * 1024);
1327 /* Power of two sized slab */
1328 for (*size
= PAGE_SIZE
; *size
<= max_size
; *size
*= 2) {
1329 *objs
= (*size
- sks_size
) / obj_size
;
1330 if (*objs
>= SPL_KMEM_CACHE_OBJ_PER_SLAB
)
1335 * Unable to satisfy target objects per slab, fall back to
1336 * allocating a maximally sized slab and assuming it can
1337 * contain the minimum objects count use it. If not fail.
1340 *objs
= (*size
- sks_size
) / obj_size
;
1341 if (*objs
>= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN
)
1349 * Make a guess at reasonable per-cpu magazine size based on the size of
1350 * each object and the cost of caching N of them in each magazine. Long
1351 * term this should really adapt based on an observed usage heuristic.
1354 spl_magazine_size(spl_kmem_cache_t
*skc
)
1356 uint32_t obj_size
= spl_obj_size(skc
);
1360 /* Per-magazine sizes below assume a 4Kib page size */
1361 if (obj_size
> (PAGE_SIZE
* 256))
1362 size
= 4; /* Minimum 4Mib per-magazine */
1363 else if (obj_size
> (PAGE_SIZE
* 32))
1364 size
= 16; /* Minimum 2Mib per-magazine */
1365 else if (obj_size
> (PAGE_SIZE
))
1366 size
= 64; /* Minimum 256Kib per-magazine */
1367 else if (obj_size
> (PAGE_SIZE
/ 4))
1368 size
= 128; /* Minimum 128Kib per-magazine */
1376 * Allocate a per-cpu magazine to associate with a specific core.
1378 static spl_kmem_magazine_t
*
1379 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
1381 spl_kmem_magazine_t
*skm
;
1382 int size
= sizeof(spl_kmem_magazine_t
) +
1383 sizeof(void *) * skc
->skc_mag_size
;
1386 skm
= kmem_alloc_node(size
, KM_SLEEP
, cpu_to_node(cpu
));
1388 skm
->skm_magic
= SKM_MAGIC
;
1390 skm
->skm_size
= skc
->skc_mag_size
;
1391 skm
->skm_refill
= skc
->skc_mag_refill
;
1392 skm
->skm_cache
= skc
;
1393 skm
->skm_age
= jiffies
;
1401 * Free a per-cpu magazine associated with a specific core.
1404 spl_magazine_free(spl_kmem_magazine_t
*skm
)
1406 int size
= sizeof(spl_kmem_magazine_t
) +
1407 sizeof(void *) * skm
->skm_size
;
1410 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1411 ASSERT(skm
->skm_avail
== 0);
1413 kmem_free(skm
, size
);
1418 * Create all pre-cpu magazines of reasonable sizes.
1421 spl_magazine_create(spl_kmem_cache_t
*skc
)
1426 skc
->skc_mag_size
= spl_magazine_size(skc
);
1427 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
1429 for_each_online_cpu(i
) {
1430 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
1431 if (!skc
->skc_mag
[i
]) {
1432 for (i
--; i
>= 0; i
--)
1433 spl_magazine_free(skc
->skc_mag
[i
]);
1443 * Destroy all pre-cpu magazines.
1446 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
1448 spl_kmem_magazine_t
*skm
;
1452 for_each_online_cpu(i
) {
1453 skm
= skc
->skc_mag
[i
];
1454 (void)spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1455 spl_magazine_free(skm
);
1462 * Create a object cache based on the following arguments:
1464 * size cache object size
1465 * align cache object alignment
1466 * ctor cache object constructor
1467 * dtor cache object destructor
1468 * reclaim cache object reclaim
1469 * priv cache private data for ctor/dtor/reclaim
1470 * vmp unused must be NULL
1472 * KMC_NOTOUCH Disable cache object aging (unsupported)
1473 * KMC_NODEBUG Disable debugging (unsupported)
1474 * KMC_NOMAGAZINE Disable magazine (unsupported)
1475 * KMC_NOHASH Disable hashing (unsupported)
1476 * KMC_QCACHE Disable qcache (unsupported)
1477 * KMC_KMEM Force kmem backed cache
1478 * KMC_VMEM Force vmem backed cache
1479 * KMC_OFFSLAB Locate objects off the slab
1482 spl_kmem_cache_create(char *name
, size_t size
, size_t align
,
1483 spl_kmem_ctor_t ctor
,
1484 spl_kmem_dtor_t dtor
,
1485 spl_kmem_reclaim_t reclaim
,
1486 void *priv
, void *vmp
, int flags
)
1488 spl_kmem_cache_t
*skc
;
1492 ASSERTF(!(flags
& KMC_NOMAGAZINE
), "Bad KMC_NOMAGAZINE (%x)\n", flags
);
1493 ASSERTF(!(flags
& KMC_NOHASH
), "Bad KMC_NOHASH (%x)\n", flags
);
1494 ASSERTF(!(flags
& KMC_QCACHE
), "Bad KMC_QCACHE (%x)\n", flags
);
1495 ASSERT(vmp
== NULL
);
1500 * Allocate memory for a new cache an initialize it. Unfortunately,
1501 * this usually ends up being a large allocation of ~32k because
1502 * we need to allocate enough memory for the worst case number of
1503 * cpus in the magazine, skc_mag[NR_CPUS]. Because of this we
1504 * explicitly pass KM_NODEBUG to suppress the kmem warning
1506 skc
= kmem_zalloc(sizeof(*skc
), KM_SLEEP
| KM_NODEBUG
);
1510 skc
->skc_magic
= SKC_MAGIC
;
1511 skc
->skc_name_size
= strlen(name
) + 1;
1512 skc
->skc_name
= (char *)kmem_alloc(skc
->skc_name_size
, KM_SLEEP
);
1513 if (skc
->skc_name
== NULL
) {
1514 kmem_free(skc
, sizeof(*skc
));
1517 strncpy(skc
->skc_name
, name
, skc
->skc_name_size
);
1519 skc
->skc_ctor
= ctor
;
1520 skc
->skc_dtor
= dtor
;
1521 skc
->skc_reclaim
= reclaim
;
1522 skc
->skc_private
= priv
;
1524 skc
->skc_flags
= flags
;
1525 skc
->skc_obj_size
= size
;
1526 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
1527 skc
->skc_delay
= SPL_KMEM_CACHE_DELAY
;
1528 skc
->skc_reap
= SPL_KMEM_CACHE_REAP
;
1529 atomic_set(&skc
->skc_ref
, 0);
1531 INIT_LIST_HEAD(&skc
->skc_list
);
1532 INIT_LIST_HEAD(&skc
->skc_complete_list
);
1533 INIT_LIST_HEAD(&skc
->skc_partial_list
);
1534 skc
->skc_emergency_tree
= RB_ROOT
;
1535 spin_lock_init(&skc
->skc_lock
);
1536 init_waitqueue_head(&skc
->skc_waitq
);
1537 skc
->skc_slab_fail
= 0;
1538 skc
->skc_slab_create
= 0;
1539 skc
->skc_slab_destroy
= 0;
1540 skc
->skc_slab_total
= 0;
1541 skc
->skc_slab_alloc
= 0;
1542 skc
->skc_slab_max
= 0;
1543 skc
->skc_obj_total
= 0;
1544 skc
->skc_obj_alloc
= 0;
1545 skc
->skc_obj_max
= 0;
1546 skc
->skc_obj_deadlock
= 0;
1547 skc
->skc_obj_emergency
= 0;
1548 skc
->skc_obj_emergency_max
= 0;
1551 VERIFY(ISP2(align
));
1552 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
); /* Min alignment */
1553 VERIFY3U(align
, <=, PAGE_SIZE
); /* Max alignment */
1554 skc
->skc_obj_align
= align
;
1557 /* If none passed select a cache type based on object size */
1558 if (!(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
))) {
1559 if (spl_obj_size(skc
) < (PAGE_SIZE
/ 8))
1560 skc
->skc_flags
|= KMC_KMEM
;
1562 skc
->skc_flags
|= KMC_VMEM
;
1565 rc
= spl_slab_size(skc
, &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
1569 rc
= spl_magazine_create(skc
);
1573 skc
->skc_taskqid
= taskq_dispatch_delay(spl_kmem_cache_taskq
,
1574 spl_cache_age
, skc
, TQ_SLEEP
,
1575 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1577 down_write(&spl_kmem_cache_sem
);
1578 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
1579 up_write(&spl_kmem_cache_sem
);
1583 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1584 kmem_free(skc
, sizeof(*skc
));
1587 EXPORT_SYMBOL(spl_kmem_cache_create
);
1590 * Register a move callback to for cache defragmentation.
1591 * XXX: Unimplemented but harmless to stub out for now.
1594 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
1595 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
1597 ASSERT(move
!= NULL
);
1599 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
1602 * Destroy a cache and all objects associated with the cache.
1605 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
1607 DECLARE_WAIT_QUEUE_HEAD(wq
);
1611 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1613 down_write(&spl_kmem_cache_sem
);
1614 list_del_init(&skc
->skc_list
);
1615 up_write(&spl_kmem_cache_sem
);
1617 /* Cancel any and wait for any pending delayed tasks */
1618 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1620 spin_lock(&skc
->skc_lock
);
1621 id
= skc
->skc_taskqid
;
1622 spin_unlock(&skc
->skc_lock
);
1624 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1626 /* Wait until all current callers complete, this is mainly
1627 * to catch the case where a low memory situation triggers a
1628 * cache reaping action which races with this destroy. */
1629 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
1631 spl_magazine_destroy(skc
);
1632 spl_slab_reclaim(skc
, 0, 1);
1633 spin_lock(&skc
->skc_lock
);
1635 /* Validate there are no objects in use and free all the
1636 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers. */
1637 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
1638 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
1639 ASSERT3U(skc
->skc_slab_total
, ==, 0);
1640 ASSERT3U(skc
->skc_obj_total
, ==, 0);
1641 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
1642 ASSERT(list_empty(&skc
->skc_complete_list
));
1644 kmem_free(skc
->skc_name
, skc
->skc_name_size
);
1645 spin_unlock(&skc
->skc_lock
);
1647 kmem_free(skc
, sizeof(*skc
));
1651 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
1654 * Allocate an object from a slab attached to the cache. This is used to
1655 * repopulate the per-cpu magazine caches in batches when they run low.
1658 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
1660 spl_kmem_obj_t
*sko
;
1662 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1663 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1664 ASSERT(spin_is_locked(&skc
->skc_lock
));
1666 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
1667 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1668 ASSERT(sko
->sko_addr
!= NULL
);
1670 /* Remove from sks_free_list */
1671 list_del_init(&sko
->sko_list
);
1673 sks
->sks_age
= jiffies
;
1675 skc
->skc_obj_alloc
++;
1677 /* Track max obj usage statistics */
1678 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
1679 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
1681 /* Track max slab usage statistics */
1682 if (sks
->sks_ref
== 1) {
1683 skc
->skc_slab_alloc
++;
1685 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
1686 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
1689 return sko
->sko_addr
;
1693 * Generic slab allocation function to run by the global work queues.
1694 * It is responsible for allocating a new slab, linking it in to the list
1695 * of partial slabs, and then waking any waiters.
1698 spl_cache_grow_work(void *data
)
1700 spl_kmem_alloc_t
*ska
=
1701 spl_get_work_data(data
, spl_kmem_alloc_t
, ska_work
.work
);
1702 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
1703 spl_kmem_slab_t
*sks
;
1705 sks
= spl_slab_alloc(skc
, ska
->ska_flags
| __GFP_NORETRY
| KM_NODEBUG
);
1706 spin_lock(&skc
->skc_lock
);
1708 skc
->skc_slab_total
++;
1709 skc
->skc_obj_total
+= sks
->sks_objs
;
1710 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1713 atomic_dec(&skc
->skc_ref
);
1714 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1715 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1716 wake_up_all(&skc
->skc_waitq
);
1717 spin_unlock(&skc
->skc_lock
);
1723 * Returns non-zero when a new slab should be available.
1726 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
1728 return !test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1732 spl_cache_reclaim_wait(void *word
)
1739 * No available objects on any slabs, create a new slab.
1742 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1747 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1752 * Before allocating a new slab wait for any reaping to complete and
1753 * then return so the local magazine can be rechecked for new objects.
1755 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1756 rc
= wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1757 spl_cache_reclaim_wait
, TASK_UNINTERRUPTIBLE
);
1758 SRETURN(rc
? rc
: -EAGAIN
);
1762 * This is handled by dispatching a work request to the global work
1763 * queue. This allows us to asynchronously allocate a new slab while
1764 * retaining the ability to safely fall back to a smaller synchronous
1765 * allocations to ensure forward progress is always maintained.
1767 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1768 spl_kmem_alloc_t
*ska
;
1770 ska
= kmalloc(sizeof(*ska
), flags
);
1772 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1773 wake_up_all(&skc
->skc_waitq
);
1777 atomic_inc(&skc
->skc_ref
);
1778 ska
->ska_cache
= skc
;
1779 ska
->ska_flags
= flags
& ~__GFP_FS
;
1780 spl_init_delayed_work(&ska
->ska_work
, spl_cache_grow_work
, ska
);
1781 schedule_delayed_work(&ska
->ska_work
, 0);
1785 * The goal here is to only detect the rare case where a virtual slab
1786 * allocation has deadlocked. We must be careful to minimize the use
1787 * of emergency objects which are more expensive to track. Therefore,
1788 * we set a very long timeout for the asynchronous allocation and if
1789 * the timeout is reached the cache is flagged as deadlocked. From
1790 * this point only new emergency objects will be allocated until the
1791 * asynchronous allocation completes and clears the deadlocked flag.
1793 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1794 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1796 remaining
= wait_event_timeout(skc
->skc_waitq
,
1797 spl_cache_grow_wait(skc
), HZ
);
1799 if (!remaining
&& test_bit(KMC_BIT_VMEM
, &skc
->skc_flags
)) {
1800 spin_lock(&skc
->skc_lock
);
1801 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1802 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1803 skc
->skc_obj_deadlock
++;
1805 spin_unlock(&skc
->skc_lock
);
1815 * Refill a per-cpu magazine with objects from the slabs for this cache.
1816 * Ideally the magazine can be repopulated using existing objects which have
1817 * been released, however if we are unable to locate enough free objects new
1818 * slabs of objects will be created. On success NULL is returned, otherwise
1819 * the address of a single emergency object is returned for use by the caller.
1822 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1824 spl_kmem_slab_t
*sks
;
1825 int count
= 0, rc
, refill
;
1829 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1830 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1832 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1833 spin_lock(&skc
->skc_lock
);
1835 while (refill
> 0) {
1836 /* No slabs available we may need to grow the cache */
1837 if (list_empty(&skc
->skc_partial_list
)) {
1838 spin_unlock(&skc
->skc_lock
);
1841 rc
= spl_cache_grow(skc
, flags
, &obj
);
1842 local_irq_disable();
1844 /* Emergency object for immediate use by caller */
1845 if (rc
== 0 && obj
!= NULL
)
1851 /* Rescheduled to different CPU skm is not local */
1852 if (skm
!= skc
->skc_mag
[smp_processor_id()])
1855 /* Potentially rescheduled to the same CPU but
1856 * allocations may have occurred from this CPU while
1857 * we were sleeping so recalculate max refill. */
1858 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
1860 spin_lock(&skc
->skc_lock
);
1864 /* Grab the next available slab */
1865 sks
= list_entry((&skc
->skc_partial_list
)->next
,
1866 spl_kmem_slab_t
, sks_list
);
1867 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1868 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
1869 ASSERT(!list_empty(&sks
->sks_free_list
));
1871 /* Consume as many objects as needed to refill the requested
1872 * cache. We must also be careful not to overfill it. */
1873 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 && ++count
) {
1874 ASSERT(skm
->skm_avail
< skm
->skm_size
);
1875 ASSERT(count
< skm
->skm_size
);
1876 skm
->skm_objs
[skm
->skm_avail
++]=spl_cache_obj(skc
,sks
);
1879 /* Move slab to skc_complete_list when full */
1880 if (sks
->sks_ref
== sks
->sks_objs
) {
1881 list_del(&sks
->sks_list
);
1882 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
1886 spin_unlock(&skc
->skc_lock
);
1892 * Release an object back to the slab from which it came.
1895 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
1897 spl_kmem_slab_t
*sks
= NULL
;
1898 spl_kmem_obj_t
*sko
= NULL
;
1901 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1902 ASSERT(spin_is_locked(&skc
->skc_lock
));
1904 sko
= spl_sko_from_obj(skc
, obj
);
1905 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1906 sks
= sko
->sko_slab
;
1907 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1908 ASSERT(sks
->sks_cache
== skc
);
1909 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
1911 sks
->sks_age
= jiffies
;
1913 skc
->skc_obj_alloc
--;
1915 /* Move slab to skc_partial_list when no longer full. Slabs
1916 * are added to the head to keep the partial list is quasi-full
1917 * sorted order. Fuller at the head, emptier at the tail. */
1918 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
1919 list_del(&sks
->sks_list
);
1920 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
1923 /* Move empty slabs to the end of the partial list so
1924 * they can be easily found and freed during reclamation. */
1925 if (sks
->sks_ref
== 0) {
1926 list_del(&sks
->sks_list
);
1927 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1928 skc
->skc_slab_alloc
--;
1935 * Release a batch of objects from a per-cpu magazine back to their
1936 * respective slabs. This occurs when we exceed the magazine size,
1937 * are under memory pressure, when the cache is idle, or during
1938 * cache cleanup. The flush argument contains the number of entries
1939 * to remove from the magazine.
1942 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
1944 int i
, count
= MIN(flush
, skm
->skm_avail
);
1947 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1948 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1951 * XXX: Currently we simply return objects from the magazine to
1952 * the slabs in fifo order. The ideal thing to do from a memory
1953 * fragmentation standpoint is to cheaply determine the set of
1954 * objects in the magazine which will result in the largest
1955 * number of free slabs if released from the magazine.
1957 spin_lock(&skc
->skc_lock
);
1958 for (i
= 0; i
< count
; i
++)
1959 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
1961 skm
->skm_avail
-= count
;
1962 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
1963 sizeof(void *) * skm
->skm_avail
);
1965 spin_unlock(&skc
->skc_lock
);
1971 * Allocate an object from the per-cpu magazine, or if the magazine
1972 * is empty directly allocate from a slab and repopulate the magazine.
1975 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
1977 spl_kmem_magazine_t
*skm
;
1978 unsigned long irq_flags
;
1982 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1983 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1984 ASSERT(flags
& KM_SLEEP
);
1985 atomic_inc(&skc
->skc_ref
);
1986 local_irq_save(irq_flags
);
1989 /* Safe to update per-cpu structure without lock, but
1990 * in the restart case we must be careful to reacquire
1991 * the local magazine since this may have changed
1992 * when we need to grow the cache. */
1993 skm
= skc
->skc_mag
[smp_processor_id()];
1994 ASSERTF(skm
->skm_magic
== SKM_MAGIC
, "%x != %x: %s/%p/%p %x/%x/%x\n",
1995 skm
->skm_magic
, SKM_MAGIC
, skc
->skc_name
, skc
, skm
,
1996 skm
->skm_size
, skm
->skm_refill
, skm
->skm_avail
);
1998 if (likely(skm
->skm_avail
)) {
1999 /* Object available in CPU cache, use it */
2000 obj
= skm
->skm_objs
[--skm
->skm_avail
];
2001 skm
->skm_age
= jiffies
;
2003 obj
= spl_cache_refill(skc
, skm
, flags
);
2005 SGOTO(restart
, obj
= NULL
);
2008 local_irq_restore(irq_flags
);
2010 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
2012 /* Pre-emptively migrate object to CPU L1 cache */
2014 atomic_dec(&skc
->skc_ref
);
2018 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
2021 * Free an object back to the local per-cpu magazine, there is no
2022 * guarantee that this is the same magazine the object was originally
2023 * allocated from. We may need to flush entire from the magazine
2024 * back to the slabs to make space.
2027 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
2029 spl_kmem_magazine_t
*skm
;
2030 unsigned long flags
;
2033 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2034 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2035 atomic_inc(&skc
->skc_ref
);
2038 * Only virtual slabs may have emergency objects and these objects
2039 * are guaranteed to have physical addresses. They must be removed
2040 * from the tree of emergency objects and the freed.
2042 if ((skc
->skc_flags
& KMC_VMEM
) && !kmem_virt(obj
))
2043 SGOTO(out
, spl_emergency_free(skc
, obj
));
2045 local_irq_save(flags
);
2047 /* Safe to update per-cpu structure without lock, but
2048 * no remote memory allocation tracking is being performed
2049 * it is entirely possible to allocate an object from one
2050 * CPU cache and return it to another. */
2051 skm
= skc
->skc_mag
[smp_processor_id()];
2052 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
2054 /* Per-CPU cache full, flush it to make space */
2055 if (unlikely(skm
->skm_avail
>= skm
->skm_size
))
2056 (void)spl_cache_flush(skc
, skm
, skm
->skm_refill
);
2058 /* Available space in cache, use it */
2059 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
2061 local_irq_restore(flags
);
2063 atomic_dec(&skc
->skc_ref
);
2067 EXPORT_SYMBOL(spl_kmem_cache_free
);
2070 * The generic shrinker function for all caches. Under Linux a shrinker
2071 * may not be tightly coupled with a slab cache. In fact Linux always
2072 * systematically tries calling all registered shrinker callbacks which
2073 * report that they contain unused objects. Because of this we only
2074 * register one shrinker function in the shim layer for all slab caches.
2075 * We always attempt to shrink all caches when this generic shrinker
2076 * is called. The shrinker should return the number of free objects
2077 * in the cache when called with nr_to_scan == 0 but not attempt to
2078 * free any objects. When nr_to_scan > 0 it is a request that nr_to_scan
2079 * objects should be freed, which differs from Solaris semantics.
2080 * Solaris semantics are to free all available objects which may (and
2081 * probably will) be more objects than the requested nr_to_scan.
2084 __spl_kmem_cache_generic_shrinker(struct shrinker
*shrink
,
2085 struct shrink_control
*sc
)
2087 spl_kmem_cache_t
*skc
;
2090 down_read(&spl_kmem_cache_sem
);
2091 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
2093 spl_kmem_cache_reap_now(skc
,
2094 MAX(sc
->nr_to_scan
>> fls64(skc
->skc_slab_objs
), 1));
2097 * Presume everything alloc'ed in reclaimable, this ensures
2098 * we are called again with nr_to_scan > 0 so can try and
2099 * reclaim. The exact number is not important either so
2100 * we forgo taking this already highly contented lock.
2102 unused
+= skc
->skc_obj_alloc
;
2104 up_read(&spl_kmem_cache_sem
);
2106 return (unused
* sysctl_vfs_cache_pressure
) / 100;
2109 SPL_SHRINKER_CALLBACK_WRAPPER(spl_kmem_cache_generic_shrinker
);
2112 * Call the registered reclaim function for a cache. Depending on how
2113 * many and which objects are released it may simply repopulate the
2114 * local magazine which will then need to age-out. Objects which cannot
2115 * fit in the magazine we will be released back to their slabs which will
2116 * also need to age out before being release. This is all just best
2117 * effort and we do not want to thrash creating and destroying slabs.
2120 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
, int count
)
2124 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
2125 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
2127 /* Prevent concurrent cache reaping when contended */
2128 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
2133 atomic_inc(&skc
->skc_ref
);
2136 * When a reclaim function is available it may be invoked repeatedly
2137 * until at least a single slab can be freed. This ensures that we
2138 * do free memory back to the system. This helps minimize the chance
2139 * of an OOM event when the bulk of memory is used by the slab.
2141 * When free slabs are already available the reclaim callback will be
2142 * skipped. Additionally, if no forward progress is detected despite
2143 * a reclaim function the cache will be skipped to avoid deadlock.
2145 * Longer term this would be the correct place to add the code which
2146 * repacks the slabs in order minimize fragmentation.
2148 if (skc
->skc_reclaim
) {
2149 uint64_t objects
= UINT64_MAX
;
2153 spin_lock(&skc
->skc_lock
);
2155 (skc
->skc_slab_total
> 0) &&
2156 ((skc
->skc_slab_total
- skc
->skc_slab_alloc
) == 0) &&
2157 (skc
->skc_obj_alloc
< objects
);
2159 objects
= skc
->skc_obj_alloc
;
2160 spin_unlock(&skc
->skc_lock
);
2163 skc
->skc_reclaim(skc
->skc_private
);
2165 } while (do_reclaim
);
2168 /* Reclaim from the cache, ignoring it's age and delay. */
2169 spl_slab_reclaim(skc
, count
, 1);
2170 clear_bit(KMC_BIT_REAPING
, &skc
->skc_flags
);
2171 smp_mb__after_clear_bit();
2172 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
2174 atomic_dec(&skc
->skc_ref
);
2178 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
2181 * Reap all free slabs from all registered caches.
2186 struct shrink_control sc
;
2188 sc
.nr_to_scan
= KMC_REAP_CHUNK
;
2189 sc
.gfp_mask
= GFP_KERNEL
;
2191 __spl_kmem_cache_generic_shrinker(NULL
, &sc
);
2193 EXPORT_SYMBOL(spl_kmem_reap
);
2195 #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING)
2197 spl_sprintf_addr(kmem_debug_t
*kd
, char *str
, int len
, int min
)
2199 int size
= ((len
- 1) < kd
->kd_size
) ? (len
- 1) : kd
->kd_size
;
2202 ASSERT(str
!= NULL
&& len
>= 17);
2203 memset(str
, 0, len
);
2205 /* Check for a fully printable string, and while we are at
2206 * it place the printable characters in the passed buffer. */
2207 for (i
= 0; i
< size
; i
++) {
2208 str
[i
] = ((char *)(kd
->kd_addr
))[i
];
2209 if (isprint(str
[i
])) {
2212 /* Minimum number of printable characters found
2213 * to make it worthwhile to print this as ascii. */
2223 sprintf(str
, "%02x%02x%02x%02x%02x%02x%02x%02x",
2224 *((uint8_t *)kd
->kd_addr
),
2225 *((uint8_t *)kd
->kd_addr
+ 2),
2226 *((uint8_t *)kd
->kd_addr
+ 4),
2227 *((uint8_t *)kd
->kd_addr
+ 6),
2228 *((uint8_t *)kd
->kd_addr
+ 8),
2229 *((uint8_t *)kd
->kd_addr
+ 10),
2230 *((uint8_t *)kd
->kd_addr
+ 12),
2231 *((uint8_t *)kd
->kd_addr
+ 14));
2238 spl_kmem_init_tracking(struct list_head
*list
, spinlock_t
*lock
, int size
)
2243 spin_lock_init(lock
);
2244 INIT_LIST_HEAD(list
);
2246 for (i
= 0; i
< size
; i
++)
2247 INIT_HLIST_HEAD(&kmem_table
[i
]);
2253 spl_kmem_fini_tracking(struct list_head
*list
, spinlock_t
*lock
)
2255 unsigned long flags
;
2260 spin_lock_irqsave(lock
, flags
);
2261 if (!list_empty(list
))
2262 printk(KERN_WARNING
"%-16s %-5s %-16s %s:%s\n", "address",
2263 "size", "data", "func", "line");
2265 list_for_each_entry(kd
, list
, kd_list
)
2266 printk(KERN_WARNING
"%p %-5d %-16s %s:%d\n", kd
->kd_addr
,
2267 (int)kd
->kd_size
, spl_sprintf_addr(kd
, str
, 17, 8),
2268 kd
->kd_func
, kd
->kd_line
);
2270 spin_unlock_irqrestore(lock
, flags
);
2273 #else /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2274 #define spl_kmem_init_tracking(list, lock, size)
2275 #define spl_kmem_fini_tracking(list, lock)
2276 #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
2279 spl_kmem_init_globals(void)
2283 /* For now all zones are includes, it may be wise to restrict
2284 * this to normal and highmem zones if we see problems. */
2285 for_each_zone(zone
) {
2287 if (!populated_zone(zone
))
2290 minfree
+= min_wmark_pages(zone
);
2291 desfree
+= low_wmark_pages(zone
);
2292 lotsfree
+= high_wmark_pages(zone
);
2295 /* Solaris default values */
2296 swapfs_minfree
= MAX(2*1024*1024 >> PAGE_SHIFT
, physmem
>> 3);
2297 swapfs_reserve
= MIN(4*1024*1024 >> PAGE_SHIFT
, physmem
>> 4);
2301 * Called at module init when it is safe to use spl_kallsyms_lookup_name()
2304 spl_kmem_init_kallsyms_lookup(void)
2306 #ifndef HAVE_GET_VMALLOC_INFO
2307 get_vmalloc_info_fn
= (get_vmalloc_info_t
)
2308 spl_kallsyms_lookup_name("get_vmalloc_info");
2309 if (!get_vmalloc_info_fn
) {
2310 printk(KERN_ERR
"Error: Unknown symbol get_vmalloc_info\n");
2313 #endif /* HAVE_GET_VMALLOC_INFO */
2315 #ifdef HAVE_PGDAT_HELPERS
2316 # ifndef HAVE_FIRST_ONLINE_PGDAT
2317 first_online_pgdat_fn
= (first_online_pgdat_t
)
2318 spl_kallsyms_lookup_name("first_online_pgdat");
2319 if (!first_online_pgdat_fn
) {
2320 printk(KERN_ERR
"Error: Unknown symbol first_online_pgdat\n");
2323 # endif /* HAVE_FIRST_ONLINE_PGDAT */
2325 # ifndef HAVE_NEXT_ONLINE_PGDAT
2326 next_online_pgdat_fn
= (next_online_pgdat_t
)
2327 spl_kallsyms_lookup_name("next_online_pgdat");
2328 if (!next_online_pgdat_fn
) {
2329 printk(KERN_ERR
"Error: Unknown symbol next_online_pgdat\n");
2332 # endif /* HAVE_NEXT_ONLINE_PGDAT */
2334 # ifndef HAVE_NEXT_ZONE
2335 next_zone_fn
= (next_zone_t
)
2336 spl_kallsyms_lookup_name("next_zone");
2337 if (!next_zone_fn
) {
2338 printk(KERN_ERR
"Error: Unknown symbol next_zone\n");
2341 # endif /* HAVE_NEXT_ZONE */
2343 #else /* HAVE_PGDAT_HELPERS */
2345 # ifndef HAVE_PGDAT_LIST
2346 pgdat_list_addr
= *(struct pglist_data
**)
2347 spl_kallsyms_lookup_name("pgdat_list");
2348 if (!pgdat_list_addr
) {
2349 printk(KERN_ERR
"Error: Unknown symbol pgdat_list\n");
2352 # endif /* HAVE_PGDAT_LIST */
2353 #endif /* HAVE_PGDAT_HELPERS */
2355 #if defined(NEED_GET_ZONE_COUNTS) && !defined(HAVE_GET_ZONE_COUNTS)
2356 get_zone_counts_fn
= (get_zone_counts_t
)
2357 spl_kallsyms_lookup_name("get_zone_counts");
2358 if (!get_zone_counts_fn
) {
2359 printk(KERN_ERR
"Error: Unknown symbol get_zone_counts\n");
2362 #endif /* NEED_GET_ZONE_COUNTS && !HAVE_GET_ZONE_COUNTS */
2365 * It is now safe to initialize the global tunings which rely on
2366 * the use of the for_each_zone() macro. This macro in turns
2367 * depends on the *_pgdat symbols which are now available.
2369 spl_kmem_init_globals();
2371 #if !defined(HAVE_INVALIDATE_INODES) && !defined(HAVE_INVALIDATE_INODES_CHECK)
2372 invalidate_inodes_fn
= (invalidate_inodes_t
)
2373 spl_kallsyms_lookup_name("invalidate_inodes");
2374 if (!invalidate_inodes_fn
) {
2375 printk(KERN_ERR
"Error: Unknown symbol invalidate_inodes\n");
2378 #endif /* !HAVE_INVALIDATE_INODES && !HAVE_INVALIDATE_INODES_CHECK */
2380 #ifndef HAVE_SHRINK_DCACHE_MEMORY
2381 /* When shrink_dcache_memory_fn == NULL support is disabled */
2382 shrink_dcache_memory_fn
= (shrink_dcache_memory_t
)
2383 spl_kallsyms_lookup_name("shrink_dcache_memory");
2384 #endif /* HAVE_SHRINK_DCACHE_MEMORY */
2386 #ifndef HAVE_SHRINK_ICACHE_MEMORY
2387 /* When shrink_icache_memory_fn == NULL support is disabled */
2388 shrink_icache_memory_fn
= (shrink_icache_memory_t
)
2389 spl_kallsyms_lookup_name("shrink_icache_memory");
2390 #endif /* HAVE_SHRINK_ICACHE_MEMORY */
2401 init_rwsem(&spl_kmem_cache_sem
);
2402 INIT_LIST_HEAD(&spl_kmem_cache_list
);
2403 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
2404 1, maxclsyspri
, 1, 32, TASKQ_PREPOPULATE
);
2406 spl_register_shrinker(&spl_kmem_cache_shrinker
);
2409 kmem_alloc_used_set(0);
2410 vmem_alloc_used_set(0);
2412 spl_kmem_init_tracking(&kmem_list
, &kmem_lock
, KMEM_TABLE_SIZE
);
2413 spl_kmem_init_tracking(&vmem_list
, &vmem_lock
, VMEM_TABLE_SIZE
);
2422 /* Display all unreclaimed memory addresses, including the
2423 * allocation size and the first few bytes of what's located
2424 * at that address to aid in debugging. Performance is not
2425 * a serious concern here since it is module unload time. */
2426 if (kmem_alloc_used_read() != 0)
2427 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2428 "kmem leaked %ld/%ld bytes\n",
2429 kmem_alloc_used_read(), kmem_alloc_max
);
2432 if (vmem_alloc_used_read() != 0)
2433 SDEBUG_LIMIT(SD_CONSOLE
| SD_WARNING
,
2434 "vmem leaked %ld/%ld bytes\n",
2435 vmem_alloc_used_read(), vmem_alloc_max
);
2437 spl_kmem_fini_tracking(&kmem_list
, &kmem_lock
);
2438 spl_kmem_fini_tracking(&vmem_list
, &vmem_lock
);
2439 #endif /* DEBUG_KMEM */
2442 spl_unregister_shrinker(&spl_kmem_cache_shrinker
);
2443 taskq_destroy(spl_kmem_cache_taskq
);