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.
10 * The SPL is free software; you can redistribute it and/or modify it
11 * under the terms of the GNU General Public License as published by the
12 * Free Software Foundation; either version 2 of the License, or (at your
13 * option) any later version.
15 * The SPL is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
20 * You should have received a copy of the GNU General Public License along
21 * with the SPL. If not, see <http://www.gnu.org/licenses/>.
24 #include <linux/percpu_compat.h>
26 #include <sys/kmem_cache.h>
27 #include <sys/taskq.h>
28 #include <sys/timer.h>
31 #include <linux/slab.h>
32 #include <linux/swap.h>
33 #include <linux/prefetch.h>
36 * Within the scope of spl-kmem.c file the kmem_cache_* definitions
37 * are removed to allow access to the real Linux slab allocator.
39 #undef kmem_cache_destroy
40 #undef kmem_cache_create
41 #undef kmem_cache_alloc
42 #undef kmem_cache_free
46 * Linux 3.16 replaced smp_mb__{before,after}_{atomic,clear}_{dec,inc,bit}()
47 * with smp_mb__{before,after}_atomic() because they were redundant. This is
48 * only used inside our SLAB allocator, so we implement an internal wrapper
49 * here to give us smp_mb__{before,after}_atomic() on older kernels.
51 #ifndef smp_mb__before_atomic
52 #define smp_mb__before_atomic(x) smp_mb__before_clear_bit(x)
55 #ifndef smp_mb__after_atomic
56 #define smp_mb__after_atomic(x) smp_mb__after_clear_bit(x)
61 * Cache magazines are an optimization designed to minimize the cost of
62 * allocating memory. They do this by keeping a per-cpu cache of recently
63 * freed objects, which can then be reallocated without taking a lock. This
64 * can improve performance on highly contended caches. However, because
65 * objects in magazines will prevent otherwise empty slabs from being
66 * immediately released this may not be ideal for low memory machines.
68 * For this reason spl_kmem_cache_magazine_size can be used to set a maximum
69 * magazine size. When this value is set to 0 the magazine size will be
70 * automatically determined based on the object size. Otherwise magazines
71 * will be limited to 2-256 objects per magazine (i.e per cpu). Magazines
72 * may never be entirely disabled in this implementation.
74 static unsigned int spl_kmem_cache_magazine_size
= 0;
75 module_param(spl_kmem_cache_magazine_size
, uint
, 0444);
76 MODULE_PARM_DESC(spl_kmem_cache_magazine_size
,
77 "Default magazine size (2-256), set automatically (0)");
79 static unsigned int spl_kmem_cache_obj_per_slab
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
80 module_param(spl_kmem_cache_obj_per_slab
, uint
, 0644);
81 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab
, "Number of objects per slab");
83 static unsigned int spl_kmem_cache_max_size
= SPL_KMEM_CACHE_MAX_SIZE
;
84 module_param(spl_kmem_cache_max_size
, uint
, 0644);
85 MODULE_PARM_DESC(spl_kmem_cache_max_size
, "Maximum size of slab in MB");
88 * For small objects the Linux slab allocator should be used to make the most
89 * efficient use of the memory. However, large objects are not supported by
90 * the Linux slab and therefore the SPL implementation is preferred. A cutoff
91 * of 16K was determined to be optimal for architectures using 4K pages and
92 * to also work well on architecutres using larger 64K page sizes.
94 static unsigned int spl_kmem_cache_slab_limit
= 16384;
95 module_param(spl_kmem_cache_slab_limit
, uint
, 0644);
96 MODULE_PARM_DESC(spl_kmem_cache_slab_limit
,
97 "Objects less than N bytes use the Linux slab");
100 * The number of threads available to allocate new slabs for caches. This
101 * should not need to be tuned but it is available for performance analysis.
103 static unsigned int spl_kmem_cache_kmem_threads
= 4;
104 module_param(spl_kmem_cache_kmem_threads
, uint
, 0444);
105 MODULE_PARM_DESC(spl_kmem_cache_kmem_threads
,
106 "Number of spl_kmem_cache threads");
110 * Slab allocation interfaces
112 * While the Linux slab implementation was inspired by the Solaris
113 * implementation I cannot use it to emulate the Solaris APIs. I
114 * require two features which are not provided by the Linux slab.
116 * 1) Constructors AND destructors. Recent versions of the Linux
117 * kernel have removed support for destructors. This is a deal
118 * breaker for the SPL which contains particularly expensive
119 * initializers for mutex's, condition variables, etc. We also
120 * require a minimal level of cleanup for these data types unlike
121 * many Linux data types which do need to be explicitly destroyed.
123 * 2) Virtual address space backed slab. Callers of the Solaris slab
124 * expect it to work well for both small are very large allocations.
125 * Because of memory fragmentation the Linux slab which is backed
126 * by kmalloc'ed memory performs very badly when confronted with
127 * large numbers of large allocations. Basing the slab on the
128 * virtual address space removes the need for contiguous pages
129 * and greatly improve performance for large allocations.
131 * For these reasons, the SPL has its own slab implementation with
132 * the needed features. It is not as highly optimized as either the
133 * Solaris or Linux slabs, but it should get me most of what is
134 * needed until it can be optimized or obsoleted by another approach.
136 * One serious concern I do have about this method is the relatively
137 * small virtual address space on 32bit arches. This will seriously
138 * constrain the size of the slab caches and their performance.
141 struct list_head spl_kmem_cache_list
; /* List of caches */
142 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
143 static taskq_t
*spl_kmem_cache_taskq
; /* Task queue for aging / reclaim */
145 static void spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
);
148 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
150 gfp_t lflags
= kmem_flags_convert(flags
);
153 ptr
= spl_vmalloc(size
, lflags
| __GFP_HIGHMEM
);
155 /* Resulting allocated memory will be page aligned */
156 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
162 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
164 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
167 * The Linux direct reclaim path uses this out of band value to
168 * determine if forward progress is being made. Normally this is
169 * incremented by kmem_freepages() which is part of the various
170 * Linux slab implementations. However, since we are using none
171 * of that infrastructure we are responsible for incrementing it.
173 if (current
->reclaim_state
)
174 #ifdef HAVE_RECLAIM_STATE_RECLAIMED
175 current
->reclaim_state
->reclaimed
+= size
>> PAGE_SHIFT
;
177 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
183 * Required space for each aligned sks.
185 static inline uint32_t
186 spl_sks_size(spl_kmem_cache_t
*skc
)
188 return (P2ROUNDUP_TYPED(sizeof (spl_kmem_slab_t
),
189 skc
->skc_obj_align
, uint32_t));
193 * Required space for each aligned object.
195 static inline uint32_t
196 spl_obj_size(spl_kmem_cache_t
*skc
)
198 uint32_t align
= skc
->skc_obj_align
;
200 return (P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
201 P2ROUNDUP_TYPED(sizeof (spl_kmem_obj_t
), align
, uint32_t));
205 spl_kmem_cache_inuse(kmem_cache_t
*cache
)
207 return (cache
->skc_obj_total
);
209 EXPORT_SYMBOL(spl_kmem_cache_inuse
);
212 spl_kmem_cache_entry_size(kmem_cache_t
*cache
)
214 return (cache
->skc_obj_size
);
216 EXPORT_SYMBOL(spl_kmem_cache_entry_size
);
219 * Lookup the spl_kmem_object_t for an object given that object.
221 static inline spl_kmem_obj_t
*
222 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
224 return (obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
225 skc
->skc_obj_align
, uint32_t));
229 * It's important that we pack the spl_kmem_obj_t structure and the
230 * actual objects in to one large address space to minimize the number
231 * of calls to the allocator. It is far better to do a few large
232 * allocations and then subdivide it ourselves. Now which allocator
233 * we use requires balancing a few trade offs.
235 * For small objects we use kmem_alloc() because as long as you are
236 * only requesting a small number of pages (ideally just one) its cheap.
237 * However, when you start requesting multiple pages with kmem_alloc()
238 * it gets increasingly expensive since it requires contiguous pages.
239 * For this reason we shift to vmem_alloc() for slabs of large objects
240 * which removes the need for contiguous pages. We do not use
241 * vmem_alloc() in all cases because there is significant locking
242 * overhead in __get_vm_area_node(). This function takes a single
243 * global lock when acquiring an available virtual address range which
244 * serializes all vmem_alloc()'s for all slab caches. Using slightly
245 * different allocation functions for small and large objects should
246 * give us the best of both worlds.
248 * +------------------------+
249 * | spl_kmem_slab_t --+-+ |
250 * | skc_obj_size <-+ | |
251 * | spl_kmem_obj_t | |
252 * | skc_obj_size <---+ |
253 * | spl_kmem_obj_t | |
255 * +------------------------+
257 static spl_kmem_slab_t
*
258 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
260 spl_kmem_slab_t
*sks
;
264 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
268 sks
= (spl_kmem_slab_t
*)base
;
269 sks
->sks_magic
= SKS_MAGIC
;
270 sks
->sks_objs
= skc
->skc_slab_objs
;
271 sks
->sks_age
= jiffies
;
272 sks
->sks_cache
= skc
;
273 INIT_LIST_HEAD(&sks
->sks_list
);
274 INIT_LIST_HEAD(&sks
->sks_free_list
);
276 obj_size
= spl_obj_size(skc
);
278 for (int i
= 0; i
< sks
->sks_objs
; i
++) {
279 void *obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
281 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
282 spl_kmem_obj_t
*sko
= spl_sko_from_obj(skc
, obj
);
284 sko
->sko_magic
= SKO_MAGIC
;
286 INIT_LIST_HEAD(&sko
->sko_list
);
287 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
294 * Remove a slab from complete or partial list, it must be called with
295 * the 'skc->skc_lock' held but the actual free must be performed
296 * outside the lock to prevent deadlocking on vmem addresses.
299 spl_slab_free(spl_kmem_slab_t
*sks
,
300 struct list_head
*sks_list
, struct list_head
*sko_list
)
302 spl_kmem_cache_t
*skc
;
304 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
305 ASSERT(sks
->sks_ref
== 0);
307 skc
= sks
->sks_cache
;
308 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
311 * Update slab/objects counters in the cache, then remove the
312 * slab from the skc->skc_partial_list. Finally add the slab
313 * and all its objects in to the private work lists where the
314 * destructors will be called and the memory freed to the system.
316 skc
->skc_obj_total
-= sks
->sks_objs
;
317 skc
->skc_slab_total
--;
318 list_del(&sks
->sks_list
);
319 list_add(&sks
->sks_list
, sks_list
);
320 list_splice_init(&sks
->sks_free_list
, sko_list
);
324 * Reclaim empty slabs at the end of the partial list.
327 spl_slab_reclaim(spl_kmem_cache_t
*skc
)
329 spl_kmem_slab_t
*sks
= NULL
, *m
= NULL
;
330 spl_kmem_obj_t
*sko
= NULL
, *n
= NULL
;
335 * Empty slabs and objects must be moved to a private list so they
336 * can be safely freed outside the spin lock. All empty slabs are
337 * at the end of skc->skc_partial_list, therefore once a non-empty
338 * slab is found we can stop scanning.
340 spin_lock(&skc
->skc_lock
);
341 list_for_each_entry_safe_reverse(sks
, m
,
342 &skc
->skc_partial_list
, sks_list
) {
344 if (sks
->sks_ref
> 0)
347 spl_slab_free(sks
, &sks_list
, &sko_list
);
349 spin_unlock(&skc
->skc_lock
);
352 * The following two loops ensure all the object destructors are run,
353 * and the slabs themselves are freed. This is all done outside the
354 * skc->skc_lock since this allows the destructor to sleep, and
355 * allows us to perform a conditional reschedule when a freeing a
356 * large number of objects and slabs back to the system.
359 list_for_each_entry_safe(sko
, n
, &sko_list
, sko_list
) {
360 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
363 list_for_each_entry_safe(sks
, m
, &sks_list
, sks_list
) {
364 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
365 kv_free(skc
, sks
, skc
->skc_slab_size
);
369 static spl_kmem_emergency_t
*
370 spl_emergency_search(struct rb_root
*root
, void *obj
)
372 struct rb_node
*node
= root
->rb_node
;
373 spl_kmem_emergency_t
*ske
;
374 unsigned long address
= (unsigned long)obj
;
377 ske
= container_of(node
, spl_kmem_emergency_t
, ske_node
);
379 if (address
< ske
->ske_obj
)
380 node
= node
->rb_left
;
381 else if (address
> ske
->ske_obj
)
382 node
= node
->rb_right
;
391 spl_emergency_insert(struct rb_root
*root
, spl_kmem_emergency_t
*ske
)
393 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
394 spl_kmem_emergency_t
*ske_tmp
;
395 unsigned long address
= ske
->ske_obj
;
398 ske_tmp
= container_of(*new, spl_kmem_emergency_t
, ske_node
);
401 if (address
< ske_tmp
->ske_obj
)
402 new = &((*new)->rb_left
);
403 else if (address
> ske_tmp
->ske_obj
)
404 new = &((*new)->rb_right
);
409 rb_link_node(&ske
->ske_node
, parent
, new);
410 rb_insert_color(&ske
->ske_node
, root
);
416 * Allocate a single emergency object and track it in a red black tree.
419 spl_emergency_alloc(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
421 gfp_t lflags
= kmem_flags_convert(flags
);
422 spl_kmem_emergency_t
*ske
;
423 int order
= get_order(skc
->skc_obj_size
);
426 /* Last chance use a partial slab if one now exists */
427 spin_lock(&skc
->skc_lock
);
428 empty
= list_empty(&skc
->skc_partial_list
);
429 spin_unlock(&skc
->skc_lock
);
433 ske
= kmalloc(sizeof (*ske
), lflags
);
437 ske
->ske_obj
= __get_free_pages(lflags
, order
);
438 if (ske
->ske_obj
== 0) {
443 spin_lock(&skc
->skc_lock
);
444 empty
= spl_emergency_insert(&skc
->skc_emergency_tree
, ske
);
446 skc
->skc_obj_total
++;
447 skc
->skc_obj_emergency
++;
448 if (skc
->skc_obj_emergency
> skc
->skc_obj_emergency_max
)
449 skc
->skc_obj_emergency_max
= skc
->skc_obj_emergency
;
451 spin_unlock(&skc
->skc_lock
);
453 if (unlikely(!empty
)) {
454 free_pages(ske
->ske_obj
, order
);
459 *obj
= (void *)ske
->ske_obj
;
465 * Locate the passed object in the red black tree and free it.
468 spl_emergency_free(spl_kmem_cache_t
*skc
, void *obj
)
470 spl_kmem_emergency_t
*ske
;
471 int order
= get_order(skc
->skc_obj_size
);
473 spin_lock(&skc
->skc_lock
);
474 ske
= spl_emergency_search(&skc
->skc_emergency_tree
, obj
);
476 rb_erase(&ske
->ske_node
, &skc
->skc_emergency_tree
);
477 skc
->skc_obj_emergency
--;
478 skc
->skc_obj_total
--;
480 spin_unlock(&skc
->skc_lock
);
485 free_pages(ske
->ske_obj
, order
);
492 * Release objects from the per-cpu magazine back to their slab. The flush
493 * argument contains the max number of entries to remove from the magazine.
496 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
498 spin_lock(&skc
->skc_lock
);
500 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
501 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
503 int count
= MIN(flush
, skm
->skm_avail
);
504 for (int i
= 0; i
< count
; i
++)
505 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
507 skm
->skm_avail
-= count
;
508 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
509 sizeof (void *) * skm
->skm_avail
);
511 spin_unlock(&skc
->skc_lock
);
515 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
516 * When on-slab we want to target spl_kmem_cache_obj_per_slab. However,
517 * for very small objects we may end up with more than this so as not
518 * to waste space in the minimal allocation of a single page.
521 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
523 uint32_t sks_size
, obj_size
, max_size
, tgt_size
, tgt_objs
;
525 sks_size
= spl_sks_size(skc
);
526 obj_size
= spl_obj_size(skc
);
527 max_size
= (spl_kmem_cache_max_size
* 1024 * 1024);
528 tgt_size
= (spl_kmem_cache_obj_per_slab
* obj_size
+ sks_size
);
530 if (tgt_size
<= max_size
) {
531 tgt_objs
= (tgt_size
- sks_size
) / obj_size
;
533 tgt_objs
= (max_size
- sks_size
) / obj_size
;
534 tgt_size
= (tgt_objs
* obj_size
) + sks_size
;
547 * Make a guess at reasonable per-cpu magazine size based on the size of
548 * each object and the cost of caching N of them in each magazine. Long
549 * term this should really adapt based on an observed usage heuristic.
552 spl_magazine_size(spl_kmem_cache_t
*skc
)
554 uint32_t obj_size
= spl_obj_size(skc
);
557 if (spl_kmem_cache_magazine_size
> 0)
558 return (MAX(MIN(spl_kmem_cache_magazine_size
, 256), 2));
560 /* Per-magazine sizes below assume a 4Kib page size */
561 if (obj_size
> (PAGE_SIZE
* 256))
562 size
= 4; /* Minimum 4Mib per-magazine */
563 else if (obj_size
> (PAGE_SIZE
* 32))
564 size
= 16; /* Minimum 2Mib per-magazine */
565 else if (obj_size
> (PAGE_SIZE
))
566 size
= 64; /* Minimum 256Kib per-magazine */
567 else if (obj_size
> (PAGE_SIZE
/ 4))
568 size
= 128; /* Minimum 128Kib per-magazine */
576 * Allocate a per-cpu magazine to associate with a specific core.
578 static spl_kmem_magazine_t
*
579 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
581 spl_kmem_magazine_t
*skm
;
582 int size
= sizeof (spl_kmem_magazine_t
) +
583 sizeof (void *) * skc
->skc_mag_size
;
585 skm
= kmalloc_node(size
, GFP_KERNEL
, cpu_to_node(cpu
));
587 skm
->skm_magic
= SKM_MAGIC
;
589 skm
->skm_size
= skc
->skc_mag_size
;
590 skm
->skm_refill
= skc
->skc_mag_refill
;
591 skm
->skm_cache
= skc
;
599 * Free a per-cpu magazine associated with a specific core.
602 spl_magazine_free(spl_kmem_magazine_t
*skm
)
604 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
605 ASSERT(skm
->skm_avail
== 0);
610 * Create all pre-cpu magazines of reasonable sizes.
613 spl_magazine_create(spl_kmem_cache_t
*skc
)
617 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
619 skc
->skc_mag
= kzalloc(sizeof (spl_kmem_magazine_t
*) *
620 num_possible_cpus(), kmem_flags_convert(KM_SLEEP
));
621 skc
->skc_mag_size
= spl_magazine_size(skc
);
622 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
624 for_each_possible_cpu(i
) {
625 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
626 if (!skc
->skc_mag
[i
]) {
627 for (i
--; i
>= 0; i
--)
628 spl_magazine_free(skc
->skc_mag
[i
]);
639 * Destroy all pre-cpu magazines.
642 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
644 spl_kmem_magazine_t
*skm
;
647 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
649 for_each_possible_cpu(i
) {
650 skm
= skc
->skc_mag
[i
];
651 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
652 spl_magazine_free(skm
);
659 * Create a object cache based on the following arguments:
661 * size cache object size
662 * align cache object alignment
663 * ctor cache object constructor
664 * dtor cache object destructor
665 * reclaim cache object reclaim
666 * priv cache private data for ctor/dtor/reclaim
667 * vmp unused must be NULL
669 * KMC_KVMEM Force kvmem backed SPL cache
670 * KMC_SLAB Force Linux slab backed cache
671 * KMC_NODEBUG Disable debugging (unsupported)
674 spl_kmem_cache_create(const char *name
, size_t size
, size_t align
,
675 spl_kmem_ctor_t ctor
, spl_kmem_dtor_t dtor
, void *reclaim
,
676 void *priv
, void *vmp
, int flags
)
678 gfp_t lflags
= kmem_flags_convert(KM_SLEEP
);
679 spl_kmem_cache_t
*skc
;
686 ASSERT(reclaim
== NULL
);
690 skc
= kzalloc(sizeof (*skc
), lflags
);
694 skc
->skc_magic
= SKC_MAGIC
;
695 skc
->skc_name_size
= strlen(name
) + 1;
696 skc
->skc_name
= kmalloc(skc
->skc_name_size
, lflags
);
697 if (skc
->skc_name
== NULL
) {
701 strlcpy(skc
->skc_name
, name
, skc
->skc_name_size
);
703 skc
->skc_ctor
= ctor
;
704 skc
->skc_dtor
= dtor
;
705 skc
->skc_private
= priv
;
707 skc
->skc_linux_cache
= NULL
;
708 skc
->skc_flags
= flags
;
709 skc
->skc_obj_size
= size
;
710 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
711 atomic_set(&skc
->skc_ref
, 0);
713 INIT_LIST_HEAD(&skc
->skc_list
);
714 INIT_LIST_HEAD(&skc
->skc_complete_list
);
715 INIT_LIST_HEAD(&skc
->skc_partial_list
);
716 skc
->skc_emergency_tree
= RB_ROOT
;
717 spin_lock_init(&skc
->skc_lock
);
718 init_waitqueue_head(&skc
->skc_waitq
);
719 skc
->skc_slab_fail
= 0;
720 skc
->skc_slab_create
= 0;
721 skc
->skc_slab_destroy
= 0;
722 skc
->skc_slab_total
= 0;
723 skc
->skc_slab_alloc
= 0;
724 skc
->skc_slab_max
= 0;
725 skc
->skc_obj_total
= 0;
726 skc
->skc_obj_alloc
= 0;
727 skc
->skc_obj_max
= 0;
728 skc
->skc_obj_deadlock
= 0;
729 skc
->skc_obj_emergency
= 0;
730 skc
->skc_obj_emergency_max
= 0;
732 rc
= percpu_counter_init_common(&skc
->skc_linux_alloc
, 0,
740 * Verify the requested alignment restriction is sane.
744 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
);
745 VERIFY3U(align
, <=, PAGE_SIZE
);
746 skc
->skc_obj_align
= align
;
750 * When no specific type of slab is requested (kmem, vmem, or
751 * linuxslab) then select a cache type based on the object size
752 * and default tunables.
754 if (!(skc
->skc_flags
& (KMC_SLAB
| KMC_KVMEM
))) {
755 if (spl_kmem_cache_slab_limit
&&
756 size
<= (size_t)spl_kmem_cache_slab_limit
) {
758 * Objects smaller than spl_kmem_cache_slab_limit can
759 * use the Linux slab for better space-efficiency.
761 skc
->skc_flags
|= KMC_SLAB
;
764 * All other objects are considered large and are
765 * placed on kvmem backed slabs.
767 skc
->skc_flags
|= KMC_KVMEM
;
772 * Given the type of slab allocate the required resources.
774 if (skc
->skc_flags
& KMC_KVMEM
) {
775 rc
= spl_slab_size(skc
,
776 &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
780 rc
= spl_magazine_create(skc
);
784 unsigned long slabflags
= 0;
786 if (size
> (SPL_MAX_KMEM_ORDER_NR_PAGES
* PAGE_SIZE
))
789 #if defined(SLAB_USERCOPY)
791 * Required for PAX-enabled kernels if the slab is to be
792 * used for copying between user and kernel space.
794 slabflags
|= SLAB_USERCOPY
;
797 #if defined(HAVE_KMEM_CACHE_CREATE_USERCOPY)
799 * Newer grsec patchset uses kmem_cache_create_usercopy()
800 * instead of SLAB_USERCOPY flag
802 skc
->skc_linux_cache
= kmem_cache_create_usercopy(
803 skc
->skc_name
, size
, align
, slabflags
, 0, size
, NULL
);
805 skc
->skc_linux_cache
= kmem_cache_create(
806 skc
->skc_name
, size
, align
, slabflags
, NULL
);
808 if (skc
->skc_linux_cache
== NULL
)
812 down_write(&spl_kmem_cache_sem
);
813 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
814 up_write(&spl_kmem_cache_sem
);
818 kfree(skc
->skc_name
);
819 percpu_counter_destroy(&skc
->skc_linux_alloc
);
823 EXPORT_SYMBOL(spl_kmem_cache_create
);
826 * Register a move callback for cache defragmentation.
827 * XXX: Unimplemented but harmless to stub out for now.
830 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
831 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
833 ASSERT(move
!= NULL
);
835 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
838 * Destroy a cache and all objects associated with the cache.
841 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
843 DECLARE_WAIT_QUEUE_HEAD(wq
);
846 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
847 ASSERT(skc
->skc_flags
& (KMC_KVMEM
| KMC_SLAB
));
849 down_write(&spl_kmem_cache_sem
);
850 list_del_init(&skc
->skc_list
);
851 up_write(&spl_kmem_cache_sem
);
853 /* Cancel any and wait for any pending delayed tasks */
854 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
856 spin_lock(&skc
->skc_lock
);
857 id
= skc
->skc_taskqid
;
858 spin_unlock(&skc
->skc_lock
);
860 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
863 * Wait until all current callers complete, this is mainly
864 * to catch the case where a low memory situation triggers a
865 * cache reaping action which races with this destroy.
867 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
869 if (skc
->skc_flags
& KMC_KVMEM
) {
870 spl_magazine_destroy(skc
);
871 spl_slab_reclaim(skc
);
873 ASSERT(skc
->skc_flags
& KMC_SLAB
);
874 kmem_cache_destroy(skc
->skc_linux_cache
);
877 spin_lock(&skc
->skc_lock
);
880 * Validate there are no objects in use and free all the
881 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers.
883 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
884 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
885 ASSERT3U(skc
->skc_slab_total
, ==, 0);
886 ASSERT3U(skc
->skc_obj_total
, ==, 0);
887 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
888 ASSERT(list_empty(&skc
->skc_complete_list
));
890 ASSERT3U(percpu_counter_sum(&skc
->skc_linux_alloc
), ==, 0);
891 percpu_counter_destroy(&skc
->skc_linux_alloc
);
893 spin_unlock(&skc
->skc_lock
);
895 kfree(skc
->skc_name
);
898 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
901 * Allocate an object from a slab attached to the cache. This is used to
902 * repopulate the per-cpu magazine caches in batches when they run low.
905 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
909 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
910 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
912 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
913 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
914 ASSERT(sko
->sko_addr
!= NULL
);
916 /* Remove from sks_free_list */
917 list_del_init(&sko
->sko_list
);
919 sks
->sks_age
= jiffies
;
921 skc
->skc_obj_alloc
++;
923 /* Track max obj usage statistics */
924 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
925 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
927 /* Track max slab usage statistics */
928 if (sks
->sks_ref
== 1) {
929 skc
->skc_slab_alloc
++;
931 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
932 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
935 return (sko
->sko_addr
);
939 * Generic slab allocation function to run by the global work queues.
940 * It is responsible for allocating a new slab, linking it in to the list
941 * of partial slabs, and then waking any waiters.
944 __spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
)
946 spl_kmem_slab_t
*sks
;
948 fstrans_cookie_t cookie
= spl_fstrans_mark();
949 sks
= spl_slab_alloc(skc
, flags
);
950 spl_fstrans_unmark(cookie
);
952 spin_lock(&skc
->skc_lock
);
954 skc
->skc_slab_total
++;
955 skc
->skc_obj_total
+= sks
->sks_objs
;
956 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
958 smp_mb__before_atomic();
959 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
960 smp_mb__after_atomic();
962 spin_unlock(&skc
->skc_lock
);
964 return (sks
== NULL
? -ENOMEM
: 0);
968 spl_cache_grow_work(void *data
)
970 spl_kmem_alloc_t
*ska
= (spl_kmem_alloc_t
*)data
;
971 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
973 int error
= __spl_cache_grow(skc
, ska
->ska_flags
);
975 atomic_dec(&skc
->skc_ref
);
976 smp_mb__before_atomic();
977 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
978 smp_mb__after_atomic();
980 wake_up_all(&skc
->skc_waitq
);
986 * Returns non-zero when a new slab should be available.
989 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
991 return (!test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
));
995 * No available objects on any slabs, create a new slab. Note that this
996 * functionality is disabled for KMC_SLAB caches which are backed by the
1000 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1002 int remaining
, rc
= 0;
1004 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1005 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1006 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
1011 * Since we can't sleep attempt an emergency allocation to satisfy
1012 * the request. The only alterative is to fail the allocation but
1013 * it's preferable try. The use of KM_NOSLEEP is expected to be rare.
1015 if (flags
& KM_NOSLEEP
)
1016 return (spl_emergency_alloc(skc
, flags
, obj
));
1021 * Before allocating a new slab wait for any reaping to complete and
1022 * then return so the local magazine can be rechecked for new objects.
1024 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1025 rc
= spl_wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1026 TASK_UNINTERRUPTIBLE
);
1027 return (rc
? rc
: -EAGAIN
);
1031 * Note: It would be nice to reduce the overhead of context switch
1032 * and improve NUMA locality, by trying to allocate a new slab in the
1033 * current process context with KM_NOSLEEP flag.
1035 * However, this can't be applied to vmem/kvmem due to a bug that
1036 * spl_vmalloc() doesn't honor gfp flags in page table allocation.
1040 * This is handled by dispatching a work request to the global work
1041 * queue. This allows us to asynchronously allocate a new slab while
1042 * retaining the ability to safely fall back to a smaller synchronous
1043 * allocations to ensure forward progress is always maintained.
1045 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1046 spl_kmem_alloc_t
*ska
;
1048 ska
= kmalloc(sizeof (*ska
), kmem_flags_convert(flags
));
1050 clear_bit_unlock(KMC_BIT_GROWING
, &skc
->skc_flags
);
1051 smp_mb__after_atomic();
1052 wake_up_all(&skc
->skc_waitq
);
1056 atomic_inc(&skc
->skc_ref
);
1057 ska
->ska_cache
= skc
;
1058 ska
->ska_flags
= flags
;
1059 taskq_init_ent(&ska
->ska_tqe
);
1060 taskq_dispatch_ent(spl_kmem_cache_taskq
,
1061 spl_cache_grow_work
, ska
, 0, &ska
->ska_tqe
);
1065 * The goal here is to only detect the rare case where a virtual slab
1066 * allocation has deadlocked. We must be careful to minimize the use
1067 * of emergency objects which are more expensive to track. Therefore,
1068 * we set a very long timeout for the asynchronous allocation and if
1069 * the timeout is reached the cache is flagged as deadlocked. From
1070 * this point only new emergency objects will be allocated until the
1071 * asynchronous allocation completes and clears the deadlocked flag.
1073 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1074 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1076 remaining
= wait_event_timeout(skc
->skc_waitq
,
1077 spl_cache_grow_wait(skc
), HZ
/ 10);
1080 spin_lock(&skc
->skc_lock
);
1081 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1082 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1083 skc
->skc_obj_deadlock
++;
1085 spin_unlock(&skc
->skc_lock
);
1095 * Refill a per-cpu magazine with objects from the slabs for this cache.
1096 * Ideally the magazine can be repopulated using existing objects which have
1097 * been released, however if we are unable to locate enough free objects new
1098 * slabs of objects will be created. On success NULL is returned, otherwise
1099 * the address of a single emergency object is returned for use by the caller.
1102 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1104 spl_kmem_slab_t
*sks
;
1105 int count
= 0, rc
, refill
;
1108 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1109 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1111 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1112 spin_lock(&skc
->skc_lock
);
1114 while (refill
> 0) {
1115 /* No slabs available we may need to grow the cache */
1116 if (list_empty(&skc
->skc_partial_list
)) {
1117 spin_unlock(&skc
->skc_lock
);
1120 rc
= spl_cache_grow(skc
, flags
, &obj
);
1121 local_irq_disable();
1123 /* Emergency object for immediate use by caller */
1124 if (rc
== 0 && obj
!= NULL
)
1130 /* Rescheduled to different CPU skm is not local */
1131 if (skm
!= skc
->skc_mag
[smp_processor_id()])
1135 * Potentially rescheduled to the same CPU but
1136 * allocations may have occurred from this CPU while
1137 * we were sleeping so recalculate max refill.
1139 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
1141 spin_lock(&skc
->skc_lock
);
1145 /* Grab the next available slab */
1146 sks
= list_entry((&skc
->skc_partial_list
)->next
,
1147 spl_kmem_slab_t
, sks_list
);
1148 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1149 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
1150 ASSERT(!list_empty(&sks
->sks_free_list
));
1153 * Consume as many objects as needed to refill the requested
1154 * cache. We must also be careful not to overfill it.
1156 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 &&
1158 ASSERT(skm
->skm_avail
< skm
->skm_size
);
1159 ASSERT(count
< skm
->skm_size
);
1160 skm
->skm_objs
[skm
->skm_avail
++] =
1161 spl_cache_obj(skc
, sks
);
1164 /* Move slab to skc_complete_list when full */
1165 if (sks
->sks_ref
== sks
->sks_objs
) {
1166 list_del(&sks
->sks_list
);
1167 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
1171 spin_unlock(&skc
->skc_lock
);
1177 * Release an object back to the slab from which it came.
1180 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
1182 spl_kmem_slab_t
*sks
= NULL
;
1183 spl_kmem_obj_t
*sko
= NULL
;
1185 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1187 sko
= spl_sko_from_obj(skc
, obj
);
1188 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1189 sks
= sko
->sko_slab
;
1190 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1191 ASSERT(sks
->sks_cache
== skc
);
1192 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
1194 sks
->sks_age
= jiffies
;
1196 skc
->skc_obj_alloc
--;
1199 * Move slab to skc_partial_list when no longer full. Slabs
1200 * are added to the head to keep the partial list is quasi-full
1201 * sorted order. Fuller at the head, emptier at the tail.
1203 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
1204 list_del(&sks
->sks_list
);
1205 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
1209 * Move empty slabs to the end of the partial list so
1210 * they can be easily found and freed during reclamation.
1212 if (sks
->sks_ref
== 0) {
1213 list_del(&sks
->sks_list
);
1214 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1215 skc
->skc_slab_alloc
--;
1220 * Allocate an object from the per-cpu magazine, or if the magazine
1221 * is empty directly allocate from a slab and repopulate the magazine.
1224 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
1226 spl_kmem_magazine_t
*skm
;
1229 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1230 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1231 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1234 * Allocate directly from a Linux slab. All optimizations are left
1235 * to the underlying cache we only need to guarantee that KM_SLEEP
1236 * callers will never fail.
1238 if (skc
->skc_flags
& KMC_SLAB
) {
1239 struct kmem_cache
*slc
= skc
->skc_linux_cache
;
1241 obj
= kmem_cache_alloc(slc
, kmem_flags_convert(flags
));
1242 } while ((obj
== NULL
) && !(flags
& KM_NOSLEEP
));
1246 * Even though we leave everything up to the
1247 * underlying cache we still keep track of
1248 * how many objects we've allocated in it for
1249 * better debuggability.
1251 percpu_counter_inc(&skc
->skc_linux_alloc
);
1256 local_irq_disable();
1260 * Safe to update per-cpu structure without lock, but
1261 * in the restart case we must be careful to reacquire
1262 * the local magazine since this may have changed
1263 * when we need to grow the cache.
1265 skm
= skc
->skc_mag
[smp_processor_id()];
1266 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1268 if (likely(skm
->skm_avail
)) {
1269 /* Object available in CPU cache, use it */
1270 obj
= skm
->skm_objs
[--skm
->skm_avail
];
1272 obj
= spl_cache_refill(skc
, skm
, flags
);
1273 if ((obj
== NULL
) && !(flags
& KM_NOSLEEP
))
1282 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
1285 /* Pre-emptively migrate object to CPU L1 cache */
1287 if (obj
&& skc
->skc_ctor
)
1288 skc
->skc_ctor(obj
, skc
->skc_private
, flags
);
1295 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
1298 * Free an object back to the local per-cpu magazine, there is no
1299 * guarantee that this is the same magazine the object was originally
1300 * allocated from. We may need to flush entire from the magazine
1301 * back to the slabs to make space.
1304 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
1306 spl_kmem_magazine_t
*skm
;
1307 unsigned long flags
;
1309 int do_emergency
= 0;
1311 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1312 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1315 * Run the destructor
1318 skc
->skc_dtor(obj
, skc
->skc_private
);
1321 * Free the object from the Linux underlying Linux slab.
1323 if (skc
->skc_flags
& KMC_SLAB
) {
1324 kmem_cache_free(skc
->skc_linux_cache
, obj
);
1325 percpu_counter_dec(&skc
->skc_linux_alloc
);
1330 * While a cache has outstanding emergency objects all freed objects
1331 * must be checked. However, since emergency objects will never use
1332 * a virtual address these objects can be safely excluded as an
1335 if (!is_vmalloc_addr(obj
)) {
1336 spin_lock(&skc
->skc_lock
);
1337 do_emergency
= (skc
->skc_obj_emergency
> 0);
1338 spin_unlock(&skc
->skc_lock
);
1340 if (do_emergency
&& (spl_emergency_free(skc
, obj
) == 0))
1344 local_irq_save(flags
);
1347 * Safe to update per-cpu structure without lock, but
1348 * no remote memory allocation tracking is being performed
1349 * it is entirely possible to allocate an object from one
1350 * CPU cache and return it to another.
1352 skm
= skc
->skc_mag
[smp_processor_id()];
1353 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1356 * Per-CPU cache full, flush it to make space for this object,
1357 * this may result in an empty slab which can be reclaimed once
1358 * interrupts are re-enabled.
1360 if (unlikely(skm
->skm_avail
>= skm
->skm_size
)) {
1361 spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1365 /* Available space in cache, use it */
1366 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
1368 local_irq_restore(flags
);
1371 spl_slab_reclaim(skc
);
1373 EXPORT_SYMBOL(spl_kmem_cache_free
);
1376 * Depending on how many and which objects are released it may simply
1377 * repopulate the local magazine which will then need to age-out. Objects
1378 * which cannot fit in the magazine will be released back to their slabs
1379 * which will also need to age out before being released. This is all just
1380 * best effort and we do not want to thrash creating and destroying slabs.
1383 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
)
1385 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1386 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1388 if (skc
->skc_flags
& KMC_SLAB
)
1391 atomic_inc(&skc
->skc_ref
);
1394 * Prevent concurrent cache reaping when contended.
1396 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
))
1399 /* Reclaim from the magazine and free all now empty slabs. */
1400 unsigned long irq_flags
;
1401 local_irq_save(irq_flags
);
1402 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
1403 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1404 local_irq_restore(irq_flags
);
1406 spl_slab_reclaim(skc
);
1407 clear_bit_unlock(KMC_BIT_REAPING
, &skc
->skc_flags
);
1408 smp_mb__after_atomic();
1409 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
1411 atomic_dec(&skc
->skc_ref
);
1413 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
1416 * This is stubbed out for code consistency with other platforms. There
1417 * is existing logic to prevent concurrent reaping so while this is ugly
1418 * it should do no harm.
1421 spl_kmem_cache_reap_active(void)
1425 EXPORT_SYMBOL(spl_kmem_cache_reap_active
);
1428 * Reap all free slabs from all registered caches.
1433 spl_kmem_cache_t
*skc
= NULL
;
1435 down_read(&spl_kmem_cache_sem
);
1436 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
1437 spl_kmem_cache_reap_now(skc
);
1439 up_read(&spl_kmem_cache_sem
);
1441 EXPORT_SYMBOL(spl_kmem_reap
);
1444 spl_kmem_cache_init(void)
1446 init_rwsem(&spl_kmem_cache_sem
);
1447 INIT_LIST_HEAD(&spl_kmem_cache_list
);
1448 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
1449 spl_kmem_cache_kmem_threads
, maxclsyspri
,
1450 spl_kmem_cache_kmem_threads
* 8, INT_MAX
,
1451 TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
1453 if (spl_kmem_cache_taskq
== NULL
)
1460 spl_kmem_cache_fini(void)
1462 taskq_destroy(spl_kmem_cache_taskq
);