2 * Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3 * Copyright (C) 2007 The Regents of the University of California.
4 * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5 * Written by Brian Behlendorf <behlendorf1@llnl.gov>.
8 * This file is part of the SPL, Solaris Porting Layer.
9 * For details, see <http://zfsonlinux.org/>.
11 * The SPL is free software; you can redistribute it and/or modify it
12 * under the terms of the GNU General Public License as published by the
13 * Free Software Foundation; either version 2 of the License, or (at your
14 * option) any later version.
16 * The SPL is distributed in the hope that it will be useful, but WITHOUT
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
21 * You should have received a copy of the GNU General Public License along
22 * with the SPL. If not, see <http://www.gnu.org/licenses/>.
26 #include <sys/kmem_cache.h>
27 #include <sys/taskq.h>
28 #include <sys/timer.h>
30 #include <linux/slab.h>
31 #include <linux/swap.h>
32 #include <linux/mm_compat.h>
33 #include <linux/wait_compat.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)
60 * Cache expiration was implemented because it was part of the default Solaris
61 * kmem_cache behavior. The idea is that per-cpu objects which haven't been
62 * accessed in several seconds should be returned to the cache. On the other
63 * hand Linux slabs never move objects back to the slabs unless there is
64 * memory pressure on the system. By default the Linux method is enabled
65 * because it has been shown to improve responsiveness on low memory systems.
66 * This policy may be changed by setting KMC_EXPIRE_AGE or KMC_EXPIRE_MEM.
68 unsigned int spl_kmem_cache_expire
= KMC_EXPIRE_MEM
;
69 EXPORT_SYMBOL(spl_kmem_cache_expire
);
70 module_param(spl_kmem_cache_expire
, uint
, 0644);
71 MODULE_PARM_DESC(spl_kmem_cache_expire
, "By age (0x1) or low memory (0x2)");
74 * Cache magazines are an optimization designed to minimize the cost of
75 * allocating memory. They do this by keeping a per-cpu cache of recently
76 * freed objects, which can then be reallocated without taking a lock. This
77 * can improve performance on highly contended caches. However, because
78 * objects in magazines will prevent otherwise empty slabs from being
79 * immediately released this may not be ideal for low memory machines.
81 * For this reason spl_kmem_cache_magazine_size can be used to set a maximum
82 * magazine size. When this value is set to 0 the magazine size will be
83 * automatically determined based on the object size. Otherwise magazines
84 * will be limited to 2-256 objects per magazine (i.e per cpu). Magazines
85 * may never be entirely disabled in this implementation.
87 unsigned int spl_kmem_cache_magazine_size
= 0;
88 module_param(spl_kmem_cache_magazine_size
, uint
, 0444);
89 MODULE_PARM_DESC(spl_kmem_cache_magazine_size
,
90 "Default magazine size (2-256), set automatically (0)\n");
93 * The default behavior is to report the number of objects remaining in the
94 * cache. This allows the Linux VM to repeatedly reclaim objects from the
95 * cache when memory is low satisfy other memory allocations. Alternately,
96 * setting this value to KMC_RECLAIM_ONCE limits how aggressively the cache
97 * is reclaimed. This may increase the likelihood of out of memory events.
99 unsigned int spl_kmem_cache_reclaim
= 0 /* KMC_RECLAIM_ONCE */;
100 module_param(spl_kmem_cache_reclaim
, uint
, 0644);
101 MODULE_PARM_DESC(spl_kmem_cache_reclaim
, "Single reclaim pass (0x1)");
103 unsigned int spl_kmem_cache_obj_per_slab
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
104 module_param(spl_kmem_cache_obj_per_slab
, uint
, 0644);
105 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab
, "Number of objects per slab");
107 unsigned int spl_kmem_cache_obj_per_slab_min
= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN
;
108 module_param(spl_kmem_cache_obj_per_slab_min
, uint
, 0644);
109 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab_min
,
110 "Minimal number of objects per slab");
112 unsigned int spl_kmem_cache_max_size
= SPL_KMEM_CACHE_MAX_SIZE
;
113 module_param(spl_kmem_cache_max_size
, uint
, 0644);
114 MODULE_PARM_DESC(spl_kmem_cache_max_size
, "Maximum size of slab in MB");
117 * For small objects the Linux slab allocator should be used to make the most
118 * efficient use of the memory. However, large objects are not supported by
119 * the Linux slab and therefore the SPL implementation is preferred. A cutoff
120 * of 16K was determined to be optimal for architectures using 4K pages.
122 #if PAGE_SIZE == 4096
123 unsigned int spl_kmem_cache_slab_limit
= 16384;
125 unsigned int spl_kmem_cache_slab_limit
= 0;
127 module_param(spl_kmem_cache_slab_limit
, uint
, 0644);
128 MODULE_PARM_DESC(spl_kmem_cache_slab_limit
,
129 "Objects less than N bytes use the Linux slab");
132 * This value defaults to a threshold designed to avoid allocations which
133 * have been deemed costly by the kernel.
135 unsigned int spl_kmem_cache_kmem_limit
=
136 ((1 << (PAGE_ALLOC_COSTLY_ORDER
- 1)) * PAGE_SIZE
) /
137 SPL_KMEM_CACHE_OBJ_PER_SLAB
;
138 module_param(spl_kmem_cache_kmem_limit
, uint
, 0644);
139 MODULE_PARM_DESC(spl_kmem_cache_kmem_limit
,
140 "Objects less than N bytes use the kmalloc");
143 * The number of threads available to allocate new slabs for caches. This
144 * should not need to be tuned but it is available for performance analysis.
146 unsigned int spl_kmem_cache_kmem_threads
= 4;
147 module_param(spl_kmem_cache_kmem_threads
, uint
, 0444);
148 MODULE_PARM_DESC(spl_kmem_cache_kmem_threads
,
149 "Number of spl_kmem_cache threads");
152 * Slab allocation interfaces
154 * While the Linux slab implementation was inspired by the Solaris
155 * implementation I cannot use it to emulate the Solaris APIs. I
156 * require two features which are not provided by the Linux slab.
158 * 1) Constructors AND destructors. Recent versions of the Linux
159 * kernel have removed support for destructors. This is a deal
160 * breaker for the SPL which contains particularly expensive
161 * initializers for mutex's, condition variables, etc. We also
162 * require a minimal level of cleanup for these data types unlike
163 * many Linux data types which do need to be explicitly destroyed.
165 * 2) Virtual address space backed slab. Callers of the Solaris slab
166 * expect it to work well for both small are very large allocations.
167 * Because of memory fragmentation the Linux slab which is backed
168 * by kmalloc'ed memory performs very badly when confronted with
169 * large numbers of large allocations. Basing the slab on the
170 * virtual address space removes the need for contiguous pages
171 * and greatly improve performance for large allocations.
173 * For these reasons, the SPL has its own slab implementation with
174 * the needed features. It is not as highly optimized as either the
175 * Solaris or Linux slabs, but it should get me most of what is
176 * needed until it can be optimized or obsoleted by another approach.
178 * One serious concern I do have about this method is the relatively
179 * small virtual address space on 32bit arches. This will seriously
180 * constrain the size of the slab caches and their performance.
183 struct list_head spl_kmem_cache_list
; /* List of caches */
184 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
185 taskq_t
*spl_kmem_cache_taskq
; /* Task queue for ageing / reclaim */
187 static void spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
);
189 SPL_SHRINKER_CALLBACK_FWD_DECLARE(spl_kmem_cache_generic_shrinker
);
190 SPL_SHRINKER_DECLARE(spl_kmem_cache_shrinker
,
191 spl_kmem_cache_generic_shrinker
, KMC_DEFAULT_SEEKS
);
194 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
196 gfp_t lflags
= kmem_flags_convert(flags
);
199 if (skc
->skc_flags
& KMC_KMEM
) {
201 ptr
= (void *)__get_free_pages(lflags
, get_order(size
));
203 ptr
= spl_vmalloc(size
, lflags
| __GFP_HIGHMEM
, PAGE_KERNEL
);
206 /* Resulting allocated memory will be page aligned */
207 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
213 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
215 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
218 * The Linux direct reclaim path uses this out of band value to
219 * determine if forward progress is being made. Normally this is
220 * incremented by kmem_freepages() which is part of the various
221 * Linux slab implementations. However, since we are using none
222 * of that infrastructure we are responsible for incrementing it.
224 if (current
->reclaim_state
)
225 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
227 if (skc
->skc_flags
& KMC_KMEM
) {
229 free_pages((unsigned long)ptr
, get_order(size
));
236 * Required space for each aligned sks.
238 static inline uint32_t
239 spl_sks_size(spl_kmem_cache_t
*skc
)
241 return (P2ROUNDUP_TYPED(sizeof (spl_kmem_slab_t
),
242 skc
->skc_obj_align
, uint32_t));
246 * Required space for each aligned object.
248 static inline uint32_t
249 spl_obj_size(spl_kmem_cache_t
*skc
)
251 uint32_t align
= skc
->skc_obj_align
;
253 return (P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
254 P2ROUNDUP_TYPED(sizeof (spl_kmem_obj_t
), align
, uint32_t));
258 * Lookup the spl_kmem_object_t for an object given that object.
260 static inline spl_kmem_obj_t
*
261 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
263 return (obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
264 skc
->skc_obj_align
, uint32_t));
268 * Required space for each offslab object taking in to account alignment
269 * restrictions and the power-of-two requirement of kv_alloc().
271 static inline uint32_t
272 spl_offslab_size(spl_kmem_cache_t
*skc
)
274 return (1UL << (fls64(spl_obj_size(skc
)) + 1));
278 * It's important that we pack the spl_kmem_obj_t structure and the
279 * actual objects in to one large address space to minimize the number
280 * of calls to the allocator. It is far better to do a few large
281 * allocations and then subdivide it ourselves. Now which allocator
282 * we use requires balancing a few trade offs.
284 * For small objects we use kmem_alloc() because as long as you are
285 * only requesting a small number of pages (ideally just one) its cheap.
286 * However, when you start requesting multiple pages with kmem_alloc()
287 * it gets increasingly expensive since it requires contiguous pages.
288 * For this reason we shift to vmem_alloc() for slabs of large objects
289 * which removes the need for contiguous pages. We do not use
290 * vmem_alloc() in all cases because there is significant locking
291 * overhead in __get_vm_area_node(). This function takes a single
292 * global lock when acquiring an available virtual address range which
293 * serializes all vmem_alloc()'s for all slab caches. Using slightly
294 * different allocation functions for small and large objects should
295 * give us the best of both worlds.
297 * KMC_ONSLAB KMC_OFFSLAB
299 * +------------------------+ +-----------------+
300 * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+
301 * | skc_obj_size <-+ | | +-----------------+ | |
302 * | spl_kmem_obj_t | | | |
303 * | skc_obj_size <---+ | +-----------------+ | |
304 * | spl_kmem_obj_t | | | skc_obj_size | <-+ |
305 * | ... v | | spl_kmem_obj_t | |
306 * +------------------------+ +-----------------+ v
308 static spl_kmem_slab_t
*
309 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
311 spl_kmem_slab_t
*sks
;
312 spl_kmem_obj_t
*sko
, *n
;
314 uint32_t obj_size
, offslab_size
= 0;
317 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
321 sks
= (spl_kmem_slab_t
*)base
;
322 sks
->sks_magic
= SKS_MAGIC
;
323 sks
->sks_objs
= skc
->skc_slab_objs
;
324 sks
->sks_age
= jiffies
;
325 sks
->sks_cache
= skc
;
326 INIT_LIST_HEAD(&sks
->sks_list
);
327 INIT_LIST_HEAD(&sks
->sks_free_list
);
329 obj_size
= spl_obj_size(skc
);
331 if (skc
->skc_flags
& KMC_OFFSLAB
)
332 offslab_size
= spl_offslab_size(skc
);
334 for (i
= 0; i
< sks
->sks_objs
; i
++) {
335 if (skc
->skc_flags
& KMC_OFFSLAB
) {
336 obj
= kv_alloc(skc
, offslab_size
, flags
);
342 obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
345 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
346 sko
= spl_sko_from_obj(skc
, obj
);
348 sko
->sko_magic
= SKO_MAGIC
;
350 INIT_LIST_HEAD(&sko
->sko_list
);
351 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
356 if (skc
->skc_flags
& KMC_OFFSLAB
)
357 list_for_each_entry_safe(sko
,
358 n
, &sks
->sks_free_list
, sko_list
)
359 kv_free(skc
, sko
->sko_addr
, offslab_size
);
361 kv_free(skc
, base
, skc
->skc_slab_size
);
369 * Remove a slab from complete or partial list, it must be called with
370 * the 'skc->skc_lock' held but the actual free must be performed
371 * outside the lock to prevent deadlocking on vmem addresses.
374 spl_slab_free(spl_kmem_slab_t
*sks
,
375 struct list_head
*sks_list
, struct list_head
*sko_list
)
377 spl_kmem_cache_t
*skc
;
379 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
380 ASSERT(sks
->sks_ref
== 0);
382 skc
= sks
->sks_cache
;
383 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
384 ASSERT(spin_is_locked(&skc
->skc_lock
));
387 * Update slab/objects counters in the cache, then remove the
388 * slab from the skc->skc_partial_list. Finally add the slab
389 * and all its objects in to the private work lists where the
390 * destructors will be called and the memory freed to the system.
392 skc
->skc_obj_total
-= sks
->sks_objs
;
393 skc
->skc_slab_total
--;
394 list_del(&sks
->sks_list
);
395 list_add(&sks
->sks_list
, sks_list
);
396 list_splice_init(&sks
->sks_free_list
, sko_list
);
400 * Reclaim empty slabs at the end of the partial list.
403 spl_slab_reclaim(spl_kmem_cache_t
*skc
)
405 spl_kmem_slab_t
*sks
, *m
;
406 spl_kmem_obj_t
*sko
, *n
;
412 * Empty slabs and objects must be moved to a private list so they
413 * can be safely freed outside the spin lock. All empty slabs are
414 * at the end of skc->skc_partial_list, therefore once a non-empty
415 * slab is found we can stop scanning.
417 spin_lock(&skc
->skc_lock
);
418 list_for_each_entry_safe_reverse(sks
, m
,
419 &skc
->skc_partial_list
, sks_list
) {
421 if (sks
->sks_ref
> 0)
424 spl_slab_free(sks
, &sks_list
, &sko_list
);
426 spin_unlock(&skc
->skc_lock
);
429 * The following two loops ensure all the object destructors are
430 * run, any offslab objects are freed, and the slabs themselves
431 * are freed. This is all done outside the skc->skc_lock since
432 * this allows the destructor to sleep, and allows us to perform
433 * a conditional reschedule when a freeing a large number of
434 * objects and slabs back to the system.
436 if (skc
->skc_flags
& KMC_OFFSLAB
)
437 size
= spl_offslab_size(skc
);
439 list_for_each_entry_safe(sko
, n
, &sko_list
, sko_list
) {
440 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
442 if (skc
->skc_flags
& KMC_OFFSLAB
)
443 kv_free(skc
, sko
->sko_addr
, size
);
446 list_for_each_entry_safe(sks
, m
, &sks_list
, sks_list
) {
447 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
448 kv_free(skc
, sks
, skc
->skc_slab_size
);
452 static spl_kmem_emergency_t
*
453 spl_emergency_search(struct rb_root
*root
, void *obj
)
455 struct rb_node
*node
= root
->rb_node
;
456 spl_kmem_emergency_t
*ske
;
457 unsigned long address
= (unsigned long)obj
;
460 ske
= container_of(node
, spl_kmem_emergency_t
, ske_node
);
462 if (address
< ske
->ske_obj
)
463 node
= node
->rb_left
;
464 else if (address
> ske
->ske_obj
)
465 node
= node
->rb_right
;
474 spl_emergency_insert(struct rb_root
*root
, spl_kmem_emergency_t
*ske
)
476 struct rb_node
**new = &(root
->rb_node
), *parent
= NULL
;
477 spl_kmem_emergency_t
*ske_tmp
;
478 unsigned long address
= ske
->ske_obj
;
481 ske_tmp
= container_of(*new, spl_kmem_emergency_t
, ske_node
);
484 if (address
< ske_tmp
->ske_obj
)
485 new = &((*new)->rb_left
);
486 else if (address
> ske_tmp
->ske_obj
)
487 new = &((*new)->rb_right
);
492 rb_link_node(&ske
->ske_node
, parent
, new);
493 rb_insert_color(&ske
->ske_node
, root
);
499 * Allocate a single emergency object and track it in a red black tree.
502 spl_emergency_alloc(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
504 gfp_t lflags
= kmem_flags_convert(flags
);
505 spl_kmem_emergency_t
*ske
;
506 int order
= get_order(skc
->skc_obj_size
);
509 /* Last chance use a partial slab if one now exists */
510 spin_lock(&skc
->skc_lock
);
511 empty
= list_empty(&skc
->skc_partial_list
);
512 spin_unlock(&skc
->skc_lock
);
516 ske
= kmalloc(sizeof (*ske
), lflags
);
520 ske
->ske_obj
= __get_free_pages(lflags
, order
);
521 if (ske
->ske_obj
== 0) {
526 spin_lock(&skc
->skc_lock
);
527 empty
= spl_emergency_insert(&skc
->skc_emergency_tree
, ske
);
529 skc
->skc_obj_total
++;
530 skc
->skc_obj_emergency
++;
531 if (skc
->skc_obj_emergency
> skc
->skc_obj_emergency_max
)
532 skc
->skc_obj_emergency_max
= skc
->skc_obj_emergency
;
534 spin_unlock(&skc
->skc_lock
);
536 if (unlikely(!empty
)) {
537 free_pages(ske
->ske_obj
, order
);
542 *obj
= (void *)ske
->ske_obj
;
548 * Locate the passed object in the red black tree and free it.
551 spl_emergency_free(spl_kmem_cache_t
*skc
, void *obj
)
553 spl_kmem_emergency_t
*ske
;
554 int order
= get_order(skc
->skc_obj_size
);
556 spin_lock(&skc
->skc_lock
);
557 ske
= spl_emergency_search(&skc
->skc_emergency_tree
, obj
);
559 rb_erase(&ske
->ske_node
, &skc
->skc_emergency_tree
);
560 skc
->skc_obj_emergency
--;
561 skc
->skc_obj_total
--;
563 spin_unlock(&skc
->skc_lock
);
568 free_pages(ske
->ske_obj
, order
);
575 * Release objects from the per-cpu magazine back to their slab. The flush
576 * argument contains the max number of entries to remove from the magazine.
579 __spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
581 int i
, count
= MIN(flush
, skm
->skm_avail
);
583 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
584 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
585 ASSERT(spin_is_locked(&skc
->skc_lock
));
587 for (i
= 0; i
< count
; i
++)
588 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
590 skm
->skm_avail
-= count
;
591 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
592 sizeof (void *) * skm
->skm_avail
);
596 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
598 spin_lock(&skc
->skc_lock
);
599 __spl_cache_flush(skc
, skm
, flush
);
600 spin_unlock(&skc
->skc_lock
);
604 spl_magazine_age(void *data
)
606 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
607 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
609 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
610 ASSERT(skm
->skm_cpu
== smp_processor_id());
611 ASSERT(irqs_disabled());
613 /* There are no available objects or they are too young to age out */
614 if ((skm
->skm_avail
== 0) ||
615 time_before(jiffies
, skm
->skm_age
+ skc
->skc_delay
* HZ
))
619 * Because we're executing in interrupt context we may have
620 * interrupted the holder of this lock. To avoid a potential
621 * deadlock return if the lock is contended.
623 if (!spin_trylock(&skc
->skc_lock
))
626 __spl_cache_flush(skc
, skm
, skm
->skm_refill
);
627 spin_unlock(&skc
->skc_lock
);
631 * Called regularly to keep a downward pressure on the cache.
633 * Objects older than skc->skc_delay seconds in the per-cpu magazines will
634 * be returned to the caches. This is done to prevent idle magazines from
635 * holding memory which could be better used elsewhere. The delay is
636 * present to prevent thrashing the magazine.
638 * The newly released objects may result in empty partial slabs. Those
639 * slabs should be released to the system. Otherwise moving the objects
640 * out of the magazines is just wasted work.
643 spl_cache_age(void *data
)
645 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
648 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
650 /* Dynamically disabled at run time */
651 if (!(spl_kmem_cache_expire
& KMC_EXPIRE_AGE
))
654 atomic_inc(&skc
->skc_ref
);
656 if (!(skc
->skc_flags
& KMC_NOMAGAZINE
))
657 on_each_cpu(spl_magazine_age
, skc
, 1);
659 spl_slab_reclaim(skc
);
661 while (!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && !id
) {
662 id
= taskq_dispatch_delay(
663 spl_kmem_cache_taskq
, spl_cache_age
, skc
, TQ_SLEEP
,
664 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
666 /* Destroy issued after dispatch immediately cancel it */
667 if (test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && id
)
668 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
671 spin_lock(&skc
->skc_lock
);
672 skc
->skc_taskqid
= id
;
673 spin_unlock(&skc
->skc_lock
);
675 atomic_dec(&skc
->skc_ref
);
679 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
680 * When on-slab we want to target spl_kmem_cache_obj_per_slab. However,
681 * for very small objects we may end up with more than this so as not
682 * to waste space in the minimal allocation of a single page. Also for
683 * very large objects we may use as few as spl_kmem_cache_obj_per_slab_min,
684 * lower than this and we will fail.
687 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
689 uint32_t sks_size
, obj_size
, max_size
, tgt_size
, tgt_objs
;
691 if (skc
->skc_flags
& KMC_OFFSLAB
) {
692 tgt_objs
= spl_kmem_cache_obj_per_slab
;
693 tgt_size
= P2ROUNDUP(sizeof (spl_kmem_slab_t
), PAGE_SIZE
);
695 if ((skc
->skc_flags
& KMC_KMEM
) &&
696 (spl_obj_size(skc
) > (SPL_MAX_ORDER_NR_PAGES
* PAGE_SIZE
)))
699 sks_size
= spl_sks_size(skc
);
700 obj_size
= spl_obj_size(skc
);
701 max_size
= (spl_kmem_cache_max_size
* 1024 * 1024);
702 tgt_size
= (spl_kmem_cache_obj_per_slab
* obj_size
+ sks_size
);
705 * KMC_KMEM slabs are allocated by __get_free_pages() which
706 * rounds up to the nearest order. Knowing this the size
707 * should be rounded up to the next power of two with a hard
708 * maximum defined by the maximum allowed allocation order.
710 if (skc
->skc_flags
& KMC_KMEM
) {
711 max_size
= SPL_MAX_ORDER_NR_PAGES
* PAGE_SIZE
;
712 tgt_size
= MIN(max_size
,
713 PAGE_SIZE
* (1 << MAX(get_order(tgt_size
) - 1, 1)));
716 if (tgt_size
<= max_size
) {
717 tgt_objs
= (tgt_size
- sks_size
) / obj_size
;
719 tgt_objs
= (max_size
- sks_size
) / obj_size
;
720 tgt_size
= (tgt_objs
* obj_size
) + sks_size
;
734 * Make a guess at reasonable per-cpu magazine size based on the size of
735 * each object and the cost of caching N of them in each magazine. Long
736 * term this should really adapt based on an observed usage heuristic.
739 spl_magazine_size(spl_kmem_cache_t
*skc
)
741 uint32_t obj_size
= spl_obj_size(skc
);
744 if (spl_kmem_cache_magazine_size
> 0)
745 return (MAX(MIN(spl_kmem_cache_magazine_size
, 256), 2));
747 /* Per-magazine sizes below assume a 4Kib page size */
748 if (obj_size
> (PAGE_SIZE
* 256))
749 size
= 4; /* Minimum 4Mib per-magazine */
750 else if (obj_size
> (PAGE_SIZE
* 32))
751 size
= 16; /* Minimum 2Mib per-magazine */
752 else if (obj_size
> (PAGE_SIZE
))
753 size
= 64; /* Minimum 256Kib per-magazine */
754 else if (obj_size
> (PAGE_SIZE
/ 4))
755 size
= 128; /* Minimum 128Kib per-magazine */
763 * Allocate a per-cpu magazine to associate with a specific core.
765 static spl_kmem_magazine_t
*
766 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
768 spl_kmem_magazine_t
*skm
;
769 int size
= sizeof (spl_kmem_magazine_t
) +
770 sizeof (void *) * skc
->skc_mag_size
;
772 skm
= kmalloc_node(size
, GFP_KERNEL
, cpu_to_node(cpu
));
774 skm
->skm_magic
= SKM_MAGIC
;
776 skm
->skm_size
= skc
->skc_mag_size
;
777 skm
->skm_refill
= skc
->skc_mag_refill
;
778 skm
->skm_cache
= skc
;
779 skm
->skm_age
= jiffies
;
787 * Free a per-cpu magazine associated with a specific core.
790 spl_magazine_free(spl_kmem_magazine_t
*skm
)
792 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
793 ASSERT(skm
->skm_avail
== 0);
798 * Create all pre-cpu magazines of reasonable sizes.
801 spl_magazine_create(spl_kmem_cache_t
*skc
)
805 if (skc
->skc_flags
& KMC_NOMAGAZINE
)
808 skc
->skc_mag
= kzalloc(sizeof (spl_kmem_magazine_t
*) *
809 num_possible_cpus(), kmem_flags_convert(KM_SLEEP
));
810 skc
->skc_mag_size
= spl_magazine_size(skc
);
811 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
813 for_each_possible_cpu(i
) {
814 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
815 if (!skc
->skc_mag
[i
]) {
816 for (i
--; i
>= 0; i
--)
817 spl_magazine_free(skc
->skc_mag
[i
]);
828 * Destroy all pre-cpu magazines.
831 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
833 spl_kmem_magazine_t
*skm
;
836 if (skc
->skc_flags
& KMC_NOMAGAZINE
)
839 for_each_possible_cpu(i
) {
840 skm
= skc
->skc_mag
[i
];
841 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
842 spl_magazine_free(skm
);
849 * Create a object cache based on the following arguments:
851 * size cache object size
852 * align cache object alignment
853 * ctor cache object constructor
854 * dtor cache object destructor
855 * reclaim cache object reclaim
856 * priv cache private data for ctor/dtor/reclaim
857 * vmp unused must be NULL
859 * KMC_NOTOUCH Disable cache object aging (unsupported)
860 * KMC_NODEBUG Disable debugging (unsupported)
861 * KMC_NOHASH Disable hashing (unsupported)
862 * KMC_QCACHE Disable qcache (unsupported)
863 * KMC_NOMAGAZINE Enabled for kmem/vmem, Disabled for Linux slab
864 * KMC_KMEM Force kmem backed cache
865 * KMC_VMEM Force vmem backed cache
866 * KMC_SLAB Force Linux slab backed cache
867 * KMC_OFFSLAB Locate objects off the slab
870 spl_kmem_cache_create(char *name
, size_t size
, size_t align
,
871 spl_kmem_ctor_t ctor
, spl_kmem_dtor_t dtor
, spl_kmem_reclaim_t reclaim
,
872 void *priv
, void *vmp
, int flags
)
874 gfp_t lflags
= kmem_flags_convert(KM_SLEEP
);
875 spl_kmem_cache_t
*skc
;
881 ASSERT0(flags
& KMC_NOMAGAZINE
);
882 ASSERT0(flags
& KMC_NOHASH
);
883 ASSERT0(flags
& KMC_QCACHE
);
888 skc
= kzalloc(sizeof (*skc
), lflags
);
892 skc
->skc_magic
= SKC_MAGIC
;
893 skc
->skc_name_size
= strlen(name
) + 1;
894 skc
->skc_name
= (char *)kmalloc(skc
->skc_name_size
, lflags
);
895 if (skc
->skc_name
== NULL
) {
899 strncpy(skc
->skc_name
, name
, skc
->skc_name_size
);
901 skc
->skc_ctor
= ctor
;
902 skc
->skc_dtor
= dtor
;
903 skc
->skc_reclaim
= reclaim
;
904 skc
->skc_private
= priv
;
906 skc
->skc_linux_cache
= NULL
;
907 skc
->skc_flags
= flags
;
908 skc
->skc_obj_size
= size
;
909 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
910 skc
->skc_delay
= SPL_KMEM_CACHE_DELAY
;
911 skc
->skc_reap
= SPL_KMEM_CACHE_REAP
;
912 atomic_set(&skc
->skc_ref
, 0);
914 INIT_LIST_HEAD(&skc
->skc_list
);
915 INIT_LIST_HEAD(&skc
->skc_complete_list
);
916 INIT_LIST_HEAD(&skc
->skc_partial_list
);
917 skc
->skc_emergency_tree
= RB_ROOT
;
918 spin_lock_init(&skc
->skc_lock
);
919 init_waitqueue_head(&skc
->skc_waitq
);
920 skc
->skc_slab_fail
= 0;
921 skc
->skc_slab_create
= 0;
922 skc
->skc_slab_destroy
= 0;
923 skc
->skc_slab_total
= 0;
924 skc
->skc_slab_alloc
= 0;
925 skc
->skc_slab_max
= 0;
926 skc
->skc_obj_total
= 0;
927 skc
->skc_obj_alloc
= 0;
928 skc
->skc_obj_max
= 0;
929 skc
->skc_obj_deadlock
= 0;
930 skc
->skc_obj_emergency
= 0;
931 skc
->skc_obj_emergency_max
= 0;
934 * Verify the requested alignment restriction is sane.
938 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
);
939 VERIFY3U(align
, <=, PAGE_SIZE
);
940 skc
->skc_obj_align
= align
;
944 * When no specific type of slab is requested (kmem, vmem, or
945 * linuxslab) then select a cache type based on the object size
946 * and default tunables.
948 if (!(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
))) {
951 * Objects smaller than spl_kmem_cache_slab_limit can
952 * use the Linux slab for better space-efficiency. By
953 * default this functionality is disabled until its
954 * performance characteristics are fully understood.
956 if (spl_kmem_cache_slab_limit
&&
957 size
<= (size_t)spl_kmem_cache_slab_limit
)
958 skc
->skc_flags
|= KMC_SLAB
;
961 * Small objects, less than spl_kmem_cache_kmem_limit per
962 * object should use kmem because their slabs are small.
964 else if (spl_obj_size(skc
) <= spl_kmem_cache_kmem_limit
)
965 skc
->skc_flags
|= KMC_KMEM
;
968 * All other objects are considered large and are placed
969 * on vmem backed slabs.
972 skc
->skc_flags
|= KMC_VMEM
;
976 * Given the type of slab allocate the required resources.
978 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
979 rc
= spl_slab_size(skc
,
980 &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
984 rc
= spl_magazine_create(skc
);
988 unsigned long slabflags
= 0;
990 if (size
> (SPL_MAX_KMEM_ORDER_NR_PAGES
* PAGE_SIZE
)) {
995 #if defined(SLAB_USERCOPY)
997 * Required for PAX-enabled kernels if the slab is to be
998 * used for coping between user and kernel space.
1000 slabflags
|= SLAB_USERCOPY
;
1003 skc
->skc_linux_cache
= kmem_cache_create(
1004 skc
->skc_name
, size
, align
, slabflags
, NULL
);
1005 if (skc
->skc_linux_cache
== NULL
) {
1010 #if defined(HAVE_KMEM_CACHE_ALLOCFLAGS)
1011 skc
->skc_linux_cache
->allocflags
|= __GFP_COMP
;
1012 #elif defined(HAVE_KMEM_CACHE_GFPFLAGS)
1013 skc
->skc_linux_cache
->gfpflags
|= __GFP_COMP
;
1015 skc
->skc_flags
|= KMC_NOMAGAZINE
;
1018 if (spl_kmem_cache_expire
& KMC_EXPIRE_AGE
)
1019 skc
->skc_taskqid
= taskq_dispatch_delay(spl_kmem_cache_taskq
,
1020 spl_cache_age
, skc
, TQ_SLEEP
,
1021 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1023 down_write(&spl_kmem_cache_sem
);
1024 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
1025 up_write(&spl_kmem_cache_sem
);
1029 kfree(skc
->skc_name
);
1033 EXPORT_SYMBOL(spl_kmem_cache_create
);
1036 * Register a move callback for cache defragmentation.
1037 * XXX: Unimplemented but harmless to stub out for now.
1040 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
1041 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
1043 ASSERT(move
!= NULL
);
1045 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
1048 * Destroy a cache and all objects associated with the cache.
1051 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
1053 DECLARE_WAIT_QUEUE_HEAD(wq
);
1056 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1057 ASSERT(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
));
1059 down_write(&spl_kmem_cache_sem
);
1060 list_del_init(&skc
->skc_list
);
1061 up_write(&spl_kmem_cache_sem
);
1063 /* Cancel any and wait for any pending delayed tasks */
1064 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1066 spin_lock(&skc
->skc_lock
);
1067 id
= skc
->skc_taskqid
;
1068 spin_unlock(&skc
->skc_lock
);
1070 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1073 * Wait until all current callers complete, this is mainly
1074 * to catch the case where a low memory situation triggers a
1075 * cache reaping action which races with this destroy.
1077 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
1079 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
1080 spl_magazine_destroy(skc
);
1081 spl_slab_reclaim(skc
);
1083 ASSERT(skc
->skc_flags
& KMC_SLAB
);
1084 kmem_cache_destroy(skc
->skc_linux_cache
);
1087 spin_lock(&skc
->skc_lock
);
1090 * Validate there are no objects in use and free all the
1091 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers.
1093 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
1094 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
1095 ASSERT3U(skc
->skc_slab_total
, ==, 0);
1096 ASSERT3U(skc
->skc_obj_total
, ==, 0);
1097 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
1098 ASSERT(list_empty(&skc
->skc_complete_list
));
1100 spin_unlock(&skc
->skc_lock
);
1102 kfree(skc
->skc_name
);
1105 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
1108 * Allocate an object from a slab attached to the cache. This is used to
1109 * repopulate the per-cpu magazine caches in batches when they run low.
1112 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
1114 spl_kmem_obj_t
*sko
;
1116 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1117 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1118 ASSERT(spin_is_locked(&skc
->skc_lock
));
1120 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
1121 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1122 ASSERT(sko
->sko_addr
!= NULL
);
1124 /* Remove from sks_free_list */
1125 list_del_init(&sko
->sko_list
);
1127 sks
->sks_age
= jiffies
;
1129 skc
->skc_obj_alloc
++;
1131 /* Track max obj usage statistics */
1132 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
1133 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
1135 /* Track max slab usage statistics */
1136 if (sks
->sks_ref
== 1) {
1137 skc
->skc_slab_alloc
++;
1139 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
1140 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
1143 return (sko
->sko_addr
);
1147 * Generic slab allocation function to run by the global work queues.
1148 * It is responsible for allocating a new slab, linking it in to the list
1149 * of partial slabs, and then waking any waiters.
1152 spl_cache_grow_work(void *data
)
1154 spl_kmem_alloc_t
*ska
= (spl_kmem_alloc_t
*)data
;
1155 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
1156 spl_kmem_slab_t
*sks
;
1158 #if defined(PF_MEMALLOC_NOIO)
1159 unsigned noio_flag
= memalloc_noio_save();
1160 sks
= spl_slab_alloc(skc
, ska
->ska_flags
);
1161 memalloc_noio_restore(noio_flag
);
1163 fstrans_cookie_t cookie
= spl_fstrans_mark();
1164 sks
= spl_slab_alloc(skc
, ska
->ska_flags
);
1165 spl_fstrans_unmark(cookie
);
1167 spin_lock(&skc
->skc_lock
);
1169 skc
->skc_slab_total
++;
1170 skc
->skc_obj_total
+= sks
->sks_objs
;
1171 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1174 atomic_dec(&skc
->skc_ref
);
1175 smp_mb__before_atomic();
1176 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1177 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1178 smp_mb__after_atomic();
1179 wake_up_all(&skc
->skc_waitq
);
1180 spin_unlock(&skc
->skc_lock
);
1186 * Returns non-zero when a new slab should be available.
1189 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
1191 return (!test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
));
1195 * No available objects on any slabs, create a new slab. Note that this
1196 * functionality is disabled for KMC_SLAB caches which are backed by the
1200 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1202 int remaining
, rc
= 0;
1204 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1205 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1206 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
1211 * Before allocating a new slab wait for any reaping to complete and
1212 * then return so the local magazine can be rechecked for new objects.
1214 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1215 rc
= spl_wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1216 TASK_UNINTERRUPTIBLE
);
1217 return (rc
? rc
: -EAGAIN
);
1221 * This is handled by dispatching a work request to the global work
1222 * queue. This allows us to asynchronously allocate a new slab while
1223 * retaining the ability to safely fall back to a smaller synchronous
1224 * allocations to ensure forward progress is always maintained.
1226 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1227 spl_kmem_alloc_t
*ska
;
1229 ska
= kmalloc(sizeof (*ska
), kmem_flags_convert(flags
));
1231 clear_bit_unlock(KMC_BIT_GROWING
, &skc
->skc_flags
);
1232 smp_mb__after_atomic();
1233 wake_up_all(&skc
->skc_waitq
);
1237 atomic_inc(&skc
->skc_ref
);
1238 ska
->ska_cache
= skc
;
1239 ska
->ska_flags
= flags
;
1240 taskq_init_ent(&ska
->ska_tqe
);
1241 taskq_dispatch_ent(spl_kmem_cache_taskq
,
1242 spl_cache_grow_work
, ska
, 0, &ska
->ska_tqe
);
1246 * The goal here is to only detect the rare case where a virtual slab
1247 * allocation has deadlocked. We must be careful to minimize the use
1248 * of emergency objects which are more expensive to track. Therefore,
1249 * we set a very long timeout for the asynchronous allocation and if
1250 * the timeout is reached the cache is flagged as deadlocked. From
1251 * this point only new emergency objects will be allocated until the
1252 * asynchronous allocation completes and clears the deadlocked flag.
1254 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1255 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1257 remaining
= wait_event_timeout(skc
->skc_waitq
,
1258 spl_cache_grow_wait(skc
), HZ
/ 10);
1261 spin_lock(&skc
->skc_lock
);
1262 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1263 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1264 skc
->skc_obj_deadlock
++;
1266 spin_unlock(&skc
->skc_lock
);
1276 * Refill a per-cpu magazine with objects from the slabs for this cache.
1277 * Ideally the magazine can be repopulated using existing objects which have
1278 * been released, however if we are unable to locate enough free objects new
1279 * slabs of objects will be created. On success NULL is returned, otherwise
1280 * the address of a single emergency object is returned for use by the caller.
1283 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1285 spl_kmem_slab_t
*sks
;
1286 int count
= 0, rc
, refill
;
1289 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1290 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1292 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1293 spin_lock(&skc
->skc_lock
);
1295 while (refill
> 0) {
1296 /* No slabs available we may need to grow the cache */
1297 if (list_empty(&skc
->skc_partial_list
)) {
1298 spin_unlock(&skc
->skc_lock
);
1301 rc
= spl_cache_grow(skc
, flags
, &obj
);
1302 local_irq_disable();
1304 /* Emergency object for immediate use by caller */
1305 if (rc
== 0 && obj
!= NULL
)
1311 /* Rescheduled to different CPU skm is not local */
1312 if (skm
!= skc
->skc_mag
[smp_processor_id()])
1316 * Potentially rescheduled to the same CPU but
1317 * allocations may have occurred from this CPU while
1318 * we were sleeping so recalculate max refill.
1320 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
1322 spin_lock(&skc
->skc_lock
);
1326 /* Grab the next available slab */
1327 sks
= list_entry((&skc
->skc_partial_list
)->next
,
1328 spl_kmem_slab_t
, sks_list
);
1329 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1330 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
1331 ASSERT(!list_empty(&sks
->sks_free_list
));
1334 * Consume as many objects as needed to refill the requested
1335 * cache. We must also be careful not to overfill it.
1337 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 &&
1339 ASSERT(skm
->skm_avail
< skm
->skm_size
);
1340 ASSERT(count
< skm
->skm_size
);
1341 skm
->skm_objs
[skm
->skm_avail
++] =
1342 spl_cache_obj(skc
, sks
);
1345 /* Move slab to skc_complete_list when full */
1346 if (sks
->sks_ref
== sks
->sks_objs
) {
1347 list_del(&sks
->sks_list
);
1348 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
1352 spin_unlock(&skc
->skc_lock
);
1358 * Release an object back to the slab from which it came.
1361 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
1363 spl_kmem_slab_t
*sks
= NULL
;
1364 spl_kmem_obj_t
*sko
= NULL
;
1366 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1367 ASSERT(spin_is_locked(&skc
->skc_lock
));
1369 sko
= spl_sko_from_obj(skc
, obj
);
1370 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1371 sks
= sko
->sko_slab
;
1372 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1373 ASSERT(sks
->sks_cache
== skc
);
1374 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
1376 sks
->sks_age
= jiffies
;
1378 skc
->skc_obj_alloc
--;
1381 * Move slab to skc_partial_list when no longer full. Slabs
1382 * are added to the head to keep the partial list is quasi-full
1383 * sorted order. Fuller at the head, emptier at the tail.
1385 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
1386 list_del(&sks
->sks_list
);
1387 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
1391 * Move empty slabs to the end of the partial list so
1392 * they can be easily found and freed during reclamation.
1394 if (sks
->sks_ref
== 0) {
1395 list_del(&sks
->sks_list
);
1396 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1397 skc
->skc_slab_alloc
--;
1402 * Allocate an object from the per-cpu magazine, or if the magazine
1403 * is empty directly allocate from a slab and repopulate the magazine.
1406 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
1408 spl_kmem_magazine_t
*skm
;
1411 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1412 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1413 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1416 * Allocate directly from a Linux slab. All optimizations are left
1417 * to the underlying cache we only need to guarantee that KM_SLEEP
1418 * callers will never fail.
1420 if (skc
->skc_flags
& KMC_SLAB
) {
1421 struct kmem_cache
*slc
= skc
->skc_linux_cache
;
1423 obj
= kmem_cache_alloc(slc
, kmem_flags_convert(flags
));
1424 } while ((obj
== NULL
) && !(flags
& KM_NOSLEEP
));
1429 local_irq_disable();
1433 * Safe to update per-cpu structure without lock, but
1434 * in the restart case we must be careful to reacquire
1435 * the local magazine since this may have changed
1436 * when we need to grow the cache.
1438 skm
= skc
->skc_mag
[smp_processor_id()];
1439 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1441 if (likely(skm
->skm_avail
)) {
1442 /* Object available in CPU cache, use it */
1443 obj
= skm
->skm_objs
[--skm
->skm_avail
];
1444 skm
->skm_age
= jiffies
;
1446 obj
= spl_cache_refill(skc
, skm
, flags
);
1447 if ((obj
== NULL
) && !(flags
& KM_NOSLEEP
))
1456 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
1459 /* Pre-emptively migrate object to CPU L1 cache */
1461 if (obj
&& skc
->skc_ctor
)
1462 skc
->skc_ctor(obj
, skc
->skc_private
, flags
);
1469 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
1472 * Free an object back to the local per-cpu magazine, there is no
1473 * guarantee that this is the same magazine the object was originally
1474 * allocated from. We may need to flush entire from the magazine
1475 * back to the slabs to make space.
1478 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
1480 spl_kmem_magazine_t
*skm
;
1481 unsigned long flags
;
1483 int do_emergency
= 0;
1485 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1486 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1489 * Run the destructor
1492 skc
->skc_dtor(obj
, skc
->skc_private
);
1495 * Free the object from the Linux underlying Linux slab.
1497 if (skc
->skc_flags
& KMC_SLAB
) {
1498 kmem_cache_free(skc
->skc_linux_cache
, obj
);
1503 * While a cache has outstanding emergency objects all freed objects
1504 * must be checked. However, since emergency objects will never use
1505 * a virtual address these objects can be safely excluded as an
1508 if (!is_vmalloc_addr(obj
)) {
1509 spin_lock(&skc
->skc_lock
);
1510 do_emergency
= (skc
->skc_obj_emergency
> 0);
1511 spin_unlock(&skc
->skc_lock
);
1513 if (do_emergency
&& (spl_emergency_free(skc
, obj
) == 0))
1517 local_irq_save(flags
);
1520 * Safe to update per-cpu structure without lock, but
1521 * no remote memory allocation tracking is being performed
1522 * it is entirely possible to allocate an object from one
1523 * CPU cache and return it to another.
1525 skm
= skc
->skc_mag
[smp_processor_id()];
1526 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1529 * Per-CPU cache full, flush it to make space for this object,
1530 * this may result in an empty slab which can be reclaimed once
1531 * interrupts are re-enabled.
1533 if (unlikely(skm
->skm_avail
>= skm
->skm_size
)) {
1534 spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1538 /* Available space in cache, use it */
1539 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
1541 local_irq_restore(flags
);
1544 spl_slab_reclaim(skc
);
1546 EXPORT_SYMBOL(spl_kmem_cache_free
);
1549 * The generic shrinker function for all caches. Under Linux a shrinker
1550 * may not be tightly coupled with a slab cache. In fact Linux always
1551 * systematically tries calling all registered shrinker callbacks which
1552 * report that they contain unused objects. Because of this we only
1553 * register one shrinker function in the shim layer for all slab caches.
1554 * We always attempt to shrink all caches when this generic shrinker
1557 * If sc->nr_to_scan is zero, the caller is requesting a query of the
1558 * number of objects which can potentially be freed. If it is nonzero,
1559 * the request is to free that many objects.
1561 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
1562 * in struct shrinker and also require the shrinker to return the number
1565 * Older kernels require the shrinker to return the number of freeable
1566 * objects following the freeing of nr_to_free.
1568 * Linux semantics differ from those under Solaris, which are to
1569 * free all available objects which may (and probably will) be more
1570 * objects than the requested nr_to_scan.
1572 static spl_shrinker_t
1573 __spl_kmem_cache_generic_shrinker(struct shrinker
*shrink
,
1574 struct shrink_control
*sc
)
1576 spl_kmem_cache_t
*skc
;
1580 * No shrinking in a transaction context. Can cause deadlocks.
1582 if (sc
->nr_to_scan
&& spl_fstrans_check())
1583 return (SHRINK_STOP
);
1585 down_read(&spl_kmem_cache_sem
);
1586 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
1587 if (sc
->nr_to_scan
) {
1588 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
1589 uint64_t oldalloc
= skc
->skc_obj_alloc
;
1590 spl_kmem_cache_reap_now(skc
,
1591 MAX(sc
->nr_to_scan
>>fls64(skc
->skc_slab_objs
), 1));
1592 if (oldalloc
> skc
->skc_obj_alloc
)
1593 alloc
+= oldalloc
- skc
->skc_obj_alloc
;
1595 spl_kmem_cache_reap_now(skc
,
1596 MAX(sc
->nr_to_scan
>>fls64(skc
->skc_slab_objs
), 1));
1597 alloc
+= skc
->skc_obj_alloc
;
1598 #endif /* HAVE_SPLIT_SHRINKER_CALLBACK */
1600 /* Request to query number of freeable objects */
1601 alloc
+= skc
->skc_obj_alloc
;
1604 up_read(&spl_kmem_cache_sem
);
1607 * When KMC_RECLAIM_ONCE is set allow only a single reclaim pass.
1608 * This functionality only exists to work around a rare issue where
1609 * shrink_slabs() is repeatedly invoked by many cores causing the
1612 if ((spl_kmem_cache_reclaim
& KMC_RECLAIM_ONCE
) && sc
->nr_to_scan
)
1613 return (SHRINK_STOP
);
1615 return (MAX(alloc
, 0));
1618 SPL_SHRINKER_CALLBACK_WRAPPER(spl_kmem_cache_generic_shrinker
);
1621 * Call the registered reclaim function for a cache. Depending on how
1622 * many and which objects are released it may simply repopulate the
1623 * local magazine which will then need to age-out. Objects which cannot
1624 * fit in the magazine we will be released back to their slabs which will
1625 * also need to age out before being release. This is all just best
1626 * effort and we do not want to thrash creating and destroying slabs.
1629 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
, int count
)
1631 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1632 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1634 atomic_inc(&skc
->skc_ref
);
1637 * Execute the registered reclaim callback if it exists.
1639 if (skc
->skc_flags
& KMC_SLAB
) {
1640 if (skc
->skc_reclaim
)
1641 skc
->skc_reclaim(skc
->skc_private
);
1646 * Prevent concurrent cache reaping when contended.
1648 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
))
1652 * When a reclaim function is available it may be invoked repeatedly
1653 * until at least a single slab can be freed. This ensures that we
1654 * do free memory back to the system. This helps minimize the chance
1655 * of an OOM event when the bulk of memory is used by the slab.
1657 * When free slabs are already available the reclaim callback will be
1658 * skipped. Additionally, if no forward progress is detected despite
1659 * a reclaim function the cache will be skipped to avoid deadlock.
1661 * Longer term this would be the correct place to add the code which
1662 * repacks the slabs in order minimize fragmentation.
1664 if (skc
->skc_reclaim
) {
1665 uint64_t objects
= UINT64_MAX
;
1669 spin_lock(&skc
->skc_lock
);
1671 (skc
->skc_slab_total
> 0) &&
1672 ((skc
->skc_slab_total
-skc
->skc_slab_alloc
) == 0) &&
1673 (skc
->skc_obj_alloc
< objects
);
1675 objects
= skc
->skc_obj_alloc
;
1676 spin_unlock(&skc
->skc_lock
);
1679 skc
->skc_reclaim(skc
->skc_private
);
1681 } while (do_reclaim
);
1684 /* Reclaim from the magazine and free all now empty slabs. */
1685 if (spl_kmem_cache_expire
& KMC_EXPIRE_MEM
) {
1686 spl_kmem_magazine_t
*skm
;
1687 unsigned long irq_flags
;
1689 local_irq_save(irq_flags
);
1690 skm
= skc
->skc_mag
[smp_processor_id()];
1691 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1692 local_irq_restore(irq_flags
);
1695 spl_slab_reclaim(skc
);
1696 clear_bit_unlock(KMC_BIT_REAPING
, &skc
->skc_flags
);
1697 smp_mb__after_atomic();
1698 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
1700 atomic_dec(&skc
->skc_ref
);
1702 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
1705 * Reap all free slabs from all registered caches.
1710 struct shrink_control sc
;
1712 sc
.nr_to_scan
= KMC_REAP_CHUNK
;
1713 sc
.gfp_mask
= GFP_KERNEL
;
1715 (void) __spl_kmem_cache_generic_shrinker(NULL
, &sc
);
1717 EXPORT_SYMBOL(spl_kmem_reap
);
1720 spl_kmem_cache_init(void)
1722 init_rwsem(&spl_kmem_cache_sem
);
1723 INIT_LIST_HEAD(&spl_kmem_cache_list
);
1724 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
1725 spl_kmem_cache_kmem_threads
, maxclsyspri
,
1726 spl_kmem_cache_kmem_threads
* 8, INT_MAX
,
1727 TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
1728 spl_register_shrinker(&spl_kmem_cache_shrinker
);
1734 spl_kmem_cache_fini(void)
1736 spl_unregister_shrinker(&spl_kmem_cache_shrinker
);
1737 taskq_destroy(spl_kmem_cache_taskq
);