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>
34 #include <linux/prefetch.h>
37 * Within the scope of spl-kmem.c file the kmem_cache_* definitions
38 * are removed to allow access to the real Linux slab allocator.
40 #undef kmem_cache_destroy
41 #undef kmem_cache_create
42 #undef kmem_cache_alloc
43 #undef kmem_cache_free
47 * Linux 3.16 replaced smp_mb__{before,after}_{atomic,clear}_{dec,inc,bit}()
48 * with smp_mb__{before,after}_atomic() because they were redundant. This is
49 * only used inside our SLAB allocator, so we implement an internal wrapper
50 * here to give us smp_mb__{before,after}_atomic() on older kernels.
52 #ifndef smp_mb__before_atomic
53 #define smp_mb__before_atomic(x) smp_mb__before_clear_bit(x)
56 #ifndef smp_mb__after_atomic
57 #define smp_mb__after_atomic(x) smp_mb__after_clear_bit(x)
61 * Cache expiration was implemented because it was part of the default Solaris
62 * kmem_cache behavior. The idea is that per-cpu objects which haven't been
63 * accessed in several seconds should be returned to the cache. On the other
64 * hand Linux slabs never move objects back to the slabs unless there is
65 * memory pressure on the system. By default the Linux method is enabled
66 * because it has been shown to improve responsiveness on low memory systems.
67 * This policy may be changed by setting KMC_EXPIRE_AGE or KMC_EXPIRE_MEM.
69 unsigned int spl_kmem_cache_expire
= KMC_EXPIRE_MEM
;
70 EXPORT_SYMBOL(spl_kmem_cache_expire
);
71 module_param(spl_kmem_cache_expire
, uint
, 0644);
72 MODULE_PARM_DESC(spl_kmem_cache_expire
, "By age (0x1) or low memory (0x2)");
75 * Cache magazines are an optimization designed to minimize the cost of
76 * allocating memory. They do this by keeping a per-cpu cache of recently
77 * freed objects, which can then be reallocated without taking a lock. This
78 * can improve performance on highly contended caches. However, because
79 * objects in magazines will prevent otherwise empty slabs from being
80 * immediately released this may not be ideal for low memory machines.
82 * For this reason spl_kmem_cache_magazine_size can be used to set a maximum
83 * magazine size. When this value is set to 0 the magazine size will be
84 * automatically determined based on the object size. Otherwise magazines
85 * will be limited to 2-256 objects per magazine (i.e per cpu). Magazines
86 * may never be entirely disabled in this implementation.
88 unsigned int spl_kmem_cache_magazine_size
= 0;
89 module_param(spl_kmem_cache_magazine_size
, uint
, 0444);
90 MODULE_PARM_DESC(spl_kmem_cache_magazine_size
,
91 "Default magazine size (2-256), set automatically (0)");
94 * The default behavior is to report the number of objects remaining in the
95 * cache. This allows the Linux VM to repeatedly reclaim objects from the
96 * cache when memory is low satisfy other memory allocations. Alternately,
97 * setting this value to KMC_RECLAIM_ONCE limits how aggressively the cache
98 * is reclaimed. This may increase the likelihood of out of memory events.
100 unsigned int spl_kmem_cache_reclaim
= 0 /* KMC_RECLAIM_ONCE */;
101 module_param(spl_kmem_cache_reclaim
, uint
, 0644);
102 MODULE_PARM_DESC(spl_kmem_cache_reclaim
, "Single reclaim pass (0x1)");
104 unsigned int spl_kmem_cache_obj_per_slab
= SPL_KMEM_CACHE_OBJ_PER_SLAB
;
105 module_param(spl_kmem_cache_obj_per_slab
, uint
, 0644);
106 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab
, "Number of objects per slab");
108 unsigned int spl_kmem_cache_obj_per_slab_min
= SPL_KMEM_CACHE_OBJ_PER_SLAB_MIN
;
109 module_param(spl_kmem_cache_obj_per_slab_min
, uint
, 0644);
110 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab_min
,
111 "Minimal number of objects per slab");
113 unsigned int spl_kmem_cache_max_size
= SPL_KMEM_CACHE_MAX_SIZE
;
114 module_param(spl_kmem_cache_max_size
, uint
, 0644);
115 MODULE_PARM_DESC(spl_kmem_cache_max_size
, "Maximum size of slab in MB");
118 * For small objects the Linux slab allocator should be used to make the most
119 * efficient use of the memory. However, large objects are not supported by
120 * the Linux slab and therefore the SPL implementation is preferred. A cutoff
121 * of 16K was determined to be optimal for architectures using 4K pages.
123 #if PAGE_SIZE == 4096
124 unsigned int spl_kmem_cache_slab_limit
= 16384;
126 unsigned int spl_kmem_cache_slab_limit
= 0;
128 module_param(spl_kmem_cache_slab_limit
, uint
, 0644);
129 MODULE_PARM_DESC(spl_kmem_cache_slab_limit
,
130 "Objects less than N bytes use the Linux slab");
133 * This value defaults to a threshold designed to avoid allocations which
134 * have been deemed costly by the kernel.
136 unsigned int spl_kmem_cache_kmem_limit
=
137 ((1 << (PAGE_ALLOC_COSTLY_ORDER
- 1)) * PAGE_SIZE
) /
138 SPL_KMEM_CACHE_OBJ_PER_SLAB
;
139 module_param(spl_kmem_cache_kmem_limit
, uint
, 0644);
140 MODULE_PARM_DESC(spl_kmem_cache_kmem_limit
,
141 "Objects less than N bytes use the kmalloc");
144 * The number of threads available to allocate new slabs for caches. This
145 * should not need to be tuned but it is available for performance analysis.
147 unsigned int spl_kmem_cache_kmem_threads
= 4;
148 module_param(spl_kmem_cache_kmem_threads
, uint
, 0444);
149 MODULE_PARM_DESC(spl_kmem_cache_kmem_threads
,
150 "Number of spl_kmem_cache threads");
153 * Slab allocation interfaces
155 * While the Linux slab implementation was inspired by the Solaris
156 * implementation I cannot use it to emulate the Solaris APIs. I
157 * require two features which are not provided by the Linux slab.
159 * 1) Constructors AND destructors. Recent versions of the Linux
160 * kernel have removed support for destructors. This is a deal
161 * breaker for the SPL which contains particularly expensive
162 * initializers for mutex's, condition variables, etc. We also
163 * require a minimal level of cleanup for these data types unlike
164 * many Linux data types which do need to be explicitly destroyed.
166 * 2) Virtual address space backed slab. Callers of the Solaris slab
167 * expect it to work well for both small are very large allocations.
168 * Because of memory fragmentation the Linux slab which is backed
169 * by kmalloc'ed memory performs very badly when confronted with
170 * large numbers of large allocations. Basing the slab on the
171 * virtual address space removes the need for contiguous pages
172 * and greatly improve performance for large allocations.
174 * For these reasons, the SPL has its own slab implementation with
175 * the needed features. It is not as highly optimized as either the
176 * Solaris or Linux slabs, but it should get me most of what is
177 * needed until it can be optimized or obsoleted by another approach.
179 * One serious concern I do have about this method is the relatively
180 * small virtual address space on 32bit arches. This will seriously
181 * constrain the size of the slab caches and their performance.
184 struct list_head spl_kmem_cache_list
; /* List of caches */
185 struct rw_semaphore spl_kmem_cache_sem
; /* Cache list lock */
186 taskq_t
*spl_kmem_cache_taskq
; /* Task queue for ageing / reclaim */
188 static void spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
);
190 SPL_SHRINKER_CALLBACK_FWD_DECLARE(spl_kmem_cache_generic_shrinker
);
191 SPL_SHRINKER_DECLARE(spl_kmem_cache_shrinker
,
192 spl_kmem_cache_generic_shrinker
, KMC_DEFAULT_SEEKS
);
195 kv_alloc(spl_kmem_cache_t
*skc
, int size
, int flags
)
197 gfp_t lflags
= kmem_flags_convert(flags
);
200 if (skc
->skc_flags
& KMC_KMEM
) {
202 ptr
= (void *)__get_free_pages(lflags
, get_order(size
));
204 ptr
= __vmalloc(size
, lflags
| __GFP_HIGHMEM
, PAGE_KERNEL
);
207 /* Resulting allocated memory will be page aligned */
208 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
214 kv_free(spl_kmem_cache_t
*skc
, void *ptr
, int size
)
216 ASSERT(IS_P2ALIGNED(ptr
, PAGE_SIZE
));
219 * The Linux direct reclaim path uses this out of band value to
220 * determine if forward progress is being made. Normally this is
221 * incremented by kmem_freepages() which is part of the various
222 * Linux slab implementations. However, since we are using none
223 * of that infrastructure we are responsible for incrementing it.
225 if (current
->reclaim_state
)
226 current
->reclaim_state
->reclaimed_slab
+= size
>> PAGE_SHIFT
;
228 if (skc
->skc_flags
& KMC_KMEM
) {
230 free_pages((unsigned long)ptr
, get_order(size
));
237 * Required space for each aligned sks.
239 static inline uint32_t
240 spl_sks_size(spl_kmem_cache_t
*skc
)
242 return (P2ROUNDUP_TYPED(sizeof (spl_kmem_slab_t
),
243 skc
->skc_obj_align
, uint32_t));
247 * Required space for each aligned object.
249 static inline uint32_t
250 spl_obj_size(spl_kmem_cache_t
*skc
)
252 uint32_t align
= skc
->skc_obj_align
;
254 return (P2ROUNDUP_TYPED(skc
->skc_obj_size
, align
, uint32_t) +
255 P2ROUNDUP_TYPED(sizeof (spl_kmem_obj_t
), align
, uint32_t));
259 * Lookup the spl_kmem_object_t for an object given that object.
261 static inline spl_kmem_obj_t
*
262 spl_sko_from_obj(spl_kmem_cache_t
*skc
, void *obj
)
264 return (obj
+ P2ROUNDUP_TYPED(skc
->skc_obj_size
,
265 skc
->skc_obj_align
, uint32_t));
269 * Required space for each offslab object taking in to account alignment
270 * restrictions and the power-of-two requirement of kv_alloc().
272 static inline uint32_t
273 spl_offslab_size(spl_kmem_cache_t
*skc
)
275 return (1UL << (fls64(spl_obj_size(skc
)) + 1));
279 * It's important that we pack the spl_kmem_obj_t structure and the
280 * actual objects in to one large address space to minimize the number
281 * of calls to the allocator. It is far better to do a few large
282 * allocations and then subdivide it ourselves. Now which allocator
283 * we use requires balancing a few trade offs.
285 * For small objects we use kmem_alloc() because as long as you are
286 * only requesting a small number of pages (ideally just one) its cheap.
287 * However, when you start requesting multiple pages with kmem_alloc()
288 * it gets increasingly expensive since it requires contiguous pages.
289 * For this reason we shift to vmem_alloc() for slabs of large objects
290 * which removes the need for contiguous pages. We do not use
291 * vmem_alloc() in all cases because there is significant locking
292 * overhead in __get_vm_area_node(). This function takes a single
293 * global lock when acquiring an available virtual address range which
294 * serializes all vmem_alloc()'s for all slab caches. Using slightly
295 * different allocation functions for small and large objects should
296 * give us the best of both worlds.
298 * KMC_ONSLAB KMC_OFFSLAB
300 * +------------------------+ +-----------------+
301 * | spl_kmem_slab_t --+-+ | | spl_kmem_slab_t |---+-+
302 * | skc_obj_size <-+ | | +-----------------+ | |
303 * | spl_kmem_obj_t | | | |
304 * | skc_obj_size <---+ | +-----------------+ | |
305 * | spl_kmem_obj_t | | | skc_obj_size | <-+ |
306 * | ... v | | spl_kmem_obj_t | |
307 * +------------------------+ +-----------------+ v
309 static spl_kmem_slab_t
*
310 spl_slab_alloc(spl_kmem_cache_t
*skc
, int flags
)
312 spl_kmem_slab_t
*sks
;
313 spl_kmem_obj_t
*sko
, *n
;
315 uint32_t obj_size
, offslab_size
= 0;
318 base
= kv_alloc(skc
, skc
->skc_slab_size
, flags
);
322 sks
= (spl_kmem_slab_t
*)base
;
323 sks
->sks_magic
= SKS_MAGIC
;
324 sks
->sks_objs
= skc
->skc_slab_objs
;
325 sks
->sks_age
= jiffies
;
326 sks
->sks_cache
= skc
;
327 INIT_LIST_HEAD(&sks
->sks_list
);
328 INIT_LIST_HEAD(&sks
->sks_free_list
);
330 obj_size
= spl_obj_size(skc
);
332 if (skc
->skc_flags
& KMC_OFFSLAB
)
333 offslab_size
= spl_offslab_size(skc
);
335 for (i
= 0; i
< sks
->sks_objs
; i
++) {
336 if (skc
->skc_flags
& KMC_OFFSLAB
) {
337 obj
= kv_alloc(skc
, offslab_size
, flags
);
343 obj
= base
+ spl_sks_size(skc
) + (i
* obj_size
);
346 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
347 sko
= spl_sko_from_obj(skc
, obj
);
349 sko
->sko_magic
= SKO_MAGIC
;
351 INIT_LIST_HEAD(&sko
->sko_list
);
352 list_add_tail(&sko
->sko_list
, &sks
->sks_free_list
);
357 if (skc
->skc_flags
& KMC_OFFSLAB
)
358 list_for_each_entry_safe(sko
,
359 n
, &sks
->sks_free_list
, sko_list
)
360 kv_free(skc
, sko
->sko_addr
, offslab_size
);
362 kv_free(skc
, base
, skc
->skc_slab_size
);
370 * Remove a slab from complete or partial list, it must be called with
371 * the 'skc->skc_lock' held but the actual free must be performed
372 * outside the lock to prevent deadlocking on vmem addresses.
375 spl_slab_free(spl_kmem_slab_t
*sks
,
376 struct list_head
*sks_list
, struct list_head
*sko_list
)
378 spl_kmem_cache_t
*skc
;
380 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
381 ASSERT(sks
->sks_ref
== 0);
383 skc
= sks
->sks_cache
;
384 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
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
);
586 for (i
= 0; i
< count
; i
++)
587 spl_cache_shrink(skc
, skm
->skm_objs
[i
]);
589 skm
->skm_avail
-= count
;
590 memmove(skm
->skm_objs
, &(skm
->skm_objs
[count
]),
591 sizeof (void *) * skm
->skm_avail
);
595 spl_cache_flush(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flush
)
597 spin_lock(&skc
->skc_lock
);
598 __spl_cache_flush(skc
, skm
, flush
);
599 spin_unlock(&skc
->skc_lock
);
603 spl_magazine_age(void *data
)
605 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
606 spl_kmem_magazine_t
*skm
= skc
->skc_mag
[smp_processor_id()];
608 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
609 ASSERT(skm
->skm_cpu
== smp_processor_id());
610 ASSERT(irqs_disabled());
612 /* There are no available objects or they are too young to age out */
613 if ((skm
->skm_avail
== 0) ||
614 time_before(jiffies
, skm
->skm_age
+ skc
->skc_delay
* HZ
))
618 * Because we're executing in interrupt context we may have
619 * interrupted the holder of this lock. To avoid a potential
620 * deadlock return if the lock is contended.
622 if (!spin_trylock(&skc
->skc_lock
))
625 __spl_cache_flush(skc
, skm
, skm
->skm_refill
);
626 spin_unlock(&skc
->skc_lock
);
630 * Called regularly to keep a downward pressure on the cache.
632 * Objects older than skc->skc_delay seconds in the per-cpu magazines will
633 * be returned to the caches. This is done to prevent idle magazines from
634 * holding memory which could be better used elsewhere. The delay is
635 * present to prevent thrashing the magazine.
637 * The newly released objects may result in empty partial slabs. Those
638 * slabs should be released to the system. Otherwise moving the objects
639 * out of the magazines is just wasted work.
642 spl_cache_age(void *data
)
644 spl_kmem_cache_t
*skc
= (spl_kmem_cache_t
*)data
;
647 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
649 /* Dynamically disabled at run time */
650 if (!(spl_kmem_cache_expire
& KMC_EXPIRE_AGE
))
653 atomic_inc(&skc
->skc_ref
);
655 if (!(skc
->skc_flags
& KMC_NOMAGAZINE
))
656 on_each_cpu(spl_magazine_age
, skc
, 1);
658 spl_slab_reclaim(skc
);
660 while (!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && !id
) {
661 id
= taskq_dispatch_delay(
662 spl_kmem_cache_taskq
, spl_cache_age
, skc
, TQ_SLEEP
,
663 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
665 /* Destroy issued after dispatch immediately cancel it */
666 if (test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
) && id
)
667 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
670 spin_lock(&skc
->skc_lock
);
671 skc
->skc_taskqid
= id
;
672 spin_unlock(&skc
->skc_lock
);
674 atomic_dec(&skc
->skc_ref
);
678 * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
679 * When on-slab we want to target spl_kmem_cache_obj_per_slab. However,
680 * for very small objects we may end up with more than this so as not
681 * to waste space in the minimal allocation of a single page. Also for
682 * very large objects we may use as few as spl_kmem_cache_obj_per_slab_min,
683 * lower than this and we will fail.
686 spl_slab_size(spl_kmem_cache_t
*skc
, uint32_t *objs
, uint32_t *size
)
688 uint32_t sks_size
, obj_size
, max_size
, tgt_size
, tgt_objs
;
690 if (skc
->skc_flags
& KMC_OFFSLAB
) {
691 tgt_objs
= spl_kmem_cache_obj_per_slab
;
692 tgt_size
= P2ROUNDUP(sizeof (spl_kmem_slab_t
), PAGE_SIZE
);
694 if ((skc
->skc_flags
& KMC_KMEM
) &&
695 (spl_obj_size(skc
) > (SPL_MAX_ORDER_NR_PAGES
* PAGE_SIZE
)))
698 sks_size
= spl_sks_size(skc
);
699 obj_size
= spl_obj_size(skc
);
700 max_size
= (spl_kmem_cache_max_size
* 1024 * 1024);
701 tgt_size
= (spl_kmem_cache_obj_per_slab
* obj_size
+ sks_size
);
704 * KMC_KMEM slabs are allocated by __get_free_pages() which
705 * rounds up to the nearest order. Knowing this the size
706 * should be rounded up to the next power of two with a hard
707 * maximum defined by the maximum allowed allocation order.
709 if (skc
->skc_flags
& KMC_KMEM
) {
710 max_size
= SPL_MAX_ORDER_NR_PAGES
* PAGE_SIZE
;
711 tgt_size
= MIN(max_size
,
712 PAGE_SIZE
* (1 << MAX(get_order(tgt_size
) - 1, 1)));
715 if (tgt_size
<= max_size
) {
716 tgt_objs
= (tgt_size
- sks_size
) / obj_size
;
718 tgt_objs
= (max_size
- sks_size
) / obj_size
;
719 tgt_size
= (tgt_objs
* obj_size
) + sks_size
;
733 * Make a guess at reasonable per-cpu magazine size based on the size of
734 * each object and the cost of caching N of them in each magazine. Long
735 * term this should really adapt based on an observed usage heuristic.
738 spl_magazine_size(spl_kmem_cache_t
*skc
)
740 uint32_t obj_size
= spl_obj_size(skc
);
743 if (spl_kmem_cache_magazine_size
> 0)
744 return (MAX(MIN(spl_kmem_cache_magazine_size
, 256), 2));
746 /* Per-magazine sizes below assume a 4Kib page size */
747 if (obj_size
> (PAGE_SIZE
* 256))
748 size
= 4; /* Minimum 4Mib per-magazine */
749 else if (obj_size
> (PAGE_SIZE
* 32))
750 size
= 16; /* Minimum 2Mib per-magazine */
751 else if (obj_size
> (PAGE_SIZE
))
752 size
= 64; /* Minimum 256Kib per-magazine */
753 else if (obj_size
> (PAGE_SIZE
/ 4))
754 size
= 128; /* Minimum 128Kib per-magazine */
762 * Allocate a per-cpu magazine to associate with a specific core.
764 static spl_kmem_magazine_t
*
765 spl_magazine_alloc(spl_kmem_cache_t
*skc
, int cpu
)
767 spl_kmem_magazine_t
*skm
;
768 int size
= sizeof (spl_kmem_magazine_t
) +
769 sizeof (void *) * skc
->skc_mag_size
;
771 skm
= kmalloc_node(size
, GFP_KERNEL
, cpu_to_node(cpu
));
773 skm
->skm_magic
= SKM_MAGIC
;
775 skm
->skm_size
= skc
->skc_mag_size
;
776 skm
->skm_refill
= skc
->skc_mag_refill
;
777 skm
->skm_cache
= skc
;
778 skm
->skm_age
= jiffies
;
786 * Free a per-cpu magazine associated with a specific core.
789 spl_magazine_free(spl_kmem_magazine_t
*skm
)
791 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
792 ASSERT(skm
->skm_avail
== 0);
797 * Create all pre-cpu magazines of reasonable sizes.
800 spl_magazine_create(spl_kmem_cache_t
*skc
)
804 if (skc
->skc_flags
& KMC_NOMAGAZINE
)
807 skc
->skc_mag
= kzalloc(sizeof (spl_kmem_magazine_t
*) *
808 num_possible_cpus(), kmem_flags_convert(KM_SLEEP
));
809 skc
->skc_mag_size
= spl_magazine_size(skc
);
810 skc
->skc_mag_refill
= (skc
->skc_mag_size
+ 1) / 2;
812 for_each_possible_cpu(i
) {
813 skc
->skc_mag
[i
] = spl_magazine_alloc(skc
, i
);
814 if (!skc
->skc_mag
[i
]) {
815 for (i
--; i
>= 0; i
--)
816 spl_magazine_free(skc
->skc_mag
[i
]);
827 * Destroy all pre-cpu magazines.
830 spl_magazine_destroy(spl_kmem_cache_t
*skc
)
832 spl_kmem_magazine_t
*skm
;
835 if (skc
->skc_flags
& KMC_NOMAGAZINE
)
838 for_each_possible_cpu(i
) {
839 skm
= skc
->skc_mag
[i
];
840 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
841 spl_magazine_free(skm
);
848 * Create a object cache based on the following arguments:
850 * size cache object size
851 * align cache object alignment
852 * ctor cache object constructor
853 * dtor cache object destructor
854 * reclaim cache object reclaim
855 * priv cache private data for ctor/dtor/reclaim
856 * vmp unused must be NULL
858 * KMC_NOTOUCH Disable cache object aging (unsupported)
859 * KMC_NODEBUG Disable debugging (unsupported)
860 * KMC_NOHASH Disable hashing (unsupported)
861 * KMC_QCACHE Disable qcache (unsupported)
862 * KMC_NOMAGAZINE Enabled for kmem/vmem, Disabled for Linux slab
863 * KMC_KMEM Force kmem backed cache
864 * KMC_VMEM Force vmem backed cache
865 * KMC_SLAB Force Linux slab backed cache
866 * KMC_OFFSLAB Locate objects off the slab
869 spl_kmem_cache_create(char *name
, size_t size
, size_t align
,
870 spl_kmem_ctor_t ctor
, spl_kmem_dtor_t dtor
, spl_kmem_reclaim_t reclaim
,
871 void *priv
, void *vmp
, int flags
)
873 gfp_t lflags
= kmem_flags_convert(KM_SLEEP
);
874 spl_kmem_cache_t
*skc
;
880 ASSERT0(flags
& KMC_NOMAGAZINE
);
881 ASSERT0(flags
& KMC_NOHASH
);
882 ASSERT0(flags
& KMC_QCACHE
);
887 skc
= kzalloc(sizeof (*skc
), lflags
);
891 skc
->skc_magic
= SKC_MAGIC
;
892 skc
->skc_name_size
= strlen(name
) + 1;
893 skc
->skc_name
= (char *)kmalloc(skc
->skc_name_size
, lflags
);
894 if (skc
->skc_name
== NULL
) {
898 strncpy(skc
->skc_name
, name
, skc
->skc_name_size
);
900 skc
->skc_ctor
= ctor
;
901 skc
->skc_dtor
= dtor
;
902 skc
->skc_reclaim
= reclaim
;
903 skc
->skc_private
= priv
;
905 skc
->skc_linux_cache
= NULL
;
906 skc
->skc_flags
= flags
;
907 skc
->skc_obj_size
= size
;
908 skc
->skc_obj_align
= SPL_KMEM_CACHE_ALIGN
;
909 skc
->skc_delay
= SPL_KMEM_CACHE_DELAY
;
910 skc
->skc_reap
= SPL_KMEM_CACHE_REAP
;
911 atomic_set(&skc
->skc_ref
, 0);
913 INIT_LIST_HEAD(&skc
->skc_list
);
914 INIT_LIST_HEAD(&skc
->skc_complete_list
);
915 INIT_LIST_HEAD(&skc
->skc_partial_list
);
916 skc
->skc_emergency_tree
= RB_ROOT
;
917 spin_lock_init(&skc
->skc_lock
);
918 init_waitqueue_head(&skc
->skc_waitq
);
919 skc
->skc_slab_fail
= 0;
920 skc
->skc_slab_create
= 0;
921 skc
->skc_slab_destroy
= 0;
922 skc
->skc_slab_total
= 0;
923 skc
->skc_slab_alloc
= 0;
924 skc
->skc_slab_max
= 0;
925 skc
->skc_obj_total
= 0;
926 skc
->skc_obj_alloc
= 0;
927 skc
->skc_obj_max
= 0;
928 skc
->skc_obj_deadlock
= 0;
929 skc
->skc_obj_emergency
= 0;
930 skc
->skc_obj_emergency_max
= 0;
933 * Verify the requested alignment restriction is sane.
937 VERIFY3U(align
, >=, SPL_KMEM_CACHE_ALIGN
);
938 VERIFY3U(align
, <=, PAGE_SIZE
);
939 skc
->skc_obj_align
= align
;
943 * When no specific type of slab is requested (kmem, vmem, or
944 * linuxslab) then select a cache type based on the object size
945 * and default tunables.
947 if (!(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
))) {
950 * Objects smaller than spl_kmem_cache_slab_limit can
951 * use the Linux slab for better space-efficiency. By
952 * default this functionality is disabled until its
953 * performance characteristics are fully understood.
955 if (spl_kmem_cache_slab_limit
&&
956 size
<= (size_t)spl_kmem_cache_slab_limit
)
957 skc
->skc_flags
|= KMC_SLAB
;
960 * Small objects, less than spl_kmem_cache_kmem_limit per
961 * object should use kmem because their slabs are small.
963 else if (spl_obj_size(skc
) <= spl_kmem_cache_kmem_limit
)
964 skc
->skc_flags
|= KMC_KMEM
;
967 * All other objects are considered large and are placed
968 * on vmem backed slabs.
971 skc
->skc_flags
|= KMC_VMEM
;
975 * Given the type of slab allocate the required resources.
977 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
978 rc
= spl_slab_size(skc
,
979 &skc
->skc_slab_objs
, &skc
->skc_slab_size
);
983 rc
= spl_magazine_create(skc
);
987 unsigned long slabflags
= 0;
989 if (size
> (SPL_MAX_KMEM_ORDER_NR_PAGES
* PAGE_SIZE
)) {
994 #if defined(SLAB_USERCOPY)
996 * Required for PAX-enabled kernels if the slab is to be
997 * used for coping between user and kernel space.
999 slabflags
|= SLAB_USERCOPY
;
1002 #if defined(HAVE_KMEM_CACHE_CREATE_USERCOPY)
1004 * Newer grsec patchset uses kmem_cache_create_usercopy()
1005 * instead of SLAB_USERCOPY flag
1007 skc
->skc_linux_cache
= kmem_cache_create_usercopy(
1008 skc
->skc_name
, size
, align
, slabflags
, 0, size
, NULL
);
1010 skc
->skc_linux_cache
= kmem_cache_create(
1011 skc
->skc_name
, size
, align
, slabflags
, NULL
);
1013 if (skc
->skc_linux_cache
== NULL
) {
1018 #if defined(HAVE_KMEM_CACHE_ALLOCFLAGS)
1019 skc
->skc_linux_cache
->allocflags
|= __GFP_COMP
;
1020 #elif defined(HAVE_KMEM_CACHE_GFPFLAGS)
1021 skc
->skc_linux_cache
->gfpflags
|= __GFP_COMP
;
1023 skc
->skc_flags
|= KMC_NOMAGAZINE
;
1026 if (spl_kmem_cache_expire
& KMC_EXPIRE_AGE
)
1027 skc
->skc_taskqid
= taskq_dispatch_delay(spl_kmem_cache_taskq
,
1028 spl_cache_age
, skc
, TQ_SLEEP
,
1029 ddi_get_lbolt() + skc
->skc_delay
/ 3 * HZ
);
1031 down_write(&spl_kmem_cache_sem
);
1032 list_add_tail(&skc
->skc_list
, &spl_kmem_cache_list
);
1033 up_write(&spl_kmem_cache_sem
);
1037 kfree(skc
->skc_name
);
1041 EXPORT_SYMBOL(spl_kmem_cache_create
);
1044 * Register a move callback for cache defragmentation.
1045 * XXX: Unimplemented but harmless to stub out for now.
1048 spl_kmem_cache_set_move(spl_kmem_cache_t
*skc
,
1049 kmem_cbrc_t (move
)(void *, void *, size_t, void *))
1051 ASSERT(move
!= NULL
);
1053 EXPORT_SYMBOL(spl_kmem_cache_set_move
);
1056 * Destroy a cache and all objects associated with the cache.
1059 spl_kmem_cache_destroy(spl_kmem_cache_t
*skc
)
1061 DECLARE_WAIT_QUEUE_HEAD(wq
);
1064 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1065 ASSERT(skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
| KMC_SLAB
));
1067 down_write(&spl_kmem_cache_sem
);
1068 list_del_init(&skc
->skc_list
);
1069 up_write(&spl_kmem_cache_sem
);
1071 /* Cancel any and wait for any pending delayed tasks */
1072 VERIFY(!test_and_set_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1074 spin_lock(&skc
->skc_lock
);
1075 id
= skc
->skc_taskqid
;
1076 spin_unlock(&skc
->skc_lock
);
1078 taskq_cancel_id(spl_kmem_cache_taskq
, id
);
1081 * Wait until all current callers complete, this is mainly
1082 * to catch the case where a low memory situation triggers a
1083 * cache reaping action which races with this destroy.
1085 wait_event(wq
, atomic_read(&skc
->skc_ref
) == 0);
1087 if (skc
->skc_flags
& (KMC_KMEM
| KMC_VMEM
)) {
1088 spl_magazine_destroy(skc
);
1089 spl_slab_reclaim(skc
);
1091 ASSERT(skc
->skc_flags
& KMC_SLAB
);
1092 kmem_cache_destroy(skc
->skc_linux_cache
);
1095 spin_lock(&skc
->skc_lock
);
1098 * Validate there are no objects in use and free all the
1099 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers.
1101 ASSERT3U(skc
->skc_slab_alloc
, ==, 0);
1102 ASSERT3U(skc
->skc_obj_alloc
, ==, 0);
1103 ASSERT3U(skc
->skc_slab_total
, ==, 0);
1104 ASSERT3U(skc
->skc_obj_total
, ==, 0);
1105 ASSERT3U(skc
->skc_obj_emergency
, ==, 0);
1106 ASSERT(list_empty(&skc
->skc_complete_list
));
1108 spin_unlock(&skc
->skc_lock
);
1110 kfree(skc
->skc_name
);
1113 EXPORT_SYMBOL(spl_kmem_cache_destroy
);
1116 * Allocate an object from a slab attached to the cache. This is used to
1117 * repopulate the per-cpu magazine caches in batches when they run low.
1120 spl_cache_obj(spl_kmem_cache_t
*skc
, spl_kmem_slab_t
*sks
)
1122 spl_kmem_obj_t
*sko
;
1124 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1125 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1127 sko
= list_entry(sks
->sks_free_list
.next
, spl_kmem_obj_t
, sko_list
);
1128 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1129 ASSERT(sko
->sko_addr
!= NULL
);
1131 /* Remove from sks_free_list */
1132 list_del_init(&sko
->sko_list
);
1134 sks
->sks_age
= jiffies
;
1136 skc
->skc_obj_alloc
++;
1138 /* Track max obj usage statistics */
1139 if (skc
->skc_obj_alloc
> skc
->skc_obj_max
)
1140 skc
->skc_obj_max
= skc
->skc_obj_alloc
;
1142 /* Track max slab usage statistics */
1143 if (sks
->sks_ref
== 1) {
1144 skc
->skc_slab_alloc
++;
1146 if (skc
->skc_slab_alloc
> skc
->skc_slab_max
)
1147 skc
->skc_slab_max
= skc
->skc_slab_alloc
;
1150 return (sko
->sko_addr
);
1154 * Generic slab allocation function to run by the global work queues.
1155 * It is responsible for allocating a new slab, linking it in to the list
1156 * of partial slabs, and then waking any waiters.
1159 __spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
)
1161 spl_kmem_slab_t
*sks
;
1163 fstrans_cookie_t cookie
= spl_fstrans_mark();
1164 sks
= spl_slab_alloc(skc
, 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
);
1173 smp_mb__before_atomic();
1174 clear_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1175 smp_mb__after_atomic();
1176 wake_up_all(&skc
->skc_waitq
);
1178 spin_unlock(&skc
->skc_lock
);
1180 return (sks
== NULL
? -ENOMEM
: 0);
1184 spl_cache_grow_work(void *data
)
1186 spl_kmem_alloc_t
*ska
= (spl_kmem_alloc_t
*)data
;
1187 spl_kmem_cache_t
*skc
= ska
->ska_cache
;
1189 (void) __spl_cache_grow(skc
, ska
->ska_flags
);
1191 atomic_dec(&skc
->skc_ref
);
1192 smp_mb__before_atomic();
1193 clear_bit(KMC_BIT_GROWING
, &skc
->skc_flags
);
1194 smp_mb__after_atomic();
1200 * Returns non-zero when a new slab should be available.
1203 spl_cache_grow_wait(spl_kmem_cache_t
*skc
)
1205 return (!test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
));
1209 * No available objects on any slabs, create a new slab. Note that this
1210 * functionality is disabled for KMC_SLAB caches which are backed by the
1214 spl_cache_grow(spl_kmem_cache_t
*skc
, int flags
, void **obj
)
1216 int remaining
, rc
= 0;
1218 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1219 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1220 ASSERT((skc
->skc_flags
& KMC_SLAB
) == 0);
1225 * Before allocating a new slab wait for any reaping to complete and
1226 * then return so the local magazine can be rechecked for new objects.
1228 if (test_bit(KMC_BIT_REAPING
, &skc
->skc_flags
)) {
1229 rc
= spl_wait_on_bit(&skc
->skc_flags
, KMC_BIT_REAPING
,
1230 TASK_UNINTERRUPTIBLE
);
1231 return (rc
? rc
: -EAGAIN
);
1235 * To reduce the overhead of context switch and improve NUMA locality,
1236 * it tries to allocate a new slab in the current process context with
1237 * KM_NOSLEEP flag. If it fails, it will launch a new taskq to do the
1240 * However, this can't be applied to KVM_VMEM due to a bug that
1241 * __vmalloc() doesn't honor gfp flags in page table allocation.
1243 if (!(skc
->skc_flags
& KMC_VMEM
)) {
1244 rc
= __spl_cache_grow(skc
, flags
| KM_NOSLEEP
);
1250 * This is handled by dispatching a work request to the global work
1251 * queue. This allows us to asynchronously allocate a new slab while
1252 * retaining the ability to safely fall back to a smaller synchronous
1253 * allocations to ensure forward progress is always maintained.
1255 if (test_and_set_bit(KMC_BIT_GROWING
, &skc
->skc_flags
) == 0) {
1256 spl_kmem_alloc_t
*ska
;
1258 ska
= kmalloc(sizeof (*ska
), kmem_flags_convert(flags
));
1260 clear_bit_unlock(KMC_BIT_GROWING
, &skc
->skc_flags
);
1261 smp_mb__after_atomic();
1262 wake_up_all(&skc
->skc_waitq
);
1266 atomic_inc(&skc
->skc_ref
);
1267 ska
->ska_cache
= skc
;
1268 ska
->ska_flags
= flags
;
1269 taskq_init_ent(&ska
->ska_tqe
);
1270 taskq_dispatch_ent(spl_kmem_cache_taskq
,
1271 spl_cache_grow_work
, ska
, 0, &ska
->ska_tqe
);
1275 * The goal here is to only detect the rare case where a virtual slab
1276 * allocation has deadlocked. We must be careful to minimize the use
1277 * of emergency objects which are more expensive to track. Therefore,
1278 * we set a very long timeout for the asynchronous allocation and if
1279 * the timeout is reached the cache is flagged as deadlocked. From
1280 * this point only new emergency objects will be allocated until the
1281 * asynchronous allocation completes and clears the deadlocked flag.
1283 if (test_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
)) {
1284 rc
= spl_emergency_alloc(skc
, flags
, obj
);
1286 remaining
= wait_event_timeout(skc
->skc_waitq
,
1287 spl_cache_grow_wait(skc
), HZ
/ 10);
1290 spin_lock(&skc
->skc_lock
);
1291 if (test_bit(KMC_BIT_GROWING
, &skc
->skc_flags
)) {
1292 set_bit(KMC_BIT_DEADLOCKED
, &skc
->skc_flags
);
1293 skc
->skc_obj_deadlock
++;
1295 spin_unlock(&skc
->skc_lock
);
1305 * Refill a per-cpu magazine with objects from the slabs for this cache.
1306 * Ideally the magazine can be repopulated using existing objects which have
1307 * been released, however if we are unable to locate enough free objects new
1308 * slabs of objects will be created. On success NULL is returned, otherwise
1309 * the address of a single emergency object is returned for use by the caller.
1312 spl_cache_refill(spl_kmem_cache_t
*skc
, spl_kmem_magazine_t
*skm
, int flags
)
1314 spl_kmem_slab_t
*sks
;
1315 int count
= 0, rc
, refill
;
1318 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1319 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1321 refill
= MIN(skm
->skm_refill
, skm
->skm_size
- skm
->skm_avail
);
1322 spin_lock(&skc
->skc_lock
);
1324 while (refill
> 0) {
1325 /* No slabs available we may need to grow the cache */
1326 if (list_empty(&skc
->skc_partial_list
)) {
1327 spin_unlock(&skc
->skc_lock
);
1330 rc
= spl_cache_grow(skc
, flags
, &obj
);
1331 local_irq_disable();
1333 /* Emergency object for immediate use by caller */
1334 if (rc
== 0 && obj
!= NULL
)
1340 /* Rescheduled to different CPU skm is not local */
1341 if (skm
!= skc
->skc_mag
[smp_processor_id()])
1345 * Potentially rescheduled to the same CPU but
1346 * allocations may have occurred from this CPU while
1347 * we were sleeping so recalculate max refill.
1349 refill
= MIN(refill
, skm
->skm_size
- skm
->skm_avail
);
1351 spin_lock(&skc
->skc_lock
);
1355 /* Grab the next available slab */
1356 sks
= list_entry((&skc
->skc_partial_list
)->next
,
1357 spl_kmem_slab_t
, sks_list
);
1358 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1359 ASSERT(sks
->sks_ref
< sks
->sks_objs
);
1360 ASSERT(!list_empty(&sks
->sks_free_list
));
1363 * Consume as many objects as needed to refill the requested
1364 * cache. We must also be careful not to overfill it.
1366 while (sks
->sks_ref
< sks
->sks_objs
&& refill
-- > 0 &&
1368 ASSERT(skm
->skm_avail
< skm
->skm_size
);
1369 ASSERT(count
< skm
->skm_size
);
1370 skm
->skm_objs
[skm
->skm_avail
++] =
1371 spl_cache_obj(skc
, sks
);
1374 /* Move slab to skc_complete_list when full */
1375 if (sks
->sks_ref
== sks
->sks_objs
) {
1376 list_del(&sks
->sks_list
);
1377 list_add(&sks
->sks_list
, &skc
->skc_complete_list
);
1381 spin_unlock(&skc
->skc_lock
);
1387 * Release an object back to the slab from which it came.
1390 spl_cache_shrink(spl_kmem_cache_t
*skc
, void *obj
)
1392 spl_kmem_slab_t
*sks
= NULL
;
1393 spl_kmem_obj_t
*sko
= NULL
;
1395 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1397 sko
= spl_sko_from_obj(skc
, obj
);
1398 ASSERT(sko
->sko_magic
== SKO_MAGIC
);
1399 sks
= sko
->sko_slab
;
1400 ASSERT(sks
->sks_magic
== SKS_MAGIC
);
1401 ASSERT(sks
->sks_cache
== skc
);
1402 list_add(&sko
->sko_list
, &sks
->sks_free_list
);
1404 sks
->sks_age
= jiffies
;
1406 skc
->skc_obj_alloc
--;
1409 * Move slab to skc_partial_list when no longer full. Slabs
1410 * are added to the head to keep the partial list is quasi-full
1411 * sorted order. Fuller at the head, emptier at the tail.
1413 if (sks
->sks_ref
== (sks
->sks_objs
- 1)) {
1414 list_del(&sks
->sks_list
);
1415 list_add(&sks
->sks_list
, &skc
->skc_partial_list
);
1419 * Move empty slabs to the end of the partial list so
1420 * they can be easily found and freed during reclamation.
1422 if (sks
->sks_ref
== 0) {
1423 list_del(&sks
->sks_list
);
1424 list_add_tail(&sks
->sks_list
, &skc
->skc_partial_list
);
1425 skc
->skc_slab_alloc
--;
1430 * Allocate an object from the per-cpu magazine, or if the magazine
1431 * is empty directly allocate from a slab and repopulate the magazine.
1434 spl_kmem_cache_alloc(spl_kmem_cache_t
*skc
, int flags
)
1436 spl_kmem_magazine_t
*skm
;
1439 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
1440 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1441 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1444 * Allocate directly from a Linux slab. All optimizations are left
1445 * to the underlying cache we only need to guarantee that KM_SLEEP
1446 * callers will never fail.
1448 if (skc
->skc_flags
& KMC_SLAB
) {
1449 struct kmem_cache
*slc
= skc
->skc_linux_cache
;
1451 obj
= kmem_cache_alloc(slc
, kmem_flags_convert(flags
));
1452 } while ((obj
== NULL
) && !(flags
& KM_NOSLEEP
));
1457 local_irq_disable();
1461 * Safe to update per-cpu structure without lock, but
1462 * in the restart case we must be careful to reacquire
1463 * the local magazine since this may have changed
1464 * when we need to grow the cache.
1466 skm
= skc
->skc_mag
[smp_processor_id()];
1467 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1469 if (likely(skm
->skm_avail
)) {
1470 /* Object available in CPU cache, use it */
1471 obj
= skm
->skm_objs
[--skm
->skm_avail
];
1472 skm
->skm_age
= jiffies
;
1474 obj
= spl_cache_refill(skc
, skm
, flags
);
1475 if ((obj
== NULL
) && !(flags
& KM_NOSLEEP
))
1484 ASSERT(IS_P2ALIGNED(obj
, skc
->skc_obj_align
));
1487 /* Pre-emptively migrate object to CPU L1 cache */
1489 if (obj
&& skc
->skc_ctor
)
1490 skc
->skc_ctor(obj
, skc
->skc_private
, flags
);
1497 EXPORT_SYMBOL(spl_kmem_cache_alloc
);
1500 * Free an object back to the local per-cpu magazine, there is no
1501 * guarantee that this is the same magazine the object was originally
1502 * allocated from. We may need to flush entire from the magazine
1503 * back to the slabs to make space.
1506 spl_kmem_cache_free(spl_kmem_cache_t
*skc
, void *obj
)
1508 spl_kmem_magazine_t
*skm
;
1509 unsigned long flags
;
1511 int do_emergency
= 0;
1513 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1514 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1517 * Run the destructor
1520 skc
->skc_dtor(obj
, skc
->skc_private
);
1523 * Free the object from the Linux underlying Linux slab.
1525 if (skc
->skc_flags
& KMC_SLAB
) {
1526 kmem_cache_free(skc
->skc_linux_cache
, obj
);
1531 * While a cache has outstanding emergency objects all freed objects
1532 * must be checked. However, since emergency objects will never use
1533 * a virtual address these objects can be safely excluded as an
1536 if (!is_vmalloc_addr(obj
)) {
1537 spin_lock(&skc
->skc_lock
);
1538 do_emergency
= (skc
->skc_obj_emergency
> 0);
1539 spin_unlock(&skc
->skc_lock
);
1541 if (do_emergency
&& (spl_emergency_free(skc
, obj
) == 0))
1545 local_irq_save(flags
);
1548 * Safe to update per-cpu structure without lock, but
1549 * no remote memory allocation tracking is being performed
1550 * it is entirely possible to allocate an object from one
1551 * CPU cache and return it to another.
1553 skm
= skc
->skc_mag
[smp_processor_id()];
1554 ASSERT(skm
->skm_magic
== SKM_MAGIC
);
1557 * Per-CPU cache full, flush it to make space for this object,
1558 * this may result in an empty slab which can be reclaimed once
1559 * interrupts are re-enabled.
1561 if (unlikely(skm
->skm_avail
>= skm
->skm_size
)) {
1562 spl_cache_flush(skc
, skm
, skm
->skm_refill
);
1566 /* Available space in cache, use it */
1567 skm
->skm_objs
[skm
->skm_avail
++] = obj
;
1569 local_irq_restore(flags
);
1572 spl_slab_reclaim(skc
);
1574 EXPORT_SYMBOL(spl_kmem_cache_free
);
1577 * The generic shrinker function for all caches. Under Linux a shrinker
1578 * may not be tightly coupled with a slab cache. In fact Linux always
1579 * systematically tries calling all registered shrinker callbacks which
1580 * report that they contain unused objects. Because of this we only
1581 * register one shrinker function in the shim layer for all slab caches.
1582 * We always attempt to shrink all caches when this generic shrinker
1585 * If sc->nr_to_scan is zero, the caller is requesting a query of the
1586 * number of objects which can potentially be freed. If it is nonzero,
1587 * the request is to free that many objects.
1589 * Linux kernels >= 3.12 have the count_objects and scan_objects callbacks
1590 * in struct shrinker and also require the shrinker to return the number
1593 * Older kernels require the shrinker to return the number of freeable
1594 * objects following the freeing of nr_to_free.
1596 * Linux semantics differ from those under Solaris, which are to
1597 * free all available objects which may (and probably will) be more
1598 * objects than the requested nr_to_scan.
1600 static spl_shrinker_t
1601 __spl_kmem_cache_generic_shrinker(struct shrinker
*shrink
,
1602 struct shrink_control
*sc
)
1604 spl_kmem_cache_t
*skc
;
1608 * No shrinking in a transaction context. Can cause deadlocks.
1610 if (sc
->nr_to_scan
&& spl_fstrans_check())
1611 return (SHRINK_STOP
);
1613 down_read(&spl_kmem_cache_sem
);
1614 list_for_each_entry(skc
, &spl_kmem_cache_list
, skc_list
) {
1615 if (sc
->nr_to_scan
) {
1616 #ifdef HAVE_SPLIT_SHRINKER_CALLBACK
1617 uint64_t oldalloc
= skc
->skc_obj_alloc
;
1618 spl_kmem_cache_reap_now(skc
,
1619 MAX(sc
->nr_to_scan
>>fls64(skc
->skc_slab_objs
), 1));
1620 if (oldalloc
> skc
->skc_obj_alloc
)
1621 alloc
+= oldalloc
- skc
->skc_obj_alloc
;
1623 spl_kmem_cache_reap_now(skc
,
1624 MAX(sc
->nr_to_scan
>>fls64(skc
->skc_slab_objs
), 1));
1625 alloc
+= skc
->skc_obj_alloc
;
1626 #endif /* HAVE_SPLIT_SHRINKER_CALLBACK */
1628 /* Request to query number of freeable objects */
1629 alloc
+= skc
->skc_obj_alloc
;
1632 up_read(&spl_kmem_cache_sem
);
1635 * When KMC_RECLAIM_ONCE is set allow only a single reclaim pass.
1636 * This functionality only exists to work around a rare issue where
1637 * shrink_slabs() is repeatedly invoked by many cores causing the
1640 if ((spl_kmem_cache_reclaim
& KMC_RECLAIM_ONCE
) && sc
->nr_to_scan
)
1641 return (SHRINK_STOP
);
1643 return (MAX(alloc
, 0));
1646 SPL_SHRINKER_CALLBACK_WRAPPER(spl_kmem_cache_generic_shrinker
);
1649 * Call the registered reclaim function for a cache. Depending on how
1650 * many and which objects are released it may simply repopulate the
1651 * local magazine which will then need to age-out. Objects which cannot
1652 * fit in the magazine we will be released back to their slabs which will
1653 * also need to age out before being release. This is all just best
1654 * effort and we do not want to thrash creating and destroying slabs.
1657 spl_kmem_cache_reap_now(spl_kmem_cache_t
*skc
, int count
)
1659 ASSERT(skc
->skc_magic
== SKC_MAGIC
);
1660 ASSERT(!test_bit(KMC_BIT_DESTROY
, &skc
->skc_flags
));
1662 atomic_inc(&skc
->skc_ref
);
1665 * Execute the registered reclaim callback if it exists.
1667 if (skc
->skc_flags
& KMC_SLAB
) {
1668 if (skc
->skc_reclaim
)
1669 skc
->skc_reclaim(skc
->skc_private
);
1674 * Prevent concurrent cache reaping when contended.
1676 if (test_and_set_bit(KMC_BIT_REAPING
, &skc
->skc_flags
))
1680 * When a reclaim function is available it may be invoked repeatedly
1681 * until at least a single slab can be freed. This ensures that we
1682 * do free memory back to the system. This helps minimize the chance
1683 * of an OOM event when the bulk of memory is used by the slab.
1685 * When free slabs are already available the reclaim callback will be
1686 * skipped. Additionally, if no forward progress is detected despite
1687 * a reclaim function the cache will be skipped to avoid deadlock.
1689 * Longer term this would be the correct place to add the code which
1690 * repacks the slabs in order minimize fragmentation.
1692 if (skc
->skc_reclaim
) {
1693 uint64_t objects
= UINT64_MAX
;
1697 spin_lock(&skc
->skc_lock
);
1699 (skc
->skc_slab_total
> 0) &&
1700 ((skc
->skc_slab_total
-skc
->skc_slab_alloc
) == 0) &&
1701 (skc
->skc_obj_alloc
< objects
);
1703 objects
= skc
->skc_obj_alloc
;
1704 spin_unlock(&skc
->skc_lock
);
1707 skc
->skc_reclaim(skc
->skc_private
);
1709 } while (do_reclaim
);
1712 /* Reclaim from the magazine and free all now empty slabs. */
1713 if (spl_kmem_cache_expire
& KMC_EXPIRE_MEM
) {
1714 spl_kmem_magazine_t
*skm
;
1715 unsigned long irq_flags
;
1717 local_irq_save(irq_flags
);
1718 skm
= skc
->skc_mag
[smp_processor_id()];
1719 spl_cache_flush(skc
, skm
, skm
->skm_avail
);
1720 local_irq_restore(irq_flags
);
1723 spl_slab_reclaim(skc
);
1724 clear_bit_unlock(KMC_BIT_REAPING
, &skc
->skc_flags
);
1725 smp_mb__after_atomic();
1726 wake_up_bit(&skc
->skc_flags
, KMC_BIT_REAPING
);
1728 atomic_dec(&skc
->skc_ref
);
1730 EXPORT_SYMBOL(spl_kmem_cache_reap_now
);
1733 * Reap all free slabs from all registered caches.
1738 struct shrink_control sc
;
1740 sc
.nr_to_scan
= KMC_REAP_CHUNK
;
1741 sc
.gfp_mask
= GFP_KERNEL
;
1743 (void) __spl_kmem_cache_generic_shrinker(NULL
, &sc
);
1745 EXPORT_SYMBOL(spl_kmem_reap
);
1748 spl_kmem_cache_init(void)
1750 init_rwsem(&spl_kmem_cache_sem
);
1751 INIT_LIST_HEAD(&spl_kmem_cache_list
);
1752 spl_kmem_cache_taskq
= taskq_create("spl_kmem_cache",
1753 spl_kmem_cache_kmem_threads
, maxclsyspri
,
1754 spl_kmem_cache_kmem_threads
* 8, INT_MAX
,
1755 TASKQ_PREPOPULATE
| TASKQ_DYNAMIC
);
1756 spl_register_shrinker(&spl_kmem_cache_shrinker
);
1762 spl_kmem_cache_fini(void)
1764 spl_unregister_shrinker(&spl_kmem_cache_shrinker
);
1765 taskq_destroy(spl_kmem_cache_taskq
);