xref: /freebsd/sys/contrib/openzfs/module/os/linux/spl/spl-kmem-cache.c (revision 1719886f6d08408b834d270c59ffcfd821c8f63a)
1 /*
2  *  Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3  *  Copyright (C) 2007 The Regents of the University of California.
4  *  Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5  *  Written by Brian Behlendorf <behlendorf1@llnl.gov>.
6  *  UCRL-CODE-235197
7  *
8  *  This file is part of the SPL, Solaris Porting Layer.
9  *
10  *  The SPL is free software; you can redistribute it and/or modify it
11  *  under the terms of the GNU General Public License as published by the
12  *  Free Software Foundation; either version 2 of the License, or (at your
13  *  option) any later version.
14  *
15  *  The SPL is distributed in the hope that it will be useful, but WITHOUT
16  *  ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17  *  FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
18  *  for more details.
19  *
20  *  You should have received a copy of the GNU General Public License along
21  *  with the SPL.  If not, see <http://www.gnu.org/licenses/>.
22  */
23 
24 #include <linux/percpu_compat.h>
25 #include <sys/kmem.h>
26 #include <sys/kmem_cache.h>
27 #include <sys/taskq.h>
28 #include <sys/timer.h>
29 #include <sys/vmem.h>
30 #include <sys/wait.h>
31 #include <sys/string.h>
32 #include <linux/slab.h>
33 #include <linux/swap.h>
34 #include <linux/prefetch.h>
35 
36 /*
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.
39  */
40 #undef kmem_cache_destroy
41 #undef kmem_cache_create
42 #undef kmem_cache_alloc
43 #undef kmem_cache_free
44 
45 
46 /*
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.
51  */
52 #ifndef smp_mb__before_atomic
53 #define	smp_mb__before_atomic(x) smp_mb__before_clear_bit(x)
54 #endif
55 
56 #ifndef smp_mb__after_atomic
57 #define	smp_mb__after_atomic(x) smp_mb__after_clear_bit(x)
58 #endif
59 
60 /* BEGIN CSTYLED */
61 /*
62  * Cache magazines are an optimization designed to minimize the cost of
63  * allocating memory.  They do this by keeping a per-cpu cache of recently
64  * freed objects, which can then be reallocated without taking a lock. This
65  * can improve performance on highly contended caches.  However, because
66  * objects in magazines will prevent otherwise empty slabs from being
67  * immediately released this may not be ideal for low memory machines.
68  *
69  * For this reason spl_kmem_cache_magazine_size can be used to set a maximum
70  * magazine size.  When this value is set to 0 the magazine size will be
71  * automatically determined based on the object size.  Otherwise magazines
72  * will be limited to 2-256 objects per magazine (i.e per cpu).  Magazines
73  * may never be entirely disabled in this implementation.
74  */
75 static unsigned int spl_kmem_cache_magazine_size = 0;
76 module_param(spl_kmem_cache_magazine_size, uint, 0444);
77 MODULE_PARM_DESC(spl_kmem_cache_magazine_size,
78 	"Default magazine size (2-256), set automatically (0)");
79 
80 static unsigned int spl_kmem_cache_obj_per_slab = SPL_KMEM_CACHE_OBJ_PER_SLAB;
81 module_param(spl_kmem_cache_obj_per_slab, uint, 0644);
82 MODULE_PARM_DESC(spl_kmem_cache_obj_per_slab, "Number of objects per slab");
83 
84 static unsigned int spl_kmem_cache_max_size = SPL_KMEM_CACHE_MAX_SIZE;
85 module_param(spl_kmem_cache_max_size, uint, 0644);
86 MODULE_PARM_DESC(spl_kmem_cache_max_size, "Maximum size of slab in MB");
87 
88 /*
89  * For small objects the Linux slab allocator should be used to make the most
90  * efficient use of the memory.  However, large objects are not supported by
91  * the Linux slab and therefore the SPL implementation is preferred.  A cutoff
92  * of 16K was determined to be optimal for architectures using 4K pages and
93  * to also work well on architecutres using larger 64K page sizes.
94  */
95 static unsigned int spl_kmem_cache_slab_limit =
96     SPL_MAX_KMEM_ORDER_NR_PAGES * PAGE_SIZE;
97 module_param(spl_kmem_cache_slab_limit, uint, 0644);
98 MODULE_PARM_DESC(spl_kmem_cache_slab_limit,
99 	"Objects less than N bytes use the Linux slab");
100 
101 /*
102  * The number of threads available to allocate new slabs for caches.  This
103  * should not need to be tuned but it is available for performance analysis.
104  */
105 static unsigned int spl_kmem_cache_kmem_threads = 4;
106 module_param(spl_kmem_cache_kmem_threads, uint, 0444);
107 MODULE_PARM_DESC(spl_kmem_cache_kmem_threads,
108 	"Number of spl_kmem_cache threads");
109 /* END CSTYLED */
110 
111 /*
112  * Slab allocation interfaces
113  *
114  * While the Linux slab implementation was inspired by the Solaris
115  * implementation I cannot use it to emulate the Solaris APIs.  I
116  * require two features which are not provided by the Linux slab.
117  *
118  * 1) Constructors AND destructors.  Recent versions of the Linux
119  *    kernel have removed support for destructors.  This is a deal
120  *    breaker for the SPL which contains particularly expensive
121  *    initializers for mutex's, condition variables, etc.  We also
122  *    require a minimal level of cleanup for these data types unlike
123  *    many Linux data types which do need to be explicitly destroyed.
124  *
125  * 2) Virtual address space backed slab.  Callers of the Solaris slab
126  *    expect it to work well for both small are very large allocations.
127  *    Because of memory fragmentation the Linux slab which is backed
128  *    by kmalloc'ed memory performs very badly when confronted with
129  *    large numbers of large allocations.  Basing the slab on the
130  *    virtual address space removes the need for contiguous pages
131  *    and greatly improve performance for large allocations.
132  *
133  * For these reasons, the SPL has its own slab implementation with
134  * the needed features.  It is not as highly optimized as either the
135  * Solaris or Linux slabs, but it should get me most of what is
136  * needed until it can be optimized or obsoleted by another approach.
137  *
138  * One serious concern I do have about this method is the relatively
139  * small virtual address space on 32bit arches.  This will seriously
140  * constrain the size of the slab caches and their performance.
141  */
142 
143 struct list_head spl_kmem_cache_list;   /* List of caches */
144 struct rw_semaphore spl_kmem_cache_sem; /* Cache list lock */
145 static taskq_t *spl_kmem_cache_taskq;   /* Task queue for aging / reclaim */
146 
147 static void spl_cache_shrink(spl_kmem_cache_t *skc, void *obj);
148 
149 static void *
150 kv_alloc(spl_kmem_cache_t *skc, int size, int flags)
151 {
152 	gfp_t lflags = kmem_flags_convert(flags);
153 	void *ptr;
154 
155 	ptr = spl_vmalloc(size, lflags | __GFP_HIGHMEM);
156 
157 	/* Resulting allocated memory will be page aligned */
158 	ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE));
159 
160 	return (ptr);
161 }
162 
163 static void
164 kv_free(spl_kmem_cache_t *skc, void *ptr, int size)
165 {
166 	ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE));
167 
168 	/*
169 	 * The Linux direct reclaim path uses this out of band value to
170 	 * determine if forward progress is being made.  Normally this is
171 	 * incremented by kmem_freepages() which is part of the various
172 	 * Linux slab implementations.  However, since we are using none
173 	 * of that infrastructure we are responsible for incrementing it.
174 	 */
175 	if (current->reclaim_state)
176 #ifdef	HAVE_RECLAIM_STATE_RECLAIMED
177 		current->reclaim_state->reclaimed += size >> PAGE_SHIFT;
178 #else
179 		current->reclaim_state->reclaimed_slab += size >> PAGE_SHIFT;
180 #endif
181 	vfree(ptr);
182 }
183 
184 /*
185  * Required space for each aligned sks.
186  */
187 static inline uint32_t
188 spl_sks_size(spl_kmem_cache_t *skc)
189 {
190 	return (P2ROUNDUP_TYPED(sizeof (spl_kmem_slab_t),
191 	    skc->skc_obj_align, uint32_t));
192 }
193 
194 /*
195  * Required space for each aligned object.
196  */
197 static inline uint32_t
198 spl_obj_size(spl_kmem_cache_t *skc)
199 {
200 	uint32_t align = skc->skc_obj_align;
201 
202 	return (P2ROUNDUP_TYPED(skc->skc_obj_size, align, uint32_t) +
203 	    P2ROUNDUP_TYPED(sizeof (spl_kmem_obj_t), align, uint32_t));
204 }
205 
206 uint64_t
207 spl_kmem_cache_inuse(kmem_cache_t *cache)
208 {
209 	return (cache->skc_obj_total);
210 }
211 EXPORT_SYMBOL(spl_kmem_cache_inuse);
212 
213 uint64_t
214 spl_kmem_cache_entry_size(kmem_cache_t *cache)
215 {
216 	return (cache->skc_obj_size);
217 }
218 EXPORT_SYMBOL(spl_kmem_cache_entry_size);
219 
220 /*
221  * Lookup the spl_kmem_object_t for an object given that object.
222  */
223 static inline spl_kmem_obj_t *
224 spl_sko_from_obj(spl_kmem_cache_t *skc, void *obj)
225 {
226 	return (obj + P2ROUNDUP_TYPED(skc->skc_obj_size,
227 	    skc->skc_obj_align, uint32_t));
228 }
229 
230 /*
231  * It's important that we pack the spl_kmem_obj_t structure and the
232  * actual objects in to one large address space to minimize the number
233  * of calls to the allocator.  It is far better to do a few large
234  * allocations and then subdivide it ourselves.  Now which allocator
235  * we use requires balancing a few trade offs.
236  *
237  * For small objects we use kmem_alloc() because as long as you are
238  * only requesting a small number of pages (ideally just one) its cheap.
239  * However, when you start requesting multiple pages with kmem_alloc()
240  * it gets increasingly expensive since it requires contiguous pages.
241  * For this reason we shift to vmem_alloc() for slabs of large objects
242  * which removes the need for contiguous pages.  We do not use
243  * vmem_alloc() in all cases because there is significant locking
244  * overhead in __get_vm_area_node().  This function takes a single
245  * global lock when acquiring an available virtual address range which
246  * serializes all vmem_alloc()'s for all slab caches.  Using slightly
247  * different allocation functions for small and large objects should
248  * give us the best of both worlds.
249  *
250  * +------------------------+
251  * | spl_kmem_slab_t --+-+  |
252  * | skc_obj_size    <-+ |  |
253  * | spl_kmem_obj_t      |  |
254  * | skc_obj_size    <---+  |
255  * | spl_kmem_obj_t      |  |
256  * | ...                 v  |
257  * +------------------------+
258  */
259 static spl_kmem_slab_t *
260 spl_slab_alloc(spl_kmem_cache_t *skc, int flags)
261 {
262 	spl_kmem_slab_t *sks;
263 	void *base;
264 	uint32_t obj_size;
265 
266 	base = kv_alloc(skc, skc->skc_slab_size, flags);
267 	if (base == NULL)
268 		return (NULL);
269 
270 	sks = (spl_kmem_slab_t *)base;
271 	sks->sks_magic = SKS_MAGIC;
272 	sks->sks_objs = skc->skc_slab_objs;
273 	sks->sks_age = jiffies;
274 	sks->sks_cache = skc;
275 	INIT_LIST_HEAD(&sks->sks_list);
276 	INIT_LIST_HEAD(&sks->sks_free_list);
277 	sks->sks_ref = 0;
278 	obj_size = spl_obj_size(skc);
279 
280 	for (int i = 0; i < sks->sks_objs; i++) {
281 		void *obj = base + spl_sks_size(skc) + (i * obj_size);
282 
283 		ASSERT(IS_P2ALIGNED(obj, skc->skc_obj_align));
284 		spl_kmem_obj_t *sko = spl_sko_from_obj(skc, obj);
285 		sko->sko_addr = obj;
286 		sko->sko_magic = SKO_MAGIC;
287 		sko->sko_slab = sks;
288 		INIT_LIST_HEAD(&sko->sko_list);
289 		list_add_tail(&sko->sko_list, &sks->sks_free_list);
290 	}
291 
292 	return (sks);
293 }
294 
295 /*
296  * Remove a slab from complete or partial list, it must be called with
297  * the 'skc->skc_lock' held but the actual free must be performed
298  * outside the lock to prevent deadlocking on vmem addresses.
299  */
300 static void
301 spl_slab_free(spl_kmem_slab_t *sks,
302     struct list_head *sks_list, struct list_head *sko_list)
303 {
304 	spl_kmem_cache_t *skc;
305 
306 	ASSERT(sks->sks_magic == SKS_MAGIC);
307 	ASSERT(sks->sks_ref == 0);
308 
309 	skc = sks->sks_cache;
310 	ASSERT(skc->skc_magic == SKC_MAGIC);
311 
312 	/*
313 	 * Update slab/objects counters in the cache, then remove the
314 	 * slab from the skc->skc_partial_list.  Finally add the slab
315 	 * and all its objects in to the private work lists where the
316 	 * destructors will be called and the memory freed to the system.
317 	 */
318 	skc->skc_obj_total -= sks->sks_objs;
319 	skc->skc_slab_total--;
320 	list_del(&sks->sks_list);
321 	list_add(&sks->sks_list, sks_list);
322 	list_splice_init(&sks->sks_free_list, sko_list);
323 }
324 
325 /*
326  * Reclaim empty slabs at the end of the partial list.
327  */
328 static void
329 spl_slab_reclaim(spl_kmem_cache_t *skc)
330 {
331 	spl_kmem_slab_t *sks = NULL, *m = NULL;
332 	spl_kmem_obj_t *sko = NULL, *n = NULL;
333 	LIST_HEAD(sks_list);
334 	LIST_HEAD(sko_list);
335 
336 	/*
337 	 * Empty slabs and objects must be moved to a private list so they
338 	 * can be safely freed outside the spin lock.  All empty slabs are
339 	 * at the end of skc->skc_partial_list, therefore once a non-empty
340 	 * slab is found we can stop scanning.
341 	 */
342 	spin_lock(&skc->skc_lock);
343 	list_for_each_entry_safe_reverse(sks, m,
344 	    &skc->skc_partial_list, sks_list) {
345 
346 		if (sks->sks_ref > 0)
347 			break;
348 
349 		spl_slab_free(sks, &sks_list, &sko_list);
350 	}
351 	spin_unlock(&skc->skc_lock);
352 
353 	/*
354 	 * The following two loops ensure all the object destructors are run,
355 	 * and the slabs themselves are freed.  This is all done outside the
356 	 * skc->skc_lock since this allows the destructor to sleep, and
357 	 * allows us to perform a conditional reschedule when a freeing a
358 	 * large number of objects and slabs back to the system.
359 	 */
360 
361 	list_for_each_entry_safe(sko, n, &sko_list, sko_list) {
362 		ASSERT(sko->sko_magic == SKO_MAGIC);
363 	}
364 
365 	list_for_each_entry_safe(sks, m, &sks_list, sks_list) {
366 		ASSERT(sks->sks_magic == SKS_MAGIC);
367 		kv_free(skc, sks, skc->skc_slab_size);
368 	}
369 }
370 
371 static spl_kmem_emergency_t *
372 spl_emergency_search(struct rb_root *root, void *obj)
373 {
374 	struct rb_node *node = root->rb_node;
375 	spl_kmem_emergency_t *ske;
376 	unsigned long address = (unsigned long)obj;
377 
378 	while (node) {
379 		ske = container_of(node, spl_kmem_emergency_t, ske_node);
380 
381 		if (address < ske->ske_obj)
382 			node = node->rb_left;
383 		else if (address > ske->ske_obj)
384 			node = node->rb_right;
385 		else
386 			return (ske);
387 	}
388 
389 	return (NULL);
390 }
391 
392 static int
393 spl_emergency_insert(struct rb_root *root, spl_kmem_emergency_t *ske)
394 {
395 	struct rb_node **new = &(root->rb_node), *parent = NULL;
396 	spl_kmem_emergency_t *ske_tmp;
397 	unsigned long address = ske->ske_obj;
398 
399 	while (*new) {
400 		ske_tmp = container_of(*new, spl_kmem_emergency_t, ske_node);
401 
402 		parent = *new;
403 		if (address < ske_tmp->ske_obj)
404 			new = &((*new)->rb_left);
405 		else if (address > ske_tmp->ske_obj)
406 			new = &((*new)->rb_right);
407 		else
408 			return (0);
409 	}
410 
411 	rb_link_node(&ske->ske_node, parent, new);
412 	rb_insert_color(&ske->ske_node, root);
413 
414 	return (1);
415 }
416 
417 /*
418  * Allocate a single emergency object and track it in a red black tree.
419  */
420 static int
421 spl_emergency_alloc(spl_kmem_cache_t *skc, int flags, void **obj)
422 {
423 	gfp_t lflags = kmem_flags_convert(flags);
424 	spl_kmem_emergency_t *ske;
425 	int order = get_order(skc->skc_obj_size);
426 	int empty;
427 
428 	/* Last chance use a partial slab if one now exists */
429 	spin_lock(&skc->skc_lock);
430 	empty = list_empty(&skc->skc_partial_list);
431 	spin_unlock(&skc->skc_lock);
432 	if (!empty)
433 		return (-EEXIST);
434 
435 	ske = kmalloc(sizeof (*ske), lflags);
436 	if (ske == NULL)
437 		return (-ENOMEM);
438 
439 	ske->ske_obj = __get_free_pages(lflags, order);
440 	if (ske->ske_obj == 0) {
441 		kfree(ske);
442 		return (-ENOMEM);
443 	}
444 
445 	spin_lock(&skc->skc_lock);
446 	empty = spl_emergency_insert(&skc->skc_emergency_tree, ske);
447 	if (likely(empty)) {
448 		skc->skc_obj_total++;
449 		skc->skc_obj_emergency++;
450 		if (skc->skc_obj_emergency > skc->skc_obj_emergency_max)
451 			skc->skc_obj_emergency_max = skc->skc_obj_emergency;
452 	}
453 	spin_unlock(&skc->skc_lock);
454 
455 	if (unlikely(!empty)) {
456 		free_pages(ske->ske_obj, order);
457 		kfree(ske);
458 		return (-EINVAL);
459 	}
460 
461 	*obj = (void *)ske->ske_obj;
462 
463 	return (0);
464 }
465 
466 /*
467  * Locate the passed object in the red black tree and free it.
468  */
469 static int
470 spl_emergency_free(spl_kmem_cache_t *skc, void *obj)
471 {
472 	spl_kmem_emergency_t *ske;
473 	int order = get_order(skc->skc_obj_size);
474 
475 	spin_lock(&skc->skc_lock);
476 	ske = spl_emergency_search(&skc->skc_emergency_tree, obj);
477 	if (ske) {
478 		rb_erase(&ske->ske_node, &skc->skc_emergency_tree);
479 		skc->skc_obj_emergency--;
480 		skc->skc_obj_total--;
481 	}
482 	spin_unlock(&skc->skc_lock);
483 
484 	if (ske == NULL)
485 		return (-ENOENT);
486 
487 	free_pages(ske->ske_obj, order);
488 	kfree(ske);
489 
490 	return (0);
491 }
492 
493 /*
494  * Release objects from the per-cpu magazine back to their slab.  The flush
495  * argument contains the max number of entries to remove from the magazine.
496  */
497 static void
498 spl_cache_flush(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flush)
499 {
500 	spin_lock(&skc->skc_lock);
501 
502 	ASSERT(skc->skc_magic == SKC_MAGIC);
503 	ASSERT(skm->skm_magic == SKM_MAGIC);
504 
505 	int count = MIN(flush, skm->skm_avail);
506 	for (int i = 0; i < count; i++)
507 		spl_cache_shrink(skc, skm->skm_objs[i]);
508 
509 	skm->skm_avail -= count;
510 	memmove(skm->skm_objs, &(skm->skm_objs[count]),
511 	    sizeof (void *) * skm->skm_avail);
512 
513 	spin_unlock(&skc->skc_lock);
514 }
515 
516 /*
517  * Size a slab based on the size of each aligned object plus spl_kmem_obj_t.
518  * When on-slab we want to target spl_kmem_cache_obj_per_slab.  However,
519  * for very small objects we may end up with more than this so as not
520  * to waste space in the minimal allocation of a single page.
521  */
522 static int
523 spl_slab_size(spl_kmem_cache_t *skc, uint32_t *objs, uint32_t *size)
524 {
525 	uint32_t sks_size, obj_size, max_size, tgt_size, tgt_objs;
526 
527 	sks_size = spl_sks_size(skc);
528 	obj_size = spl_obj_size(skc);
529 	max_size = (spl_kmem_cache_max_size * 1024 * 1024);
530 	tgt_size = (spl_kmem_cache_obj_per_slab * obj_size + sks_size);
531 
532 	if (tgt_size <= max_size) {
533 		tgt_objs = (tgt_size - sks_size) / obj_size;
534 	} else {
535 		tgt_objs = (max_size - sks_size) / obj_size;
536 		tgt_size = (tgt_objs * obj_size) + sks_size;
537 	}
538 
539 	if (tgt_objs == 0)
540 		return (-ENOSPC);
541 
542 	*objs = tgt_objs;
543 	*size = tgt_size;
544 
545 	return (0);
546 }
547 
548 /*
549  * Make a guess at reasonable per-cpu magazine size based on the size of
550  * each object and the cost of caching N of them in each magazine.  Long
551  * term this should really adapt based on an observed usage heuristic.
552  */
553 static int
554 spl_magazine_size(spl_kmem_cache_t *skc)
555 {
556 	uint32_t obj_size = spl_obj_size(skc);
557 	int size;
558 
559 	if (spl_kmem_cache_magazine_size > 0)
560 		return (MAX(MIN(spl_kmem_cache_magazine_size, 256), 2));
561 
562 	/* Per-magazine sizes below assume a 4Kib page size */
563 	if (obj_size > (PAGE_SIZE * 256))
564 		size = 4;  /* Minimum 4Mib per-magazine */
565 	else if (obj_size > (PAGE_SIZE * 32))
566 		size = 16; /* Minimum 2Mib per-magazine */
567 	else if (obj_size > (PAGE_SIZE))
568 		size = 64; /* Minimum 256Kib per-magazine */
569 	else if (obj_size > (PAGE_SIZE / 4))
570 		size = 128; /* Minimum 128Kib per-magazine */
571 	else
572 		size = 256;
573 
574 	return (size);
575 }
576 
577 /*
578  * Allocate a per-cpu magazine to associate with a specific core.
579  */
580 static spl_kmem_magazine_t *
581 spl_magazine_alloc(spl_kmem_cache_t *skc, int cpu)
582 {
583 	spl_kmem_magazine_t *skm;
584 	int size = sizeof (spl_kmem_magazine_t) +
585 	    sizeof (void *) * skc->skc_mag_size;
586 
587 	skm = kmalloc_node(size, GFP_KERNEL, cpu_to_node(cpu));
588 	if (skm) {
589 		skm->skm_magic = SKM_MAGIC;
590 		skm->skm_avail = 0;
591 		skm->skm_size = skc->skc_mag_size;
592 		skm->skm_refill = skc->skc_mag_refill;
593 		skm->skm_cache = skc;
594 		skm->skm_cpu = cpu;
595 	}
596 
597 	return (skm);
598 }
599 
600 /*
601  * Free a per-cpu magazine associated with a specific core.
602  */
603 static void
604 spl_magazine_free(spl_kmem_magazine_t *skm)
605 {
606 	ASSERT(skm->skm_magic == SKM_MAGIC);
607 	ASSERT(skm->skm_avail == 0);
608 	kfree(skm);
609 }
610 
611 /*
612  * Create all pre-cpu magazines of reasonable sizes.
613  */
614 static int
615 spl_magazine_create(spl_kmem_cache_t *skc)
616 {
617 	int i = 0;
618 
619 	ASSERT((skc->skc_flags & KMC_SLAB) == 0);
620 
621 	skc->skc_mag = kzalloc(sizeof (spl_kmem_magazine_t *) *
622 	    num_possible_cpus(), kmem_flags_convert(KM_SLEEP));
623 	skc->skc_mag_size = spl_magazine_size(skc);
624 	skc->skc_mag_refill = (skc->skc_mag_size + 1) / 2;
625 
626 	for_each_possible_cpu(i) {
627 		skc->skc_mag[i] = spl_magazine_alloc(skc, i);
628 		if (!skc->skc_mag[i]) {
629 			for (i--; i >= 0; i--)
630 				spl_magazine_free(skc->skc_mag[i]);
631 
632 			kfree(skc->skc_mag);
633 			return (-ENOMEM);
634 		}
635 	}
636 
637 	return (0);
638 }
639 
640 /*
641  * Destroy all pre-cpu magazines.
642  */
643 static void
644 spl_magazine_destroy(spl_kmem_cache_t *skc)
645 {
646 	spl_kmem_magazine_t *skm;
647 	int i = 0;
648 
649 	ASSERT((skc->skc_flags & KMC_SLAB) == 0);
650 
651 	for_each_possible_cpu(i) {
652 		skm = skc->skc_mag[i];
653 		spl_cache_flush(skc, skm, skm->skm_avail);
654 		spl_magazine_free(skm);
655 	}
656 
657 	kfree(skc->skc_mag);
658 }
659 
660 /*
661  * Create a object cache based on the following arguments:
662  * name		cache name
663  * size		cache object size
664  * align	cache object alignment
665  * ctor		cache object constructor
666  * dtor		cache object destructor
667  * reclaim	cache object reclaim
668  * priv		cache private data for ctor/dtor/reclaim
669  * vmp		unused must be NULL
670  * flags
671  *	KMC_KVMEM       Force kvmem backed SPL cache
672  *	KMC_SLAB        Force Linux slab backed cache
673  *	KMC_NODEBUG	Disable debugging (unsupported)
674  */
675 spl_kmem_cache_t *
676 spl_kmem_cache_create(const char *name, size_t size, size_t align,
677     spl_kmem_ctor_t ctor, spl_kmem_dtor_t dtor, void *reclaim,
678     void *priv, void *vmp, int flags)
679 {
680 	gfp_t lflags = kmem_flags_convert(KM_SLEEP);
681 	spl_kmem_cache_t *skc;
682 	int rc;
683 
684 	/*
685 	 * Unsupported flags
686 	 */
687 	ASSERT(vmp == NULL);
688 	ASSERT(reclaim == NULL);
689 
690 	might_sleep();
691 
692 	skc = kzalloc(sizeof (*skc), lflags);
693 	if (skc == NULL)
694 		return (NULL);
695 
696 	skc->skc_magic = SKC_MAGIC;
697 	skc->skc_name_size = strlen(name) + 1;
698 	skc->skc_name = kmalloc(skc->skc_name_size, lflags);
699 	if (skc->skc_name == NULL) {
700 		kfree(skc);
701 		return (NULL);
702 	}
703 	strlcpy(skc->skc_name, name, skc->skc_name_size);
704 
705 	skc->skc_ctor = ctor;
706 	skc->skc_dtor = dtor;
707 	skc->skc_private = priv;
708 	skc->skc_vmp = vmp;
709 	skc->skc_linux_cache = NULL;
710 	skc->skc_flags = flags;
711 	skc->skc_obj_size = size;
712 	skc->skc_obj_align = SPL_KMEM_CACHE_ALIGN;
713 	atomic_set(&skc->skc_ref, 0);
714 
715 	INIT_LIST_HEAD(&skc->skc_list);
716 	INIT_LIST_HEAD(&skc->skc_complete_list);
717 	INIT_LIST_HEAD(&skc->skc_partial_list);
718 	skc->skc_emergency_tree = RB_ROOT;
719 	spin_lock_init(&skc->skc_lock);
720 	init_waitqueue_head(&skc->skc_waitq);
721 	skc->skc_slab_fail = 0;
722 	skc->skc_slab_create = 0;
723 	skc->skc_slab_destroy = 0;
724 	skc->skc_slab_total = 0;
725 	skc->skc_slab_alloc = 0;
726 	skc->skc_slab_max = 0;
727 	skc->skc_obj_total = 0;
728 	skc->skc_obj_alloc = 0;
729 	skc->skc_obj_max = 0;
730 	skc->skc_obj_deadlock = 0;
731 	skc->skc_obj_emergency = 0;
732 	skc->skc_obj_emergency_max = 0;
733 
734 	rc = percpu_counter_init_common(&skc->skc_linux_alloc, 0,
735 	    GFP_KERNEL);
736 	if (rc != 0) {
737 		kfree(skc);
738 		return (NULL);
739 	}
740 
741 	/*
742 	 * Verify the requested alignment restriction is sane.
743 	 */
744 	if (align) {
745 		VERIFY(ISP2(align));
746 		VERIFY3U(align, >=, SPL_KMEM_CACHE_ALIGN);
747 		VERIFY3U(align, <=, PAGE_SIZE);
748 		skc->skc_obj_align = align;
749 	}
750 
751 	/*
752 	 * When no specific type of slab is requested (kmem, vmem, or
753 	 * linuxslab) then select a cache type based on the object size
754 	 * and default tunables.
755 	 */
756 	if (!(skc->skc_flags & (KMC_SLAB | KMC_KVMEM))) {
757 		if (spl_kmem_cache_slab_limit &&
758 		    size <= (size_t)spl_kmem_cache_slab_limit) {
759 			/*
760 			 * Objects smaller than spl_kmem_cache_slab_limit can
761 			 * use the Linux slab for better space-efficiency.
762 			 */
763 			skc->skc_flags |= KMC_SLAB;
764 		} else {
765 			/*
766 			 * All other objects are considered large and are
767 			 * placed on kvmem backed slabs.
768 			 */
769 			skc->skc_flags |= KMC_KVMEM;
770 		}
771 	}
772 
773 	/*
774 	 * Given the type of slab allocate the required resources.
775 	 */
776 	if (skc->skc_flags & KMC_KVMEM) {
777 		rc = spl_slab_size(skc,
778 		    &skc->skc_slab_objs, &skc->skc_slab_size);
779 		if (rc)
780 			goto out;
781 
782 		rc = spl_magazine_create(skc);
783 		if (rc)
784 			goto out;
785 	} else {
786 		unsigned long slabflags = 0;
787 
788 		if (size > spl_kmem_cache_slab_limit)
789 			goto out;
790 
791 #if defined(SLAB_USERCOPY)
792 		/*
793 		 * Required for PAX-enabled kernels if the slab is to be
794 		 * used for copying between user and kernel space.
795 		 */
796 		slabflags |= SLAB_USERCOPY;
797 #endif
798 
799 #if defined(HAVE_KMEM_CACHE_CREATE_USERCOPY)
800 		/*
801 		 * Newer grsec patchset uses kmem_cache_create_usercopy()
802 		 * instead of SLAB_USERCOPY flag
803 		 */
804 		skc->skc_linux_cache = kmem_cache_create_usercopy(
805 		    skc->skc_name, size, align, slabflags, 0, size, NULL);
806 #else
807 		skc->skc_linux_cache = kmem_cache_create(
808 		    skc->skc_name, size, align, slabflags, NULL);
809 #endif
810 		if (skc->skc_linux_cache == NULL)
811 			goto out;
812 	}
813 
814 	down_write(&spl_kmem_cache_sem);
815 	list_add_tail(&skc->skc_list, &spl_kmem_cache_list);
816 	up_write(&spl_kmem_cache_sem);
817 
818 	return (skc);
819 out:
820 	kfree(skc->skc_name);
821 	percpu_counter_destroy(&skc->skc_linux_alloc);
822 	kfree(skc);
823 	return (NULL);
824 }
825 EXPORT_SYMBOL(spl_kmem_cache_create);
826 
827 /*
828  * Register a move callback for cache defragmentation.
829  * XXX: Unimplemented but harmless to stub out for now.
830  */
831 void
832 spl_kmem_cache_set_move(spl_kmem_cache_t *skc,
833     kmem_cbrc_t (move)(void *, void *, size_t, void *))
834 {
835 	ASSERT(move != NULL);
836 }
837 EXPORT_SYMBOL(spl_kmem_cache_set_move);
838 
839 /*
840  * Destroy a cache and all objects associated with the cache.
841  */
842 void
843 spl_kmem_cache_destroy(spl_kmem_cache_t *skc)
844 {
845 	DECLARE_WAIT_QUEUE_HEAD(wq);
846 	taskqid_t id;
847 
848 	ASSERT(skc->skc_magic == SKC_MAGIC);
849 	ASSERT(skc->skc_flags & (KMC_KVMEM | KMC_SLAB));
850 
851 	down_write(&spl_kmem_cache_sem);
852 	list_del_init(&skc->skc_list);
853 	up_write(&spl_kmem_cache_sem);
854 
855 	/* Cancel any and wait for any pending delayed tasks */
856 	VERIFY(!test_and_set_bit(KMC_BIT_DESTROY, &skc->skc_flags));
857 
858 	spin_lock(&skc->skc_lock);
859 	id = skc->skc_taskqid;
860 	spin_unlock(&skc->skc_lock);
861 
862 	taskq_cancel_id(spl_kmem_cache_taskq, id);
863 
864 	/*
865 	 * Wait until all current callers complete, this is mainly
866 	 * to catch the case where a low memory situation triggers a
867 	 * cache reaping action which races with this destroy.
868 	 */
869 	wait_event(wq, atomic_read(&skc->skc_ref) == 0);
870 
871 	if (skc->skc_flags & KMC_KVMEM) {
872 		spl_magazine_destroy(skc);
873 		spl_slab_reclaim(skc);
874 	} else {
875 		ASSERT(skc->skc_flags & KMC_SLAB);
876 		kmem_cache_destroy(skc->skc_linux_cache);
877 	}
878 
879 	spin_lock(&skc->skc_lock);
880 
881 	/*
882 	 * Validate there are no objects in use and free all the
883 	 * spl_kmem_slab_t, spl_kmem_obj_t, and object buffers.
884 	 */
885 	ASSERT3U(skc->skc_slab_alloc, ==, 0);
886 	ASSERT3U(skc->skc_obj_alloc, ==, 0);
887 	ASSERT3U(skc->skc_slab_total, ==, 0);
888 	ASSERT3U(skc->skc_obj_total, ==, 0);
889 	ASSERT3U(skc->skc_obj_emergency, ==, 0);
890 	ASSERT(list_empty(&skc->skc_complete_list));
891 
892 	ASSERT3U(percpu_counter_sum(&skc->skc_linux_alloc), ==, 0);
893 	percpu_counter_destroy(&skc->skc_linux_alloc);
894 
895 	spin_unlock(&skc->skc_lock);
896 
897 	kfree(skc->skc_name);
898 	kfree(skc);
899 }
900 EXPORT_SYMBOL(spl_kmem_cache_destroy);
901 
902 /*
903  * Allocate an object from a slab attached to the cache.  This is used to
904  * repopulate the per-cpu magazine caches in batches when they run low.
905  */
906 static void *
907 spl_cache_obj(spl_kmem_cache_t *skc, spl_kmem_slab_t *sks)
908 {
909 	spl_kmem_obj_t *sko;
910 
911 	ASSERT(skc->skc_magic == SKC_MAGIC);
912 	ASSERT(sks->sks_magic == SKS_MAGIC);
913 
914 	sko = list_entry(sks->sks_free_list.next, spl_kmem_obj_t, sko_list);
915 	ASSERT(sko->sko_magic == SKO_MAGIC);
916 	ASSERT(sko->sko_addr != NULL);
917 
918 	/* Remove from sks_free_list */
919 	list_del_init(&sko->sko_list);
920 
921 	sks->sks_age = jiffies;
922 	sks->sks_ref++;
923 	skc->skc_obj_alloc++;
924 
925 	/* Track max obj usage statistics */
926 	if (skc->skc_obj_alloc > skc->skc_obj_max)
927 		skc->skc_obj_max = skc->skc_obj_alloc;
928 
929 	/* Track max slab usage statistics */
930 	if (sks->sks_ref == 1) {
931 		skc->skc_slab_alloc++;
932 
933 		if (skc->skc_slab_alloc > skc->skc_slab_max)
934 			skc->skc_slab_max = skc->skc_slab_alloc;
935 	}
936 
937 	return (sko->sko_addr);
938 }
939 
940 /*
941  * Generic slab allocation function to run by the global work queues.
942  * It is responsible for allocating a new slab, linking it in to the list
943  * of partial slabs, and then waking any waiters.
944  */
945 static int
946 __spl_cache_grow(spl_kmem_cache_t *skc, int flags)
947 {
948 	spl_kmem_slab_t *sks;
949 
950 	fstrans_cookie_t cookie = spl_fstrans_mark();
951 	sks = spl_slab_alloc(skc, flags);
952 	spl_fstrans_unmark(cookie);
953 
954 	spin_lock(&skc->skc_lock);
955 	if (sks) {
956 		skc->skc_slab_total++;
957 		skc->skc_obj_total += sks->sks_objs;
958 		list_add_tail(&sks->sks_list, &skc->skc_partial_list);
959 
960 		smp_mb__before_atomic();
961 		clear_bit(KMC_BIT_DEADLOCKED, &skc->skc_flags);
962 		smp_mb__after_atomic();
963 	}
964 	spin_unlock(&skc->skc_lock);
965 
966 	return (sks == NULL ? -ENOMEM : 0);
967 }
968 
969 static void
970 spl_cache_grow_work(void *data)
971 {
972 	spl_kmem_alloc_t *ska = (spl_kmem_alloc_t *)data;
973 	spl_kmem_cache_t *skc = ska->ska_cache;
974 
975 	int error = __spl_cache_grow(skc, ska->ska_flags);
976 
977 	atomic_dec(&skc->skc_ref);
978 	smp_mb__before_atomic();
979 	clear_bit(KMC_BIT_GROWING, &skc->skc_flags);
980 	smp_mb__after_atomic();
981 	if (error == 0)
982 		wake_up_all(&skc->skc_waitq);
983 
984 	kfree(ska);
985 }
986 
987 /*
988  * Returns non-zero when a new slab should be available.
989  */
990 static int
991 spl_cache_grow_wait(spl_kmem_cache_t *skc)
992 {
993 	return (!test_bit(KMC_BIT_GROWING, &skc->skc_flags));
994 }
995 
996 /*
997  * No available objects on any slabs, create a new slab.  Note that this
998  * functionality is disabled for KMC_SLAB caches which are backed by the
999  * Linux slab.
1000  */
1001 static int
1002 spl_cache_grow(spl_kmem_cache_t *skc, int flags, void **obj)
1003 {
1004 	int remaining, rc = 0;
1005 
1006 	ASSERT0(flags & ~KM_PUBLIC_MASK);
1007 	ASSERT(skc->skc_magic == SKC_MAGIC);
1008 	ASSERT((skc->skc_flags & KMC_SLAB) == 0);
1009 
1010 	*obj = NULL;
1011 
1012 	/*
1013 	 * Since we can't sleep attempt an emergency allocation to satisfy
1014 	 * the request.  The only alterative is to fail the allocation but
1015 	 * it's preferable try.  The use of KM_NOSLEEP is expected to be rare.
1016 	 */
1017 	if (flags & KM_NOSLEEP)
1018 		return (spl_emergency_alloc(skc, flags, obj));
1019 
1020 	might_sleep();
1021 
1022 	/*
1023 	 * Before allocating a new slab wait for any reaping to complete and
1024 	 * then return so the local magazine can be rechecked for new objects.
1025 	 */
1026 	if (test_bit(KMC_BIT_REAPING, &skc->skc_flags)) {
1027 		rc = spl_wait_on_bit(&skc->skc_flags, KMC_BIT_REAPING,
1028 		    TASK_UNINTERRUPTIBLE);
1029 		return (rc ? rc : -EAGAIN);
1030 	}
1031 
1032 	/*
1033 	 * Note: It would be nice to reduce the overhead of context switch
1034 	 * and improve NUMA locality, by trying to allocate a new slab in the
1035 	 * current process context with KM_NOSLEEP flag.
1036 	 *
1037 	 * However, this can't be applied to vmem/kvmem due to a bug that
1038 	 * spl_vmalloc() doesn't honor gfp flags in page table allocation.
1039 	 */
1040 
1041 	/*
1042 	 * This is handled by dispatching a work request to the global work
1043 	 * queue.  This allows us to asynchronously allocate a new slab while
1044 	 * retaining the ability to safely fall back to a smaller synchronous
1045 	 * allocations to ensure forward progress is always maintained.
1046 	 */
1047 	if (test_and_set_bit(KMC_BIT_GROWING, &skc->skc_flags) == 0) {
1048 		spl_kmem_alloc_t *ska;
1049 
1050 		ska = kmalloc(sizeof (*ska), kmem_flags_convert(flags));
1051 		if (ska == NULL) {
1052 			clear_bit_unlock(KMC_BIT_GROWING, &skc->skc_flags);
1053 			smp_mb__after_atomic();
1054 			wake_up_all(&skc->skc_waitq);
1055 			return (-ENOMEM);
1056 		}
1057 
1058 		atomic_inc(&skc->skc_ref);
1059 		ska->ska_cache = skc;
1060 		ska->ska_flags = flags;
1061 		taskq_init_ent(&ska->ska_tqe);
1062 		taskq_dispatch_ent(spl_kmem_cache_taskq,
1063 		    spl_cache_grow_work, ska, 0, &ska->ska_tqe);
1064 	}
1065 
1066 	/*
1067 	 * The goal here is to only detect the rare case where a virtual slab
1068 	 * allocation has deadlocked.  We must be careful to minimize the use
1069 	 * of emergency objects which are more expensive to track.  Therefore,
1070 	 * we set a very long timeout for the asynchronous allocation and if
1071 	 * the timeout is reached the cache is flagged as deadlocked.  From
1072 	 * this point only new emergency objects will be allocated until the
1073 	 * asynchronous allocation completes and clears the deadlocked flag.
1074 	 */
1075 	if (test_bit(KMC_BIT_DEADLOCKED, &skc->skc_flags)) {
1076 		rc = spl_emergency_alloc(skc, flags, obj);
1077 	} else {
1078 		remaining = wait_event_timeout(skc->skc_waitq,
1079 		    spl_cache_grow_wait(skc), HZ / 10);
1080 
1081 		if (!remaining) {
1082 			spin_lock(&skc->skc_lock);
1083 			if (test_bit(KMC_BIT_GROWING, &skc->skc_flags)) {
1084 				set_bit(KMC_BIT_DEADLOCKED, &skc->skc_flags);
1085 				skc->skc_obj_deadlock++;
1086 			}
1087 			spin_unlock(&skc->skc_lock);
1088 		}
1089 
1090 		rc = -ENOMEM;
1091 	}
1092 
1093 	return (rc);
1094 }
1095 
1096 /*
1097  * Refill a per-cpu magazine with objects from the slabs for this cache.
1098  * Ideally the magazine can be repopulated using existing objects which have
1099  * been released, however if we are unable to locate enough free objects new
1100  * slabs of objects will be created.  On success NULL is returned, otherwise
1101  * the address of a single emergency object is returned for use by the caller.
1102  */
1103 static void *
1104 spl_cache_refill(spl_kmem_cache_t *skc, spl_kmem_magazine_t *skm, int flags)
1105 {
1106 	spl_kmem_slab_t *sks;
1107 	int count = 0, rc, refill;
1108 	void *obj = NULL;
1109 
1110 	ASSERT(skc->skc_magic == SKC_MAGIC);
1111 	ASSERT(skm->skm_magic == SKM_MAGIC);
1112 
1113 	refill = MIN(skm->skm_refill, skm->skm_size - skm->skm_avail);
1114 	spin_lock(&skc->skc_lock);
1115 
1116 	while (refill > 0) {
1117 		/* No slabs available we may need to grow the cache */
1118 		if (list_empty(&skc->skc_partial_list)) {
1119 			spin_unlock(&skc->skc_lock);
1120 
1121 			local_irq_enable();
1122 			rc = spl_cache_grow(skc, flags, &obj);
1123 			local_irq_disable();
1124 
1125 			/* Emergency object for immediate use by caller */
1126 			if (rc == 0 && obj != NULL)
1127 				return (obj);
1128 
1129 			if (rc)
1130 				goto out;
1131 
1132 			/* Rescheduled to different CPU skm is not local */
1133 			if (skm != skc->skc_mag[smp_processor_id()])
1134 				goto out;
1135 
1136 			/*
1137 			 * Potentially rescheduled to the same CPU but
1138 			 * allocations may have occurred from this CPU while
1139 			 * we were sleeping so recalculate max refill.
1140 			 */
1141 			refill = MIN(refill, skm->skm_size - skm->skm_avail);
1142 
1143 			spin_lock(&skc->skc_lock);
1144 			continue;
1145 		}
1146 
1147 		/* Grab the next available slab */
1148 		sks = list_entry((&skc->skc_partial_list)->next,
1149 		    spl_kmem_slab_t, sks_list);
1150 		ASSERT(sks->sks_magic == SKS_MAGIC);
1151 		ASSERT(sks->sks_ref < sks->sks_objs);
1152 		ASSERT(!list_empty(&sks->sks_free_list));
1153 
1154 		/*
1155 		 * Consume as many objects as needed to refill the requested
1156 		 * cache.  We must also be careful not to overfill it.
1157 		 */
1158 		while (sks->sks_ref < sks->sks_objs && refill-- > 0 &&
1159 		    ++count) {
1160 			ASSERT(skm->skm_avail < skm->skm_size);
1161 			ASSERT(count < skm->skm_size);
1162 			skm->skm_objs[skm->skm_avail++] =
1163 			    spl_cache_obj(skc, sks);
1164 		}
1165 
1166 		/* Move slab to skc_complete_list when full */
1167 		if (sks->sks_ref == sks->sks_objs) {
1168 			list_del(&sks->sks_list);
1169 			list_add(&sks->sks_list, &skc->skc_complete_list);
1170 		}
1171 	}
1172 
1173 	spin_unlock(&skc->skc_lock);
1174 out:
1175 	return (NULL);
1176 }
1177 
1178 /*
1179  * Release an object back to the slab from which it came.
1180  */
1181 static void
1182 spl_cache_shrink(spl_kmem_cache_t *skc, void *obj)
1183 {
1184 	spl_kmem_slab_t *sks = NULL;
1185 	spl_kmem_obj_t *sko = NULL;
1186 
1187 	ASSERT(skc->skc_magic == SKC_MAGIC);
1188 
1189 	sko = spl_sko_from_obj(skc, obj);
1190 	ASSERT(sko->sko_magic == SKO_MAGIC);
1191 	sks = sko->sko_slab;
1192 	ASSERT(sks->sks_magic == SKS_MAGIC);
1193 	ASSERT(sks->sks_cache == skc);
1194 	list_add(&sko->sko_list, &sks->sks_free_list);
1195 
1196 	sks->sks_age = jiffies;
1197 	sks->sks_ref--;
1198 	skc->skc_obj_alloc--;
1199 
1200 	/*
1201 	 * Move slab to skc_partial_list when no longer full.  Slabs
1202 	 * are added to the head to keep the partial list is quasi-full
1203 	 * sorted order.  Fuller at the head, emptier at the tail.
1204 	 */
1205 	if (sks->sks_ref == (sks->sks_objs - 1)) {
1206 		list_del(&sks->sks_list);
1207 		list_add(&sks->sks_list, &skc->skc_partial_list);
1208 	}
1209 
1210 	/*
1211 	 * Move empty slabs to the end of the partial list so
1212 	 * they can be easily found and freed during reclamation.
1213 	 */
1214 	if (sks->sks_ref == 0) {
1215 		list_del(&sks->sks_list);
1216 		list_add_tail(&sks->sks_list, &skc->skc_partial_list);
1217 		skc->skc_slab_alloc--;
1218 	}
1219 }
1220 
1221 /*
1222  * Allocate an object from the per-cpu magazine, or if the magazine
1223  * is empty directly allocate from a slab and repopulate the magazine.
1224  */
1225 void *
1226 spl_kmem_cache_alloc(spl_kmem_cache_t *skc, int flags)
1227 {
1228 	spl_kmem_magazine_t *skm;
1229 	void *obj = NULL;
1230 
1231 	ASSERT0(flags & ~KM_PUBLIC_MASK);
1232 	ASSERT(skc->skc_magic == SKC_MAGIC);
1233 	ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags));
1234 
1235 	/*
1236 	 * Allocate directly from a Linux slab.  All optimizations are left
1237 	 * to the underlying cache we only need to guarantee that KM_SLEEP
1238 	 * callers will never fail.
1239 	 */
1240 	if (skc->skc_flags & KMC_SLAB) {
1241 		struct kmem_cache *slc = skc->skc_linux_cache;
1242 		do {
1243 			obj = kmem_cache_alloc(slc, kmem_flags_convert(flags));
1244 		} while ((obj == NULL) && !(flags & KM_NOSLEEP));
1245 
1246 		if (obj != NULL) {
1247 			/*
1248 			 * Even though we leave everything up to the
1249 			 * underlying cache we still keep track of
1250 			 * how many objects we've allocated in it for
1251 			 * better debuggability.
1252 			 */
1253 			percpu_counter_inc(&skc->skc_linux_alloc);
1254 		}
1255 		goto ret;
1256 	}
1257 
1258 	local_irq_disable();
1259 
1260 restart:
1261 	/*
1262 	 * Safe to update per-cpu structure without lock, but
1263 	 * in the restart case we must be careful to reacquire
1264 	 * the local magazine since this may have changed
1265 	 * when we need to grow the cache.
1266 	 */
1267 	skm = skc->skc_mag[smp_processor_id()];
1268 	ASSERT(skm->skm_magic == SKM_MAGIC);
1269 
1270 	if (likely(skm->skm_avail)) {
1271 		/* Object available in CPU cache, use it */
1272 		obj = skm->skm_objs[--skm->skm_avail];
1273 	} else {
1274 		obj = spl_cache_refill(skc, skm, flags);
1275 		if ((obj == NULL) && !(flags & KM_NOSLEEP))
1276 			goto restart;
1277 
1278 		local_irq_enable();
1279 		goto ret;
1280 	}
1281 
1282 	local_irq_enable();
1283 	ASSERT(obj);
1284 	ASSERT(IS_P2ALIGNED(obj, skc->skc_obj_align));
1285 
1286 ret:
1287 	/* Pre-emptively migrate object to CPU L1 cache */
1288 	if (obj) {
1289 		if (obj && skc->skc_ctor)
1290 			skc->skc_ctor(obj, skc->skc_private, flags);
1291 		else
1292 			prefetchw(obj);
1293 	}
1294 
1295 	return (obj);
1296 }
1297 EXPORT_SYMBOL(spl_kmem_cache_alloc);
1298 
1299 /*
1300  * Free an object back to the local per-cpu magazine, there is no
1301  * guarantee that this is the same magazine the object was originally
1302  * allocated from.  We may need to flush entire from the magazine
1303  * back to the slabs to make space.
1304  */
1305 void
1306 spl_kmem_cache_free(spl_kmem_cache_t *skc, void *obj)
1307 {
1308 	spl_kmem_magazine_t *skm;
1309 	unsigned long flags;
1310 	int do_reclaim = 0;
1311 	int do_emergency = 0;
1312 
1313 	ASSERT(skc->skc_magic == SKC_MAGIC);
1314 	ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags));
1315 
1316 	/*
1317 	 * Run the destructor
1318 	 */
1319 	if (skc->skc_dtor)
1320 		skc->skc_dtor(obj, skc->skc_private);
1321 
1322 	/*
1323 	 * Free the object from the Linux underlying Linux slab.
1324 	 */
1325 	if (skc->skc_flags & KMC_SLAB) {
1326 		kmem_cache_free(skc->skc_linux_cache, obj);
1327 		percpu_counter_dec(&skc->skc_linux_alloc);
1328 		return;
1329 	}
1330 
1331 	/*
1332 	 * While a cache has outstanding emergency objects all freed objects
1333 	 * must be checked.  However, since emergency objects will never use
1334 	 * a virtual address these objects can be safely excluded as an
1335 	 * optimization.
1336 	 */
1337 	if (!is_vmalloc_addr(obj)) {
1338 		spin_lock(&skc->skc_lock);
1339 		do_emergency = (skc->skc_obj_emergency > 0);
1340 		spin_unlock(&skc->skc_lock);
1341 
1342 		if (do_emergency && (spl_emergency_free(skc, obj) == 0))
1343 			return;
1344 	}
1345 
1346 	local_irq_save(flags);
1347 
1348 	/*
1349 	 * Safe to update per-cpu structure without lock, but
1350 	 * no remote memory allocation tracking is being performed
1351 	 * it is entirely possible to allocate an object from one
1352 	 * CPU cache and return it to another.
1353 	 */
1354 	skm = skc->skc_mag[smp_processor_id()];
1355 	ASSERT(skm->skm_magic == SKM_MAGIC);
1356 
1357 	/*
1358 	 * Per-CPU cache full, flush it to make space for this object,
1359 	 * this may result in an empty slab which can be reclaimed once
1360 	 * interrupts are re-enabled.
1361 	 */
1362 	if (unlikely(skm->skm_avail >= skm->skm_size)) {
1363 		spl_cache_flush(skc, skm, skm->skm_refill);
1364 		do_reclaim = 1;
1365 	}
1366 
1367 	/* Available space in cache, use it */
1368 	skm->skm_objs[skm->skm_avail++] = obj;
1369 
1370 	local_irq_restore(flags);
1371 
1372 	if (do_reclaim)
1373 		spl_slab_reclaim(skc);
1374 }
1375 EXPORT_SYMBOL(spl_kmem_cache_free);
1376 
1377 /*
1378  * Depending on how many and which objects are released it may simply
1379  * repopulate the local magazine which will then need to age-out.  Objects
1380  * which cannot fit in the magazine will be released back to their slabs
1381  * which will also need to age out before being released.  This is all just
1382  * best effort and we do not want to thrash creating and destroying slabs.
1383  */
1384 void
1385 spl_kmem_cache_reap_now(spl_kmem_cache_t *skc)
1386 {
1387 	ASSERT(skc->skc_magic == SKC_MAGIC);
1388 	ASSERT(!test_bit(KMC_BIT_DESTROY, &skc->skc_flags));
1389 
1390 	if (skc->skc_flags & KMC_SLAB)
1391 		return;
1392 
1393 	atomic_inc(&skc->skc_ref);
1394 
1395 	/*
1396 	 * Prevent concurrent cache reaping when contended.
1397 	 */
1398 	if (test_and_set_bit(KMC_BIT_REAPING, &skc->skc_flags))
1399 		goto out;
1400 
1401 	/* Reclaim from the magazine and free all now empty slabs. */
1402 	unsigned long irq_flags;
1403 	local_irq_save(irq_flags);
1404 	spl_kmem_magazine_t *skm = skc->skc_mag[smp_processor_id()];
1405 	spl_cache_flush(skc, skm, skm->skm_avail);
1406 	local_irq_restore(irq_flags);
1407 
1408 	spl_slab_reclaim(skc);
1409 	clear_bit_unlock(KMC_BIT_REAPING, &skc->skc_flags);
1410 	smp_mb__after_atomic();
1411 	wake_up_bit(&skc->skc_flags, KMC_BIT_REAPING);
1412 out:
1413 	atomic_dec(&skc->skc_ref);
1414 }
1415 EXPORT_SYMBOL(spl_kmem_cache_reap_now);
1416 
1417 /*
1418  * This is stubbed out for code consistency with other platforms.  There
1419  * is existing logic to prevent concurrent reaping so while this is ugly
1420  * it should do no harm.
1421  */
1422 int
1423 spl_kmem_cache_reap_active(void)
1424 {
1425 	return (0);
1426 }
1427 EXPORT_SYMBOL(spl_kmem_cache_reap_active);
1428 
1429 /*
1430  * Reap all free slabs from all registered caches.
1431  */
1432 void
1433 spl_kmem_reap(void)
1434 {
1435 	spl_kmem_cache_t *skc = NULL;
1436 
1437 	down_read(&spl_kmem_cache_sem);
1438 	list_for_each_entry(skc, &spl_kmem_cache_list, skc_list) {
1439 		spl_kmem_cache_reap_now(skc);
1440 	}
1441 	up_read(&spl_kmem_cache_sem);
1442 }
1443 EXPORT_SYMBOL(spl_kmem_reap);
1444 
1445 int
1446 spl_kmem_cache_init(void)
1447 {
1448 	init_rwsem(&spl_kmem_cache_sem);
1449 	INIT_LIST_HEAD(&spl_kmem_cache_list);
1450 	spl_kmem_cache_taskq = taskq_create("spl_kmem_cache",
1451 	    spl_kmem_cache_kmem_threads, maxclsyspri,
1452 	    spl_kmem_cache_kmem_threads * 8, INT_MAX,
1453 	    TASKQ_PREPOPULATE | TASKQ_DYNAMIC);
1454 
1455 	if (spl_kmem_cache_taskq == NULL)
1456 		return (-ENOMEM);
1457 
1458 	return (0);
1459 }
1460 
1461 void
1462 spl_kmem_cache_fini(void)
1463 {
1464 	taskq_destroy(spl_kmem_cache_taskq);
1465 }
1466