Loading...
1/*
2 * zswap.c - zswap driver file
3 *
4 * zswap is a backend for frontswap that takes pages that are in the process
5 * of being swapped out and attempts to compress and store them in a
6 * RAM-based memory pool. This can result in a significant I/O reduction on
7 * the swap device and, in the case where decompressing from RAM is faster
8 * than reading from the swap device, can also improve workload performance.
9 *
10 * Copyright (C) 2012 Seth Jennings <sjenning@linux.vnet.ibm.com>
11 *
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version 2
15 * of the License, or (at your option) any later version.
16 *
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
21*/
22
23#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
24
25#include <linux/module.h>
26#include <linux/cpu.h>
27#include <linux/highmem.h>
28#include <linux/slab.h>
29#include <linux/spinlock.h>
30#include <linux/types.h>
31#include <linux/atomic.h>
32#include <linux/frontswap.h>
33#include <linux/rbtree.h>
34#include <linux/swap.h>
35#include <linux/crypto.h>
36#include <linux/mempool.h>
37#include <linux/zpool.h>
38
39#include <linux/mm_types.h>
40#include <linux/page-flags.h>
41#include <linux/swapops.h>
42#include <linux/writeback.h>
43#include <linux/pagemap.h>
44
45/*********************************
46* statistics
47**********************************/
48/* Total bytes used by the compressed storage */
49static u64 zswap_pool_total_size;
50/* The number of compressed pages currently stored in zswap */
51static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
52
53/*
54 * The statistics below are not protected from concurrent access for
55 * performance reasons so they may not be a 100% accurate. However,
56 * they do provide useful information on roughly how many times a
57 * certain event is occurring.
58*/
59
60/* Pool limit was hit (see zswap_max_pool_percent) */
61static u64 zswap_pool_limit_hit;
62/* Pages written back when pool limit was reached */
63static u64 zswap_written_back_pages;
64/* Store failed due to a reclaim failure after pool limit was reached */
65static u64 zswap_reject_reclaim_fail;
66/* Compressed page was too big for the allocator to (optimally) store */
67static u64 zswap_reject_compress_poor;
68/* Store failed because underlying allocator could not get memory */
69static u64 zswap_reject_alloc_fail;
70/* Store failed because the entry metadata could not be allocated (rare) */
71static u64 zswap_reject_kmemcache_fail;
72/* Duplicate store was encountered (rare) */
73static u64 zswap_duplicate_entry;
74
75/*********************************
76* tunables
77**********************************/
78
79/* Enable/disable zswap (disabled by default) */
80static bool zswap_enabled;
81module_param_named(enabled, zswap_enabled, bool, 0644);
82
83/* Crypto compressor to use */
84#define ZSWAP_COMPRESSOR_DEFAULT "lzo"
85static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
86static int zswap_compressor_param_set(const char *,
87 const struct kernel_param *);
88static struct kernel_param_ops zswap_compressor_param_ops = {
89 .set = zswap_compressor_param_set,
90 .get = param_get_charp,
91 .free = param_free_charp,
92};
93module_param_cb(compressor, &zswap_compressor_param_ops,
94 &zswap_compressor, 0644);
95
96/* Compressed storage zpool to use */
97#define ZSWAP_ZPOOL_DEFAULT "zbud"
98static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
99static int zswap_zpool_param_set(const char *, const struct kernel_param *);
100static struct kernel_param_ops zswap_zpool_param_ops = {
101 .set = zswap_zpool_param_set,
102 .get = param_get_charp,
103 .free = param_free_charp,
104};
105module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
106
107/* The maximum percentage of memory that the compressed pool can occupy */
108static unsigned int zswap_max_pool_percent = 20;
109module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
110
111/*********************************
112* data structures
113**********************************/
114
115struct zswap_pool {
116 struct zpool *zpool;
117 struct crypto_comp * __percpu *tfm;
118 struct kref kref;
119 struct list_head list;
120 struct rcu_head rcu_head;
121 struct notifier_block notifier;
122 char tfm_name[CRYPTO_MAX_ALG_NAME];
123};
124
125/*
126 * struct zswap_entry
127 *
128 * This structure contains the metadata for tracking a single compressed
129 * page within zswap.
130 *
131 * rbnode - links the entry into red-black tree for the appropriate swap type
132 * offset - the swap offset for the entry. Index into the red-black tree.
133 * refcount - the number of outstanding reference to the entry. This is needed
134 * to protect against premature freeing of the entry by code
135 * concurrent calls to load, invalidate, and writeback. The lock
136 * for the zswap_tree structure that contains the entry must
137 * be held while changing the refcount. Since the lock must
138 * be held, there is no reason to also make refcount atomic.
139 * length - the length in bytes of the compressed page data. Needed during
140 * decompression
141 * pool - the zswap_pool the entry's data is in
142 * handle - zpool allocation handle that stores the compressed page data
143 */
144struct zswap_entry {
145 struct rb_node rbnode;
146 pgoff_t offset;
147 int refcount;
148 unsigned int length;
149 struct zswap_pool *pool;
150 unsigned long handle;
151};
152
153struct zswap_header {
154 swp_entry_t swpentry;
155};
156
157/*
158 * The tree lock in the zswap_tree struct protects a few things:
159 * - the rbtree
160 * - the refcount field of each entry in the tree
161 */
162struct zswap_tree {
163 struct rb_root rbroot;
164 spinlock_t lock;
165};
166
167static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
168
169/* RCU-protected iteration */
170static LIST_HEAD(zswap_pools);
171/* protects zswap_pools list modification */
172static DEFINE_SPINLOCK(zswap_pools_lock);
173/* pool counter to provide unique names to zpool */
174static atomic_t zswap_pools_count = ATOMIC_INIT(0);
175
176/* used by param callback function */
177static bool zswap_init_started;
178
179/*********************************
180* helpers and fwd declarations
181**********************************/
182
183#define zswap_pool_debug(msg, p) \
184 pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name, \
185 zpool_get_type((p)->zpool))
186
187static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
188static int zswap_pool_get(struct zswap_pool *pool);
189static void zswap_pool_put(struct zswap_pool *pool);
190
191static const struct zpool_ops zswap_zpool_ops = {
192 .evict = zswap_writeback_entry
193};
194
195static bool zswap_is_full(void)
196{
197 return totalram_pages * zswap_max_pool_percent / 100 <
198 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
199}
200
201static void zswap_update_total_size(void)
202{
203 struct zswap_pool *pool;
204 u64 total = 0;
205
206 rcu_read_lock();
207
208 list_for_each_entry_rcu(pool, &zswap_pools, list)
209 total += zpool_get_total_size(pool->zpool);
210
211 rcu_read_unlock();
212
213 zswap_pool_total_size = total;
214}
215
216/*********************************
217* zswap entry functions
218**********************************/
219static struct kmem_cache *zswap_entry_cache;
220
221static int __init zswap_entry_cache_create(void)
222{
223 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
224 return zswap_entry_cache == NULL;
225}
226
227static void __init zswap_entry_cache_destroy(void)
228{
229 kmem_cache_destroy(zswap_entry_cache);
230}
231
232static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
233{
234 struct zswap_entry *entry;
235 entry = kmem_cache_alloc(zswap_entry_cache, gfp);
236 if (!entry)
237 return NULL;
238 entry->refcount = 1;
239 RB_CLEAR_NODE(&entry->rbnode);
240 return entry;
241}
242
243static void zswap_entry_cache_free(struct zswap_entry *entry)
244{
245 kmem_cache_free(zswap_entry_cache, entry);
246}
247
248/*********************************
249* rbtree functions
250**********************************/
251static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
252{
253 struct rb_node *node = root->rb_node;
254 struct zswap_entry *entry;
255
256 while (node) {
257 entry = rb_entry(node, struct zswap_entry, rbnode);
258 if (entry->offset > offset)
259 node = node->rb_left;
260 else if (entry->offset < offset)
261 node = node->rb_right;
262 else
263 return entry;
264 }
265 return NULL;
266}
267
268/*
269 * In the case that a entry with the same offset is found, a pointer to
270 * the existing entry is stored in dupentry and the function returns -EEXIST
271 */
272static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
273 struct zswap_entry **dupentry)
274{
275 struct rb_node **link = &root->rb_node, *parent = NULL;
276 struct zswap_entry *myentry;
277
278 while (*link) {
279 parent = *link;
280 myentry = rb_entry(parent, struct zswap_entry, rbnode);
281 if (myentry->offset > entry->offset)
282 link = &(*link)->rb_left;
283 else if (myentry->offset < entry->offset)
284 link = &(*link)->rb_right;
285 else {
286 *dupentry = myentry;
287 return -EEXIST;
288 }
289 }
290 rb_link_node(&entry->rbnode, parent, link);
291 rb_insert_color(&entry->rbnode, root);
292 return 0;
293}
294
295static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
296{
297 if (!RB_EMPTY_NODE(&entry->rbnode)) {
298 rb_erase(&entry->rbnode, root);
299 RB_CLEAR_NODE(&entry->rbnode);
300 }
301}
302
303/*
304 * Carries out the common pattern of freeing and entry's zpool allocation,
305 * freeing the entry itself, and decrementing the number of stored pages.
306 */
307static void zswap_free_entry(struct zswap_entry *entry)
308{
309 zpool_free(entry->pool->zpool, entry->handle);
310 zswap_pool_put(entry->pool);
311 zswap_entry_cache_free(entry);
312 atomic_dec(&zswap_stored_pages);
313 zswap_update_total_size();
314}
315
316/* caller must hold the tree lock */
317static void zswap_entry_get(struct zswap_entry *entry)
318{
319 entry->refcount++;
320}
321
322/* caller must hold the tree lock
323* remove from the tree and free it, if nobody reference the entry
324*/
325static void zswap_entry_put(struct zswap_tree *tree,
326 struct zswap_entry *entry)
327{
328 int refcount = --entry->refcount;
329
330 BUG_ON(refcount < 0);
331 if (refcount == 0) {
332 zswap_rb_erase(&tree->rbroot, entry);
333 zswap_free_entry(entry);
334 }
335}
336
337/* caller must hold the tree lock */
338static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
339 pgoff_t offset)
340{
341 struct zswap_entry *entry;
342
343 entry = zswap_rb_search(root, offset);
344 if (entry)
345 zswap_entry_get(entry);
346
347 return entry;
348}
349
350/*********************************
351* per-cpu code
352**********************************/
353static DEFINE_PER_CPU(u8 *, zswap_dstmem);
354
355static int __zswap_cpu_dstmem_notifier(unsigned long action, unsigned long cpu)
356{
357 u8 *dst;
358
359 switch (action) {
360 case CPU_UP_PREPARE:
361 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
362 if (!dst) {
363 pr_err("can't allocate compressor buffer\n");
364 return NOTIFY_BAD;
365 }
366 per_cpu(zswap_dstmem, cpu) = dst;
367 break;
368 case CPU_DEAD:
369 case CPU_UP_CANCELED:
370 dst = per_cpu(zswap_dstmem, cpu);
371 kfree(dst);
372 per_cpu(zswap_dstmem, cpu) = NULL;
373 break;
374 default:
375 break;
376 }
377 return NOTIFY_OK;
378}
379
380static int zswap_cpu_dstmem_notifier(struct notifier_block *nb,
381 unsigned long action, void *pcpu)
382{
383 return __zswap_cpu_dstmem_notifier(action, (unsigned long)pcpu);
384}
385
386static struct notifier_block zswap_dstmem_notifier = {
387 .notifier_call = zswap_cpu_dstmem_notifier,
388};
389
390static int __init zswap_cpu_dstmem_init(void)
391{
392 unsigned long cpu;
393
394 cpu_notifier_register_begin();
395 for_each_online_cpu(cpu)
396 if (__zswap_cpu_dstmem_notifier(CPU_UP_PREPARE, cpu) ==
397 NOTIFY_BAD)
398 goto cleanup;
399 __register_cpu_notifier(&zswap_dstmem_notifier);
400 cpu_notifier_register_done();
401 return 0;
402
403cleanup:
404 for_each_online_cpu(cpu)
405 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
406 cpu_notifier_register_done();
407 return -ENOMEM;
408}
409
410static void zswap_cpu_dstmem_destroy(void)
411{
412 unsigned long cpu;
413
414 cpu_notifier_register_begin();
415 for_each_online_cpu(cpu)
416 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
417 __unregister_cpu_notifier(&zswap_dstmem_notifier);
418 cpu_notifier_register_done();
419}
420
421static int __zswap_cpu_comp_notifier(struct zswap_pool *pool,
422 unsigned long action, unsigned long cpu)
423{
424 struct crypto_comp *tfm;
425
426 switch (action) {
427 case CPU_UP_PREPARE:
428 if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
429 break;
430 tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
431 if (IS_ERR_OR_NULL(tfm)) {
432 pr_err("could not alloc crypto comp %s : %ld\n",
433 pool->tfm_name, PTR_ERR(tfm));
434 return NOTIFY_BAD;
435 }
436 *per_cpu_ptr(pool->tfm, cpu) = tfm;
437 break;
438 case CPU_DEAD:
439 case CPU_UP_CANCELED:
440 tfm = *per_cpu_ptr(pool->tfm, cpu);
441 if (!IS_ERR_OR_NULL(tfm))
442 crypto_free_comp(tfm);
443 *per_cpu_ptr(pool->tfm, cpu) = NULL;
444 break;
445 default:
446 break;
447 }
448 return NOTIFY_OK;
449}
450
451static int zswap_cpu_comp_notifier(struct notifier_block *nb,
452 unsigned long action, void *pcpu)
453{
454 unsigned long cpu = (unsigned long)pcpu;
455 struct zswap_pool *pool = container_of(nb, typeof(*pool), notifier);
456
457 return __zswap_cpu_comp_notifier(pool, action, cpu);
458}
459
460static int zswap_cpu_comp_init(struct zswap_pool *pool)
461{
462 unsigned long cpu;
463
464 memset(&pool->notifier, 0, sizeof(pool->notifier));
465 pool->notifier.notifier_call = zswap_cpu_comp_notifier;
466
467 cpu_notifier_register_begin();
468 for_each_online_cpu(cpu)
469 if (__zswap_cpu_comp_notifier(pool, CPU_UP_PREPARE, cpu) ==
470 NOTIFY_BAD)
471 goto cleanup;
472 __register_cpu_notifier(&pool->notifier);
473 cpu_notifier_register_done();
474 return 0;
475
476cleanup:
477 for_each_online_cpu(cpu)
478 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
479 cpu_notifier_register_done();
480 return -ENOMEM;
481}
482
483static void zswap_cpu_comp_destroy(struct zswap_pool *pool)
484{
485 unsigned long cpu;
486
487 cpu_notifier_register_begin();
488 for_each_online_cpu(cpu)
489 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
490 __unregister_cpu_notifier(&pool->notifier);
491 cpu_notifier_register_done();
492}
493
494/*********************************
495* pool functions
496**********************************/
497
498static struct zswap_pool *__zswap_pool_current(void)
499{
500 struct zswap_pool *pool;
501
502 pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
503 WARN_ON(!pool);
504
505 return pool;
506}
507
508static struct zswap_pool *zswap_pool_current(void)
509{
510 assert_spin_locked(&zswap_pools_lock);
511
512 return __zswap_pool_current();
513}
514
515static struct zswap_pool *zswap_pool_current_get(void)
516{
517 struct zswap_pool *pool;
518
519 rcu_read_lock();
520
521 pool = __zswap_pool_current();
522 if (!pool || !zswap_pool_get(pool))
523 pool = NULL;
524
525 rcu_read_unlock();
526
527 return pool;
528}
529
530static struct zswap_pool *zswap_pool_last_get(void)
531{
532 struct zswap_pool *pool, *last = NULL;
533
534 rcu_read_lock();
535
536 list_for_each_entry_rcu(pool, &zswap_pools, list)
537 last = pool;
538 if (!WARN_ON(!last) && !zswap_pool_get(last))
539 last = NULL;
540
541 rcu_read_unlock();
542
543 return last;
544}
545
546/* type and compressor must be null-terminated */
547static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
548{
549 struct zswap_pool *pool;
550
551 assert_spin_locked(&zswap_pools_lock);
552
553 list_for_each_entry_rcu(pool, &zswap_pools, list) {
554 if (strcmp(pool->tfm_name, compressor))
555 continue;
556 if (strcmp(zpool_get_type(pool->zpool), type))
557 continue;
558 /* if we can't get it, it's about to be destroyed */
559 if (!zswap_pool_get(pool))
560 continue;
561 return pool;
562 }
563
564 return NULL;
565}
566
567static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
568{
569 struct zswap_pool *pool;
570 char name[38]; /* 'zswap' + 32 char (max) num + \0 */
571 gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
572
573 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
574 if (!pool) {
575 pr_err("pool alloc failed\n");
576 return NULL;
577 }
578
579 /* unique name for each pool specifically required by zsmalloc */
580 snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
581
582 pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
583 if (!pool->zpool) {
584 pr_err("%s zpool not available\n", type);
585 goto error;
586 }
587 pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
588
589 strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
590 pool->tfm = alloc_percpu(struct crypto_comp *);
591 if (!pool->tfm) {
592 pr_err("percpu alloc failed\n");
593 goto error;
594 }
595
596 if (zswap_cpu_comp_init(pool))
597 goto error;
598 pr_debug("using %s compressor\n", pool->tfm_name);
599
600 /* being the current pool takes 1 ref; this func expects the
601 * caller to always add the new pool as the current pool
602 */
603 kref_init(&pool->kref);
604 INIT_LIST_HEAD(&pool->list);
605
606 zswap_pool_debug("created", pool);
607
608 return pool;
609
610error:
611 free_percpu(pool->tfm);
612 if (pool->zpool)
613 zpool_destroy_pool(pool->zpool);
614 kfree(pool);
615 return NULL;
616}
617
618static __init struct zswap_pool *__zswap_pool_create_fallback(void)
619{
620 if (!crypto_has_comp(zswap_compressor, 0, 0)) {
621 if (!strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
622 pr_err("default compressor %s not available\n",
623 zswap_compressor);
624 return NULL;
625 }
626 pr_err("compressor %s not available, using default %s\n",
627 zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
628 param_free_charp(&zswap_compressor);
629 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
630 }
631 if (!zpool_has_pool(zswap_zpool_type)) {
632 if (!strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
633 pr_err("default zpool %s not available\n",
634 zswap_zpool_type);
635 return NULL;
636 }
637 pr_err("zpool %s not available, using default %s\n",
638 zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
639 param_free_charp(&zswap_zpool_type);
640 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
641 }
642
643 return zswap_pool_create(zswap_zpool_type, zswap_compressor);
644}
645
646static void zswap_pool_destroy(struct zswap_pool *pool)
647{
648 zswap_pool_debug("destroying", pool);
649
650 zswap_cpu_comp_destroy(pool);
651 free_percpu(pool->tfm);
652 zpool_destroy_pool(pool->zpool);
653 kfree(pool);
654}
655
656static int __must_check zswap_pool_get(struct zswap_pool *pool)
657{
658 return kref_get_unless_zero(&pool->kref);
659}
660
661static void __zswap_pool_release(struct rcu_head *head)
662{
663 struct zswap_pool *pool = container_of(head, typeof(*pool), rcu_head);
664
665 /* nobody should have been able to get a kref... */
666 WARN_ON(kref_get_unless_zero(&pool->kref));
667
668 /* pool is now off zswap_pools list and has no references. */
669 zswap_pool_destroy(pool);
670}
671
672static void __zswap_pool_empty(struct kref *kref)
673{
674 struct zswap_pool *pool;
675
676 pool = container_of(kref, typeof(*pool), kref);
677
678 spin_lock(&zswap_pools_lock);
679
680 WARN_ON(pool == zswap_pool_current());
681
682 list_del_rcu(&pool->list);
683 call_rcu(&pool->rcu_head, __zswap_pool_release);
684
685 spin_unlock(&zswap_pools_lock);
686}
687
688static void zswap_pool_put(struct zswap_pool *pool)
689{
690 kref_put(&pool->kref, __zswap_pool_empty);
691}
692
693/*********************************
694* param callbacks
695**********************************/
696
697/* val must be a null-terminated string */
698static int __zswap_param_set(const char *val, const struct kernel_param *kp,
699 char *type, char *compressor)
700{
701 struct zswap_pool *pool, *put_pool = NULL;
702 char *s = strstrip((char *)val);
703 int ret;
704
705 /* no change required */
706 if (!strcmp(s, *(char **)kp->arg))
707 return 0;
708
709 /* if this is load-time (pre-init) param setting,
710 * don't create a pool; that's done during init.
711 */
712 if (!zswap_init_started)
713 return param_set_charp(s, kp);
714
715 if (!type) {
716 if (!zpool_has_pool(s)) {
717 pr_err("zpool %s not available\n", s);
718 return -ENOENT;
719 }
720 type = s;
721 } else if (!compressor) {
722 if (!crypto_has_comp(s, 0, 0)) {
723 pr_err("compressor %s not available\n", s);
724 return -ENOENT;
725 }
726 compressor = s;
727 } else {
728 WARN_ON(1);
729 return -EINVAL;
730 }
731
732 spin_lock(&zswap_pools_lock);
733
734 pool = zswap_pool_find_get(type, compressor);
735 if (pool) {
736 zswap_pool_debug("using existing", pool);
737 list_del_rcu(&pool->list);
738 } else {
739 spin_unlock(&zswap_pools_lock);
740 pool = zswap_pool_create(type, compressor);
741 spin_lock(&zswap_pools_lock);
742 }
743
744 if (pool)
745 ret = param_set_charp(s, kp);
746 else
747 ret = -EINVAL;
748
749 if (!ret) {
750 put_pool = zswap_pool_current();
751 list_add_rcu(&pool->list, &zswap_pools);
752 } else if (pool) {
753 /* add the possibly pre-existing pool to the end of the pools
754 * list; if it's new (and empty) then it'll be removed and
755 * destroyed by the put after we drop the lock
756 */
757 list_add_tail_rcu(&pool->list, &zswap_pools);
758 put_pool = pool;
759 }
760
761 spin_unlock(&zswap_pools_lock);
762
763 /* drop the ref from either the old current pool,
764 * or the new pool we failed to add
765 */
766 if (put_pool)
767 zswap_pool_put(put_pool);
768
769 return ret;
770}
771
772static int zswap_compressor_param_set(const char *val,
773 const struct kernel_param *kp)
774{
775 return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
776}
777
778static int zswap_zpool_param_set(const char *val,
779 const struct kernel_param *kp)
780{
781 return __zswap_param_set(val, kp, NULL, zswap_compressor);
782}
783
784/*********************************
785* writeback code
786**********************************/
787/* return enum for zswap_get_swap_cache_page */
788enum zswap_get_swap_ret {
789 ZSWAP_SWAPCACHE_NEW,
790 ZSWAP_SWAPCACHE_EXIST,
791 ZSWAP_SWAPCACHE_FAIL,
792};
793
794/*
795 * zswap_get_swap_cache_page
796 *
797 * This is an adaption of read_swap_cache_async()
798 *
799 * This function tries to find a page with the given swap entry
800 * in the swapper_space address space (the swap cache). If the page
801 * is found, it is returned in retpage. Otherwise, a page is allocated,
802 * added to the swap cache, and returned in retpage.
803 *
804 * If success, the swap cache page is returned in retpage
805 * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
806 * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
807 * the new page is added to swapcache and locked
808 * Returns ZSWAP_SWAPCACHE_FAIL on error
809 */
810static int zswap_get_swap_cache_page(swp_entry_t entry,
811 struct page **retpage)
812{
813 bool page_was_allocated;
814
815 *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
816 NULL, 0, &page_was_allocated);
817 if (page_was_allocated)
818 return ZSWAP_SWAPCACHE_NEW;
819 if (!*retpage)
820 return ZSWAP_SWAPCACHE_FAIL;
821 return ZSWAP_SWAPCACHE_EXIST;
822}
823
824/*
825 * Attempts to free an entry by adding a page to the swap cache,
826 * decompressing the entry data into the page, and issuing a
827 * bio write to write the page back to the swap device.
828 *
829 * This can be thought of as a "resumed writeback" of the page
830 * to the swap device. We are basically resuming the same swap
831 * writeback path that was intercepted with the frontswap_store()
832 * in the first place. After the page has been decompressed into
833 * the swap cache, the compressed version stored by zswap can be
834 * freed.
835 */
836static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
837{
838 struct zswap_header *zhdr;
839 swp_entry_t swpentry;
840 struct zswap_tree *tree;
841 pgoff_t offset;
842 struct zswap_entry *entry;
843 struct page *page;
844 struct crypto_comp *tfm;
845 u8 *src, *dst;
846 unsigned int dlen;
847 int ret;
848 struct writeback_control wbc = {
849 .sync_mode = WB_SYNC_NONE,
850 };
851
852 /* extract swpentry from data */
853 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
854 swpentry = zhdr->swpentry; /* here */
855 zpool_unmap_handle(pool, handle);
856 tree = zswap_trees[swp_type(swpentry)];
857 offset = swp_offset(swpentry);
858
859 /* find and ref zswap entry */
860 spin_lock(&tree->lock);
861 entry = zswap_entry_find_get(&tree->rbroot, offset);
862 if (!entry) {
863 /* entry was invalidated */
864 spin_unlock(&tree->lock);
865 return 0;
866 }
867 spin_unlock(&tree->lock);
868 BUG_ON(offset != entry->offset);
869
870 /* try to allocate swap cache page */
871 switch (zswap_get_swap_cache_page(swpentry, &page)) {
872 case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
873 ret = -ENOMEM;
874 goto fail;
875
876 case ZSWAP_SWAPCACHE_EXIST:
877 /* page is already in the swap cache, ignore for now */
878 put_page(page);
879 ret = -EEXIST;
880 goto fail;
881
882 case ZSWAP_SWAPCACHE_NEW: /* page is locked */
883 /* decompress */
884 dlen = PAGE_SIZE;
885 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
886 ZPOOL_MM_RO) + sizeof(struct zswap_header);
887 dst = kmap_atomic(page);
888 tfm = *get_cpu_ptr(entry->pool->tfm);
889 ret = crypto_comp_decompress(tfm, src, entry->length,
890 dst, &dlen);
891 put_cpu_ptr(entry->pool->tfm);
892 kunmap_atomic(dst);
893 zpool_unmap_handle(entry->pool->zpool, entry->handle);
894 BUG_ON(ret);
895 BUG_ON(dlen != PAGE_SIZE);
896
897 /* page is up to date */
898 SetPageUptodate(page);
899 }
900
901 /* move it to the tail of the inactive list after end_writeback */
902 SetPageReclaim(page);
903
904 /* start writeback */
905 __swap_writepage(page, &wbc, end_swap_bio_write);
906 put_page(page);
907 zswap_written_back_pages++;
908
909 spin_lock(&tree->lock);
910 /* drop local reference */
911 zswap_entry_put(tree, entry);
912
913 /*
914 * There are two possible situations for entry here:
915 * (1) refcount is 1(normal case), entry is valid and on the tree
916 * (2) refcount is 0, entry is freed and not on the tree
917 * because invalidate happened during writeback
918 * search the tree and free the entry if find entry
919 */
920 if (entry == zswap_rb_search(&tree->rbroot, offset))
921 zswap_entry_put(tree, entry);
922 spin_unlock(&tree->lock);
923
924 goto end;
925
926 /*
927 * if we get here due to ZSWAP_SWAPCACHE_EXIST
928 * a load may happening concurrently
929 * it is safe and okay to not free the entry
930 * if we free the entry in the following put
931 * it it either okay to return !0
932 */
933fail:
934 spin_lock(&tree->lock);
935 zswap_entry_put(tree, entry);
936 spin_unlock(&tree->lock);
937
938end:
939 return ret;
940}
941
942static int zswap_shrink(void)
943{
944 struct zswap_pool *pool;
945 int ret;
946
947 pool = zswap_pool_last_get();
948 if (!pool)
949 return -ENOENT;
950
951 ret = zpool_shrink(pool->zpool, 1, NULL);
952
953 zswap_pool_put(pool);
954
955 return ret;
956}
957
958/*********************************
959* frontswap hooks
960**********************************/
961/* attempts to compress and store an single page */
962static int zswap_frontswap_store(unsigned type, pgoff_t offset,
963 struct page *page)
964{
965 struct zswap_tree *tree = zswap_trees[type];
966 struct zswap_entry *entry, *dupentry;
967 struct crypto_comp *tfm;
968 int ret;
969 unsigned int dlen = PAGE_SIZE, len;
970 unsigned long handle;
971 char *buf;
972 u8 *src, *dst;
973 struct zswap_header *zhdr;
974
975 if (!zswap_enabled || !tree) {
976 ret = -ENODEV;
977 goto reject;
978 }
979
980 /* reclaim space if needed */
981 if (zswap_is_full()) {
982 zswap_pool_limit_hit++;
983 if (zswap_shrink()) {
984 zswap_reject_reclaim_fail++;
985 ret = -ENOMEM;
986 goto reject;
987 }
988 }
989
990 /* allocate entry */
991 entry = zswap_entry_cache_alloc(GFP_KERNEL);
992 if (!entry) {
993 zswap_reject_kmemcache_fail++;
994 ret = -ENOMEM;
995 goto reject;
996 }
997
998 /* if entry is successfully added, it keeps the reference */
999 entry->pool = zswap_pool_current_get();
1000 if (!entry->pool) {
1001 ret = -EINVAL;
1002 goto freepage;
1003 }
1004
1005 /* compress */
1006 dst = get_cpu_var(zswap_dstmem);
1007 tfm = *get_cpu_ptr(entry->pool->tfm);
1008 src = kmap_atomic(page);
1009 ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
1010 kunmap_atomic(src);
1011 put_cpu_ptr(entry->pool->tfm);
1012 if (ret) {
1013 ret = -EINVAL;
1014 goto put_dstmem;
1015 }
1016
1017 /* store */
1018 len = dlen + sizeof(struct zswap_header);
1019 ret = zpool_malloc(entry->pool->zpool, len,
1020 __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
1021 &handle);
1022 if (ret == -ENOSPC) {
1023 zswap_reject_compress_poor++;
1024 goto put_dstmem;
1025 }
1026 if (ret) {
1027 zswap_reject_alloc_fail++;
1028 goto put_dstmem;
1029 }
1030 zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1031 zhdr->swpentry = swp_entry(type, offset);
1032 buf = (u8 *)(zhdr + 1);
1033 memcpy(buf, dst, dlen);
1034 zpool_unmap_handle(entry->pool->zpool, handle);
1035 put_cpu_var(zswap_dstmem);
1036
1037 /* populate entry */
1038 entry->offset = offset;
1039 entry->handle = handle;
1040 entry->length = dlen;
1041
1042 /* map */
1043 spin_lock(&tree->lock);
1044 do {
1045 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1046 if (ret == -EEXIST) {
1047 zswap_duplicate_entry++;
1048 /* remove from rbtree */
1049 zswap_rb_erase(&tree->rbroot, dupentry);
1050 zswap_entry_put(tree, dupentry);
1051 }
1052 } while (ret == -EEXIST);
1053 spin_unlock(&tree->lock);
1054
1055 /* update stats */
1056 atomic_inc(&zswap_stored_pages);
1057 zswap_update_total_size();
1058
1059 return 0;
1060
1061put_dstmem:
1062 put_cpu_var(zswap_dstmem);
1063 zswap_pool_put(entry->pool);
1064freepage:
1065 zswap_entry_cache_free(entry);
1066reject:
1067 return ret;
1068}
1069
1070/*
1071 * returns 0 if the page was successfully decompressed
1072 * return -1 on entry not found or error
1073*/
1074static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1075 struct page *page)
1076{
1077 struct zswap_tree *tree = zswap_trees[type];
1078 struct zswap_entry *entry;
1079 struct crypto_comp *tfm;
1080 u8 *src, *dst;
1081 unsigned int dlen;
1082 int ret;
1083
1084 /* find */
1085 spin_lock(&tree->lock);
1086 entry = zswap_entry_find_get(&tree->rbroot, offset);
1087 if (!entry) {
1088 /* entry was written back */
1089 spin_unlock(&tree->lock);
1090 return -1;
1091 }
1092 spin_unlock(&tree->lock);
1093
1094 /* decompress */
1095 dlen = PAGE_SIZE;
1096 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1097 ZPOOL_MM_RO) + sizeof(struct zswap_header);
1098 dst = kmap_atomic(page);
1099 tfm = *get_cpu_ptr(entry->pool->tfm);
1100 ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1101 put_cpu_ptr(entry->pool->tfm);
1102 kunmap_atomic(dst);
1103 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1104 BUG_ON(ret);
1105
1106 spin_lock(&tree->lock);
1107 zswap_entry_put(tree, entry);
1108 spin_unlock(&tree->lock);
1109
1110 return 0;
1111}
1112
1113/* frees an entry in zswap */
1114static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1115{
1116 struct zswap_tree *tree = zswap_trees[type];
1117 struct zswap_entry *entry;
1118
1119 /* find */
1120 spin_lock(&tree->lock);
1121 entry = zswap_rb_search(&tree->rbroot, offset);
1122 if (!entry) {
1123 /* entry was written back */
1124 spin_unlock(&tree->lock);
1125 return;
1126 }
1127
1128 /* remove from rbtree */
1129 zswap_rb_erase(&tree->rbroot, entry);
1130
1131 /* drop the initial reference from entry creation */
1132 zswap_entry_put(tree, entry);
1133
1134 spin_unlock(&tree->lock);
1135}
1136
1137/* frees all zswap entries for the given swap type */
1138static void zswap_frontswap_invalidate_area(unsigned type)
1139{
1140 struct zswap_tree *tree = zswap_trees[type];
1141 struct zswap_entry *entry, *n;
1142
1143 if (!tree)
1144 return;
1145
1146 /* walk the tree and free everything */
1147 spin_lock(&tree->lock);
1148 rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1149 zswap_free_entry(entry);
1150 tree->rbroot = RB_ROOT;
1151 spin_unlock(&tree->lock);
1152 kfree(tree);
1153 zswap_trees[type] = NULL;
1154}
1155
1156static void zswap_frontswap_init(unsigned type)
1157{
1158 struct zswap_tree *tree;
1159
1160 tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1161 if (!tree) {
1162 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1163 return;
1164 }
1165
1166 tree->rbroot = RB_ROOT;
1167 spin_lock_init(&tree->lock);
1168 zswap_trees[type] = tree;
1169}
1170
1171static struct frontswap_ops zswap_frontswap_ops = {
1172 .store = zswap_frontswap_store,
1173 .load = zswap_frontswap_load,
1174 .invalidate_page = zswap_frontswap_invalidate_page,
1175 .invalidate_area = zswap_frontswap_invalidate_area,
1176 .init = zswap_frontswap_init
1177};
1178
1179/*********************************
1180* debugfs functions
1181**********************************/
1182#ifdef CONFIG_DEBUG_FS
1183#include <linux/debugfs.h>
1184
1185static struct dentry *zswap_debugfs_root;
1186
1187static int __init zswap_debugfs_init(void)
1188{
1189 if (!debugfs_initialized())
1190 return -ENODEV;
1191
1192 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1193 if (!zswap_debugfs_root)
1194 return -ENOMEM;
1195
1196 debugfs_create_u64("pool_limit_hit", S_IRUGO,
1197 zswap_debugfs_root, &zswap_pool_limit_hit);
1198 debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1199 zswap_debugfs_root, &zswap_reject_reclaim_fail);
1200 debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1201 zswap_debugfs_root, &zswap_reject_alloc_fail);
1202 debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1203 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1204 debugfs_create_u64("reject_compress_poor", S_IRUGO,
1205 zswap_debugfs_root, &zswap_reject_compress_poor);
1206 debugfs_create_u64("written_back_pages", S_IRUGO,
1207 zswap_debugfs_root, &zswap_written_back_pages);
1208 debugfs_create_u64("duplicate_entry", S_IRUGO,
1209 zswap_debugfs_root, &zswap_duplicate_entry);
1210 debugfs_create_u64("pool_total_size", S_IRUGO,
1211 zswap_debugfs_root, &zswap_pool_total_size);
1212 debugfs_create_atomic_t("stored_pages", S_IRUGO,
1213 zswap_debugfs_root, &zswap_stored_pages);
1214
1215 return 0;
1216}
1217
1218static void __exit zswap_debugfs_exit(void)
1219{
1220 debugfs_remove_recursive(zswap_debugfs_root);
1221}
1222#else
1223static int __init zswap_debugfs_init(void)
1224{
1225 return 0;
1226}
1227
1228static void __exit zswap_debugfs_exit(void) { }
1229#endif
1230
1231/*********************************
1232* module init and exit
1233**********************************/
1234static int __init init_zswap(void)
1235{
1236 struct zswap_pool *pool;
1237
1238 zswap_init_started = true;
1239
1240 if (zswap_entry_cache_create()) {
1241 pr_err("entry cache creation failed\n");
1242 goto cache_fail;
1243 }
1244
1245 if (zswap_cpu_dstmem_init()) {
1246 pr_err("dstmem alloc failed\n");
1247 goto dstmem_fail;
1248 }
1249
1250 pool = __zswap_pool_create_fallback();
1251 if (!pool) {
1252 pr_err("pool creation failed\n");
1253 goto pool_fail;
1254 }
1255 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1256 zpool_get_type(pool->zpool));
1257
1258 list_add(&pool->list, &zswap_pools);
1259
1260 frontswap_register_ops(&zswap_frontswap_ops);
1261 if (zswap_debugfs_init())
1262 pr_warn("debugfs initialization failed\n");
1263 return 0;
1264
1265pool_fail:
1266 zswap_cpu_dstmem_destroy();
1267dstmem_fail:
1268 zswap_entry_cache_destroy();
1269cache_fail:
1270 return -ENOMEM;
1271}
1272/* must be late so crypto has time to come up */
1273late_initcall(init_zswap);
1274
1275MODULE_LICENSE("GPL");
1276MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1277MODULE_DESCRIPTION("Compressed cache for swap pages");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * zswap.c - zswap driver file
4 *
5 * zswap is a cache that takes pages that are in the process
6 * of being swapped out and attempts to compress and store them in a
7 * RAM-based memory pool. This can result in a significant I/O reduction on
8 * the swap device and, in the case where decompressing from RAM is faster
9 * than reading from the swap device, can also improve workload performance.
10 *
11 * Copyright (C) 2012 Seth Jennings <sjenning@linux.vnet.ibm.com>
12*/
13
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16#include <linux/module.h>
17#include <linux/cpu.h>
18#include <linux/highmem.h>
19#include <linux/slab.h>
20#include <linux/spinlock.h>
21#include <linux/types.h>
22#include <linux/atomic.h>
23#include <linux/rbtree.h>
24#include <linux/swap.h>
25#include <linux/crypto.h>
26#include <linux/scatterlist.h>
27#include <linux/mempolicy.h>
28#include <linux/mempool.h>
29#include <linux/zpool.h>
30#include <crypto/acompress.h>
31#include <linux/zswap.h>
32#include <linux/mm_types.h>
33#include <linux/page-flags.h>
34#include <linux/swapops.h>
35#include <linux/writeback.h>
36#include <linux/pagemap.h>
37#include <linux/workqueue.h>
38#include <linux/list_lru.h>
39
40#include "swap.h"
41#include "internal.h"
42
43/*********************************
44* statistics
45**********************************/
46/* Total bytes used by the compressed storage */
47u64 zswap_pool_total_size;
48/* The number of compressed pages currently stored in zswap */
49atomic_t zswap_stored_pages = ATOMIC_INIT(0);
50/* The number of same-value filled pages currently stored in zswap */
51static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
52
53/*
54 * The statistics below are not protected from concurrent access for
55 * performance reasons so they may not be a 100% accurate. However,
56 * they do provide useful information on roughly how many times a
57 * certain event is occurring.
58*/
59
60/* Pool limit was hit (see zswap_max_pool_percent) */
61static u64 zswap_pool_limit_hit;
62/* Pages written back when pool limit was reached */
63static u64 zswap_written_back_pages;
64/* Store failed due to a reclaim failure after pool limit was reached */
65static u64 zswap_reject_reclaim_fail;
66/* Store failed due to compression algorithm failure */
67static u64 zswap_reject_compress_fail;
68/* Compressed page was too big for the allocator to (optimally) store */
69static u64 zswap_reject_compress_poor;
70/* Store failed because underlying allocator could not get memory */
71static u64 zswap_reject_alloc_fail;
72/* Store failed because the entry metadata could not be allocated (rare) */
73static u64 zswap_reject_kmemcache_fail;
74
75/* Shrinker work queue */
76static struct workqueue_struct *shrink_wq;
77/* Pool limit was hit, we need to calm down */
78static bool zswap_pool_reached_full;
79
80/*********************************
81* tunables
82**********************************/
83
84#define ZSWAP_PARAM_UNSET ""
85
86static int zswap_setup(void);
87
88/* Enable/disable zswap */
89static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
90static int zswap_enabled_param_set(const char *,
91 const struct kernel_param *);
92static const struct kernel_param_ops zswap_enabled_param_ops = {
93 .set = zswap_enabled_param_set,
94 .get = param_get_bool,
95};
96module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
97
98/* Crypto compressor to use */
99static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
100static int zswap_compressor_param_set(const char *,
101 const struct kernel_param *);
102static const struct kernel_param_ops zswap_compressor_param_ops = {
103 .set = zswap_compressor_param_set,
104 .get = param_get_charp,
105 .free = param_free_charp,
106};
107module_param_cb(compressor, &zswap_compressor_param_ops,
108 &zswap_compressor, 0644);
109
110/* Compressed storage zpool to use */
111static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
112static int zswap_zpool_param_set(const char *, const struct kernel_param *);
113static const struct kernel_param_ops zswap_zpool_param_ops = {
114 .set = zswap_zpool_param_set,
115 .get = param_get_charp,
116 .free = param_free_charp,
117};
118module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
119
120/* The maximum percentage of memory that the compressed pool can occupy */
121static unsigned int zswap_max_pool_percent = 20;
122module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
123
124/* The threshold for accepting new pages after the max_pool_percent was hit */
125static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
126module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
127 uint, 0644);
128
129/*
130 * Enable/disable handling same-value filled pages (enabled by default).
131 * If disabled every page is considered non-same-value filled.
132 */
133static bool zswap_same_filled_pages_enabled = true;
134module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
135 bool, 0644);
136
137/* Enable/disable handling non-same-value filled pages (enabled by default) */
138static bool zswap_non_same_filled_pages_enabled = true;
139module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
140 bool, 0644);
141
142/* Number of zpools in zswap_pool (empirically determined for scalability) */
143#define ZSWAP_NR_ZPOOLS 32
144
145/* Enable/disable memory pressure-based shrinker. */
146static bool zswap_shrinker_enabled = IS_ENABLED(
147 CONFIG_ZSWAP_SHRINKER_DEFAULT_ON);
148module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
149
150bool is_zswap_enabled(void)
151{
152 return zswap_enabled;
153}
154
155/*********************************
156* data structures
157**********************************/
158
159struct crypto_acomp_ctx {
160 struct crypto_acomp *acomp;
161 struct acomp_req *req;
162 struct crypto_wait wait;
163 u8 *buffer;
164 struct mutex mutex;
165 bool is_sleepable;
166};
167
168/*
169 * The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
170 * The only case where lru_lock is not acquired while holding tree.lock is
171 * when a zswap_entry is taken off the lru for writeback, in that case it
172 * needs to be verified that it's still valid in the tree.
173 */
174struct zswap_pool {
175 struct zpool *zpools[ZSWAP_NR_ZPOOLS];
176 struct crypto_acomp_ctx __percpu *acomp_ctx;
177 struct percpu_ref ref;
178 struct list_head list;
179 struct work_struct release_work;
180 struct hlist_node node;
181 char tfm_name[CRYPTO_MAX_ALG_NAME];
182};
183
184/* Global LRU lists shared by all zswap pools. */
185static struct list_lru zswap_list_lru;
186/* counter of pages stored in all zswap pools. */
187static atomic_t zswap_nr_stored = ATOMIC_INIT(0);
188
189/* The lock protects zswap_next_shrink updates. */
190static DEFINE_SPINLOCK(zswap_shrink_lock);
191static struct mem_cgroup *zswap_next_shrink;
192static struct work_struct zswap_shrink_work;
193static struct shrinker *zswap_shrinker;
194
195/*
196 * struct zswap_entry
197 *
198 * This structure contains the metadata for tracking a single compressed
199 * page within zswap.
200 *
201 * rbnode - links the entry into red-black tree for the appropriate swap type
202 * swpentry - associated swap entry, the offset indexes into the red-black tree
203 * length - the length in bytes of the compressed page data. Needed during
204 * decompression. For a same value filled page length is 0, and both
205 * pool and lru are invalid and must be ignored.
206 * pool - the zswap_pool the entry's data is in
207 * handle - zpool allocation handle that stores the compressed page data
208 * value - value of the same-value filled pages which have same content
209 * objcg - the obj_cgroup that the compressed memory is charged to
210 * lru - handle to the pool's lru used to evict pages.
211 */
212struct zswap_entry {
213 struct rb_node rbnode;
214 swp_entry_t swpentry;
215 unsigned int length;
216 struct zswap_pool *pool;
217 union {
218 unsigned long handle;
219 unsigned long value;
220 };
221 struct obj_cgroup *objcg;
222 struct list_head lru;
223};
224
225struct zswap_tree {
226 struct rb_root rbroot;
227 spinlock_t lock;
228};
229
230static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
231static unsigned int nr_zswap_trees[MAX_SWAPFILES];
232
233/* RCU-protected iteration */
234static LIST_HEAD(zswap_pools);
235/* protects zswap_pools list modification */
236static DEFINE_SPINLOCK(zswap_pools_lock);
237/* pool counter to provide unique names to zpool */
238static atomic_t zswap_pools_count = ATOMIC_INIT(0);
239
240enum zswap_init_type {
241 ZSWAP_UNINIT,
242 ZSWAP_INIT_SUCCEED,
243 ZSWAP_INIT_FAILED
244};
245
246static enum zswap_init_type zswap_init_state;
247
248/* used to ensure the integrity of initialization */
249static DEFINE_MUTEX(zswap_init_lock);
250
251/* init completed, but couldn't create the initial pool */
252static bool zswap_has_pool;
253
254/*********************************
255* helpers and fwd declarations
256**********************************/
257
258static inline struct zswap_tree *swap_zswap_tree(swp_entry_t swp)
259{
260 return &zswap_trees[swp_type(swp)][swp_offset(swp)
261 >> SWAP_ADDRESS_SPACE_SHIFT];
262}
263
264#define zswap_pool_debug(msg, p) \
265 pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name, \
266 zpool_get_type((p)->zpools[0]))
267
268static bool zswap_is_full(void)
269{
270 return totalram_pages() * zswap_max_pool_percent / 100 <
271 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
272}
273
274static bool zswap_can_accept(void)
275{
276 return totalram_pages() * zswap_accept_thr_percent / 100 *
277 zswap_max_pool_percent / 100 >
278 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
279}
280
281static u64 get_zswap_pool_size(struct zswap_pool *pool)
282{
283 u64 pool_size = 0;
284 int i;
285
286 for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
287 pool_size += zpool_get_total_size(pool->zpools[i]);
288
289 return pool_size;
290}
291
292static void zswap_update_total_size(void)
293{
294 struct zswap_pool *pool;
295 u64 total = 0;
296
297 rcu_read_lock();
298
299 list_for_each_entry_rcu(pool, &zswap_pools, list)
300 total += get_zswap_pool_size(pool);
301
302 rcu_read_unlock();
303
304 zswap_pool_total_size = total;
305}
306
307/*********************************
308* pool functions
309**********************************/
310static void __zswap_pool_empty(struct percpu_ref *ref);
311
312static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
313{
314 int i;
315 struct zswap_pool *pool;
316 char name[38]; /* 'zswap' + 32 char (max) num + \0 */
317 gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
318 int ret;
319
320 if (!zswap_has_pool) {
321 /* if either are unset, pool initialization failed, and we
322 * need both params to be set correctly before trying to
323 * create a pool.
324 */
325 if (!strcmp(type, ZSWAP_PARAM_UNSET))
326 return NULL;
327 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
328 return NULL;
329 }
330
331 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
332 if (!pool)
333 return NULL;
334
335 for (i = 0; i < ZSWAP_NR_ZPOOLS; i++) {
336 /* unique name for each pool specifically required by zsmalloc */
337 snprintf(name, 38, "zswap%x",
338 atomic_inc_return(&zswap_pools_count));
339
340 pool->zpools[i] = zpool_create_pool(type, name, gfp);
341 if (!pool->zpools[i]) {
342 pr_err("%s zpool not available\n", type);
343 goto error;
344 }
345 }
346 pr_debug("using %s zpool\n", zpool_get_type(pool->zpools[0]));
347
348 strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
349
350 pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
351 if (!pool->acomp_ctx) {
352 pr_err("percpu alloc failed\n");
353 goto error;
354 }
355
356 ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
357 &pool->node);
358 if (ret)
359 goto error;
360
361 /* being the current pool takes 1 ref; this func expects the
362 * caller to always add the new pool as the current pool
363 */
364 ret = percpu_ref_init(&pool->ref, __zswap_pool_empty,
365 PERCPU_REF_ALLOW_REINIT, GFP_KERNEL);
366 if (ret)
367 goto ref_fail;
368 INIT_LIST_HEAD(&pool->list);
369
370 zswap_pool_debug("created", pool);
371
372 return pool;
373
374ref_fail:
375 cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
376error:
377 if (pool->acomp_ctx)
378 free_percpu(pool->acomp_ctx);
379 while (i--)
380 zpool_destroy_pool(pool->zpools[i]);
381 kfree(pool);
382 return NULL;
383}
384
385static struct zswap_pool *__zswap_pool_create_fallback(void)
386{
387 bool has_comp, has_zpool;
388
389 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
390 if (!has_comp && strcmp(zswap_compressor,
391 CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
392 pr_err("compressor %s not available, using default %s\n",
393 zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
394 param_free_charp(&zswap_compressor);
395 zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
396 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
397 }
398 if (!has_comp) {
399 pr_err("default compressor %s not available\n",
400 zswap_compressor);
401 param_free_charp(&zswap_compressor);
402 zswap_compressor = ZSWAP_PARAM_UNSET;
403 }
404
405 has_zpool = zpool_has_pool(zswap_zpool_type);
406 if (!has_zpool && strcmp(zswap_zpool_type,
407 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
408 pr_err("zpool %s not available, using default %s\n",
409 zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
410 param_free_charp(&zswap_zpool_type);
411 zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
412 has_zpool = zpool_has_pool(zswap_zpool_type);
413 }
414 if (!has_zpool) {
415 pr_err("default zpool %s not available\n",
416 zswap_zpool_type);
417 param_free_charp(&zswap_zpool_type);
418 zswap_zpool_type = ZSWAP_PARAM_UNSET;
419 }
420
421 if (!has_comp || !has_zpool)
422 return NULL;
423
424 return zswap_pool_create(zswap_zpool_type, zswap_compressor);
425}
426
427static void zswap_pool_destroy(struct zswap_pool *pool)
428{
429 int i;
430
431 zswap_pool_debug("destroying", pool);
432
433 cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
434 free_percpu(pool->acomp_ctx);
435
436 for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
437 zpool_destroy_pool(pool->zpools[i]);
438 kfree(pool);
439}
440
441static void __zswap_pool_release(struct work_struct *work)
442{
443 struct zswap_pool *pool = container_of(work, typeof(*pool),
444 release_work);
445
446 synchronize_rcu();
447
448 /* nobody should have been able to get a ref... */
449 WARN_ON(!percpu_ref_is_zero(&pool->ref));
450 percpu_ref_exit(&pool->ref);
451
452 /* pool is now off zswap_pools list and has no references. */
453 zswap_pool_destroy(pool);
454}
455
456static struct zswap_pool *zswap_pool_current(void);
457
458static void __zswap_pool_empty(struct percpu_ref *ref)
459{
460 struct zswap_pool *pool;
461
462 pool = container_of(ref, typeof(*pool), ref);
463
464 spin_lock_bh(&zswap_pools_lock);
465
466 WARN_ON(pool == zswap_pool_current());
467
468 list_del_rcu(&pool->list);
469
470 INIT_WORK(&pool->release_work, __zswap_pool_release);
471 schedule_work(&pool->release_work);
472
473 spin_unlock_bh(&zswap_pools_lock);
474}
475
476static int __must_check zswap_pool_get(struct zswap_pool *pool)
477{
478 if (!pool)
479 return 0;
480
481 return percpu_ref_tryget(&pool->ref);
482}
483
484static void zswap_pool_put(struct zswap_pool *pool)
485{
486 percpu_ref_put(&pool->ref);
487}
488
489static struct zswap_pool *__zswap_pool_current(void)
490{
491 struct zswap_pool *pool;
492
493 pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
494 WARN_ONCE(!pool && zswap_has_pool,
495 "%s: no page storage pool!\n", __func__);
496
497 return pool;
498}
499
500static struct zswap_pool *zswap_pool_current(void)
501{
502 assert_spin_locked(&zswap_pools_lock);
503
504 return __zswap_pool_current();
505}
506
507static struct zswap_pool *zswap_pool_current_get(void)
508{
509 struct zswap_pool *pool;
510
511 rcu_read_lock();
512
513 pool = __zswap_pool_current();
514 if (!zswap_pool_get(pool))
515 pool = NULL;
516
517 rcu_read_unlock();
518
519 return pool;
520}
521
522/* type and compressor must be null-terminated */
523static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
524{
525 struct zswap_pool *pool;
526
527 assert_spin_locked(&zswap_pools_lock);
528
529 list_for_each_entry_rcu(pool, &zswap_pools, list) {
530 if (strcmp(pool->tfm_name, compressor))
531 continue;
532 /* all zpools share the same type */
533 if (strcmp(zpool_get_type(pool->zpools[0]), type))
534 continue;
535 /* if we can't get it, it's about to be destroyed */
536 if (!zswap_pool_get(pool))
537 continue;
538 return pool;
539 }
540
541 return NULL;
542}
543
544/*********************************
545* param callbacks
546**********************************/
547
548static bool zswap_pool_changed(const char *s, const struct kernel_param *kp)
549{
550 /* no change required */
551 if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
552 return false;
553 return true;
554}
555
556/* val must be a null-terminated string */
557static int __zswap_param_set(const char *val, const struct kernel_param *kp,
558 char *type, char *compressor)
559{
560 struct zswap_pool *pool, *put_pool = NULL;
561 char *s = strstrip((char *)val);
562 int ret = 0;
563 bool new_pool = false;
564
565 mutex_lock(&zswap_init_lock);
566 switch (zswap_init_state) {
567 case ZSWAP_UNINIT:
568 /* if this is load-time (pre-init) param setting,
569 * don't create a pool; that's done during init.
570 */
571 ret = param_set_charp(s, kp);
572 break;
573 case ZSWAP_INIT_SUCCEED:
574 new_pool = zswap_pool_changed(s, kp);
575 break;
576 case ZSWAP_INIT_FAILED:
577 pr_err("can't set param, initialization failed\n");
578 ret = -ENODEV;
579 }
580 mutex_unlock(&zswap_init_lock);
581
582 /* no need to create a new pool, return directly */
583 if (!new_pool)
584 return ret;
585
586 if (!type) {
587 if (!zpool_has_pool(s)) {
588 pr_err("zpool %s not available\n", s);
589 return -ENOENT;
590 }
591 type = s;
592 } else if (!compressor) {
593 if (!crypto_has_acomp(s, 0, 0)) {
594 pr_err("compressor %s not available\n", s);
595 return -ENOENT;
596 }
597 compressor = s;
598 } else {
599 WARN_ON(1);
600 return -EINVAL;
601 }
602
603 spin_lock_bh(&zswap_pools_lock);
604
605 pool = zswap_pool_find_get(type, compressor);
606 if (pool) {
607 zswap_pool_debug("using existing", pool);
608 WARN_ON(pool == zswap_pool_current());
609 list_del_rcu(&pool->list);
610 }
611
612 spin_unlock_bh(&zswap_pools_lock);
613
614 if (!pool)
615 pool = zswap_pool_create(type, compressor);
616 else {
617 /*
618 * Restore the initial ref dropped by percpu_ref_kill()
619 * when the pool was decommissioned and switch it again
620 * to percpu mode.
621 */
622 percpu_ref_resurrect(&pool->ref);
623
624 /* Drop the ref from zswap_pool_find_get(). */
625 zswap_pool_put(pool);
626 }
627
628 if (pool)
629 ret = param_set_charp(s, kp);
630 else
631 ret = -EINVAL;
632
633 spin_lock_bh(&zswap_pools_lock);
634
635 if (!ret) {
636 put_pool = zswap_pool_current();
637 list_add_rcu(&pool->list, &zswap_pools);
638 zswap_has_pool = true;
639 } else if (pool) {
640 /* add the possibly pre-existing pool to the end of the pools
641 * list; if it's new (and empty) then it'll be removed and
642 * destroyed by the put after we drop the lock
643 */
644 list_add_tail_rcu(&pool->list, &zswap_pools);
645 put_pool = pool;
646 }
647
648 spin_unlock_bh(&zswap_pools_lock);
649
650 if (!zswap_has_pool && !pool) {
651 /* if initial pool creation failed, and this pool creation also
652 * failed, maybe both compressor and zpool params were bad.
653 * Allow changing this param, so pool creation will succeed
654 * when the other param is changed. We already verified this
655 * param is ok in the zpool_has_pool() or crypto_has_acomp()
656 * checks above.
657 */
658 ret = param_set_charp(s, kp);
659 }
660
661 /* drop the ref from either the old current pool,
662 * or the new pool we failed to add
663 */
664 if (put_pool)
665 percpu_ref_kill(&put_pool->ref);
666
667 return ret;
668}
669
670static int zswap_compressor_param_set(const char *val,
671 const struct kernel_param *kp)
672{
673 return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
674}
675
676static int zswap_zpool_param_set(const char *val,
677 const struct kernel_param *kp)
678{
679 return __zswap_param_set(val, kp, NULL, zswap_compressor);
680}
681
682static int zswap_enabled_param_set(const char *val,
683 const struct kernel_param *kp)
684{
685 int ret = -ENODEV;
686
687 /* if this is load-time (pre-init) param setting, only set param. */
688 if (system_state != SYSTEM_RUNNING)
689 return param_set_bool(val, kp);
690
691 mutex_lock(&zswap_init_lock);
692 switch (zswap_init_state) {
693 case ZSWAP_UNINIT:
694 if (zswap_setup())
695 break;
696 fallthrough;
697 case ZSWAP_INIT_SUCCEED:
698 if (!zswap_has_pool)
699 pr_err("can't enable, no pool configured\n");
700 else
701 ret = param_set_bool(val, kp);
702 break;
703 case ZSWAP_INIT_FAILED:
704 pr_err("can't enable, initialization failed\n");
705 }
706 mutex_unlock(&zswap_init_lock);
707
708 return ret;
709}
710
711/*********************************
712* lru functions
713**********************************/
714
715/* should be called under RCU */
716#ifdef CONFIG_MEMCG
717static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
718{
719 return entry->objcg ? obj_cgroup_memcg(entry->objcg) : NULL;
720}
721#else
722static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
723{
724 return NULL;
725}
726#endif
727
728static inline int entry_to_nid(struct zswap_entry *entry)
729{
730 return page_to_nid(virt_to_page(entry));
731}
732
733static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
734{
735 atomic_long_t *nr_zswap_protected;
736 unsigned long lru_size, old, new;
737 int nid = entry_to_nid(entry);
738 struct mem_cgroup *memcg;
739 struct lruvec *lruvec;
740
741 /*
742 * Note that it is safe to use rcu_read_lock() here, even in the face of
743 * concurrent memcg offlining. Thanks to the memcg->kmemcg_id indirection
744 * used in list_lru lookup, only two scenarios are possible:
745 *
746 * 1. list_lru_add() is called before memcg->kmemcg_id is updated. The
747 * new entry will be reparented to memcg's parent's list_lru.
748 * 2. list_lru_add() is called after memcg->kmemcg_id is updated. The
749 * new entry will be added directly to memcg's parent's list_lru.
750 *
751 * Similar reasoning holds for list_lru_del().
752 */
753 rcu_read_lock();
754 memcg = mem_cgroup_from_entry(entry);
755 /* will always succeed */
756 list_lru_add(list_lru, &entry->lru, nid, memcg);
757
758 /* Update the protection area */
759 lru_size = list_lru_count_one(list_lru, nid, memcg);
760 lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
761 nr_zswap_protected = &lruvec->zswap_lruvec_state.nr_zswap_protected;
762 old = atomic_long_inc_return(nr_zswap_protected);
763 /*
764 * Decay to avoid overflow and adapt to changing workloads.
765 * This is based on LRU reclaim cost decaying heuristics.
766 */
767 do {
768 new = old > lru_size / 4 ? old / 2 : old;
769 } while (!atomic_long_try_cmpxchg(nr_zswap_protected, &old, new));
770 rcu_read_unlock();
771}
772
773static void zswap_lru_del(struct list_lru *list_lru, struct zswap_entry *entry)
774{
775 int nid = entry_to_nid(entry);
776 struct mem_cgroup *memcg;
777
778 rcu_read_lock();
779 memcg = mem_cgroup_from_entry(entry);
780 /* will always succeed */
781 list_lru_del(list_lru, &entry->lru, nid, memcg);
782 rcu_read_unlock();
783}
784
785void zswap_lruvec_state_init(struct lruvec *lruvec)
786{
787 atomic_long_set(&lruvec->zswap_lruvec_state.nr_zswap_protected, 0);
788}
789
790void zswap_folio_swapin(struct folio *folio)
791{
792 struct lruvec *lruvec;
793
794 if (folio) {
795 lruvec = folio_lruvec(folio);
796 atomic_long_inc(&lruvec->zswap_lruvec_state.nr_zswap_protected);
797 }
798}
799
800void zswap_memcg_offline_cleanup(struct mem_cgroup *memcg)
801{
802 /* lock out zswap shrinker walking memcg tree */
803 spin_lock(&zswap_shrink_lock);
804 if (zswap_next_shrink == memcg)
805 zswap_next_shrink = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
806 spin_unlock(&zswap_shrink_lock);
807}
808
809/*********************************
810* rbtree functions
811**********************************/
812static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
813{
814 struct rb_node *node = root->rb_node;
815 struct zswap_entry *entry;
816 pgoff_t entry_offset;
817
818 while (node) {
819 entry = rb_entry(node, struct zswap_entry, rbnode);
820 entry_offset = swp_offset(entry->swpentry);
821 if (entry_offset > offset)
822 node = node->rb_left;
823 else if (entry_offset < offset)
824 node = node->rb_right;
825 else
826 return entry;
827 }
828 return NULL;
829}
830
831/*
832 * In the case that a entry with the same offset is found, a pointer to
833 * the existing entry is stored in dupentry and the function returns -EEXIST
834 */
835static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
836 struct zswap_entry **dupentry)
837{
838 struct rb_node **link = &root->rb_node, *parent = NULL;
839 struct zswap_entry *myentry;
840 pgoff_t myentry_offset, entry_offset = swp_offset(entry->swpentry);
841
842 while (*link) {
843 parent = *link;
844 myentry = rb_entry(parent, struct zswap_entry, rbnode);
845 myentry_offset = swp_offset(myentry->swpentry);
846 if (myentry_offset > entry_offset)
847 link = &(*link)->rb_left;
848 else if (myentry_offset < entry_offset)
849 link = &(*link)->rb_right;
850 else {
851 *dupentry = myentry;
852 return -EEXIST;
853 }
854 }
855 rb_link_node(&entry->rbnode, parent, link);
856 rb_insert_color(&entry->rbnode, root);
857 return 0;
858}
859
860static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
861{
862 rb_erase(&entry->rbnode, root);
863 RB_CLEAR_NODE(&entry->rbnode);
864}
865
866/*********************************
867* zswap entry functions
868**********************************/
869static struct kmem_cache *zswap_entry_cache;
870
871static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp, int nid)
872{
873 struct zswap_entry *entry;
874 entry = kmem_cache_alloc_node(zswap_entry_cache, gfp, nid);
875 if (!entry)
876 return NULL;
877 RB_CLEAR_NODE(&entry->rbnode);
878 return entry;
879}
880
881static void zswap_entry_cache_free(struct zswap_entry *entry)
882{
883 kmem_cache_free(zswap_entry_cache, entry);
884}
885
886static struct zpool *zswap_find_zpool(struct zswap_entry *entry)
887{
888 int i = 0;
889
890 if (ZSWAP_NR_ZPOOLS > 1)
891 i = hash_ptr(entry, ilog2(ZSWAP_NR_ZPOOLS));
892
893 return entry->pool->zpools[i];
894}
895
896/*
897 * Carries out the common pattern of freeing and entry's zpool allocation,
898 * freeing the entry itself, and decrementing the number of stored pages.
899 */
900static void zswap_entry_free(struct zswap_entry *entry)
901{
902 if (!entry->length)
903 atomic_dec(&zswap_same_filled_pages);
904 else {
905 zswap_lru_del(&zswap_list_lru, entry);
906 zpool_free(zswap_find_zpool(entry), entry->handle);
907 atomic_dec(&zswap_nr_stored);
908 zswap_pool_put(entry->pool);
909 }
910 if (entry->objcg) {
911 obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
912 obj_cgroup_put(entry->objcg);
913 }
914 zswap_entry_cache_free(entry);
915 atomic_dec(&zswap_stored_pages);
916 zswap_update_total_size();
917}
918
919/*
920 * The caller hold the tree lock and search the entry from the tree,
921 * so it must be on the tree, remove it from the tree and free it.
922 */
923static void zswap_invalidate_entry(struct zswap_tree *tree,
924 struct zswap_entry *entry)
925{
926 zswap_rb_erase(&tree->rbroot, entry);
927 zswap_entry_free(entry);
928}
929
930/*********************************
931* compressed storage functions
932**********************************/
933static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
934{
935 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
936 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
937 struct crypto_acomp *acomp;
938 struct acomp_req *req;
939 int ret;
940
941 mutex_init(&acomp_ctx->mutex);
942
943 acomp_ctx->buffer = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
944 if (!acomp_ctx->buffer)
945 return -ENOMEM;
946
947 acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
948 if (IS_ERR(acomp)) {
949 pr_err("could not alloc crypto acomp %s : %ld\n",
950 pool->tfm_name, PTR_ERR(acomp));
951 ret = PTR_ERR(acomp);
952 goto acomp_fail;
953 }
954 acomp_ctx->acomp = acomp;
955 acomp_ctx->is_sleepable = acomp_is_async(acomp);
956
957 req = acomp_request_alloc(acomp_ctx->acomp);
958 if (!req) {
959 pr_err("could not alloc crypto acomp_request %s\n",
960 pool->tfm_name);
961 ret = -ENOMEM;
962 goto req_fail;
963 }
964 acomp_ctx->req = req;
965
966 crypto_init_wait(&acomp_ctx->wait);
967 /*
968 * if the backend of acomp is async zip, crypto_req_done() will wakeup
969 * crypto_wait_req(); if the backend of acomp is scomp, the callback
970 * won't be called, crypto_wait_req() will return without blocking.
971 */
972 acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
973 crypto_req_done, &acomp_ctx->wait);
974
975 return 0;
976
977req_fail:
978 crypto_free_acomp(acomp_ctx->acomp);
979acomp_fail:
980 kfree(acomp_ctx->buffer);
981 return ret;
982}
983
984static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
985{
986 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
987 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
988
989 if (!IS_ERR_OR_NULL(acomp_ctx)) {
990 if (!IS_ERR_OR_NULL(acomp_ctx->req))
991 acomp_request_free(acomp_ctx->req);
992 if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
993 crypto_free_acomp(acomp_ctx->acomp);
994 kfree(acomp_ctx->buffer);
995 }
996
997 return 0;
998}
999
1000static bool zswap_compress(struct folio *folio, struct zswap_entry *entry)
1001{
1002 struct crypto_acomp_ctx *acomp_ctx;
1003 struct scatterlist input, output;
1004 int comp_ret = 0, alloc_ret = 0;
1005 unsigned int dlen = PAGE_SIZE;
1006 unsigned long handle;
1007 struct zpool *zpool;
1008 char *buf;
1009 gfp_t gfp;
1010 u8 *dst;
1011
1012 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1013
1014 mutex_lock(&acomp_ctx->mutex);
1015
1016 dst = acomp_ctx->buffer;
1017 sg_init_table(&input, 1);
1018 sg_set_page(&input, &folio->page, PAGE_SIZE, 0);
1019
1020 /*
1021 * We need PAGE_SIZE * 2 here since there maybe over-compression case,
1022 * and hardware-accelerators may won't check the dst buffer size, so
1023 * giving the dst buffer with enough length to avoid buffer overflow.
1024 */
1025 sg_init_one(&output, dst, PAGE_SIZE * 2);
1026 acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1027
1028 /*
1029 * it maybe looks a little bit silly that we send an asynchronous request,
1030 * then wait for its completion synchronously. This makes the process look
1031 * synchronous in fact.
1032 * Theoretically, acomp supports users send multiple acomp requests in one
1033 * acomp instance, then get those requests done simultaneously. but in this
1034 * case, zswap actually does store and load page by page, there is no
1035 * existing method to send the second page before the first page is done
1036 * in one thread doing zwap.
1037 * but in different threads running on different cpu, we have different
1038 * acomp instance, so multiple threads can do (de)compression in parallel.
1039 */
1040 comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1041 dlen = acomp_ctx->req->dlen;
1042 if (comp_ret)
1043 goto unlock;
1044
1045 zpool = zswap_find_zpool(entry);
1046 gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1047 if (zpool_malloc_support_movable(zpool))
1048 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1049 alloc_ret = zpool_malloc(zpool, dlen, gfp, &handle);
1050 if (alloc_ret)
1051 goto unlock;
1052
1053 buf = zpool_map_handle(zpool, handle, ZPOOL_MM_WO);
1054 memcpy(buf, dst, dlen);
1055 zpool_unmap_handle(zpool, handle);
1056
1057 entry->handle = handle;
1058 entry->length = dlen;
1059
1060unlock:
1061 if (comp_ret == -ENOSPC || alloc_ret == -ENOSPC)
1062 zswap_reject_compress_poor++;
1063 else if (comp_ret)
1064 zswap_reject_compress_fail++;
1065 else if (alloc_ret)
1066 zswap_reject_alloc_fail++;
1067
1068 mutex_unlock(&acomp_ctx->mutex);
1069 return comp_ret == 0 && alloc_ret == 0;
1070}
1071
1072static void zswap_decompress(struct zswap_entry *entry, struct page *page)
1073{
1074 struct zpool *zpool = zswap_find_zpool(entry);
1075 struct scatterlist input, output;
1076 struct crypto_acomp_ctx *acomp_ctx;
1077 u8 *src;
1078
1079 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1080 mutex_lock(&acomp_ctx->mutex);
1081
1082 src = zpool_map_handle(zpool, entry->handle, ZPOOL_MM_RO);
1083 /*
1084 * If zpool_map_handle is atomic, we cannot reliably utilize its mapped buffer
1085 * to do crypto_acomp_decompress() which might sleep. In such cases, we must
1086 * resort to copying the buffer to a temporary one.
1087 * Meanwhile, zpool_map_handle() might return a non-linearly mapped buffer,
1088 * such as a kmap address of high memory or even ever a vmap address.
1089 * However, sg_init_one is only equipped to handle linearly mapped low memory.
1090 * In such cases, we also must copy the buffer to a temporary and lowmem one.
1091 */
1092 if ((acomp_ctx->is_sleepable && !zpool_can_sleep_mapped(zpool)) ||
1093 !virt_addr_valid(src)) {
1094 memcpy(acomp_ctx->buffer, src, entry->length);
1095 src = acomp_ctx->buffer;
1096 zpool_unmap_handle(zpool, entry->handle);
1097 }
1098
1099 sg_init_one(&input, src, entry->length);
1100 sg_init_table(&output, 1);
1101 sg_set_page(&output, page, PAGE_SIZE, 0);
1102 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
1103 BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
1104 BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
1105 mutex_unlock(&acomp_ctx->mutex);
1106
1107 if (src != acomp_ctx->buffer)
1108 zpool_unmap_handle(zpool, entry->handle);
1109}
1110
1111/*********************************
1112* writeback code
1113**********************************/
1114/*
1115 * Attempts to free an entry by adding a folio to the swap cache,
1116 * decompressing the entry data into the folio, and issuing a
1117 * bio write to write the folio back to the swap device.
1118 *
1119 * This can be thought of as a "resumed writeback" of the folio
1120 * to the swap device. We are basically resuming the same swap
1121 * writeback path that was intercepted with the zswap_store()
1122 * in the first place. After the folio has been decompressed into
1123 * the swap cache, the compressed version stored by zswap can be
1124 * freed.
1125 */
1126static int zswap_writeback_entry(struct zswap_entry *entry,
1127 swp_entry_t swpentry)
1128{
1129 struct zswap_tree *tree;
1130 struct folio *folio;
1131 struct mempolicy *mpol;
1132 bool folio_was_allocated;
1133 struct writeback_control wbc = {
1134 .sync_mode = WB_SYNC_NONE,
1135 };
1136
1137 /* try to allocate swap cache folio */
1138 mpol = get_task_policy(current);
1139 folio = __read_swap_cache_async(swpentry, GFP_KERNEL, mpol,
1140 NO_INTERLEAVE_INDEX, &folio_was_allocated, true);
1141 if (!folio)
1142 return -ENOMEM;
1143
1144 /*
1145 * Found an existing folio, we raced with swapin or concurrent
1146 * shrinker. We generally writeback cold folios from zswap, and
1147 * swapin means the folio just became hot, so skip this folio.
1148 * For unlikely concurrent shrinker case, it will be unlinked
1149 * and freed when invalidated by the concurrent shrinker anyway.
1150 */
1151 if (!folio_was_allocated) {
1152 folio_put(folio);
1153 return -EEXIST;
1154 }
1155
1156 /*
1157 * folio is locked, and the swapcache is now secured against
1158 * concurrent swapping to and from the slot, and concurrent
1159 * swapoff so we can safely dereference the zswap tree here.
1160 * Verify that the swap entry hasn't been invalidated and recycled
1161 * behind our backs, to avoid overwriting a new swap folio with
1162 * old compressed data. Only when this is successful can the entry
1163 * be dereferenced.
1164 */
1165 tree = swap_zswap_tree(swpentry);
1166 spin_lock(&tree->lock);
1167 if (zswap_rb_search(&tree->rbroot, swp_offset(swpentry)) != entry) {
1168 spin_unlock(&tree->lock);
1169 delete_from_swap_cache(folio);
1170 folio_unlock(folio);
1171 folio_put(folio);
1172 return -ENOMEM;
1173 }
1174
1175 /* Safe to deref entry after the entry is verified above. */
1176 zswap_rb_erase(&tree->rbroot, entry);
1177 spin_unlock(&tree->lock);
1178
1179 zswap_decompress(entry, &folio->page);
1180
1181 count_vm_event(ZSWPWB);
1182 if (entry->objcg)
1183 count_objcg_event(entry->objcg, ZSWPWB);
1184
1185 zswap_entry_free(entry);
1186
1187 /* folio is up to date */
1188 folio_mark_uptodate(folio);
1189
1190 /* move it to the tail of the inactive list after end_writeback */
1191 folio_set_reclaim(folio);
1192
1193 /* start writeback */
1194 __swap_writepage(folio, &wbc);
1195 folio_put(folio);
1196
1197 return 0;
1198}
1199
1200/*********************************
1201* shrinker functions
1202**********************************/
1203static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
1204 spinlock_t *lock, void *arg)
1205{
1206 struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
1207 bool *encountered_page_in_swapcache = (bool *)arg;
1208 swp_entry_t swpentry;
1209 enum lru_status ret = LRU_REMOVED_RETRY;
1210 int writeback_result;
1211
1212 /*
1213 * As soon as we drop the LRU lock, the entry can be freed by
1214 * a concurrent invalidation. This means the following:
1215 *
1216 * 1. We extract the swp_entry_t to the stack, allowing
1217 * zswap_writeback_entry() to pin the swap entry and
1218 * then validate the zwap entry against that swap entry's
1219 * tree using pointer value comparison. Only when that
1220 * is successful can the entry be dereferenced.
1221 *
1222 * 2. Usually, objects are taken off the LRU for reclaim. In
1223 * this case this isn't possible, because if reclaim fails
1224 * for whatever reason, we have no means of knowing if the
1225 * entry is alive to put it back on the LRU.
1226 *
1227 * So rotate it before dropping the lock. If the entry is
1228 * written back or invalidated, the free path will unlink
1229 * it. For failures, rotation is the right thing as well.
1230 *
1231 * Temporary failures, where the same entry should be tried
1232 * again immediately, almost never happen for this shrinker.
1233 * We don't do any trylocking; -ENOMEM comes closest,
1234 * but that's extremely rare and doesn't happen spuriously
1235 * either. Don't bother distinguishing this case.
1236 */
1237 list_move_tail(item, &l->list);
1238
1239 /*
1240 * Once the lru lock is dropped, the entry might get freed. The
1241 * swpentry is copied to the stack, and entry isn't deref'd again
1242 * until the entry is verified to still be alive in the tree.
1243 */
1244 swpentry = entry->swpentry;
1245
1246 /*
1247 * It's safe to drop the lock here because we return either
1248 * LRU_REMOVED_RETRY or LRU_RETRY.
1249 */
1250 spin_unlock(lock);
1251
1252 writeback_result = zswap_writeback_entry(entry, swpentry);
1253
1254 if (writeback_result) {
1255 zswap_reject_reclaim_fail++;
1256 ret = LRU_RETRY;
1257
1258 /*
1259 * Encountering a page already in swap cache is a sign that we are shrinking
1260 * into the warmer region. We should terminate shrinking (if we're in the dynamic
1261 * shrinker context).
1262 */
1263 if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
1264 ret = LRU_STOP;
1265 *encountered_page_in_swapcache = true;
1266 }
1267 } else {
1268 zswap_written_back_pages++;
1269 }
1270
1271 spin_lock(lock);
1272 return ret;
1273}
1274
1275static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
1276 struct shrink_control *sc)
1277{
1278 struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
1279 unsigned long shrink_ret, nr_protected, lru_size;
1280 bool encountered_page_in_swapcache = false;
1281
1282 if (!zswap_shrinker_enabled ||
1283 !mem_cgroup_zswap_writeback_enabled(sc->memcg)) {
1284 sc->nr_scanned = 0;
1285 return SHRINK_STOP;
1286 }
1287
1288 nr_protected =
1289 atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
1290 lru_size = list_lru_shrink_count(&zswap_list_lru, sc);
1291
1292 /*
1293 * Abort if we are shrinking into the protected region.
1294 *
1295 * This short-circuiting is necessary because if we have too many multiple
1296 * concurrent reclaimers getting the freeable zswap object counts at the
1297 * same time (before any of them made reasonable progress), the total
1298 * number of reclaimed objects might be more than the number of unprotected
1299 * objects (i.e the reclaimers will reclaim into the protected area of the
1300 * zswap LRU).
1301 */
1302 if (nr_protected >= lru_size - sc->nr_to_scan) {
1303 sc->nr_scanned = 0;
1304 return SHRINK_STOP;
1305 }
1306
1307 shrink_ret = list_lru_shrink_walk(&zswap_list_lru, sc, &shrink_memcg_cb,
1308 &encountered_page_in_swapcache);
1309
1310 if (encountered_page_in_swapcache)
1311 return SHRINK_STOP;
1312
1313 return shrink_ret ? shrink_ret : SHRINK_STOP;
1314}
1315
1316static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
1317 struct shrink_control *sc)
1318{
1319 struct mem_cgroup *memcg = sc->memcg;
1320 struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
1321 unsigned long nr_backing, nr_stored, nr_freeable, nr_protected;
1322
1323 if (!zswap_shrinker_enabled || !mem_cgroup_zswap_writeback_enabled(memcg))
1324 return 0;
1325
1326 /*
1327 * The shrinker resumes swap writeback, which will enter block
1328 * and may enter fs. XXX: Harmonize with vmscan.c __GFP_FS
1329 * rules (may_enter_fs()), which apply on a per-folio basis.
1330 */
1331 if (!gfp_has_io_fs(sc->gfp_mask))
1332 return 0;
1333
1334 /*
1335 * For memcg, use the cgroup-wide ZSWAP stats since we don't
1336 * have them per-node and thus per-lruvec. Careful if memcg is
1337 * runtime-disabled: we can get sc->memcg == NULL, which is ok
1338 * for the lruvec, but not for memcg_page_state().
1339 *
1340 * Without memcg, use the zswap pool-wide metrics.
1341 */
1342 if (!mem_cgroup_disabled()) {
1343 mem_cgroup_flush_stats(memcg);
1344 nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
1345 nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
1346 } else {
1347 nr_backing = zswap_pool_total_size >> PAGE_SHIFT;
1348 nr_stored = atomic_read(&zswap_nr_stored);
1349 }
1350
1351 if (!nr_stored)
1352 return 0;
1353
1354 nr_protected =
1355 atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
1356 nr_freeable = list_lru_shrink_count(&zswap_list_lru, sc);
1357 /*
1358 * Subtract the lru size by an estimate of the number of pages
1359 * that should be protected.
1360 */
1361 nr_freeable = nr_freeable > nr_protected ? nr_freeable - nr_protected : 0;
1362
1363 /*
1364 * Scale the number of freeable pages by the memory saving factor.
1365 * This ensures that the better zswap compresses memory, the fewer
1366 * pages we will evict to swap (as it will otherwise incur IO for
1367 * relatively small memory saving).
1368 */
1369 return mult_frac(nr_freeable, nr_backing, nr_stored);
1370}
1371
1372static struct shrinker *zswap_alloc_shrinker(void)
1373{
1374 struct shrinker *shrinker;
1375
1376 shrinker =
1377 shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
1378 if (!shrinker)
1379 return NULL;
1380
1381 shrinker->scan_objects = zswap_shrinker_scan;
1382 shrinker->count_objects = zswap_shrinker_count;
1383 shrinker->batch = 0;
1384 shrinker->seeks = DEFAULT_SEEKS;
1385 return shrinker;
1386}
1387
1388static int shrink_memcg(struct mem_cgroup *memcg)
1389{
1390 int nid, shrunk = 0;
1391
1392 if (!mem_cgroup_zswap_writeback_enabled(memcg))
1393 return -EINVAL;
1394
1395 /*
1396 * Skip zombies because their LRUs are reparented and we would be
1397 * reclaiming from the parent instead of the dead memcg.
1398 */
1399 if (memcg && !mem_cgroup_online(memcg))
1400 return -ENOENT;
1401
1402 for_each_node_state(nid, N_NORMAL_MEMORY) {
1403 unsigned long nr_to_walk = 1;
1404
1405 shrunk += list_lru_walk_one(&zswap_list_lru, nid, memcg,
1406 &shrink_memcg_cb, NULL, &nr_to_walk);
1407 }
1408 return shrunk ? 0 : -EAGAIN;
1409}
1410
1411static void shrink_worker(struct work_struct *w)
1412{
1413 struct mem_cgroup *memcg;
1414 int ret, failures = 0;
1415
1416 /* global reclaim will select cgroup in a round-robin fashion. */
1417 do {
1418 spin_lock(&zswap_shrink_lock);
1419 zswap_next_shrink = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
1420 memcg = zswap_next_shrink;
1421
1422 /*
1423 * We need to retry if we have gone through a full round trip, or if we
1424 * got an offline memcg (or else we risk undoing the effect of the
1425 * zswap memcg offlining cleanup callback). This is not catastrophic
1426 * per se, but it will keep the now offlined memcg hostage for a while.
1427 *
1428 * Note that if we got an online memcg, we will keep the extra
1429 * reference in case the original reference obtained by mem_cgroup_iter
1430 * is dropped by the zswap memcg offlining callback, ensuring that the
1431 * memcg is not killed when we are reclaiming.
1432 */
1433 if (!memcg) {
1434 spin_unlock(&zswap_shrink_lock);
1435 if (++failures == MAX_RECLAIM_RETRIES)
1436 break;
1437
1438 goto resched;
1439 }
1440
1441 if (!mem_cgroup_tryget_online(memcg)) {
1442 /* drop the reference from mem_cgroup_iter() */
1443 mem_cgroup_iter_break(NULL, memcg);
1444 zswap_next_shrink = NULL;
1445 spin_unlock(&zswap_shrink_lock);
1446
1447 if (++failures == MAX_RECLAIM_RETRIES)
1448 break;
1449
1450 goto resched;
1451 }
1452 spin_unlock(&zswap_shrink_lock);
1453
1454 ret = shrink_memcg(memcg);
1455 /* drop the extra reference */
1456 mem_cgroup_put(memcg);
1457
1458 if (ret == -EINVAL)
1459 break;
1460 if (ret && ++failures == MAX_RECLAIM_RETRIES)
1461 break;
1462
1463resched:
1464 cond_resched();
1465 } while (!zswap_can_accept());
1466}
1467
1468static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1469{
1470 unsigned long *page;
1471 unsigned long val;
1472 unsigned int pos, last_pos = PAGE_SIZE / sizeof(*page) - 1;
1473
1474 page = (unsigned long *)ptr;
1475 val = page[0];
1476
1477 if (val != page[last_pos])
1478 return 0;
1479
1480 for (pos = 1; pos < last_pos; pos++) {
1481 if (val != page[pos])
1482 return 0;
1483 }
1484
1485 *value = val;
1486
1487 return 1;
1488}
1489
1490static void zswap_fill_page(void *ptr, unsigned long value)
1491{
1492 unsigned long *page;
1493
1494 page = (unsigned long *)ptr;
1495 memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1496}
1497
1498bool zswap_store(struct folio *folio)
1499{
1500 swp_entry_t swp = folio->swap;
1501 pgoff_t offset = swp_offset(swp);
1502 struct zswap_tree *tree = swap_zswap_tree(swp);
1503 struct zswap_entry *entry, *dupentry;
1504 struct obj_cgroup *objcg = NULL;
1505 struct mem_cgroup *memcg = NULL;
1506
1507 VM_WARN_ON_ONCE(!folio_test_locked(folio));
1508 VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
1509
1510 /* Large folios aren't supported */
1511 if (folio_test_large(folio))
1512 return false;
1513
1514 if (!zswap_enabled)
1515 goto check_old;
1516
1517 objcg = get_obj_cgroup_from_folio(folio);
1518 if (objcg && !obj_cgroup_may_zswap(objcg)) {
1519 memcg = get_mem_cgroup_from_objcg(objcg);
1520 if (shrink_memcg(memcg)) {
1521 mem_cgroup_put(memcg);
1522 goto reject;
1523 }
1524 mem_cgroup_put(memcg);
1525 }
1526
1527 /* reclaim space if needed */
1528 if (zswap_is_full()) {
1529 zswap_pool_limit_hit++;
1530 zswap_pool_reached_full = true;
1531 goto shrink;
1532 }
1533
1534 if (zswap_pool_reached_full) {
1535 if (!zswap_can_accept())
1536 goto shrink;
1537 else
1538 zswap_pool_reached_full = false;
1539 }
1540
1541 /* allocate entry */
1542 entry = zswap_entry_cache_alloc(GFP_KERNEL, folio_nid(folio));
1543 if (!entry) {
1544 zswap_reject_kmemcache_fail++;
1545 goto reject;
1546 }
1547
1548 if (zswap_same_filled_pages_enabled) {
1549 unsigned long value;
1550 u8 *src;
1551
1552 src = kmap_local_folio(folio, 0);
1553 if (zswap_is_page_same_filled(src, &value)) {
1554 kunmap_local(src);
1555 entry->length = 0;
1556 entry->value = value;
1557 atomic_inc(&zswap_same_filled_pages);
1558 goto insert_entry;
1559 }
1560 kunmap_local(src);
1561 }
1562
1563 if (!zswap_non_same_filled_pages_enabled)
1564 goto freepage;
1565
1566 /* if entry is successfully added, it keeps the reference */
1567 entry->pool = zswap_pool_current_get();
1568 if (!entry->pool)
1569 goto freepage;
1570
1571 if (objcg) {
1572 memcg = get_mem_cgroup_from_objcg(objcg);
1573 if (memcg_list_lru_alloc(memcg, &zswap_list_lru, GFP_KERNEL)) {
1574 mem_cgroup_put(memcg);
1575 goto put_pool;
1576 }
1577 mem_cgroup_put(memcg);
1578 }
1579
1580 if (!zswap_compress(folio, entry))
1581 goto put_pool;
1582
1583insert_entry:
1584 entry->swpentry = swp;
1585 entry->objcg = objcg;
1586 if (objcg) {
1587 obj_cgroup_charge_zswap(objcg, entry->length);
1588 /* Account before objcg ref is moved to tree */
1589 count_objcg_event(objcg, ZSWPOUT);
1590 }
1591
1592 /* map */
1593 spin_lock(&tree->lock);
1594 /*
1595 * The folio may have been dirtied again, invalidate the
1596 * possibly stale entry before inserting the new entry.
1597 */
1598 if (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST) {
1599 zswap_invalidate_entry(tree, dupentry);
1600 WARN_ON(zswap_rb_insert(&tree->rbroot, entry, &dupentry));
1601 }
1602 if (entry->length) {
1603 INIT_LIST_HEAD(&entry->lru);
1604 zswap_lru_add(&zswap_list_lru, entry);
1605 atomic_inc(&zswap_nr_stored);
1606 }
1607 spin_unlock(&tree->lock);
1608
1609 /* update stats */
1610 atomic_inc(&zswap_stored_pages);
1611 zswap_update_total_size();
1612 count_vm_event(ZSWPOUT);
1613
1614 return true;
1615
1616put_pool:
1617 zswap_pool_put(entry->pool);
1618freepage:
1619 zswap_entry_cache_free(entry);
1620reject:
1621 if (objcg)
1622 obj_cgroup_put(objcg);
1623check_old:
1624 /*
1625 * If the zswap store fails or zswap is disabled, we must invalidate the
1626 * possibly stale entry which was previously stored at this offset.
1627 * Otherwise, writeback could overwrite the new data in the swapfile.
1628 */
1629 spin_lock(&tree->lock);
1630 entry = zswap_rb_search(&tree->rbroot, offset);
1631 if (entry)
1632 zswap_invalidate_entry(tree, entry);
1633 spin_unlock(&tree->lock);
1634 return false;
1635
1636shrink:
1637 queue_work(shrink_wq, &zswap_shrink_work);
1638 goto reject;
1639}
1640
1641bool zswap_load(struct folio *folio)
1642{
1643 swp_entry_t swp = folio->swap;
1644 pgoff_t offset = swp_offset(swp);
1645 struct page *page = &folio->page;
1646 bool swapcache = folio_test_swapcache(folio);
1647 struct zswap_tree *tree = swap_zswap_tree(swp);
1648 struct zswap_entry *entry;
1649 u8 *dst;
1650
1651 VM_WARN_ON_ONCE(!folio_test_locked(folio));
1652
1653 spin_lock(&tree->lock);
1654 entry = zswap_rb_search(&tree->rbroot, offset);
1655 if (!entry) {
1656 spin_unlock(&tree->lock);
1657 return false;
1658 }
1659 /*
1660 * When reading into the swapcache, invalidate our entry. The
1661 * swapcache can be the authoritative owner of the page and
1662 * its mappings, and the pressure that results from having two
1663 * in-memory copies outweighs any benefits of caching the
1664 * compression work.
1665 *
1666 * (Most swapins go through the swapcache. The notable
1667 * exception is the singleton fault on SWP_SYNCHRONOUS_IO
1668 * files, which reads into a private page and may free it if
1669 * the fault fails. We remain the primary owner of the entry.)
1670 */
1671 if (swapcache)
1672 zswap_rb_erase(&tree->rbroot, entry);
1673 spin_unlock(&tree->lock);
1674
1675 if (entry->length)
1676 zswap_decompress(entry, page);
1677 else {
1678 dst = kmap_local_page(page);
1679 zswap_fill_page(dst, entry->value);
1680 kunmap_local(dst);
1681 }
1682
1683 count_vm_event(ZSWPIN);
1684 if (entry->objcg)
1685 count_objcg_event(entry->objcg, ZSWPIN);
1686
1687 if (swapcache) {
1688 zswap_entry_free(entry);
1689 folio_mark_dirty(folio);
1690 }
1691
1692 return true;
1693}
1694
1695void zswap_invalidate(swp_entry_t swp)
1696{
1697 pgoff_t offset = swp_offset(swp);
1698 struct zswap_tree *tree = swap_zswap_tree(swp);
1699 struct zswap_entry *entry;
1700
1701 spin_lock(&tree->lock);
1702 entry = zswap_rb_search(&tree->rbroot, offset);
1703 if (entry)
1704 zswap_invalidate_entry(tree, entry);
1705 spin_unlock(&tree->lock);
1706}
1707
1708int zswap_swapon(int type, unsigned long nr_pages)
1709{
1710 struct zswap_tree *trees, *tree;
1711 unsigned int nr, i;
1712
1713 nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
1714 trees = kvcalloc(nr, sizeof(*tree), GFP_KERNEL);
1715 if (!trees) {
1716 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1717 return -ENOMEM;
1718 }
1719
1720 for (i = 0; i < nr; i++) {
1721 tree = trees + i;
1722 tree->rbroot = RB_ROOT;
1723 spin_lock_init(&tree->lock);
1724 }
1725
1726 nr_zswap_trees[type] = nr;
1727 zswap_trees[type] = trees;
1728 return 0;
1729}
1730
1731void zswap_swapoff(int type)
1732{
1733 struct zswap_tree *trees = zswap_trees[type];
1734 unsigned int i;
1735
1736 if (!trees)
1737 return;
1738
1739 /* try_to_unuse() invalidated all the entries already */
1740 for (i = 0; i < nr_zswap_trees[type]; i++)
1741 WARN_ON_ONCE(!RB_EMPTY_ROOT(&trees[i].rbroot));
1742
1743 kvfree(trees);
1744 nr_zswap_trees[type] = 0;
1745 zswap_trees[type] = NULL;
1746}
1747
1748/*********************************
1749* debugfs functions
1750**********************************/
1751#ifdef CONFIG_DEBUG_FS
1752#include <linux/debugfs.h>
1753
1754static struct dentry *zswap_debugfs_root;
1755
1756static int zswap_debugfs_init(void)
1757{
1758 if (!debugfs_initialized())
1759 return -ENODEV;
1760
1761 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1762
1763 debugfs_create_u64("pool_limit_hit", 0444,
1764 zswap_debugfs_root, &zswap_pool_limit_hit);
1765 debugfs_create_u64("reject_reclaim_fail", 0444,
1766 zswap_debugfs_root, &zswap_reject_reclaim_fail);
1767 debugfs_create_u64("reject_alloc_fail", 0444,
1768 zswap_debugfs_root, &zswap_reject_alloc_fail);
1769 debugfs_create_u64("reject_kmemcache_fail", 0444,
1770 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1771 debugfs_create_u64("reject_compress_fail", 0444,
1772 zswap_debugfs_root, &zswap_reject_compress_fail);
1773 debugfs_create_u64("reject_compress_poor", 0444,
1774 zswap_debugfs_root, &zswap_reject_compress_poor);
1775 debugfs_create_u64("written_back_pages", 0444,
1776 zswap_debugfs_root, &zswap_written_back_pages);
1777 debugfs_create_u64("pool_total_size", 0444,
1778 zswap_debugfs_root, &zswap_pool_total_size);
1779 debugfs_create_atomic_t("stored_pages", 0444,
1780 zswap_debugfs_root, &zswap_stored_pages);
1781 debugfs_create_atomic_t("same_filled_pages", 0444,
1782 zswap_debugfs_root, &zswap_same_filled_pages);
1783
1784 return 0;
1785}
1786#else
1787static int zswap_debugfs_init(void)
1788{
1789 return 0;
1790}
1791#endif
1792
1793/*********************************
1794* module init and exit
1795**********************************/
1796static int zswap_setup(void)
1797{
1798 struct zswap_pool *pool;
1799 int ret;
1800
1801 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
1802 if (!zswap_entry_cache) {
1803 pr_err("entry cache creation failed\n");
1804 goto cache_fail;
1805 }
1806
1807 ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1808 "mm/zswap_pool:prepare",
1809 zswap_cpu_comp_prepare,
1810 zswap_cpu_comp_dead);
1811 if (ret)
1812 goto hp_fail;
1813
1814 shrink_wq = alloc_workqueue("zswap-shrink",
1815 WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
1816 if (!shrink_wq)
1817 goto shrink_wq_fail;
1818
1819 zswap_shrinker = zswap_alloc_shrinker();
1820 if (!zswap_shrinker)
1821 goto shrinker_fail;
1822 if (list_lru_init_memcg(&zswap_list_lru, zswap_shrinker))
1823 goto lru_fail;
1824 shrinker_register(zswap_shrinker);
1825
1826 INIT_WORK(&zswap_shrink_work, shrink_worker);
1827
1828 pool = __zswap_pool_create_fallback();
1829 if (pool) {
1830 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1831 zpool_get_type(pool->zpools[0]));
1832 list_add(&pool->list, &zswap_pools);
1833 zswap_has_pool = true;
1834 } else {
1835 pr_err("pool creation failed\n");
1836 zswap_enabled = false;
1837 }
1838
1839 if (zswap_debugfs_init())
1840 pr_warn("debugfs initialization failed\n");
1841 zswap_init_state = ZSWAP_INIT_SUCCEED;
1842 return 0;
1843
1844lru_fail:
1845 shrinker_free(zswap_shrinker);
1846shrinker_fail:
1847 destroy_workqueue(shrink_wq);
1848shrink_wq_fail:
1849 cpuhp_remove_multi_state(CPUHP_MM_ZSWP_POOL_PREPARE);
1850hp_fail:
1851 kmem_cache_destroy(zswap_entry_cache);
1852cache_fail:
1853 /* if built-in, we aren't unloaded on failure; don't allow use */
1854 zswap_init_state = ZSWAP_INIT_FAILED;
1855 zswap_enabled = false;
1856 return -ENOMEM;
1857}
1858
1859static int __init zswap_init(void)
1860{
1861 if (!zswap_enabled)
1862 return 0;
1863 return zswap_setup();
1864}
1865/* must be late so crypto has time to come up */
1866late_initcall(zswap_init);
1867
1868MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1869MODULE_DESCRIPTION("Compressed cache for swap pages");