Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * zswap.c - zswap driver file
4 *
5 * zswap is a backend for frontswap 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/frontswap.h>
24#include <linux/rbtree.h>
25#include <linux/swap.h>
26#include <linux/crypto.h>
27#include <linux/scatterlist.h>
28#include <linux/mempool.h>
29#include <linux/zpool.h>
30#include <crypto/acompress.h>
31
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
39#include "swap.h"
40
41/*********************************
42* statistics
43**********************************/
44/* Total bytes used by the compressed storage */
45u64 zswap_pool_total_size;
46/* The number of compressed pages currently stored in zswap */
47atomic_t zswap_stored_pages = ATOMIC_INIT(0);
48/* The number of same-value filled pages currently stored in zswap */
49static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
50
51/*
52 * The statistics below are not protected from concurrent access for
53 * performance reasons so they may not be a 100% accurate. However,
54 * they do provide useful information on roughly how many times a
55 * certain event is occurring.
56*/
57
58/* Pool limit was hit (see zswap_max_pool_percent) */
59static u64 zswap_pool_limit_hit;
60/* Pages written back when pool limit was reached */
61static u64 zswap_written_back_pages;
62/* Store failed due to a reclaim failure after pool limit was reached */
63static u64 zswap_reject_reclaim_fail;
64/* Compressed page was too big for the allocator to (optimally) store */
65static u64 zswap_reject_compress_poor;
66/* Store failed because underlying allocator could not get memory */
67static u64 zswap_reject_alloc_fail;
68/* Store failed because the entry metadata could not be allocated (rare) */
69static u64 zswap_reject_kmemcache_fail;
70/* Duplicate store was encountered (rare) */
71static u64 zswap_duplicate_entry;
72
73/* Shrinker work queue */
74static struct workqueue_struct *shrink_wq;
75/* Pool limit was hit, we need to calm down */
76static bool zswap_pool_reached_full;
77
78/*********************************
79* tunables
80**********************************/
81
82#define ZSWAP_PARAM_UNSET ""
83
84/* Enable/disable zswap */
85static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
86static int zswap_enabled_param_set(const char *,
87 const struct kernel_param *);
88static const struct kernel_param_ops zswap_enabled_param_ops = {
89 .set = zswap_enabled_param_set,
90 .get = param_get_bool,
91};
92module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
93
94/* Crypto compressor to use */
95static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
96static int zswap_compressor_param_set(const char *,
97 const struct kernel_param *);
98static const struct kernel_param_ops zswap_compressor_param_ops = {
99 .set = zswap_compressor_param_set,
100 .get = param_get_charp,
101 .free = param_free_charp,
102};
103module_param_cb(compressor, &zswap_compressor_param_ops,
104 &zswap_compressor, 0644);
105
106/* Compressed storage zpool to use */
107static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
108static int zswap_zpool_param_set(const char *, const struct kernel_param *);
109static const struct kernel_param_ops zswap_zpool_param_ops = {
110 .set = zswap_zpool_param_set,
111 .get = param_get_charp,
112 .free = param_free_charp,
113};
114module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
115
116/* The maximum percentage of memory that the compressed pool can occupy */
117static unsigned int zswap_max_pool_percent = 20;
118module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
119
120/* The threshold for accepting new pages after the max_pool_percent was hit */
121static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
122module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
123 uint, 0644);
124
125/*
126 * Enable/disable handling same-value filled pages (enabled by default).
127 * If disabled every page is considered non-same-value filled.
128 */
129static bool zswap_same_filled_pages_enabled = true;
130module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
131 bool, 0644);
132
133/* Enable/disable handling non-same-value filled pages (enabled by default) */
134static bool zswap_non_same_filled_pages_enabled = true;
135module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
136 bool, 0644);
137
138/*********************************
139* data structures
140**********************************/
141
142struct crypto_acomp_ctx {
143 struct crypto_acomp *acomp;
144 struct acomp_req *req;
145 struct crypto_wait wait;
146 u8 *dstmem;
147 struct mutex *mutex;
148};
149
150struct zswap_pool {
151 struct zpool *zpool;
152 struct crypto_acomp_ctx __percpu *acomp_ctx;
153 struct kref kref;
154 struct list_head list;
155 struct work_struct release_work;
156 struct work_struct shrink_work;
157 struct hlist_node node;
158 char tfm_name[CRYPTO_MAX_ALG_NAME];
159};
160
161/*
162 * struct zswap_entry
163 *
164 * This structure contains the metadata for tracking a single compressed
165 * page within zswap.
166 *
167 * rbnode - links the entry into red-black tree for the appropriate swap type
168 * offset - the swap offset for the entry. Index into the red-black tree.
169 * refcount - the number of outstanding reference to the entry. This is needed
170 * to protect against premature freeing of the entry by code
171 * concurrent calls to load, invalidate, and writeback. The lock
172 * for the zswap_tree structure that contains the entry must
173 * be held while changing the refcount. Since the lock must
174 * be held, there is no reason to also make refcount atomic.
175 * length - the length in bytes of the compressed page data. Needed during
176 * decompression. For a same value filled page length is 0.
177 * pool - the zswap_pool the entry's data is in
178 * handle - zpool allocation handle that stores the compressed page data
179 * value - value of the same-value filled pages which have same content
180 */
181struct zswap_entry {
182 struct rb_node rbnode;
183 pgoff_t offset;
184 int refcount;
185 unsigned int length;
186 struct zswap_pool *pool;
187 union {
188 unsigned long handle;
189 unsigned long value;
190 };
191 struct obj_cgroup *objcg;
192};
193
194struct zswap_header {
195 swp_entry_t swpentry;
196};
197
198/*
199 * The tree lock in the zswap_tree struct protects a few things:
200 * - the rbtree
201 * - the refcount field of each entry in the tree
202 */
203struct zswap_tree {
204 struct rb_root rbroot;
205 spinlock_t lock;
206};
207
208static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
209
210/* RCU-protected iteration */
211static LIST_HEAD(zswap_pools);
212/* protects zswap_pools list modification */
213static DEFINE_SPINLOCK(zswap_pools_lock);
214/* pool counter to provide unique names to zpool */
215static atomic_t zswap_pools_count = ATOMIC_INIT(0);
216
217/* used by param callback function */
218static bool zswap_init_started;
219
220/* fatal error during init */
221static bool zswap_init_failed;
222
223/* init completed, but couldn't create the initial pool */
224static bool zswap_has_pool;
225
226/*********************************
227* helpers and fwd declarations
228**********************************/
229
230#define zswap_pool_debug(msg, p) \
231 pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name, \
232 zpool_get_type((p)->zpool))
233
234static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
235static int zswap_pool_get(struct zswap_pool *pool);
236static void zswap_pool_put(struct zswap_pool *pool);
237
238static const struct zpool_ops zswap_zpool_ops = {
239 .evict = zswap_writeback_entry
240};
241
242static bool zswap_is_full(void)
243{
244 return totalram_pages() * zswap_max_pool_percent / 100 <
245 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
246}
247
248static bool zswap_can_accept(void)
249{
250 return totalram_pages() * zswap_accept_thr_percent / 100 *
251 zswap_max_pool_percent / 100 >
252 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
253}
254
255static void zswap_update_total_size(void)
256{
257 struct zswap_pool *pool;
258 u64 total = 0;
259
260 rcu_read_lock();
261
262 list_for_each_entry_rcu(pool, &zswap_pools, list)
263 total += zpool_get_total_size(pool->zpool);
264
265 rcu_read_unlock();
266
267 zswap_pool_total_size = total;
268}
269
270/*********************************
271* zswap entry functions
272**********************************/
273static struct kmem_cache *zswap_entry_cache;
274
275static int __init zswap_entry_cache_create(void)
276{
277 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
278 return zswap_entry_cache == NULL;
279}
280
281static void __init zswap_entry_cache_destroy(void)
282{
283 kmem_cache_destroy(zswap_entry_cache);
284}
285
286static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
287{
288 struct zswap_entry *entry;
289 entry = kmem_cache_alloc(zswap_entry_cache, gfp);
290 if (!entry)
291 return NULL;
292 entry->refcount = 1;
293 RB_CLEAR_NODE(&entry->rbnode);
294 return entry;
295}
296
297static void zswap_entry_cache_free(struct zswap_entry *entry)
298{
299 kmem_cache_free(zswap_entry_cache, entry);
300}
301
302/*********************************
303* rbtree functions
304**********************************/
305static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
306{
307 struct rb_node *node = root->rb_node;
308 struct zswap_entry *entry;
309
310 while (node) {
311 entry = rb_entry(node, struct zswap_entry, rbnode);
312 if (entry->offset > offset)
313 node = node->rb_left;
314 else if (entry->offset < offset)
315 node = node->rb_right;
316 else
317 return entry;
318 }
319 return NULL;
320}
321
322/*
323 * In the case that a entry with the same offset is found, a pointer to
324 * the existing entry is stored in dupentry and the function returns -EEXIST
325 */
326static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
327 struct zswap_entry **dupentry)
328{
329 struct rb_node **link = &root->rb_node, *parent = NULL;
330 struct zswap_entry *myentry;
331
332 while (*link) {
333 parent = *link;
334 myentry = rb_entry(parent, struct zswap_entry, rbnode);
335 if (myentry->offset > entry->offset)
336 link = &(*link)->rb_left;
337 else if (myentry->offset < entry->offset)
338 link = &(*link)->rb_right;
339 else {
340 *dupentry = myentry;
341 return -EEXIST;
342 }
343 }
344 rb_link_node(&entry->rbnode, parent, link);
345 rb_insert_color(&entry->rbnode, root);
346 return 0;
347}
348
349static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
350{
351 if (!RB_EMPTY_NODE(&entry->rbnode)) {
352 rb_erase(&entry->rbnode, root);
353 RB_CLEAR_NODE(&entry->rbnode);
354 }
355}
356
357/*
358 * Carries out the common pattern of freeing and entry's zpool allocation,
359 * freeing the entry itself, and decrementing the number of stored pages.
360 */
361static void zswap_free_entry(struct zswap_entry *entry)
362{
363 if (entry->objcg) {
364 obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
365 obj_cgroup_put(entry->objcg);
366 }
367 if (!entry->length)
368 atomic_dec(&zswap_same_filled_pages);
369 else {
370 zpool_free(entry->pool->zpool, entry->handle);
371 zswap_pool_put(entry->pool);
372 }
373 zswap_entry_cache_free(entry);
374 atomic_dec(&zswap_stored_pages);
375 zswap_update_total_size();
376}
377
378/* caller must hold the tree lock */
379static void zswap_entry_get(struct zswap_entry *entry)
380{
381 entry->refcount++;
382}
383
384/* caller must hold the tree lock
385* remove from the tree and free it, if nobody reference the entry
386*/
387static void zswap_entry_put(struct zswap_tree *tree,
388 struct zswap_entry *entry)
389{
390 int refcount = --entry->refcount;
391
392 BUG_ON(refcount < 0);
393 if (refcount == 0) {
394 zswap_rb_erase(&tree->rbroot, entry);
395 zswap_free_entry(entry);
396 }
397}
398
399/* caller must hold the tree lock */
400static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
401 pgoff_t offset)
402{
403 struct zswap_entry *entry;
404
405 entry = zswap_rb_search(root, offset);
406 if (entry)
407 zswap_entry_get(entry);
408
409 return entry;
410}
411
412/*********************************
413* per-cpu code
414**********************************/
415static DEFINE_PER_CPU(u8 *, zswap_dstmem);
416/*
417 * If users dynamically change the zpool type and compressor at runtime, i.e.
418 * zswap is running, zswap can have more than one zpool on one cpu, but they
419 * are sharing dtsmem. So we need this mutex to be per-cpu.
420 */
421static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
422
423static int zswap_dstmem_prepare(unsigned int cpu)
424{
425 struct mutex *mutex;
426 u8 *dst;
427
428 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
429 if (!dst)
430 return -ENOMEM;
431
432 mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
433 if (!mutex) {
434 kfree(dst);
435 return -ENOMEM;
436 }
437
438 mutex_init(mutex);
439 per_cpu(zswap_dstmem, cpu) = dst;
440 per_cpu(zswap_mutex, cpu) = mutex;
441 return 0;
442}
443
444static int zswap_dstmem_dead(unsigned int cpu)
445{
446 struct mutex *mutex;
447 u8 *dst;
448
449 mutex = per_cpu(zswap_mutex, cpu);
450 kfree(mutex);
451 per_cpu(zswap_mutex, cpu) = NULL;
452
453 dst = per_cpu(zswap_dstmem, cpu);
454 kfree(dst);
455 per_cpu(zswap_dstmem, cpu) = NULL;
456
457 return 0;
458}
459
460static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
461{
462 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
463 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
464 struct crypto_acomp *acomp;
465 struct acomp_req *req;
466
467 acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
468 if (IS_ERR(acomp)) {
469 pr_err("could not alloc crypto acomp %s : %ld\n",
470 pool->tfm_name, PTR_ERR(acomp));
471 return PTR_ERR(acomp);
472 }
473 acomp_ctx->acomp = acomp;
474
475 req = acomp_request_alloc(acomp_ctx->acomp);
476 if (!req) {
477 pr_err("could not alloc crypto acomp_request %s\n",
478 pool->tfm_name);
479 crypto_free_acomp(acomp_ctx->acomp);
480 return -ENOMEM;
481 }
482 acomp_ctx->req = req;
483
484 crypto_init_wait(&acomp_ctx->wait);
485 /*
486 * if the backend of acomp is async zip, crypto_req_done() will wakeup
487 * crypto_wait_req(); if the backend of acomp is scomp, the callback
488 * won't be called, crypto_wait_req() will return without blocking.
489 */
490 acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
491 crypto_req_done, &acomp_ctx->wait);
492
493 acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
494 acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
495
496 return 0;
497}
498
499static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
500{
501 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
502 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
503
504 if (!IS_ERR_OR_NULL(acomp_ctx)) {
505 if (!IS_ERR_OR_NULL(acomp_ctx->req))
506 acomp_request_free(acomp_ctx->req);
507 if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
508 crypto_free_acomp(acomp_ctx->acomp);
509 }
510
511 return 0;
512}
513
514/*********************************
515* pool functions
516**********************************/
517
518static struct zswap_pool *__zswap_pool_current(void)
519{
520 struct zswap_pool *pool;
521
522 pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
523 WARN_ONCE(!pool && zswap_has_pool,
524 "%s: no page storage pool!\n", __func__);
525
526 return pool;
527}
528
529static struct zswap_pool *zswap_pool_current(void)
530{
531 assert_spin_locked(&zswap_pools_lock);
532
533 return __zswap_pool_current();
534}
535
536static struct zswap_pool *zswap_pool_current_get(void)
537{
538 struct zswap_pool *pool;
539
540 rcu_read_lock();
541
542 pool = __zswap_pool_current();
543 if (!zswap_pool_get(pool))
544 pool = NULL;
545
546 rcu_read_unlock();
547
548 return pool;
549}
550
551static struct zswap_pool *zswap_pool_last_get(void)
552{
553 struct zswap_pool *pool, *last = NULL;
554
555 rcu_read_lock();
556
557 list_for_each_entry_rcu(pool, &zswap_pools, list)
558 last = pool;
559 WARN_ONCE(!last && zswap_has_pool,
560 "%s: no page storage pool!\n", __func__);
561 if (!zswap_pool_get(last))
562 last = NULL;
563
564 rcu_read_unlock();
565
566 return last;
567}
568
569/* type and compressor must be null-terminated */
570static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
571{
572 struct zswap_pool *pool;
573
574 assert_spin_locked(&zswap_pools_lock);
575
576 list_for_each_entry_rcu(pool, &zswap_pools, list) {
577 if (strcmp(pool->tfm_name, compressor))
578 continue;
579 if (strcmp(zpool_get_type(pool->zpool), type))
580 continue;
581 /* if we can't get it, it's about to be destroyed */
582 if (!zswap_pool_get(pool))
583 continue;
584 return pool;
585 }
586
587 return NULL;
588}
589
590static void shrink_worker(struct work_struct *w)
591{
592 struct zswap_pool *pool = container_of(w, typeof(*pool),
593 shrink_work);
594
595 if (zpool_shrink(pool->zpool, 1, NULL))
596 zswap_reject_reclaim_fail++;
597 zswap_pool_put(pool);
598}
599
600static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
601{
602 struct zswap_pool *pool;
603 char name[38]; /* 'zswap' + 32 char (max) num + \0 */
604 gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
605 int ret;
606
607 if (!zswap_has_pool) {
608 /* if either are unset, pool initialization failed, and we
609 * need both params to be set correctly before trying to
610 * create a pool.
611 */
612 if (!strcmp(type, ZSWAP_PARAM_UNSET))
613 return NULL;
614 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
615 return NULL;
616 }
617
618 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
619 if (!pool)
620 return NULL;
621
622 /* unique name for each pool specifically required by zsmalloc */
623 snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
624
625 pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
626 if (!pool->zpool) {
627 pr_err("%s zpool not available\n", type);
628 goto error;
629 }
630 pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
631
632 strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
633
634 pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
635 if (!pool->acomp_ctx) {
636 pr_err("percpu alloc failed\n");
637 goto error;
638 }
639
640 ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
641 &pool->node);
642 if (ret)
643 goto error;
644 pr_debug("using %s compressor\n", pool->tfm_name);
645
646 /* being the current pool takes 1 ref; this func expects the
647 * caller to always add the new pool as the current pool
648 */
649 kref_init(&pool->kref);
650 INIT_LIST_HEAD(&pool->list);
651 INIT_WORK(&pool->shrink_work, shrink_worker);
652
653 zswap_pool_debug("created", pool);
654
655 return pool;
656
657error:
658 if (pool->acomp_ctx)
659 free_percpu(pool->acomp_ctx);
660 if (pool->zpool)
661 zpool_destroy_pool(pool->zpool);
662 kfree(pool);
663 return NULL;
664}
665
666static __init struct zswap_pool *__zswap_pool_create_fallback(void)
667{
668 bool has_comp, has_zpool;
669
670 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
671 if (!has_comp && strcmp(zswap_compressor,
672 CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
673 pr_err("compressor %s not available, using default %s\n",
674 zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
675 param_free_charp(&zswap_compressor);
676 zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
677 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
678 }
679 if (!has_comp) {
680 pr_err("default compressor %s not available\n",
681 zswap_compressor);
682 param_free_charp(&zswap_compressor);
683 zswap_compressor = ZSWAP_PARAM_UNSET;
684 }
685
686 has_zpool = zpool_has_pool(zswap_zpool_type);
687 if (!has_zpool && strcmp(zswap_zpool_type,
688 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
689 pr_err("zpool %s not available, using default %s\n",
690 zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
691 param_free_charp(&zswap_zpool_type);
692 zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
693 has_zpool = zpool_has_pool(zswap_zpool_type);
694 }
695 if (!has_zpool) {
696 pr_err("default zpool %s not available\n",
697 zswap_zpool_type);
698 param_free_charp(&zswap_zpool_type);
699 zswap_zpool_type = ZSWAP_PARAM_UNSET;
700 }
701
702 if (!has_comp || !has_zpool)
703 return NULL;
704
705 return zswap_pool_create(zswap_zpool_type, zswap_compressor);
706}
707
708static void zswap_pool_destroy(struct zswap_pool *pool)
709{
710 zswap_pool_debug("destroying", pool);
711
712 cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
713 free_percpu(pool->acomp_ctx);
714 zpool_destroy_pool(pool->zpool);
715 kfree(pool);
716}
717
718static int __must_check zswap_pool_get(struct zswap_pool *pool)
719{
720 if (!pool)
721 return 0;
722
723 return kref_get_unless_zero(&pool->kref);
724}
725
726static void __zswap_pool_release(struct work_struct *work)
727{
728 struct zswap_pool *pool = container_of(work, typeof(*pool),
729 release_work);
730
731 synchronize_rcu();
732
733 /* nobody should have been able to get a kref... */
734 WARN_ON(kref_get_unless_zero(&pool->kref));
735
736 /* pool is now off zswap_pools list and has no references. */
737 zswap_pool_destroy(pool);
738}
739
740static void __zswap_pool_empty(struct kref *kref)
741{
742 struct zswap_pool *pool;
743
744 pool = container_of(kref, typeof(*pool), kref);
745
746 spin_lock(&zswap_pools_lock);
747
748 WARN_ON(pool == zswap_pool_current());
749
750 list_del_rcu(&pool->list);
751
752 INIT_WORK(&pool->release_work, __zswap_pool_release);
753 schedule_work(&pool->release_work);
754
755 spin_unlock(&zswap_pools_lock);
756}
757
758static void zswap_pool_put(struct zswap_pool *pool)
759{
760 kref_put(&pool->kref, __zswap_pool_empty);
761}
762
763/*********************************
764* param callbacks
765**********************************/
766
767/* val must be a null-terminated string */
768static int __zswap_param_set(const char *val, const struct kernel_param *kp,
769 char *type, char *compressor)
770{
771 struct zswap_pool *pool, *put_pool = NULL;
772 char *s = strstrip((char *)val);
773 int ret;
774
775 if (zswap_init_failed) {
776 pr_err("can't set param, initialization failed\n");
777 return -ENODEV;
778 }
779
780 /* no change required */
781 if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
782 return 0;
783
784 /* if this is load-time (pre-init) param setting,
785 * don't create a pool; that's done during init.
786 */
787 if (!zswap_init_started)
788 return param_set_charp(s, kp);
789
790 if (!type) {
791 if (!zpool_has_pool(s)) {
792 pr_err("zpool %s not available\n", s);
793 return -ENOENT;
794 }
795 type = s;
796 } else if (!compressor) {
797 if (!crypto_has_acomp(s, 0, 0)) {
798 pr_err("compressor %s not available\n", s);
799 return -ENOENT;
800 }
801 compressor = s;
802 } else {
803 WARN_ON(1);
804 return -EINVAL;
805 }
806
807 spin_lock(&zswap_pools_lock);
808
809 pool = zswap_pool_find_get(type, compressor);
810 if (pool) {
811 zswap_pool_debug("using existing", pool);
812 WARN_ON(pool == zswap_pool_current());
813 list_del_rcu(&pool->list);
814 }
815
816 spin_unlock(&zswap_pools_lock);
817
818 if (!pool)
819 pool = zswap_pool_create(type, compressor);
820
821 if (pool)
822 ret = param_set_charp(s, kp);
823 else
824 ret = -EINVAL;
825
826 spin_lock(&zswap_pools_lock);
827
828 if (!ret) {
829 put_pool = zswap_pool_current();
830 list_add_rcu(&pool->list, &zswap_pools);
831 zswap_has_pool = true;
832 } else if (pool) {
833 /* add the possibly pre-existing pool to the end of the pools
834 * list; if it's new (and empty) then it'll be removed and
835 * destroyed by the put after we drop the lock
836 */
837 list_add_tail_rcu(&pool->list, &zswap_pools);
838 put_pool = pool;
839 }
840
841 spin_unlock(&zswap_pools_lock);
842
843 if (!zswap_has_pool && !pool) {
844 /* if initial pool creation failed, and this pool creation also
845 * failed, maybe both compressor and zpool params were bad.
846 * Allow changing this param, so pool creation will succeed
847 * when the other param is changed. We already verified this
848 * param is ok in the zpool_has_pool() or crypto_has_acomp()
849 * checks above.
850 */
851 ret = param_set_charp(s, kp);
852 }
853
854 /* drop the ref from either the old current pool,
855 * or the new pool we failed to add
856 */
857 if (put_pool)
858 zswap_pool_put(put_pool);
859
860 return ret;
861}
862
863static int zswap_compressor_param_set(const char *val,
864 const struct kernel_param *kp)
865{
866 return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
867}
868
869static int zswap_zpool_param_set(const char *val,
870 const struct kernel_param *kp)
871{
872 return __zswap_param_set(val, kp, NULL, zswap_compressor);
873}
874
875static int zswap_enabled_param_set(const char *val,
876 const struct kernel_param *kp)
877{
878 if (zswap_init_failed) {
879 pr_err("can't enable, initialization failed\n");
880 return -ENODEV;
881 }
882 if (!zswap_has_pool && zswap_init_started) {
883 pr_err("can't enable, no pool configured\n");
884 return -ENODEV;
885 }
886
887 return param_set_bool(val, kp);
888}
889
890/*********************************
891* writeback code
892**********************************/
893/* return enum for zswap_get_swap_cache_page */
894enum zswap_get_swap_ret {
895 ZSWAP_SWAPCACHE_NEW,
896 ZSWAP_SWAPCACHE_EXIST,
897 ZSWAP_SWAPCACHE_FAIL,
898};
899
900/*
901 * zswap_get_swap_cache_page
902 *
903 * This is an adaption of read_swap_cache_async()
904 *
905 * This function tries to find a page with the given swap entry
906 * in the swapper_space address space (the swap cache). If the page
907 * is found, it is returned in retpage. Otherwise, a page is allocated,
908 * added to the swap cache, and returned in retpage.
909 *
910 * If success, the swap cache page is returned in retpage
911 * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
912 * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
913 * the new page is added to swapcache and locked
914 * Returns ZSWAP_SWAPCACHE_FAIL on error
915 */
916static int zswap_get_swap_cache_page(swp_entry_t entry,
917 struct page **retpage)
918{
919 bool page_was_allocated;
920
921 *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
922 NULL, 0, &page_was_allocated);
923 if (page_was_allocated)
924 return ZSWAP_SWAPCACHE_NEW;
925 if (!*retpage)
926 return ZSWAP_SWAPCACHE_FAIL;
927 return ZSWAP_SWAPCACHE_EXIST;
928}
929
930/*
931 * Attempts to free an entry by adding a page to the swap cache,
932 * decompressing the entry data into the page, and issuing a
933 * bio write to write the page back to the swap device.
934 *
935 * This can be thought of as a "resumed writeback" of the page
936 * to the swap device. We are basically resuming the same swap
937 * writeback path that was intercepted with the frontswap_store()
938 * in the first place. After the page has been decompressed into
939 * the swap cache, the compressed version stored by zswap can be
940 * freed.
941 */
942static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
943{
944 struct zswap_header *zhdr;
945 swp_entry_t swpentry;
946 struct zswap_tree *tree;
947 pgoff_t offset;
948 struct zswap_entry *entry;
949 struct page *page;
950 struct scatterlist input, output;
951 struct crypto_acomp_ctx *acomp_ctx;
952
953 u8 *src, *tmp = NULL;
954 unsigned int dlen;
955 int ret;
956 struct writeback_control wbc = {
957 .sync_mode = WB_SYNC_NONE,
958 };
959
960 if (!zpool_can_sleep_mapped(pool)) {
961 tmp = kmalloc(PAGE_SIZE, GFP_KERNEL);
962 if (!tmp)
963 return -ENOMEM;
964 }
965
966 /* extract swpentry from data */
967 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
968 swpentry = zhdr->swpentry; /* here */
969 tree = zswap_trees[swp_type(swpentry)];
970 offset = swp_offset(swpentry);
971 zpool_unmap_handle(pool, handle);
972
973 /* find and ref zswap entry */
974 spin_lock(&tree->lock);
975 entry = zswap_entry_find_get(&tree->rbroot, offset);
976 if (!entry) {
977 /* entry was invalidated */
978 spin_unlock(&tree->lock);
979 kfree(tmp);
980 return 0;
981 }
982 spin_unlock(&tree->lock);
983 BUG_ON(offset != entry->offset);
984
985 /* try to allocate swap cache page */
986 switch (zswap_get_swap_cache_page(swpentry, &page)) {
987 case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
988 ret = -ENOMEM;
989 goto fail;
990
991 case ZSWAP_SWAPCACHE_EXIST:
992 /* page is already in the swap cache, ignore for now */
993 put_page(page);
994 ret = -EEXIST;
995 goto fail;
996
997 case ZSWAP_SWAPCACHE_NEW: /* page is locked */
998 /* decompress */
999 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1000 dlen = PAGE_SIZE;
1001
1002 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
1003 src = (u8 *)zhdr + sizeof(struct zswap_header);
1004 if (!zpool_can_sleep_mapped(pool)) {
1005 memcpy(tmp, src, entry->length);
1006 src = tmp;
1007 zpool_unmap_handle(pool, handle);
1008 }
1009
1010 mutex_lock(acomp_ctx->mutex);
1011 sg_init_one(&input, src, entry->length);
1012 sg_init_table(&output, 1);
1013 sg_set_page(&output, page, PAGE_SIZE, 0);
1014 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1015 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1016 dlen = acomp_ctx->req->dlen;
1017 mutex_unlock(acomp_ctx->mutex);
1018
1019 if (!zpool_can_sleep_mapped(pool))
1020 kfree(tmp);
1021 else
1022 zpool_unmap_handle(pool, handle);
1023
1024 BUG_ON(ret);
1025 BUG_ON(dlen != PAGE_SIZE);
1026
1027 /* page is up to date */
1028 SetPageUptodate(page);
1029 }
1030
1031 /* move it to the tail of the inactive list after end_writeback */
1032 SetPageReclaim(page);
1033
1034 /* start writeback */
1035 __swap_writepage(page, &wbc);
1036 put_page(page);
1037 zswap_written_back_pages++;
1038
1039 spin_lock(&tree->lock);
1040 /* drop local reference */
1041 zswap_entry_put(tree, entry);
1042
1043 /*
1044 * There are two possible situations for entry here:
1045 * (1) refcount is 1(normal case), entry is valid and on the tree
1046 * (2) refcount is 0, entry is freed and not on the tree
1047 * because invalidate happened during writeback
1048 * search the tree and free the entry if find entry
1049 */
1050 if (entry == zswap_rb_search(&tree->rbroot, offset))
1051 zswap_entry_put(tree, entry);
1052 spin_unlock(&tree->lock);
1053
1054 return ret;
1055
1056fail:
1057 if (!zpool_can_sleep_mapped(pool))
1058 kfree(tmp);
1059
1060 /*
1061 * if we get here due to ZSWAP_SWAPCACHE_EXIST
1062 * a load may be happening concurrently.
1063 * it is safe and okay to not free the entry.
1064 * if we free the entry in the following put
1065 * it is also okay to return !0
1066 */
1067 spin_lock(&tree->lock);
1068 zswap_entry_put(tree, entry);
1069 spin_unlock(&tree->lock);
1070
1071 return ret;
1072}
1073
1074static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1075{
1076 unsigned int pos;
1077 unsigned long *page;
1078
1079 page = (unsigned long *)ptr;
1080 for (pos = 1; pos < PAGE_SIZE / sizeof(*page); pos++) {
1081 if (page[pos] != page[0])
1082 return 0;
1083 }
1084 *value = page[0];
1085 return 1;
1086}
1087
1088static void zswap_fill_page(void *ptr, unsigned long value)
1089{
1090 unsigned long *page;
1091
1092 page = (unsigned long *)ptr;
1093 memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1094}
1095
1096/*********************************
1097* frontswap hooks
1098**********************************/
1099/* attempts to compress and store an single page */
1100static int zswap_frontswap_store(unsigned type, pgoff_t offset,
1101 struct page *page)
1102{
1103 struct zswap_tree *tree = zswap_trees[type];
1104 struct zswap_entry *entry, *dupentry;
1105 struct scatterlist input, output;
1106 struct crypto_acomp_ctx *acomp_ctx;
1107 struct obj_cgroup *objcg = NULL;
1108 struct zswap_pool *pool;
1109 int ret;
1110 unsigned int hlen, dlen = PAGE_SIZE;
1111 unsigned long handle, value;
1112 char *buf;
1113 u8 *src, *dst;
1114 struct zswap_header zhdr = { .swpentry = swp_entry(type, offset) };
1115 gfp_t gfp;
1116
1117 /* THP isn't supported */
1118 if (PageTransHuge(page)) {
1119 ret = -EINVAL;
1120 goto reject;
1121 }
1122
1123 if (!zswap_enabled || !tree) {
1124 ret = -ENODEV;
1125 goto reject;
1126 }
1127
1128 objcg = get_obj_cgroup_from_page(page);
1129 if (objcg && !obj_cgroup_may_zswap(objcg))
1130 goto shrink;
1131
1132 /* reclaim space if needed */
1133 if (zswap_is_full()) {
1134 zswap_pool_limit_hit++;
1135 zswap_pool_reached_full = true;
1136 goto shrink;
1137 }
1138
1139 if (zswap_pool_reached_full) {
1140 if (!zswap_can_accept()) {
1141 ret = -ENOMEM;
1142 goto reject;
1143 } else
1144 zswap_pool_reached_full = false;
1145 }
1146
1147 /* allocate entry */
1148 entry = zswap_entry_cache_alloc(GFP_KERNEL);
1149 if (!entry) {
1150 zswap_reject_kmemcache_fail++;
1151 ret = -ENOMEM;
1152 goto reject;
1153 }
1154
1155 if (zswap_same_filled_pages_enabled) {
1156 src = kmap_atomic(page);
1157 if (zswap_is_page_same_filled(src, &value)) {
1158 kunmap_atomic(src);
1159 entry->offset = offset;
1160 entry->length = 0;
1161 entry->value = value;
1162 atomic_inc(&zswap_same_filled_pages);
1163 goto insert_entry;
1164 }
1165 kunmap_atomic(src);
1166 }
1167
1168 if (!zswap_non_same_filled_pages_enabled) {
1169 ret = -EINVAL;
1170 goto freepage;
1171 }
1172
1173 /* if entry is successfully added, it keeps the reference */
1174 entry->pool = zswap_pool_current_get();
1175 if (!entry->pool) {
1176 ret = -EINVAL;
1177 goto freepage;
1178 }
1179
1180 /* compress */
1181 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1182
1183 mutex_lock(acomp_ctx->mutex);
1184
1185 dst = acomp_ctx->dstmem;
1186 sg_init_table(&input, 1);
1187 sg_set_page(&input, page, PAGE_SIZE, 0);
1188
1189 /* zswap_dstmem is of size (PAGE_SIZE * 2). Reflect same in sg_list */
1190 sg_init_one(&output, dst, PAGE_SIZE * 2);
1191 acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1192 /*
1193 * it maybe looks a little bit silly that we send an asynchronous request,
1194 * then wait for its completion synchronously. This makes the process look
1195 * synchronous in fact.
1196 * Theoretically, acomp supports users send multiple acomp requests in one
1197 * acomp instance, then get those requests done simultaneously. but in this
1198 * case, frontswap actually does store and load page by page, there is no
1199 * existing method to send the second page before the first page is done
1200 * in one thread doing frontswap.
1201 * but in different threads running on different cpu, we have different
1202 * acomp instance, so multiple threads can do (de)compression in parallel.
1203 */
1204 ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1205 dlen = acomp_ctx->req->dlen;
1206
1207 if (ret) {
1208 ret = -EINVAL;
1209 goto put_dstmem;
1210 }
1211
1212 /* store */
1213 hlen = zpool_evictable(entry->pool->zpool) ? sizeof(zhdr) : 0;
1214 gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1215 if (zpool_malloc_support_movable(entry->pool->zpool))
1216 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1217 ret = zpool_malloc(entry->pool->zpool, hlen + dlen, gfp, &handle);
1218 if (ret == -ENOSPC) {
1219 zswap_reject_compress_poor++;
1220 goto put_dstmem;
1221 }
1222 if (ret) {
1223 zswap_reject_alloc_fail++;
1224 goto put_dstmem;
1225 }
1226 buf = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_WO);
1227 memcpy(buf, &zhdr, hlen);
1228 memcpy(buf + hlen, dst, dlen);
1229 zpool_unmap_handle(entry->pool->zpool, handle);
1230 mutex_unlock(acomp_ctx->mutex);
1231
1232 /* populate entry */
1233 entry->offset = offset;
1234 entry->handle = handle;
1235 entry->length = dlen;
1236
1237insert_entry:
1238 entry->objcg = objcg;
1239 if (objcg) {
1240 obj_cgroup_charge_zswap(objcg, entry->length);
1241 /* Account before objcg ref is moved to tree */
1242 count_objcg_event(objcg, ZSWPOUT);
1243 }
1244
1245 /* map */
1246 spin_lock(&tree->lock);
1247 do {
1248 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1249 if (ret == -EEXIST) {
1250 zswap_duplicate_entry++;
1251 /* remove from rbtree */
1252 zswap_rb_erase(&tree->rbroot, dupentry);
1253 zswap_entry_put(tree, dupentry);
1254 }
1255 } while (ret == -EEXIST);
1256 spin_unlock(&tree->lock);
1257
1258 /* update stats */
1259 atomic_inc(&zswap_stored_pages);
1260 zswap_update_total_size();
1261 count_vm_event(ZSWPOUT);
1262
1263 return 0;
1264
1265put_dstmem:
1266 mutex_unlock(acomp_ctx->mutex);
1267 zswap_pool_put(entry->pool);
1268freepage:
1269 zswap_entry_cache_free(entry);
1270reject:
1271 if (objcg)
1272 obj_cgroup_put(objcg);
1273 return ret;
1274
1275shrink:
1276 pool = zswap_pool_last_get();
1277 if (pool)
1278 queue_work(shrink_wq, &pool->shrink_work);
1279 ret = -ENOMEM;
1280 goto reject;
1281}
1282
1283/*
1284 * returns 0 if the page was successfully decompressed
1285 * return -1 on entry not found or error
1286*/
1287static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1288 struct page *page)
1289{
1290 struct zswap_tree *tree = zswap_trees[type];
1291 struct zswap_entry *entry;
1292 struct scatterlist input, output;
1293 struct crypto_acomp_ctx *acomp_ctx;
1294 u8 *src, *dst, *tmp;
1295 unsigned int dlen;
1296 int ret;
1297
1298 /* find */
1299 spin_lock(&tree->lock);
1300 entry = zswap_entry_find_get(&tree->rbroot, offset);
1301 if (!entry) {
1302 /* entry was written back */
1303 spin_unlock(&tree->lock);
1304 return -1;
1305 }
1306 spin_unlock(&tree->lock);
1307
1308 if (!entry->length) {
1309 dst = kmap_atomic(page);
1310 zswap_fill_page(dst, entry->value);
1311 kunmap_atomic(dst);
1312 ret = 0;
1313 goto stats;
1314 }
1315
1316 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1317 tmp = kmalloc(entry->length, GFP_KERNEL);
1318 if (!tmp) {
1319 ret = -ENOMEM;
1320 goto freeentry;
1321 }
1322 }
1323
1324 /* decompress */
1325 dlen = PAGE_SIZE;
1326 src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
1327 if (zpool_evictable(entry->pool->zpool))
1328 src += sizeof(struct zswap_header);
1329
1330 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1331 memcpy(tmp, src, entry->length);
1332 src = tmp;
1333 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1334 }
1335
1336 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1337 mutex_lock(acomp_ctx->mutex);
1338 sg_init_one(&input, src, entry->length);
1339 sg_init_table(&output, 1);
1340 sg_set_page(&output, page, PAGE_SIZE, 0);
1341 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1342 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1343 mutex_unlock(acomp_ctx->mutex);
1344
1345 if (zpool_can_sleep_mapped(entry->pool->zpool))
1346 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1347 else
1348 kfree(tmp);
1349
1350 BUG_ON(ret);
1351stats:
1352 count_vm_event(ZSWPIN);
1353 if (entry->objcg)
1354 count_objcg_event(entry->objcg, ZSWPIN);
1355freeentry:
1356 spin_lock(&tree->lock);
1357 zswap_entry_put(tree, entry);
1358 spin_unlock(&tree->lock);
1359
1360 return ret;
1361}
1362
1363/* frees an entry in zswap */
1364static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1365{
1366 struct zswap_tree *tree = zswap_trees[type];
1367 struct zswap_entry *entry;
1368
1369 /* find */
1370 spin_lock(&tree->lock);
1371 entry = zswap_rb_search(&tree->rbroot, offset);
1372 if (!entry) {
1373 /* entry was written back */
1374 spin_unlock(&tree->lock);
1375 return;
1376 }
1377
1378 /* remove from rbtree */
1379 zswap_rb_erase(&tree->rbroot, entry);
1380
1381 /* drop the initial reference from entry creation */
1382 zswap_entry_put(tree, entry);
1383
1384 spin_unlock(&tree->lock);
1385}
1386
1387/* frees all zswap entries for the given swap type */
1388static void zswap_frontswap_invalidate_area(unsigned type)
1389{
1390 struct zswap_tree *tree = zswap_trees[type];
1391 struct zswap_entry *entry, *n;
1392
1393 if (!tree)
1394 return;
1395
1396 /* walk the tree and free everything */
1397 spin_lock(&tree->lock);
1398 rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1399 zswap_free_entry(entry);
1400 tree->rbroot = RB_ROOT;
1401 spin_unlock(&tree->lock);
1402 kfree(tree);
1403 zswap_trees[type] = NULL;
1404}
1405
1406static void zswap_frontswap_init(unsigned type)
1407{
1408 struct zswap_tree *tree;
1409
1410 tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1411 if (!tree) {
1412 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1413 return;
1414 }
1415
1416 tree->rbroot = RB_ROOT;
1417 spin_lock_init(&tree->lock);
1418 zswap_trees[type] = tree;
1419}
1420
1421static const struct frontswap_ops zswap_frontswap_ops = {
1422 .store = zswap_frontswap_store,
1423 .load = zswap_frontswap_load,
1424 .invalidate_page = zswap_frontswap_invalidate_page,
1425 .invalidate_area = zswap_frontswap_invalidate_area,
1426 .init = zswap_frontswap_init
1427};
1428
1429/*********************************
1430* debugfs functions
1431**********************************/
1432#ifdef CONFIG_DEBUG_FS
1433#include <linux/debugfs.h>
1434
1435static struct dentry *zswap_debugfs_root;
1436
1437static int __init zswap_debugfs_init(void)
1438{
1439 if (!debugfs_initialized())
1440 return -ENODEV;
1441
1442 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1443
1444 debugfs_create_u64("pool_limit_hit", 0444,
1445 zswap_debugfs_root, &zswap_pool_limit_hit);
1446 debugfs_create_u64("reject_reclaim_fail", 0444,
1447 zswap_debugfs_root, &zswap_reject_reclaim_fail);
1448 debugfs_create_u64("reject_alloc_fail", 0444,
1449 zswap_debugfs_root, &zswap_reject_alloc_fail);
1450 debugfs_create_u64("reject_kmemcache_fail", 0444,
1451 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1452 debugfs_create_u64("reject_compress_poor", 0444,
1453 zswap_debugfs_root, &zswap_reject_compress_poor);
1454 debugfs_create_u64("written_back_pages", 0444,
1455 zswap_debugfs_root, &zswap_written_back_pages);
1456 debugfs_create_u64("duplicate_entry", 0444,
1457 zswap_debugfs_root, &zswap_duplicate_entry);
1458 debugfs_create_u64("pool_total_size", 0444,
1459 zswap_debugfs_root, &zswap_pool_total_size);
1460 debugfs_create_atomic_t("stored_pages", 0444,
1461 zswap_debugfs_root, &zswap_stored_pages);
1462 debugfs_create_atomic_t("same_filled_pages", 0444,
1463 zswap_debugfs_root, &zswap_same_filled_pages);
1464
1465 return 0;
1466}
1467#else
1468static int __init zswap_debugfs_init(void)
1469{
1470 return 0;
1471}
1472#endif
1473
1474/*********************************
1475* module init and exit
1476**********************************/
1477static int __init init_zswap(void)
1478{
1479 struct zswap_pool *pool;
1480 int ret;
1481
1482 zswap_init_started = true;
1483
1484 if (zswap_entry_cache_create()) {
1485 pr_err("entry cache creation failed\n");
1486 goto cache_fail;
1487 }
1488
1489 ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1490 zswap_dstmem_prepare, zswap_dstmem_dead);
1491 if (ret) {
1492 pr_err("dstmem alloc failed\n");
1493 goto dstmem_fail;
1494 }
1495
1496 ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1497 "mm/zswap_pool:prepare",
1498 zswap_cpu_comp_prepare,
1499 zswap_cpu_comp_dead);
1500 if (ret)
1501 goto hp_fail;
1502
1503 pool = __zswap_pool_create_fallback();
1504 if (pool) {
1505 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1506 zpool_get_type(pool->zpool));
1507 list_add(&pool->list, &zswap_pools);
1508 zswap_has_pool = true;
1509 } else {
1510 pr_err("pool creation failed\n");
1511 zswap_enabled = false;
1512 }
1513
1514 shrink_wq = create_workqueue("zswap-shrink");
1515 if (!shrink_wq)
1516 goto fallback_fail;
1517
1518 ret = frontswap_register_ops(&zswap_frontswap_ops);
1519 if (ret)
1520 goto destroy_wq;
1521 if (zswap_debugfs_init())
1522 pr_warn("debugfs initialization failed\n");
1523 return 0;
1524
1525destroy_wq:
1526 destroy_workqueue(shrink_wq);
1527fallback_fail:
1528 if (pool)
1529 zswap_pool_destroy(pool);
1530hp_fail:
1531 cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1532dstmem_fail:
1533 zswap_entry_cache_destroy();
1534cache_fail:
1535 /* if built-in, we aren't unloaded on failure; don't allow use */
1536 zswap_init_failed = true;
1537 zswap_enabled = false;
1538 return -ENOMEM;
1539}
1540/* must be late so crypto has time to come up */
1541late_initcall(init_zswap);
1542
1543MODULE_LICENSE("GPL");
1544MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1545MODULE_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");