1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * zswap.c - zswap driver file
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.
11 * Copyright (C) 2012 Seth Jennings <sjenning@linux.vnet.ibm.com>
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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>
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>
39 /*********************************
41 **********************************/
42 /* Total bytes used by the compressed storage */
43 static u64 zswap_pool_total_size;
44 /* The number of compressed pages currently stored in zswap */
45 static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
46 /* The number of same-value filled pages currently stored in zswap */
47 static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
50 * The statistics below are not protected from concurrent access for
51 * performance reasons so they may not be a 100% accurate. However,
52 * they do provide useful information on roughly how many times a
53 * certain event is occurring.
56 /* Pool limit was hit (see zswap_max_pool_percent) */
57 static u64 zswap_pool_limit_hit;
58 /* Pages written back when pool limit was reached */
59 static u64 zswap_written_back_pages;
60 /* Store failed due to a reclaim failure after pool limit was reached */
61 static u64 zswap_reject_reclaim_fail;
62 /* Compressed page was too big for the allocator to (optimally) store */
63 static u64 zswap_reject_compress_poor;
64 /* Store failed because underlying allocator could not get memory */
65 static u64 zswap_reject_alloc_fail;
66 /* Store failed because the entry metadata could not be allocated (rare) */
67 static u64 zswap_reject_kmemcache_fail;
68 /* Duplicate store was encountered (rare) */
69 static u64 zswap_duplicate_entry;
71 /* Shrinker work queue */
72 static struct workqueue_struct *shrink_wq;
73 /* Pool limit was hit, we need to calm down */
74 static bool zswap_pool_reached_full;
76 /*********************************
78 **********************************/
80 #define ZSWAP_PARAM_UNSET ""
82 /* Enable/disable zswap */
83 static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
84 static int zswap_enabled_param_set(const char *,
85 const struct kernel_param *);
86 static const struct kernel_param_ops zswap_enabled_param_ops = {
87 .set = zswap_enabled_param_set,
88 .get = param_get_bool,
90 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
92 /* Crypto compressor to use */
93 static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
94 static int zswap_compressor_param_set(const char *,
95 const struct kernel_param *);
96 static const struct kernel_param_ops zswap_compressor_param_ops = {
97 .set = zswap_compressor_param_set,
98 .get = param_get_charp,
99 .free = param_free_charp,
101 module_param_cb(compressor, &zswap_compressor_param_ops,
102 &zswap_compressor, 0644);
104 /* Compressed storage zpool to use */
105 static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
106 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
107 static const struct kernel_param_ops zswap_zpool_param_ops = {
108 .set = zswap_zpool_param_set,
109 .get = param_get_charp,
110 .free = param_free_charp,
112 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
114 /* The maximum percentage of memory that the compressed pool can occupy */
115 static unsigned int zswap_max_pool_percent = 20;
116 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
118 /* The threshold for accepting new pages after the max_pool_percent was hit */
119 static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
120 module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
123 /* Enable/disable handling same-value filled pages (enabled by default) */
124 static bool zswap_same_filled_pages_enabled = true;
125 module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
128 /*********************************
130 **********************************/
132 struct crypto_acomp_ctx {
133 struct crypto_acomp *acomp;
134 struct acomp_req *req;
135 struct crypto_wait wait;
142 struct crypto_acomp_ctx __percpu *acomp_ctx;
144 struct list_head list;
145 struct work_struct release_work;
146 struct work_struct shrink_work;
147 struct hlist_node node;
148 char tfm_name[CRYPTO_MAX_ALG_NAME];
154 * This structure contains the metadata for tracking a single compressed
157 * rbnode - links the entry into red-black tree for the appropriate swap type
158 * offset - the swap offset for the entry. Index into the red-black tree.
159 * refcount - the number of outstanding reference to the entry. This is needed
160 * to protect against premature freeing of the entry by code
161 * concurrent calls to load, invalidate, and writeback. The lock
162 * for the zswap_tree structure that contains the entry must
163 * be held while changing the refcount. Since the lock must
164 * be held, there is no reason to also make refcount atomic.
165 * length - the length in bytes of the compressed page data. Needed during
166 * decompression. For a same value filled page length is 0.
167 * pool - the zswap_pool the entry's data is in
168 * handle - zpool allocation handle that stores the compressed page data
169 * value - value of the same-value filled pages which have same content
172 struct rb_node rbnode;
176 struct zswap_pool *pool;
178 unsigned long handle;
183 struct zswap_header {
184 swp_entry_t swpentry;
188 * The tree lock in the zswap_tree struct protects a few things:
190 * - the refcount field of each entry in the tree
193 struct rb_root rbroot;
197 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
199 /* RCU-protected iteration */
200 static LIST_HEAD(zswap_pools);
201 /* protects zswap_pools list modification */
202 static DEFINE_SPINLOCK(zswap_pools_lock);
203 /* pool counter to provide unique names to zpool */
204 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
206 /* used by param callback function */
207 static bool zswap_init_started;
209 /* fatal error during init */
210 static bool zswap_init_failed;
212 /* init completed, but couldn't create the initial pool */
213 static bool zswap_has_pool;
215 /*********************************
216 * helpers and fwd declarations
217 **********************************/
219 #define zswap_pool_debug(msg, p) \
220 pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name, \
221 zpool_get_type((p)->zpool))
223 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
224 static int zswap_pool_get(struct zswap_pool *pool);
225 static void zswap_pool_put(struct zswap_pool *pool);
227 static const struct zpool_ops zswap_zpool_ops = {
228 .evict = zswap_writeback_entry
231 static bool zswap_is_full(void)
233 return totalram_pages() * zswap_max_pool_percent / 100 <
234 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
237 static bool zswap_can_accept(void)
239 return totalram_pages() * zswap_accept_thr_percent / 100 *
240 zswap_max_pool_percent / 100 >
241 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
244 static void zswap_update_total_size(void)
246 struct zswap_pool *pool;
251 list_for_each_entry_rcu(pool, &zswap_pools, list)
252 total += zpool_get_total_size(pool->zpool);
256 zswap_pool_total_size = total;
259 /*********************************
260 * zswap entry functions
261 **********************************/
262 static struct kmem_cache *zswap_entry_cache;
264 static int __init zswap_entry_cache_create(void)
266 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
267 return zswap_entry_cache == NULL;
270 static void __init zswap_entry_cache_destroy(void)
272 kmem_cache_destroy(zswap_entry_cache);
275 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
277 struct zswap_entry *entry;
278 entry = kmem_cache_alloc(zswap_entry_cache, gfp);
282 RB_CLEAR_NODE(&entry->rbnode);
286 static void zswap_entry_cache_free(struct zswap_entry *entry)
288 kmem_cache_free(zswap_entry_cache, entry);
291 /*********************************
293 **********************************/
294 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
296 struct rb_node *node = root->rb_node;
297 struct zswap_entry *entry;
300 entry = rb_entry(node, struct zswap_entry, rbnode);
301 if (entry->offset > offset)
302 node = node->rb_left;
303 else if (entry->offset < offset)
304 node = node->rb_right;
312 * In the case that a entry with the same offset is found, a pointer to
313 * the existing entry is stored in dupentry and the function returns -EEXIST
315 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
316 struct zswap_entry **dupentry)
318 struct rb_node **link = &root->rb_node, *parent = NULL;
319 struct zswap_entry *myentry;
323 myentry = rb_entry(parent, struct zswap_entry, rbnode);
324 if (myentry->offset > entry->offset)
325 link = &(*link)->rb_left;
326 else if (myentry->offset < entry->offset)
327 link = &(*link)->rb_right;
333 rb_link_node(&entry->rbnode, parent, link);
334 rb_insert_color(&entry->rbnode, root);
338 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
340 if (!RB_EMPTY_NODE(&entry->rbnode)) {
341 rb_erase(&entry->rbnode, root);
342 RB_CLEAR_NODE(&entry->rbnode);
347 * Carries out the common pattern of freeing and entry's zpool allocation,
348 * freeing the entry itself, and decrementing the number of stored pages.
350 static void zswap_free_entry(struct zswap_entry *entry)
353 atomic_dec(&zswap_same_filled_pages);
355 zpool_free(entry->pool->zpool, entry->handle);
356 zswap_pool_put(entry->pool);
358 zswap_entry_cache_free(entry);
359 atomic_dec(&zswap_stored_pages);
360 zswap_update_total_size();
363 /* caller must hold the tree lock */
364 static void zswap_entry_get(struct zswap_entry *entry)
369 /* caller must hold the tree lock
370 * remove from the tree and free it, if nobody reference the entry
372 static void zswap_entry_put(struct zswap_tree *tree,
373 struct zswap_entry *entry)
375 int refcount = --entry->refcount;
377 BUG_ON(refcount < 0);
379 zswap_rb_erase(&tree->rbroot, entry);
380 zswap_free_entry(entry);
384 /* caller must hold the tree lock */
385 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
388 struct zswap_entry *entry;
390 entry = zswap_rb_search(root, offset);
392 zswap_entry_get(entry);
397 /*********************************
399 **********************************/
400 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
402 * If users dynamically change the zpool type and compressor at runtime, i.e.
403 * zswap is running, zswap can have more than one zpool on one cpu, but they
404 * are sharing dtsmem. So we need this mutex to be per-cpu.
406 static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
408 static int zswap_dstmem_prepare(unsigned int cpu)
413 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
417 mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
424 per_cpu(zswap_dstmem, cpu) = dst;
425 per_cpu(zswap_mutex, cpu) = mutex;
429 static int zswap_dstmem_dead(unsigned int cpu)
434 mutex = per_cpu(zswap_mutex, cpu);
436 per_cpu(zswap_mutex, cpu) = NULL;
438 dst = per_cpu(zswap_dstmem, cpu);
440 per_cpu(zswap_dstmem, cpu) = NULL;
445 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
447 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
448 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
449 struct crypto_acomp *acomp;
450 struct acomp_req *req;
452 acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
454 pr_err("could not alloc crypto acomp %s : %ld\n",
455 pool->tfm_name, PTR_ERR(acomp));
456 return PTR_ERR(acomp);
458 acomp_ctx->acomp = acomp;
460 req = acomp_request_alloc(acomp_ctx->acomp);
462 pr_err("could not alloc crypto acomp_request %s\n",
464 crypto_free_acomp(acomp_ctx->acomp);
467 acomp_ctx->req = req;
469 crypto_init_wait(&acomp_ctx->wait);
471 * if the backend of acomp is async zip, crypto_req_done() will wakeup
472 * crypto_wait_req(); if the backend of acomp is scomp, the callback
473 * won't be called, crypto_wait_req() will return without blocking.
475 acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
476 crypto_req_done, &acomp_ctx->wait);
478 acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
479 acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
484 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
486 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
487 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
489 if (!IS_ERR_OR_NULL(acomp_ctx)) {
490 if (!IS_ERR_OR_NULL(acomp_ctx->req))
491 acomp_request_free(acomp_ctx->req);
492 if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
493 crypto_free_acomp(acomp_ctx->acomp);
499 /*********************************
501 **********************************/
503 static struct zswap_pool *__zswap_pool_current(void)
505 struct zswap_pool *pool;
507 pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
508 WARN_ONCE(!pool && zswap_has_pool,
509 "%s: no page storage pool!\n", __func__);
514 static struct zswap_pool *zswap_pool_current(void)
516 assert_spin_locked(&zswap_pools_lock);
518 return __zswap_pool_current();
521 static struct zswap_pool *zswap_pool_current_get(void)
523 struct zswap_pool *pool;
527 pool = __zswap_pool_current();
528 if (!zswap_pool_get(pool))
536 static struct zswap_pool *zswap_pool_last_get(void)
538 struct zswap_pool *pool, *last = NULL;
542 list_for_each_entry_rcu(pool, &zswap_pools, list)
544 WARN_ONCE(!last && zswap_has_pool,
545 "%s: no page storage pool!\n", __func__);
546 if (!zswap_pool_get(last))
554 /* type and compressor must be null-terminated */
555 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
557 struct zswap_pool *pool;
559 assert_spin_locked(&zswap_pools_lock);
561 list_for_each_entry_rcu(pool, &zswap_pools, list) {
562 if (strcmp(pool->tfm_name, compressor))
564 if (strcmp(zpool_get_type(pool->zpool), type))
566 /* if we can't get it, it's about to be destroyed */
567 if (!zswap_pool_get(pool))
575 static void shrink_worker(struct work_struct *w)
577 struct zswap_pool *pool = container_of(w, typeof(*pool),
580 if (zpool_shrink(pool->zpool, 1, NULL))
581 zswap_reject_reclaim_fail++;
582 zswap_pool_put(pool);
585 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
587 struct zswap_pool *pool;
588 char name[38]; /* 'zswap' + 32 char (max) num + \0 */
589 gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
592 if (!zswap_has_pool) {
593 /* if either are unset, pool initialization failed, and we
594 * need both params to be set correctly before trying to
597 if (!strcmp(type, ZSWAP_PARAM_UNSET))
599 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
603 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
607 /* unique name for each pool specifically required by zsmalloc */
608 snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
610 pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
612 pr_err("%s zpool not available\n", type);
615 pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
617 strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
619 pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
620 if (!pool->acomp_ctx) {
621 pr_err("percpu alloc failed\n");
625 ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
629 pr_debug("using %s compressor\n", pool->tfm_name);
631 /* being the current pool takes 1 ref; this func expects the
632 * caller to always add the new pool as the current pool
634 kref_init(&pool->kref);
635 INIT_LIST_HEAD(&pool->list);
636 INIT_WORK(&pool->shrink_work, shrink_worker);
638 zswap_pool_debug("created", pool);
644 free_percpu(pool->acomp_ctx);
646 zpool_destroy_pool(pool->zpool);
651 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
653 bool has_comp, has_zpool;
655 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
656 if (!has_comp && strcmp(zswap_compressor,
657 CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
658 pr_err("compressor %s not available, using default %s\n",
659 zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
660 param_free_charp(&zswap_compressor);
661 zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
662 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
665 pr_err("default compressor %s not available\n",
667 param_free_charp(&zswap_compressor);
668 zswap_compressor = ZSWAP_PARAM_UNSET;
671 has_zpool = zpool_has_pool(zswap_zpool_type);
672 if (!has_zpool && strcmp(zswap_zpool_type,
673 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
674 pr_err("zpool %s not available, using default %s\n",
675 zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
676 param_free_charp(&zswap_zpool_type);
677 zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
678 has_zpool = zpool_has_pool(zswap_zpool_type);
681 pr_err("default zpool %s not available\n",
683 param_free_charp(&zswap_zpool_type);
684 zswap_zpool_type = ZSWAP_PARAM_UNSET;
687 if (!has_comp || !has_zpool)
690 return zswap_pool_create(zswap_zpool_type, zswap_compressor);
693 static void zswap_pool_destroy(struct zswap_pool *pool)
695 zswap_pool_debug("destroying", pool);
697 cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
698 free_percpu(pool->acomp_ctx);
699 zpool_destroy_pool(pool->zpool);
703 static int __must_check zswap_pool_get(struct zswap_pool *pool)
708 return kref_get_unless_zero(&pool->kref);
711 static void __zswap_pool_release(struct work_struct *work)
713 struct zswap_pool *pool = container_of(work, typeof(*pool),
718 /* nobody should have been able to get a kref... */
719 WARN_ON(kref_get_unless_zero(&pool->kref));
721 /* pool is now off zswap_pools list and has no references. */
722 zswap_pool_destroy(pool);
725 static void __zswap_pool_empty(struct kref *kref)
727 struct zswap_pool *pool;
729 pool = container_of(kref, typeof(*pool), kref);
731 spin_lock(&zswap_pools_lock);
733 WARN_ON(pool == zswap_pool_current());
735 list_del_rcu(&pool->list);
737 INIT_WORK(&pool->release_work, __zswap_pool_release);
738 schedule_work(&pool->release_work);
740 spin_unlock(&zswap_pools_lock);
743 static void zswap_pool_put(struct zswap_pool *pool)
745 kref_put(&pool->kref, __zswap_pool_empty);
748 /*********************************
750 **********************************/
752 /* val must be a null-terminated string */
753 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
754 char *type, char *compressor)
756 struct zswap_pool *pool, *put_pool = NULL;
757 char *s = strstrip((char *)val);
760 if (zswap_init_failed) {
761 pr_err("can't set param, initialization failed\n");
765 /* no change required */
766 if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
769 /* if this is load-time (pre-init) param setting,
770 * don't create a pool; that's done during init.
772 if (!zswap_init_started)
773 return param_set_charp(s, kp);
776 if (!zpool_has_pool(s)) {
777 pr_err("zpool %s not available\n", s);
781 } else if (!compressor) {
782 if (!crypto_has_acomp(s, 0, 0)) {
783 pr_err("compressor %s not available\n", s);
792 spin_lock(&zswap_pools_lock);
794 pool = zswap_pool_find_get(type, compressor);
796 zswap_pool_debug("using existing", pool);
797 WARN_ON(pool == zswap_pool_current());
798 list_del_rcu(&pool->list);
801 spin_unlock(&zswap_pools_lock);
804 pool = zswap_pool_create(type, compressor);
807 ret = param_set_charp(s, kp);
811 spin_lock(&zswap_pools_lock);
814 put_pool = zswap_pool_current();
815 list_add_rcu(&pool->list, &zswap_pools);
816 zswap_has_pool = true;
818 /* add the possibly pre-existing pool to the end of the pools
819 * list; if it's new (and empty) then it'll be removed and
820 * destroyed by the put after we drop the lock
822 list_add_tail_rcu(&pool->list, &zswap_pools);
826 spin_unlock(&zswap_pools_lock);
828 if (!zswap_has_pool && !pool) {
829 /* if initial pool creation failed, and this pool creation also
830 * failed, maybe both compressor and zpool params were bad.
831 * Allow changing this param, so pool creation will succeed
832 * when the other param is changed. We already verified this
833 * param is ok in the zpool_has_pool() or crypto_has_acomp()
836 ret = param_set_charp(s, kp);
839 /* drop the ref from either the old current pool,
840 * or the new pool we failed to add
843 zswap_pool_put(put_pool);
848 static int zswap_compressor_param_set(const char *val,
849 const struct kernel_param *kp)
851 return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
854 static int zswap_zpool_param_set(const char *val,
855 const struct kernel_param *kp)
857 return __zswap_param_set(val, kp, NULL, zswap_compressor);
860 static int zswap_enabled_param_set(const char *val,
861 const struct kernel_param *kp)
863 if (zswap_init_failed) {
864 pr_err("can't enable, initialization failed\n");
867 if (!zswap_has_pool && zswap_init_started) {
868 pr_err("can't enable, no pool configured\n");
872 return param_set_bool(val, kp);
875 /*********************************
877 **********************************/
878 /* return enum for zswap_get_swap_cache_page */
879 enum zswap_get_swap_ret {
881 ZSWAP_SWAPCACHE_EXIST,
882 ZSWAP_SWAPCACHE_FAIL,
886 * zswap_get_swap_cache_page
888 * This is an adaption of read_swap_cache_async()
890 * This function tries to find a page with the given swap entry
891 * in the swapper_space address space (the swap cache). If the page
892 * is found, it is returned in retpage. Otherwise, a page is allocated,
893 * added to the swap cache, and returned in retpage.
895 * If success, the swap cache page is returned in retpage
896 * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
897 * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
898 * the new page is added to swapcache and locked
899 * Returns ZSWAP_SWAPCACHE_FAIL on error
901 static int zswap_get_swap_cache_page(swp_entry_t entry,
902 struct page **retpage)
904 bool page_was_allocated;
906 *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
907 NULL, 0, &page_was_allocated);
908 if (page_was_allocated)
909 return ZSWAP_SWAPCACHE_NEW;
911 return ZSWAP_SWAPCACHE_FAIL;
912 return ZSWAP_SWAPCACHE_EXIST;
916 * Attempts to free an entry by adding a page to the swap cache,
917 * decompressing the entry data into the page, and issuing a
918 * bio write to write the page back to the swap device.
920 * This can be thought of as a "resumed writeback" of the page
921 * to the swap device. We are basically resuming the same swap
922 * writeback path that was intercepted with the frontswap_store()
923 * in the first place. After the page has been decompressed into
924 * the swap cache, the compressed version stored by zswap can be
927 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
929 struct zswap_header *zhdr;
930 swp_entry_t swpentry;
931 struct zswap_tree *tree;
933 struct zswap_entry *entry;
935 struct scatterlist input, output;
936 struct crypto_acomp_ctx *acomp_ctx;
938 u8 *src, *tmp = NULL;
941 struct writeback_control wbc = {
942 .sync_mode = WB_SYNC_NONE,
945 if (!zpool_can_sleep_mapped(pool)) {
946 tmp = kmalloc(PAGE_SIZE, GFP_ATOMIC);
951 /* extract swpentry from data */
952 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
953 swpentry = zhdr->swpentry; /* here */
954 tree = zswap_trees[swp_type(swpentry)];
955 offset = swp_offset(swpentry);
957 /* find and ref zswap entry */
958 spin_lock(&tree->lock);
959 entry = zswap_entry_find_get(&tree->rbroot, offset);
961 /* entry was invalidated */
962 spin_unlock(&tree->lock);
963 zpool_unmap_handle(pool, handle);
967 spin_unlock(&tree->lock);
968 BUG_ON(offset != entry->offset);
970 src = (u8 *)zhdr + sizeof(struct zswap_header);
971 if (!zpool_can_sleep_mapped(pool)) {
972 memcpy(tmp, src, entry->length);
974 zpool_unmap_handle(pool, handle);
977 /* try to allocate swap cache page */
978 switch (zswap_get_swap_cache_page(swpentry, &page)) {
979 case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
983 case ZSWAP_SWAPCACHE_EXIST:
984 /* page is already in the swap cache, ignore for now */
989 case ZSWAP_SWAPCACHE_NEW: /* page is locked */
991 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
994 mutex_lock(acomp_ctx->mutex);
995 sg_init_one(&input, src, entry->length);
996 sg_init_table(&output, 1);
997 sg_set_page(&output, page, PAGE_SIZE, 0);
998 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
999 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1000 dlen = acomp_ctx->req->dlen;
1001 mutex_unlock(acomp_ctx->mutex);
1004 BUG_ON(dlen != PAGE_SIZE);
1006 /* page is up to date */
1007 SetPageUptodate(page);
1010 /* move it to the tail of the inactive list after end_writeback */
1011 SetPageReclaim(page);
1013 /* start writeback */
1014 __swap_writepage(page, &wbc, end_swap_bio_write);
1016 zswap_written_back_pages++;
1018 spin_lock(&tree->lock);
1019 /* drop local reference */
1020 zswap_entry_put(tree, entry);
1023 * There are two possible situations for entry here:
1024 * (1) refcount is 1(normal case), entry is valid and on the tree
1025 * (2) refcount is 0, entry is freed and not on the tree
1026 * because invalidate happened during writeback
1027 * search the tree and free the entry if find entry
1029 if (entry == zswap_rb_search(&tree->rbroot, offset))
1030 zswap_entry_put(tree, entry);
1031 spin_unlock(&tree->lock);
1036 * if we get here due to ZSWAP_SWAPCACHE_EXIST
1037 * a load may be happening concurrently.
1038 * it is safe and okay to not free the entry.
1039 * if we free the entry in the following put
1040 * it is also okay to return !0
1043 spin_lock(&tree->lock);
1044 zswap_entry_put(tree, entry);
1045 spin_unlock(&tree->lock);
1048 if (zpool_can_sleep_mapped(pool))
1049 zpool_unmap_handle(pool, handle);
1056 static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1059 unsigned long *page;
1061 page = (unsigned long *)ptr;
1062 for (pos = 1; pos < PAGE_SIZE / sizeof(*page); pos++) {
1063 if (page[pos] != page[0])
1070 static void zswap_fill_page(void *ptr, unsigned long value)
1072 unsigned long *page;
1074 page = (unsigned long *)ptr;
1075 memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1078 /*********************************
1080 **********************************/
1081 /* attempts to compress and store an single page */
1082 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
1085 struct zswap_tree *tree = zswap_trees[type];
1086 struct zswap_entry *entry, *dupentry;
1087 struct scatterlist input, output;
1088 struct crypto_acomp_ctx *acomp_ctx;
1090 unsigned int hlen, dlen = PAGE_SIZE;
1091 unsigned long handle, value;
1094 struct zswap_header zhdr = { .swpentry = swp_entry(type, offset) };
1097 /* THP isn't supported */
1098 if (PageTransHuge(page)) {
1103 if (!zswap_enabled || !tree) {
1108 /* reclaim space if needed */
1109 if (zswap_is_full()) {
1110 struct zswap_pool *pool;
1112 zswap_pool_limit_hit++;
1113 zswap_pool_reached_full = true;
1114 pool = zswap_pool_last_get();
1116 queue_work(shrink_wq, &pool->shrink_work);
1121 if (zswap_pool_reached_full) {
1122 if (!zswap_can_accept()) {
1126 zswap_pool_reached_full = false;
1129 /* allocate entry */
1130 entry = zswap_entry_cache_alloc(GFP_KERNEL);
1132 zswap_reject_kmemcache_fail++;
1137 if (zswap_same_filled_pages_enabled) {
1138 src = kmap_atomic(page);
1139 if (zswap_is_page_same_filled(src, &value)) {
1141 entry->offset = offset;
1143 entry->value = value;
1144 atomic_inc(&zswap_same_filled_pages);
1150 /* if entry is successfully added, it keeps the reference */
1151 entry->pool = zswap_pool_current_get();
1158 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1160 mutex_lock(acomp_ctx->mutex);
1162 dst = acomp_ctx->dstmem;
1163 sg_init_table(&input, 1);
1164 sg_set_page(&input, page, PAGE_SIZE, 0);
1166 /* zswap_dstmem is of size (PAGE_SIZE * 2). Reflect same in sg_list */
1167 sg_init_one(&output, dst, PAGE_SIZE * 2);
1168 acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1170 * it maybe looks a little bit silly that we send an asynchronous request,
1171 * then wait for its completion synchronously. This makes the process look
1172 * synchronous in fact.
1173 * Theoretically, acomp supports users send multiple acomp requests in one
1174 * acomp instance, then get those requests done simultaneously. but in this
1175 * case, frontswap actually does store and load page by page, there is no
1176 * existing method to send the second page before the first page is done
1177 * in one thread doing frontswap.
1178 * but in different threads running on different cpu, we have different
1179 * acomp instance, so multiple threads can do (de)compression in parallel.
1181 ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1182 dlen = acomp_ctx->req->dlen;
1190 hlen = zpool_evictable(entry->pool->zpool) ? sizeof(zhdr) : 0;
1191 gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1192 if (zpool_malloc_support_movable(entry->pool->zpool))
1193 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1194 ret = zpool_malloc(entry->pool->zpool, hlen + dlen, gfp, &handle);
1195 if (ret == -ENOSPC) {
1196 zswap_reject_compress_poor++;
1200 zswap_reject_alloc_fail++;
1203 buf = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_WO);
1204 memcpy(buf, &zhdr, hlen);
1205 memcpy(buf + hlen, dst, dlen);
1206 zpool_unmap_handle(entry->pool->zpool, handle);
1207 mutex_unlock(acomp_ctx->mutex);
1209 /* populate entry */
1210 entry->offset = offset;
1211 entry->handle = handle;
1212 entry->length = dlen;
1216 spin_lock(&tree->lock);
1218 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1219 if (ret == -EEXIST) {
1220 zswap_duplicate_entry++;
1221 /* remove from rbtree */
1222 zswap_rb_erase(&tree->rbroot, dupentry);
1223 zswap_entry_put(tree, dupentry);
1225 } while (ret == -EEXIST);
1226 spin_unlock(&tree->lock);
1229 atomic_inc(&zswap_stored_pages);
1230 zswap_update_total_size();
1235 mutex_unlock(acomp_ctx->mutex);
1236 zswap_pool_put(entry->pool);
1238 zswap_entry_cache_free(entry);
1244 * returns 0 if the page was successfully decompressed
1245 * return -1 on entry not found or error
1247 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1250 struct zswap_tree *tree = zswap_trees[type];
1251 struct zswap_entry *entry;
1252 struct scatterlist input, output;
1253 struct crypto_acomp_ctx *acomp_ctx;
1254 u8 *src, *dst, *tmp;
1259 spin_lock(&tree->lock);
1260 entry = zswap_entry_find_get(&tree->rbroot, offset);
1262 /* entry was written back */
1263 spin_unlock(&tree->lock);
1266 spin_unlock(&tree->lock);
1268 if (!entry->length) {
1269 dst = kmap_atomic(page);
1270 zswap_fill_page(dst, entry->value);
1276 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1278 tmp = kmalloc(entry->length, GFP_ATOMIC);
1287 src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
1288 if (zpool_evictable(entry->pool->zpool))
1289 src += sizeof(struct zswap_header);
1291 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1293 memcpy(tmp, src, entry->length);
1296 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1299 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1300 mutex_lock(acomp_ctx->mutex);
1301 sg_init_one(&input, src, entry->length);
1302 sg_init_table(&output, 1);
1303 sg_set_page(&output, page, PAGE_SIZE, 0);
1304 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1305 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1306 mutex_unlock(acomp_ctx->mutex);
1308 if (zpool_can_sleep_mapped(entry->pool->zpool))
1309 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1316 spin_lock(&tree->lock);
1317 zswap_entry_put(tree, entry);
1318 spin_unlock(&tree->lock);
1323 /* frees an entry in zswap */
1324 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1326 struct zswap_tree *tree = zswap_trees[type];
1327 struct zswap_entry *entry;
1330 spin_lock(&tree->lock);
1331 entry = zswap_rb_search(&tree->rbroot, offset);
1333 /* entry was written back */
1334 spin_unlock(&tree->lock);
1338 /* remove from rbtree */
1339 zswap_rb_erase(&tree->rbroot, entry);
1341 /* drop the initial reference from entry creation */
1342 zswap_entry_put(tree, entry);
1344 spin_unlock(&tree->lock);
1347 /* frees all zswap entries for the given swap type */
1348 static void zswap_frontswap_invalidate_area(unsigned type)
1350 struct zswap_tree *tree = zswap_trees[type];
1351 struct zswap_entry *entry, *n;
1356 /* walk the tree and free everything */
1357 spin_lock(&tree->lock);
1358 rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1359 zswap_free_entry(entry);
1360 tree->rbroot = RB_ROOT;
1361 spin_unlock(&tree->lock);
1363 zswap_trees[type] = NULL;
1366 static void zswap_frontswap_init(unsigned type)
1368 struct zswap_tree *tree;
1370 tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1372 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1376 tree->rbroot = RB_ROOT;
1377 spin_lock_init(&tree->lock);
1378 zswap_trees[type] = tree;
1381 static const struct frontswap_ops zswap_frontswap_ops = {
1382 .store = zswap_frontswap_store,
1383 .load = zswap_frontswap_load,
1384 .invalidate_page = zswap_frontswap_invalidate_page,
1385 .invalidate_area = zswap_frontswap_invalidate_area,
1386 .init = zswap_frontswap_init
1389 /*********************************
1391 **********************************/
1392 #ifdef CONFIG_DEBUG_FS
1393 #include <linux/debugfs.h>
1395 static struct dentry *zswap_debugfs_root;
1397 static int __init zswap_debugfs_init(void)
1399 if (!debugfs_initialized())
1402 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1404 debugfs_create_u64("pool_limit_hit", 0444,
1405 zswap_debugfs_root, &zswap_pool_limit_hit);
1406 debugfs_create_u64("reject_reclaim_fail", 0444,
1407 zswap_debugfs_root, &zswap_reject_reclaim_fail);
1408 debugfs_create_u64("reject_alloc_fail", 0444,
1409 zswap_debugfs_root, &zswap_reject_alloc_fail);
1410 debugfs_create_u64("reject_kmemcache_fail", 0444,
1411 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1412 debugfs_create_u64("reject_compress_poor", 0444,
1413 zswap_debugfs_root, &zswap_reject_compress_poor);
1414 debugfs_create_u64("written_back_pages", 0444,
1415 zswap_debugfs_root, &zswap_written_back_pages);
1416 debugfs_create_u64("duplicate_entry", 0444,
1417 zswap_debugfs_root, &zswap_duplicate_entry);
1418 debugfs_create_u64("pool_total_size", 0444,
1419 zswap_debugfs_root, &zswap_pool_total_size);
1420 debugfs_create_atomic_t("stored_pages", 0444,
1421 zswap_debugfs_root, &zswap_stored_pages);
1422 debugfs_create_atomic_t("same_filled_pages", 0444,
1423 zswap_debugfs_root, &zswap_same_filled_pages);
1428 static int __init zswap_debugfs_init(void)
1434 /*********************************
1435 * module init and exit
1436 **********************************/
1437 static int __init init_zswap(void)
1439 struct zswap_pool *pool;
1442 zswap_init_started = true;
1444 if (zswap_entry_cache_create()) {
1445 pr_err("entry cache creation failed\n");
1449 ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1450 zswap_dstmem_prepare, zswap_dstmem_dead);
1452 pr_err("dstmem alloc failed\n");
1456 ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1457 "mm/zswap_pool:prepare",
1458 zswap_cpu_comp_prepare,
1459 zswap_cpu_comp_dead);
1463 pool = __zswap_pool_create_fallback();
1465 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1466 zpool_get_type(pool->zpool));
1467 list_add(&pool->list, &zswap_pools);
1468 zswap_has_pool = true;
1470 pr_err("pool creation failed\n");
1471 zswap_enabled = false;
1474 shrink_wq = create_workqueue("zswap-shrink");
1478 ret = frontswap_register_ops(&zswap_frontswap_ops);
1481 if (zswap_debugfs_init())
1482 pr_warn("debugfs initialization failed\n");
1486 destroy_workqueue(shrink_wq);
1489 zswap_pool_destroy(pool);
1491 cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1493 zswap_entry_cache_destroy();
1495 /* if built-in, we aren't unloaded on failure; don't allow use */
1496 zswap_init_failed = true;
1497 zswap_enabled = false;
1500 /* must be late so crypto has time to come up */
1501 late_initcall(init_zswap);
1503 MODULE_LICENSE("GPL");
1504 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1505 MODULE_DESCRIPTION("Compressed cache for swap pages");