eedc27894b1013187cc81831e956e4c8344b5176
[muen/linux.git] / mm / zswap.c
1 /*
2  * zswap.c - zswap driver file
3  *
4  * zswap is a backend for frontswap that takes pages that are in the process
5  * of being swapped out and attempts to compress and store them in a
6  * RAM-based memory pool.  This can result in a significant I/O reduction on
7  * the swap device and, in the case where decompressing from RAM is faster
8  * than reading from the swap device, can also improve workload performance.
9  *
10  * Copyright (C) 2012  Seth Jennings <sjenning@linux.vnet.ibm.com>
11  *
12  * This program is free software; you can redistribute it and/or
13  * modify it under the terms of the GNU General Public License
14  * as published by the Free Software Foundation; either version 2
15  * of the License, or (at your option) any later version.
16  *
17  * This program is distributed in the hope that it will be useful,
18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
20  * GNU General Public License for more details.
21 */
22
23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
24
25 #include <linux/module.h>
26 #include <linux/cpu.h>
27 #include <linux/highmem.h>
28 #include <linux/slab.h>
29 #include <linux/spinlock.h>
30 #include <linux/types.h>
31 #include <linux/atomic.h>
32 #include <linux/frontswap.h>
33 #include <linux/rbtree.h>
34 #include <linux/swap.h>
35 #include <linux/crypto.h>
36 #include <linux/mempool.h>
37 #include <linux/zpool.h>
38
39 #include <linux/mm_types.h>
40 #include <linux/page-flags.h>
41 #include <linux/swapops.h>
42 #include <linux/writeback.h>
43 #include <linux/pagemap.h>
44
45 /*********************************
46 * statistics
47 **********************************/
48 /* Total bytes used by the compressed storage */
49 static u64 zswap_pool_total_size;
50 /* The number of compressed pages currently stored in zswap */
51 static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
52
53 /*
54  * The statistics below are not protected from concurrent access for
55  * performance reasons so they may not be a 100% accurate.  However,
56  * they do provide useful information on roughly how many times a
57  * certain event is occurring.
58 */
59
60 /* Pool limit was hit (see zswap_max_pool_percent) */
61 static u64 zswap_pool_limit_hit;
62 /* Pages written back when pool limit was reached */
63 static u64 zswap_written_back_pages;
64 /* Store failed due to a reclaim failure after pool limit was reached */
65 static u64 zswap_reject_reclaim_fail;
66 /* Compressed page was too big for the allocator to (optimally) store */
67 static u64 zswap_reject_compress_poor;
68 /* Store failed because underlying allocator could not get memory */
69 static u64 zswap_reject_alloc_fail;
70 /* Store failed because the entry metadata could not be allocated (rare) */
71 static u64 zswap_reject_kmemcache_fail;
72 /* Duplicate store was encountered (rare) */
73 static u64 zswap_duplicate_entry;
74
75 /*********************************
76 * tunables
77 **********************************/
78
79 #define ZSWAP_PARAM_UNSET ""
80
81 /* Enable/disable zswap (disabled by default) */
82 static bool zswap_enabled;
83 static int zswap_enabled_param_set(const char *,
84                                    const struct kernel_param *);
85 static struct kernel_param_ops zswap_enabled_param_ops = {
86         .set =          zswap_enabled_param_set,
87         .get =          param_get_bool,
88 };
89 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
90
91 /* Crypto compressor to use */
92 #define ZSWAP_COMPRESSOR_DEFAULT "lzo"
93 static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
94 static int zswap_compressor_param_set(const char *,
95                                       const struct kernel_param *);
96 static struct kernel_param_ops zswap_compressor_param_ops = {
97         .set =          zswap_compressor_param_set,
98         .get =          param_get_charp,
99         .free =         param_free_charp,
100 };
101 module_param_cb(compressor, &zswap_compressor_param_ops,
102                 &zswap_compressor, 0644);
103
104 /* Compressed storage zpool to use */
105 #define ZSWAP_ZPOOL_DEFAULT "zbud"
106 static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
107 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
108 static struct kernel_param_ops zswap_zpool_param_ops = {
109         .set =          zswap_zpool_param_set,
110         .get =          param_get_charp,
111         .free =         param_free_charp,
112 };
113 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
114
115 /* The maximum percentage of memory that the compressed pool can occupy */
116 static unsigned int zswap_max_pool_percent = 20;
117 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
118
119 /*********************************
120 * data structures
121 **********************************/
122
123 struct zswap_pool {
124         struct zpool *zpool;
125         struct crypto_comp * __percpu *tfm;
126         struct kref kref;
127         struct list_head list;
128         struct work_struct work;
129         struct hlist_node node;
130         char tfm_name[CRYPTO_MAX_ALG_NAME];
131 };
132
133 /*
134  * struct zswap_entry
135  *
136  * This structure contains the metadata for tracking a single compressed
137  * page within zswap.
138  *
139  * rbnode - links the entry into red-black tree for the appropriate swap type
140  * offset - the swap offset for the entry.  Index into the red-black tree.
141  * refcount - the number of outstanding reference to the entry. This is needed
142  *            to protect against premature freeing of the entry by code
143  *            concurrent calls to load, invalidate, and writeback.  The lock
144  *            for the zswap_tree structure that contains the entry must
145  *            be held while changing the refcount.  Since the lock must
146  *            be held, there is no reason to also make refcount atomic.
147  * length - the length in bytes of the compressed page data.  Needed during
148  *          decompression
149  * pool - the zswap_pool the entry's data is in
150  * handle - zpool allocation handle that stores the compressed page data
151  */
152 struct zswap_entry {
153         struct rb_node rbnode;
154         pgoff_t offset;
155         int refcount;
156         unsigned int length;
157         struct zswap_pool *pool;
158         unsigned long handle;
159 };
160
161 struct zswap_header {
162         swp_entry_t swpentry;
163 };
164
165 /*
166  * The tree lock in the zswap_tree struct protects a few things:
167  * - the rbtree
168  * - the refcount field of each entry in the tree
169  */
170 struct zswap_tree {
171         struct rb_root rbroot;
172         spinlock_t lock;
173 };
174
175 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
176
177 /* RCU-protected iteration */
178 static LIST_HEAD(zswap_pools);
179 /* protects zswap_pools list modification */
180 static DEFINE_SPINLOCK(zswap_pools_lock);
181 /* pool counter to provide unique names to zpool */
182 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
183
184 /* used by param callback function */
185 static bool zswap_init_started;
186
187 /* fatal error during init */
188 static bool zswap_init_failed;
189
190 /* init completed, but couldn't create the initial pool */
191 static bool zswap_has_pool;
192
193 /*********************************
194 * helpers and fwd declarations
195 **********************************/
196
197 #define zswap_pool_debug(msg, p)                                \
198         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
199                  zpool_get_type((p)->zpool))
200
201 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
202 static int zswap_pool_get(struct zswap_pool *pool);
203 static void zswap_pool_put(struct zswap_pool *pool);
204
205 static const struct zpool_ops zswap_zpool_ops = {
206         .evict = zswap_writeback_entry
207 };
208
209 static bool zswap_is_full(void)
210 {
211         return totalram_pages * zswap_max_pool_percent / 100 <
212                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
213 }
214
215 static void zswap_update_total_size(void)
216 {
217         struct zswap_pool *pool;
218         u64 total = 0;
219
220         rcu_read_lock();
221
222         list_for_each_entry_rcu(pool, &zswap_pools, list)
223                 total += zpool_get_total_size(pool->zpool);
224
225         rcu_read_unlock();
226
227         zswap_pool_total_size = total;
228 }
229
230 /*********************************
231 * zswap entry functions
232 **********************************/
233 static struct kmem_cache *zswap_entry_cache;
234
235 static int __init zswap_entry_cache_create(void)
236 {
237         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
238         return zswap_entry_cache == NULL;
239 }
240
241 static void __init zswap_entry_cache_destroy(void)
242 {
243         kmem_cache_destroy(zswap_entry_cache);
244 }
245
246 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
247 {
248         struct zswap_entry *entry;
249         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
250         if (!entry)
251                 return NULL;
252         entry->refcount = 1;
253         RB_CLEAR_NODE(&entry->rbnode);
254         return entry;
255 }
256
257 static void zswap_entry_cache_free(struct zswap_entry *entry)
258 {
259         kmem_cache_free(zswap_entry_cache, entry);
260 }
261
262 /*********************************
263 * rbtree functions
264 **********************************/
265 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
266 {
267         struct rb_node *node = root->rb_node;
268         struct zswap_entry *entry;
269
270         while (node) {
271                 entry = rb_entry(node, struct zswap_entry, rbnode);
272                 if (entry->offset > offset)
273                         node = node->rb_left;
274                 else if (entry->offset < offset)
275                         node = node->rb_right;
276                 else
277                         return entry;
278         }
279         return NULL;
280 }
281
282 /*
283  * In the case that a entry with the same offset is found, a pointer to
284  * the existing entry is stored in dupentry and the function returns -EEXIST
285  */
286 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
287                         struct zswap_entry **dupentry)
288 {
289         struct rb_node **link = &root->rb_node, *parent = NULL;
290         struct zswap_entry *myentry;
291
292         while (*link) {
293                 parent = *link;
294                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
295                 if (myentry->offset > entry->offset)
296                         link = &(*link)->rb_left;
297                 else if (myentry->offset < entry->offset)
298                         link = &(*link)->rb_right;
299                 else {
300                         *dupentry = myentry;
301                         return -EEXIST;
302                 }
303         }
304         rb_link_node(&entry->rbnode, parent, link);
305         rb_insert_color(&entry->rbnode, root);
306         return 0;
307 }
308
309 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
310 {
311         if (!RB_EMPTY_NODE(&entry->rbnode)) {
312                 rb_erase(&entry->rbnode, root);
313                 RB_CLEAR_NODE(&entry->rbnode);
314         }
315 }
316
317 /*
318  * Carries out the common pattern of freeing and entry's zpool allocation,
319  * freeing the entry itself, and decrementing the number of stored pages.
320  */
321 static void zswap_free_entry(struct zswap_entry *entry)
322 {
323         zpool_free(entry->pool->zpool, entry->handle);
324         zswap_pool_put(entry->pool);
325         zswap_entry_cache_free(entry);
326         atomic_dec(&zswap_stored_pages);
327         zswap_update_total_size();
328 }
329
330 /* caller must hold the tree lock */
331 static void zswap_entry_get(struct zswap_entry *entry)
332 {
333         entry->refcount++;
334 }
335
336 /* caller must hold the tree lock
337 * remove from the tree and free it, if nobody reference the entry
338 */
339 static void zswap_entry_put(struct zswap_tree *tree,
340                         struct zswap_entry *entry)
341 {
342         int refcount = --entry->refcount;
343
344         BUG_ON(refcount < 0);
345         if (refcount == 0) {
346                 zswap_rb_erase(&tree->rbroot, entry);
347                 zswap_free_entry(entry);
348         }
349 }
350
351 /* caller must hold the tree lock */
352 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
353                                 pgoff_t offset)
354 {
355         struct zswap_entry *entry;
356
357         entry = zswap_rb_search(root, offset);
358         if (entry)
359                 zswap_entry_get(entry);
360
361         return entry;
362 }
363
364 /*********************************
365 * per-cpu code
366 **********************************/
367 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
368
369 static int zswap_dstmem_prepare(unsigned int cpu)
370 {
371         u8 *dst;
372
373         dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
374         if (!dst) {
375                 pr_err("can't allocate compressor buffer\n");
376                 return -ENOMEM;
377         }
378         per_cpu(zswap_dstmem, cpu) = dst;
379         return 0;
380 }
381
382 static int zswap_dstmem_dead(unsigned int cpu)
383 {
384         u8 *dst;
385
386         dst = per_cpu(zswap_dstmem, cpu);
387         kfree(dst);
388         per_cpu(zswap_dstmem, cpu) = NULL;
389
390         return 0;
391 }
392
393 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
394 {
395         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
396         struct crypto_comp *tfm;
397
398         if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
399                 return 0;
400
401         tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
402         if (IS_ERR_OR_NULL(tfm)) {
403                 pr_err("could not alloc crypto comp %s : %ld\n",
404                        pool->tfm_name, PTR_ERR(tfm));
405                 return -ENOMEM;
406         }
407         *per_cpu_ptr(pool->tfm, cpu) = tfm;
408         return 0;
409 }
410
411 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
412 {
413         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
414         struct crypto_comp *tfm;
415
416         tfm = *per_cpu_ptr(pool->tfm, cpu);
417         if (!IS_ERR_OR_NULL(tfm))
418                 crypto_free_comp(tfm);
419         *per_cpu_ptr(pool->tfm, cpu) = NULL;
420         return 0;
421 }
422
423 /*********************************
424 * pool functions
425 **********************************/
426
427 static struct zswap_pool *__zswap_pool_current(void)
428 {
429         struct zswap_pool *pool;
430
431         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
432         WARN_ONCE(!pool && zswap_has_pool,
433                   "%s: no page storage pool!\n", __func__);
434
435         return pool;
436 }
437
438 static struct zswap_pool *zswap_pool_current(void)
439 {
440         assert_spin_locked(&zswap_pools_lock);
441
442         return __zswap_pool_current();
443 }
444
445 static struct zswap_pool *zswap_pool_current_get(void)
446 {
447         struct zswap_pool *pool;
448
449         rcu_read_lock();
450
451         pool = __zswap_pool_current();
452         if (!zswap_pool_get(pool))
453                 pool = NULL;
454
455         rcu_read_unlock();
456
457         return pool;
458 }
459
460 static struct zswap_pool *zswap_pool_last_get(void)
461 {
462         struct zswap_pool *pool, *last = NULL;
463
464         rcu_read_lock();
465
466         list_for_each_entry_rcu(pool, &zswap_pools, list)
467                 last = pool;
468         WARN_ONCE(!last && zswap_has_pool,
469                   "%s: no page storage pool!\n", __func__);
470         if (!zswap_pool_get(last))
471                 last = NULL;
472
473         rcu_read_unlock();
474
475         return last;
476 }
477
478 /* type and compressor must be null-terminated */
479 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
480 {
481         struct zswap_pool *pool;
482
483         assert_spin_locked(&zswap_pools_lock);
484
485         list_for_each_entry_rcu(pool, &zswap_pools, list) {
486                 if (strcmp(pool->tfm_name, compressor))
487                         continue;
488                 if (strcmp(zpool_get_type(pool->zpool), type))
489                         continue;
490                 /* if we can't get it, it's about to be destroyed */
491                 if (!zswap_pool_get(pool))
492                         continue;
493                 return pool;
494         }
495
496         return NULL;
497 }
498
499 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
500 {
501         struct zswap_pool *pool;
502         char name[38]; /* 'zswap' + 32 char (max) num + \0 */
503         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
504         int ret;
505
506         if (!zswap_has_pool) {
507                 /* if either are unset, pool initialization failed, and we
508                  * need both params to be set correctly before trying to
509                  * create a pool.
510                  */
511                 if (!strcmp(type, ZSWAP_PARAM_UNSET))
512                         return NULL;
513                 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
514                         return NULL;
515         }
516
517         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
518         if (!pool) {
519                 pr_err("pool alloc failed\n");
520                 return NULL;
521         }
522
523         /* unique name for each pool specifically required by zsmalloc */
524         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
525
526         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
527         if (!pool->zpool) {
528                 pr_err("%s zpool not available\n", type);
529                 goto error;
530         }
531         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
532
533         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
534         pool->tfm = alloc_percpu(struct crypto_comp *);
535         if (!pool->tfm) {
536                 pr_err("percpu alloc failed\n");
537                 goto error;
538         }
539
540         ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
541                                        &pool->node);
542         if (ret)
543                 goto error;
544         pr_debug("using %s compressor\n", pool->tfm_name);
545
546         /* being the current pool takes 1 ref; this func expects the
547          * caller to always add the new pool as the current pool
548          */
549         kref_init(&pool->kref);
550         INIT_LIST_HEAD(&pool->list);
551
552         zswap_pool_debug("created", pool);
553
554         return pool;
555
556 error:
557         free_percpu(pool->tfm);
558         if (pool->zpool)
559                 zpool_destroy_pool(pool->zpool);
560         kfree(pool);
561         return NULL;
562 }
563
564 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
565 {
566         bool has_comp, has_zpool;
567
568         has_comp = crypto_has_comp(zswap_compressor, 0, 0);
569         if (!has_comp && strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
570                 pr_err("compressor %s not available, using default %s\n",
571                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
572                 param_free_charp(&zswap_compressor);
573                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
574                 has_comp = crypto_has_comp(zswap_compressor, 0, 0);
575         }
576         if (!has_comp) {
577                 pr_err("default compressor %s not available\n",
578                        zswap_compressor);
579                 param_free_charp(&zswap_compressor);
580                 zswap_compressor = ZSWAP_PARAM_UNSET;
581         }
582
583         has_zpool = zpool_has_pool(zswap_zpool_type);
584         if (!has_zpool && strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
585                 pr_err("zpool %s not available, using default %s\n",
586                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
587                 param_free_charp(&zswap_zpool_type);
588                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
589                 has_zpool = zpool_has_pool(zswap_zpool_type);
590         }
591         if (!has_zpool) {
592                 pr_err("default zpool %s not available\n",
593                        zswap_zpool_type);
594                 param_free_charp(&zswap_zpool_type);
595                 zswap_zpool_type = ZSWAP_PARAM_UNSET;
596         }
597
598         if (!has_comp || !has_zpool)
599                 return NULL;
600
601         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
602 }
603
604 static void zswap_pool_destroy(struct zswap_pool *pool)
605 {
606         zswap_pool_debug("destroying", pool);
607
608         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
609         free_percpu(pool->tfm);
610         zpool_destroy_pool(pool->zpool);
611         kfree(pool);
612 }
613
614 static int __must_check zswap_pool_get(struct zswap_pool *pool)
615 {
616         if (!pool)
617                 return 0;
618
619         return kref_get_unless_zero(&pool->kref);
620 }
621
622 static void __zswap_pool_release(struct work_struct *work)
623 {
624         struct zswap_pool *pool = container_of(work, typeof(*pool), work);
625
626         synchronize_rcu();
627
628         /* nobody should have been able to get a kref... */
629         WARN_ON(kref_get_unless_zero(&pool->kref));
630
631         /* pool is now off zswap_pools list and has no references. */
632         zswap_pool_destroy(pool);
633 }
634
635 static void __zswap_pool_empty(struct kref *kref)
636 {
637         struct zswap_pool *pool;
638
639         pool = container_of(kref, typeof(*pool), kref);
640
641         spin_lock(&zswap_pools_lock);
642
643         WARN_ON(pool == zswap_pool_current());
644
645         list_del_rcu(&pool->list);
646
647         INIT_WORK(&pool->work, __zswap_pool_release);
648         schedule_work(&pool->work);
649
650         spin_unlock(&zswap_pools_lock);
651 }
652
653 static void zswap_pool_put(struct zswap_pool *pool)
654 {
655         kref_put(&pool->kref, __zswap_pool_empty);
656 }
657
658 /*********************************
659 * param callbacks
660 **********************************/
661
662 /* val must be a null-terminated string */
663 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
664                              char *type, char *compressor)
665 {
666         struct zswap_pool *pool, *put_pool = NULL;
667         char *s = strstrip((char *)val);
668         int ret;
669
670         if (zswap_init_failed) {
671                 pr_err("can't set param, initialization failed\n");
672                 return -ENODEV;
673         }
674
675         /* no change required */
676         if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
677                 return 0;
678
679         /* if this is load-time (pre-init) param setting,
680          * don't create a pool; that's done during init.
681          */
682         if (!zswap_init_started)
683                 return param_set_charp(s, kp);
684
685         if (!type) {
686                 if (!zpool_has_pool(s)) {
687                         pr_err("zpool %s not available\n", s);
688                         return -ENOENT;
689                 }
690                 type = s;
691         } else if (!compressor) {
692                 if (!crypto_has_comp(s, 0, 0)) {
693                         pr_err("compressor %s not available\n", s);
694                         return -ENOENT;
695                 }
696                 compressor = s;
697         } else {
698                 WARN_ON(1);
699                 return -EINVAL;
700         }
701
702         spin_lock(&zswap_pools_lock);
703
704         pool = zswap_pool_find_get(type, compressor);
705         if (pool) {
706                 zswap_pool_debug("using existing", pool);
707                 WARN_ON(pool == zswap_pool_current());
708                 list_del_rcu(&pool->list);
709         }
710
711         spin_unlock(&zswap_pools_lock);
712
713         if (!pool)
714                 pool = zswap_pool_create(type, compressor);
715
716         if (pool)
717                 ret = param_set_charp(s, kp);
718         else
719                 ret = -EINVAL;
720
721         spin_lock(&zswap_pools_lock);
722
723         if (!ret) {
724                 put_pool = zswap_pool_current();
725                 list_add_rcu(&pool->list, &zswap_pools);
726                 zswap_has_pool = true;
727         } else if (pool) {
728                 /* add the possibly pre-existing pool to the end of the pools
729                  * list; if it's new (and empty) then it'll be removed and
730                  * destroyed by the put after we drop the lock
731                  */
732                 list_add_tail_rcu(&pool->list, &zswap_pools);
733                 put_pool = pool;
734         }
735
736         spin_unlock(&zswap_pools_lock);
737
738         if (!zswap_has_pool && !pool) {
739                 /* if initial pool creation failed, and this pool creation also
740                  * failed, maybe both compressor and zpool params were bad.
741                  * Allow changing this param, so pool creation will succeed
742                  * when the other param is changed. We already verified this
743                  * param is ok in the zpool_has_pool() or crypto_has_comp()
744                  * checks above.
745                  */
746                 ret = param_set_charp(s, kp);
747         }
748
749         /* drop the ref from either the old current pool,
750          * or the new pool we failed to add
751          */
752         if (put_pool)
753                 zswap_pool_put(put_pool);
754
755         return ret;
756 }
757
758 static int zswap_compressor_param_set(const char *val,
759                                       const struct kernel_param *kp)
760 {
761         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
762 }
763
764 static int zswap_zpool_param_set(const char *val,
765                                  const struct kernel_param *kp)
766 {
767         return __zswap_param_set(val, kp, NULL, zswap_compressor);
768 }
769
770 static int zswap_enabled_param_set(const char *val,
771                                    const struct kernel_param *kp)
772 {
773         if (zswap_init_failed) {
774                 pr_err("can't enable, initialization failed\n");
775                 return -ENODEV;
776         }
777         if (!zswap_has_pool && zswap_init_started) {
778                 pr_err("can't enable, no pool configured\n");
779                 return -ENODEV;
780         }
781
782         return param_set_bool(val, kp);
783 }
784
785 /*********************************
786 * writeback code
787 **********************************/
788 /* return enum for zswap_get_swap_cache_page */
789 enum zswap_get_swap_ret {
790         ZSWAP_SWAPCACHE_NEW,
791         ZSWAP_SWAPCACHE_EXIST,
792         ZSWAP_SWAPCACHE_FAIL,
793 };
794
795 /*
796  * zswap_get_swap_cache_page
797  *
798  * This is an adaption of read_swap_cache_async()
799  *
800  * This function tries to find a page with the given swap entry
801  * in the swapper_space address space (the swap cache).  If the page
802  * is found, it is returned in retpage.  Otherwise, a page is allocated,
803  * added to the swap cache, and returned in retpage.
804  *
805  * If success, the swap cache page is returned in retpage
806  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
807  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
808  *     the new page is added to swapcache and locked
809  * Returns ZSWAP_SWAPCACHE_FAIL on error
810  */
811 static int zswap_get_swap_cache_page(swp_entry_t entry,
812                                 struct page **retpage)
813 {
814         bool page_was_allocated;
815
816         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
817                         NULL, 0, &page_was_allocated);
818         if (page_was_allocated)
819                 return ZSWAP_SWAPCACHE_NEW;
820         if (!*retpage)
821                 return ZSWAP_SWAPCACHE_FAIL;
822         return ZSWAP_SWAPCACHE_EXIST;
823 }
824
825 /*
826  * Attempts to free an entry by adding a page to the swap cache,
827  * decompressing the entry data into the page, and issuing a
828  * bio write to write the page back to the swap device.
829  *
830  * This can be thought of as a "resumed writeback" of the page
831  * to the swap device.  We are basically resuming the same swap
832  * writeback path that was intercepted with the frontswap_store()
833  * in the first place.  After the page has been decompressed into
834  * the swap cache, the compressed version stored by zswap can be
835  * freed.
836  */
837 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
838 {
839         struct zswap_header *zhdr;
840         swp_entry_t swpentry;
841         struct zswap_tree *tree;
842         pgoff_t offset;
843         struct zswap_entry *entry;
844         struct page *page;
845         struct crypto_comp *tfm;
846         u8 *src, *dst;
847         unsigned int dlen;
848         int ret;
849         struct writeback_control wbc = {
850                 .sync_mode = WB_SYNC_NONE,
851         };
852
853         /* extract swpentry from data */
854         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
855         swpentry = zhdr->swpentry; /* here */
856         zpool_unmap_handle(pool, handle);
857         tree = zswap_trees[swp_type(swpentry)];
858         offset = swp_offset(swpentry);
859
860         /* find and ref zswap entry */
861         spin_lock(&tree->lock);
862         entry = zswap_entry_find_get(&tree->rbroot, offset);
863         if (!entry) {
864                 /* entry was invalidated */
865                 spin_unlock(&tree->lock);
866                 return 0;
867         }
868         spin_unlock(&tree->lock);
869         BUG_ON(offset != entry->offset);
870
871         /* try to allocate swap cache page */
872         switch (zswap_get_swap_cache_page(swpentry, &page)) {
873         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
874                 ret = -ENOMEM;
875                 goto fail;
876
877         case ZSWAP_SWAPCACHE_EXIST:
878                 /* page is already in the swap cache, ignore for now */
879                 put_page(page);
880                 ret = -EEXIST;
881                 goto fail;
882
883         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
884                 /* decompress */
885                 dlen = PAGE_SIZE;
886                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
887                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
888                 dst = kmap_atomic(page);
889                 tfm = *get_cpu_ptr(entry->pool->tfm);
890                 ret = crypto_comp_decompress(tfm, src, entry->length,
891                                              dst, &dlen);
892                 put_cpu_ptr(entry->pool->tfm);
893                 kunmap_atomic(dst);
894                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
895                 BUG_ON(ret);
896                 BUG_ON(dlen != PAGE_SIZE);
897
898                 /* page is up to date */
899                 SetPageUptodate(page);
900         }
901
902         /* move it to the tail of the inactive list after end_writeback */
903         SetPageReclaim(page);
904
905         /* start writeback */
906         __swap_writepage(page, &wbc, end_swap_bio_write);
907         put_page(page);
908         zswap_written_back_pages++;
909
910         spin_lock(&tree->lock);
911         /* drop local reference */
912         zswap_entry_put(tree, entry);
913
914         /*
915         * There are two possible situations for entry here:
916         * (1) refcount is 1(normal case),  entry is valid and on the tree
917         * (2) refcount is 0, entry is freed and not on the tree
918         *     because invalidate happened during writeback
919         *  search the tree and free the entry if find entry
920         */
921         if (entry == zswap_rb_search(&tree->rbroot, offset))
922                 zswap_entry_put(tree, entry);
923         spin_unlock(&tree->lock);
924
925         goto end;
926
927         /*
928         * if we get here due to ZSWAP_SWAPCACHE_EXIST
929         * a load may happening concurrently
930         * it is safe and okay to not free the entry
931         * if we free the entry in the following put
932         * it it either okay to return !0
933         */
934 fail:
935         spin_lock(&tree->lock);
936         zswap_entry_put(tree, entry);
937         spin_unlock(&tree->lock);
938
939 end:
940         return ret;
941 }
942
943 static int zswap_shrink(void)
944 {
945         struct zswap_pool *pool;
946         int ret;
947
948         pool = zswap_pool_last_get();
949         if (!pool)
950                 return -ENOENT;
951
952         ret = zpool_shrink(pool->zpool, 1, NULL);
953
954         zswap_pool_put(pool);
955
956         return ret;
957 }
958
959 /*********************************
960 * frontswap hooks
961 **********************************/
962 /* attempts to compress and store an single page */
963 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
964                                 struct page *page)
965 {
966         struct zswap_tree *tree = zswap_trees[type];
967         struct zswap_entry *entry, *dupentry;
968         struct crypto_comp *tfm;
969         int ret;
970         unsigned int dlen = PAGE_SIZE, len;
971         unsigned long handle;
972         char *buf;
973         u8 *src, *dst;
974         struct zswap_header *zhdr;
975
976         if (!zswap_enabled || !tree) {
977                 ret = -ENODEV;
978                 goto reject;
979         }
980
981         /* reclaim space if needed */
982         if (zswap_is_full()) {
983                 zswap_pool_limit_hit++;
984                 if (zswap_shrink()) {
985                         zswap_reject_reclaim_fail++;
986                         ret = -ENOMEM;
987                         goto reject;
988                 }
989         }
990
991         /* allocate entry */
992         entry = zswap_entry_cache_alloc(GFP_KERNEL);
993         if (!entry) {
994                 zswap_reject_kmemcache_fail++;
995                 ret = -ENOMEM;
996                 goto reject;
997         }
998
999         /* if entry is successfully added, it keeps the reference */
1000         entry->pool = zswap_pool_current_get();
1001         if (!entry->pool) {
1002                 ret = -EINVAL;
1003                 goto freepage;
1004         }
1005
1006         /* compress */
1007         dst = get_cpu_var(zswap_dstmem);
1008         tfm = *get_cpu_ptr(entry->pool->tfm);
1009         src = kmap_atomic(page);
1010         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
1011         kunmap_atomic(src);
1012         put_cpu_ptr(entry->pool->tfm);
1013         if (ret) {
1014                 ret = -EINVAL;
1015                 goto put_dstmem;
1016         }
1017
1018         /* store */
1019         len = dlen + sizeof(struct zswap_header);
1020         ret = zpool_malloc(entry->pool->zpool, len,
1021                            __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
1022                            &handle);
1023         if (ret == -ENOSPC) {
1024                 zswap_reject_compress_poor++;
1025                 goto put_dstmem;
1026         }
1027         if (ret) {
1028                 zswap_reject_alloc_fail++;
1029                 goto put_dstmem;
1030         }
1031         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1032         zhdr->swpentry = swp_entry(type, offset);
1033         buf = (u8 *)(zhdr + 1);
1034         memcpy(buf, dst, dlen);
1035         zpool_unmap_handle(entry->pool->zpool, handle);
1036         put_cpu_var(zswap_dstmem);
1037
1038         /* populate entry */
1039         entry->offset = offset;
1040         entry->handle = handle;
1041         entry->length = dlen;
1042
1043         /* map */
1044         spin_lock(&tree->lock);
1045         do {
1046                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1047                 if (ret == -EEXIST) {
1048                         zswap_duplicate_entry++;
1049                         /* remove from rbtree */
1050                         zswap_rb_erase(&tree->rbroot, dupentry);
1051                         zswap_entry_put(tree, dupentry);
1052                 }
1053         } while (ret == -EEXIST);
1054         spin_unlock(&tree->lock);
1055
1056         /* update stats */
1057         atomic_inc(&zswap_stored_pages);
1058         zswap_update_total_size();
1059
1060         return 0;
1061
1062 put_dstmem:
1063         put_cpu_var(zswap_dstmem);
1064         zswap_pool_put(entry->pool);
1065 freepage:
1066         zswap_entry_cache_free(entry);
1067 reject:
1068         return ret;
1069 }
1070
1071 /*
1072  * returns 0 if the page was successfully decompressed
1073  * return -1 on entry not found or error
1074 */
1075 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1076                                 struct page *page)
1077 {
1078         struct zswap_tree *tree = zswap_trees[type];
1079         struct zswap_entry *entry;
1080         struct crypto_comp *tfm;
1081         u8 *src, *dst;
1082         unsigned int dlen;
1083         int ret;
1084
1085         /* find */
1086         spin_lock(&tree->lock);
1087         entry = zswap_entry_find_get(&tree->rbroot, offset);
1088         if (!entry) {
1089                 /* entry was written back */
1090                 spin_unlock(&tree->lock);
1091                 return -1;
1092         }
1093         spin_unlock(&tree->lock);
1094
1095         /* decompress */
1096         dlen = PAGE_SIZE;
1097         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1098                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1099         dst = kmap_atomic(page);
1100         tfm = *get_cpu_ptr(entry->pool->tfm);
1101         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1102         put_cpu_ptr(entry->pool->tfm);
1103         kunmap_atomic(dst);
1104         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1105         BUG_ON(ret);
1106
1107         spin_lock(&tree->lock);
1108         zswap_entry_put(tree, entry);
1109         spin_unlock(&tree->lock);
1110
1111         return 0;
1112 }
1113
1114 /* frees an entry in zswap */
1115 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1116 {
1117         struct zswap_tree *tree = zswap_trees[type];
1118         struct zswap_entry *entry;
1119
1120         /* find */
1121         spin_lock(&tree->lock);
1122         entry = zswap_rb_search(&tree->rbroot, offset);
1123         if (!entry) {
1124                 /* entry was written back */
1125                 spin_unlock(&tree->lock);
1126                 return;
1127         }
1128
1129         /* remove from rbtree */
1130         zswap_rb_erase(&tree->rbroot, entry);
1131
1132         /* drop the initial reference from entry creation */
1133         zswap_entry_put(tree, entry);
1134
1135         spin_unlock(&tree->lock);
1136 }
1137
1138 /* frees all zswap entries for the given swap type */
1139 static void zswap_frontswap_invalidate_area(unsigned type)
1140 {
1141         struct zswap_tree *tree = zswap_trees[type];
1142         struct zswap_entry *entry, *n;
1143
1144         if (!tree)
1145                 return;
1146
1147         /* walk the tree and free everything */
1148         spin_lock(&tree->lock);
1149         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1150                 zswap_free_entry(entry);
1151         tree->rbroot = RB_ROOT;
1152         spin_unlock(&tree->lock);
1153         kfree(tree);
1154         zswap_trees[type] = NULL;
1155 }
1156
1157 static void zswap_frontswap_init(unsigned type)
1158 {
1159         struct zswap_tree *tree;
1160
1161         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1162         if (!tree) {
1163                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1164                 return;
1165         }
1166
1167         tree->rbroot = RB_ROOT;
1168         spin_lock_init(&tree->lock);
1169         zswap_trees[type] = tree;
1170 }
1171
1172 static struct frontswap_ops zswap_frontswap_ops = {
1173         .store = zswap_frontswap_store,
1174         .load = zswap_frontswap_load,
1175         .invalidate_page = zswap_frontswap_invalidate_page,
1176         .invalidate_area = zswap_frontswap_invalidate_area,
1177         .init = zswap_frontswap_init
1178 };
1179
1180 /*********************************
1181 * debugfs functions
1182 **********************************/
1183 #ifdef CONFIG_DEBUG_FS
1184 #include <linux/debugfs.h>
1185
1186 static struct dentry *zswap_debugfs_root;
1187
1188 static int __init zswap_debugfs_init(void)
1189 {
1190         if (!debugfs_initialized())
1191                 return -ENODEV;
1192
1193         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1194         if (!zswap_debugfs_root)
1195                 return -ENOMEM;
1196
1197         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1198                         zswap_debugfs_root, &zswap_pool_limit_hit);
1199         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1200                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1201         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1202                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1203         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1204                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1205         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1206                         zswap_debugfs_root, &zswap_reject_compress_poor);
1207         debugfs_create_u64("written_back_pages", S_IRUGO,
1208                         zswap_debugfs_root, &zswap_written_back_pages);
1209         debugfs_create_u64("duplicate_entry", S_IRUGO,
1210                         zswap_debugfs_root, &zswap_duplicate_entry);
1211         debugfs_create_u64("pool_total_size", S_IRUGO,
1212                         zswap_debugfs_root, &zswap_pool_total_size);
1213         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1214                         zswap_debugfs_root, &zswap_stored_pages);
1215
1216         return 0;
1217 }
1218
1219 static void __exit zswap_debugfs_exit(void)
1220 {
1221         debugfs_remove_recursive(zswap_debugfs_root);
1222 }
1223 #else
1224 static int __init zswap_debugfs_init(void)
1225 {
1226         return 0;
1227 }
1228
1229 static void __exit zswap_debugfs_exit(void) { }
1230 #endif
1231
1232 /*********************************
1233 * module init and exit
1234 **********************************/
1235 static int __init init_zswap(void)
1236 {
1237         struct zswap_pool *pool;
1238         int ret;
1239
1240         zswap_init_started = true;
1241
1242         if (zswap_entry_cache_create()) {
1243                 pr_err("entry cache creation failed\n");
1244                 goto cache_fail;
1245         }
1246
1247         ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1248                                 zswap_dstmem_prepare, zswap_dstmem_dead);
1249         if (ret) {
1250                 pr_err("dstmem alloc failed\n");
1251                 goto dstmem_fail;
1252         }
1253
1254         ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1255                                       "mm/zswap_pool:prepare",
1256                                       zswap_cpu_comp_prepare,
1257                                       zswap_cpu_comp_dead);
1258         if (ret)
1259                 goto hp_fail;
1260
1261         pool = __zswap_pool_create_fallback();
1262         if (pool) {
1263                 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1264                         zpool_get_type(pool->zpool));
1265                 list_add(&pool->list, &zswap_pools);
1266                 zswap_has_pool = true;
1267         } else {
1268                 pr_err("pool creation failed\n");
1269                 zswap_enabled = false;
1270         }
1271
1272         frontswap_register_ops(&zswap_frontswap_ops);
1273         if (zswap_debugfs_init())
1274                 pr_warn("debugfs initialization failed\n");
1275         return 0;
1276
1277 hp_fail:
1278         cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1279 dstmem_fail:
1280         zswap_entry_cache_destroy();
1281 cache_fail:
1282         /* if built-in, we aren't unloaded on failure; don't allow use */
1283         zswap_init_failed = true;
1284         zswap_enabled = false;
1285         return -ENOMEM;
1286 }
1287 /* must be late so crypto has time to come up */
1288 late_initcall(init_zswap);
1289
1290 MODULE_LICENSE("GPL");
1291 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1292 MODULE_DESCRIPTION("Compressed cache for swap pages");