mm/zswap.c: improve a size determination in zswap_frontswap_init()
[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                 return NULL;
520
521         /* unique name for each pool specifically required by zsmalloc */
522         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
523
524         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
525         if (!pool->zpool) {
526                 pr_err("%s zpool not available\n", type);
527                 goto error;
528         }
529         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
530
531         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
532         pool->tfm = alloc_percpu(struct crypto_comp *);
533         if (!pool->tfm) {
534                 pr_err("percpu alloc failed\n");
535                 goto error;
536         }
537
538         ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
539                                        &pool->node);
540         if (ret)
541                 goto error;
542         pr_debug("using %s compressor\n", pool->tfm_name);
543
544         /* being the current pool takes 1 ref; this func expects the
545          * caller to always add the new pool as the current pool
546          */
547         kref_init(&pool->kref);
548         INIT_LIST_HEAD(&pool->list);
549
550         zswap_pool_debug("created", pool);
551
552         return pool;
553
554 error:
555         free_percpu(pool->tfm);
556         if (pool->zpool)
557                 zpool_destroy_pool(pool->zpool);
558         kfree(pool);
559         return NULL;
560 }
561
562 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
563 {
564         bool has_comp, has_zpool;
565
566         has_comp = crypto_has_comp(zswap_compressor, 0, 0);
567         if (!has_comp && strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
568                 pr_err("compressor %s not available, using default %s\n",
569                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
570                 param_free_charp(&zswap_compressor);
571                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
572                 has_comp = crypto_has_comp(zswap_compressor, 0, 0);
573         }
574         if (!has_comp) {
575                 pr_err("default compressor %s not available\n",
576                        zswap_compressor);
577                 param_free_charp(&zswap_compressor);
578                 zswap_compressor = ZSWAP_PARAM_UNSET;
579         }
580
581         has_zpool = zpool_has_pool(zswap_zpool_type);
582         if (!has_zpool && strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
583                 pr_err("zpool %s not available, using default %s\n",
584                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
585                 param_free_charp(&zswap_zpool_type);
586                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
587                 has_zpool = zpool_has_pool(zswap_zpool_type);
588         }
589         if (!has_zpool) {
590                 pr_err("default zpool %s not available\n",
591                        zswap_zpool_type);
592                 param_free_charp(&zswap_zpool_type);
593                 zswap_zpool_type = ZSWAP_PARAM_UNSET;
594         }
595
596         if (!has_comp || !has_zpool)
597                 return NULL;
598
599         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
600 }
601
602 static void zswap_pool_destroy(struct zswap_pool *pool)
603 {
604         zswap_pool_debug("destroying", pool);
605
606         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
607         free_percpu(pool->tfm);
608         zpool_destroy_pool(pool->zpool);
609         kfree(pool);
610 }
611
612 static int __must_check zswap_pool_get(struct zswap_pool *pool)
613 {
614         if (!pool)
615                 return 0;
616
617         return kref_get_unless_zero(&pool->kref);
618 }
619
620 static void __zswap_pool_release(struct work_struct *work)
621 {
622         struct zswap_pool *pool = container_of(work, typeof(*pool), work);
623
624         synchronize_rcu();
625
626         /* nobody should have been able to get a kref... */
627         WARN_ON(kref_get_unless_zero(&pool->kref));
628
629         /* pool is now off zswap_pools list and has no references. */
630         zswap_pool_destroy(pool);
631 }
632
633 static void __zswap_pool_empty(struct kref *kref)
634 {
635         struct zswap_pool *pool;
636
637         pool = container_of(kref, typeof(*pool), kref);
638
639         spin_lock(&zswap_pools_lock);
640
641         WARN_ON(pool == zswap_pool_current());
642
643         list_del_rcu(&pool->list);
644
645         INIT_WORK(&pool->work, __zswap_pool_release);
646         schedule_work(&pool->work);
647
648         spin_unlock(&zswap_pools_lock);
649 }
650
651 static void zswap_pool_put(struct zswap_pool *pool)
652 {
653         kref_put(&pool->kref, __zswap_pool_empty);
654 }
655
656 /*********************************
657 * param callbacks
658 **********************************/
659
660 /* val must be a null-terminated string */
661 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
662                              char *type, char *compressor)
663 {
664         struct zswap_pool *pool, *put_pool = NULL;
665         char *s = strstrip((char *)val);
666         int ret;
667
668         if (zswap_init_failed) {
669                 pr_err("can't set param, initialization failed\n");
670                 return -ENODEV;
671         }
672
673         /* no change required */
674         if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
675                 return 0;
676
677         /* if this is load-time (pre-init) param setting,
678          * don't create a pool; that's done during init.
679          */
680         if (!zswap_init_started)
681                 return param_set_charp(s, kp);
682
683         if (!type) {
684                 if (!zpool_has_pool(s)) {
685                         pr_err("zpool %s not available\n", s);
686                         return -ENOENT;
687                 }
688                 type = s;
689         } else if (!compressor) {
690                 if (!crypto_has_comp(s, 0, 0)) {
691                         pr_err("compressor %s not available\n", s);
692                         return -ENOENT;
693                 }
694                 compressor = s;
695         } else {
696                 WARN_ON(1);
697                 return -EINVAL;
698         }
699
700         spin_lock(&zswap_pools_lock);
701
702         pool = zswap_pool_find_get(type, compressor);
703         if (pool) {
704                 zswap_pool_debug("using existing", pool);
705                 WARN_ON(pool == zswap_pool_current());
706                 list_del_rcu(&pool->list);
707         }
708
709         spin_unlock(&zswap_pools_lock);
710
711         if (!pool)
712                 pool = zswap_pool_create(type, compressor);
713
714         if (pool)
715                 ret = param_set_charp(s, kp);
716         else
717                 ret = -EINVAL;
718
719         spin_lock(&zswap_pools_lock);
720
721         if (!ret) {
722                 put_pool = zswap_pool_current();
723                 list_add_rcu(&pool->list, &zswap_pools);
724                 zswap_has_pool = true;
725         } else if (pool) {
726                 /* add the possibly pre-existing pool to the end of the pools
727                  * list; if it's new (and empty) then it'll be removed and
728                  * destroyed by the put after we drop the lock
729                  */
730                 list_add_tail_rcu(&pool->list, &zswap_pools);
731                 put_pool = pool;
732         }
733
734         spin_unlock(&zswap_pools_lock);
735
736         if (!zswap_has_pool && !pool) {
737                 /* if initial pool creation failed, and this pool creation also
738                  * failed, maybe both compressor and zpool params were bad.
739                  * Allow changing this param, so pool creation will succeed
740                  * when the other param is changed. We already verified this
741                  * param is ok in the zpool_has_pool() or crypto_has_comp()
742                  * checks above.
743                  */
744                 ret = param_set_charp(s, kp);
745         }
746
747         /* drop the ref from either the old current pool,
748          * or the new pool we failed to add
749          */
750         if (put_pool)
751                 zswap_pool_put(put_pool);
752
753         return ret;
754 }
755
756 static int zswap_compressor_param_set(const char *val,
757                                       const struct kernel_param *kp)
758 {
759         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
760 }
761
762 static int zswap_zpool_param_set(const char *val,
763                                  const struct kernel_param *kp)
764 {
765         return __zswap_param_set(val, kp, NULL, zswap_compressor);
766 }
767
768 static int zswap_enabled_param_set(const char *val,
769                                    const struct kernel_param *kp)
770 {
771         if (zswap_init_failed) {
772                 pr_err("can't enable, initialization failed\n");
773                 return -ENODEV;
774         }
775         if (!zswap_has_pool && zswap_init_started) {
776                 pr_err("can't enable, no pool configured\n");
777                 return -ENODEV;
778         }
779
780         return param_set_bool(val, kp);
781 }
782
783 /*********************************
784 * writeback code
785 **********************************/
786 /* return enum for zswap_get_swap_cache_page */
787 enum zswap_get_swap_ret {
788         ZSWAP_SWAPCACHE_NEW,
789         ZSWAP_SWAPCACHE_EXIST,
790         ZSWAP_SWAPCACHE_FAIL,
791 };
792
793 /*
794  * zswap_get_swap_cache_page
795  *
796  * This is an adaption of read_swap_cache_async()
797  *
798  * This function tries to find a page with the given swap entry
799  * in the swapper_space address space (the swap cache).  If the page
800  * is found, it is returned in retpage.  Otherwise, a page is allocated,
801  * added to the swap cache, and returned in retpage.
802  *
803  * If success, the swap cache page is returned in retpage
804  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
805  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
806  *     the new page is added to swapcache and locked
807  * Returns ZSWAP_SWAPCACHE_FAIL on error
808  */
809 static int zswap_get_swap_cache_page(swp_entry_t entry,
810                                 struct page **retpage)
811 {
812         bool page_was_allocated;
813
814         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
815                         NULL, 0, &page_was_allocated);
816         if (page_was_allocated)
817                 return ZSWAP_SWAPCACHE_NEW;
818         if (!*retpage)
819                 return ZSWAP_SWAPCACHE_FAIL;
820         return ZSWAP_SWAPCACHE_EXIST;
821 }
822
823 /*
824  * Attempts to free an entry by adding a page to the swap cache,
825  * decompressing the entry data into the page, and issuing a
826  * bio write to write the page back to the swap device.
827  *
828  * This can be thought of as a "resumed writeback" of the page
829  * to the swap device.  We are basically resuming the same swap
830  * writeback path that was intercepted with the frontswap_store()
831  * in the first place.  After the page has been decompressed into
832  * the swap cache, the compressed version stored by zswap can be
833  * freed.
834  */
835 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
836 {
837         struct zswap_header *zhdr;
838         swp_entry_t swpentry;
839         struct zswap_tree *tree;
840         pgoff_t offset;
841         struct zswap_entry *entry;
842         struct page *page;
843         struct crypto_comp *tfm;
844         u8 *src, *dst;
845         unsigned int dlen;
846         int ret;
847         struct writeback_control wbc = {
848                 .sync_mode = WB_SYNC_NONE,
849         };
850
851         /* extract swpentry from data */
852         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
853         swpentry = zhdr->swpentry; /* here */
854         zpool_unmap_handle(pool, handle);
855         tree = zswap_trees[swp_type(swpentry)];
856         offset = swp_offset(swpentry);
857
858         /* find and ref zswap entry */
859         spin_lock(&tree->lock);
860         entry = zswap_entry_find_get(&tree->rbroot, offset);
861         if (!entry) {
862                 /* entry was invalidated */
863                 spin_unlock(&tree->lock);
864                 return 0;
865         }
866         spin_unlock(&tree->lock);
867         BUG_ON(offset != entry->offset);
868
869         /* try to allocate swap cache page */
870         switch (zswap_get_swap_cache_page(swpentry, &page)) {
871         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
872                 ret = -ENOMEM;
873                 goto fail;
874
875         case ZSWAP_SWAPCACHE_EXIST:
876                 /* page is already in the swap cache, ignore for now */
877                 put_page(page);
878                 ret = -EEXIST;
879                 goto fail;
880
881         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
882                 /* decompress */
883                 dlen = PAGE_SIZE;
884                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
885                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
886                 dst = kmap_atomic(page);
887                 tfm = *get_cpu_ptr(entry->pool->tfm);
888                 ret = crypto_comp_decompress(tfm, src, entry->length,
889                                              dst, &dlen);
890                 put_cpu_ptr(entry->pool->tfm);
891                 kunmap_atomic(dst);
892                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
893                 BUG_ON(ret);
894                 BUG_ON(dlen != PAGE_SIZE);
895
896                 /* page is up to date */
897                 SetPageUptodate(page);
898         }
899
900         /* move it to the tail of the inactive list after end_writeback */
901         SetPageReclaim(page);
902
903         /* start writeback */
904         __swap_writepage(page, &wbc, end_swap_bio_write);
905         put_page(page);
906         zswap_written_back_pages++;
907
908         spin_lock(&tree->lock);
909         /* drop local reference */
910         zswap_entry_put(tree, entry);
911
912         /*
913         * There are two possible situations for entry here:
914         * (1) refcount is 1(normal case),  entry is valid and on the tree
915         * (2) refcount is 0, entry is freed and not on the tree
916         *     because invalidate happened during writeback
917         *  search the tree and free the entry if find entry
918         */
919         if (entry == zswap_rb_search(&tree->rbroot, offset))
920                 zswap_entry_put(tree, entry);
921         spin_unlock(&tree->lock);
922
923         goto end;
924
925         /*
926         * if we get here due to ZSWAP_SWAPCACHE_EXIST
927         * a load may happening concurrently
928         * it is safe and okay to not free the entry
929         * if we free the entry in the following put
930         * it it either okay to return !0
931         */
932 fail:
933         spin_lock(&tree->lock);
934         zswap_entry_put(tree, entry);
935         spin_unlock(&tree->lock);
936
937 end:
938         return ret;
939 }
940
941 static int zswap_shrink(void)
942 {
943         struct zswap_pool *pool;
944         int ret;
945
946         pool = zswap_pool_last_get();
947         if (!pool)
948                 return -ENOENT;
949
950         ret = zpool_shrink(pool->zpool, 1, NULL);
951
952         zswap_pool_put(pool);
953
954         return ret;
955 }
956
957 /*********************************
958 * frontswap hooks
959 **********************************/
960 /* attempts to compress and store an single page */
961 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
962                                 struct page *page)
963 {
964         struct zswap_tree *tree = zswap_trees[type];
965         struct zswap_entry *entry, *dupentry;
966         struct crypto_comp *tfm;
967         int ret;
968         unsigned int dlen = PAGE_SIZE, len;
969         unsigned long handle;
970         char *buf;
971         u8 *src, *dst;
972         struct zswap_header *zhdr;
973
974         if (!zswap_enabled || !tree) {
975                 ret = -ENODEV;
976                 goto reject;
977         }
978
979         /* reclaim space if needed */
980         if (zswap_is_full()) {
981                 zswap_pool_limit_hit++;
982                 if (zswap_shrink()) {
983                         zswap_reject_reclaim_fail++;
984                         ret = -ENOMEM;
985                         goto reject;
986                 }
987         }
988
989         /* allocate entry */
990         entry = zswap_entry_cache_alloc(GFP_KERNEL);
991         if (!entry) {
992                 zswap_reject_kmemcache_fail++;
993                 ret = -ENOMEM;
994                 goto reject;
995         }
996
997         /* if entry is successfully added, it keeps the reference */
998         entry->pool = zswap_pool_current_get();
999         if (!entry->pool) {
1000                 ret = -EINVAL;
1001                 goto freepage;
1002         }
1003
1004         /* compress */
1005         dst = get_cpu_var(zswap_dstmem);
1006         tfm = *get_cpu_ptr(entry->pool->tfm);
1007         src = kmap_atomic(page);
1008         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
1009         kunmap_atomic(src);
1010         put_cpu_ptr(entry->pool->tfm);
1011         if (ret) {
1012                 ret = -EINVAL;
1013                 goto put_dstmem;
1014         }
1015
1016         /* store */
1017         len = dlen + sizeof(struct zswap_header);
1018         ret = zpool_malloc(entry->pool->zpool, len,
1019                            __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
1020                            &handle);
1021         if (ret == -ENOSPC) {
1022                 zswap_reject_compress_poor++;
1023                 goto put_dstmem;
1024         }
1025         if (ret) {
1026                 zswap_reject_alloc_fail++;
1027                 goto put_dstmem;
1028         }
1029         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1030         zhdr->swpentry = swp_entry(type, offset);
1031         buf = (u8 *)(zhdr + 1);
1032         memcpy(buf, dst, dlen);
1033         zpool_unmap_handle(entry->pool->zpool, handle);
1034         put_cpu_var(zswap_dstmem);
1035
1036         /* populate entry */
1037         entry->offset = offset;
1038         entry->handle = handle;
1039         entry->length = dlen;
1040
1041         /* map */
1042         spin_lock(&tree->lock);
1043         do {
1044                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1045                 if (ret == -EEXIST) {
1046                         zswap_duplicate_entry++;
1047                         /* remove from rbtree */
1048                         zswap_rb_erase(&tree->rbroot, dupentry);
1049                         zswap_entry_put(tree, dupentry);
1050                 }
1051         } while (ret == -EEXIST);
1052         spin_unlock(&tree->lock);
1053
1054         /* update stats */
1055         atomic_inc(&zswap_stored_pages);
1056         zswap_update_total_size();
1057
1058         return 0;
1059
1060 put_dstmem:
1061         put_cpu_var(zswap_dstmem);
1062         zswap_pool_put(entry->pool);
1063 freepage:
1064         zswap_entry_cache_free(entry);
1065 reject:
1066         return ret;
1067 }
1068
1069 /*
1070  * returns 0 if the page was successfully decompressed
1071  * return -1 on entry not found or error
1072 */
1073 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1074                                 struct page *page)
1075 {
1076         struct zswap_tree *tree = zswap_trees[type];
1077         struct zswap_entry *entry;
1078         struct crypto_comp *tfm;
1079         u8 *src, *dst;
1080         unsigned int dlen;
1081         int ret;
1082
1083         /* find */
1084         spin_lock(&tree->lock);
1085         entry = zswap_entry_find_get(&tree->rbroot, offset);
1086         if (!entry) {
1087                 /* entry was written back */
1088                 spin_unlock(&tree->lock);
1089                 return -1;
1090         }
1091         spin_unlock(&tree->lock);
1092
1093         /* decompress */
1094         dlen = PAGE_SIZE;
1095         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1096                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1097         dst = kmap_atomic(page);
1098         tfm = *get_cpu_ptr(entry->pool->tfm);
1099         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1100         put_cpu_ptr(entry->pool->tfm);
1101         kunmap_atomic(dst);
1102         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1103         BUG_ON(ret);
1104
1105         spin_lock(&tree->lock);
1106         zswap_entry_put(tree, entry);
1107         spin_unlock(&tree->lock);
1108
1109         return 0;
1110 }
1111
1112 /* frees an entry in zswap */
1113 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1114 {
1115         struct zswap_tree *tree = zswap_trees[type];
1116         struct zswap_entry *entry;
1117
1118         /* find */
1119         spin_lock(&tree->lock);
1120         entry = zswap_rb_search(&tree->rbroot, offset);
1121         if (!entry) {
1122                 /* entry was written back */
1123                 spin_unlock(&tree->lock);
1124                 return;
1125         }
1126
1127         /* remove from rbtree */
1128         zswap_rb_erase(&tree->rbroot, entry);
1129
1130         /* drop the initial reference from entry creation */
1131         zswap_entry_put(tree, entry);
1132
1133         spin_unlock(&tree->lock);
1134 }
1135
1136 /* frees all zswap entries for the given swap type */
1137 static void zswap_frontswap_invalidate_area(unsigned type)
1138 {
1139         struct zswap_tree *tree = zswap_trees[type];
1140         struct zswap_entry *entry, *n;
1141
1142         if (!tree)
1143                 return;
1144
1145         /* walk the tree and free everything */
1146         spin_lock(&tree->lock);
1147         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1148                 zswap_free_entry(entry);
1149         tree->rbroot = RB_ROOT;
1150         spin_unlock(&tree->lock);
1151         kfree(tree);
1152         zswap_trees[type] = NULL;
1153 }
1154
1155 static void zswap_frontswap_init(unsigned type)
1156 {
1157         struct zswap_tree *tree;
1158
1159         tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1160         if (!tree) {
1161                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1162                 return;
1163         }
1164
1165         tree->rbroot = RB_ROOT;
1166         spin_lock_init(&tree->lock);
1167         zswap_trees[type] = tree;
1168 }
1169
1170 static struct frontswap_ops zswap_frontswap_ops = {
1171         .store = zswap_frontswap_store,
1172         .load = zswap_frontswap_load,
1173         .invalidate_page = zswap_frontswap_invalidate_page,
1174         .invalidate_area = zswap_frontswap_invalidate_area,
1175         .init = zswap_frontswap_init
1176 };
1177
1178 /*********************************
1179 * debugfs functions
1180 **********************************/
1181 #ifdef CONFIG_DEBUG_FS
1182 #include <linux/debugfs.h>
1183
1184 static struct dentry *zswap_debugfs_root;
1185
1186 static int __init zswap_debugfs_init(void)
1187 {
1188         if (!debugfs_initialized())
1189                 return -ENODEV;
1190
1191         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1192         if (!zswap_debugfs_root)
1193                 return -ENOMEM;
1194
1195         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1196                         zswap_debugfs_root, &zswap_pool_limit_hit);
1197         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1198                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1199         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1200                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1201         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1202                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1203         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1204                         zswap_debugfs_root, &zswap_reject_compress_poor);
1205         debugfs_create_u64("written_back_pages", S_IRUGO,
1206                         zswap_debugfs_root, &zswap_written_back_pages);
1207         debugfs_create_u64("duplicate_entry", S_IRUGO,
1208                         zswap_debugfs_root, &zswap_duplicate_entry);
1209         debugfs_create_u64("pool_total_size", S_IRUGO,
1210                         zswap_debugfs_root, &zswap_pool_total_size);
1211         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1212                         zswap_debugfs_root, &zswap_stored_pages);
1213
1214         return 0;
1215 }
1216
1217 static void __exit zswap_debugfs_exit(void)
1218 {
1219         debugfs_remove_recursive(zswap_debugfs_root);
1220 }
1221 #else
1222 static int __init zswap_debugfs_init(void)
1223 {
1224         return 0;
1225 }
1226
1227 static void __exit zswap_debugfs_exit(void) { }
1228 #endif
1229
1230 /*********************************
1231 * module init and exit
1232 **********************************/
1233 static int __init init_zswap(void)
1234 {
1235         struct zswap_pool *pool;
1236         int ret;
1237
1238         zswap_init_started = true;
1239
1240         if (zswap_entry_cache_create()) {
1241                 pr_err("entry cache creation failed\n");
1242                 goto cache_fail;
1243         }
1244
1245         ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1246                                 zswap_dstmem_prepare, zswap_dstmem_dead);
1247         if (ret) {
1248                 pr_err("dstmem alloc failed\n");
1249                 goto dstmem_fail;
1250         }
1251
1252         ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1253                                       "mm/zswap_pool:prepare",
1254                                       zswap_cpu_comp_prepare,
1255                                       zswap_cpu_comp_dead);
1256         if (ret)
1257                 goto hp_fail;
1258
1259         pool = __zswap_pool_create_fallback();
1260         if (pool) {
1261                 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1262                         zpool_get_type(pool->zpool));
1263                 list_add(&pool->list, &zswap_pools);
1264                 zswap_has_pool = true;
1265         } else {
1266                 pr_err("pool creation failed\n");
1267                 zswap_enabled = false;
1268         }
1269
1270         frontswap_register_ops(&zswap_frontswap_ops);
1271         if (zswap_debugfs_init())
1272                 pr_warn("debugfs initialization failed\n");
1273         return 0;
1274
1275 hp_fail:
1276         cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1277 dstmem_fail:
1278         zswap_entry_cache_destroy();
1279 cache_fail:
1280         /* if built-in, we aren't unloaded on failure; don't allow use */
1281         zswap_init_failed = true;
1282         zswap_enabled = false;
1283         return -ENOMEM;
1284 }
1285 /* must be late so crypto has time to come up */
1286 late_initcall(init_zswap);
1287
1288 MODULE_LICENSE("GPL");
1289 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1290 MODULE_DESCRIPTION("Compressed cache for swap pages");