Loading...
1/*
2 * zswap.c - zswap driver file
3 *
4 * zswap is a backend for frontswap that takes pages that are in the process
5 * of being swapped out and attempts to compress and store them in a
6 * RAM-based memory pool. This can result in a significant I/O reduction on
7 * the swap device and, in the case where decompressing from RAM is faster
8 * than reading from the swap device, can also improve workload performance.
9 *
10 * Copyright (C) 2012 Seth Jennings <sjenning@linux.vnet.ibm.com>
11 *
12 * This program is free software; you can redistribute it and/or
13 * modify it under the terms of the GNU General Public License
14 * as published by the Free Software Foundation; either version 2
15 * of the License, or (at your option) any later version.
16 *
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
21*/
22
23#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
24
25#include <linux/module.h>
26#include <linux/cpu.h>
27#include <linux/highmem.h>
28#include <linux/slab.h>
29#include <linux/spinlock.h>
30#include <linux/types.h>
31#include <linux/atomic.h>
32#include <linux/frontswap.h>
33#include <linux/rbtree.h>
34#include <linux/swap.h>
35#include <linux/crypto.h>
36#include <linux/mempool.h>
37#include <linux/zbud.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/* Number of memory pages used by the compressed pool */
49static u64 zswap_pool_pages;
50/* The number of compressed pages currently stored in zswap */
51static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
52
53/*
54 * The statistics below are not protected from concurrent access for
55 * performance reasons so they may not be a 100% accurate. However,
56 * they do provide useful information on roughly how many times a
57 * certain event is occurring.
58*/
59
60/* Pool limit was hit (see zswap_max_pool_percent) */
61static u64 zswap_pool_limit_hit;
62/* Pages written back when pool limit was reached */
63static u64 zswap_written_back_pages;
64/* Store failed due to a reclaim failure after pool limit was reached */
65static u64 zswap_reject_reclaim_fail;
66/* Compressed page was too big for the allocator to (optimally) store */
67static u64 zswap_reject_compress_poor;
68/* Store failed because underlying allocator could not get memory */
69static u64 zswap_reject_alloc_fail;
70/* Store failed because the entry metadata could not be allocated (rare) */
71static u64 zswap_reject_kmemcache_fail;
72/* Duplicate store was encountered (rare) */
73static u64 zswap_duplicate_entry;
74
75/*********************************
76* tunables
77**********************************/
78/* Enable/disable zswap (disabled by default, fixed at boot for now) */
79static bool zswap_enabled __read_mostly;
80module_param_named(enabled, zswap_enabled, bool, 0444);
81
82/* Compressor to be used by zswap (fixed at boot for now) */
83#define ZSWAP_COMPRESSOR_DEFAULT "lzo"
84static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
85module_param_named(compressor, zswap_compressor, charp, 0444);
86
87/* The maximum percentage of memory that the compressed pool can occupy */
88static unsigned int zswap_max_pool_percent = 20;
89module_param_named(max_pool_percent,
90 zswap_max_pool_percent, uint, 0644);
91
92/* zbud_pool is shared by all of zswap backend */
93static struct zbud_pool *zswap_pool;
94
95/*********************************
96* compression functions
97**********************************/
98/* per-cpu compression transforms */
99static struct crypto_comp * __percpu *zswap_comp_pcpu_tfms;
100
101enum comp_op {
102 ZSWAP_COMPOP_COMPRESS,
103 ZSWAP_COMPOP_DECOMPRESS
104};
105
106static int zswap_comp_op(enum comp_op op, const u8 *src, unsigned int slen,
107 u8 *dst, unsigned int *dlen)
108{
109 struct crypto_comp *tfm;
110 int ret;
111
112 tfm = *per_cpu_ptr(zswap_comp_pcpu_tfms, get_cpu());
113 switch (op) {
114 case ZSWAP_COMPOP_COMPRESS:
115 ret = crypto_comp_compress(tfm, src, slen, dst, dlen);
116 break;
117 case ZSWAP_COMPOP_DECOMPRESS:
118 ret = crypto_comp_decompress(tfm, src, slen, dst, dlen);
119 break;
120 default:
121 ret = -EINVAL;
122 }
123
124 put_cpu();
125 return ret;
126}
127
128static int __init zswap_comp_init(void)
129{
130 if (!crypto_has_comp(zswap_compressor, 0, 0)) {
131 pr_info("%s compressor not available\n", zswap_compressor);
132 /* fall back to default compressor */
133 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
134 if (!crypto_has_comp(zswap_compressor, 0, 0))
135 /* can't even load the default compressor */
136 return -ENODEV;
137 }
138 pr_info("using %s compressor\n", zswap_compressor);
139
140 /* alloc percpu transforms */
141 zswap_comp_pcpu_tfms = alloc_percpu(struct crypto_comp *);
142 if (!zswap_comp_pcpu_tfms)
143 return -ENOMEM;
144 return 0;
145}
146
147static void zswap_comp_exit(void)
148{
149 /* free percpu transforms */
150 if (zswap_comp_pcpu_tfms)
151 free_percpu(zswap_comp_pcpu_tfms);
152}
153
154/*********************************
155* data structures
156**********************************/
157/*
158 * struct zswap_entry
159 *
160 * This structure contains the metadata for tracking a single compressed
161 * page within zswap.
162 *
163 * rbnode - links the entry into red-black tree for the appropriate swap type
164 * refcount - the number of outstanding reference to the entry. This is needed
165 * to protect against premature freeing of the entry by code
166 * concurrent calls to load, invalidate, and writeback. The lock
167 * for the zswap_tree structure that contains the entry must
168 * be held while changing the refcount. Since the lock must
169 * be held, there is no reason to also make refcount atomic.
170 * offset - the swap offset for the entry. Index into the red-black tree.
171 * handle - zbud allocation handle that stores the compressed page data
172 * length - the length in bytes of the compressed page data. Needed during
173 * decompression
174 */
175struct zswap_entry {
176 struct rb_node rbnode;
177 pgoff_t offset;
178 int refcount;
179 unsigned int length;
180 unsigned long handle;
181};
182
183struct zswap_header {
184 swp_entry_t swpentry;
185};
186
187/*
188 * The tree lock in the zswap_tree struct protects a few things:
189 * - the rbtree
190 * - the refcount field of each entry in the tree
191 */
192struct zswap_tree {
193 struct rb_root rbroot;
194 spinlock_t lock;
195};
196
197static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
198
199/*********************************
200* zswap entry functions
201**********************************/
202static struct kmem_cache *zswap_entry_cache;
203
204static int zswap_entry_cache_create(void)
205{
206 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
207 return zswap_entry_cache == NULL;
208}
209
210static void zswap_entry_cache_destory(void)
211{
212 kmem_cache_destroy(zswap_entry_cache);
213}
214
215static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
216{
217 struct zswap_entry *entry;
218 entry = kmem_cache_alloc(zswap_entry_cache, gfp);
219 if (!entry)
220 return NULL;
221 entry->refcount = 1;
222 RB_CLEAR_NODE(&entry->rbnode);
223 return entry;
224}
225
226static void zswap_entry_cache_free(struct zswap_entry *entry)
227{
228 kmem_cache_free(zswap_entry_cache, entry);
229}
230
231/*********************************
232* rbtree functions
233**********************************/
234static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
235{
236 struct rb_node *node = root->rb_node;
237 struct zswap_entry *entry;
238
239 while (node) {
240 entry = rb_entry(node, struct zswap_entry, rbnode);
241 if (entry->offset > offset)
242 node = node->rb_left;
243 else if (entry->offset < offset)
244 node = node->rb_right;
245 else
246 return entry;
247 }
248 return NULL;
249}
250
251/*
252 * In the case that a entry with the same offset is found, a pointer to
253 * the existing entry is stored in dupentry and the function returns -EEXIST
254 */
255static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
256 struct zswap_entry **dupentry)
257{
258 struct rb_node **link = &root->rb_node, *parent = NULL;
259 struct zswap_entry *myentry;
260
261 while (*link) {
262 parent = *link;
263 myentry = rb_entry(parent, struct zswap_entry, rbnode);
264 if (myentry->offset > entry->offset)
265 link = &(*link)->rb_left;
266 else if (myentry->offset < entry->offset)
267 link = &(*link)->rb_right;
268 else {
269 *dupentry = myentry;
270 return -EEXIST;
271 }
272 }
273 rb_link_node(&entry->rbnode, parent, link);
274 rb_insert_color(&entry->rbnode, root);
275 return 0;
276}
277
278static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
279{
280 if (!RB_EMPTY_NODE(&entry->rbnode)) {
281 rb_erase(&entry->rbnode, root);
282 RB_CLEAR_NODE(&entry->rbnode);
283 }
284}
285
286/*
287 * Carries out the common pattern of freeing and entry's zbud allocation,
288 * freeing the entry itself, and decrementing the number of stored pages.
289 */
290static void zswap_free_entry(struct zswap_entry *entry)
291{
292 zbud_free(zswap_pool, entry->handle);
293 zswap_entry_cache_free(entry);
294 atomic_dec(&zswap_stored_pages);
295 zswap_pool_pages = zbud_get_pool_size(zswap_pool);
296}
297
298/* caller must hold the tree lock */
299static void zswap_entry_get(struct zswap_entry *entry)
300{
301 entry->refcount++;
302}
303
304/* caller must hold the tree lock
305* remove from the tree and free it, if nobody reference the entry
306*/
307static void zswap_entry_put(struct zswap_tree *tree,
308 struct zswap_entry *entry)
309{
310 int refcount = --entry->refcount;
311
312 BUG_ON(refcount < 0);
313 if (refcount == 0) {
314 zswap_rb_erase(&tree->rbroot, entry);
315 zswap_free_entry(entry);
316 }
317}
318
319/* caller must hold the tree lock */
320static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
321 pgoff_t offset)
322{
323 struct zswap_entry *entry = NULL;
324
325 entry = zswap_rb_search(root, offset);
326 if (entry)
327 zswap_entry_get(entry);
328
329 return entry;
330}
331
332/*********************************
333* per-cpu code
334**********************************/
335static DEFINE_PER_CPU(u8 *, zswap_dstmem);
336
337static int __zswap_cpu_notifier(unsigned long action, unsigned long cpu)
338{
339 struct crypto_comp *tfm;
340 u8 *dst;
341
342 switch (action) {
343 case CPU_UP_PREPARE:
344 tfm = crypto_alloc_comp(zswap_compressor, 0, 0);
345 if (IS_ERR(tfm)) {
346 pr_err("can't allocate compressor transform\n");
347 return NOTIFY_BAD;
348 }
349 *per_cpu_ptr(zswap_comp_pcpu_tfms, cpu) = tfm;
350 dst = kmalloc(PAGE_SIZE * 2, GFP_KERNEL);
351 if (!dst) {
352 pr_err("can't allocate compressor buffer\n");
353 crypto_free_comp(tfm);
354 *per_cpu_ptr(zswap_comp_pcpu_tfms, cpu) = NULL;
355 return NOTIFY_BAD;
356 }
357 per_cpu(zswap_dstmem, cpu) = dst;
358 break;
359 case CPU_DEAD:
360 case CPU_UP_CANCELED:
361 tfm = *per_cpu_ptr(zswap_comp_pcpu_tfms, cpu);
362 if (tfm) {
363 crypto_free_comp(tfm);
364 *per_cpu_ptr(zswap_comp_pcpu_tfms, cpu) = NULL;
365 }
366 dst = per_cpu(zswap_dstmem, cpu);
367 kfree(dst);
368 per_cpu(zswap_dstmem, cpu) = NULL;
369 break;
370 default:
371 break;
372 }
373 return NOTIFY_OK;
374}
375
376static int zswap_cpu_notifier(struct notifier_block *nb,
377 unsigned long action, void *pcpu)
378{
379 unsigned long cpu = (unsigned long)pcpu;
380 return __zswap_cpu_notifier(action, cpu);
381}
382
383static struct notifier_block zswap_cpu_notifier_block = {
384 .notifier_call = zswap_cpu_notifier
385};
386
387static int zswap_cpu_init(void)
388{
389 unsigned long cpu;
390
391 cpu_notifier_register_begin();
392 for_each_online_cpu(cpu)
393 if (__zswap_cpu_notifier(CPU_UP_PREPARE, cpu) != NOTIFY_OK)
394 goto cleanup;
395 __register_cpu_notifier(&zswap_cpu_notifier_block);
396 cpu_notifier_register_done();
397 return 0;
398
399cleanup:
400 for_each_online_cpu(cpu)
401 __zswap_cpu_notifier(CPU_UP_CANCELED, cpu);
402 cpu_notifier_register_done();
403 return -ENOMEM;
404}
405
406/*********************************
407* helpers
408**********************************/
409static bool zswap_is_full(void)
410{
411 return totalram_pages * zswap_max_pool_percent / 100 <
412 zswap_pool_pages;
413}
414
415/*********************************
416* writeback code
417**********************************/
418/* return enum for zswap_get_swap_cache_page */
419enum zswap_get_swap_ret {
420 ZSWAP_SWAPCACHE_NEW,
421 ZSWAP_SWAPCACHE_EXIST,
422 ZSWAP_SWAPCACHE_FAIL,
423};
424
425/*
426 * zswap_get_swap_cache_page
427 *
428 * This is an adaption of read_swap_cache_async()
429 *
430 * This function tries to find a page with the given swap entry
431 * in the swapper_space address space (the swap cache). If the page
432 * is found, it is returned in retpage. Otherwise, a page is allocated,
433 * added to the swap cache, and returned in retpage.
434 *
435 * If success, the swap cache page is returned in retpage
436 * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
437 * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
438 * the new page is added to swapcache and locked
439 * Returns ZSWAP_SWAPCACHE_FAIL on error
440 */
441static int zswap_get_swap_cache_page(swp_entry_t entry,
442 struct page **retpage)
443{
444 struct page *found_page, *new_page = NULL;
445 struct address_space *swapper_space = swap_address_space(entry);
446 int err;
447
448 *retpage = NULL;
449 do {
450 /*
451 * First check the swap cache. Since this is normally
452 * called after lookup_swap_cache() failed, re-calling
453 * that would confuse statistics.
454 */
455 found_page = find_get_page(swapper_space, entry.val);
456 if (found_page)
457 break;
458
459 /*
460 * Get a new page to read into from swap.
461 */
462 if (!new_page) {
463 new_page = alloc_page(GFP_KERNEL);
464 if (!new_page)
465 break; /* Out of memory */
466 }
467
468 /*
469 * call radix_tree_preload() while we can wait.
470 */
471 err = radix_tree_preload(GFP_KERNEL);
472 if (err)
473 break;
474
475 /*
476 * Swap entry may have been freed since our caller observed it.
477 */
478 err = swapcache_prepare(entry);
479 if (err == -EEXIST) { /* seems racy */
480 radix_tree_preload_end();
481 continue;
482 }
483 if (err) { /* swp entry is obsolete ? */
484 radix_tree_preload_end();
485 break;
486 }
487
488 /* May fail (-ENOMEM) if radix-tree node allocation failed. */
489 __set_page_locked(new_page);
490 SetPageSwapBacked(new_page);
491 err = __add_to_swap_cache(new_page, entry);
492 if (likely(!err)) {
493 radix_tree_preload_end();
494 lru_cache_add_anon(new_page);
495 *retpage = new_page;
496 return ZSWAP_SWAPCACHE_NEW;
497 }
498 radix_tree_preload_end();
499 ClearPageSwapBacked(new_page);
500 __clear_page_locked(new_page);
501 /*
502 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
503 * clear SWAP_HAS_CACHE flag.
504 */
505 swapcache_free(entry, NULL);
506 } while (err != -ENOMEM);
507
508 if (new_page)
509 page_cache_release(new_page);
510 if (!found_page)
511 return ZSWAP_SWAPCACHE_FAIL;
512 *retpage = found_page;
513 return ZSWAP_SWAPCACHE_EXIST;
514}
515
516/*
517 * Attempts to free an entry by adding a page to the swap cache,
518 * decompressing the entry data into the page, and issuing a
519 * bio write to write the page back to the swap device.
520 *
521 * This can be thought of as a "resumed writeback" of the page
522 * to the swap device. We are basically resuming the same swap
523 * writeback path that was intercepted with the frontswap_store()
524 * in the first place. After the page has been decompressed into
525 * the swap cache, the compressed version stored by zswap can be
526 * freed.
527 */
528static int zswap_writeback_entry(struct zbud_pool *pool, unsigned long handle)
529{
530 struct zswap_header *zhdr;
531 swp_entry_t swpentry;
532 struct zswap_tree *tree;
533 pgoff_t offset;
534 struct zswap_entry *entry;
535 struct page *page;
536 u8 *src, *dst;
537 unsigned int dlen;
538 int ret;
539 struct writeback_control wbc = {
540 .sync_mode = WB_SYNC_NONE,
541 };
542
543 /* extract swpentry from data */
544 zhdr = zbud_map(pool, handle);
545 swpentry = zhdr->swpentry; /* here */
546 zbud_unmap(pool, handle);
547 tree = zswap_trees[swp_type(swpentry)];
548 offset = swp_offset(swpentry);
549
550 /* find and ref zswap entry */
551 spin_lock(&tree->lock);
552 entry = zswap_entry_find_get(&tree->rbroot, offset);
553 if (!entry) {
554 /* entry was invalidated */
555 spin_unlock(&tree->lock);
556 return 0;
557 }
558 spin_unlock(&tree->lock);
559 BUG_ON(offset != entry->offset);
560
561 /* try to allocate swap cache page */
562 switch (zswap_get_swap_cache_page(swpentry, &page)) {
563 case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
564 ret = -ENOMEM;
565 goto fail;
566
567 case ZSWAP_SWAPCACHE_EXIST:
568 /* page is already in the swap cache, ignore for now */
569 page_cache_release(page);
570 ret = -EEXIST;
571 goto fail;
572
573 case ZSWAP_SWAPCACHE_NEW: /* page is locked */
574 /* decompress */
575 dlen = PAGE_SIZE;
576 src = (u8 *)zbud_map(zswap_pool, entry->handle) +
577 sizeof(struct zswap_header);
578 dst = kmap_atomic(page);
579 ret = zswap_comp_op(ZSWAP_COMPOP_DECOMPRESS, src,
580 entry->length, dst, &dlen);
581 kunmap_atomic(dst);
582 zbud_unmap(zswap_pool, entry->handle);
583 BUG_ON(ret);
584 BUG_ON(dlen != PAGE_SIZE);
585
586 /* page is up to date */
587 SetPageUptodate(page);
588 }
589
590 /* move it to the tail of the inactive list after end_writeback */
591 SetPageReclaim(page);
592
593 /* start writeback */
594 __swap_writepage(page, &wbc, end_swap_bio_write);
595 page_cache_release(page);
596 zswap_written_back_pages++;
597
598 spin_lock(&tree->lock);
599 /* drop local reference */
600 zswap_entry_put(tree, entry);
601
602 /*
603 * There are two possible situations for entry here:
604 * (1) refcount is 1(normal case), entry is valid and on the tree
605 * (2) refcount is 0, entry is freed and not on the tree
606 * because invalidate happened during writeback
607 * search the tree and free the entry if find entry
608 */
609 if (entry == zswap_rb_search(&tree->rbroot, offset))
610 zswap_entry_put(tree, entry);
611 spin_unlock(&tree->lock);
612
613 goto end;
614
615 /*
616 * if we get here due to ZSWAP_SWAPCACHE_EXIST
617 * a load may happening concurrently
618 * it is safe and okay to not free the entry
619 * if we free the entry in the following put
620 * it it either okay to return !0
621 */
622fail:
623 spin_lock(&tree->lock);
624 zswap_entry_put(tree, entry);
625 spin_unlock(&tree->lock);
626
627end:
628 return ret;
629}
630
631/*********************************
632* frontswap hooks
633**********************************/
634/* attempts to compress and store an single page */
635static int zswap_frontswap_store(unsigned type, pgoff_t offset,
636 struct page *page)
637{
638 struct zswap_tree *tree = zswap_trees[type];
639 struct zswap_entry *entry, *dupentry;
640 int ret;
641 unsigned int dlen = PAGE_SIZE, len;
642 unsigned long handle;
643 char *buf;
644 u8 *src, *dst;
645 struct zswap_header *zhdr;
646
647 if (!tree) {
648 ret = -ENODEV;
649 goto reject;
650 }
651
652 /* reclaim space if needed */
653 if (zswap_is_full()) {
654 zswap_pool_limit_hit++;
655 if (zbud_reclaim_page(zswap_pool, 8)) {
656 zswap_reject_reclaim_fail++;
657 ret = -ENOMEM;
658 goto reject;
659 }
660 }
661
662 /* allocate entry */
663 entry = zswap_entry_cache_alloc(GFP_KERNEL);
664 if (!entry) {
665 zswap_reject_kmemcache_fail++;
666 ret = -ENOMEM;
667 goto reject;
668 }
669
670 /* compress */
671 dst = get_cpu_var(zswap_dstmem);
672 src = kmap_atomic(page);
673 ret = zswap_comp_op(ZSWAP_COMPOP_COMPRESS, src, PAGE_SIZE, dst, &dlen);
674 kunmap_atomic(src);
675 if (ret) {
676 ret = -EINVAL;
677 goto freepage;
678 }
679
680 /* store */
681 len = dlen + sizeof(struct zswap_header);
682 ret = zbud_alloc(zswap_pool, len, __GFP_NORETRY | __GFP_NOWARN,
683 &handle);
684 if (ret == -ENOSPC) {
685 zswap_reject_compress_poor++;
686 goto freepage;
687 }
688 if (ret) {
689 zswap_reject_alloc_fail++;
690 goto freepage;
691 }
692 zhdr = zbud_map(zswap_pool, handle);
693 zhdr->swpentry = swp_entry(type, offset);
694 buf = (u8 *)(zhdr + 1);
695 memcpy(buf, dst, dlen);
696 zbud_unmap(zswap_pool, handle);
697 put_cpu_var(zswap_dstmem);
698
699 /* populate entry */
700 entry->offset = offset;
701 entry->handle = handle;
702 entry->length = dlen;
703
704 /* map */
705 spin_lock(&tree->lock);
706 do {
707 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
708 if (ret == -EEXIST) {
709 zswap_duplicate_entry++;
710 /* remove from rbtree */
711 zswap_rb_erase(&tree->rbroot, dupentry);
712 zswap_entry_put(tree, dupentry);
713 }
714 } while (ret == -EEXIST);
715 spin_unlock(&tree->lock);
716
717 /* update stats */
718 atomic_inc(&zswap_stored_pages);
719 zswap_pool_pages = zbud_get_pool_size(zswap_pool);
720
721 return 0;
722
723freepage:
724 put_cpu_var(zswap_dstmem);
725 zswap_entry_cache_free(entry);
726reject:
727 return ret;
728}
729
730/*
731 * returns 0 if the page was successfully decompressed
732 * return -1 on entry not found or error
733*/
734static int zswap_frontswap_load(unsigned type, pgoff_t offset,
735 struct page *page)
736{
737 struct zswap_tree *tree = zswap_trees[type];
738 struct zswap_entry *entry;
739 u8 *src, *dst;
740 unsigned int dlen;
741 int ret;
742
743 /* find */
744 spin_lock(&tree->lock);
745 entry = zswap_entry_find_get(&tree->rbroot, offset);
746 if (!entry) {
747 /* entry was written back */
748 spin_unlock(&tree->lock);
749 return -1;
750 }
751 spin_unlock(&tree->lock);
752
753 /* decompress */
754 dlen = PAGE_SIZE;
755 src = (u8 *)zbud_map(zswap_pool, entry->handle) +
756 sizeof(struct zswap_header);
757 dst = kmap_atomic(page);
758 ret = zswap_comp_op(ZSWAP_COMPOP_DECOMPRESS, src, entry->length,
759 dst, &dlen);
760 kunmap_atomic(dst);
761 zbud_unmap(zswap_pool, entry->handle);
762 BUG_ON(ret);
763
764 spin_lock(&tree->lock);
765 zswap_entry_put(tree, entry);
766 spin_unlock(&tree->lock);
767
768 return 0;
769}
770
771/* frees an entry in zswap */
772static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
773{
774 struct zswap_tree *tree = zswap_trees[type];
775 struct zswap_entry *entry;
776
777 /* find */
778 spin_lock(&tree->lock);
779 entry = zswap_rb_search(&tree->rbroot, offset);
780 if (!entry) {
781 /* entry was written back */
782 spin_unlock(&tree->lock);
783 return;
784 }
785
786 /* remove from rbtree */
787 zswap_rb_erase(&tree->rbroot, entry);
788
789 /* drop the initial reference from entry creation */
790 zswap_entry_put(tree, entry);
791
792 spin_unlock(&tree->lock);
793}
794
795/* frees all zswap entries for the given swap type */
796static void zswap_frontswap_invalidate_area(unsigned type)
797{
798 struct zswap_tree *tree = zswap_trees[type];
799 struct zswap_entry *entry, *n;
800
801 if (!tree)
802 return;
803
804 /* walk the tree and free everything */
805 spin_lock(&tree->lock);
806 rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
807 zswap_free_entry(entry);
808 tree->rbroot = RB_ROOT;
809 spin_unlock(&tree->lock);
810 kfree(tree);
811 zswap_trees[type] = NULL;
812}
813
814static struct zbud_ops zswap_zbud_ops = {
815 .evict = zswap_writeback_entry
816};
817
818static void zswap_frontswap_init(unsigned type)
819{
820 struct zswap_tree *tree;
821
822 tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
823 if (!tree) {
824 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
825 return;
826 }
827
828 tree->rbroot = RB_ROOT;
829 spin_lock_init(&tree->lock);
830 zswap_trees[type] = tree;
831}
832
833static struct frontswap_ops zswap_frontswap_ops = {
834 .store = zswap_frontswap_store,
835 .load = zswap_frontswap_load,
836 .invalidate_page = zswap_frontswap_invalidate_page,
837 .invalidate_area = zswap_frontswap_invalidate_area,
838 .init = zswap_frontswap_init
839};
840
841/*********************************
842* debugfs functions
843**********************************/
844#ifdef CONFIG_DEBUG_FS
845#include <linux/debugfs.h>
846
847static struct dentry *zswap_debugfs_root;
848
849static int __init zswap_debugfs_init(void)
850{
851 if (!debugfs_initialized())
852 return -ENODEV;
853
854 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
855 if (!zswap_debugfs_root)
856 return -ENOMEM;
857
858 debugfs_create_u64("pool_limit_hit", S_IRUGO,
859 zswap_debugfs_root, &zswap_pool_limit_hit);
860 debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
861 zswap_debugfs_root, &zswap_reject_reclaim_fail);
862 debugfs_create_u64("reject_alloc_fail", S_IRUGO,
863 zswap_debugfs_root, &zswap_reject_alloc_fail);
864 debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
865 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
866 debugfs_create_u64("reject_compress_poor", S_IRUGO,
867 zswap_debugfs_root, &zswap_reject_compress_poor);
868 debugfs_create_u64("written_back_pages", S_IRUGO,
869 zswap_debugfs_root, &zswap_written_back_pages);
870 debugfs_create_u64("duplicate_entry", S_IRUGO,
871 zswap_debugfs_root, &zswap_duplicate_entry);
872 debugfs_create_u64("pool_pages", S_IRUGO,
873 zswap_debugfs_root, &zswap_pool_pages);
874 debugfs_create_atomic_t("stored_pages", S_IRUGO,
875 zswap_debugfs_root, &zswap_stored_pages);
876
877 return 0;
878}
879
880static void __exit zswap_debugfs_exit(void)
881{
882 debugfs_remove_recursive(zswap_debugfs_root);
883}
884#else
885static int __init zswap_debugfs_init(void)
886{
887 return 0;
888}
889
890static void __exit zswap_debugfs_exit(void) { }
891#endif
892
893/*********************************
894* module init and exit
895**********************************/
896static int __init init_zswap(void)
897{
898 if (!zswap_enabled)
899 return 0;
900
901 pr_info("loading zswap\n");
902
903 zswap_pool = zbud_create_pool(GFP_KERNEL, &zswap_zbud_ops);
904 if (!zswap_pool) {
905 pr_err("zbud pool creation failed\n");
906 goto error;
907 }
908
909 if (zswap_entry_cache_create()) {
910 pr_err("entry cache creation failed\n");
911 goto cachefail;
912 }
913 if (zswap_comp_init()) {
914 pr_err("compressor initialization failed\n");
915 goto compfail;
916 }
917 if (zswap_cpu_init()) {
918 pr_err("per-cpu initialization failed\n");
919 goto pcpufail;
920 }
921
922 frontswap_register_ops(&zswap_frontswap_ops);
923 if (zswap_debugfs_init())
924 pr_warn("debugfs initialization failed\n");
925 return 0;
926pcpufail:
927 zswap_comp_exit();
928compfail:
929 zswap_entry_cache_destory();
930cachefail:
931 zbud_destroy_pool(zswap_pool);
932error:
933 return -ENOMEM;
934}
935/* must be late so crypto has time to come up */
936late_initcall(init_zswap);
937
938MODULE_LICENSE("GPL");
939MODULE_AUTHOR("Seth Jennings <sjenning@linux.vnet.ibm.com>");
940MODULE_DESCRIPTION("Compressed cache for swap pages");
1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * zswap.c - zswap driver file
4 *
5 * zswap is a backend for frontswap that takes pages that are in the process
6 * of being swapped out and attempts to compress and store them in a
7 * RAM-based memory pool. This can result in a significant I/O reduction on
8 * the swap device and, in the case where decompressing from RAM is faster
9 * than reading from the swap device, can also improve workload performance.
10 *
11 * Copyright (C) 2012 Seth Jennings <sjenning@linux.vnet.ibm.com>
12*/
13
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
16#include <linux/module.h>
17#include <linux/cpu.h>
18#include <linux/highmem.h>
19#include <linux/slab.h>
20#include <linux/spinlock.h>
21#include <linux/types.h>
22#include <linux/atomic.h>
23#include <linux/frontswap.h>
24#include <linux/rbtree.h>
25#include <linux/swap.h>
26#include <linux/crypto.h>
27#include <linux/scatterlist.h>
28#include <linux/mempool.h>
29#include <linux/zpool.h>
30#include <crypto/acompress.h>
31
32#include <linux/mm_types.h>
33#include <linux/page-flags.h>
34#include <linux/swapops.h>
35#include <linux/writeback.h>
36#include <linux/pagemap.h>
37#include <linux/workqueue.h>
38
39#include "swap.h"
40
41/*********************************
42* statistics
43**********************************/
44/* Total bytes used by the compressed storage */
45u64 zswap_pool_total_size;
46/* The number of compressed pages currently stored in zswap */
47atomic_t zswap_stored_pages = ATOMIC_INIT(0);
48/* The number of same-value filled pages currently stored in zswap */
49static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
50
51/*
52 * The statistics below are not protected from concurrent access for
53 * performance reasons so they may not be a 100% accurate. However,
54 * they do provide useful information on roughly how many times a
55 * certain event is occurring.
56*/
57
58/* Pool limit was hit (see zswap_max_pool_percent) */
59static u64 zswap_pool_limit_hit;
60/* Pages written back when pool limit was reached */
61static u64 zswap_written_back_pages;
62/* Store failed due to a reclaim failure after pool limit was reached */
63static u64 zswap_reject_reclaim_fail;
64/* Compressed page was too big for the allocator to (optimally) store */
65static u64 zswap_reject_compress_poor;
66/* Store failed because underlying allocator could not get memory */
67static u64 zswap_reject_alloc_fail;
68/* Store failed because the entry metadata could not be allocated (rare) */
69static u64 zswap_reject_kmemcache_fail;
70/* Duplicate store was encountered (rare) */
71static u64 zswap_duplicate_entry;
72
73/* Shrinker work queue */
74static struct workqueue_struct *shrink_wq;
75/* Pool limit was hit, we need to calm down */
76static bool zswap_pool_reached_full;
77
78/*********************************
79* tunables
80**********************************/
81
82#define ZSWAP_PARAM_UNSET ""
83
84/* Enable/disable zswap */
85static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
86static int zswap_enabled_param_set(const char *,
87 const struct kernel_param *);
88static const struct kernel_param_ops zswap_enabled_param_ops = {
89 .set = zswap_enabled_param_set,
90 .get = param_get_bool,
91};
92module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
93
94/* Crypto compressor to use */
95static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
96static int zswap_compressor_param_set(const char *,
97 const struct kernel_param *);
98static const struct kernel_param_ops zswap_compressor_param_ops = {
99 .set = zswap_compressor_param_set,
100 .get = param_get_charp,
101 .free = param_free_charp,
102};
103module_param_cb(compressor, &zswap_compressor_param_ops,
104 &zswap_compressor, 0644);
105
106/* Compressed storage zpool to use */
107static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
108static int zswap_zpool_param_set(const char *, const struct kernel_param *);
109static const struct kernel_param_ops zswap_zpool_param_ops = {
110 .set = zswap_zpool_param_set,
111 .get = param_get_charp,
112 .free = param_free_charp,
113};
114module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
115
116/* The maximum percentage of memory that the compressed pool can occupy */
117static unsigned int zswap_max_pool_percent = 20;
118module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
119
120/* The threshold for accepting new pages after the max_pool_percent was hit */
121static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
122module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
123 uint, 0644);
124
125/*
126 * Enable/disable handling same-value filled pages (enabled by default).
127 * If disabled every page is considered non-same-value filled.
128 */
129static bool zswap_same_filled_pages_enabled = true;
130module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
131 bool, 0644);
132
133/* Enable/disable handling non-same-value filled pages (enabled by default) */
134static bool zswap_non_same_filled_pages_enabled = true;
135module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
136 bool, 0644);
137
138/*********************************
139* data structures
140**********************************/
141
142struct crypto_acomp_ctx {
143 struct crypto_acomp *acomp;
144 struct acomp_req *req;
145 struct crypto_wait wait;
146 u8 *dstmem;
147 struct mutex *mutex;
148};
149
150struct zswap_pool {
151 struct zpool *zpool;
152 struct crypto_acomp_ctx __percpu *acomp_ctx;
153 struct kref kref;
154 struct list_head list;
155 struct work_struct release_work;
156 struct work_struct shrink_work;
157 struct hlist_node node;
158 char tfm_name[CRYPTO_MAX_ALG_NAME];
159};
160
161/*
162 * struct zswap_entry
163 *
164 * This structure contains the metadata for tracking a single compressed
165 * page within zswap.
166 *
167 * rbnode - links the entry into red-black tree for the appropriate swap type
168 * offset - the swap offset for the entry. Index into the red-black tree.
169 * refcount - the number of outstanding reference to the entry. This is needed
170 * to protect against premature freeing of the entry by code
171 * concurrent calls to load, invalidate, and writeback. The lock
172 * for the zswap_tree structure that contains the entry must
173 * be held while changing the refcount. Since the lock must
174 * be held, there is no reason to also make refcount atomic.
175 * length - the length in bytes of the compressed page data. Needed during
176 * decompression. For a same value filled page length is 0.
177 * pool - the zswap_pool the entry's data is in
178 * handle - zpool allocation handle that stores the compressed page data
179 * value - value of the same-value filled pages which have same content
180 */
181struct zswap_entry {
182 struct rb_node rbnode;
183 pgoff_t offset;
184 int refcount;
185 unsigned int length;
186 struct zswap_pool *pool;
187 union {
188 unsigned long handle;
189 unsigned long value;
190 };
191 struct obj_cgroup *objcg;
192};
193
194struct zswap_header {
195 swp_entry_t swpentry;
196};
197
198/*
199 * The tree lock in the zswap_tree struct protects a few things:
200 * - the rbtree
201 * - the refcount field of each entry in the tree
202 */
203struct zswap_tree {
204 struct rb_root rbroot;
205 spinlock_t lock;
206};
207
208static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
209
210/* RCU-protected iteration */
211static LIST_HEAD(zswap_pools);
212/* protects zswap_pools list modification */
213static DEFINE_SPINLOCK(zswap_pools_lock);
214/* pool counter to provide unique names to zpool */
215static atomic_t zswap_pools_count = ATOMIC_INIT(0);
216
217/* used by param callback function */
218static bool zswap_init_started;
219
220/* fatal error during init */
221static bool zswap_init_failed;
222
223/* init completed, but couldn't create the initial pool */
224static bool zswap_has_pool;
225
226/*********************************
227* helpers and fwd declarations
228**********************************/
229
230#define zswap_pool_debug(msg, p) \
231 pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name, \
232 zpool_get_type((p)->zpool))
233
234static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
235static int zswap_pool_get(struct zswap_pool *pool);
236static void zswap_pool_put(struct zswap_pool *pool);
237
238static const struct zpool_ops zswap_zpool_ops = {
239 .evict = zswap_writeback_entry
240};
241
242static bool zswap_is_full(void)
243{
244 return totalram_pages() * zswap_max_pool_percent / 100 <
245 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
246}
247
248static bool zswap_can_accept(void)
249{
250 return totalram_pages() * zswap_accept_thr_percent / 100 *
251 zswap_max_pool_percent / 100 >
252 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
253}
254
255static void zswap_update_total_size(void)
256{
257 struct zswap_pool *pool;
258 u64 total = 0;
259
260 rcu_read_lock();
261
262 list_for_each_entry_rcu(pool, &zswap_pools, list)
263 total += zpool_get_total_size(pool->zpool);
264
265 rcu_read_unlock();
266
267 zswap_pool_total_size = total;
268}
269
270/*********************************
271* zswap entry functions
272**********************************/
273static struct kmem_cache *zswap_entry_cache;
274
275static int __init zswap_entry_cache_create(void)
276{
277 zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
278 return zswap_entry_cache == NULL;
279}
280
281static void __init zswap_entry_cache_destroy(void)
282{
283 kmem_cache_destroy(zswap_entry_cache);
284}
285
286static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
287{
288 struct zswap_entry *entry;
289 entry = kmem_cache_alloc(zswap_entry_cache, gfp);
290 if (!entry)
291 return NULL;
292 entry->refcount = 1;
293 RB_CLEAR_NODE(&entry->rbnode);
294 return entry;
295}
296
297static void zswap_entry_cache_free(struct zswap_entry *entry)
298{
299 kmem_cache_free(zswap_entry_cache, entry);
300}
301
302/*********************************
303* rbtree functions
304**********************************/
305static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
306{
307 struct rb_node *node = root->rb_node;
308 struct zswap_entry *entry;
309
310 while (node) {
311 entry = rb_entry(node, struct zswap_entry, rbnode);
312 if (entry->offset > offset)
313 node = node->rb_left;
314 else if (entry->offset < offset)
315 node = node->rb_right;
316 else
317 return entry;
318 }
319 return NULL;
320}
321
322/*
323 * In the case that a entry with the same offset is found, a pointer to
324 * the existing entry is stored in dupentry and the function returns -EEXIST
325 */
326static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
327 struct zswap_entry **dupentry)
328{
329 struct rb_node **link = &root->rb_node, *parent = NULL;
330 struct zswap_entry *myentry;
331
332 while (*link) {
333 parent = *link;
334 myentry = rb_entry(parent, struct zswap_entry, rbnode);
335 if (myentry->offset > entry->offset)
336 link = &(*link)->rb_left;
337 else if (myentry->offset < entry->offset)
338 link = &(*link)->rb_right;
339 else {
340 *dupentry = myentry;
341 return -EEXIST;
342 }
343 }
344 rb_link_node(&entry->rbnode, parent, link);
345 rb_insert_color(&entry->rbnode, root);
346 return 0;
347}
348
349static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
350{
351 if (!RB_EMPTY_NODE(&entry->rbnode)) {
352 rb_erase(&entry->rbnode, root);
353 RB_CLEAR_NODE(&entry->rbnode);
354 }
355}
356
357/*
358 * Carries out the common pattern of freeing and entry's zpool allocation,
359 * freeing the entry itself, and decrementing the number of stored pages.
360 */
361static void zswap_free_entry(struct zswap_entry *entry)
362{
363 if (entry->objcg) {
364 obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
365 obj_cgroup_put(entry->objcg);
366 }
367 if (!entry->length)
368 atomic_dec(&zswap_same_filled_pages);
369 else {
370 zpool_free(entry->pool->zpool, entry->handle);
371 zswap_pool_put(entry->pool);
372 }
373 zswap_entry_cache_free(entry);
374 atomic_dec(&zswap_stored_pages);
375 zswap_update_total_size();
376}
377
378/* caller must hold the tree lock */
379static void zswap_entry_get(struct zswap_entry *entry)
380{
381 entry->refcount++;
382}
383
384/* caller must hold the tree lock
385* remove from the tree and free it, if nobody reference the entry
386*/
387static void zswap_entry_put(struct zswap_tree *tree,
388 struct zswap_entry *entry)
389{
390 int refcount = --entry->refcount;
391
392 BUG_ON(refcount < 0);
393 if (refcount == 0) {
394 zswap_rb_erase(&tree->rbroot, entry);
395 zswap_free_entry(entry);
396 }
397}
398
399/* caller must hold the tree lock */
400static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
401 pgoff_t offset)
402{
403 struct zswap_entry *entry;
404
405 entry = zswap_rb_search(root, offset);
406 if (entry)
407 zswap_entry_get(entry);
408
409 return entry;
410}
411
412/*********************************
413* per-cpu code
414**********************************/
415static DEFINE_PER_CPU(u8 *, zswap_dstmem);
416/*
417 * If users dynamically change the zpool type and compressor at runtime, i.e.
418 * zswap is running, zswap can have more than one zpool on one cpu, but they
419 * are sharing dtsmem. So we need this mutex to be per-cpu.
420 */
421static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
422
423static int zswap_dstmem_prepare(unsigned int cpu)
424{
425 struct mutex *mutex;
426 u8 *dst;
427
428 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
429 if (!dst)
430 return -ENOMEM;
431
432 mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
433 if (!mutex) {
434 kfree(dst);
435 return -ENOMEM;
436 }
437
438 mutex_init(mutex);
439 per_cpu(zswap_dstmem, cpu) = dst;
440 per_cpu(zswap_mutex, cpu) = mutex;
441 return 0;
442}
443
444static int zswap_dstmem_dead(unsigned int cpu)
445{
446 struct mutex *mutex;
447 u8 *dst;
448
449 mutex = per_cpu(zswap_mutex, cpu);
450 kfree(mutex);
451 per_cpu(zswap_mutex, cpu) = NULL;
452
453 dst = per_cpu(zswap_dstmem, cpu);
454 kfree(dst);
455 per_cpu(zswap_dstmem, cpu) = NULL;
456
457 return 0;
458}
459
460static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
461{
462 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
463 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
464 struct crypto_acomp *acomp;
465 struct acomp_req *req;
466
467 acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
468 if (IS_ERR(acomp)) {
469 pr_err("could not alloc crypto acomp %s : %ld\n",
470 pool->tfm_name, PTR_ERR(acomp));
471 return PTR_ERR(acomp);
472 }
473 acomp_ctx->acomp = acomp;
474
475 req = acomp_request_alloc(acomp_ctx->acomp);
476 if (!req) {
477 pr_err("could not alloc crypto acomp_request %s\n",
478 pool->tfm_name);
479 crypto_free_acomp(acomp_ctx->acomp);
480 return -ENOMEM;
481 }
482 acomp_ctx->req = req;
483
484 crypto_init_wait(&acomp_ctx->wait);
485 /*
486 * if the backend of acomp is async zip, crypto_req_done() will wakeup
487 * crypto_wait_req(); if the backend of acomp is scomp, the callback
488 * won't be called, crypto_wait_req() will return without blocking.
489 */
490 acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
491 crypto_req_done, &acomp_ctx->wait);
492
493 acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
494 acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
495
496 return 0;
497}
498
499static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
500{
501 struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
502 struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
503
504 if (!IS_ERR_OR_NULL(acomp_ctx)) {
505 if (!IS_ERR_OR_NULL(acomp_ctx->req))
506 acomp_request_free(acomp_ctx->req);
507 if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
508 crypto_free_acomp(acomp_ctx->acomp);
509 }
510
511 return 0;
512}
513
514/*********************************
515* pool functions
516**********************************/
517
518static struct zswap_pool *__zswap_pool_current(void)
519{
520 struct zswap_pool *pool;
521
522 pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
523 WARN_ONCE(!pool && zswap_has_pool,
524 "%s: no page storage pool!\n", __func__);
525
526 return pool;
527}
528
529static struct zswap_pool *zswap_pool_current(void)
530{
531 assert_spin_locked(&zswap_pools_lock);
532
533 return __zswap_pool_current();
534}
535
536static struct zswap_pool *zswap_pool_current_get(void)
537{
538 struct zswap_pool *pool;
539
540 rcu_read_lock();
541
542 pool = __zswap_pool_current();
543 if (!zswap_pool_get(pool))
544 pool = NULL;
545
546 rcu_read_unlock();
547
548 return pool;
549}
550
551static struct zswap_pool *zswap_pool_last_get(void)
552{
553 struct zswap_pool *pool, *last = NULL;
554
555 rcu_read_lock();
556
557 list_for_each_entry_rcu(pool, &zswap_pools, list)
558 last = pool;
559 WARN_ONCE(!last && zswap_has_pool,
560 "%s: no page storage pool!\n", __func__);
561 if (!zswap_pool_get(last))
562 last = NULL;
563
564 rcu_read_unlock();
565
566 return last;
567}
568
569/* type and compressor must be null-terminated */
570static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
571{
572 struct zswap_pool *pool;
573
574 assert_spin_locked(&zswap_pools_lock);
575
576 list_for_each_entry_rcu(pool, &zswap_pools, list) {
577 if (strcmp(pool->tfm_name, compressor))
578 continue;
579 if (strcmp(zpool_get_type(pool->zpool), type))
580 continue;
581 /* if we can't get it, it's about to be destroyed */
582 if (!zswap_pool_get(pool))
583 continue;
584 return pool;
585 }
586
587 return NULL;
588}
589
590static void shrink_worker(struct work_struct *w)
591{
592 struct zswap_pool *pool = container_of(w, typeof(*pool),
593 shrink_work);
594
595 if (zpool_shrink(pool->zpool, 1, NULL))
596 zswap_reject_reclaim_fail++;
597 zswap_pool_put(pool);
598}
599
600static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
601{
602 struct zswap_pool *pool;
603 char name[38]; /* 'zswap' + 32 char (max) num + \0 */
604 gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
605 int ret;
606
607 if (!zswap_has_pool) {
608 /* if either are unset, pool initialization failed, and we
609 * need both params to be set correctly before trying to
610 * create a pool.
611 */
612 if (!strcmp(type, ZSWAP_PARAM_UNSET))
613 return NULL;
614 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
615 return NULL;
616 }
617
618 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
619 if (!pool)
620 return NULL;
621
622 /* unique name for each pool specifically required by zsmalloc */
623 snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
624
625 pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
626 if (!pool->zpool) {
627 pr_err("%s zpool not available\n", type);
628 goto error;
629 }
630 pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
631
632 strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
633
634 pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
635 if (!pool->acomp_ctx) {
636 pr_err("percpu alloc failed\n");
637 goto error;
638 }
639
640 ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
641 &pool->node);
642 if (ret)
643 goto error;
644 pr_debug("using %s compressor\n", pool->tfm_name);
645
646 /* being the current pool takes 1 ref; this func expects the
647 * caller to always add the new pool as the current pool
648 */
649 kref_init(&pool->kref);
650 INIT_LIST_HEAD(&pool->list);
651 INIT_WORK(&pool->shrink_work, shrink_worker);
652
653 zswap_pool_debug("created", pool);
654
655 return pool;
656
657error:
658 if (pool->acomp_ctx)
659 free_percpu(pool->acomp_ctx);
660 if (pool->zpool)
661 zpool_destroy_pool(pool->zpool);
662 kfree(pool);
663 return NULL;
664}
665
666static __init struct zswap_pool *__zswap_pool_create_fallback(void)
667{
668 bool has_comp, has_zpool;
669
670 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
671 if (!has_comp && strcmp(zswap_compressor,
672 CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
673 pr_err("compressor %s not available, using default %s\n",
674 zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
675 param_free_charp(&zswap_compressor);
676 zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
677 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
678 }
679 if (!has_comp) {
680 pr_err("default compressor %s not available\n",
681 zswap_compressor);
682 param_free_charp(&zswap_compressor);
683 zswap_compressor = ZSWAP_PARAM_UNSET;
684 }
685
686 has_zpool = zpool_has_pool(zswap_zpool_type);
687 if (!has_zpool && strcmp(zswap_zpool_type,
688 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
689 pr_err("zpool %s not available, using default %s\n",
690 zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
691 param_free_charp(&zswap_zpool_type);
692 zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
693 has_zpool = zpool_has_pool(zswap_zpool_type);
694 }
695 if (!has_zpool) {
696 pr_err("default zpool %s not available\n",
697 zswap_zpool_type);
698 param_free_charp(&zswap_zpool_type);
699 zswap_zpool_type = ZSWAP_PARAM_UNSET;
700 }
701
702 if (!has_comp || !has_zpool)
703 return NULL;
704
705 return zswap_pool_create(zswap_zpool_type, zswap_compressor);
706}
707
708static void zswap_pool_destroy(struct zswap_pool *pool)
709{
710 zswap_pool_debug("destroying", pool);
711
712 cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
713 free_percpu(pool->acomp_ctx);
714 zpool_destroy_pool(pool->zpool);
715 kfree(pool);
716}
717
718static int __must_check zswap_pool_get(struct zswap_pool *pool)
719{
720 if (!pool)
721 return 0;
722
723 return kref_get_unless_zero(&pool->kref);
724}
725
726static void __zswap_pool_release(struct work_struct *work)
727{
728 struct zswap_pool *pool = container_of(work, typeof(*pool),
729 release_work);
730
731 synchronize_rcu();
732
733 /* nobody should have been able to get a kref... */
734 WARN_ON(kref_get_unless_zero(&pool->kref));
735
736 /* pool is now off zswap_pools list and has no references. */
737 zswap_pool_destroy(pool);
738}
739
740static void __zswap_pool_empty(struct kref *kref)
741{
742 struct zswap_pool *pool;
743
744 pool = container_of(kref, typeof(*pool), kref);
745
746 spin_lock(&zswap_pools_lock);
747
748 WARN_ON(pool == zswap_pool_current());
749
750 list_del_rcu(&pool->list);
751
752 INIT_WORK(&pool->release_work, __zswap_pool_release);
753 schedule_work(&pool->release_work);
754
755 spin_unlock(&zswap_pools_lock);
756}
757
758static void zswap_pool_put(struct zswap_pool *pool)
759{
760 kref_put(&pool->kref, __zswap_pool_empty);
761}
762
763/*********************************
764* param callbacks
765**********************************/
766
767/* val must be a null-terminated string */
768static int __zswap_param_set(const char *val, const struct kernel_param *kp,
769 char *type, char *compressor)
770{
771 struct zswap_pool *pool, *put_pool = NULL;
772 char *s = strstrip((char *)val);
773 int ret;
774
775 if (zswap_init_failed) {
776 pr_err("can't set param, initialization failed\n");
777 return -ENODEV;
778 }
779
780 /* no change required */
781 if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
782 return 0;
783
784 /* if this is load-time (pre-init) param setting,
785 * don't create a pool; that's done during init.
786 */
787 if (!zswap_init_started)
788 return param_set_charp(s, kp);
789
790 if (!type) {
791 if (!zpool_has_pool(s)) {
792 pr_err("zpool %s not available\n", s);
793 return -ENOENT;
794 }
795 type = s;
796 } else if (!compressor) {
797 if (!crypto_has_acomp(s, 0, 0)) {
798 pr_err("compressor %s not available\n", s);
799 return -ENOENT;
800 }
801 compressor = s;
802 } else {
803 WARN_ON(1);
804 return -EINVAL;
805 }
806
807 spin_lock(&zswap_pools_lock);
808
809 pool = zswap_pool_find_get(type, compressor);
810 if (pool) {
811 zswap_pool_debug("using existing", pool);
812 WARN_ON(pool == zswap_pool_current());
813 list_del_rcu(&pool->list);
814 }
815
816 spin_unlock(&zswap_pools_lock);
817
818 if (!pool)
819 pool = zswap_pool_create(type, compressor);
820
821 if (pool)
822 ret = param_set_charp(s, kp);
823 else
824 ret = -EINVAL;
825
826 spin_lock(&zswap_pools_lock);
827
828 if (!ret) {
829 put_pool = zswap_pool_current();
830 list_add_rcu(&pool->list, &zswap_pools);
831 zswap_has_pool = true;
832 } else if (pool) {
833 /* add the possibly pre-existing pool to the end of the pools
834 * list; if it's new (and empty) then it'll be removed and
835 * destroyed by the put after we drop the lock
836 */
837 list_add_tail_rcu(&pool->list, &zswap_pools);
838 put_pool = pool;
839 }
840
841 spin_unlock(&zswap_pools_lock);
842
843 if (!zswap_has_pool && !pool) {
844 /* if initial pool creation failed, and this pool creation also
845 * failed, maybe both compressor and zpool params were bad.
846 * Allow changing this param, so pool creation will succeed
847 * when the other param is changed. We already verified this
848 * param is ok in the zpool_has_pool() or crypto_has_acomp()
849 * checks above.
850 */
851 ret = param_set_charp(s, kp);
852 }
853
854 /* drop the ref from either the old current pool,
855 * or the new pool we failed to add
856 */
857 if (put_pool)
858 zswap_pool_put(put_pool);
859
860 return ret;
861}
862
863static int zswap_compressor_param_set(const char *val,
864 const struct kernel_param *kp)
865{
866 return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
867}
868
869static int zswap_zpool_param_set(const char *val,
870 const struct kernel_param *kp)
871{
872 return __zswap_param_set(val, kp, NULL, zswap_compressor);
873}
874
875static int zswap_enabled_param_set(const char *val,
876 const struct kernel_param *kp)
877{
878 if (zswap_init_failed) {
879 pr_err("can't enable, initialization failed\n");
880 return -ENODEV;
881 }
882 if (!zswap_has_pool && zswap_init_started) {
883 pr_err("can't enable, no pool configured\n");
884 return -ENODEV;
885 }
886
887 return param_set_bool(val, kp);
888}
889
890/*********************************
891* writeback code
892**********************************/
893/* return enum for zswap_get_swap_cache_page */
894enum zswap_get_swap_ret {
895 ZSWAP_SWAPCACHE_NEW,
896 ZSWAP_SWAPCACHE_EXIST,
897 ZSWAP_SWAPCACHE_FAIL,
898};
899
900/*
901 * zswap_get_swap_cache_page
902 *
903 * This is an adaption of read_swap_cache_async()
904 *
905 * This function tries to find a page with the given swap entry
906 * in the swapper_space address space (the swap cache). If the page
907 * is found, it is returned in retpage. Otherwise, a page is allocated,
908 * added to the swap cache, and returned in retpage.
909 *
910 * If success, the swap cache page is returned in retpage
911 * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
912 * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
913 * the new page is added to swapcache and locked
914 * Returns ZSWAP_SWAPCACHE_FAIL on error
915 */
916static int zswap_get_swap_cache_page(swp_entry_t entry,
917 struct page **retpage)
918{
919 bool page_was_allocated;
920
921 *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
922 NULL, 0, &page_was_allocated);
923 if (page_was_allocated)
924 return ZSWAP_SWAPCACHE_NEW;
925 if (!*retpage)
926 return ZSWAP_SWAPCACHE_FAIL;
927 return ZSWAP_SWAPCACHE_EXIST;
928}
929
930/*
931 * Attempts to free an entry by adding a page to the swap cache,
932 * decompressing the entry data into the page, and issuing a
933 * bio write to write the page back to the swap device.
934 *
935 * This can be thought of as a "resumed writeback" of the page
936 * to the swap device. We are basically resuming the same swap
937 * writeback path that was intercepted with the frontswap_store()
938 * in the first place. After the page has been decompressed into
939 * the swap cache, the compressed version stored by zswap can be
940 * freed.
941 */
942static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
943{
944 struct zswap_header *zhdr;
945 swp_entry_t swpentry;
946 struct zswap_tree *tree;
947 pgoff_t offset;
948 struct zswap_entry *entry;
949 struct page *page;
950 struct scatterlist input, output;
951 struct crypto_acomp_ctx *acomp_ctx;
952
953 u8 *src, *tmp = NULL;
954 unsigned int dlen;
955 int ret;
956 struct writeback_control wbc = {
957 .sync_mode = WB_SYNC_NONE,
958 };
959
960 if (!zpool_can_sleep_mapped(pool)) {
961 tmp = kmalloc(PAGE_SIZE, GFP_KERNEL);
962 if (!tmp)
963 return -ENOMEM;
964 }
965
966 /* extract swpentry from data */
967 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
968 swpentry = zhdr->swpentry; /* here */
969 tree = zswap_trees[swp_type(swpentry)];
970 offset = swp_offset(swpentry);
971 zpool_unmap_handle(pool, handle);
972
973 /* find and ref zswap entry */
974 spin_lock(&tree->lock);
975 entry = zswap_entry_find_get(&tree->rbroot, offset);
976 if (!entry) {
977 /* entry was invalidated */
978 spin_unlock(&tree->lock);
979 kfree(tmp);
980 return 0;
981 }
982 spin_unlock(&tree->lock);
983 BUG_ON(offset != entry->offset);
984
985 /* try to allocate swap cache page */
986 switch (zswap_get_swap_cache_page(swpentry, &page)) {
987 case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
988 ret = -ENOMEM;
989 goto fail;
990
991 case ZSWAP_SWAPCACHE_EXIST:
992 /* page is already in the swap cache, ignore for now */
993 put_page(page);
994 ret = -EEXIST;
995 goto fail;
996
997 case ZSWAP_SWAPCACHE_NEW: /* page is locked */
998 /* decompress */
999 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1000 dlen = PAGE_SIZE;
1001
1002 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
1003 src = (u8 *)zhdr + sizeof(struct zswap_header);
1004 if (!zpool_can_sleep_mapped(pool)) {
1005 memcpy(tmp, src, entry->length);
1006 src = tmp;
1007 zpool_unmap_handle(pool, handle);
1008 }
1009
1010 mutex_lock(acomp_ctx->mutex);
1011 sg_init_one(&input, src, entry->length);
1012 sg_init_table(&output, 1);
1013 sg_set_page(&output, page, PAGE_SIZE, 0);
1014 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1015 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1016 dlen = acomp_ctx->req->dlen;
1017 mutex_unlock(acomp_ctx->mutex);
1018
1019 if (!zpool_can_sleep_mapped(pool))
1020 kfree(tmp);
1021 else
1022 zpool_unmap_handle(pool, handle);
1023
1024 BUG_ON(ret);
1025 BUG_ON(dlen != PAGE_SIZE);
1026
1027 /* page is up to date */
1028 SetPageUptodate(page);
1029 }
1030
1031 /* move it to the tail of the inactive list after end_writeback */
1032 SetPageReclaim(page);
1033
1034 /* start writeback */
1035 __swap_writepage(page, &wbc);
1036 put_page(page);
1037 zswap_written_back_pages++;
1038
1039 spin_lock(&tree->lock);
1040 /* drop local reference */
1041 zswap_entry_put(tree, entry);
1042
1043 /*
1044 * There are two possible situations for entry here:
1045 * (1) refcount is 1(normal case), entry is valid and on the tree
1046 * (2) refcount is 0, entry is freed and not on the tree
1047 * because invalidate happened during writeback
1048 * search the tree and free the entry if find entry
1049 */
1050 if (entry == zswap_rb_search(&tree->rbroot, offset))
1051 zswap_entry_put(tree, entry);
1052 spin_unlock(&tree->lock);
1053
1054 return ret;
1055
1056fail:
1057 if (!zpool_can_sleep_mapped(pool))
1058 kfree(tmp);
1059
1060 /*
1061 * if we get here due to ZSWAP_SWAPCACHE_EXIST
1062 * a load may be happening concurrently.
1063 * it is safe and okay to not free the entry.
1064 * if we free the entry in the following put
1065 * it is also okay to return !0
1066 */
1067 spin_lock(&tree->lock);
1068 zswap_entry_put(tree, entry);
1069 spin_unlock(&tree->lock);
1070
1071 return ret;
1072}
1073
1074static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1075{
1076 unsigned int pos;
1077 unsigned long *page;
1078
1079 page = (unsigned long *)ptr;
1080 for (pos = 1; pos < PAGE_SIZE / sizeof(*page); pos++) {
1081 if (page[pos] != page[0])
1082 return 0;
1083 }
1084 *value = page[0];
1085 return 1;
1086}
1087
1088static void zswap_fill_page(void *ptr, unsigned long value)
1089{
1090 unsigned long *page;
1091
1092 page = (unsigned long *)ptr;
1093 memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1094}
1095
1096/*********************************
1097* frontswap hooks
1098**********************************/
1099/* attempts to compress and store an single page */
1100static int zswap_frontswap_store(unsigned type, pgoff_t offset,
1101 struct page *page)
1102{
1103 struct zswap_tree *tree = zswap_trees[type];
1104 struct zswap_entry *entry, *dupentry;
1105 struct scatterlist input, output;
1106 struct crypto_acomp_ctx *acomp_ctx;
1107 struct obj_cgroup *objcg = NULL;
1108 struct zswap_pool *pool;
1109 int ret;
1110 unsigned int hlen, dlen = PAGE_SIZE;
1111 unsigned long handle, value;
1112 char *buf;
1113 u8 *src, *dst;
1114 struct zswap_header zhdr = { .swpentry = swp_entry(type, offset) };
1115 gfp_t gfp;
1116
1117 /* THP isn't supported */
1118 if (PageTransHuge(page)) {
1119 ret = -EINVAL;
1120 goto reject;
1121 }
1122
1123 if (!zswap_enabled || !tree) {
1124 ret = -ENODEV;
1125 goto reject;
1126 }
1127
1128 objcg = get_obj_cgroup_from_page(page);
1129 if (objcg && !obj_cgroup_may_zswap(objcg))
1130 goto shrink;
1131
1132 /* reclaim space if needed */
1133 if (zswap_is_full()) {
1134 zswap_pool_limit_hit++;
1135 zswap_pool_reached_full = true;
1136 goto shrink;
1137 }
1138
1139 if (zswap_pool_reached_full) {
1140 if (!zswap_can_accept()) {
1141 ret = -ENOMEM;
1142 goto reject;
1143 } else
1144 zswap_pool_reached_full = false;
1145 }
1146
1147 /* allocate entry */
1148 entry = zswap_entry_cache_alloc(GFP_KERNEL);
1149 if (!entry) {
1150 zswap_reject_kmemcache_fail++;
1151 ret = -ENOMEM;
1152 goto reject;
1153 }
1154
1155 if (zswap_same_filled_pages_enabled) {
1156 src = kmap_atomic(page);
1157 if (zswap_is_page_same_filled(src, &value)) {
1158 kunmap_atomic(src);
1159 entry->offset = offset;
1160 entry->length = 0;
1161 entry->value = value;
1162 atomic_inc(&zswap_same_filled_pages);
1163 goto insert_entry;
1164 }
1165 kunmap_atomic(src);
1166 }
1167
1168 if (!zswap_non_same_filled_pages_enabled) {
1169 ret = -EINVAL;
1170 goto freepage;
1171 }
1172
1173 /* if entry is successfully added, it keeps the reference */
1174 entry->pool = zswap_pool_current_get();
1175 if (!entry->pool) {
1176 ret = -EINVAL;
1177 goto freepage;
1178 }
1179
1180 /* compress */
1181 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1182
1183 mutex_lock(acomp_ctx->mutex);
1184
1185 dst = acomp_ctx->dstmem;
1186 sg_init_table(&input, 1);
1187 sg_set_page(&input, page, PAGE_SIZE, 0);
1188
1189 /* zswap_dstmem is of size (PAGE_SIZE * 2). Reflect same in sg_list */
1190 sg_init_one(&output, dst, PAGE_SIZE * 2);
1191 acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1192 /*
1193 * it maybe looks a little bit silly that we send an asynchronous request,
1194 * then wait for its completion synchronously. This makes the process look
1195 * synchronous in fact.
1196 * Theoretically, acomp supports users send multiple acomp requests in one
1197 * acomp instance, then get those requests done simultaneously. but in this
1198 * case, frontswap actually does store and load page by page, there is no
1199 * existing method to send the second page before the first page is done
1200 * in one thread doing frontswap.
1201 * but in different threads running on different cpu, we have different
1202 * acomp instance, so multiple threads can do (de)compression in parallel.
1203 */
1204 ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1205 dlen = acomp_ctx->req->dlen;
1206
1207 if (ret) {
1208 ret = -EINVAL;
1209 goto put_dstmem;
1210 }
1211
1212 /* store */
1213 hlen = zpool_evictable(entry->pool->zpool) ? sizeof(zhdr) : 0;
1214 gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1215 if (zpool_malloc_support_movable(entry->pool->zpool))
1216 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1217 ret = zpool_malloc(entry->pool->zpool, hlen + dlen, gfp, &handle);
1218 if (ret == -ENOSPC) {
1219 zswap_reject_compress_poor++;
1220 goto put_dstmem;
1221 }
1222 if (ret) {
1223 zswap_reject_alloc_fail++;
1224 goto put_dstmem;
1225 }
1226 buf = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_WO);
1227 memcpy(buf, &zhdr, hlen);
1228 memcpy(buf + hlen, dst, dlen);
1229 zpool_unmap_handle(entry->pool->zpool, handle);
1230 mutex_unlock(acomp_ctx->mutex);
1231
1232 /* populate entry */
1233 entry->offset = offset;
1234 entry->handle = handle;
1235 entry->length = dlen;
1236
1237insert_entry:
1238 entry->objcg = objcg;
1239 if (objcg) {
1240 obj_cgroup_charge_zswap(objcg, entry->length);
1241 /* Account before objcg ref is moved to tree */
1242 count_objcg_event(objcg, ZSWPOUT);
1243 }
1244
1245 /* map */
1246 spin_lock(&tree->lock);
1247 do {
1248 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1249 if (ret == -EEXIST) {
1250 zswap_duplicate_entry++;
1251 /* remove from rbtree */
1252 zswap_rb_erase(&tree->rbroot, dupentry);
1253 zswap_entry_put(tree, dupentry);
1254 }
1255 } while (ret == -EEXIST);
1256 spin_unlock(&tree->lock);
1257
1258 /* update stats */
1259 atomic_inc(&zswap_stored_pages);
1260 zswap_update_total_size();
1261 count_vm_event(ZSWPOUT);
1262
1263 return 0;
1264
1265put_dstmem:
1266 mutex_unlock(acomp_ctx->mutex);
1267 zswap_pool_put(entry->pool);
1268freepage:
1269 zswap_entry_cache_free(entry);
1270reject:
1271 if (objcg)
1272 obj_cgroup_put(objcg);
1273 return ret;
1274
1275shrink:
1276 pool = zswap_pool_last_get();
1277 if (pool)
1278 queue_work(shrink_wq, &pool->shrink_work);
1279 ret = -ENOMEM;
1280 goto reject;
1281}
1282
1283/*
1284 * returns 0 if the page was successfully decompressed
1285 * return -1 on entry not found or error
1286*/
1287static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1288 struct page *page)
1289{
1290 struct zswap_tree *tree = zswap_trees[type];
1291 struct zswap_entry *entry;
1292 struct scatterlist input, output;
1293 struct crypto_acomp_ctx *acomp_ctx;
1294 u8 *src, *dst, *tmp;
1295 unsigned int dlen;
1296 int ret;
1297
1298 /* find */
1299 spin_lock(&tree->lock);
1300 entry = zswap_entry_find_get(&tree->rbroot, offset);
1301 if (!entry) {
1302 /* entry was written back */
1303 spin_unlock(&tree->lock);
1304 return -1;
1305 }
1306 spin_unlock(&tree->lock);
1307
1308 if (!entry->length) {
1309 dst = kmap_atomic(page);
1310 zswap_fill_page(dst, entry->value);
1311 kunmap_atomic(dst);
1312 ret = 0;
1313 goto stats;
1314 }
1315
1316 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1317 tmp = kmalloc(entry->length, GFP_KERNEL);
1318 if (!tmp) {
1319 ret = -ENOMEM;
1320 goto freeentry;
1321 }
1322 }
1323
1324 /* decompress */
1325 dlen = PAGE_SIZE;
1326 src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
1327 if (zpool_evictable(entry->pool->zpool))
1328 src += sizeof(struct zswap_header);
1329
1330 if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1331 memcpy(tmp, src, entry->length);
1332 src = tmp;
1333 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1334 }
1335
1336 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1337 mutex_lock(acomp_ctx->mutex);
1338 sg_init_one(&input, src, entry->length);
1339 sg_init_table(&output, 1);
1340 sg_set_page(&output, page, PAGE_SIZE, 0);
1341 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1342 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1343 mutex_unlock(acomp_ctx->mutex);
1344
1345 if (zpool_can_sleep_mapped(entry->pool->zpool))
1346 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1347 else
1348 kfree(tmp);
1349
1350 BUG_ON(ret);
1351stats:
1352 count_vm_event(ZSWPIN);
1353 if (entry->objcg)
1354 count_objcg_event(entry->objcg, ZSWPIN);
1355freeentry:
1356 spin_lock(&tree->lock);
1357 zswap_entry_put(tree, entry);
1358 spin_unlock(&tree->lock);
1359
1360 return ret;
1361}
1362
1363/* frees an entry in zswap */
1364static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1365{
1366 struct zswap_tree *tree = zswap_trees[type];
1367 struct zswap_entry *entry;
1368
1369 /* find */
1370 spin_lock(&tree->lock);
1371 entry = zswap_rb_search(&tree->rbroot, offset);
1372 if (!entry) {
1373 /* entry was written back */
1374 spin_unlock(&tree->lock);
1375 return;
1376 }
1377
1378 /* remove from rbtree */
1379 zswap_rb_erase(&tree->rbroot, entry);
1380
1381 /* drop the initial reference from entry creation */
1382 zswap_entry_put(tree, entry);
1383
1384 spin_unlock(&tree->lock);
1385}
1386
1387/* frees all zswap entries for the given swap type */
1388static void zswap_frontswap_invalidate_area(unsigned type)
1389{
1390 struct zswap_tree *tree = zswap_trees[type];
1391 struct zswap_entry *entry, *n;
1392
1393 if (!tree)
1394 return;
1395
1396 /* walk the tree and free everything */
1397 spin_lock(&tree->lock);
1398 rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1399 zswap_free_entry(entry);
1400 tree->rbroot = RB_ROOT;
1401 spin_unlock(&tree->lock);
1402 kfree(tree);
1403 zswap_trees[type] = NULL;
1404}
1405
1406static void zswap_frontswap_init(unsigned type)
1407{
1408 struct zswap_tree *tree;
1409
1410 tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1411 if (!tree) {
1412 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1413 return;
1414 }
1415
1416 tree->rbroot = RB_ROOT;
1417 spin_lock_init(&tree->lock);
1418 zswap_trees[type] = tree;
1419}
1420
1421static const struct frontswap_ops zswap_frontswap_ops = {
1422 .store = zswap_frontswap_store,
1423 .load = zswap_frontswap_load,
1424 .invalidate_page = zswap_frontswap_invalidate_page,
1425 .invalidate_area = zswap_frontswap_invalidate_area,
1426 .init = zswap_frontswap_init
1427};
1428
1429/*********************************
1430* debugfs functions
1431**********************************/
1432#ifdef CONFIG_DEBUG_FS
1433#include <linux/debugfs.h>
1434
1435static struct dentry *zswap_debugfs_root;
1436
1437static int __init zswap_debugfs_init(void)
1438{
1439 if (!debugfs_initialized())
1440 return -ENODEV;
1441
1442 zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1443
1444 debugfs_create_u64("pool_limit_hit", 0444,
1445 zswap_debugfs_root, &zswap_pool_limit_hit);
1446 debugfs_create_u64("reject_reclaim_fail", 0444,
1447 zswap_debugfs_root, &zswap_reject_reclaim_fail);
1448 debugfs_create_u64("reject_alloc_fail", 0444,
1449 zswap_debugfs_root, &zswap_reject_alloc_fail);
1450 debugfs_create_u64("reject_kmemcache_fail", 0444,
1451 zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1452 debugfs_create_u64("reject_compress_poor", 0444,
1453 zswap_debugfs_root, &zswap_reject_compress_poor);
1454 debugfs_create_u64("written_back_pages", 0444,
1455 zswap_debugfs_root, &zswap_written_back_pages);
1456 debugfs_create_u64("duplicate_entry", 0444,
1457 zswap_debugfs_root, &zswap_duplicate_entry);
1458 debugfs_create_u64("pool_total_size", 0444,
1459 zswap_debugfs_root, &zswap_pool_total_size);
1460 debugfs_create_atomic_t("stored_pages", 0444,
1461 zswap_debugfs_root, &zswap_stored_pages);
1462 debugfs_create_atomic_t("same_filled_pages", 0444,
1463 zswap_debugfs_root, &zswap_same_filled_pages);
1464
1465 return 0;
1466}
1467#else
1468static int __init zswap_debugfs_init(void)
1469{
1470 return 0;
1471}
1472#endif
1473
1474/*********************************
1475* module init and exit
1476**********************************/
1477static int __init init_zswap(void)
1478{
1479 struct zswap_pool *pool;
1480 int ret;
1481
1482 zswap_init_started = true;
1483
1484 if (zswap_entry_cache_create()) {
1485 pr_err("entry cache creation failed\n");
1486 goto cache_fail;
1487 }
1488
1489 ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1490 zswap_dstmem_prepare, zswap_dstmem_dead);
1491 if (ret) {
1492 pr_err("dstmem alloc failed\n");
1493 goto dstmem_fail;
1494 }
1495
1496 ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1497 "mm/zswap_pool:prepare",
1498 zswap_cpu_comp_prepare,
1499 zswap_cpu_comp_dead);
1500 if (ret)
1501 goto hp_fail;
1502
1503 pool = __zswap_pool_create_fallback();
1504 if (pool) {
1505 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1506 zpool_get_type(pool->zpool));
1507 list_add(&pool->list, &zswap_pools);
1508 zswap_has_pool = true;
1509 } else {
1510 pr_err("pool creation failed\n");
1511 zswap_enabled = false;
1512 }
1513
1514 shrink_wq = create_workqueue("zswap-shrink");
1515 if (!shrink_wq)
1516 goto fallback_fail;
1517
1518 ret = frontswap_register_ops(&zswap_frontswap_ops);
1519 if (ret)
1520 goto destroy_wq;
1521 if (zswap_debugfs_init())
1522 pr_warn("debugfs initialization failed\n");
1523 return 0;
1524
1525destroy_wq:
1526 destroy_workqueue(shrink_wq);
1527fallback_fail:
1528 if (pool)
1529 zswap_pool_destroy(pool);
1530hp_fail:
1531 cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1532dstmem_fail:
1533 zswap_entry_cache_destroy();
1534cache_fail:
1535 /* if built-in, we aren't unloaded on failure; don't allow use */
1536 zswap_init_failed = true;
1537 zswap_enabled = false;
1538 return -ENOMEM;
1539}
1540/* must be late so crypto has time to come up */
1541late_initcall(init_zswap);
1542
1543MODULE_LICENSE("GPL");
1544MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1545MODULE_DESCRIPTION("Compressed cache for swap pages");