xref: /linux/mm/zswap.c (revision da51bbcdbace8f43adf6066934c3926b656376e5)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * zswap.c - zswap driver file
4  *
5  * zswap is a cache that takes pages that are in the process
6  * of being swapped out and attempts to compress and store them in a
7  * RAM-based memory pool.  This can result in a significant I/O reduction on
8  * the swap device and, in the case where decompressing from RAM is faster
9  * than reading from the swap device, can also improve workload performance.
10  *
11  * Copyright (C) 2012  Seth Jennings <sjenning@linux.vnet.ibm.com>
12 */
13 
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 
16 #include <linux/module.h>
17 #include <linux/cpu.h>
18 #include <linux/highmem.h>
19 #include <linux/slab.h>
20 #include <linux/spinlock.h>
21 #include <linux/types.h>
22 #include <linux/atomic.h>
23 #include <linux/rbtree.h>
24 #include <linux/swap.h>
25 #include <linux/crypto.h>
26 #include <linux/scatterlist.h>
27 #include <linux/mempolicy.h>
28 #include <linux/mempool.h>
29 #include <linux/zpool.h>
30 #include <crypto/acompress.h>
31 #include <linux/zswap.h>
32 #include <linux/mm_types.h>
33 #include <linux/page-flags.h>
34 #include <linux/swapops.h>
35 #include <linux/writeback.h>
36 #include <linux/pagemap.h>
37 #include <linux/workqueue.h>
38 #include <linux/list_lru.h>
39 
40 #include "swap.h"
41 #include "internal.h"
42 
43 /*********************************
44 * statistics
45 **********************************/
46 /* Total bytes used by the compressed storage */
47 u64 zswap_pool_total_size;
48 /* The number of compressed pages currently stored in zswap */
49 atomic_t zswap_stored_pages = ATOMIC_INIT(0);
50 /* The number of same-value filled pages currently stored in zswap */
51 static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
52 
53 /*
54  * The statistics below are not protected from concurrent access for
55  * performance reasons so they may not be a 100% accurate.  However,
56  * they do provide useful information on roughly how many times a
57  * certain event is occurring.
58 */
59 
60 /* Pool limit was hit (see zswap_max_pool_percent) */
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 /* Store failed due to compression algorithm failure */
67 static u64 zswap_reject_compress_fail;
68 /* Compressed page was too big for the allocator to (optimally) store */
69 static u64 zswap_reject_compress_poor;
70 /* Store failed because underlying allocator could not get memory */
71 static u64 zswap_reject_alloc_fail;
72 /* Store failed because the entry metadata could not be allocated (rare) */
73 static u64 zswap_reject_kmemcache_fail;
74 
75 /* Shrinker work queue */
76 static struct workqueue_struct *shrink_wq;
77 /* Pool limit was hit, we need to calm down */
78 static bool zswap_pool_reached_full;
79 
80 /*********************************
81 * tunables
82 **********************************/
83 
84 #define ZSWAP_PARAM_UNSET ""
85 
86 static int zswap_setup(void);
87 
88 /* Enable/disable zswap */
89 static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
90 static int zswap_enabled_param_set(const char *,
91 				   const struct kernel_param *);
92 static const struct kernel_param_ops zswap_enabled_param_ops = {
93 	.set =		zswap_enabled_param_set,
94 	.get =		param_get_bool,
95 };
96 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
97 
98 /* Crypto compressor to use */
99 static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
100 static int zswap_compressor_param_set(const char *,
101 				      const struct kernel_param *);
102 static const struct kernel_param_ops zswap_compressor_param_ops = {
103 	.set =		zswap_compressor_param_set,
104 	.get =		param_get_charp,
105 	.free =		param_free_charp,
106 };
107 module_param_cb(compressor, &zswap_compressor_param_ops,
108 		&zswap_compressor, 0644);
109 
110 /* Compressed storage zpool to use */
111 static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
112 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
113 static const struct kernel_param_ops zswap_zpool_param_ops = {
114 	.set =		zswap_zpool_param_set,
115 	.get =		param_get_charp,
116 	.free =		param_free_charp,
117 };
118 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
119 
120 /* The maximum percentage of memory that the compressed pool can occupy */
121 static unsigned int zswap_max_pool_percent = 20;
122 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
123 
124 /* The threshold for accepting new pages after the max_pool_percent was hit */
125 static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
126 module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
127 		   uint, 0644);
128 
129 /*
130  * Enable/disable handling same-value filled pages (enabled by default).
131  * If disabled every page is considered non-same-value filled.
132  */
133 static bool zswap_same_filled_pages_enabled = true;
134 module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
135 		   bool, 0644);
136 
137 /* Enable/disable handling non-same-value filled pages (enabled by default) */
138 static bool zswap_non_same_filled_pages_enabled = true;
139 module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
140 		   bool, 0644);
141 
142 /* Number of zpools in zswap_pool (empirically determined for scalability) */
143 #define ZSWAP_NR_ZPOOLS 32
144 
145 /* Enable/disable memory pressure-based shrinker. */
146 static bool zswap_shrinker_enabled = IS_ENABLED(
147 		CONFIG_ZSWAP_SHRINKER_DEFAULT_ON);
148 module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
149 
150 bool is_zswap_enabled(void)
151 {
152 	return zswap_enabled;
153 }
154 
155 /*********************************
156 * data structures
157 **********************************/
158 
159 struct crypto_acomp_ctx {
160 	struct crypto_acomp *acomp;
161 	struct acomp_req *req;
162 	struct crypto_wait wait;
163 	u8 *buffer;
164 	struct mutex mutex;
165 	bool is_sleepable;
166 };
167 
168 /*
169  * The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
170  * The only case where lru_lock is not acquired while holding tree.lock is
171  * when a zswap_entry is taken off the lru for writeback, in that case it
172  * needs to be verified that it's still valid in the tree.
173  */
174 struct zswap_pool {
175 	struct zpool *zpools[ZSWAP_NR_ZPOOLS];
176 	struct crypto_acomp_ctx __percpu *acomp_ctx;
177 	struct percpu_ref ref;
178 	struct list_head list;
179 	struct work_struct release_work;
180 	struct hlist_node node;
181 	char tfm_name[CRYPTO_MAX_ALG_NAME];
182 };
183 
184 /* Global LRU lists shared by all zswap pools. */
185 static struct list_lru zswap_list_lru;
186 /* counter of pages stored in all zswap pools. */
187 static atomic_t zswap_nr_stored = ATOMIC_INIT(0);
188 
189 /* The lock protects zswap_next_shrink updates. */
190 static DEFINE_SPINLOCK(zswap_shrink_lock);
191 static struct mem_cgroup *zswap_next_shrink;
192 static struct work_struct zswap_shrink_work;
193 static struct shrinker *zswap_shrinker;
194 
195 /*
196  * struct zswap_entry
197  *
198  * This structure contains the metadata for tracking a single compressed
199  * page within zswap.
200  *
201  * rbnode - links the entry into red-black tree for the appropriate swap type
202  * swpentry - associated swap entry, the offset indexes into the red-black tree
203  * length - the length in bytes of the compressed page data.  Needed during
204  *          decompression. For a same value filled page length is 0, and both
205  *          pool and lru are invalid and must be ignored.
206  * pool - the zswap_pool the entry's data is in
207  * handle - zpool allocation handle that stores the compressed page data
208  * value - value of the same-value filled pages which have same content
209  * objcg - the obj_cgroup that the compressed memory is charged to
210  * lru - handle to the pool's lru used to evict pages.
211  */
212 struct zswap_entry {
213 	struct rb_node rbnode;
214 	swp_entry_t swpentry;
215 	unsigned int length;
216 	struct zswap_pool *pool;
217 	union {
218 		unsigned long handle;
219 		unsigned long value;
220 	};
221 	struct obj_cgroup *objcg;
222 	struct list_head lru;
223 };
224 
225 struct zswap_tree {
226 	struct rb_root rbroot;
227 	spinlock_t lock;
228 };
229 
230 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
231 static unsigned int nr_zswap_trees[MAX_SWAPFILES];
232 
233 /* RCU-protected iteration */
234 static LIST_HEAD(zswap_pools);
235 /* protects zswap_pools list modification */
236 static DEFINE_SPINLOCK(zswap_pools_lock);
237 /* pool counter to provide unique names to zpool */
238 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
239 
240 enum zswap_init_type {
241 	ZSWAP_UNINIT,
242 	ZSWAP_INIT_SUCCEED,
243 	ZSWAP_INIT_FAILED
244 };
245 
246 static enum zswap_init_type zswap_init_state;
247 
248 /* used to ensure the integrity of initialization */
249 static DEFINE_MUTEX(zswap_init_lock);
250 
251 /* init completed, but couldn't create the initial pool */
252 static bool zswap_has_pool;
253 
254 /*********************************
255 * helpers and fwd declarations
256 **********************************/
257 
258 static inline struct zswap_tree *swap_zswap_tree(swp_entry_t swp)
259 {
260 	return &zswap_trees[swp_type(swp)][swp_offset(swp)
261 		>> SWAP_ADDRESS_SPACE_SHIFT];
262 }
263 
264 #define zswap_pool_debug(msg, p)				\
265 	pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,		\
266 		 zpool_get_type((p)->zpools[0]))
267 
268 static bool zswap_is_full(void)
269 {
270 	return totalram_pages() * zswap_max_pool_percent / 100 <
271 			DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
272 }
273 
274 static bool zswap_can_accept(void)
275 {
276 	return totalram_pages() * zswap_accept_thr_percent / 100 *
277 				zswap_max_pool_percent / 100 >
278 			DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
279 }
280 
281 static u64 get_zswap_pool_size(struct zswap_pool *pool)
282 {
283 	u64 pool_size = 0;
284 	int i;
285 
286 	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
287 		pool_size += zpool_get_total_size(pool->zpools[i]);
288 
289 	return pool_size;
290 }
291 
292 static void zswap_update_total_size(void)
293 {
294 	struct zswap_pool *pool;
295 	u64 total = 0;
296 
297 	rcu_read_lock();
298 
299 	list_for_each_entry_rcu(pool, &zswap_pools, list)
300 		total += get_zswap_pool_size(pool);
301 
302 	rcu_read_unlock();
303 
304 	zswap_pool_total_size = total;
305 }
306 
307 /*********************************
308 * pool functions
309 **********************************/
310 static void __zswap_pool_empty(struct percpu_ref *ref);
311 
312 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
313 {
314 	int i;
315 	struct zswap_pool *pool;
316 	char name[38]; /* 'zswap' + 32 char (max) num + \0 */
317 	gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
318 	int ret;
319 
320 	if (!zswap_has_pool) {
321 		/* if either are unset, pool initialization failed, and we
322 		 * need both params to be set correctly before trying to
323 		 * create a pool.
324 		 */
325 		if (!strcmp(type, ZSWAP_PARAM_UNSET))
326 			return NULL;
327 		if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
328 			return NULL;
329 	}
330 
331 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
332 	if (!pool)
333 		return NULL;
334 
335 	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++) {
336 		/* unique name for each pool specifically required by zsmalloc */
337 		snprintf(name, 38, "zswap%x",
338 			 atomic_inc_return(&zswap_pools_count));
339 
340 		pool->zpools[i] = zpool_create_pool(type, name, gfp);
341 		if (!pool->zpools[i]) {
342 			pr_err("%s zpool not available\n", type);
343 			goto error;
344 		}
345 	}
346 	pr_debug("using %s zpool\n", zpool_get_type(pool->zpools[0]));
347 
348 	strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
349 
350 	pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
351 	if (!pool->acomp_ctx) {
352 		pr_err("percpu alloc failed\n");
353 		goto error;
354 	}
355 
356 	ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
357 				       &pool->node);
358 	if (ret)
359 		goto error;
360 
361 	/* being the current pool takes 1 ref; this func expects the
362 	 * caller to always add the new pool as the current pool
363 	 */
364 	ret = percpu_ref_init(&pool->ref, __zswap_pool_empty,
365 			      PERCPU_REF_ALLOW_REINIT, GFP_KERNEL);
366 	if (ret)
367 		goto ref_fail;
368 	INIT_LIST_HEAD(&pool->list);
369 
370 	zswap_pool_debug("created", pool);
371 
372 	return pool;
373 
374 ref_fail:
375 	cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
376 error:
377 	if (pool->acomp_ctx)
378 		free_percpu(pool->acomp_ctx);
379 	while (i--)
380 		zpool_destroy_pool(pool->zpools[i]);
381 	kfree(pool);
382 	return NULL;
383 }
384 
385 static struct zswap_pool *__zswap_pool_create_fallback(void)
386 {
387 	bool has_comp, has_zpool;
388 
389 	has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
390 	if (!has_comp && strcmp(zswap_compressor,
391 				CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
392 		pr_err("compressor %s not available, using default %s\n",
393 		       zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
394 		param_free_charp(&zswap_compressor);
395 		zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
396 		has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
397 	}
398 	if (!has_comp) {
399 		pr_err("default compressor %s not available\n",
400 		       zswap_compressor);
401 		param_free_charp(&zswap_compressor);
402 		zswap_compressor = ZSWAP_PARAM_UNSET;
403 	}
404 
405 	has_zpool = zpool_has_pool(zswap_zpool_type);
406 	if (!has_zpool && strcmp(zswap_zpool_type,
407 				 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
408 		pr_err("zpool %s not available, using default %s\n",
409 		       zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
410 		param_free_charp(&zswap_zpool_type);
411 		zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
412 		has_zpool = zpool_has_pool(zswap_zpool_type);
413 	}
414 	if (!has_zpool) {
415 		pr_err("default zpool %s not available\n",
416 		       zswap_zpool_type);
417 		param_free_charp(&zswap_zpool_type);
418 		zswap_zpool_type = ZSWAP_PARAM_UNSET;
419 	}
420 
421 	if (!has_comp || !has_zpool)
422 		return NULL;
423 
424 	return zswap_pool_create(zswap_zpool_type, zswap_compressor);
425 }
426 
427 static void zswap_pool_destroy(struct zswap_pool *pool)
428 {
429 	int i;
430 
431 	zswap_pool_debug("destroying", pool);
432 
433 	cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
434 	free_percpu(pool->acomp_ctx);
435 
436 	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
437 		zpool_destroy_pool(pool->zpools[i]);
438 	kfree(pool);
439 }
440 
441 static void __zswap_pool_release(struct work_struct *work)
442 {
443 	struct zswap_pool *pool = container_of(work, typeof(*pool),
444 						release_work);
445 
446 	synchronize_rcu();
447 
448 	/* nobody should have been able to get a ref... */
449 	WARN_ON(!percpu_ref_is_zero(&pool->ref));
450 	percpu_ref_exit(&pool->ref);
451 
452 	/* pool is now off zswap_pools list and has no references. */
453 	zswap_pool_destroy(pool);
454 }
455 
456 static struct zswap_pool *zswap_pool_current(void);
457 
458 static void __zswap_pool_empty(struct percpu_ref *ref)
459 {
460 	struct zswap_pool *pool;
461 
462 	pool = container_of(ref, typeof(*pool), ref);
463 
464 	spin_lock_bh(&zswap_pools_lock);
465 
466 	WARN_ON(pool == zswap_pool_current());
467 
468 	list_del_rcu(&pool->list);
469 
470 	INIT_WORK(&pool->release_work, __zswap_pool_release);
471 	schedule_work(&pool->release_work);
472 
473 	spin_unlock_bh(&zswap_pools_lock);
474 }
475 
476 static int __must_check zswap_pool_get(struct zswap_pool *pool)
477 {
478 	if (!pool)
479 		return 0;
480 
481 	return percpu_ref_tryget(&pool->ref);
482 }
483 
484 static void zswap_pool_put(struct zswap_pool *pool)
485 {
486 	percpu_ref_put(&pool->ref);
487 }
488 
489 static struct zswap_pool *__zswap_pool_current(void)
490 {
491 	struct zswap_pool *pool;
492 
493 	pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
494 	WARN_ONCE(!pool && zswap_has_pool,
495 		  "%s: no page storage pool!\n", __func__);
496 
497 	return pool;
498 }
499 
500 static struct zswap_pool *zswap_pool_current(void)
501 {
502 	assert_spin_locked(&zswap_pools_lock);
503 
504 	return __zswap_pool_current();
505 }
506 
507 static struct zswap_pool *zswap_pool_current_get(void)
508 {
509 	struct zswap_pool *pool;
510 
511 	rcu_read_lock();
512 
513 	pool = __zswap_pool_current();
514 	if (!zswap_pool_get(pool))
515 		pool = NULL;
516 
517 	rcu_read_unlock();
518 
519 	return pool;
520 }
521 
522 /* type and compressor must be null-terminated */
523 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
524 {
525 	struct zswap_pool *pool;
526 
527 	assert_spin_locked(&zswap_pools_lock);
528 
529 	list_for_each_entry_rcu(pool, &zswap_pools, list) {
530 		if (strcmp(pool->tfm_name, compressor))
531 			continue;
532 		/* all zpools share the same type */
533 		if (strcmp(zpool_get_type(pool->zpools[0]), type))
534 			continue;
535 		/* if we can't get it, it's about to be destroyed */
536 		if (!zswap_pool_get(pool))
537 			continue;
538 		return pool;
539 	}
540 
541 	return NULL;
542 }
543 
544 /*********************************
545 * param callbacks
546 **********************************/
547 
548 static bool zswap_pool_changed(const char *s, const struct kernel_param *kp)
549 {
550 	/* no change required */
551 	if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
552 		return false;
553 	return true;
554 }
555 
556 /* val must be a null-terminated string */
557 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
558 			     char *type, char *compressor)
559 {
560 	struct zswap_pool *pool, *put_pool = NULL;
561 	char *s = strstrip((char *)val);
562 	int ret = 0;
563 	bool new_pool = false;
564 
565 	mutex_lock(&zswap_init_lock);
566 	switch (zswap_init_state) {
567 	case ZSWAP_UNINIT:
568 		/* if this is load-time (pre-init) param setting,
569 		 * don't create a pool; that's done during init.
570 		 */
571 		ret = param_set_charp(s, kp);
572 		break;
573 	case ZSWAP_INIT_SUCCEED:
574 		new_pool = zswap_pool_changed(s, kp);
575 		break;
576 	case ZSWAP_INIT_FAILED:
577 		pr_err("can't set param, initialization failed\n");
578 		ret = -ENODEV;
579 	}
580 	mutex_unlock(&zswap_init_lock);
581 
582 	/* no need to create a new pool, return directly */
583 	if (!new_pool)
584 		return ret;
585 
586 	if (!type) {
587 		if (!zpool_has_pool(s)) {
588 			pr_err("zpool %s not available\n", s);
589 			return -ENOENT;
590 		}
591 		type = s;
592 	} else if (!compressor) {
593 		if (!crypto_has_acomp(s, 0, 0)) {
594 			pr_err("compressor %s not available\n", s);
595 			return -ENOENT;
596 		}
597 		compressor = s;
598 	} else {
599 		WARN_ON(1);
600 		return -EINVAL;
601 	}
602 
603 	spin_lock_bh(&zswap_pools_lock);
604 
605 	pool = zswap_pool_find_get(type, compressor);
606 	if (pool) {
607 		zswap_pool_debug("using existing", pool);
608 		WARN_ON(pool == zswap_pool_current());
609 		list_del_rcu(&pool->list);
610 	}
611 
612 	spin_unlock_bh(&zswap_pools_lock);
613 
614 	if (!pool)
615 		pool = zswap_pool_create(type, compressor);
616 	else {
617 		/*
618 		 * Restore the initial ref dropped by percpu_ref_kill()
619 		 * when the pool was decommissioned and switch it again
620 		 * to percpu mode.
621 		 */
622 		percpu_ref_resurrect(&pool->ref);
623 
624 		/* Drop the ref from zswap_pool_find_get(). */
625 		zswap_pool_put(pool);
626 	}
627 
628 	if (pool)
629 		ret = param_set_charp(s, kp);
630 	else
631 		ret = -EINVAL;
632 
633 	spin_lock_bh(&zswap_pools_lock);
634 
635 	if (!ret) {
636 		put_pool = zswap_pool_current();
637 		list_add_rcu(&pool->list, &zswap_pools);
638 		zswap_has_pool = true;
639 	} else if (pool) {
640 		/* add the possibly pre-existing pool to the end of the pools
641 		 * list; if it's new (and empty) then it'll be removed and
642 		 * destroyed by the put after we drop the lock
643 		 */
644 		list_add_tail_rcu(&pool->list, &zswap_pools);
645 		put_pool = pool;
646 	}
647 
648 	spin_unlock_bh(&zswap_pools_lock);
649 
650 	if (!zswap_has_pool && !pool) {
651 		/* if initial pool creation failed, and this pool creation also
652 		 * failed, maybe both compressor and zpool params were bad.
653 		 * Allow changing this param, so pool creation will succeed
654 		 * when the other param is changed. We already verified this
655 		 * param is ok in the zpool_has_pool() or crypto_has_acomp()
656 		 * checks above.
657 		 */
658 		ret = param_set_charp(s, kp);
659 	}
660 
661 	/* drop the ref from either the old current pool,
662 	 * or the new pool we failed to add
663 	 */
664 	if (put_pool)
665 		percpu_ref_kill(&put_pool->ref);
666 
667 	return ret;
668 }
669 
670 static int zswap_compressor_param_set(const char *val,
671 				      const struct kernel_param *kp)
672 {
673 	return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
674 }
675 
676 static int zswap_zpool_param_set(const char *val,
677 				 const struct kernel_param *kp)
678 {
679 	return __zswap_param_set(val, kp, NULL, zswap_compressor);
680 }
681 
682 static int zswap_enabled_param_set(const char *val,
683 				   const struct kernel_param *kp)
684 {
685 	int ret = -ENODEV;
686 
687 	/* if this is load-time (pre-init) param setting, only set param. */
688 	if (system_state != SYSTEM_RUNNING)
689 		return param_set_bool(val, kp);
690 
691 	mutex_lock(&zswap_init_lock);
692 	switch (zswap_init_state) {
693 	case ZSWAP_UNINIT:
694 		if (zswap_setup())
695 			break;
696 		fallthrough;
697 	case ZSWAP_INIT_SUCCEED:
698 		if (!zswap_has_pool)
699 			pr_err("can't enable, no pool configured\n");
700 		else
701 			ret = param_set_bool(val, kp);
702 		break;
703 	case ZSWAP_INIT_FAILED:
704 		pr_err("can't enable, initialization failed\n");
705 	}
706 	mutex_unlock(&zswap_init_lock);
707 
708 	return ret;
709 }
710 
711 /*********************************
712 * lru functions
713 **********************************/
714 
715 /* should be called under RCU */
716 #ifdef CONFIG_MEMCG
717 static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
718 {
719 	return entry->objcg ? obj_cgroup_memcg(entry->objcg) : NULL;
720 }
721 #else
722 static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
723 {
724 	return NULL;
725 }
726 #endif
727 
728 static inline int entry_to_nid(struct zswap_entry *entry)
729 {
730 	return page_to_nid(virt_to_page(entry));
731 }
732 
733 static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
734 {
735 	atomic_long_t *nr_zswap_protected;
736 	unsigned long lru_size, old, new;
737 	int nid = entry_to_nid(entry);
738 	struct mem_cgroup *memcg;
739 	struct lruvec *lruvec;
740 
741 	/*
742 	 * Note that it is safe to use rcu_read_lock() here, even in the face of
743 	 * concurrent memcg offlining. Thanks to the memcg->kmemcg_id indirection
744 	 * used in list_lru lookup, only two scenarios are possible:
745 	 *
746 	 * 1. list_lru_add() is called before memcg->kmemcg_id is updated. The
747 	 *    new entry will be reparented to memcg's parent's list_lru.
748 	 * 2. list_lru_add() is called after memcg->kmemcg_id is updated. The
749 	 *    new entry will be added directly to memcg's parent's list_lru.
750 	 *
751 	 * Similar reasoning holds for list_lru_del().
752 	 */
753 	rcu_read_lock();
754 	memcg = mem_cgroup_from_entry(entry);
755 	/* will always succeed */
756 	list_lru_add(list_lru, &entry->lru, nid, memcg);
757 
758 	/* Update the protection area */
759 	lru_size = list_lru_count_one(list_lru, nid, memcg);
760 	lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
761 	nr_zswap_protected = &lruvec->zswap_lruvec_state.nr_zswap_protected;
762 	old = atomic_long_inc_return(nr_zswap_protected);
763 	/*
764 	 * Decay to avoid overflow and adapt to changing workloads.
765 	 * This is based on LRU reclaim cost decaying heuristics.
766 	 */
767 	do {
768 		new = old > lru_size / 4 ? old / 2 : old;
769 	} while (!atomic_long_try_cmpxchg(nr_zswap_protected, &old, new));
770 	rcu_read_unlock();
771 }
772 
773 static void zswap_lru_del(struct list_lru *list_lru, struct zswap_entry *entry)
774 {
775 	int nid = entry_to_nid(entry);
776 	struct mem_cgroup *memcg;
777 
778 	rcu_read_lock();
779 	memcg = mem_cgroup_from_entry(entry);
780 	/* will always succeed */
781 	list_lru_del(list_lru, &entry->lru, nid, memcg);
782 	rcu_read_unlock();
783 }
784 
785 void zswap_lruvec_state_init(struct lruvec *lruvec)
786 {
787 	atomic_long_set(&lruvec->zswap_lruvec_state.nr_zswap_protected, 0);
788 }
789 
790 void zswap_folio_swapin(struct folio *folio)
791 {
792 	struct lruvec *lruvec;
793 
794 	if (folio) {
795 		lruvec = folio_lruvec(folio);
796 		atomic_long_inc(&lruvec->zswap_lruvec_state.nr_zswap_protected);
797 	}
798 }
799 
800 void zswap_memcg_offline_cleanup(struct mem_cgroup *memcg)
801 {
802 	/* lock out zswap shrinker walking memcg tree */
803 	spin_lock(&zswap_shrink_lock);
804 	if (zswap_next_shrink == memcg)
805 		zswap_next_shrink = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
806 	spin_unlock(&zswap_shrink_lock);
807 }
808 
809 /*********************************
810 * rbtree functions
811 **********************************/
812 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
813 {
814 	struct rb_node *node = root->rb_node;
815 	struct zswap_entry *entry;
816 	pgoff_t entry_offset;
817 
818 	while (node) {
819 		entry = rb_entry(node, struct zswap_entry, rbnode);
820 		entry_offset = swp_offset(entry->swpentry);
821 		if (entry_offset > offset)
822 			node = node->rb_left;
823 		else if (entry_offset < offset)
824 			node = node->rb_right;
825 		else
826 			return entry;
827 	}
828 	return NULL;
829 }
830 
831 /*
832  * In the case that a entry with the same offset is found, a pointer to
833  * the existing entry is stored in dupentry and the function returns -EEXIST
834  */
835 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
836 			struct zswap_entry **dupentry)
837 {
838 	struct rb_node **link = &root->rb_node, *parent = NULL;
839 	struct zswap_entry *myentry;
840 	pgoff_t myentry_offset, entry_offset = swp_offset(entry->swpentry);
841 
842 	while (*link) {
843 		parent = *link;
844 		myentry = rb_entry(parent, struct zswap_entry, rbnode);
845 		myentry_offset = swp_offset(myentry->swpentry);
846 		if (myentry_offset > entry_offset)
847 			link = &(*link)->rb_left;
848 		else if (myentry_offset < entry_offset)
849 			link = &(*link)->rb_right;
850 		else {
851 			*dupentry = myentry;
852 			return -EEXIST;
853 		}
854 	}
855 	rb_link_node(&entry->rbnode, parent, link);
856 	rb_insert_color(&entry->rbnode, root);
857 	return 0;
858 }
859 
860 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
861 {
862 	rb_erase(&entry->rbnode, root);
863 	RB_CLEAR_NODE(&entry->rbnode);
864 }
865 
866 /*********************************
867 * zswap entry functions
868 **********************************/
869 static struct kmem_cache *zswap_entry_cache;
870 
871 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp, int nid)
872 {
873 	struct zswap_entry *entry;
874 	entry = kmem_cache_alloc_node(zswap_entry_cache, gfp, nid);
875 	if (!entry)
876 		return NULL;
877 	RB_CLEAR_NODE(&entry->rbnode);
878 	return entry;
879 }
880 
881 static void zswap_entry_cache_free(struct zswap_entry *entry)
882 {
883 	kmem_cache_free(zswap_entry_cache, entry);
884 }
885 
886 static struct zpool *zswap_find_zpool(struct zswap_entry *entry)
887 {
888 	int i = 0;
889 
890 	if (ZSWAP_NR_ZPOOLS > 1)
891 		i = hash_ptr(entry, ilog2(ZSWAP_NR_ZPOOLS));
892 
893 	return entry->pool->zpools[i];
894 }
895 
896 /*
897  * Carries out the common pattern of freeing and entry's zpool allocation,
898  * freeing the entry itself, and decrementing the number of stored pages.
899  */
900 static void zswap_entry_free(struct zswap_entry *entry)
901 {
902 	if (!entry->length)
903 		atomic_dec(&zswap_same_filled_pages);
904 	else {
905 		zswap_lru_del(&zswap_list_lru, entry);
906 		zpool_free(zswap_find_zpool(entry), entry->handle);
907 		atomic_dec(&zswap_nr_stored);
908 		zswap_pool_put(entry->pool);
909 	}
910 	if (entry->objcg) {
911 		obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
912 		obj_cgroup_put(entry->objcg);
913 	}
914 	zswap_entry_cache_free(entry);
915 	atomic_dec(&zswap_stored_pages);
916 	zswap_update_total_size();
917 }
918 
919 /*
920  * The caller hold the tree lock and search the entry from the tree,
921  * so it must be on the tree, remove it from the tree and free it.
922  */
923 static void zswap_invalidate_entry(struct zswap_tree *tree,
924 				   struct zswap_entry *entry)
925 {
926 	zswap_rb_erase(&tree->rbroot, entry);
927 	zswap_entry_free(entry);
928 }
929 
930 /*********************************
931 * compressed storage functions
932 **********************************/
933 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
934 {
935 	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
936 	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
937 	struct crypto_acomp *acomp;
938 	struct acomp_req *req;
939 	int ret;
940 
941 	mutex_init(&acomp_ctx->mutex);
942 
943 	acomp_ctx->buffer = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
944 	if (!acomp_ctx->buffer)
945 		return -ENOMEM;
946 
947 	acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
948 	if (IS_ERR(acomp)) {
949 		pr_err("could not alloc crypto acomp %s : %ld\n",
950 				pool->tfm_name, PTR_ERR(acomp));
951 		ret = PTR_ERR(acomp);
952 		goto acomp_fail;
953 	}
954 	acomp_ctx->acomp = acomp;
955 	acomp_ctx->is_sleepable = acomp_is_async(acomp);
956 
957 	req = acomp_request_alloc(acomp_ctx->acomp);
958 	if (!req) {
959 		pr_err("could not alloc crypto acomp_request %s\n",
960 		       pool->tfm_name);
961 		ret = -ENOMEM;
962 		goto req_fail;
963 	}
964 	acomp_ctx->req = req;
965 
966 	crypto_init_wait(&acomp_ctx->wait);
967 	/*
968 	 * if the backend of acomp is async zip, crypto_req_done() will wakeup
969 	 * crypto_wait_req(); if the backend of acomp is scomp, the callback
970 	 * won't be called, crypto_wait_req() will return without blocking.
971 	 */
972 	acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
973 				   crypto_req_done, &acomp_ctx->wait);
974 
975 	return 0;
976 
977 req_fail:
978 	crypto_free_acomp(acomp_ctx->acomp);
979 acomp_fail:
980 	kfree(acomp_ctx->buffer);
981 	return ret;
982 }
983 
984 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
985 {
986 	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
987 	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
988 
989 	if (!IS_ERR_OR_NULL(acomp_ctx)) {
990 		if (!IS_ERR_OR_NULL(acomp_ctx->req))
991 			acomp_request_free(acomp_ctx->req);
992 		if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
993 			crypto_free_acomp(acomp_ctx->acomp);
994 		kfree(acomp_ctx->buffer);
995 	}
996 
997 	return 0;
998 }
999 
1000 static bool zswap_compress(struct folio *folio, struct zswap_entry *entry)
1001 {
1002 	struct crypto_acomp_ctx *acomp_ctx;
1003 	struct scatterlist input, output;
1004 	int comp_ret = 0, alloc_ret = 0;
1005 	unsigned int dlen = PAGE_SIZE;
1006 	unsigned long handle;
1007 	struct zpool *zpool;
1008 	char *buf;
1009 	gfp_t gfp;
1010 	u8 *dst;
1011 
1012 	acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1013 
1014 	mutex_lock(&acomp_ctx->mutex);
1015 
1016 	dst = acomp_ctx->buffer;
1017 	sg_init_table(&input, 1);
1018 	sg_set_page(&input, &folio->page, PAGE_SIZE, 0);
1019 
1020 	/*
1021 	 * We need PAGE_SIZE * 2 here since there maybe over-compression case,
1022 	 * and hardware-accelerators may won't check the dst buffer size, so
1023 	 * giving the dst buffer with enough length to avoid buffer overflow.
1024 	 */
1025 	sg_init_one(&output, dst, PAGE_SIZE * 2);
1026 	acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1027 
1028 	/*
1029 	 * it maybe looks a little bit silly that we send an asynchronous request,
1030 	 * then wait for its completion synchronously. This makes the process look
1031 	 * synchronous in fact.
1032 	 * Theoretically, acomp supports users send multiple acomp requests in one
1033 	 * acomp instance, then get those requests done simultaneously. but in this
1034 	 * case, zswap actually does store and load page by page, there is no
1035 	 * existing method to send the second page before the first page is done
1036 	 * in one thread doing zwap.
1037 	 * but in different threads running on different cpu, we have different
1038 	 * acomp instance, so multiple threads can do (de)compression in parallel.
1039 	 */
1040 	comp_ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1041 	dlen = acomp_ctx->req->dlen;
1042 	if (comp_ret)
1043 		goto unlock;
1044 
1045 	zpool = zswap_find_zpool(entry);
1046 	gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1047 	if (zpool_malloc_support_movable(zpool))
1048 		gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1049 	alloc_ret = zpool_malloc(zpool, dlen, gfp, &handle);
1050 	if (alloc_ret)
1051 		goto unlock;
1052 
1053 	buf = zpool_map_handle(zpool, handle, ZPOOL_MM_WO);
1054 	memcpy(buf, dst, dlen);
1055 	zpool_unmap_handle(zpool, handle);
1056 
1057 	entry->handle = handle;
1058 	entry->length = dlen;
1059 
1060 unlock:
1061 	if (comp_ret == -ENOSPC || alloc_ret == -ENOSPC)
1062 		zswap_reject_compress_poor++;
1063 	else if (comp_ret)
1064 		zswap_reject_compress_fail++;
1065 	else if (alloc_ret)
1066 		zswap_reject_alloc_fail++;
1067 
1068 	mutex_unlock(&acomp_ctx->mutex);
1069 	return comp_ret == 0 && alloc_ret == 0;
1070 }
1071 
1072 static void zswap_decompress(struct zswap_entry *entry, struct page *page)
1073 {
1074 	struct zpool *zpool = zswap_find_zpool(entry);
1075 	struct scatterlist input, output;
1076 	struct crypto_acomp_ctx *acomp_ctx;
1077 	u8 *src;
1078 
1079 	acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1080 	mutex_lock(&acomp_ctx->mutex);
1081 
1082 	src = zpool_map_handle(zpool, entry->handle, ZPOOL_MM_RO);
1083 	if (acomp_ctx->is_sleepable && !zpool_can_sleep_mapped(zpool)) {
1084 		memcpy(acomp_ctx->buffer, src, entry->length);
1085 		src = acomp_ctx->buffer;
1086 		zpool_unmap_handle(zpool, entry->handle);
1087 	}
1088 
1089 	sg_init_one(&input, src, entry->length);
1090 	sg_init_table(&output, 1);
1091 	sg_set_page(&output, page, PAGE_SIZE, 0);
1092 	acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
1093 	BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
1094 	BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
1095 	mutex_unlock(&acomp_ctx->mutex);
1096 
1097 	if (!acomp_ctx->is_sleepable || zpool_can_sleep_mapped(zpool))
1098 		zpool_unmap_handle(zpool, entry->handle);
1099 }
1100 
1101 /*********************************
1102 * writeback code
1103 **********************************/
1104 /*
1105  * Attempts to free an entry by adding a folio to the swap cache,
1106  * decompressing the entry data into the folio, and issuing a
1107  * bio write to write the folio back to the swap device.
1108  *
1109  * This can be thought of as a "resumed writeback" of the folio
1110  * to the swap device.  We are basically resuming the same swap
1111  * writeback path that was intercepted with the zswap_store()
1112  * in the first place.  After the folio has been decompressed into
1113  * the swap cache, the compressed version stored by zswap can be
1114  * freed.
1115  */
1116 static int zswap_writeback_entry(struct zswap_entry *entry,
1117 				 swp_entry_t swpentry)
1118 {
1119 	struct zswap_tree *tree;
1120 	struct folio *folio;
1121 	struct mempolicy *mpol;
1122 	bool folio_was_allocated;
1123 	struct writeback_control wbc = {
1124 		.sync_mode = WB_SYNC_NONE,
1125 	};
1126 
1127 	/* try to allocate swap cache folio */
1128 	mpol = get_task_policy(current);
1129 	folio = __read_swap_cache_async(swpentry, GFP_KERNEL, mpol,
1130 				NO_INTERLEAVE_INDEX, &folio_was_allocated, true);
1131 	if (!folio)
1132 		return -ENOMEM;
1133 
1134 	/*
1135 	 * Found an existing folio, we raced with swapin or concurrent
1136 	 * shrinker. We generally writeback cold folios from zswap, and
1137 	 * swapin means the folio just became hot, so skip this folio.
1138 	 * For unlikely concurrent shrinker case, it will be unlinked
1139 	 * and freed when invalidated by the concurrent shrinker anyway.
1140 	 */
1141 	if (!folio_was_allocated) {
1142 		folio_put(folio);
1143 		return -EEXIST;
1144 	}
1145 
1146 	/*
1147 	 * folio is locked, and the swapcache is now secured against
1148 	 * concurrent swapping to and from the slot, and concurrent
1149 	 * swapoff so we can safely dereference the zswap tree here.
1150 	 * Verify that the swap entry hasn't been invalidated and recycled
1151 	 * behind our backs, to avoid overwriting a new swap folio with
1152 	 * old compressed data. Only when this is successful can the entry
1153 	 * be dereferenced.
1154 	 */
1155 	tree = swap_zswap_tree(swpentry);
1156 	spin_lock(&tree->lock);
1157 	if (zswap_rb_search(&tree->rbroot, swp_offset(swpentry)) != entry) {
1158 		spin_unlock(&tree->lock);
1159 		delete_from_swap_cache(folio);
1160 		folio_unlock(folio);
1161 		folio_put(folio);
1162 		return -ENOMEM;
1163 	}
1164 
1165 	/* Safe to deref entry after the entry is verified above. */
1166 	zswap_rb_erase(&tree->rbroot, entry);
1167 	spin_unlock(&tree->lock);
1168 
1169 	zswap_decompress(entry, &folio->page);
1170 
1171 	count_vm_event(ZSWPWB);
1172 	if (entry->objcg)
1173 		count_objcg_event(entry->objcg, ZSWPWB);
1174 
1175 	zswap_entry_free(entry);
1176 
1177 	/* folio is up to date */
1178 	folio_mark_uptodate(folio);
1179 
1180 	/* move it to the tail of the inactive list after end_writeback */
1181 	folio_set_reclaim(folio);
1182 
1183 	/* start writeback */
1184 	__swap_writepage(folio, &wbc);
1185 	folio_put(folio);
1186 
1187 	return 0;
1188 }
1189 
1190 /*********************************
1191 * shrinker functions
1192 **********************************/
1193 static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
1194 				       spinlock_t *lock, void *arg)
1195 {
1196 	struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
1197 	bool *encountered_page_in_swapcache = (bool *)arg;
1198 	swp_entry_t swpentry;
1199 	enum lru_status ret = LRU_REMOVED_RETRY;
1200 	int writeback_result;
1201 
1202 	/*
1203 	 * As soon as we drop the LRU lock, the entry can be freed by
1204 	 * a concurrent invalidation. This means the following:
1205 	 *
1206 	 * 1. We extract the swp_entry_t to the stack, allowing
1207 	 *    zswap_writeback_entry() to pin the swap entry and
1208 	 *    then validate the zwap entry against that swap entry's
1209 	 *    tree using pointer value comparison. Only when that
1210 	 *    is successful can the entry be dereferenced.
1211 	 *
1212 	 * 2. Usually, objects are taken off the LRU for reclaim. In
1213 	 *    this case this isn't possible, because if reclaim fails
1214 	 *    for whatever reason, we have no means of knowing if the
1215 	 *    entry is alive to put it back on the LRU.
1216 	 *
1217 	 *    So rotate it before dropping the lock. If the entry is
1218 	 *    written back or invalidated, the free path will unlink
1219 	 *    it. For failures, rotation is the right thing as well.
1220 	 *
1221 	 *    Temporary failures, where the same entry should be tried
1222 	 *    again immediately, almost never happen for this shrinker.
1223 	 *    We don't do any trylocking; -ENOMEM comes closest,
1224 	 *    but that's extremely rare and doesn't happen spuriously
1225 	 *    either. Don't bother distinguishing this case.
1226 	 */
1227 	list_move_tail(item, &l->list);
1228 
1229 	/*
1230 	 * Once the lru lock is dropped, the entry might get freed. The
1231 	 * swpentry is copied to the stack, and entry isn't deref'd again
1232 	 * until the entry is verified to still be alive in the tree.
1233 	 */
1234 	swpentry = entry->swpentry;
1235 
1236 	/*
1237 	 * It's safe to drop the lock here because we return either
1238 	 * LRU_REMOVED_RETRY or LRU_RETRY.
1239 	 */
1240 	spin_unlock(lock);
1241 
1242 	writeback_result = zswap_writeback_entry(entry, swpentry);
1243 
1244 	if (writeback_result) {
1245 		zswap_reject_reclaim_fail++;
1246 		ret = LRU_RETRY;
1247 
1248 		/*
1249 		 * Encountering a page already in swap cache is a sign that we are shrinking
1250 		 * into the warmer region. We should terminate shrinking (if we're in the dynamic
1251 		 * shrinker context).
1252 		 */
1253 		if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
1254 			ret = LRU_STOP;
1255 			*encountered_page_in_swapcache = true;
1256 		}
1257 	} else {
1258 		zswap_written_back_pages++;
1259 	}
1260 
1261 	spin_lock(lock);
1262 	return ret;
1263 }
1264 
1265 static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
1266 		struct shrink_control *sc)
1267 {
1268 	struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
1269 	unsigned long shrink_ret, nr_protected, lru_size;
1270 	bool encountered_page_in_swapcache = false;
1271 
1272 	if (!zswap_shrinker_enabled ||
1273 			!mem_cgroup_zswap_writeback_enabled(sc->memcg)) {
1274 		sc->nr_scanned = 0;
1275 		return SHRINK_STOP;
1276 	}
1277 
1278 	nr_protected =
1279 		atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
1280 	lru_size = list_lru_shrink_count(&zswap_list_lru, sc);
1281 
1282 	/*
1283 	 * Abort if we are shrinking into the protected region.
1284 	 *
1285 	 * This short-circuiting is necessary because if we have too many multiple
1286 	 * concurrent reclaimers getting the freeable zswap object counts at the
1287 	 * same time (before any of them made reasonable progress), the total
1288 	 * number of reclaimed objects might be more than the number of unprotected
1289 	 * objects (i.e the reclaimers will reclaim into the protected area of the
1290 	 * zswap LRU).
1291 	 */
1292 	if (nr_protected >= lru_size - sc->nr_to_scan) {
1293 		sc->nr_scanned = 0;
1294 		return SHRINK_STOP;
1295 	}
1296 
1297 	shrink_ret = list_lru_shrink_walk(&zswap_list_lru, sc, &shrink_memcg_cb,
1298 		&encountered_page_in_swapcache);
1299 
1300 	if (encountered_page_in_swapcache)
1301 		return SHRINK_STOP;
1302 
1303 	return shrink_ret ? shrink_ret : SHRINK_STOP;
1304 }
1305 
1306 static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
1307 		struct shrink_control *sc)
1308 {
1309 	struct mem_cgroup *memcg = sc->memcg;
1310 	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
1311 	unsigned long nr_backing, nr_stored, nr_freeable, nr_protected;
1312 
1313 	if (!zswap_shrinker_enabled || !mem_cgroup_zswap_writeback_enabled(memcg))
1314 		return 0;
1315 
1316 #ifdef CONFIG_MEMCG_KMEM
1317 	mem_cgroup_flush_stats(memcg);
1318 	nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
1319 	nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
1320 #else
1321 	/* use pool stats instead of memcg stats */
1322 	nr_backing = zswap_pool_total_size >> PAGE_SHIFT;
1323 	nr_stored = atomic_read(&zswap_nr_stored);
1324 #endif
1325 
1326 	if (!nr_stored)
1327 		return 0;
1328 
1329 	nr_protected =
1330 		atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
1331 	nr_freeable = list_lru_shrink_count(&zswap_list_lru, sc);
1332 	/*
1333 	 * Subtract the lru size by an estimate of the number of pages
1334 	 * that should be protected.
1335 	 */
1336 	nr_freeable = nr_freeable > nr_protected ? nr_freeable - nr_protected : 0;
1337 
1338 	/*
1339 	 * Scale the number of freeable pages by the memory saving factor.
1340 	 * This ensures that the better zswap compresses memory, the fewer
1341 	 * pages we will evict to swap (as it will otherwise incur IO for
1342 	 * relatively small memory saving).
1343 	 */
1344 	return mult_frac(nr_freeable, nr_backing, nr_stored);
1345 }
1346 
1347 static struct shrinker *zswap_alloc_shrinker(void)
1348 {
1349 	struct shrinker *shrinker;
1350 
1351 	shrinker =
1352 		shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
1353 	if (!shrinker)
1354 		return NULL;
1355 
1356 	shrinker->scan_objects = zswap_shrinker_scan;
1357 	shrinker->count_objects = zswap_shrinker_count;
1358 	shrinker->batch = 0;
1359 	shrinker->seeks = DEFAULT_SEEKS;
1360 	return shrinker;
1361 }
1362 
1363 static int shrink_memcg(struct mem_cgroup *memcg)
1364 {
1365 	int nid, shrunk = 0;
1366 
1367 	if (!mem_cgroup_zswap_writeback_enabled(memcg))
1368 		return -EINVAL;
1369 
1370 	/*
1371 	 * Skip zombies because their LRUs are reparented and we would be
1372 	 * reclaiming from the parent instead of the dead memcg.
1373 	 */
1374 	if (memcg && !mem_cgroup_online(memcg))
1375 		return -ENOENT;
1376 
1377 	for_each_node_state(nid, N_NORMAL_MEMORY) {
1378 		unsigned long nr_to_walk = 1;
1379 
1380 		shrunk += list_lru_walk_one(&zswap_list_lru, nid, memcg,
1381 					    &shrink_memcg_cb, NULL, &nr_to_walk);
1382 	}
1383 	return shrunk ? 0 : -EAGAIN;
1384 }
1385 
1386 static void shrink_worker(struct work_struct *w)
1387 {
1388 	struct mem_cgroup *memcg;
1389 	int ret, failures = 0;
1390 
1391 	/* global reclaim will select cgroup in a round-robin fashion. */
1392 	do {
1393 		spin_lock(&zswap_shrink_lock);
1394 		zswap_next_shrink = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
1395 		memcg = zswap_next_shrink;
1396 
1397 		/*
1398 		 * We need to retry if we have gone through a full round trip, or if we
1399 		 * got an offline memcg (or else we risk undoing the effect of the
1400 		 * zswap memcg offlining cleanup callback). This is not catastrophic
1401 		 * per se, but it will keep the now offlined memcg hostage for a while.
1402 		 *
1403 		 * Note that if we got an online memcg, we will keep the extra
1404 		 * reference in case the original reference obtained by mem_cgroup_iter
1405 		 * is dropped by the zswap memcg offlining callback, ensuring that the
1406 		 * memcg is not killed when we are reclaiming.
1407 		 */
1408 		if (!memcg) {
1409 			spin_unlock(&zswap_shrink_lock);
1410 			if (++failures == MAX_RECLAIM_RETRIES)
1411 				break;
1412 
1413 			goto resched;
1414 		}
1415 
1416 		if (!mem_cgroup_tryget_online(memcg)) {
1417 			/* drop the reference from mem_cgroup_iter() */
1418 			mem_cgroup_iter_break(NULL, memcg);
1419 			zswap_next_shrink = NULL;
1420 			spin_unlock(&zswap_shrink_lock);
1421 
1422 			if (++failures == MAX_RECLAIM_RETRIES)
1423 				break;
1424 
1425 			goto resched;
1426 		}
1427 		spin_unlock(&zswap_shrink_lock);
1428 
1429 		ret = shrink_memcg(memcg);
1430 		/* drop the extra reference */
1431 		mem_cgroup_put(memcg);
1432 
1433 		if (ret == -EINVAL)
1434 			break;
1435 		if (ret && ++failures == MAX_RECLAIM_RETRIES)
1436 			break;
1437 
1438 resched:
1439 		cond_resched();
1440 	} while (!zswap_can_accept());
1441 }
1442 
1443 static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1444 {
1445 	unsigned long *page;
1446 	unsigned long val;
1447 	unsigned int pos, last_pos = PAGE_SIZE / sizeof(*page) - 1;
1448 
1449 	page = (unsigned long *)ptr;
1450 	val = page[0];
1451 
1452 	if (val != page[last_pos])
1453 		return 0;
1454 
1455 	for (pos = 1; pos < last_pos; pos++) {
1456 		if (val != page[pos])
1457 			return 0;
1458 	}
1459 
1460 	*value = val;
1461 
1462 	return 1;
1463 }
1464 
1465 static void zswap_fill_page(void *ptr, unsigned long value)
1466 {
1467 	unsigned long *page;
1468 
1469 	page = (unsigned long *)ptr;
1470 	memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1471 }
1472 
1473 bool zswap_store(struct folio *folio)
1474 {
1475 	swp_entry_t swp = folio->swap;
1476 	pgoff_t offset = swp_offset(swp);
1477 	struct zswap_tree *tree = swap_zswap_tree(swp);
1478 	struct zswap_entry *entry, *dupentry;
1479 	struct obj_cgroup *objcg = NULL;
1480 	struct mem_cgroup *memcg = NULL;
1481 
1482 	VM_WARN_ON_ONCE(!folio_test_locked(folio));
1483 	VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
1484 
1485 	/* Large folios aren't supported */
1486 	if (folio_test_large(folio))
1487 		return false;
1488 
1489 	if (!zswap_enabled)
1490 		goto check_old;
1491 
1492 	objcg = get_obj_cgroup_from_folio(folio);
1493 	if (objcg && !obj_cgroup_may_zswap(objcg)) {
1494 		memcg = get_mem_cgroup_from_objcg(objcg);
1495 		if (shrink_memcg(memcg)) {
1496 			mem_cgroup_put(memcg);
1497 			goto reject;
1498 		}
1499 		mem_cgroup_put(memcg);
1500 	}
1501 
1502 	/* reclaim space if needed */
1503 	if (zswap_is_full()) {
1504 		zswap_pool_limit_hit++;
1505 		zswap_pool_reached_full = true;
1506 		goto shrink;
1507 	}
1508 
1509 	if (zswap_pool_reached_full) {
1510 	       if (!zswap_can_accept())
1511 			goto shrink;
1512 		else
1513 			zswap_pool_reached_full = false;
1514 	}
1515 
1516 	/* allocate entry */
1517 	entry = zswap_entry_cache_alloc(GFP_KERNEL, folio_nid(folio));
1518 	if (!entry) {
1519 		zswap_reject_kmemcache_fail++;
1520 		goto reject;
1521 	}
1522 
1523 	if (zswap_same_filled_pages_enabled) {
1524 		unsigned long value;
1525 		u8 *src;
1526 
1527 		src = kmap_local_folio(folio, 0);
1528 		if (zswap_is_page_same_filled(src, &value)) {
1529 			kunmap_local(src);
1530 			entry->length = 0;
1531 			entry->value = value;
1532 			atomic_inc(&zswap_same_filled_pages);
1533 			goto insert_entry;
1534 		}
1535 		kunmap_local(src);
1536 	}
1537 
1538 	if (!zswap_non_same_filled_pages_enabled)
1539 		goto freepage;
1540 
1541 	/* if entry is successfully added, it keeps the reference */
1542 	entry->pool = zswap_pool_current_get();
1543 	if (!entry->pool)
1544 		goto freepage;
1545 
1546 	if (objcg) {
1547 		memcg = get_mem_cgroup_from_objcg(objcg);
1548 		if (memcg_list_lru_alloc(memcg, &zswap_list_lru, GFP_KERNEL)) {
1549 			mem_cgroup_put(memcg);
1550 			goto put_pool;
1551 		}
1552 		mem_cgroup_put(memcg);
1553 	}
1554 
1555 	if (!zswap_compress(folio, entry))
1556 		goto put_pool;
1557 
1558 insert_entry:
1559 	entry->swpentry = swp;
1560 	entry->objcg = objcg;
1561 	if (objcg) {
1562 		obj_cgroup_charge_zswap(objcg, entry->length);
1563 		/* Account before objcg ref is moved to tree */
1564 		count_objcg_event(objcg, ZSWPOUT);
1565 	}
1566 
1567 	/* map */
1568 	spin_lock(&tree->lock);
1569 	/*
1570 	 * The folio may have been dirtied again, invalidate the
1571 	 * possibly stale entry before inserting the new entry.
1572 	 */
1573 	if (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST) {
1574 		zswap_invalidate_entry(tree, dupentry);
1575 		WARN_ON(zswap_rb_insert(&tree->rbroot, entry, &dupentry));
1576 	}
1577 	if (entry->length) {
1578 		INIT_LIST_HEAD(&entry->lru);
1579 		zswap_lru_add(&zswap_list_lru, entry);
1580 		atomic_inc(&zswap_nr_stored);
1581 	}
1582 	spin_unlock(&tree->lock);
1583 
1584 	/* update stats */
1585 	atomic_inc(&zswap_stored_pages);
1586 	zswap_update_total_size();
1587 	count_vm_event(ZSWPOUT);
1588 
1589 	return true;
1590 
1591 put_pool:
1592 	zswap_pool_put(entry->pool);
1593 freepage:
1594 	zswap_entry_cache_free(entry);
1595 reject:
1596 	if (objcg)
1597 		obj_cgroup_put(objcg);
1598 check_old:
1599 	/*
1600 	 * If the zswap store fails or zswap is disabled, we must invalidate the
1601 	 * possibly stale entry which was previously stored at this offset.
1602 	 * Otherwise, writeback could overwrite the new data in the swapfile.
1603 	 */
1604 	spin_lock(&tree->lock);
1605 	entry = zswap_rb_search(&tree->rbroot, offset);
1606 	if (entry)
1607 		zswap_invalidate_entry(tree, entry);
1608 	spin_unlock(&tree->lock);
1609 	return false;
1610 
1611 shrink:
1612 	queue_work(shrink_wq, &zswap_shrink_work);
1613 	goto reject;
1614 }
1615 
1616 bool zswap_load(struct folio *folio)
1617 {
1618 	swp_entry_t swp = folio->swap;
1619 	pgoff_t offset = swp_offset(swp);
1620 	struct page *page = &folio->page;
1621 	struct zswap_tree *tree = swap_zswap_tree(swp);
1622 	struct zswap_entry *entry;
1623 	u8 *dst;
1624 
1625 	VM_WARN_ON_ONCE(!folio_test_locked(folio));
1626 
1627 	spin_lock(&tree->lock);
1628 	entry = zswap_rb_search(&tree->rbroot, offset);
1629 	if (!entry) {
1630 		spin_unlock(&tree->lock);
1631 		return false;
1632 	}
1633 	zswap_rb_erase(&tree->rbroot, entry);
1634 	spin_unlock(&tree->lock);
1635 
1636 	if (entry->length)
1637 		zswap_decompress(entry, page);
1638 	else {
1639 		dst = kmap_local_page(page);
1640 		zswap_fill_page(dst, entry->value);
1641 		kunmap_local(dst);
1642 	}
1643 
1644 	count_vm_event(ZSWPIN);
1645 	if (entry->objcg)
1646 		count_objcg_event(entry->objcg, ZSWPIN);
1647 
1648 	zswap_entry_free(entry);
1649 
1650 	folio_mark_dirty(folio);
1651 
1652 	return true;
1653 }
1654 
1655 void zswap_invalidate(swp_entry_t swp)
1656 {
1657 	pgoff_t offset = swp_offset(swp);
1658 	struct zswap_tree *tree = swap_zswap_tree(swp);
1659 	struct zswap_entry *entry;
1660 
1661 	spin_lock(&tree->lock);
1662 	entry = zswap_rb_search(&tree->rbroot, offset);
1663 	if (entry)
1664 		zswap_invalidate_entry(tree, entry);
1665 	spin_unlock(&tree->lock);
1666 }
1667 
1668 int zswap_swapon(int type, unsigned long nr_pages)
1669 {
1670 	struct zswap_tree *trees, *tree;
1671 	unsigned int nr, i;
1672 
1673 	nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
1674 	trees = kvcalloc(nr, sizeof(*tree), GFP_KERNEL);
1675 	if (!trees) {
1676 		pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1677 		return -ENOMEM;
1678 	}
1679 
1680 	for (i = 0; i < nr; i++) {
1681 		tree = trees + i;
1682 		tree->rbroot = RB_ROOT;
1683 		spin_lock_init(&tree->lock);
1684 	}
1685 
1686 	nr_zswap_trees[type] = nr;
1687 	zswap_trees[type] = trees;
1688 	return 0;
1689 }
1690 
1691 void zswap_swapoff(int type)
1692 {
1693 	struct zswap_tree *trees = zswap_trees[type];
1694 	unsigned int i;
1695 
1696 	if (!trees)
1697 		return;
1698 
1699 	/* try_to_unuse() invalidated all the entries already */
1700 	for (i = 0; i < nr_zswap_trees[type]; i++)
1701 		WARN_ON_ONCE(!RB_EMPTY_ROOT(&trees[i].rbroot));
1702 
1703 	kvfree(trees);
1704 	nr_zswap_trees[type] = 0;
1705 	zswap_trees[type] = NULL;
1706 }
1707 
1708 /*********************************
1709 * debugfs functions
1710 **********************************/
1711 #ifdef CONFIG_DEBUG_FS
1712 #include <linux/debugfs.h>
1713 
1714 static struct dentry *zswap_debugfs_root;
1715 
1716 static int zswap_debugfs_init(void)
1717 {
1718 	if (!debugfs_initialized())
1719 		return -ENODEV;
1720 
1721 	zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1722 
1723 	debugfs_create_u64("pool_limit_hit", 0444,
1724 			   zswap_debugfs_root, &zswap_pool_limit_hit);
1725 	debugfs_create_u64("reject_reclaim_fail", 0444,
1726 			   zswap_debugfs_root, &zswap_reject_reclaim_fail);
1727 	debugfs_create_u64("reject_alloc_fail", 0444,
1728 			   zswap_debugfs_root, &zswap_reject_alloc_fail);
1729 	debugfs_create_u64("reject_kmemcache_fail", 0444,
1730 			   zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1731 	debugfs_create_u64("reject_compress_fail", 0444,
1732 			   zswap_debugfs_root, &zswap_reject_compress_fail);
1733 	debugfs_create_u64("reject_compress_poor", 0444,
1734 			   zswap_debugfs_root, &zswap_reject_compress_poor);
1735 	debugfs_create_u64("written_back_pages", 0444,
1736 			   zswap_debugfs_root, &zswap_written_back_pages);
1737 	debugfs_create_u64("pool_total_size", 0444,
1738 			   zswap_debugfs_root, &zswap_pool_total_size);
1739 	debugfs_create_atomic_t("stored_pages", 0444,
1740 				zswap_debugfs_root, &zswap_stored_pages);
1741 	debugfs_create_atomic_t("same_filled_pages", 0444,
1742 				zswap_debugfs_root, &zswap_same_filled_pages);
1743 
1744 	return 0;
1745 }
1746 #else
1747 static int zswap_debugfs_init(void)
1748 {
1749 	return 0;
1750 }
1751 #endif
1752 
1753 /*********************************
1754 * module init and exit
1755 **********************************/
1756 static int zswap_setup(void)
1757 {
1758 	struct zswap_pool *pool;
1759 	int ret;
1760 
1761 	zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
1762 	if (!zswap_entry_cache) {
1763 		pr_err("entry cache creation failed\n");
1764 		goto cache_fail;
1765 	}
1766 
1767 	ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1768 				      "mm/zswap_pool:prepare",
1769 				      zswap_cpu_comp_prepare,
1770 				      zswap_cpu_comp_dead);
1771 	if (ret)
1772 		goto hp_fail;
1773 
1774 	shrink_wq = alloc_workqueue("zswap-shrink",
1775 			WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
1776 	if (!shrink_wq)
1777 		goto shrink_wq_fail;
1778 
1779 	zswap_shrinker = zswap_alloc_shrinker();
1780 	if (!zswap_shrinker)
1781 		goto shrinker_fail;
1782 	if (list_lru_init_memcg(&zswap_list_lru, zswap_shrinker))
1783 		goto lru_fail;
1784 	shrinker_register(zswap_shrinker);
1785 
1786 	INIT_WORK(&zswap_shrink_work, shrink_worker);
1787 
1788 	pool = __zswap_pool_create_fallback();
1789 	if (pool) {
1790 		pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1791 			zpool_get_type(pool->zpools[0]));
1792 		list_add(&pool->list, &zswap_pools);
1793 		zswap_has_pool = true;
1794 	} else {
1795 		pr_err("pool creation failed\n");
1796 		zswap_enabled = false;
1797 	}
1798 
1799 	if (zswap_debugfs_init())
1800 		pr_warn("debugfs initialization failed\n");
1801 	zswap_init_state = ZSWAP_INIT_SUCCEED;
1802 	return 0;
1803 
1804 lru_fail:
1805 	shrinker_free(zswap_shrinker);
1806 shrinker_fail:
1807 	destroy_workqueue(shrink_wq);
1808 shrink_wq_fail:
1809 	cpuhp_remove_multi_state(CPUHP_MM_ZSWP_POOL_PREPARE);
1810 hp_fail:
1811 	kmem_cache_destroy(zswap_entry_cache);
1812 cache_fail:
1813 	/* if built-in, we aren't unloaded on failure; don't allow use */
1814 	zswap_init_state = ZSWAP_INIT_FAILED;
1815 	zswap_enabled = false;
1816 	return -ENOMEM;
1817 }
1818 
1819 static int __init zswap_init(void)
1820 {
1821 	if (!zswap_enabled)
1822 		return 0;
1823 	return zswap_setup();
1824 }
1825 /* must be late so crypto has time to come up */
1826 late_initcall(zswap_init);
1827 
1828 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1829 MODULE_DESCRIPTION("Compressed cache for swap pages");
1830