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/swap.h>
24	#include <linux/crypto.h>
25	#include <linux/scatterlist.h>
26	#include <linux/mempolicy.h>
27	#include <linux/mempool.h>
28	#include <crypto/acompress.h>
29	#include <linux/zswap.h>
30	#include <linux/mm_types.h>
31	#include <linux/page-flags.h>
32	#include <linux/swapops.h>
33	#include <linux/writeback.h>
34	#include <linux/pagemap.h>
35	#include <linux/workqueue.h>
36	#include <linux/list_lru.h>
37	#include <linux/zsmalloc.h>
38
39	#include "swap.h"
40	#include "internal.h"
41
42	/*********************************
43	* statistics
44	**********************************/
45	/* The number of pages currently stored in zswap */
46	atomic_long_t zswap_stored_pages = ATOMIC_LONG_INIT(0);
47	/* The number of incompressible pages currently stored in zswap */
48	static atomic_long_t zswap_stored_incompressible_pages = ATOMIC_LONG_INIT(0);
49
50	/*
51	* The statistics below are not protected from concurrent access for
52	* performance reasons so they may not be a 100% accurate. However,
53	* they do provide useful information on roughly how many times a
54	* certain event is occurring.
55	*/
56
57	/* Pool limit was hit (see zswap_max_pool_percent) */
58	static u64 zswap_pool_limit_hit;
59	/* Pages written back when pool limit was reached */
60	static u64 zswap_written_back_pages;
61	/* Store failed due to a reclaim failure after pool limit was reached */
62	static u64 zswap_reject_reclaim_fail;
63	/* Store failed due to compression algorithm failure */
64	static u64 zswap_reject_compress_fail;
65	/* Compressed page was too big for the allocator to (optimally) store */
66	static u64 zswap_reject_compress_poor;
67	/* Load or writeback failed due to decompression failure */
68	static u64 zswap_decompress_fail;
69	/* Store failed because underlying allocator could not get memory */
70	static u64 zswap_reject_alloc_fail;
71	/* Store failed because the entry metadata could not be allocated (rare) */
72	static u64 zswap_reject_kmemcache_fail;
73
74	/* Shrinker work queue */
75	static struct workqueue_struct *shrink_wq;
76	/* Pool limit was hit, we need to calm down */
77	static bool zswap_pool_reached_full;
78
79	/*********************************
80	* tunables
81	**********************************/
82
83	#define ZSWAP_PARAM_UNSET ""
84
85	static int zswap_setup(void);
86
87	/* Enable/disable zswap */
88	static DEFINE_STATIC_KEY_MAYBE(CONFIG_ZSWAP_DEFAULT_ON, zswap_ever_enabled);
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	/* The maximum percentage of memory that the compressed pool can occupy */
111	static unsigned int zswap_max_pool_percent = 20;
112	module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
113
114	/* The threshold for accepting new pages after the max_pool_percent was hit */
115	static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
116	module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
117	uint, 0644);
118
119	/* Enable/disable memory pressure-based shrinker. */
120	static bool zswap_shrinker_enabled = IS_ENABLED(
121	CONFIG_ZSWAP_SHRINKER_DEFAULT_ON);
122	module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
123
124	bool zswap_is_enabled(void)
125	{
126	return zswap_enabled;
127	}
128
129	bool zswap_never_enabled(void)
130	{
131	return !static_branch_maybe(CONFIG_ZSWAP_DEFAULT_ON, &zswap_ever_enabled);
132	}
133
134	/*********************************
135	* data structures
136	**********************************/
137
138	struct crypto_acomp_ctx {
139	struct crypto_acomp *acomp;
140	struct acomp_req *req;
141	struct crypto_wait wait;
142	u8 *buffer;
143	struct mutex mutex;
144	bool is_sleepable;
145	};
146
147	/*
148	* The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
149	* The only case where lru_lock is not acquired while holding tree.lock is
150	* when a zswap_entry is taken off the lru for writeback, in that case it
151	* needs to be verified that it's still valid in the tree.
152	*/
153	struct zswap_pool {
154	struct zs_pool *zs_pool;
155	struct crypto_acomp_ctx __percpu *acomp_ctx;
156	struct percpu_ref ref;
157	struct list_head list;
158	struct work_struct release_work;
159	struct hlist_node node;
160	char tfm_name[CRYPTO_MAX_ALG_NAME];
161	};
162
163	/* Global LRU lists shared by all zswap pools. */
164	static struct list_lru zswap_list_lru;
165
166	/* The lock protects zswap_next_shrink updates. */
167	static DEFINE_SPINLOCK(zswap_shrink_lock);
168	static struct mem_cgroup *zswap_next_shrink;
169	static struct work_struct zswap_shrink_work;
170	static struct shrinker *zswap_shrinker;
171
172	/*
173	* struct zswap_entry
174	*
175	* This structure contains the metadata for tracking a single compressed
176	* page within zswap.
177	*
178	* swpentry - associated swap entry, the offset indexes into the xarray
179	* length - the length in bytes of the compressed page data. Needed during
180	* decompression.
181	* referenced - true if the entry recently entered the zswap pool. Unset by the
182	* writeback logic. The entry is only reclaimed by the writeback
183	* logic if referenced is unset. See comments in the shrinker
184	* section for context.
185	* pool - the zswap_pool the entry's data is in
186	* handle - zsmalloc allocation handle that stores the compressed page data
187	* objcg - the obj_cgroup that the compressed memory is charged to
188	* lru - handle to the pool's lru used to evict pages.
189	*/
190	struct zswap_entry {
191	swp_entry_t swpentry;
192	unsigned int length;
193	bool referenced;
194	struct zswap_pool *pool;
195	unsigned long handle;
196	struct obj_cgroup *objcg;
197	struct list_head lru;
198	};
199
200	static struct xarray *zswap_trees[MAX_SWAPFILES];
201	static unsigned int nr_zswap_trees[MAX_SWAPFILES];
202
203	/* RCU-protected iteration */
204	static LIST_HEAD(zswap_pools);
205	/* protects zswap_pools list modification */
206	static DEFINE_SPINLOCK(zswap_pools_lock);
207	/* pool counter to provide unique names to zsmalloc */
208	static atomic_t zswap_pools_count = ATOMIC_INIT(0);
209
210	enum zswap_init_type {
211	ZSWAP_UNINIT,
212	ZSWAP_INIT_SUCCEED,
213	ZSWAP_INIT_FAILED
214	};
215
216	static enum zswap_init_type zswap_init_state;
217
218	/* used to ensure the integrity of initialization */
219	static DEFINE_MUTEX(zswap_init_lock);
220
221	/* init completed, but couldn't create the initial pool */
222	static bool zswap_has_pool;
223
224	/*********************************
225	* helpers and fwd declarations
226	**********************************/
227
228	/* One swap address space for each 64M swap space */
229	#define ZSWAP_ADDRESS_SPACE_SHIFT 14
230	#define ZSWAP_ADDRESS_SPACE_PAGES (1 << ZSWAP_ADDRESS_SPACE_SHIFT)
231	static inline struct xarray *swap_zswap_tree(swp_entry_t swp)
232	{
233	return &zswap_trees[swp_type(entry: swp)][swp_offset(entry: swp)
234	>> ZSWAP_ADDRESS_SPACE_SHIFT];
235	}
236
237	#define zswap_pool_debug(msg, p) \
238	pr_debug("%s pool %s\n", msg, (p)->tfm_name)
239
240	/*********************************
241	* pool functions
242	**********************************/
243	static void __zswap_pool_empty(struct percpu_ref *ref);
244
245	static struct zswap_pool *zswap_pool_create(char *compressor)
246	{
247	struct zswap_pool *pool;
248	char name[38]; /* 'zswap' + 32 char (max) num + \0 */
249	int ret, cpu;
250
251	if (!zswap_has_pool && !strcmp(compressor, ZSWAP_PARAM_UNSET))
252	return NULL;
253
254	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
255	if (!pool)
256	return NULL;
257
258	/* unique name for each pool specifically required by zsmalloc */
259	snprintf(buf: name, size: 38, fmt: "zswap%x", atomic_inc_return(v: &zswap_pools_count));
260	pool->zs_pool = zs_create_pool(name);
261	if (!pool->zs_pool)
262	goto error;
263
264	strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
265
266	pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
267	if (!pool->acomp_ctx) {
268	pr_err("percpu alloc failed\n");
269	goto error;
270	}
271
272	for_each_possible_cpu(cpu)
273	mutex_init(&per_cpu_ptr(pool->acomp_ctx, cpu)->mutex);
274
275	ret = cpuhp_state_add_instance(state: CPUHP_MM_ZSWP_POOL_PREPARE,
276	node: &pool->node);
277	if (ret)
278	goto error;
279
280	/* being the current pool takes 1 ref; this func expects the
281	* caller to always add the new pool as the current pool
282	*/
283	ret = percpu_ref_init(ref: &pool->ref, release: __zswap_pool_empty,
284	flags: PERCPU_REF_ALLOW_REINIT, GFP_KERNEL);
285	if (ret)
286	goto ref_fail;
287	INIT_LIST_HEAD(list: &pool->list);
288
289	zswap_pool_debug("created", pool);
290
291	return pool;
292
293	ref_fail:
294	cpuhp_state_remove_instance(state: CPUHP_MM_ZSWP_POOL_PREPARE, node: &pool->node);
295	error:
296	if (pool->acomp_ctx)
297	free_percpu(pdata: pool->acomp_ctx);
298	if (pool->zs_pool)
299	zs_destroy_pool(pool: pool->zs_pool);
300	kfree(objp: pool);
301	return NULL;
302	}
303
304	static struct zswap_pool *__zswap_pool_create_fallback(void)
305	{
306	if (!crypto_has_acomp(alg_name: zswap_compressor, type: 0, mask: 0) &&
307	strcmp(zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
308	pr_err("compressor %s not available, using default %s\n",
309	zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
310	param_free_charp(arg: &zswap_compressor);
311	zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
312	}
313
314	/* Default compressor should be available. Kconfig bug? */
315	if (WARN_ON_ONCE(!crypto_has_acomp(zswap_compressor, 0, 0))) {
316	zswap_compressor = ZSWAP_PARAM_UNSET;
317	return NULL;
318	}
319
320	return zswap_pool_create(compressor: zswap_compressor);
321	}
322
323	static void zswap_pool_destroy(struct zswap_pool *pool)
324	{
325	zswap_pool_debug("destroying", pool);
326
327	cpuhp_state_remove_instance(state: CPUHP_MM_ZSWP_POOL_PREPARE, node: &pool->node);
328	free_percpu(pdata: pool->acomp_ctx);
329
330	zs_destroy_pool(pool: pool->zs_pool);
331	kfree(objp: pool);
332	}
333
334	static void __zswap_pool_release(struct work_struct *work)
335	{
336	struct zswap_pool *pool = container_of(work, typeof(*pool),
337	release_work);
338
339	synchronize_rcu();
340
341	/* nobody should have been able to get a ref... */
342	WARN_ON(!percpu_ref_is_zero(&pool->ref));
343	percpu_ref_exit(ref: &pool->ref);
344
345	/* pool is now off zswap_pools list and has no references. */
346	zswap_pool_destroy(pool);
347	}
348
349	static struct zswap_pool *zswap_pool_current(void);
350
351	static void __zswap_pool_empty(struct percpu_ref *ref)
352	{
353	struct zswap_pool *pool;
354
355	pool = container_of(ref, typeof(*pool), ref);
356
357	spin_lock_bh(lock: &zswap_pools_lock);
358
359	WARN_ON(pool == zswap_pool_current());
360
361	list_del_rcu(entry: &pool->list);
362
363	INIT_WORK(&pool->release_work, __zswap_pool_release);
364	schedule_work(work: &pool->release_work);
365
366	spin_unlock_bh(lock: &zswap_pools_lock);
367	}
368
369	static int __must_check zswap_pool_tryget(struct zswap_pool *pool)
370	{
371	if (!pool)
372	return 0;
373
374	return percpu_ref_tryget(ref: &pool->ref);
375	}
376
377	/* The caller must already have a reference. */
378	static void zswap_pool_get(struct zswap_pool *pool)
379	{
380	percpu_ref_get(ref: &pool->ref);
381	}
382
383	static void zswap_pool_put(struct zswap_pool *pool)
384	{
385	percpu_ref_put(ref: &pool->ref);
386	}
387
388	static struct zswap_pool *__zswap_pool_current(void)
389	{
390	struct zswap_pool *pool;
391
392	pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
393	WARN_ONCE(!pool && zswap_has_pool,
394	"%s: no page storage pool!\n", __func__);
395
396	return pool;
397	}
398
399	static struct zswap_pool *zswap_pool_current(void)
400	{
401	assert_spin_locked(&zswap_pools_lock);
402
403	return __zswap_pool_current();
404	}
405
406	static struct zswap_pool *zswap_pool_current_get(void)
407	{
408	struct zswap_pool *pool;
409
410	rcu_read_lock();
411
412	pool = __zswap_pool_current();
413	if (!zswap_pool_tryget(pool))
414	pool = NULL;
415
416	rcu_read_unlock();
417
418	return pool;
419	}
420
421	/* type and compressor must be null-terminated */
422	static struct zswap_pool *zswap_pool_find_get(char *compressor)
423	{
424	struct zswap_pool *pool;
425
426	assert_spin_locked(&zswap_pools_lock);
427
428	list_for_each_entry_rcu(pool, &zswap_pools, list) {
429	if (strcmp(pool->tfm_name, compressor))
430	continue;
431	/* if we can't get it, it's about to be destroyed */
432	if (!zswap_pool_tryget(pool))
433	continue;
434	return pool;
435	}
436
437	return NULL;
438	}
439
440	static unsigned long zswap_max_pages(void)
441	{
442	return totalram_pages() * zswap_max_pool_percent / 100;
443	}
444
445	static unsigned long zswap_accept_thr_pages(void)
446	{
447	return zswap_max_pages() * zswap_accept_thr_percent / 100;
448	}
449
450	unsigned long zswap_total_pages(void)
451	{
452	struct zswap_pool *pool;
453	unsigned long total = 0;
454
455	rcu_read_lock();
456	list_for_each_entry_rcu(pool, &zswap_pools, list)
457	total += zs_get_total_pages(pool: pool->zs_pool);
458	rcu_read_unlock();
459
460	return total;
461	}
462
463	static bool zswap_check_limits(void)
464	{
465	unsigned long cur_pages = zswap_total_pages();
466	unsigned long max_pages = zswap_max_pages();
467
468	if (cur_pages >= max_pages) {
469	zswap_pool_limit_hit++;
470	zswap_pool_reached_full = true;
471	} else if (zswap_pool_reached_full &&
472	cur_pages <= zswap_accept_thr_pages()) {
473	zswap_pool_reached_full = false;
474	}
475	return zswap_pool_reached_full;
476	}
477
478	/*********************************
479	* param callbacks
480	**********************************/
481
482	static int zswap_compressor_param_set(const char *val, const struct kernel_param *kp)
483	{
484	struct zswap_pool *pool, *put_pool = NULL;
485	char *s = strstrip(str: (char *)val);
486	bool create_pool = false;
487	int ret = 0;
488
489	mutex_lock(&zswap_init_lock);
490	switch (zswap_init_state) {
491	case ZSWAP_UNINIT:
492	/* Handled in zswap_setup() */
493	ret = param_set_charp(val: s, kp);
494	break;
495	case ZSWAP_INIT_SUCCEED:
496	if (!zswap_has_pool || strcmp(s, *(char **)kp->arg))
497	create_pool = true;
498	break;
499	case ZSWAP_INIT_FAILED:
500	pr_err("can't set param, initialization failed\n");
501	ret = -ENODEV;
502	}
503	mutex_unlock(lock: &zswap_init_lock);
504
505	if (!create_pool)
506	return ret;
507
508	if (!crypto_has_acomp(alg_name: s, type: 0, mask: 0)) {
509	pr_err("compressor %s not available\n", s);
510	return -ENOENT;
511	}
512
513	spin_lock_bh(lock: &zswap_pools_lock);
514
515	pool = zswap_pool_find_get(compressor: s);
516	if (pool) {
517	zswap_pool_debug("using existing", pool);
518	WARN_ON(pool == zswap_pool_current());
519	list_del_rcu(entry: &pool->list);
520	}
521
522	spin_unlock_bh(lock: &zswap_pools_lock);
523
524	if (!pool)
525	pool = zswap_pool_create(compressor: s);
526	else {
527	/*
528	* Restore the initial ref dropped by percpu_ref_kill()
529	* when the pool was decommissioned and switch it again
530	* to percpu mode.
531	*/
532	percpu_ref_resurrect(ref: &pool->ref);
533
534	/* Drop the ref from zswap_pool_find_get(). */
535	zswap_pool_put(pool);
536	}
537
538	if (pool)
539	ret = param_set_charp(val: s, kp);
540	else
541	ret = -EINVAL;
542
543	spin_lock_bh(lock: &zswap_pools_lock);
544
545	if (!ret) {
546	put_pool = zswap_pool_current();
547	list_add_rcu(new: &pool->list, head: &zswap_pools);
548	zswap_has_pool = true;
549	} else if (pool) {
550	/*
551	* Add the possibly pre-existing pool to the end of the pools
552	* list; if it's new (and empty) then it'll be removed and
553	* destroyed by the put after we drop the lock
554	*/
555	list_add_tail_rcu(new: &pool->list, head: &zswap_pools);
556	put_pool = pool;
557	}
558
559	spin_unlock_bh(lock: &zswap_pools_lock);
560
561	/*
562	* Drop the ref from either the old current pool,
563	* or the new pool we failed to add
564	*/
565	if (put_pool)
566	percpu_ref_kill(ref: &put_pool->ref);
567
568	return ret;
569	}
570
571	static int zswap_enabled_param_set(const char *val,
572	const struct kernel_param *kp)
573	{
574	int ret = -ENODEV;
575
576	/* if this is load-time (pre-init) param setting, only set param. */
577	if (system_state != SYSTEM_RUNNING)
578	return param_set_bool(val, kp);
579
580	mutex_lock(&zswap_init_lock);
581	switch (zswap_init_state) {
582	case ZSWAP_UNINIT:
583	if (zswap_setup())
584	break;
585	fallthrough;
586	case ZSWAP_INIT_SUCCEED:
587	if (!zswap_has_pool)
588	pr_err("can't enable, no pool configured\n");
589	else
590	ret = param_set_bool(val, kp);
591	break;
592	case ZSWAP_INIT_FAILED:
593	pr_err("can't enable, initialization failed\n");
594	}
595	mutex_unlock(lock: &zswap_init_lock);
596
597	return ret;
598	}
599
600	/*********************************
601	* lru functions
602	**********************************/
603
604	/* should be called under RCU */
605	#ifdef CONFIG_MEMCG
606	static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
607	{
608	return entry->objcg ? obj_cgroup_memcg(objcg: entry->objcg) : NULL;
609	}
610	#else
611	static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
612	{
613	return NULL;
614	}
615	#endif
616
617	static inline int entry_to_nid(struct zswap_entry *entry)
618	{
619	return page_to_nid(virt_to_page(entry));
620	}
621
622	static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
623	{
624	int nid = entry_to_nid(entry);
625	struct mem_cgroup *memcg;
626
627	/*
628	* Note that it is safe to use rcu_read_lock() here, even in the face of
629	* concurrent memcg offlining:
630	*
631	* 1. list_lru_add() is called before list_lru_one is dead. The
632	* new entry will be reparented to memcg's parent's list_lru.
633	* 2. list_lru_add() is called after list_lru_one is dead. The
634	* new entry will be added directly to memcg's parent's list_lru.
635	*
636	* Similar reasoning holds for list_lru_del().
637	*/
638	rcu_read_lock();
639	memcg = mem_cgroup_from_entry(entry);
640	/* will always succeed */
641	list_lru_add(lru: list_lru, item: &entry->lru, nid, memcg);
642	rcu_read_unlock();
643	}
644
645	static void zswap_lru_del(struct list_lru *list_lru, struct zswap_entry *entry)
646	{
647	int nid = entry_to_nid(entry);
648	struct mem_cgroup *memcg;
649
650	rcu_read_lock();
651	memcg = mem_cgroup_from_entry(entry);
652	/* will always succeed */
653	list_lru_del(lru: list_lru, item: &entry->lru, nid, memcg);
654	rcu_read_unlock();
655	}
656
657	void zswap_lruvec_state_init(struct lruvec *lruvec)
658	{
659	atomic_long_set(v: &lruvec->zswap_lruvec_state.nr_disk_swapins, i: 0);
660	}
661
662	void zswap_folio_swapin(struct folio *folio)
663	{
664	struct lruvec *lruvec;
665
666	if (folio) {
667	lruvec = folio_lruvec(folio);
668	atomic_long_inc(v: &lruvec->zswap_lruvec_state.nr_disk_swapins);
669	}
670	}
671
672	/*
673	* This function should be called when a memcg is being offlined.
674	*
675	* Since the global shrinker shrink_worker() may hold a reference
676	* of the memcg, we must check and release the reference in
677	* zswap_next_shrink.
678	*
679	* shrink_worker() must handle the case where this function releases
680	* the reference of memcg being shrunk.
681	*/
682	void zswap_memcg_offline_cleanup(struct mem_cgroup *memcg)
683	{
684	/* lock out zswap shrinker walking memcg tree */
685	spin_lock(lock: &zswap_shrink_lock);
686	if (zswap_next_shrink == memcg) {
687	do {
688	zswap_next_shrink = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
689	} while (zswap_next_shrink && !mem_cgroup_online(memcg: zswap_next_shrink));
690	}
691	spin_unlock(lock: &zswap_shrink_lock);
692	}
693
694	/*********************************
695	* zswap entry functions
696	**********************************/
697	static struct kmem_cache *zswap_entry_cache;
698
699	static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp, int nid)
700	{
701	struct zswap_entry *entry;
702	entry = kmem_cache_alloc_node(zswap_entry_cache, gfp, nid);
703	if (!entry)
704	return NULL;
705	return entry;
706	}
707
708	static void zswap_entry_cache_free(struct zswap_entry *entry)
709	{
710	kmem_cache_free(s: zswap_entry_cache, objp: entry);
711	}
712
713	/*
714	* Carries out the common pattern of freeing an entry's zsmalloc allocation,
715	* freeing the entry itself, and decrementing the number of stored pages.
716	*/
717	static void zswap_entry_free(struct zswap_entry *entry)
718	{
719	zswap_lru_del(list_lru: &zswap_list_lru, entry);
720	zs_free(pool: entry->pool->zs_pool, obj: entry->handle);
721	zswap_pool_put(pool: entry->pool);
722	if (entry->objcg) {
723	obj_cgroup_uncharge_zswap(objcg: entry->objcg, size: entry->length);
724	obj_cgroup_put(objcg: entry->objcg);
725	}
726	if (entry->length == PAGE_SIZE)
727	atomic_long_dec(v: &zswap_stored_incompressible_pages);
728	zswap_entry_cache_free(entry);
729	atomic_long_dec(v: &zswap_stored_pages);
730	}
731
732	/*********************************
733	* compressed storage functions
734	**********************************/
735	static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
736	{
737	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
738	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
739	struct crypto_acomp *acomp = NULL;
740	struct acomp_req *req = NULL;
741	u8 *buffer = NULL;
742	int ret;
743
744	buffer = kmalloc_node(PAGE_SIZE, GFP_KERNEL, cpu_to_node(cpu));
745	if (!buffer) {
746	ret = -ENOMEM;
747	goto fail;
748	}
749
750	acomp = crypto_alloc_acomp_node(alg_name: pool->tfm_name, type: 0, mask: 0, cpu_to_node(cpu));
751	if (IS_ERR(ptr: acomp)) {
752	pr_err("could not alloc crypto acomp %s : %ld\n",
753	pool->tfm_name, PTR_ERR(acomp));
754	ret = PTR_ERR(ptr: acomp);
755	goto fail;
756	}
757
758	req = acomp_request_alloc(acomp);
759	if (!req) {
760	pr_err("could not alloc crypto acomp_request %s\n",
761	pool->tfm_name);
762	ret = -ENOMEM;
763	goto fail;
764	}
765
766	/*
767	* Only hold the mutex after completing allocations, otherwise we may
768	* recurse into zswap through reclaim and attempt to hold the mutex
769	* again resulting in a deadlock.
770	*/
771	mutex_lock(&acomp_ctx->mutex);
772	crypto_init_wait(wait: &acomp_ctx->wait);
773
774	/*
775	* if the backend of acomp is async zip, crypto_req_done() will wakeup
776	* crypto_wait_req(); if the backend of acomp is scomp, the callback
777	* won't be called, crypto_wait_req() will return without blocking.
778	*/
779	acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
780	cmpl: crypto_req_done, data: &acomp_ctx->wait);
781
782	acomp_ctx->buffer = buffer;
783	acomp_ctx->acomp = acomp;
784	acomp_ctx->is_sleepable = acomp_is_async(tfm: acomp);
785	acomp_ctx->req = req;
786	mutex_unlock(lock: &acomp_ctx->mutex);
787	return 0;
788
789	fail:
790	if (!IS_ERR_OR_NULL(ptr: acomp))
791	crypto_free_acomp(tfm: acomp);
792	kfree(objp: buffer);
793	return ret;
794	}
795
796	static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
797	{
798	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
799	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
800	struct acomp_req *req;
801	struct crypto_acomp *acomp;
802	u8 *buffer;
803
804	if (IS_ERR_OR_NULL(ptr: acomp_ctx))
805	return 0;
806
807	mutex_lock(&acomp_ctx->mutex);
808	req = acomp_ctx->req;
809	acomp = acomp_ctx->acomp;
810	buffer = acomp_ctx->buffer;
811	acomp_ctx->req = NULL;
812	acomp_ctx->acomp = NULL;
813	acomp_ctx->buffer = NULL;
814	mutex_unlock(lock: &acomp_ctx->mutex);
815
816	/*
817	* Do the actual freeing after releasing the mutex to avoid subtle
818	* locking dependencies causing deadlocks.
819	*/
820	if (!IS_ERR_OR_NULL(ptr: req))
821	acomp_request_free(req);
822	if (!IS_ERR_OR_NULL(ptr: acomp))
823	crypto_free_acomp(tfm: acomp);
824	kfree(objp: buffer);
825
826	return 0;
827	}
828
829	static struct crypto_acomp_ctx *acomp_ctx_get_cpu_lock(struct zswap_pool *pool)
830	{
831	struct crypto_acomp_ctx *acomp_ctx;
832
833	for (;;) {
834	acomp_ctx = raw_cpu_ptr(pool->acomp_ctx);
835	mutex_lock(&acomp_ctx->mutex);
836	if (likely(acomp_ctx->req))
837	return acomp_ctx;
838	/*
839	* It is possible that we were migrated to a different CPU after
840	* getting the per-CPU ctx but before the mutex was acquired. If
841	* the old CPU got offlined, zswap_cpu_comp_dead() could have
842	* already freed ctx->req (among other things) and set it to
843	* NULL. Just try again on the new CPU that we ended up on.
844	*/
845	mutex_unlock(lock: &acomp_ctx->mutex);
846	}
847	}
848
849	static void acomp_ctx_put_unlock(struct crypto_acomp_ctx *acomp_ctx)
850	{
851	mutex_unlock(lock: &acomp_ctx->mutex);
852	}
853
854	static bool zswap_compress(struct page *page, struct zswap_entry *entry,
855	struct zswap_pool *pool)
856	{
857	struct crypto_acomp_ctx *acomp_ctx;
858	struct scatterlist input, output;
859	int comp_ret = 0, alloc_ret = 0;
860	unsigned int dlen = PAGE_SIZE;
861	unsigned long handle;
862	gfp_t gfp;
863	u8 *dst;
864	bool mapped = false;
865
866	acomp_ctx = acomp_ctx_get_cpu_lock(pool);
867	dst = acomp_ctx->buffer;
868	sg_init_table(&input, 1);
869	sg_set_page(sg: &input, page, PAGE_SIZE, offset: 0);
870
871	sg_init_one(&output, dst, PAGE_SIZE);
872	acomp_request_set_params(req: acomp_ctx->req, src: &input, dst: &output, PAGE_SIZE, dlen);
873
874	/*
875	* it maybe looks a little bit silly that we send an asynchronous request,
876	* then wait for its completion synchronously. This makes the process look
877	* synchronous in fact.
878	* Theoretically, acomp supports users send multiple acomp requests in one
879	* acomp instance, then get those requests done simultaneously. but in this
880	* case, zswap actually does store and load page by page, there is no
881	* existing method to send the second page before the first page is done
882	* in one thread doing zswap.
883	* but in different threads running on different cpu, we have different
884	* acomp instance, so multiple threads can do (de)compression in parallel.
885	*/
886	comp_ret = crypto_wait_req(err: crypto_acomp_compress(req: acomp_ctx->req), wait: &acomp_ctx->wait);
887	dlen = acomp_ctx->req->dlen;
888
889	/*
890	* If a page cannot be compressed into a size smaller than PAGE_SIZE,
891	* save the content as is without a compression, to keep the LRU order
892	* of writebacks. If writeback is disabled, reject the page since it
893	* only adds metadata overhead. swap_writeout() will put the page back
894	* to the active LRU list in the case.
895	*/
896	if (comp_ret || !dlen || dlen >= PAGE_SIZE) {
897	if (!mem_cgroup_zswap_writeback_enabled(
898	memcg: folio_memcg(page_folio(page)))) {
899	comp_ret = comp_ret ? comp_ret : -EINVAL;
900	goto unlock;
901	}
902	comp_ret = 0;
903	dlen = PAGE_SIZE;
904	dst = kmap_local_page(page);
905	mapped = true;
906	}
907
908	gfp = GFP_NOWAIT | __GFP_NORETRY | __GFP_HIGHMEM | __GFP_MOVABLE;
909	handle = zs_malloc(pool: pool->zs_pool, size: dlen, flags: gfp, nid: page_to_nid(page));
910	if (IS_ERR_VALUE(handle)) {
911	alloc_ret = PTR_ERR(ptr: (void *)handle);
912	goto unlock;
913	}
914
915	zs_obj_write(pool: pool->zs_pool, handle, handle_mem: dst, mem_len: dlen);
916	entry->handle = handle;
917	entry->length = dlen;
918
919	unlock:
920	if (mapped)
921	kunmap_local(dst);
922	if (comp_ret == -ENOSPC || alloc_ret == -ENOSPC)
923	zswap_reject_compress_poor++;
924	else if (comp_ret)
925	zswap_reject_compress_fail++;
926	else if (alloc_ret)
927	zswap_reject_alloc_fail++;
928
929	acomp_ctx_put_unlock(acomp_ctx);
930	return comp_ret == 0 && alloc_ret == 0;
931	}
932
933	static bool zswap_decompress(struct zswap_entry *entry, struct folio *folio)
934	{
935	struct zswap_pool *pool = entry->pool;
936	struct scatterlist input, output;
937	struct crypto_acomp_ctx *acomp_ctx;
938	int decomp_ret = 0, dlen = PAGE_SIZE;
939	u8 *src, *obj;
940
941	acomp_ctx = acomp_ctx_get_cpu_lock(pool);
942	obj = zs_obj_read_begin(pool: pool->zs_pool, handle: entry->handle, local_copy: acomp_ctx->buffer);
943
944	/* zswap entries of length PAGE_SIZE are not compressed. */
945	if (entry->length == PAGE_SIZE) {
946	memcpy_to_folio(folio, offset: 0, from: obj, len: entry->length);
947	goto read_done;
948	}
949
950	/*
951	* zs_obj_read_begin() might return a kmap address of highmem when
952	* acomp_ctx->buffer is not used. However, sg_init_one() does not
953	* handle highmem addresses, so copy the object to acomp_ctx->buffer.
954	*/
955	if (virt_addr_valid(obj)) {
956	src = obj;
957	} else {
958	WARN_ON_ONCE(obj == acomp_ctx->buffer);
959	memcpy(acomp_ctx->buffer, obj, entry->length);
960	src = acomp_ctx->buffer;
961	}
962
963	sg_init_one(&input, src, entry->length);
964	sg_init_table(&output, 1);
965	sg_set_folio(sg: &output, folio, PAGE_SIZE, offset: 0);
966	acomp_request_set_params(req: acomp_ctx->req, src: &input, dst: &output, slen: entry->length, PAGE_SIZE);
967	decomp_ret = crypto_wait_req(err: crypto_acomp_decompress(req: acomp_ctx->req), wait: &acomp_ctx->wait);
968	dlen = acomp_ctx->req->dlen;
969
970	read_done:
971	zs_obj_read_end(pool: pool->zs_pool, handle: entry->handle, handle_mem: obj);
972	acomp_ctx_put_unlock(acomp_ctx);
973
974	if (!decomp_ret && dlen == PAGE_SIZE)
975	return true;
976
977	zswap_decompress_fail++;
978	pr_alert_ratelimited("Decompression error from zswap (%d:%lu %s %u->%d)\n",
979	swp_type(entry->swpentry),
980	swp_offset(entry->swpentry),
981	entry->pool->tfm_name, entry->length, dlen);
982	return false;
983	}
984
985	/*********************************
986	* writeback code
987	**********************************/
988	/*
989	* Attempts to free an entry by adding a folio to the swap cache,
990	* decompressing the entry data into the folio, and issuing a
991	* bio write to write the folio back to the swap device.
992	*
993	* This can be thought of as a "resumed writeback" of the folio
994	* to the swap device. We are basically resuming the same swap
995	* writeback path that was intercepted with the zswap_store()
996	* in the first place. After the folio has been decompressed into
997	* the swap cache, the compressed version stored by zswap can be
998	* freed.
999	*/
1000	static int zswap_writeback_entry(struct zswap_entry *entry,
1001	swp_entry_t swpentry)
1002	{
1003	struct xarray *tree;
1004	pgoff_t offset = swp_offset(entry: swpentry);
1005	struct folio *folio;
1006	struct mempolicy *mpol;
1007	bool folio_was_allocated;
1008	struct swap_info_struct *si;
1009	int ret = 0;
1010
1011	/* try to allocate swap cache folio */
1012	si = get_swap_device(entry: swpentry);
1013	if (!si)
1014	return -EEXIST;
1015
1016	mpol = get_task_policy(current);
1017	folio = __read_swap_cache_async(entry: swpentry, GFP_KERNEL, mpol,
1018	NO_INTERLEAVE_INDEX, new_page_allocated: &folio_was_allocated, skip_if_exists: true);
1019	put_swap_device(si);
1020	if (!folio)
1021	return -ENOMEM;
1022
1023	/*
1024	* Found an existing folio, we raced with swapin or concurrent
1025	* shrinker. We generally writeback cold folios from zswap, and
1026	* swapin means the folio just became hot, so skip this folio.
1027	* For unlikely concurrent shrinker case, it will be unlinked
1028	* and freed when invalidated by the concurrent shrinker anyway.
1029	*/
1030	if (!folio_was_allocated) {
1031	ret = -EEXIST;
1032	goto out;
1033	}
1034
1035	/*
1036	* folio is locked, and the swapcache is now secured against
1037	* concurrent swapping to and from the slot, and concurrent
1038	* swapoff so we can safely dereference the zswap tree here.
1039	* Verify that the swap entry hasn't been invalidated and recycled
1040	* behind our backs, to avoid overwriting a new swap folio with
1041	* old compressed data. Only when this is successful can the entry
1042	* be dereferenced.
1043	*/
1044	tree = swap_zswap_tree(swp: swpentry);
1045	if (entry != xa_load(tree, index: offset)) {
1046	ret = -ENOMEM;
1047	goto out;
1048	}
1049
1050	if (!zswap_decompress(entry, folio)) {
1051	ret = -EIO;
1052	goto out;
1053	}
1054
1055	xa_erase(tree, index: offset);
1056
1057	count_vm_event(item: ZSWPWB);
1058	if (entry->objcg)
1059	count_objcg_events(objcg: entry->objcg, idx: ZSWPWB, count: 1);
1060
1061	zswap_entry_free(entry);
1062
1063	/* folio is up to date */
1064	folio_mark_uptodate(folio);
1065
1066	/* move it to the tail of the inactive list after end_writeback */
1067	folio_set_reclaim(folio);
1068
1069	/* start writeback */
1070	__swap_writepage(folio, NULL);
1071
1072	out:
1073	if (ret && ret != -EEXIST) {
1074	swap_cache_del_folio(folio);
1075	folio_unlock(folio);
1076	}
1077	folio_put(folio);
1078	return ret;
1079	}
1080
1081	/*********************************
1082	* shrinker functions
1083	**********************************/
1084	/*
1085	* The dynamic shrinker is modulated by the following factors:
1086	*
1087	* 1. Each zswap entry has a referenced bit, which the shrinker unsets (giving
1088	* the entry a second chance) before rotating it in the LRU list. If the
1089	* entry is considered again by the shrinker, with its referenced bit unset,
1090	* it is written back. The writeback rate as a result is dynamically
1091	* adjusted by the pool activities - if the pool is dominated by new entries
1092	* (i.e lots of recent zswapouts), these entries will be protected and
1093	* the writeback rate will slow down. On the other hand, if the pool has a
1094	* lot of stagnant entries, these entries will be reclaimed immediately,
1095	* effectively increasing the writeback rate.
1096	*
1097	* 2. Swapins counter: If we observe swapins, it is a sign that we are
1098	* overshrinking and should slow down. We maintain a swapins counter, which
1099	* is consumed and subtract from the number of eligible objects on the LRU
1100	* in zswap_shrinker_count().
1101	*
1102	* 3. Compression ratio. The better the workload compresses, the less gains we
1103	* can expect from writeback. We scale down the number of objects available
1104	* for reclaim by this ratio.
1105	*/
1106	static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
1107	void *arg)
1108	{
1109	struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
1110	bool *encountered_page_in_swapcache = (bool *)arg;
1111	swp_entry_t swpentry;
1112	enum lru_status ret = LRU_REMOVED_RETRY;
1113	int writeback_result;
1114
1115	/*
1116	* Second chance algorithm: if the entry has its referenced bit set, give it
1117	* a second chance. Only clear the referenced bit and rotate it in the
1118	* zswap's LRU list.
1119	*/
1120	if (entry->referenced) {
1121	entry->referenced = false;
1122	return LRU_ROTATE;
1123	}
1124
1125	/*
1126	* As soon as we drop the LRU lock, the entry can be freed by
1127	* a concurrent invalidation. This means the following:
1128	*
1129	* 1. We extract the swp_entry_t to the stack, allowing
1130	* zswap_writeback_entry() to pin the swap entry and
1131	* then validate the zswap entry against that swap entry's
1132	* tree using pointer value comparison. Only when that
1133	* is successful can the entry be dereferenced.
1134	*
1135	* 2. Usually, objects are taken off the LRU for reclaim. In
1136	* this case this isn't possible, because if reclaim fails
1137	* for whatever reason, we have no means of knowing if the
1138	* entry is alive to put it back on the LRU.
1139	*
1140	* So rotate it before dropping the lock. If the entry is
1141	* written back or invalidated, the free path will unlink
1142	* it. For failures, rotation is the right thing as well.
1143	*
1144	* Temporary failures, where the same entry should be tried
1145	* again immediately, almost never happen for this shrinker.
1146	* We don't do any trylocking; -ENOMEM comes closest,
1147	* but that's extremely rare and doesn't happen spuriously
1148	* either. Don't bother distinguishing this case.
1149	*/
1150	list_move_tail(list: item, head: &l->list);
1151
1152	/*
1153	* Once the lru lock is dropped, the entry might get freed. The
1154	* swpentry is copied to the stack, and entry isn't deref'd again
1155	* until the entry is verified to still be alive in the tree.
1156	*/
1157	swpentry = entry->swpentry;
1158
1159	/*
1160	* It's safe to drop the lock here because we return either
1161	* LRU_REMOVED_RETRY, LRU_RETRY or LRU_STOP.
1162	*/
1163	spin_unlock(lock: &l->lock);
1164
1165	writeback_result = zswap_writeback_entry(entry, swpentry);
1166
1167	if (writeback_result) {
1168	zswap_reject_reclaim_fail++;
1169	ret = LRU_RETRY;
1170
1171	/*
1172	* Encountering a page already in swap cache is a sign that we are shrinking
1173	* into the warmer region. We should terminate shrinking (if we're in the dynamic
1174	* shrinker context).
1175	*/
1176	if (writeback_result == -EEXIST && encountered_page_in_swapcache) {
1177	ret = LRU_STOP;
1178	*encountered_page_in_swapcache = true;
1179	}
1180	} else {
1181	zswap_written_back_pages++;
1182	}
1183
1184	return ret;
1185	}
1186
1187	static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
1188	struct shrink_control *sc)
1189	{
1190	unsigned long shrink_ret;
1191	bool encountered_page_in_swapcache = false;
1192
1193	if (!zswap_shrinker_enabled ||
1194	!mem_cgroup_zswap_writeback_enabled(memcg: sc->memcg)) {
1195	sc->nr_scanned = 0;
1196	return SHRINK_STOP;
1197	}
1198
1199	shrink_ret = list_lru_shrink_walk(lru: &zswap_list_lru, sc, isolate: &shrink_memcg_cb,
1200	cb_arg: &encountered_page_in_swapcache);
1201
1202	if (encountered_page_in_swapcache)
1203	return SHRINK_STOP;
1204
1205	return shrink_ret ? shrink_ret : SHRINK_STOP;
1206	}
1207
1208	static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
1209	struct shrink_control *sc)
1210	{
1211	struct mem_cgroup *memcg = sc->memcg;
1212	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
1213	atomic_long_t *nr_disk_swapins =
1214	&lruvec->zswap_lruvec_state.nr_disk_swapins;
1215	unsigned long nr_backing, nr_stored, nr_freeable, nr_disk_swapins_cur,
1216	nr_remain;
1217
1218	if (!zswap_shrinker_enabled || !mem_cgroup_zswap_writeback_enabled(memcg))
1219	return 0;
1220
1221	/*
1222	* The shrinker resumes swap writeback, which will enter block
1223	* and may enter fs. XXX: Harmonize with vmscan.c __GFP_FS
1224	* rules (may_enter_fs()), which apply on a per-folio basis.
1225	*/
1226	if (!gfp_has_io_fs(gfp: sc->gfp_mask))
1227	return 0;
1228
1229	/*
1230	* For memcg, use the cgroup-wide ZSWAP stats since we don't
1231	* have them per-node and thus per-lruvec. Careful if memcg is
1232	* runtime-disabled: we can get sc->memcg == NULL, which is ok
1233	* for the lruvec, but not for memcg_page_state().
1234	*
1235	* Without memcg, use the zswap pool-wide metrics.
1236	*/
1237	if (!mem_cgroup_disabled()) {
1238	mem_cgroup_flush_stats(memcg);
1239	nr_backing = memcg_page_state(memcg, idx: MEMCG_ZSWAP_B) >> PAGE_SHIFT;
1240	nr_stored = memcg_page_state(memcg, idx: MEMCG_ZSWAPPED);
1241	} else {
1242	nr_backing = zswap_total_pages();
1243	nr_stored = atomic_long_read(v: &zswap_stored_pages);
1244	}
1245
1246	if (!nr_stored)
1247	return 0;
1248
1249	nr_freeable = list_lru_shrink_count(lru: &zswap_list_lru, sc);
1250	if (!nr_freeable)
1251	return 0;
1252
1253	/*
1254	* Subtract from the lru size the number of pages that are recently swapped
1255	* in from disk. The idea is that had we protect the zswap's LRU by this
1256	* amount of pages, these disk swapins would not have happened.
1257	*/
1258	nr_disk_swapins_cur = atomic_long_read(v: nr_disk_swapins);
1259	do {
1260	if (nr_freeable >= nr_disk_swapins_cur)
1261	nr_remain = 0;
1262	else
1263	nr_remain = nr_disk_swapins_cur - nr_freeable;
1264	} while (!atomic_long_try_cmpxchg(
1265	v: nr_disk_swapins, old: &nr_disk_swapins_cur, new: nr_remain));
1266
1267	nr_freeable -= nr_disk_swapins_cur - nr_remain;
1268	if (!nr_freeable)
1269	return 0;
1270
1271	/*
1272	* Scale the number of freeable pages by the memory saving factor.
1273	* This ensures that the better zswap compresses memory, the fewer
1274	* pages we will evict to swap (as it will otherwise incur IO for
1275	* relatively small memory saving).
1276	*/
1277	return mult_frac(nr_freeable, nr_backing, nr_stored);
1278	}
1279
1280	static struct shrinker *zswap_alloc_shrinker(void)
1281	{
1282	struct shrinker *shrinker;
1283
1284	shrinker =
1285	shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, fmt: "mm-zswap");
1286	if (!shrinker)
1287	return NULL;
1288
1289	shrinker->scan_objects = zswap_shrinker_scan;
1290	shrinker->count_objects = zswap_shrinker_count;
1291	shrinker->batch = 0;
1292	shrinker->seeks = DEFAULT_SEEKS;
1293	return shrinker;
1294	}
1295
1296	static int shrink_memcg(struct mem_cgroup *memcg)
1297	{
1298	int nid, shrunk = 0, scanned = 0;
1299
1300	if (!mem_cgroup_zswap_writeback_enabled(memcg))
1301	return -ENOENT;
1302
1303	/*
1304	* Skip zombies because their LRUs are reparented and we would be
1305	* reclaiming from the parent instead of the dead memcg.
1306	*/
1307	if (memcg && !mem_cgroup_online(memcg))
1308	return -ENOENT;
1309
1310	for_each_node_state(nid, N_NORMAL_MEMORY) {
1311	unsigned long nr_to_walk = 1;
1312
1313	shrunk += list_lru_walk_one(lru: &zswap_list_lru, nid, memcg,
1314	isolate: &shrink_memcg_cb, NULL, nr_to_walk: &nr_to_walk);
1315	scanned += 1 - nr_to_walk;
1316	}
1317
1318	if (!scanned)
1319	return -ENOENT;
1320
1321	return shrunk ? 0 : -EAGAIN;
1322	}
1323
1324	static void shrink_worker(struct work_struct *w)
1325	{
1326	struct mem_cgroup *memcg;
1327	int ret, failures = 0, attempts = 0;
1328	unsigned long thr;
1329
1330	/* Reclaim down to the accept threshold */
1331	thr = zswap_accept_thr_pages();
1332
1333	/*
1334	* Global reclaim will select cgroup in a round-robin fashion from all
1335	* online memcgs, but memcgs that have no pages in zswap and
1336	* writeback-disabled memcgs (memory.zswap.writeback=0) are not
1337	* candidates for shrinking.
1338	*
1339	* Shrinking will be aborted if we encounter the following
1340	* MAX_RECLAIM_RETRIES times:
1341	* - No writeback-candidate memcgs found in a memcg tree walk.
1342	* - Shrinking a writeback-candidate memcg failed.
1343	*
1344	* We save iteration cursor memcg into zswap_next_shrink,
1345	* which can be modified by the offline memcg cleaner
1346	* zswap_memcg_offline_cleanup().
1347	*
1348	* Since the offline cleaner is called only once, we cannot leave an
1349	* offline memcg reference in zswap_next_shrink.
1350	* We can rely on the cleaner only if we get online memcg under lock.
1351	*
1352	* If we get an offline memcg, we cannot determine if the cleaner has
1353	* already been called or will be called later. We must put back the
1354	* reference before returning from this function. Otherwise, the
1355	* offline memcg left in zswap_next_shrink will hold the reference
1356	* until the next run of shrink_worker().
1357	*/
1358	do {
1359	/*
1360	* Start shrinking from the next memcg after zswap_next_shrink.
1361	* When the offline cleaner has already advanced the cursor,
1362	* advancing the cursor here overlooks one memcg, but this
1363	* should be negligibly rare.
1364	*
1365	* If we get an online memcg, keep the extra reference in case
1366	* the original one obtained by mem_cgroup_iter() is dropped by
1367	* zswap_memcg_offline_cleanup() while we are shrinking the
1368	* memcg.
1369	*/
1370	spin_lock(lock: &zswap_shrink_lock);
1371	do {
1372	memcg = mem_cgroup_iter(NULL, zswap_next_shrink, NULL);
1373	zswap_next_shrink = memcg;
1374	} while (memcg && !mem_cgroup_tryget_online(memcg));
1375	spin_unlock(lock: &zswap_shrink_lock);
1376
1377	if (!memcg) {
1378	/*
1379	* Continue shrinking without incrementing failures if
1380	* we found candidate memcgs in the last tree walk.
1381	*/
1382	if (!attempts && ++failures == MAX_RECLAIM_RETRIES)
1383	break;
1384
1385	attempts = 0;
1386	goto resched;
1387	}
1388
1389	ret = shrink_memcg(memcg);
1390	/* drop the extra reference */
1391	mem_cgroup_put(memcg);
1392
1393	/*
1394	* There are no writeback-candidate pages in the memcg.
1395	* This is not an issue as long as we can find another memcg
1396	* with pages in zswap. Skip this without incrementing attempts
1397	* and failures.
1398	*/
1399	if (ret == -ENOENT)
1400	continue;
1401	++attempts;
1402
1403	if (ret && ++failures == MAX_RECLAIM_RETRIES)
1404	break;
1405	resched:
1406	cond_resched();
1407	} while (zswap_total_pages() > thr);
1408	}
1409
1410	/*********************************
1411	* main API
1412	**********************************/
1413
1414	static bool zswap_store_page(struct page *page,
1415	struct obj_cgroup *objcg,
1416	struct zswap_pool *pool)
1417	{
1418	swp_entry_t page_swpentry = page_swap_entry(page);
1419	struct zswap_entry *entry, *old;
1420
1421	/* allocate entry */
1422	entry = zswap_entry_cache_alloc(GFP_KERNEL, nid: page_to_nid(page));
1423	if (!entry) {
1424	zswap_reject_kmemcache_fail++;
1425	return false;
1426	}
1427
1428	if (!zswap_compress(page, entry, pool))
1429	goto compress_failed;
1430
1431	old = xa_store(swap_zswap_tree(swp: page_swpentry),
1432	index: swp_offset(entry: page_swpentry),
1433	entry, GFP_KERNEL);
1434	if (xa_is_err(entry: old)) {
1435	int err = xa_err(entry: old);
1436
1437	WARN_ONCE(err != -ENOMEM, "unexpected xarray error: %d\n", err);
1438	zswap_reject_alloc_fail++;
1439	goto store_failed;
1440	}
1441
1442	/*
1443	* We may have had an existing entry that became stale when
1444	* the folio was redirtied and now the new version is being
1445	* swapped out. Get rid of the old.
1446	*/
1447	if (old)
1448	zswap_entry_free(entry: old);
1449
1450	/*
1451	* The entry is successfully compressed and stored in the tree, there is
1452	* no further possibility of failure. Grab refs to the pool and objcg,
1453	* charge zswap memory, and increment zswap_stored_pages.
1454	* The opposite actions will be performed by zswap_entry_free()
1455	* when the entry is removed from the tree.
1456	*/
1457	zswap_pool_get(pool);
1458	if (objcg) {
1459	obj_cgroup_get(objcg);
1460	obj_cgroup_charge_zswap(objcg, size: entry->length);
1461	}
1462	atomic_long_inc(v: &zswap_stored_pages);
1463	if (entry->length == PAGE_SIZE)
1464	atomic_long_inc(v: &zswap_stored_incompressible_pages);
1465
1466	/*
1467	* We finish initializing the entry while it's already in xarray.
1468	* This is safe because:
1469	*
1470	* 1. Concurrent stores and invalidations are excluded by folio lock.
1471	*
1472	* 2. Writeback is excluded by the entry not being on the LRU yet.
1473	* The publishing order matters to prevent writeback from seeing
1474	* an incoherent entry.
1475	*/
1476	entry->pool = pool;
1477	entry->swpentry = page_swpentry;
1478	entry->objcg = objcg;
1479	entry->referenced = true;
1480	if (entry->length) {
1481	INIT_LIST_HEAD(list: &entry->lru);
1482	zswap_lru_add(list_lru: &zswap_list_lru, entry);
1483	}
1484
1485	return true;
1486
1487	store_failed:
1488	zs_free(pool: pool->zs_pool, obj: entry->handle);
1489	compress_failed:
1490	zswap_entry_cache_free(entry);
1491	return false;
1492	}
1493
1494	bool zswap_store(struct folio *folio)
1495	{
1496	long nr_pages = folio_nr_pages(folio);
1497	swp_entry_t swp = folio->swap;
1498	struct obj_cgroup *objcg = NULL;
1499	struct mem_cgroup *memcg = NULL;
1500	struct zswap_pool *pool;
1501	bool ret = false;
1502	long index;
1503
1504	VM_WARN_ON_ONCE(!folio_test_locked(folio));
1505	VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
1506
1507	if (!zswap_enabled)
1508	goto check_old;
1509
1510	objcg = get_obj_cgroup_from_folio(folio);
1511	if (objcg && !obj_cgroup_may_zswap(objcg)) {
1512	memcg = get_mem_cgroup_from_objcg(objcg);
1513	if (shrink_memcg(memcg)) {
1514	mem_cgroup_put(memcg);
1515	goto put_objcg;
1516	}
1517	mem_cgroup_put(memcg);
1518	}
1519
1520	if (zswap_check_limits())
1521	goto put_objcg;
1522
1523	pool = zswap_pool_current_get();
1524	if (!pool)
1525	goto put_objcg;
1526
1527	if (objcg) {
1528	memcg = get_mem_cgroup_from_objcg(objcg);
1529	if (memcg_list_lru_alloc(memcg, lru: &zswap_list_lru, GFP_KERNEL)) {
1530	mem_cgroup_put(memcg);
1531	goto put_pool;
1532	}
1533	mem_cgroup_put(memcg);
1534	}
1535
1536	for (index = 0; index < nr_pages; ++index) {
1537	struct page *page = folio_page(folio, index);
1538
1539	if (!zswap_store_page(page, objcg, pool))
1540	goto put_pool;
1541	}
1542
1543	if (objcg)
1544	count_objcg_events(objcg, idx: ZSWPOUT, count: nr_pages);
1545
1546	count_vm_events(item: ZSWPOUT, delta: nr_pages);
1547
1548	ret = true;
1549
1550	put_pool:
1551	zswap_pool_put(pool);
1552	put_objcg:
1553	obj_cgroup_put(objcg);
1554	if (!ret && zswap_pool_reached_full)
1555	queue_work(wq: shrink_wq, work: &zswap_shrink_work);
1556	check_old:
1557	/*
1558	* If the zswap store fails or zswap is disabled, we must invalidate
1559	* the possibly stale entries which were previously stored at the
1560	* offsets corresponding to each page of the folio. Otherwise,
1561	* writeback could overwrite the new data in the swapfile.
1562	*/
1563	if (!ret) {
1564	unsigned type = swp_type(entry: swp);
1565	pgoff_t offset = swp_offset(entry: swp);
1566	struct zswap_entry *entry;
1567	struct xarray *tree;
1568
1569	for (index = 0; index < nr_pages; ++index) {
1570	tree = swap_zswap_tree(swp: swp_entry(type, offset: offset + index));
1571	entry = xa_erase(tree, index: offset + index);
1572	if (entry)
1573	zswap_entry_free(entry);
1574	}
1575	}
1576
1577	return ret;
1578	}
1579
1580	/**
1581	* zswap_load() - load a folio from zswap
1582	* @folio: folio to load
1583	*
1584	* Return: 0 on success, with the folio unlocked and marked up-to-date, or one
1585	* of the following error codes:
1586	*
1587	* -EIO: if the swapped out content was in zswap, but could not be loaded
1588	* into the page due to a decompression failure. The folio is unlocked, but
1589	* NOT marked up-to-date, so that an IO error is emitted (e.g. do_swap_page()
1590	* will SIGBUS).
1591	*
1592	* -EINVAL: if the swapped out content was in zswap, but the page belongs
1593	* to a large folio, which is not supported by zswap. The folio is unlocked,
1594	* but NOT marked up-to-date, so that an IO error is emitted (e.g.
1595	* do_swap_page() will SIGBUS).
1596	*
1597	* -ENOENT: if the swapped out content was not in zswap. The folio remains
1598	* locked on return.
1599	*/
1600	int zswap_load(struct folio *folio)
1601	{
1602	swp_entry_t swp = folio->swap;
1603	pgoff_t offset = swp_offset(entry: swp);
1604	bool swapcache = folio_test_swapcache(folio);
1605	struct xarray *tree = swap_zswap_tree(swp);
1606	struct zswap_entry *entry;
1607
1608	VM_WARN_ON_ONCE(!folio_test_locked(folio));
1609
1610	if (zswap_never_enabled())
1611	return -ENOENT;
1612
1613	/*
1614	* Large folios should not be swapped in while zswap is being used, as
1615	* they are not properly handled. Zswap does not properly load large
1616	* folios, and a large folio may only be partially in zswap.
1617	*/
1618	if (WARN_ON_ONCE(folio_test_large(folio))) {
1619	folio_unlock(folio);
1620	return -EINVAL;
1621	}
1622
1623	entry = xa_load(tree, index: offset);
1624	if (!entry)
1625	return -ENOENT;
1626
1627	if (!zswap_decompress(entry, folio)) {
1628	folio_unlock(folio);
1629	return -EIO;
1630	}
1631
1632	folio_mark_uptodate(folio);
1633
1634	count_vm_event(item: ZSWPIN);
1635	if (entry->objcg)
1636	count_objcg_events(objcg: entry->objcg, idx: ZSWPIN, count: 1);
1637
1638	/*
1639	* When reading into the swapcache, invalidate our entry. The
1640	* swapcache can be the authoritative owner of the page and
1641	* its mappings, and the pressure that results from having two
1642	* in-memory copies outweighs any benefits of caching the
1643	* compression work.
1644	*
1645	* (Most swapins go through the swapcache. The notable
1646	* exception is the singleton fault on SWP_SYNCHRONOUS_IO
1647	* files, which reads into a private page and may free it if
1648	* the fault fails. We remain the primary owner of the entry.)
1649	*/
1650	if (swapcache) {
1651	folio_mark_dirty(folio);
1652	xa_erase(tree, index: offset);
1653	zswap_entry_free(entry);
1654	}
1655
1656	folio_unlock(folio);
1657	return 0;
1658	}
1659
1660	void zswap_invalidate(swp_entry_t swp)
1661	{
1662	pgoff_t offset = swp_offset(entry: swp);
1663	struct xarray *tree = swap_zswap_tree(swp);
1664	struct zswap_entry *entry;
1665
1666	if (xa_empty(xa: tree))
1667	return;
1668
1669	entry = xa_erase(tree, index: offset);
1670	if (entry)
1671	zswap_entry_free(entry);
1672	}
1673
1674	int zswap_swapon(int type, unsigned long nr_pages)
1675	{
1676	struct xarray *trees, *tree;
1677	unsigned int nr, i;
1678
1679	nr = DIV_ROUND_UP(nr_pages, ZSWAP_ADDRESS_SPACE_PAGES);
1680	trees = kvcalloc(nr, sizeof(*tree), GFP_KERNEL);
1681	if (!trees) {
1682	pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1683	return -ENOMEM;
1684	}
1685
1686	for (i = 0; i < nr; i++)
1687	xa_init(xa: trees + i);
1688
1689	nr_zswap_trees[type] = nr;
1690	zswap_trees[type] = trees;
1691	return 0;
1692	}
1693
1694	void zswap_swapoff(int type)
1695	{
1696	struct xarray *trees = zswap_trees[type];
1697	unsigned int i;
1698
1699	if (!trees)
1700	return;
1701
1702	/* try_to_unuse() invalidated all the entries already */
1703	for (i = 0; i < nr_zswap_trees[type]; i++)
1704	WARN_ON_ONCE(!xa_empty(trees + i));
1705
1706	kvfree(addr: trees);
1707	nr_zswap_trees[type] = 0;
1708	zswap_trees[type] = NULL;
1709	}
1710
1711	/*********************************
1712	* debugfs functions
1713	**********************************/
1714	#ifdef CONFIG_DEBUG_FS
1715	#include <linux/debugfs.h>
1716
1717	static struct dentry *zswap_debugfs_root;
1718
1719	static int debugfs_get_total_size(void *data, u64 *val)
1720	{
1721	*val = zswap_total_pages() * PAGE_SIZE;
1722	return 0;
1723	}
1724	DEFINE_DEBUGFS_ATTRIBUTE(total_size_fops, debugfs_get_total_size, NULL, "%llu\n");
1725
1726	static int debugfs_get_stored_pages(void *data, u64 *val)
1727	{
1728	*val = atomic_long_read(v: &zswap_stored_pages);
1729	return 0;
1730	}
1731	DEFINE_DEBUGFS_ATTRIBUTE(stored_pages_fops, debugfs_get_stored_pages, NULL, "%llu\n");
1732
1733	static int debugfs_get_stored_incompressible_pages(void *data, u64 *val)
1734	{
1735	*val = atomic_long_read(v: &zswap_stored_incompressible_pages);
1736	return 0;
1737	}
1738	DEFINE_DEBUGFS_ATTRIBUTE(stored_incompressible_pages_fops,
1739	debugfs_get_stored_incompressible_pages, NULL, "%llu\n");
1740
1741	static int zswap_debugfs_init(void)
1742	{
1743	if (!debugfs_initialized())
1744	return -ENODEV;
1745
1746	zswap_debugfs_root = debugfs_create_dir(name: "zswap", NULL);
1747
1748	debugfs_create_u64(name: "pool_limit_hit", mode: 0444,
1749	parent: zswap_debugfs_root, value: &zswap_pool_limit_hit);
1750	debugfs_create_u64(name: "reject_reclaim_fail", mode: 0444,
1751	parent: zswap_debugfs_root, value: &zswap_reject_reclaim_fail);
1752	debugfs_create_u64(name: "reject_alloc_fail", mode: 0444,
1753	parent: zswap_debugfs_root, value: &zswap_reject_alloc_fail);
1754	debugfs_create_u64(name: "reject_kmemcache_fail", mode: 0444,
1755	parent: zswap_debugfs_root, value: &zswap_reject_kmemcache_fail);
1756	debugfs_create_u64(name: "reject_compress_fail", mode: 0444,
1757	parent: zswap_debugfs_root, value: &zswap_reject_compress_fail);
1758	debugfs_create_u64(name: "reject_compress_poor", mode: 0444,
1759	parent: zswap_debugfs_root, value: &zswap_reject_compress_poor);
1760	debugfs_create_u64(name: "decompress_fail", mode: 0444,
1761	parent: zswap_debugfs_root, value: &zswap_decompress_fail);
1762	debugfs_create_u64(name: "written_back_pages", mode: 0444,
1763	parent: zswap_debugfs_root, value: &zswap_written_back_pages);
1764	debugfs_create_file("pool_total_size", 0444,
1765	zswap_debugfs_root, NULL, &total_size_fops);
1766	debugfs_create_file("stored_pages", 0444,
1767	zswap_debugfs_root, NULL, &stored_pages_fops);
1768	debugfs_create_file("stored_incompressible_pages", 0444,
1769	zswap_debugfs_root, NULL,
1770	&stored_incompressible_pages_fops);
1771
1772	return 0;
1773	}
1774	#else
1775	static int zswap_debugfs_init(void)
1776	{
1777	return 0;
1778	}
1779	#endif
1780
1781	/*********************************
1782	* module init and exit
1783	**********************************/
1784	static int zswap_setup(void)
1785	{
1786	struct zswap_pool *pool;
1787	int ret;
1788
1789	zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
1790	if (!zswap_entry_cache) {
1791	pr_err("entry cache creation failed\n");
1792	goto cache_fail;
1793	}
1794
1795	ret = cpuhp_setup_state_multi(state: CPUHP_MM_ZSWP_POOL_PREPARE,
1796	name: "mm/zswap_pool:prepare",
1797	startup: zswap_cpu_comp_prepare,
1798	teardown: zswap_cpu_comp_dead);
1799	if (ret)
1800	goto hp_fail;
1801
1802	shrink_wq = alloc_workqueue("zswap-shrink",
1803	WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
1804	if (!shrink_wq)
1805	goto shrink_wq_fail;
1806
1807	zswap_shrinker = zswap_alloc_shrinker();
1808	if (!zswap_shrinker)
1809	goto shrinker_fail;
1810	if (list_lru_init_memcg(&zswap_list_lru, zswap_shrinker))
1811	goto lru_fail;
1812	shrinker_register(shrinker: zswap_shrinker);
1813
1814	INIT_WORK(&zswap_shrink_work, shrink_worker);
1815
1816	pool = __zswap_pool_create_fallback();
1817	if (pool) {
1818	pr_info("loaded using pool %s\n", pool->tfm_name);
1819	list_add(new: &pool->list, head: &zswap_pools);
1820	zswap_has_pool = true;
1821	static_branch_enable(&zswap_ever_enabled);
1822	} else {
1823	pr_err("pool creation failed\n");
1824	zswap_enabled = false;
1825	}
1826
1827	if (zswap_debugfs_init())
1828	pr_warn("debugfs initialization failed\n");
1829	zswap_init_state = ZSWAP_INIT_SUCCEED;
1830	return 0;
1831
1832	lru_fail:
1833	shrinker_free(shrinker: zswap_shrinker);
1834	shrinker_fail:
1835	destroy_workqueue(wq: shrink_wq);
1836	shrink_wq_fail:
1837	cpuhp_remove_multi_state(state: CPUHP_MM_ZSWP_POOL_PREPARE);
1838	hp_fail:
1839	kmem_cache_destroy(s: zswap_entry_cache);
1840	cache_fail:
1841	/* if built-in, we aren't unloaded on failure; don't allow use */
1842	zswap_init_state = ZSWAP_INIT_FAILED;
1843	zswap_enabled = false;
1844	return -ENOMEM;
1845	}
1846
1847	static int __init zswap_init(void)
1848	{
1849	if (!zswap_enabled)
1850	return 0;
1851	return zswap_setup();
1852	}
1853	/* must be late so crypto has time to come up */
1854	late_initcall(zswap_init);
1855
1856	MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1857	MODULE_DESCRIPTION("Compressed cache for swap pages");
1858