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 */
  47u64 zswap_pool_total_size;
  48/* The number of compressed pages currently stored in zswap */
  49atomic_t zswap_stored_pages = ATOMIC_INIT(0);
  50/* The number of same-value filled pages currently stored in zswap */
  51static 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) */
  61static u64 zswap_pool_limit_hit;
  62/* Pages written back when pool limit was reached */
  63static u64 zswap_written_back_pages;
  64/* Store failed due to a reclaim failure after pool limit was reached */
  65static u64 zswap_reject_reclaim_fail;
  66/* Store failed due to compression algorithm failure */
  67static u64 zswap_reject_compress_fail;
  68/* Compressed page was too big for the allocator to (optimally) store */
  69static u64 zswap_reject_compress_poor;
  70/* Store failed because underlying allocator could not get memory */
  71static u64 zswap_reject_alloc_fail;
  72/* Store failed because the entry metadata could not be allocated (rare) */
  73static u64 zswap_reject_kmemcache_fail;
  74/* Duplicate store was encountered (rare) */
  75static u64 zswap_duplicate_entry;
  76
  77/* Shrinker work queue */
  78static struct workqueue_struct *shrink_wq;
  79/* Pool limit was hit, we need to calm down */
  80static bool zswap_pool_reached_full;
  81
  82/*********************************
  83* tunables
  84**********************************/
  85
  86#define ZSWAP_PARAM_UNSET ""
  87
  88static int zswap_setup(void);
  89
  90/* Enable/disable zswap */
  91static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
  92static int zswap_enabled_param_set(const char *,
  93				   const struct kernel_param *);
  94static const struct kernel_param_ops zswap_enabled_param_ops = {
  95	.set =		zswap_enabled_param_set,
  96	.get =		param_get_bool,
  97};
  98module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
  99
 100/* Crypto compressor to use */
 101static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
 
 102static int zswap_compressor_param_set(const char *,
 103				      const struct kernel_param *);
 104static const struct kernel_param_ops zswap_compressor_param_ops = {
 105	.set =		zswap_compressor_param_set,
 106	.get =		param_get_charp,
 107	.free =		param_free_charp,
 108};
 109module_param_cb(compressor, &zswap_compressor_param_ops,
 110		&zswap_compressor, 0644);
 111
 112/* Compressed storage zpool to use */
 113static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
 
 114static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 115static const struct kernel_param_ops zswap_zpool_param_ops = {
 116	.set =		zswap_zpool_param_set,
 117	.get =		param_get_charp,
 118	.free =		param_free_charp,
 119};
 120module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
 121
 122/* The maximum percentage of memory that the compressed pool can occupy */
 123static unsigned int zswap_max_pool_percent = 20;
 124module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
 125
 126/* The threshold for accepting new pages after the max_pool_percent was hit */
 127static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
 128module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
 129		   uint, 0644);
 130
 131/*
 132 * Enable/disable handling same-value filled pages (enabled by default).
 133 * If disabled every page is considered non-same-value filled.
 134 */
 135static bool zswap_same_filled_pages_enabled = true;
 136module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
 137		   bool, 0644);
 138
 139/* Enable/disable handling non-same-value filled pages (enabled by default) */
 140static bool zswap_non_same_filled_pages_enabled = true;
 141module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
 142		   bool, 0644);
 143
 144static bool zswap_exclusive_loads_enabled = IS_ENABLED(
 145		CONFIG_ZSWAP_EXCLUSIVE_LOADS_DEFAULT_ON);
 146module_param_named(exclusive_loads, zswap_exclusive_loads_enabled, bool, 0644);
 147
 148/* Number of zpools in zswap_pool (empirically determined for scalability) */
 149#define ZSWAP_NR_ZPOOLS 32
 150
 151/* Enable/disable memory pressure-based shrinker. */
 152static bool zswap_shrinker_enabled = IS_ENABLED(
 153		CONFIG_ZSWAP_SHRINKER_DEFAULT_ON);
 154module_param_named(shrinker_enabled, zswap_shrinker_enabled, bool, 0644);
 155
 156bool is_zswap_enabled(void)
 157{
 158	return zswap_enabled;
 159}
 160
 161/*********************************
 162* data structures
 163**********************************/
 164
 165struct crypto_acomp_ctx {
 166	struct crypto_acomp *acomp;
 167	struct acomp_req *req;
 168	struct crypto_wait wait;
 169	u8 *buffer;
 170	struct mutex mutex;
 171};
 172
 173/*
 174 * The lock ordering is zswap_tree.lock -> zswap_pool.lru_lock.
 175 * The only case where lru_lock is not acquired while holding tree.lock is
 176 * when a zswap_entry is taken off the lru for writeback, in that case it
 177 * needs to be verified that it's still valid in the tree.
 178 */
 179struct zswap_pool {
 180	struct zpool *zpools[ZSWAP_NR_ZPOOLS];
 181	struct crypto_acomp_ctx __percpu *acomp_ctx;
 182	struct kref kref;
 183	struct list_head list;
 184	struct work_struct release_work;
 185	struct work_struct shrink_work;
 186	struct hlist_node node;
 187	char tfm_name[CRYPTO_MAX_ALG_NAME];
 188	struct list_lru list_lru;
 189	struct mem_cgroup *next_shrink;
 190	struct shrinker *shrinker;
 191	atomic_t nr_stored;
 192};
 193
 194/*
 195 * struct zswap_entry
 196 *
 197 * This structure contains the metadata for tracking a single compressed
 198 * page within zswap.
 199 *
 200 * rbnode - links the entry into red-black tree for the appropriate swap type
 201 * swpentry - associated swap entry, the offset indexes into the red-black tree
 202 * refcount - the number of outstanding reference to the entry. This is needed
 203 *            to protect against premature freeing of the entry by code
 204 *            concurrent calls to load, invalidate, and writeback.  The lock
 205 *            for the zswap_tree structure that contains the entry must
 206 *            be held while changing the refcount.  Since the lock must
 207 *            be held, there is no reason to also make refcount atomic.
 208 * length - the length in bytes of the compressed page data.  Needed during
 209 *          decompression. For a same value filled page length is 0, and both
 210 *          pool and lru are invalid and must be ignored.
 211 * pool - the zswap_pool the entry's data is in
 212 * handle - zpool allocation handle that stores the compressed page data
 213 * value - value of the same-value filled pages which have same content
 214 * objcg - the obj_cgroup that the compressed memory is charged to
 215 * lru - handle to the pool's lru used to evict pages.
 216 */
 217struct zswap_entry {
 218	struct rb_node rbnode;
 219	swp_entry_t swpentry;
 220	int refcount;
 221	unsigned int length;
 222	struct zswap_pool *pool;
 223	union {
 224		unsigned long handle;
 225		unsigned long value;
 226	};
 227	struct obj_cgroup *objcg;
 228	struct list_head lru;
 229};
 230
 231/*
 232 * The tree lock in the zswap_tree struct protects a few things:
 233 * - the rbtree
 234 * - the refcount field of each entry in the tree
 235 */
 236struct zswap_tree {
 237	struct rb_root rbroot;
 238	spinlock_t lock;
 239};
 240
 241static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 242
 243/* RCU-protected iteration */
 244static LIST_HEAD(zswap_pools);
 245/* protects zswap_pools list modification */
 246static DEFINE_SPINLOCK(zswap_pools_lock);
 247/* pool counter to provide unique names to zpool */
 248static atomic_t zswap_pools_count = ATOMIC_INIT(0);
 249
 250enum zswap_init_type {
 251	ZSWAP_UNINIT,
 252	ZSWAP_INIT_SUCCEED,
 253	ZSWAP_INIT_FAILED
 254};
 255
 256static enum zswap_init_type zswap_init_state;
 257
 258/* used to ensure the integrity of initialization */
 259static DEFINE_MUTEX(zswap_init_lock);
 260
 261/* init completed, but couldn't create the initial pool */
 262static bool zswap_has_pool;
 263
 264/*********************************
 265* helpers and fwd declarations
 266**********************************/
 267
 268#define zswap_pool_debug(msg, p)				\
 269	pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,		\
 270		 zpool_get_type((p)->zpools[0]))
 271
 272static int zswap_writeback_entry(struct zswap_entry *entry,
 273				 struct zswap_tree *tree);
 274static int zswap_pool_get(struct zswap_pool *pool);
 275static void zswap_pool_put(struct zswap_pool *pool);
 276
 277static bool zswap_is_full(void)
 278{
 279	return totalram_pages() * zswap_max_pool_percent / 100 <
 280			DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 281}
 282
 283static bool zswap_can_accept(void)
 284{
 285	return totalram_pages() * zswap_accept_thr_percent / 100 *
 286				zswap_max_pool_percent / 100 >
 287			DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 288}
 289
 290static u64 get_zswap_pool_size(struct zswap_pool *pool)
 291{
 292	u64 pool_size = 0;
 293	int i;
 294
 295	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
 296		pool_size += zpool_get_total_size(pool->zpools[i]);
 297
 298	return pool_size;
 299}
 300
 301static void zswap_update_total_size(void)
 302{
 303	struct zswap_pool *pool;
 304	u64 total = 0;
 305
 306	rcu_read_lock();
 307
 308	list_for_each_entry_rcu(pool, &zswap_pools, list)
 309		total += get_zswap_pool_size(pool);
 310
 311	rcu_read_unlock();
 312
 313	zswap_pool_total_size = total;
 314}
 315
 316/* should be called under RCU */
 317#ifdef CONFIG_MEMCG
 318static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
 319{
 320	return entry->objcg ? obj_cgroup_memcg(entry->objcg) : NULL;
 321}
 322#else
 323static inline struct mem_cgroup *mem_cgroup_from_entry(struct zswap_entry *entry)
 324{
 325	return NULL;
 326}
 327#endif
 328
 329static inline int entry_to_nid(struct zswap_entry *entry)
 330{
 331	return page_to_nid(virt_to_page(entry));
 
 332}
 333
 334void zswap_memcg_offline_cleanup(struct mem_cgroup *memcg)
 335{
 336	struct zswap_pool *pool;
 337
 338	/* lock out zswap pools list modification */
 339	spin_lock(&zswap_pools_lock);
 340	list_for_each_entry(pool, &zswap_pools, list) {
 341		if (pool->next_shrink == memcg)
 342			pool->next_shrink = mem_cgroup_iter(NULL, pool->next_shrink, NULL);
 343	}
 344	spin_unlock(&zswap_pools_lock);
 345}
 346
 347/*********************************
 348* zswap entry functions
 349**********************************/
 350static struct kmem_cache *zswap_entry_cache;
 351
 352static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp, int nid)
 353{
 354	struct zswap_entry *entry;
 355	entry = kmem_cache_alloc_node(zswap_entry_cache, gfp, nid);
 356	if (!entry)
 357		return NULL;
 358	entry->refcount = 1;
 359	RB_CLEAR_NODE(&entry->rbnode);
 360	return entry;
 361}
 362
 363static void zswap_entry_cache_free(struct zswap_entry *entry)
 364{
 365	kmem_cache_free(zswap_entry_cache, entry);
 366}
 367
 368/*********************************
 369* zswap lruvec functions
 370**********************************/
 371void zswap_lruvec_state_init(struct lruvec *lruvec)
 372{
 373	atomic_long_set(&lruvec->zswap_lruvec_state.nr_zswap_protected, 0);
 374}
 375
 376void zswap_folio_swapin(struct folio *folio)
 377{
 378	struct lruvec *lruvec;
 379
 380	VM_WARN_ON_ONCE(!folio_test_locked(folio));
 381	lruvec = folio_lruvec(folio);
 382	atomic_long_inc(&lruvec->zswap_lruvec_state.nr_zswap_protected);
 383}
 384
 385/*********************************
 386* lru functions
 387**********************************/
 388static void zswap_lru_add(struct list_lru *list_lru, struct zswap_entry *entry)
 389{
 390	atomic_long_t *nr_zswap_protected;
 391	unsigned long lru_size, old, new;
 392	int nid = entry_to_nid(entry);
 393	struct mem_cgroup *memcg;
 394	struct lruvec *lruvec;
 395
 396	/*
 397	 * Note that it is safe to use rcu_read_lock() here, even in the face of
 398	 * concurrent memcg offlining. Thanks to the memcg->kmemcg_id indirection
 399	 * used in list_lru lookup, only two scenarios are possible:
 400	 *
 401	 * 1. list_lru_add() is called before memcg->kmemcg_id is updated. The
 402	 *    new entry will be reparented to memcg's parent's list_lru.
 403	 * 2. list_lru_add() is called after memcg->kmemcg_id is updated. The
 404	 *    new entry will be added directly to memcg's parent's list_lru.
 405	 *
 406	 * Similar reasoning holds for list_lru_del() and list_lru_putback().
 407	 */
 408	rcu_read_lock();
 409	memcg = mem_cgroup_from_entry(entry);
 410	/* will always succeed */
 411	list_lru_add(list_lru, &entry->lru, nid, memcg);
 412
 413	/* Update the protection area */
 414	lru_size = list_lru_count_one(list_lru, nid, memcg);
 415	lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(nid));
 416	nr_zswap_protected = &lruvec->zswap_lruvec_state.nr_zswap_protected;
 417	old = atomic_long_inc_return(nr_zswap_protected);
 418	/*
 419	 * Decay to avoid overflow and adapt to changing workloads.
 420	 * This is based on LRU reclaim cost decaying heuristics.
 421	 */
 422	do {
 423		new = old > lru_size / 4 ? old / 2 : old;
 424	} while (!atomic_long_try_cmpxchg(nr_zswap_protected, &old, new));
 425	rcu_read_unlock();
 426}
 427
 428static void zswap_lru_del(struct list_lru *list_lru, struct zswap_entry *entry)
 429{
 430	int nid = entry_to_nid(entry);
 431	struct mem_cgroup *memcg;
 432
 433	rcu_read_lock();
 434	memcg = mem_cgroup_from_entry(entry);
 435	/* will always succeed */
 436	list_lru_del(list_lru, &entry->lru, nid, memcg);
 437	rcu_read_unlock();
 438}
 439
 440static void zswap_lru_putback(struct list_lru *list_lru,
 441		struct zswap_entry *entry)
 442{
 443	int nid = entry_to_nid(entry);
 444	spinlock_t *lock = &list_lru->node[nid].lock;
 445	struct mem_cgroup *memcg;
 446	struct lruvec *lruvec;
 447
 448	rcu_read_lock();
 449	memcg = mem_cgroup_from_entry(entry);
 450	spin_lock(lock);
 451	/* we cannot use list_lru_add here, because it increments node's lru count */
 452	list_lru_putback(list_lru, &entry->lru, nid, memcg);
 453	spin_unlock(lock);
 454
 455	lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(entry_to_nid(entry)));
 456	/* increment the protection area to account for the LRU rotation. */
 457	atomic_long_inc(&lruvec->zswap_lruvec_state.nr_zswap_protected);
 458	rcu_read_unlock();
 459}
 460
 461/*********************************
 462* rbtree functions
 463**********************************/
 464static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 465{
 466	struct rb_node *node = root->rb_node;
 467	struct zswap_entry *entry;
 468	pgoff_t entry_offset;
 469
 470	while (node) {
 471		entry = rb_entry(node, struct zswap_entry, rbnode);
 472		entry_offset = swp_offset(entry->swpentry);
 473		if (entry_offset > offset)
 474			node = node->rb_left;
 475		else if (entry_offset < offset)
 476			node = node->rb_right;
 477		else
 478			return entry;
 479	}
 480	return NULL;
 481}
 482
 483/*
 484 * In the case that a entry with the same offset is found, a pointer to
 485 * the existing entry is stored in dupentry and the function returns -EEXIST
 486 */
 487static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 488			struct zswap_entry **dupentry)
 489{
 490	struct rb_node **link = &root->rb_node, *parent = NULL;
 491	struct zswap_entry *myentry;
 492	pgoff_t myentry_offset, entry_offset = swp_offset(entry->swpentry);
 493
 494	while (*link) {
 495		parent = *link;
 496		myentry = rb_entry(parent, struct zswap_entry, rbnode);
 497		myentry_offset = swp_offset(myentry->swpentry);
 498		if (myentry_offset > entry_offset)
 499			link = &(*link)->rb_left;
 500		else if (myentry_offset < entry_offset)
 501			link = &(*link)->rb_right;
 502		else {
 503			*dupentry = myentry;
 504			return -EEXIST;
 505		}
 506	}
 507	rb_link_node(&entry->rbnode, parent, link);
 508	rb_insert_color(&entry->rbnode, root);
 509	return 0;
 510}
 511
 512static bool zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 513{
 514	if (!RB_EMPTY_NODE(&entry->rbnode)) {
 515		rb_erase(&entry->rbnode, root);
 516		RB_CLEAR_NODE(&entry->rbnode);
 517		return true;
 518	}
 519	return false;
 520}
 521
 522static struct zpool *zswap_find_zpool(struct zswap_entry *entry)
 523{
 524	int i = 0;
 525
 526	if (ZSWAP_NR_ZPOOLS > 1)
 527		i = hash_ptr(entry, ilog2(ZSWAP_NR_ZPOOLS));
 528
 529	return entry->pool->zpools[i];
 530}
 531
 532/*
 533 * Carries out the common pattern of freeing and entry's zpool allocation,
 534 * freeing the entry itself, and decrementing the number of stored pages.
 535 */
 536static void zswap_free_entry(struct zswap_entry *entry)
 537{
 538	if (!entry->length)
 539		atomic_dec(&zswap_same_filled_pages);
 540	else {
 541		zswap_lru_del(&entry->pool->list_lru, entry);
 542		zpool_free(zswap_find_zpool(entry), entry->handle);
 543		atomic_dec(&entry->pool->nr_stored);
 544		zswap_pool_put(entry->pool);
 545	}
 546	if (entry->objcg) {
 547		obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
 548		obj_cgroup_put(entry->objcg);
 549	}
 550	zswap_entry_cache_free(entry);
 551	atomic_dec(&zswap_stored_pages);
 552	zswap_update_total_size();
 553}
 554
 555/* caller must hold the tree lock */
 556static void zswap_entry_get(struct zswap_entry *entry)
 557{
 558	entry->refcount++;
 559}
 560
 561/* caller must hold the tree lock
 562* remove from the tree and free it, if nobody reference the entry
 563*/
 564static void zswap_entry_put(struct zswap_tree *tree,
 565			struct zswap_entry *entry)
 566{
 567	int refcount = --entry->refcount;
 568
 569	WARN_ON_ONCE(refcount < 0);
 570	if (refcount == 0) {
 571		WARN_ON_ONCE(!RB_EMPTY_NODE(&entry->rbnode));
 572		zswap_free_entry(entry);
 573	}
 574}
 575
 576/* caller must hold the tree lock */
 577static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 578				pgoff_t offset)
 579{
 580	struct zswap_entry *entry;
 581
 582	entry = zswap_rb_search(root, offset);
 583	if (entry)
 584		zswap_entry_get(entry);
 585
 586	return entry;
 587}
 588
 589/*********************************
 590* shrinker functions
 591**********************************/
 592static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
 593				       spinlock_t *lock, void *arg);
 594
 595static unsigned long zswap_shrinker_scan(struct shrinker *shrinker,
 596		struct shrink_control *sc)
 597{
 598	struct lruvec *lruvec = mem_cgroup_lruvec(sc->memcg, NODE_DATA(sc->nid));
 599	unsigned long shrink_ret, nr_protected, lru_size;
 600	struct zswap_pool *pool = shrinker->private_data;
 601	bool encountered_page_in_swapcache = false;
 602
 603	if (!zswap_shrinker_enabled ||
 604			!mem_cgroup_zswap_writeback_enabled(sc->memcg)) {
 605		sc->nr_scanned = 0;
 606		return SHRINK_STOP;
 607	}
 608
 609	nr_protected =
 610		atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
 611	lru_size = list_lru_shrink_count(&pool->list_lru, sc);
 612
 613	/*
 614	 * Abort if we are shrinking into the protected region.
 615	 *
 616	 * This short-circuiting is necessary because if we have too many multiple
 617	 * concurrent reclaimers getting the freeable zswap object counts at the
 618	 * same time (before any of them made reasonable progress), the total
 619	 * number of reclaimed objects might be more than the number of unprotected
 620	 * objects (i.e the reclaimers will reclaim into the protected area of the
 621	 * zswap LRU).
 622	 */
 623	if (nr_protected >= lru_size - sc->nr_to_scan) {
 624		sc->nr_scanned = 0;
 625		return SHRINK_STOP;
 
 
 
 
 626	}
 627
 628	shrink_ret = list_lru_shrink_walk(&pool->list_lru, sc, &shrink_memcg_cb,
 629		&encountered_page_in_swapcache);
 630
 631	if (encountered_page_in_swapcache)
 632		return SHRINK_STOP;
 633
 634	return shrink_ret ? shrink_ret : SHRINK_STOP;
 635}
 636
 637static unsigned long zswap_shrinker_count(struct shrinker *shrinker,
 638		struct shrink_control *sc)
 639{
 640	struct zswap_pool *pool = shrinker->private_data;
 641	struct mem_cgroup *memcg = sc->memcg;
 642	struct lruvec *lruvec = mem_cgroup_lruvec(memcg, NODE_DATA(sc->nid));
 643	unsigned long nr_backing, nr_stored, nr_freeable, nr_protected;
 644
 645	if (!zswap_shrinker_enabled || !mem_cgroup_zswap_writeback_enabled(memcg))
 646		return 0;
 647
 648#ifdef CONFIG_MEMCG_KMEM
 649	mem_cgroup_flush_stats(memcg);
 650	nr_backing = memcg_page_state(memcg, MEMCG_ZSWAP_B) >> PAGE_SHIFT;
 651	nr_stored = memcg_page_state(memcg, MEMCG_ZSWAPPED);
 652#else
 653	/* use pool stats instead of memcg stats */
 654	nr_backing = get_zswap_pool_size(pool) >> PAGE_SHIFT;
 655	nr_stored = atomic_read(&pool->nr_stored);
 656#endif
 657
 658	if (!nr_stored)
 659		return 0;
 
 660
 661	nr_protected =
 662		atomic_long_read(&lruvec->zswap_lruvec_state.nr_zswap_protected);
 663	nr_freeable = list_lru_shrink_count(&pool->list_lru, sc);
 664	/*
 665	 * Subtract the lru size by an estimate of the number of pages
 666	 * that should be protected.
 667	 */
 668	nr_freeable = nr_freeable > nr_protected ? nr_freeable - nr_protected : 0;
 669
 670	/*
 671	 * Scale the number of freeable pages by the memory saving factor.
 672	 * This ensures that the better zswap compresses memory, the fewer
 673	 * pages we will evict to swap (as it will otherwise incur IO for
 674	 * relatively small memory saving).
 675	 */
 676	return mult_frac(nr_freeable, nr_backing, nr_stored);
 677}
 678
 679static void zswap_alloc_shrinker(struct zswap_pool *pool)
 680{
 681	pool->shrinker =
 682		shrinker_alloc(SHRINKER_NUMA_AWARE | SHRINKER_MEMCG_AWARE, "mm-zswap");
 683	if (!pool->shrinker)
 684		return;
 685
 686	pool->shrinker->private_data = pool;
 687	pool->shrinker->scan_objects = zswap_shrinker_scan;
 688	pool->shrinker->count_objects = zswap_shrinker_count;
 689	pool->shrinker->batch = 0;
 690	pool->shrinker->seeks = DEFAULT_SEEKS;
 691}
 692
 693/*********************************
 694* per-cpu code
 695**********************************/
 696static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 697{
 698	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 699	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
 700	struct crypto_acomp *acomp;
 701	struct acomp_req *req;
 702	int ret;
 703
 704	mutex_init(&acomp_ctx->mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 705
 706	acomp_ctx->buffer = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 707	if (!acomp_ctx->buffer)
 708		return -ENOMEM;
 
 
 709
 710	acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
 711	if (IS_ERR(acomp)) {
 712		pr_err("could not alloc crypto acomp %s : %ld\n",
 713				pool->tfm_name, PTR_ERR(acomp));
 714		ret = PTR_ERR(acomp);
 715		goto acomp_fail;
 716	}
 717	acomp_ctx->acomp = acomp;
 718
 719	req = acomp_request_alloc(acomp_ctx->acomp);
 720	if (!req) {
 721		pr_err("could not alloc crypto acomp_request %s\n",
 722		       pool->tfm_name);
 723		ret = -ENOMEM;
 724		goto req_fail;
 725	}
 726	acomp_ctx->req = req;
 727
 728	crypto_init_wait(&acomp_ctx->wait);
 729	/*
 730	 * if the backend of acomp is async zip, crypto_req_done() will wakeup
 731	 * crypto_wait_req(); if the backend of acomp is scomp, the callback
 732	 * won't be called, crypto_wait_req() will return without blocking.
 733	 */
 734	acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 735				   crypto_req_done, &acomp_ctx->wait);
 736
 
 
 
 
 
 
 
 
 
 
 737	return 0;
 738
 739req_fail:
 740	crypto_free_acomp(acomp_ctx->acomp);
 741acomp_fail:
 742	kfree(acomp_ctx->buffer);
 743	return ret;
 744}
 745
 746static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
 747{
 748	struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 749	struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
 750
 751	if (!IS_ERR_OR_NULL(acomp_ctx)) {
 752		if (!IS_ERR_OR_NULL(acomp_ctx->req))
 753			acomp_request_free(acomp_ctx->req);
 754		if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
 755			crypto_free_acomp(acomp_ctx->acomp);
 756		kfree(acomp_ctx->buffer);
 757	}
 758
 759	return 0;
 
 
 
 
 760}
 761
 762/*********************************
 763* pool functions
 764**********************************/
 765
 766static struct zswap_pool *__zswap_pool_current(void)
 767{
 768	struct zswap_pool *pool;
 769
 770	pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 771	WARN_ONCE(!pool && zswap_has_pool,
 772		  "%s: no page storage pool!\n", __func__);
 773
 774	return pool;
 775}
 776
 777static struct zswap_pool *zswap_pool_current(void)
 778{
 779	assert_spin_locked(&zswap_pools_lock);
 780
 781	return __zswap_pool_current();
 782}
 783
 784static struct zswap_pool *zswap_pool_current_get(void)
 785{
 786	struct zswap_pool *pool;
 787
 788	rcu_read_lock();
 789
 790	pool = __zswap_pool_current();
 791	if (!zswap_pool_get(pool))
 792		pool = NULL;
 793
 794	rcu_read_unlock();
 795
 796	return pool;
 797}
 798
 799static struct zswap_pool *zswap_pool_last_get(void)
 800{
 801	struct zswap_pool *pool, *last = NULL;
 802
 803	rcu_read_lock();
 804
 805	list_for_each_entry_rcu(pool, &zswap_pools, list)
 806		last = pool;
 807	WARN_ONCE(!last && zswap_has_pool,
 808		  "%s: no page storage pool!\n", __func__);
 809	if (!zswap_pool_get(last))
 810		last = NULL;
 811
 812	rcu_read_unlock();
 813
 814	return last;
 815}
 816
 817/* type and compressor must be null-terminated */
 818static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 819{
 820	struct zswap_pool *pool;
 821
 822	assert_spin_locked(&zswap_pools_lock);
 823
 824	list_for_each_entry_rcu(pool, &zswap_pools, list) {
 825		if (strcmp(pool->tfm_name, compressor))
 826			continue;
 827		/* all zpools share the same type */
 828		if (strcmp(zpool_get_type(pool->zpools[0]), type))
 829			continue;
 830		/* if we can't get it, it's about to be destroyed */
 831		if (!zswap_pool_get(pool))
 832			continue;
 833		return pool;
 834	}
 835
 836	return NULL;
 837}
 838
 839/*
 840 * If the entry is still valid in the tree, drop the initial ref and remove it
 841 * from the tree. This function must be called with an additional ref held,
 842 * otherwise it may race with another invalidation freeing the entry.
 843 */
 844static void zswap_invalidate_entry(struct zswap_tree *tree,
 845				   struct zswap_entry *entry)
 846{
 847	if (zswap_rb_erase(&tree->rbroot, entry))
 848		zswap_entry_put(tree, entry);
 849}
 850
 851static enum lru_status shrink_memcg_cb(struct list_head *item, struct list_lru_one *l,
 852				       spinlock_t *lock, void *arg)
 853{
 854	struct zswap_entry *entry = container_of(item, struct zswap_entry, lru);
 855	bool *encountered_page_in_swapcache = (bool *)arg;
 856	struct zswap_tree *tree;
 857	pgoff_t swpoffset;
 858	enum lru_status ret = LRU_REMOVED_RETRY;
 859	int writeback_result;
 860
 861	/*
 862	 * Once the lru lock is dropped, the entry might get freed. The
 863	 * swpoffset is copied to the stack, and entry isn't deref'd again
 864	 * until the entry is verified to still be alive in the tree.
 865	 */
 866	swpoffset = swp_offset(entry->swpentry);
 867	tree = zswap_trees[swp_type(entry->swpentry)];
 868	list_lru_isolate(l, item);
 869	/*
 870	 * It's safe to drop the lock here because we return either
 871	 * LRU_REMOVED_RETRY or LRU_RETRY.
 872	 */
 873	spin_unlock(lock);
 874
 875	/* Check for invalidate() race */
 876	spin_lock(&tree->lock);
 877	if (entry != zswap_rb_search(&tree->rbroot, swpoffset))
 878		goto unlock;
 879
 880	/* Hold a reference to prevent a free during writeback */
 881	zswap_entry_get(entry);
 882	spin_unlock(&tree->lock);
 883
 884	writeback_result = zswap_writeback_entry(entry, tree);
 885
 886	spin_lock(&tree->lock);
 887	if (writeback_result) {
 888		zswap_reject_reclaim_fail++;
 889		zswap_lru_putback(&entry->pool->list_lru, entry);
 890		ret = LRU_RETRY;
 891
 892		/*
 893		 * Encountering a page already in swap cache is a sign that we are shrinking
 894		 * into the warmer region. We should terminate shrinking (if we're in the dynamic
 895		 * shrinker context).
 896		 */
 897		if (writeback_result == -EEXIST && encountered_page_in_swapcache)
 898			*encountered_page_in_swapcache = true;
 899
 900		goto put_unlock;
 901	}
 902	zswap_written_back_pages++;
 903
 904	if (entry->objcg)
 905		count_objcg_event(entry->objcg, ZSWPWB);
 906
 907	count_vm_event(ZSWPWB);
 908	/*
 909	 * Writeback started successfully, the page now belongs to the
 910	 * swapcache. Drop the entry from zswap - unless invalidate already
 911	 * took it out while we had the tree->lock released for IO.
 912	 */
 913	zswap_invalidate_entry(tree, entry);
 914
 915put_unlock:
 916	/* Drop local reference */
 917	zswap_entry_put(tree, entry);
 918unlock:
 919	spin_unlock(&tree->lock);
 920	spin_lock(lock);
 921	return ret;
 922}
 923
 924static int shrink_memcg(struct mem_cgroup *memcg)
 925{
 926	struct zswap_pool *pool;
 927	int nid, shrunk = 0;
 928
 929	if (!mem_cgroup_zswap_writeback_enabled(memcg))
 930		return -EINVAL;
 931
 932	/*
 933	 * Skip zombies because their LRUs are reparented and we would be
 934	 * reclaiming from the parent instead of the dead memcg.
 935	 */
 936	if (memcg && !mem_cgroup_online(memcg))
 937		return -ENOENT;
 938
 939	pool = zswap_pool_current_get();
 940	if (!pool)
 941		return -EINVAL;
 942
 943	for_each_node_state(nid, N_NORMAL_MEMORY) {
 944		unsigned long nr_to_walk = 1;
 945
 946		shrunk += list_lru_walk_one(&pool->list_lru, nid, memcg,
 947					    &shrink_memcg_cb, NULL, &nr_to_walk);
 948	}
 949	zswap_pool_put(pool);
 950	return shrunk ? 0 : -EAGAIN;
 951}
 952
 953static void shrink_worker(struct work_struct *w)
 954{
 955	struct zswap_pool *pool = container_of(w, typeof(*pool),
 956						shrink_work);
 957	struct mem_cgroup *memcg;
 958	int ret, failures = 0;
 959
 960	/* global reclaim will select cgroup in a round-robin fashion. */
 961	do {
 962		spin_lock(&zswap_pools_lock);
 963		pool->next_shrink = mem_cgroup_iter(NULL, pool->next_shrink, NULL);
 964		memcg = pool->next_shrink;
 965
 966		/*
 967		 * We need to retry if we have gone through a full round trip, or if we
 968		 * got an offline memcg (or else we risk undoing the effect of the
 969		 * zswap memcg offlining cleanup callback). This is not catastrophic
 970		 * per se, but it will keep the now offlined memcg hostage for a while.
 971		 *
 972		 * Note that if we got an online memcg, we will keep the extra
 973		 * reference in case the original reference obtained by mem_cgroup_iter
 974		 * is dropped by the zswap memcg offlining callback, ensuring that the
 975		 * memcg is not killed when we are reclaiming.
 976		 */
 977		if (!memcg) {
 978			spin_unlock(&zswap_pools_lock);
 979			if (++failures == MAX_RECLAIM_RETRIES)
 980				break;
 981
 982			goto resched;
 983		}
 984
 985		if (!mem_cgroup_tryget_online(memcg)) {
 986			/* drop the reference from mem_cgroup_iter() */
 987			mem_cgroup_iter_break(NULL, memcg);
 988			pool->next_shrink = NULL;
 989			spin_unlock(&zswap_pools_lock);
 990
 991			if (++failures == MAX_RECLAIM_RETRIES)
 992				break;
 993
 994			goto resched;
 995		}
 996		spin_unlock(&zswap_pools_lock);
 997
 998		ret = shrink_memcg(memcg);
 999		/* drop the extra reference */
1000		mem_cgroup_put(memcg);
1001
1002		if (ret == -EINVAL)
1003			break;
1004		if (ret && ++failures == MAX_RECLAIM_RETRIES)
1005			break;
1006
1007resched:
1008		cond_resched();
1009	} while (!zswap_can_accept());
1010	zswap_pool_put(pool);
1011}
1012
1013static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
1014{
1015	int i;
1016	struct zswap_pool *pool;
1017	char name[38]; /* 'zswap' + 32 char (max) num + \0 */
1018	gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1019	int ret;
1020
1021	if (!zswap_has_pool) {
1022		/* if either are unset, pool initialization failed, and we
1023		 * need both params to be set correctly before trying to
1024		 * create a pool.
1025		 */
1026		if (!strcmp(type, ZSWAP_PARAM_UNSET))
1027			return NULL;
1028		if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
1029			return NULL;
1030	}
1031
1032	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1033	if (!pool)
 
1034		return NULL;
1035
1036	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++) {
1037		/* unique name for each pool specifically required by zsmalloc */
1038		snprintf(name, 38, "zswap%x",
1039			 atomic_inc_return(&zswap_pools_count));
1040
1041		pool->zpools[i] = zpool_create_pool(type, name, gfp);
1042		if (!pool->zpools[i]) {
1043			pr_err("%s zpool not available\n", type);
1044			goto error;
1045		}
1046	}
1047	pr_debug("using %s zpool\n", zpool_get_type(pool->zpools[0]));
1048
1049	strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 
1050
1051	pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
1052	if (!pool->acomp_ctx) {
1053		pr_err("percpu alloc failed\n");
1054		goto error;
1055	}
 
1056
1057	ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
1058				       &pool->node);
1059	if (ret)
 
1060		goto error;
 
1061
1062	zswap_alloc_shrinker(pool);
1063	if (!pool->shrinker)
1064		goto error;
1065
1066	pr_debug("using %s compressor\n", pool->tfm_name);
1067
1068	/* being the current pool takes 1 ref; this func expects the
1069	 * caller to always add the new pool as the current pool
1070	 */
1071	kref_init(&pool->kref);
1072	INIT_LIST_HEAD(&pool->list);
1073	if (list_lru_init_memcg(&pool->list_lru, pool->shrinker))
1074		goto lru_fail;
1075	shrinker_register(pool->shrinker);
1076	INIT_WORK(&pool->shrink_work, shrink_worker);
1077	atomic_set(&pool->nr_stored, 0);
1078
1079	zswap_pool_debug("created", pool);
1080
1081	return pool;
1082
1083lru_fail:
1084	list_lru_destroy(&pool->list_lru);
1085	shrinker_free(pool->shrinker);
1086error:
1087	if (pool->acomp_ctx)
1088		free_percpu(pool->acomp_ctx);
1089	while (i--)
1090		zpool_destroy_pool(pool->zpools[i]);
1091	kfree(pool);
1092	return NULL;
1093}
1094
1095static struct zswap_pool *__zswap_pool_create_fallback(void)
1096{
1097	bool has_comp, has_zpool;
1098
1099	has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
1100	if (!has_comp && strcmp(zswap_compressor,
1101				CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
 
1102		pr_err("compressor %s not available, using default %s\n",
1103		       zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
1104		param_free_charp(&zswap_compressor);
1105		zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
1106		has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
1107	}
1108	if (!has_comp) {
1109		pr_err("default compressor %s not available\n",
1110		       zswap_compressor);
1111		param_free_charp(&zswap_compressor);
1112		zswap_compressor = ZSWAP_PARAM_UNSET;
1113	}
1114
1115	has_zpool = zpool_has_pool(zswap_zpool_type);
1116	if (!has_zpool && strcmp(zswap_zpool_type,
1117				 CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
 
 
1118		pr_err("zpool %s not available, using default %s\n",
1119		       zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
1120		param_free_charp(&zswap_zpool_type);
1121		zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
1122		has_zpool = zpool_has_pool(zswap_zpool_type);
1123	}
1124	if (!has_zpool) {
1125		pr_err("default zpool %s not available\n",
1126		       zswap_zpool_type);
1127		param_free_charp(&zswap_zpool_type);
1128		zswap_zpool_type = ZSWAP_PARAM_UNSET;
1129	}
1130
1131	if (!has_comp || !has_zpool)
1132		return NULL;
1133
1134	return zswap_pool_create(zswap_zpool_type, zswap_compressor);
1135}
1136
1137static void zswap_pool_destroy(struct zswap_pool *pool)
1138{
1139	int i;
1140
1141	zswap_pool_debug("destroying", pool);
1142
1143	shrinker_free(pool->shrinker);
1144	cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
1145	free_percpu(pool->acomp_ctx);
1146	list_lru_destroy(&pool->list_lru);
1147
1148	spin_lock(&zswap_pools_lock);
1149	mem_cgroup_iter_break(NULL, pool->next_shrink);
1150	pool->next_shrink = NULL;
1151	spin_unlock(&zswap_pools_lock);
1152
1153	for (i = 0; i < ZSWAP_NR_ZPOOLS; i++)
1154		zpool_destroy_pool(pool->zpools[i]);
1155	kfree(pool);
1156}
1157
1158static int __must_check zswap_pool_get(struct zswap_pool *pool)
1159{
1160	if (!pool)
1161		return 0;
1162
1163	return kref_get_unless_zero(&pool->kref);
1164}
1165
1166static void __zswap_pool_release(struct work_struct *work)
1167{
1168	struct zswap_pool *pool = container_of(work, typeof(*pool),
1169						release_work);
1170
1171	synchronize_rcu();
1172
1173	/* nobody should have been able to get a kref... */
1174	WARN_ON(kref_get_unless_zero(&pool->kref));
1175
1176	/* pool is now off zswap_pools list and has no references. */
1177	zswap_pool_destroy(pool);
1178}
1179
1180static void __zswap_pool_empty(struct kref *kref)
1181{
1182	struct zswap_pool *pool;
1183
1184	pool = container_of(kref, typeof(*pool), kref);
1185
1186	spin_lock(&zswap_pools_lock);
1187
1188	WARN_ON(pool == zswap_pool_current());
1189
1190	list_del_rcu(&pool->list);
1191
1192	INIT_WORK(&pool->release_work, __zswap_pool_release);
1193	schedule_work(&pool->release_work);
1194
1195	spin_unlock(&zswap_pools_lock);
1196}
1197
1198static void zswap_pool_put(struct zswap_pool *pool)
1199{
1200	kref_put(&pool->kref, __zswap_pool_empty);
1201}
1202
1203/*********************************
1204* param callbacks
1205**********************************/
1206
1207static bool zswap_pool_changed(const char *s, const struct kernel_param *kp)
1208{
1209	/* no change required */
1210	if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
1211		return false;
1212	return true;
1213}
1214
1215/* val must be a null-terminated string */
1216static int __zswap_param_set(const char *val, const struct kernel_param *kp,
1217			     char *type, char *compressor)
1218{
1219	struct zswap_pool *pool, *put_pool = NULL;
1220	char *s = strstrip((char *)val);
1221	int ret = 0;
1222	bool new_pool = false;
1223
1224	mutex_lock(&zswap_init_lock);
1225	switch (zswap_init_state) {
1226	case ZSWAP_UNINIT:
1227		/* if this is load-time (pre-init) param setting,
1228		 * don't create a pool; that's done during init.
1229		 */
1230		ret = param_set_charp(s, kp);
1231		break;
1232	case ZSWAP_INIT_SUCCEED:
1233		new_pool = zswap_pool_changed(s, kp);
1234		break;
1235	case ZSWAP_INIT_FAILED:
1236		pr_err("can't set param, initialization failed\n");
1237		ret = -ENODEV;
1238	}
1239	mutex_unlock(&zswap_init_lock);
1240
1241	/* no need to create a new pool, return directly */
1242	if (!new_pool)
1243		return ret;
 
 
1244
1245	if (!type) {
1246		if (!zpool_has_pool(s)) {
1247			pr_err("zpool %s not available\n", s);
1248			return -ENOENT;
1249		}
1250		type = s;
1251	} else if (!compressor) {
1252		if (!crypto_has_acomp(s, 0, 0)) {
1253			pr_err("compressor %s not available\n", s);
1254			return -ENOENT;
1255		}
1256		compressor = s;
1257	} else {
1258		WARN_ON(1);
1259		return -EINVAL;
1260	}
1261
1262	spin_lock(&zswap_pools_lock);
1263
1264	pool = zswap_pool_find_get(type, compressor);
1265	if (pool) {
1266		zswap_pool_debug("using existing", pool);
1267		WARN_ON(pool == zswap_pool_current());
1268		list_del_rcu(&pool->list);
1269	}
1270
1271	spin_unlock(&zswap_pools_lock);
1272
1273	if (!pool)
1274		pool = zswap_pool_create(type, compressor);
 
 
1275
1276	if (pool)
1277		ret = param_set_charp(s, kp);
1278	else
1279		ret = -EINVAL;
1280
1281	spin_lock(&zswap_pools_lock);
1282
1283	if (!ret) {
1284		put_pool = zswap_pool_current();
1285		list_add_rcu(&pool->list, &zswap_pools);
1286		zswap_has_pool = true;
1287	} else if (pool) {
1288		/* add the possibly pre-existing pool to the end of the pools
1289		 * list; if it's new (and empty) then it'll be removed and
1290		 * destroyed by the put after we drop the lock
1291		 */
1292		list_add_tail_rcu(&pool->list, &zswap_pools);
1293		put_pool = pool;
1294	}
1295
1296	spin_unlock(&zswap_pools_lock);
1297
1298	if (!zswap_has_pool && !pool) {
1299		/* if initial pool creation failed, and this pool creation also
1300		 * failed, maybe both compressor and zpool params were bad.
1301		 * Allow changing this param, so pool creation will succeed
1302		 * when the other param is changed. We already verified this
1303		 * param is ok in the zpool_has_pool() or crypto_has_acomp()
1304		 * checks above.
1305		 */
1306		ret = param_set_charp(s, kp);
1307	}
1308
1309	/* drop the ref from either the old current pool,
1310	 * or the new pool we failed to add
1311	 */
1312	if (put_pool)
1313		zswap_pool_put(put_pool);
1314
1315	return ret;
1316}
1317
1318static int zswap_compressor_param_set(const char *val,
1319				      const struct kernel_param *kp)
1320{
1321	return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
1322}
1323
1324static int zswap_zpool_param_set(const char *val,
1325				 const struct kernel_param *kp)
1326{
1327	return __zswap_param_set(val, kp, NULL, zswap_compressor);
1328}
1329
1330static int zswap_enabled_param_set(const char *val,
1331				   const struct kernel_param *kp)
1332{
1333	int ret = -ENODEV;
1334
1335	/* if this is load-time (pre-init) param setting, only set param. */
1336	if (system_state != SYSTEM_RUNNING)
1337		return param_set_bool(val, kp);
1338
1339	mutex_lock(&zswap_init_lock);
1340	switch (zswap_init_state) {
1341	case ZSWAP_UNINIT:
1342		if (zswap_setup())
1343			break;
1344		fallthrough;
1345	case ZSWAP_INIT_SUCCEED:
1346		if (!zswap_has_pool)
1347			pr_err("can't enable, no pool configured\n");
1348		else
1349			ret = param_set_bool(val, kp);
1350		break;
1351	case ZSWAP_INIT_FAILED:
1352		pr_err("can't enable, initialization failed\n");
1353	}
1354	mutex_unlock(&zswap_init_lock);
1355
1356	return ret;
1357}
1358
1359static void __zswap_load(struct zswap_entry *entry, struct page *page)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1360{
1361	struct zpool *zpool = zswap_find_zpool(entry);
1362	struct scatterlist input, output;
1363	struct crypto_acomp_ctx *acomp_ctx;
1364	u8 *src;
1365
1366	acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1367	mutex_lock(&acomp_ctx->mutex);
1368
1369	src = zpool_map_handle(zpool, entry->handle, ZPOOL_MM_RO);
1370	if (!zpool_can_sleep_mapped(zpool)) {
1371		memcpy(acomp_ctx->buffer, src, entry->length);
1372		src = acomp_ctx->buffer;
1373		zpool_unmap_handle(zpool, entry->handle);
1374	}
1375
1376	sg_init_one(&input, src, entry->length);
1377	sg_init_table(&output, 1);
1378	sg_set_page(&output, page, PAGE_SIZE, 0);
1379	acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, PAGE_SIZE);
1380	BUG_ON(crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait));
1381	BUG_ON(acomp_ctx->req->dlen != PAGE_SIZE);
1382	mutex_unlock(&acomp_ctx->mutex);
1383
1384	if (zpool_can_sleep_mapped(zpool))
1385		zpool_unmap_handle(zpool, entry->handle);
 
 
 
 
 
1386}
1387
1388/*********************************
1389* writeback code
1390**********************************/
1391/*
1392 * Attempts to free an entry by adding a folio to the swap cache,
1393 * decompressing the entry data into the folio, and issuing a
1394 * bio write to write the folio back to the swap device.
1395 *
1396 * This can be thought of as a "resumed writeback" of the folio
1397 * to the swap device.  We are basically resuming the same swap
1398 * writeback path that was intercepted with the zswap_store()
1399 * in the first place.  After the folio has been decompressed into
1400 * the swap cache, the compressed version stored by zswap can be
1401 * freed.
1402 */
1403static int zswap_writeback_entry(struct zswap_entry *entry,
1404				 struct zswap_tree *tree)
1405{
1406	swp_entry_t swpentry = entry->swpentry;
1407	struct folio *folio;
1408	struct mempolicy *mpol;
1409	bool folio_was_allocated;
 
 
 
 
 
 
1410	struct writeback_control wbc = {
1411		.sync_mode = WB_SYNC_NONE,
1412	};
1413
1414	/* try to allocate swap cache folio */
1415	mpol = get_task_policy(current);
1416	folio = __read_swap_cache_async(swpentry, GFP_KERNEL, mpol,
1417				NO_INTERLEAVE_INDEX, &folio_was_allocated, true);
1418	if (!folio)
1419		return -ENOMEM;
1420
1421	/*
1422	 * Found an existing folio, we raced with load/swapin. We generally
1423	 * writeback cold folios from zswap, and swapin means the folio just
1424	 * became hot. Skip this folio and let the caller find another one.
1425	 */
1426	if (!folio_was_allocated) {
1427		folio_put(folio);
1428		return -EEXIST;
1429	}
1430
1431	/*
1432	 * folio is locked, and the swapcache is now secured against
1433	 * concurrent swapping to and from the slot. Verify that the
1434	 * swap entry hasn't been invalidated and recycled behind our
1435	 * backs (our zswap_entry reference doesn't prevent that), to
1436	 * avoid overwriting a new swap folio with old compressed data.
1437	 */
1438	spin_lock(&tree->lock);
1439	if (zswap_rb_search(&tree->rbroot, swp_offset(entry->swpentry)) != entry) {
 
 
1440		spin_unlock(&tree->lock);
1441		delete_from_swap_cache(folio);
1442		folio_unlock(folio);
1443		folio_put(folio);
1444		return -ENOMEM;
1445	}
1446	spin_unlock(&tree->lock);
 
 
 
 
 
 
 
1447
1448	__zswap_load(entry, &folio->page);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1449
1450	/* folio is up to date */
1451	folio_mark_uptodate(folio);
 
1452
1453	/* move it to the tail of the inactive list after end_writeback */
1454	folio_set_reclaim(folio);
1455
1456	/* start writeback */
1457	__swap_writepage(folio, &wbc);
1458	folio_put(folio);
1459
1460	return 0;
1461}
1462
1463static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1464{
1465	unsigned long *page;
1466	unsigned long val;
1467	unsigned int pos, last_pos = PAGE_SIZE / sizeof(*page) - 1;
1468
1469	page = (unsigned long *)ptr;
1470	val = page[0];
 
1471
1472	if (val != page[last_pos])
1473		return 0;
 
 
 
 
 
 
 
 
1474
1475	for (pos = 1; pos < last_pos; pos++) {
1476		if (val != page[pos])
1477			return 0;
1478	}
1479
1480	*value = val;
 
 
 
 
 
 
 
 
 
 
1481
1482	return 1;
 
1483}
1484
1485static void zswap_fill_page(void *ptr, unsigned long value)
1486{
1487	unsigned long *page;
 
 
 
 
 
1488
1489	page = (unsigned long *)ptr;
1490	memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
 
 
 
1491}
1492
1493bool zswap_store(struct folio *folio)
 
 
 
 
 
1494{
1495	swp_entry_t swp = folio->swap;
1496	int type = swp_type(swp);
1497	pgoff_t offset = swp_offset(swp);
1498	struct page *page = &folio->page;
1499	struct zswap_tree *tree = zswap_trees[type];
1500	struct zswap_entry *entry, *dupentry;
1501	struct scatterlist input, output;
1502	struct crypto_acomp_ctx *acomp_ctx;
1503	struct obj_cgroup *objcg = NULL;
1504	struct mem_cgroup *memcg = NULL;
1505	struct zswap_pool *pool;
1506	struct zpool *zpool;
1507	unsigned int dlen = PAGE_SIZE;
1508	unsigned long handle, value;
1509	char *buf;
1510	u8 *src, *dst;
1511	gfp_t gfp;
1512	int ret;
1513
1514	VM_WARN_ON_ONCE(!folio_test_locked(folio));
1515	VM_WARN_ON_ONCE(!folio_test_swapcache(folio));
1516
1517	/* Large folios aren't supported */
1518	if (folio_test_large(folio))
1519		return false;
1520
1521	if (!tree)
1522		return false;
1523
1524	/*
1525	 * If this is a duplicate, it must be removed before attempting to store
1526	 * it, otherwise, if the store fails the old page won't be removed from
1527	 * the tree, and it might be written back overriding the new data.
1528	 */
1529	spin_lock(&tree->lock);
1530	dupentry = zswap_rb_search(&tree->rbroot, offset);
1531	if (dupentry) {
1532		zswap_duplicate_entry++;
1533		zswap_invalidate_entry(tree, dupentry);
1534	}
1535	spin_unlock(&tree->lock);
1536
1537	if (!zswap_enabled)
1538		return false;
1539
1540	objcg = get_obj_cgroup_from_folio(folio);
1541	if (objcg && !obj_cgroup_may_zswap(objcg)) {
1542		memcg = get_mem_cgroup_from_objcg(objcg);
1543		if (shrink_memcg(memcg)) {
1544			mem_cgroup_put(memcg);
1545			goto reject;
1546		}
1547		mem_cgroup_put(memcg);
1548	}
1549
1550	/* reclaim space if needed */
1551	if (zswap_is_full()) {
1552		zswap_pool_limit_hit++;
1553		zswap_pool_reached_full = true;
1554		goto shrink;
1555	}
1556
1557	if (zswap_pool_reached_full) {
1558	       if (!zswap_can_accept())
1559			goto shrink;
1560		else
1561			zswap_pool_reached_full = false;
1562	}
1563
1564	/* allocate entry */
1565	entry = zswap_entry_cache_alloc(GFP_KERNEL, page_to_nid(page));
1566	if (!entry) {
1567		zswap_reject_kmemcache_fail++;
 
1568		goto reject;
1569	}
1570
1571	if (zswap_same_filled_pages_enabled) {
1572		src = kmap_local_page(page);
1573		if (zswap_is_page_same_filled(src, &value)) {
1574			kunmap_local(src);
1575			entry->swpentry = swp_entry(type, offset);
1576			entry->length = 0;
1577			entry->value = value;
1578			atomic_inc(&zswap_same_filled_pages);
1579			goto insert_entry;
1580		}
1581		kunmap_local(src);
1582	}
1583
1584	if (!zswap_non_same_filled_pages_enabled)
1585		goto freepage;
1586
1587	/* if entry is successfully added, it keeps the reference */
1588	entry->pool = zswap_pool_current_get();
1589	if (!entry->pool)
 
1590		goto freepage;
1591
1592	if (objcg) {
1593		memcg = get_mem_cgroup_from_objcg(objcg);
1594		if (memcg_list_lru_alloc(memcg, &entry->pool->list_lru, GFP_KERNEL)) {
1595			mem_cgroup_put(memcg);
1596			goto put_pool;
1597		}
1598		mem_cgroup_put(memcg);
1599	}
1600
1601	/* compress */
1602	acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1603
1604	mutex_lock(&acomp_ctx->mutex);
1605
1606	dst = acomp_ctx->buffer;
1607	sg_init_table(&input, 1);
1608	sg_set_page(&input, &folio->page, PAGE_SIZE, 0);
1609
1610	/*
1611	 * We need PAGE_SIZE * 2 here since there maybe over-compression case,
1612	 * and hardware-accelerators may won't check the dst buffer size, so
1613	 * giving the dst buffer with enough length to avoid buffer overflow.
1614	 */
1615	sg_init_one(&output, dst, PAGE_SIZE * 2);
1616	acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1617	/*
1618	 * it maybe looks a little bit silly that we send an asynchronous request,
1619	 * then wait for its completion synchronously. This makes the process look
1620	 * synchronous in fact.
1621	 * Theoretically, acomp supports users send multiple acomp requests in one
1622	 * acomp instance, then get those requests done simultaneously. but in this
1623	 * case, zswap actually does store and load page by page, there is no
1624	 * existing method to send the second page before the first page is done
1625	 * in one thread doing zwap.
1626	 * but in different threads running on different cpu, we have different
1627	 * acomp instance, so multiple threads can do (de)compression in parallel.
1628	 */
1629	ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1630	dlen = acomp_ctx->req->dlen;
1631
1632	if (ret) {
1633		zswap_reject_compress_fail++;
1634		goto put_dstmem;
1635	}
1636
1637	/* store */
1638	zpool = zswap_find_zpool(entry);
1639	gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1640	if (zpool_malloc_support_movable(zpool))
1641		gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1642	ret = zpool_malloc(zpool, dlen, gfp, &handle);
1643	if (ret == -ENOSPC) {
1644		zswap_reject_compress_poor++;
1645		goto put_dstmem;
1646	}
1647	if (ret) {
1648		zswap_reject_alloc_fail++;
1649		goto put_dstmem;
1650	}
1651	buf = zpool_map_handle(zpool, handle, ZPOOL_MM_WO);
 
 
1652	memcpy(buf, dst, dlen);
1653	zpool_unmap_handle(zpool, handle);
1654	mutex_unlock(&acomp_ctx->mutex);
1655
1656	/* populate entry */
1657	entry->swpentry = swp_entry(type, offset);
1658	entry->handle = handle;
1659	entry->length = dlen;
1660
1661insert_entry:
1662	entry->objcg = objcg;
1663	if (objcg) {
1664		obj_cgroup_charge_zswap(objcg, entry->length);
1665		/* Account before objcg ref is moved to tree */
1666		count_objcg_event(objcg, ZSWPOUT);
1667	}
1668
1669	/* map */
1670	spin_lock(&tree->lock);
1671	/*
1672	 * A duplicate entry should have been removed at the beginning of this
1673	 * function. Since the swap entry should be pinned, if a duplicate is
1674	 * found again here it means that something went wrong in the swap
1675	 * cache.
1676	 */
1677	while (zswap_rb_insert(&tree->rbroot, entry, &dupentry) == -EEXIST) {
1678		WARN_ON(1);
1679		zswap_duplicate_entry++;
1680		zswap_invalidate_entry(tree, dupentry);
1681	}
1682	if (entry->length) {
1683		INIT_LIST_HEAD(&entry->lru);
1684		zswap_lru_add(&entry->pool->list_lru, entry);
1685		atomic_inc(&entry->pool->nr_stored);
1686	}
1687	spin_unlock(&tree->lock);
1688
1689	/* update stats */
1690	atomic_inc(&zswap_stored_pages);
1691	zswap_update_total_size();
1692	count_vm_event(ZSWPOUT);
1693
1694	return true;
1695
1696put_dstmem:
1697	mutex_unlock(&acomp_ctx->mutex);
1698put_pool:
1699	zswap_pool_put(entry->pool);
1700freepage:
1701	zswap_entry_cache_free(entry);
1702reject:
1703	if (objcg)
1704		obj_cgroup_put(objcg);
1705	return false;
1706
1707shrink:
1708	pool = zswap_pool_last_get();
1709	if (pool && !queue_work(shrink_wq, &pool->shrink_work))
1710		zswap_pool_put(pool);
1711	goto reject;
1712}
1713
1714bool zswap_load(struct folio *folio)
 
 
 
 
 
1715{
1716	swp_entry_t swp = folio->swap;
1717	int type = swp_type(swp);
1718	pgoff_t offset = swp_offset(swp);
1719	struct page *page = &folio->page;
1720	struct zswap_tree *tree = zswap_trees[type];
1721	struct zswap_entry *entry;
1722	u8 *dst;
1723
1724	VM_WARN_ON_ONCE(!folio_test_locked(folio));
 
1725
1726	/* find */
1727	spin_lock(&tree->lock);
1728	entry = zswap_entry_find_get(&tree->rbroot, offset);
1729	if (!entry) {
 
1730		spin_unlock(&tree->lock);
1731		return false;
1732	}
1733	spin_unlock(&tree->lock);
1734
1735	if (entry->length)
1736		__zswap_load(entry, page);
1737	else {
1738		dst = kmap_local_page(page);
1739		zswap_fill_page(dst, entry->value);
1740		kunmap_local(dst);
1741	}
1742
1743	count_vm_event(ZSWPIN);
1744	if (entry->objcg)
1745		count_objcg_event(entry->objcg, ZSWPIN);
1746
1747	spin_lock(&tree->lock);
1748	if (zswap_exclusive_loads_enabled) {
1749		zswap_invalidate_entry(tree, entry);
1750		folio_mark_dirty(folio);
1751	} else if (entry->length) {
1752		zswap_lru_del(&entry->pool->list_lru, entry);
1753		zswap_lru_add(&entry->pool->list_lru, entry);
1754	}
1755	zswap_entry_put(tree, entry);
1756	spin_unlock(&tree->lock);
1757
1758	return true;
1759}
1760
1761void zswap_invalidate(int type, pgoff_t offset)
 
1762{
1763	struct zswap_tree *tree = zswap_trees[type];
1764	struct zswap_entry *entry;
1765
1766	/* find */
1767	spin_lock(&tree->lock);
1768	entry = zswap_rb_search(&tree->rbroot, offset);
1769	if (!entry) {
1770		/* entry was written back */
1771		spin_unlock(&tree->lock);
1772		return;
1773	}
1774	zswap_invalidate_entry(tree, entry);
1775	spin_unlock(&tree->lock);
1776}
1777
1778void zswap_swapon(int type)
1779{
1780	struct zswap_tree *tree;
1781
1782	tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1783	if (!tree) {
1784		pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1785		return;
1786	}
1787
1788	tree->rbroot = RB_ROOT;
1789	spin_lock_init(&tree->lock);
1790	zswap_trees[type] = tree;
1791}
1792
1793void zswap_swapoff(int type)
 
1794{
1795	struct zswap_tree *tree = zswap_trees[type];
1796	struct zswap_entry *entry, *n;
1797
1798	if (!tree)
1799		return;
1800
1801	/* walk the tree and free everything */
1802	spin_lock(&tree->lock);
1803	rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1804		zswap_free_entry(entry);
1805	tree->rbroot = RB_ROOT;
1806	spin_unlock(&tree->lock);
1807	kfree(tree);
1808	zswap_trees[type] = NULL;
1809}
1810
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1811/*********************************
1812* debugfs functions
1813**********************************/
1814#ifdef CONFIG_DEBUG_FS
1815#include <linux/debugfs.h>
1816
1817static struct dentry *zswap_debugfs_root;
1818
1819static int zswap_debugfs_init(void)
1820{
1821	if (!debugfs_initialized())
1822		return -ENODEV;
1823
1824	zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
 
 
1825
1826	debugfs_create_u64("pool_limit_hit", 0444,
1827			   zswap_debugfs_root, &zswap_pool_limit_hit);
1828	debugfs_create_u64("reject_reclaim_fail", 0444,
1829			   zswap_debugfs_root, &zswap_reject_reclaim_fail);
1830	debugfs_create_u64("reject_alloc_fail", 0444,
1831			   zswap_debugfs_root, &zswap_reject_alloc_fail);
1832	debugfs_create_u64("reject_kmemcache_fail", 0444,
1833			   zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1834	debugfs_create_u64("reject_compress_fail", 0444,
1835			   zswap_debugfs_root, &zswap_reject_compress_fail);
1836	debugfs_create_u64("reject_compress_poor", 0444,
1837			   zswap_debugfs_root, &zswap_reject_compress_poor);
1838	debugfs_create_u64("written_back_pages", 0444,
1839			   zswap_debugfs_root, &zswap_written_back_pages);
1840	debugfs_create_u64("duplicate_entry", 0444,
1841			   zswap_debugfs_root, &zswap_duplicate_entry);
1842	debugfs_create_u64("pool_total_size", 0444,
1843			   zswap_debugfs_root, &zswap_pool_total_size);
1844	debugfs_create_atomic_t("stored_pages", 0444,
1845				zswap_debugfs_root, &zswap_stored_pages);
1846	debugfs_create_atomic_t("same_filled_pages", 0444,
1847				zswap_debugfs_root, &zswap_same_filled_pages);
1848
1849	return 0;
1850}
 
 
 
 
 
1851#else
1852static int zswap_debugfs_init(void)
1853{
1854	return 0;
1855}
 
 
1856#endif
1857
1858/*********************************
1859* module init and exit
1860**********************************/
1861static int zswap_setup(void)
1862{
1863	struct zswap_pool *pool;
1864	int ret;
1865
1866	zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
1867	if (!zswap_entry_cache) {
 
1868		pr_err("entry cache creation failed\n");
1869		goto cache_fail;
1870	}
1871
1872	ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1873				      "mm/zswap_pool:prepare",
1874				      zswap_cpu_comp_prepare,
1875				      zswap_cpu_comp_dead);
1876	if (ret)
1877		goto hp_fail;
1878
1879	pool = __zswap_pool_create_fallback();
1880	if (pool) {
1881		pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1882			zpool_get_type(pool->zpools[0]));
1883		list_add(&pool->list, &zswap_pools);
1884		zswap_has_pool = true;
1885	} else {
1886		pr_err("pool creation failed\n");
1887		zswap_enabled = false;
1888	}
 
 
1889
1890	shrink_wq = create_workqueue("zswap-shrink");
1891	if (!shrink_wq)
1892		goto fallback_fail;
1893
 
1894	if (zswap_debugfs_init())
1895		pr_warn("debugfs initialization failed\n");
1896	zswap_init_state = ZSWAP_INIT_SUCCEED;
1897	return 0;
1898
1899fallback_fail:
1900	if (pool)
1901		zswap_pool_destroy(pool);
1902hp_fail:
1903	kmem_cache_destroy(zswap_entry_cache);
1904cache_fail:
1905	/* if built-in, we aren't unloaded on failure; don't allow use */
1906	zswap_init_state = ZSWAP_INIT_FAILED;
1907	zswap_enabled = false;
1908	return -ENOMEM;
1909}
1910
1911static int __init zswap_init(void)
1912{
1913	if (!zswap_enabled)
1914		return 0;
1915	return zswap_setup();
1916}
1917/* must be late so crypto has time to come up */
1918late_initcall(zswap_init);
1919
 
1920MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1921MODULE_DESCRIPTION("Compressed cache for swap pages");