mm/zswap.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * zswap.c - zswap driver file
   4  *
   5  * zswap is a backend for frontswap that takes pages that are in the process
   6  * of being swapped out and attempts to compress and store them in a
   7  * RAM-based memory pool.  This can result in a significant I/O reduction on
   8  * the swap device and, in the case where decompressing from RAM is faster
   9  * than reading from the swap device, can also improve workload performance.
  10  *
  11  * Copyright (C) 2012  Seth Jennings <sjenning@linux.vnet.ibm.com>
  12 */
  13
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15
  16 #include <linux/module.h>
  17 #include <linux/cpu.h>
  18 #include <linux/highmem.h>
  19 #include <linux/slab.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/types.h>
  22 #include <linux/atomic.h>
  23 #include <linux/frontswap.h>
  24 #include <linux/rbtree.h>
  25 #include <linux/swap.h>
  26 #include <linux/crypto.h>
  27 #include <linux/scatterlist.h>
  28 #include <linux/mempool.h>
  29 #include <linux/zpool.h>
  30 #include <crypto/acompress.h>
  31
  32 #include <linux/mm_types.h>
  33 #include <linux/page-flags.h>
  34 #include <linux/swapops.h>
  35 #include <linux/writeback.h>
  36 #include <linux/pagemap.h>
  37 #include <linux/workqueue.h>
  38
  39 #include "swap.h"
  40
  41 /*********************************
  42 * statistics
  43 **********************************/
  44 /* Total bytes used by the compressed storage */
  45 u64 zswap_pool_total_size;
  46 /* The number of compressed pages currently stored in zswap */
  47 atomic_t zswap_stored_pages = ATOMIC_INIT(0);
  48 /* The number of same-value filled pages currently stored in zswap */
  49 static atomic_t zswap_same_filled_pages = ATOMIC_INIT(0);
  50
  51 /*
  52  * The statistics below are not protected from concurrent access for
  53  * performance reasons so they may not be a 100% accurate.  However,
  54  * they do provide useful information on roughly how many times a
  55  * certain event is occurring.
  56 */
  57
  58 /* Pool limit was hit (see zswap_max_pool_percent) */
  59 static u64 zswap_pool_limit_hit;
  60 /* Pages written back when pool limit was reached */
  61 static u64 zswap_written_back_pages;
  62 /* Store failed due to a reclaim failure after pool limit was reached */
  63 static u64 zswap_reject_reclaim_fail;
  64 /* Compressed page was too big for the allocator to (optimally) store */
  65 static u64 zswap_reject_compress_poor;
  66 /* Store failed because underlying allocator could not get memory */
  67 static u64 zswap_reject_alloc_fail;
  68 /* Store failed because the entry metadata could not be allocated (rare) */
  69 static u64 zswap_reject_kmemcache_fail;
  70 /* Duplicate store was encountered (rare) */
  71 static u64 zswap_duplicate_entry;
  72
  73 /* Shrinker work queue */
  74 static struct workqueue_struct *shrink_wq;
  75 /* Pool limit was hit, we need to calm down */
  76 static bool zswap_pool_reached_full;
  77
  78 /*********************************
  79 * tunables
  80 **********************************/
  81
  82 #define ZSWAP_PARAM_UNSET ""
  83
  84 static int zswap_setup(void);
  85
  86 /* Enable/disable zswap */
  87 static bool zswap_enabled = IS_ENABLED(CONFIG_ZSWAP_DEFAULT_ON);
  88 static int zswap_enabled_param_set(const char *,
  89                                    const struct kernel_param *);
  90 static const struct kernel_param_ops zswap_enabled_param_ops = {
  91         .set =          zswap_enabled_param_set,
  92         .get =          param_get_bool,
  93 };
  94 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
  95
  96 /* Crypto compressor to use */
  97 static char *zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
  98 static int zswap_compressor_param_set(const char *,
  99                                       const struct kernel_param *);
 100 static const struct kernel_param_ops zswap_compressor_param_ops = {
 101         .set =          zswap_compressor_param_set,
 102         .get =          param_get_charp,
 103         .free =         param_free_charp,
 104 };
 105 module_param_cb(compressor, &zswap_compressor_param_ops,
 106                 &zswap_compressor, 0644);
 107
 108 /* Compressed storage zpool to use */
 109 static char *zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
 110 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 111 static const struct kernel_param_ops zswap_zpool_param_ops = {
 112         .set =          zswap_zpool_param_set,
 113         .get =          param_get_charp,
 114         .free =         param_free_charp,
 115 };
 116 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
 117
 118 /* The maximum percentage of memory that the compressed pool can occupy */
 119 static unsigned int zswap_max_pool_percent = 20;
 120 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
 121
 122 /* The threshold for accepting new pages after the max_pool_percent was hit */
 123 static unsigned int zswap_accept_thr_percent = 90; /* of max pool size */
 124 module_param_named(accept_threshold_percent, zswap_accept_thr_percent,
 125                    uint, 0644);
 126
 127 /*
 128  * Enable/disable handling same-value filled pages (enabled by default).
 129  * If disabled every page is considered non-same-value filled.
 130  */
 131 static bool zswap_same_filled_pages_enabled = true;
 132 module_param_named(same_filled_pages_enabled, zswap_same_filled_pages_enabled,
 133                    bool, 0644);
 134
 135 /* Enable/disable handling non-same-value filled pages (enabled by default) */
 136 static bool zswap_non_same_filled_pages_enabled = true;
 137 module_param_named(non_same_filled_pages_enabled, zswap_non_same_filled_pages_enabled,
 138                    bool, 0644);
 139
 140 /*********************************
 141 * data structures
 142 **********************************/
 143
 144 struct crypto_acomp_ctx {
 145         struct crypto_acomp *acomp;
 146         struct acomp_req *req;
 147         struct crypto_wait wait;
 148         u8 *dstmem;
 149         struct mutex *mutex;
 150 };
 151
 152 struct zswap_pool {
 153         struct zpool *zpool;
 154         struct crypto_acomp_ctx __percpu *acomp_ctx;
 155         struct kref kref;
 156         struct list_head list;
 157         struct work_struct release_work;
 158         struct work_struct shrink_work;
 159         struct hlist_node node;
 160         char tfm_name[CRYPTO_MAX_ALG_NAME];
 161 };
 162
 163 /*
 164  * struct zswap_entry
 165  *
 166  * This structure contains the metadata for tracking a single compressed
 167  * page within zswap.
 168  *
 169  * rbnode - links the entry into red-black tree for the appropriate swap type
 170  * offset - the swap offset for the entry.  Index into the red-black tree.
 171  * refcount - the number of outstanding reference to the entry. This is needed
 172  *            to protect against premature freeing of the entry by code
 173  *            concurrent calls to load, invalidate, and writeback.  The lock
 174  *            for the zswap_tree structure that contains the entry must
 175  *            be held while changing the refcount.  Since the lock must
 176  *            be held, there is no reason to also make refcount atomic.
 177  * length - the length in bytes of the compressed page data.  Needed during
 178  *          decompression. For a same value filled page length is 0.
 179  * pool - the zswap_pool the entry's data is in
 180  * handle - zpool allocation handle that stores the compressed page data
 181  * value - value of the same-value filled pages which have same content
 182  */
 183 struct zswap_entry {
 184         struct rb_node rbnode;
 185         pgoff_t offset;
 186         int refcount;
 187         unsigned int length;
 188         struct zswap_pool *pool;
 189         union {
 190                 unsigned long handle;
 191                 unsigned long value;
 192         };
 193         struct obj_cgroup *objcg;
 194 };
 195
 196 struct zswap_header {
 197         swp_entry_t swpentry;
 198 };
 199
 200 /*
 201  * The tree lock in the zswap_tree struct protects a few things:
 202  * - the rbtree
 203  * - the refcount field of each entry in the tree
 204  */
 205 struct zswap_tree {
 206         struct rb_root rbroot;
 207         spinlock_t lock;
 208 };
 209
 210 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 211
 212 /* RCU-protected iteration */
 213 static LIST_HEAD(zswap_pools);
 214 /* protects zswap_pools list modification */
 215 static DEFINE_SPINLOCK(zswap_pools_lock);
 216 /* pool counter to provide unique names to zpool */
 217 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
 218
 219 enum zswap_init_type {
 220         ZSWAP_UNINIT,
 221         ZSWAP_INIT_SUCCEED,
 222         ZSWAP_INIT_FAILED
 223 };
 224
 225 static enum zswap_init_type zswap_init_state;
 226
 227 /* used to ensure the integrity of initialization */
 228 static DEFINE_MUTEX(zswap_init_lock);
 229
 230 /* init completed, but couldn't create the initial pool */
 231 static bool zswap_has_pool;
 232
 233 /*********************************
 234 * helpers and fwd declarations
 235 **********************************/
 236
 237 #define zswap_pool_debug(msg, p)                                \
 238         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 239                  zpool_get_type((p)->zpool))
 240
 241 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 242 static int zswap_pool_get(struct zswap_pool *pool);
 243 static void zswap_pool_put(struct zswap_pool *pool);
 244
 245 static const struct zpool_ops zswap_zpool_ops = {
 246         .evict = zswap_writeback_entry
 247 };
 248
 249 static bool zswap_is_full(void)
 250 {
 251         return totalram_pages() * zswap_max_pool_percent / 100 <
 252                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 253 }
 254
 255 static bool zswap_can_accept(void)
 256 {
 257         return totalram_pages() * zswap_accept_thr_percent / 100 *
 258                                 zswap_max_pool_percent / 100 >
 259                         DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 260 }
 261
 262 static void zswap_update_total_size(void)
 263 {
 264         struct zswap_pool *pool;
 265         u64 total = 0;
 266
 267         rcu_read_lock();
 268
 269         list_for_each_entry_rcu(pool, &zswap_pools, list)
 270                 total += zpool_get_total_size(pool->zpool);
 271
 272         rcu_read_unlock();
 273
 274         zswap_pool_total_size = total;
 275 }
 276
 277 /*********************************
 278 * zswap entry functions
 279 **********************************/
 280 static struct kmem_cache *zswap_entry_cache;
 281
 282 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 283 {
 284         struct zswap_entry *entry;
 285         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 286         if (!entry)
 287                 return NULL;
 288         entry->refcount = 1;
 289         RB_CLEAR_NODE(&entry->rbnode);
 290         return entry;
 291 }
 292
 293 static void zswap_entry_cache_free(struct zswap_entry *entry)
 294 {
 295         kmem_cache_free(zswap_entry_cache, entry);
 296 }
 297
 298 /*********************************
 299 * rbtree functions
 300 **********************************/
 301 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 302 {
 303         struct rb_node *node = root->rb_node;
 304         struct zswap_entry *entry;
 305
 306         while (node) {
 307                 entry = rb_entry(node, struct zswap_entry, rbnode);
 308                 if (entry->offset > offset)
 309                         node = node->rb_left;
 310                 else if (entry->offset < offset)
 311                         node = node->rb_right;
 312                 else
 313                         return entry;
 314         }
 315         return NULL;
 316 }
 317
 318 /*
 319  * In the case that a entry with the same offset is found, a pointer to
 320  * the existing entry is stored in dupentry and the function returns -EEXIST
 321  */
 322 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 323                         struct zswap_entry **dupentry)
 324 {
 325         struct rb_node **link = &root->rb_node, *parent = NULL;
 326         struct zswap_entry *myentry;
 327
 328         while (*link) {
 329                 parent = *link;
 330                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 331                 if (myentry->offset > entry->offset)
 332                         link = &(*link)->rb_left;
 333                 else if (myentry->offset < entry->offset)
 334                         link = &(*link)->rb_right;
 335                 else {
 336                         *dupentry = myentry;
 337                         return -EEXIST;
 338                 }
 339         }
 340         rb_link_node(&entry->rbnode, parent, link);
 341         rb_insert_color(&entry->rbnode, root);
 342         return 0;
 343 }
 344
 345 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 346 {
 347         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 348                 rb_erase(&entry->rbnode, root);
 349                 RB_CLEAR_NODE(&entry->rbnode);
 350         }
 351 }
 352
 353 /*
 354  * Carries out the common pattern of freeing and entry's zpool allocation,
 355  * freeing the entry itself, and decrementing the number of stored pages.
 356  */
 357 static void zswap_free_entry(struct zswap_entry *entry)
 358 {
 359         if (entry->objcg) {
 360                 obj_cgroup_uncharge_zswap(entry->objcg, entry->length);
 361                 obj_cgroup_put(entry->objcg);
 362         }
 363         if (!entry->length)
 364                 atomic_dec(&zswap_same_filled_pages);
 365         else {
 366                 zpool_free(entry->pool->zpool, entry->handle);
 367                 zswap_pool_put(entry->pool);
 368         }
 369         zswap_entry_cache_free(entry);
 370         atomic_dec(&zswap_stored_pages);
 371         zswap_update_total_size();
 372 }
 373
 374 /* caller must hold the tree lock */
 375 static void zswap_entry_get(struct zswap_entry *entry)
 376 {
 377         entry->refcount++;
 378 }
 379
 380 /* caller must hold the tree lock
 381 * remove from the tree and free it, if nobody reference the entry
 382 */
 383 static void zswap_entry_put(struct zswap_tree *tree,
 384                         struct zswap_entry *entry)
 385 {
 386         int refcount = --entry->refcount;
 387
 388         BUG_ON(refcount < 0);
 389         if (refcount == 0) {
 390                 zswap_rb_erase(&tree->rbroot, entry);
 391                 zswap_free_entry(entry);
 392         }
 393 }
 394
 395 /* caller must hold the tree lock */
 396 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 397                                 pgoff_t offset)
 398 {
 399         struct zswap_entry *entry;
 400
 401         entry = zswap_rb_search(root, offset);
 402         if (entry)
 403                 zswap_entry_get(entry);
 404
 405         return entry;
 406 }
 407
 408 /*********************************
 409 * per-cpu code
 410 **********************************/
 411 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 412 /*
 413  * If users dynamically change the zpool type and compressor at runtime, i.e.
 414  * zswap is running, zswap can have more than one zpool on one cpu, but they
 415  * are sharing dtsmem. So we need this mutex to be per-cpu.
 416  */
 417 static DEFINE_PER_CPU(struct mutex *, zswap_mutex);
 418
 419 static int zswap_dstmem_prepare(unsigned int cpu)
 420 {
 421         struct mutex *mutex;
 422         u8 *dst;
 423
 424         dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 425         if (!dst)
 426                 return -ENOMEM;
 427
 428         mutex = kmalloc_node(sizeof(*mutex), GFP_KERNEL, cpu_to_node(cpu));
 429         if (!mutex) {
 430                 kfree(dst);
 431                 return -ENOMEM;
 432         }
 433
 434         mutex_init(mutex);
 435         per_cpu(zswap_dstmem, cpu) = dst;
 436         per_cpu(zswap_mutex, cpu) = mutex;
 437         return 0;
 438 }
 439
 440 static int zswap_dstmem_dead(unsigned int cpu)
 441 {
 442         struct mutex *mutex;
 443         u8 *dst;
 444
 445         mutex = per_cpu(zswap_mutex, cpu);
 446         kfree(mutex);
 447         per_cpu(zswap_mutex, cpu) = NULL;
 448
 449         dst = per_cpu(zswap_dstmem, cpu);
 450         kfree(dst);
 451         per_cpu(zswap_dstmem, cpu) = NULL;
 452
 453         return 0;
 454 }
 455
 456 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 457 {
 458         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 459         struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
 460         struct crypto_acomp *acomp;
 461         struct acomp_req *req;
 462
 463         acomp = crypto_alloc_acomp_node(pool->tfm_name, 0, 0, cpu_to_node(cpu));
 464         if (IS_ERR(acomp)) {
 465                 pr_err("could not alloc crypto acomp %s : %ld\n",
 466                                 pool->tfm_name, PTR_ERR(acomp));
 467                 return PTR_ERR(acomp);
 468         }
 469         acomp_ctx->acomp = acomp;
 470
 471         req = acomp_request_alloc(acomp_ctx->acomp);
 472         if (!req) {
 473                 pr_err("could not alloc crypto acomp_request %s\n",
 474                        pool->tfm_name);
 475                 crypto_free_acomp(acomp_ctx->acomp);
 476                 return -ENOMEM;
 477         }
 478         acomp_ctx->req = req;
 479
 480         crypto_init_wait(&acomp_ctx->wait);
 481         /*
 482          * if the backend of acomp is async zip, crypto_req_done() will wakeup
 483          * crypto_wait_req(); if the backend of acomp is scomp, the callback
 484          * won't be called, crypto_wait_req() will return without blocking.
 485          */
 486         acomp_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 487                                    crypto_req_done, &acomp_ctx->wait);
 488
 489         acomp_ctx->mutex = per_cpu(zswap_mutex, cpu);
 490         acomp_ctx->dstmem = per_cpu(zswap_dstmem, cpu);
 491
 492         return 0;
 493 }
 494
 495 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
 496 {
 497         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 498         struct crypto_acomp_ctx *acomp_ctx = per_cpu_ptr(pool->acomp_ctx, cpu);
 499
 500         if (!IS_ERR_OR_NULL(acomp_ctx)) {
 501                 if (!IS_ERR_OR_NULL(acomp_ctx->req))
 502                         acomp_request_free(acomp_ctx->req);
 503                 if (!IS_ERR_OR_NULL(acomp_ctx->acomp))
 504                         crypto_free_acomp(acomp_ctx->acomp);
 505         }
 506
 507         return 0;
 508 }
 509
 510 /*********************************
 511 * pool functions
 512 **********************************/
 513
 514 static struct zswap_pool *__zswap_pool_current(void)
 515 {
 516         struct zswap_pool *pool;
 517
 518         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 519         WARN_ONCE(!pool && zswap_has_pool,
 520                   "%s: no page storage pool!\n", __func__);
 521
 522         return pool;
 523 }
 524
 525 static struct zswap_pool *zswap_pool_current(void)
 526 {
 527         assert_spin_locked(&zswap_pools_lock);
 528
 529         return __zswap_pool_current();
 530 }
 531
 532 static struct zswap_pool *zswap_pool_current_get(void)
 533 {
 534         struct zswap_pool *pool;
 535
 536         rcu_read_lock();
 537
 538         pool = __zswap_pool_current();
 539         if (!zswap_pool_get(pool))
 540                 pool = NULL;
 541
 542         rcu_read_unlock();
 543
 544         return pool;
 545 }
 546
 547 static struct zswap_pool *zswap_pool_last_get(void)
 548 {
 549         struct zswap_pool *pool, *last = NULL;
 550
 551         rcu_read_lock();
 552
 553         list_for_each_entry_rcu(pool, &zswap_pools, list)
 554                 last = pool;
 555         WARN_ONCE(!last && zswap_has_pool,
 556                   "%s: no page storage pool!\n", __func__);
 557         if (!zswap_pool_get(last))
 558                 last = NULL;
 559
 560         rcu_read_unlock();
 561
 562         return last;
 563 }
 564
 565 /* type and compressor must be null-terminated */
 566 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 567 {
 568         struct zswap_pool *pool;
 569
 570         assert_spin_locked(&zswap_pools_lock);
 571
 572         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 573                 if (strcmp(pool->tfm_name, compressor))
 574                         continue;
 575                 if (strcmp(zpool_get_type(pool->zpool), type))
 576                         continue;
 577                 /* if we can't get it, it's about to be destroyed */
 578                 if (!zswap_pool_get(pool))
 579                         continue;
 580                 return pool;
 581         }
 582
 583         return NULL;
 584 }
 585
 586 static void shrink_worker(struct work_struct *w)
 587 {
 588         struct zswap_pool *pool = container_of(w, typeof(*pool),
 589                                                 shrink_work);
 590
 591         if (zpool_shrink(pool->zpool, 1, NULL))
 592                 zswap_reject_reclaim_fail++;
 593         zswap_pool_put(pool);
 594 }
 595
 596 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 597 {
 598         struct zswap_pool *pool;
 599         char name[38]; /* 'zswap' + 32 char (max) num + \0 */
 600         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
 601         int ret;
 602
 603         if (!zswap_has_pool) {
 604                 /* if either are unset, pool initialization failed, and we
 605                  * need both params to be set correctly before trying to
 606                  * create a pool.
 607                  */
 608                 if (!strcmp(type, ZSWAP_PARAM_UNSET))
 609                         return NULL;
 610                 if (!strcmp(compressor, ZSWAP_PARAM_UNSET))
 611                         return NULL;
 612         }
 613
 614         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 615         if (!pool)
 616                 return NULL;
 617
 618         /* unique name for each pool specifically required by zsmalloc */
 619         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
 620
 621         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
 622         if (!pool->zpool) {
 623                 pr_err("%s zpool not available\n", type);
 624                 goto error;
 625         }
 626         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 627
 628         strscpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 629
 630         pool->acomp_ctx = alloc_percpu(*pool->acomp_ctx);
 631         if (!pool->acomp_ctx) {
 632                 pr_err("percpu alloc failed\n");
 633                 goto error;
 634         }
 635
 636         ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
 637                                        &pool->node);
 638         if (ret)
 639                 goto error;
 640         pr_debug("using %s compressor\n", pool->tfm_name);
 641
 642         /* being the current pool takes 1 ref; this func expects the
 643          * caller to always add the new pool as the current pool
 644          */
 645         kref_init(&pool->kref);
 646         INIT_LIST_HEAD(&pool->list);
 647         INIT_WORK(&pool->shrink_work, shrink_worker);
 648
 649         zswap_pool_debug("created", pool);
 650
 651         return pool;
 652
 653 error:
 654         if (pool->acomp_ctx)
 655                 free_percpu(pool->acomp_ctx);
 656         if (pool->zpool)
 657                 zpool_destroy_pool(pool->zpool);
 658         kfree(pool);
 659         return NULL;
 660 }
 661
 662 static struct zswap_pool *__zswap_pool_create_fallback(void)
 663 {
 664         bool has_comp, has_zpool;
 665
 666         has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
 667         if (!has_comp && strcmp(zswap_compressor,
 668                                 CONFIG_ZSWAP_COMPRESSOR_DEFAULT)) {
 669                 pr_err("compressor %s not available, using default %s\n",
 670                        zswap_compressor, CONFIG_ZSWAP_COMPRESSOR_DEFAULT);
 671                 param_free_charp(&zswap_compressor);
 672                 zswap_compressor = CONFIG_ZSWAP_COMPRESSOR_DEFAULT;
 673                 has_comp = crypto_has_acomp(zswap_compressor, 0, 0);
 674         }
 675         if (!has_comp) {
 676                 pr_err("default compressor %s not available\n",
 677                        zswap_compressor);
 678                 param_free_charp(&zswap_compressor);
 679                 zswap_compressor = ZSWAP_PARAM_UNSET;
 680         }
 681
 682         has_zpool = zpool_has_pool(zswap_zpool_type);
 683         if (!has_zpool && strcmp(zswap_zpool_type,
 684                                  CONFIG_ZSWAP_ZPOOL_DEFAULT)) {
 685                 pr_err("zpool %s not available, using default %s\n",
 686                        zswap_zpool_type, CONFIG_ZSWAP_ZPOOL_DEFAULT);
 687                 param_free_charp(&zswap_zpool_type);
 688                 zswap_zpool_type = CONFIG_ZSWAP_ZPOOL_DEFAULT;
 689                 has_zpool = zpool_has_pool(zswap_zpool_type);
 690         }
 691         if (!has_zpool) {
 692                 pr_err("default zpool %s not available\n",
 693                        zswap_zpool_type);
 694                 param_free_charp(&zswap_zpool_type);
 695                 zswap_zpool_type = ZSWAP_PARAM_UNSET;
 696         }
 697
 698         if (!has_comp || !has_zpool)
 699                 return NULL;
 700
 701         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 702 }
 703
 704 static void zswap_pool_destroy(struct zswap_pool *pool)
 705 {
 706         zswap_pool_debug("destroying", pool);
 707
 708         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
 709         free_percpu(pool->acomp_ctx);
 710         zpool_destroy_pool(pool->zpool);
 711         kfree(pool);
 712 }
 713
 714 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 715 {
 716         if (!pool)
 717                 return 0;
 718
 719         return kref_get_unless_zero(&pool->kref);
 720 }
 721
 722 static void __zswap_pool_release(struct work_struct *work)
 723 {
 724         struct zswap_pool *pool = container_of(work, typeof(*pool),
 725                                                 release_work);
 726
 727         synchronize_rcu();
 728
 729         /* nobody should have been able to get a kref... */
 730         WARN_ON(kref_get_unless_zero(&pool->kref));
 731
 732         /* pool is now off zswap_pools list and has no references. */
 733         zswap_pool_destroy(pool);
 734 }
 735
 736 static void __zswap_pool_empty(struct kref *kref)
 737 {
 738         struct zswap_pool *pool;
 739
 740         pool = container_of(kref, typeof(*pool), kref);
 741
 742         spin_lock(&zswap_pools_lock);
 743
 744         WARN_ON(pool == zswap_pool_current());
 745
 746         list_del_rcu(&pool->list);
 747
 748         INIT_WORK(&pool->release_work, __zswap_pool_release);
 749         schedule_work(&pool->release_work);
 750
 751         spin_unlock(&zswap_pools_lock);
 752 }
 753
 754 static void zswap_pool_put(struct zswap_pool *pool)
 755 {
 756         kref_put(&pool->kref, __zswap_pool_empty);
 757 }
 758
 759 /*********************************
 760 * param callbacks
 761 **********************************/
 762
 763 static bool zswap_pool_changed(const char *s, const struct kernel_param *kp)
 764 {
 765         /* no change required */
 766         if (!strcmp(s, *(char **)kp->arg) && zswap_has_pool)
 767                 return false;
 768         return true;
 769 }
 770
 771 /* val must be a null-terminated string */
 772 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 773                              char *type, char *compressor)
 774 {
 775         struct zswap_pool *pool, *put_pool = NULL;
 776         char *s = strstrip((char *)val);
 777         int ret = 0;
 778         bool new_pool = false;
 779
 780         mutex_lock(&zswap_init_lock);
 781         switch (zswap_init_state) {
 782         case ZSWAP_UNINIT:
 783                 /* if this is load-time (pre-init) param setting,
 784                  * don't create a pool; that's done during init.
 785                  */
 786                 ret = param_set_charp(s, kp);
 787                 break;
 788         case ZSWAP_INIT_SUCCEED:
 789                 new_pool = zswap_pool_changed(s, kp);
 790                 break;
 791         case ZSWAP_INIT_FAILED:
 792                 pr_err("can't set param, initialization failed\n");
 793                 ret = -ENODEV;
 794         }
 795         mutex_unlock(&zswap_init_lock);
 796
 797         /* no need to create a new pool, return directly */
 798         if (!new_pool)
 799                 return ret;
 800
 801         if (!type) {
 802                 if (!zpool_has_pool(s)) {
 803                         pr_err("zpool %s not available\n", s);
 804                         return -ENOENT;
 805                 }
 806                 type = s;
 807         } else if (!compressor) {
 808                 if (!crypto_has_acomp(s, 0, 0)) {
 809                         pr_err("compressor %s not available\n", s);
 810                         return -ENOENT;
 811                 }
 812                 compressor = s;
 813         } else {
 814                 WARN_ON(1);
 815                 return -EINVAL;
 816         }
 817
 818         spin_lock(&zswap_pools_lock);
 819
 820         pool = zswap_pool_find_get(type, compressor);
 821         if (pool) {
 822                 zswap_pool_debug("using existing", pool);
 823                 WARN_ON(pool == zswap_pool_current());
 824                 list_del_rcu(&pool->list);
 825         }
 826
 827         spin_unlock(&zswap_pools_lock);
 828
 829         if (!pool)
 830                 pool = zswap_pool_create(type, compressor);
 831
 832         if (pool)
 833                 ret = param_set_charp(s, kp);
 834         else
 835                 ret = -EINVAL;
 836
 837         spin_lock(&zswap_pools_lock);
 838
 839         if (!ret) {
 840                 put_pool = zswap_pool_current();
 841                 list_add_rcu(&pool->list, &zswap_pools);
 842                 zswap_has_pool = true;
 843         } else if (pool) {
 844                 /* add the possibly pre-existing pool to the end of the pools
 845                  * list; if it's new (and empty) then it'll be removed and
 846                  * destroyed by the put after we drop the lock
 847                  */
 848                 list_add_tail_rcu(&pool->list, &zswap_pools);
 849                 put_pool = pool;
 850         }
 851
 852         spin_unlock(&zswap_pools_lock);
 853
 854         if (!zswap_has_pool && !pool) {
 855                 /* if initial pool creation failed, and this pool creation also
 856                  * failed, maybe both compressor and zpool params were bad.
 857                  * Allow changing this param, so pool creation will succeed
 858                  * when the other param is changed. We already verified this
 859                  * param is ok in the zpool_has_pool() or crypto_has_acomp()
 860                  * checks above.
 861                  */
 862                 ret = param_set_charp(s, kp);
 863         }
 864
 865         /* drop the ref from either the old current pool,
 866          * or the new pool we failed to add
 867          */
 868         if (put_pool)
 869                 zswap_pool_put(put_pool);
 870
 871         return ret;
 872 }
 873
 874 static int zswap_compressor_param_set(const char *val,
 875                                       const struct kernel_param *kp)
 876 {
 877         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
 878 }
 879
 880 static int zswap_zpool_param_set(const char *val,
 881                                  const struct kernel_param *kp)
 882 {
 883         return __zswap_param_set(val, kp, NULL, zswap_compressor);
 884 }
 885
 886 static int zswap_enabled_param_set(const char *val,
 887                                    const struct kernel_param *kp)
 888 {
 889         int ret = -ENODEV;
 890
 891         /* if this is load-time (pre-init) param setting, only set param. */
 892         if (system_state != SYSTEM_RUNNING)
 893                 return param_set_bool(val, kp);
 894
 895         mutex_lock(&zswap_init_lock);
 896         switch (zswap_init_state) {
 897         case ZSWAP_UNINIT:
 898                 if (zswap_setup())
 899                         break;
 900                 fallthrough;
 901         case ZSWAP_INIT_SUCCEED:
 902                 if (!zswap_has_pool)
 903                         pr_err("can't enable, no pool configured\n");
 904                 else
 905                         ret = param_set_bool(val, kp);
 906                 break;
 907         case ZSWAP_INIT_FAILED:
 908                 pr_err("can't enable, initialization failed\n");
 909         }
 910         mutex_unlock(&zswap_init_lock);
 911
 912         return ret;
 913 }
 914
 915 /*********************************
 916 * writeback code
 917 **********************************/
 918 /* return enum for zswap_get_swap_cache_page */
 919 enum zswap_get_swap_ret {
 920         ZSWAP_SWAPCACHE_NEW,
 921         ZSWAP_SWAPCACHE_EXIST,
 922         ZSWAP_SWAPCACHE_FAIL,
 923 };
 924
 925 /*
 926  * zswap_get_swap_cache_page
 927  *
 928  * This is an adaption of read_swap_cache_async()
 929  *
 930  * This function tries to find a page with the given swap entry
 931  * in the swapper_space address space (the swap cache).  If the page
 932  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 933  * added to the swap cache, and returned in retpage.
 934  *
 935  * If success, the swap cache page is returned in retpage
 936  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 937  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 938  *     the new page is added to swapcache and locked
 939  * Returns ZSWAP_SWAPCACHE_FAIL on error
 940  */
 941 static int zswap_get_swap_cache_page(swp_entry_t entry,
 942                                 struct page **retpage)
 943 {
 944         bool page_was_allocated;
 945
 946         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 947                         NULL, 0, &page_was_allocated);
 948         if (page_was_allocated)
 949                 return ZSWAP_SWAPCACHE_NEW;
 950         if (!*retpage)
 951                 return ZSWAP_SWAPCACHE_FAIL;
 952         return ZSWAP_SWAPCACHE_EXIST;
 953 }
 954
 955 /*
 956  * Attempts to free an entry by adding a page to the swap cache,
 957  * decompressing the entry data into the page, and issuing a
 958  * bio write to write the page back to the swap device.
 959  *
 960  * This can be thought of as a "resumed writeback" of the page
 961  * to the swap device.  We are basically resuming the same swap
 962  * writeback path that was intercepted with the frontswap_store()
 963  * in the first place.  After the page has been decompressed into
 964  * the swap cache, the compressed version stored by zswap can be
 965  * freed.
 966  */
 967 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 968 {
 969         struct zswap_header *zhdr;
 970         swp_entry_t swpentry;
 971         struct zswap_tree *tree;
 972         pgoff_t offset;
 973         struct zswap_entry *entry;
 974         struct page *page;
 975         struct scatterlist input, output;
 976         struct crypto_acomp_ctx *acomp_ctx;
 977
 978         u8 *src, *tmp = NULL;
 979         unsigned int dlen;
 980         int ret;
 981         struct writeback_control wbc = {
 982                 .sync_mode = WB_SYNC_NONE,
 983         };
 984
 985         if (!zpool_can_sleep_mapped(pool)) {
 986                 tmp = kmalloc(PAGE_SIZE, GFP_KERNEL);
 987                 if (!tmp)
 988                         return -ENOMEM;
 989         }
 990
 991         /* extract swpentry from data */
 992         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 993         swpentry = zhdr->swpentry; /* here */
 994         tree = zswap_trees[swp_type(swpentry)];
 995         offset = swp_offset(swpentry);
 996         zpool_unmap_handle(pool, handle);
 997
 998         /* find and ref zswap entry */
 999         spin_lock(&tree->lock);
1000         entry = zswap_entry_find_get(&tree->rbroot, offset);
1001         if (!entry) {
1002                 /* entry was invalidated */
1003                 spin_unlock(&tree->lock);
1004                 kfree(tmp);
1005                 return 0;
1006         }
1007         spin_unlock(&tree->lock);
1008         BUG_ON(offset != entry->offset);
1009
1010         /* try to allocate swap cache page */
1011         switch (zswap_get_swap_cache_page(swpentry, &page)) {
1012         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
1013                 ret = -ENOMEM;
1014                 goto fail;
1015
1016         case ZSWAP_SWAPCACHE_EXIST:
1017                 /* page is already in the swap cache, ignore for now */
1018                 put_page(page);
1019                 ret = -EEXIST;
1020                 goto fail;
1021
1022         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
1023                 /*
1024                  * Having a local reference to the zswap entry doesn't exclude
1025                  * swapping from invalidating and recycling the swap slot. Once
1026                  * the swapcache is secured against concurrent swapping to and
1027                  * from the slot, recheck that the entry is still current before
1028                  * writing.
1029                  */
1030                 spin_lock(&tree->lock);
1031                 if (zswap_rb_search(&tree->rbroot, entry->offset) != entry) {
1032                         spin_unlock(&tree->lock);
1033                         delete_from_swap_cache(page_folio(page));
1034                         ret = -ENOMEM;
1035                         goto fail;
1036                 }
1037                 spin_unlock(&tree->lock);
1038
1039                 /* decompress */
1040                 acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1041                 dlen = PAGE_SIZE;
1042
1043                 zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
1044                 src = (u8 *)zhdr + sizeof(struct zswap_header);
1045                 if (!zpool_can_sleep_mapped(pool)) {
1046                         memcpy(tmp, src, entry->length);
1047                         src = tmp;
1048                         zpool_unmap_handle(pool, handle);
1049                 }
1050
1051                 mutex_lock(acomp_ctx->mutex);
1052                 sg_init_one(&input, src, entry->length);
1053                 sg_init_table(&output, 1);
1054                 sg_set_page(&output, page, PAGE_SIZE, 0);
1055                 acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1056                 ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1057                 dlen = acomp_ctx->req->dlen;
1058                 mutex_unlock(acomp_ctx->mutex);
1059
1060                 if (!zpool_can_sleep_mapped(pool))
1061                         kfree(tmp);
1062                 else
1063                         zpool_unmap_handle(pool, handle);
1064
1065                 BUG_ON(ret);
1066                 BUG_ON(dlen != PAGE_SIZE);
1067
1068                 /* page is up to date */
1069                 SetPageUptodate(page);
1070         }
1071
1072         /* move it to the tail of the inactive list after end_writeback */
1073         SetPageReclaim(page);
1074
1075         /* start writeback */
1076         __swap_writepage(page, &wbc);
1077         put_page(page);
1078         zswap_written_back_pages++;
1079
1080         spin_lock(&tree->lock);
1081         /* drop local reference */
1082         zswap_entry_put(tree, entry);
1083
1084         /*
1085         * There are two possible situations for entry here:
1086         * (1) refcount is 1(normal case),  entry is valid and on the tree
1087         * (2) refcount is 0, entry is freed and not on the tree
1088         *     because invalidate happened during writeback
1089         *  search the tree and free the entry if find entry
1090         */
1091         if (entry == zswap_rb_search(&tree->rbroot, offset))
1092                 zswap_entry_put(tree, entry);
1093         spin_unlock(&tree->lock);
1094
1095         return ret;
1096
1097 fail:
1098         if (!zpool_can_sleep_mapped(pool))
1099                 kfree(tmp);
1100
1101         /*
1102         * if we get here due to ZSWAP_SWAPCACHE_EXIST
1103         * a load may be happening concurrently.
1104         * it is safe and okay to not free the entry.
1105         * if we free the entry in the following put
1106         * it is also okay to return !0
1107         */
1108         spin_lock(&tree->lock);
1109         zswap_entry_put(tree, entry);
1110         spin_unlock(&tree->lock);
1111
1112         return ret;
1113 }
1114
1115 static int zswap_is_page_same_filled(void *ptr, unsigned long *value)
1116 {
1117         unsigned long *page;
1118         unsigned long val;
1119         unsigned int pos, last_pos = PAGE_SIZE / sizeof(*page) - 1;
1120
1121         page = (unsigned long *)ptr;
1122         val = page[0];
1123
1124         if (val != page[last_pos])
1125                 return 0;
1126
1127         for (pos = 1; pos < last_pos; pos++) {
1128                 if (val != page[pos])
1129                         return 0;
1130         }
1131
1132         *value = val;
1133
1134         return 1;
1135 }
1136
1137 static void zswap_fill_page(void *ptr, unsigned long value)
1138 {
1139         unsigned long *page;
1140
1141         page = (unsigned long *)ptr;
1142         memset_l(page, value, PAGE_SIZE / sizeof(unsigned long));
1143 }
1144
1145 /*********************************
1146 * frontswap hooks
1147 **********************************/
1148 /* attempts to compress and store an single page */
1149 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
1150                                 struct page *page)
1151 {
1152         struct zswap_tree *tree = zswap_trees[type];
1153         struct zswap_entry *entry, *dupentry;
1154         struct scatterlist input, output;
1155         struct crypto_acomp_ctx *acomp_ctx;
1156         struct obj_cgroup *objcg = NULL;
1157         struct zswap_pool *pool;
1158         int ret;
1159         unsigned int hlen, dlen = PAGE_SIZE;
1160         unsigned long handle, value;
1161         char *buf;
1162         u8 *src, *dst;
1163         struct zswap_header zhdr = { .swpentry = swp_entry(type, offset) };
1164         gfp_t gfp;
1165
1166         /* THP isn't supported */
1167         if (PageTransHuge(page)) {
1168                 ret = -EINVAL;
1169                 goto reject;
1170         }
1171
1172         if (!zswap_enabled || !tree) {
1173                 ret = -ENODEV;
1174                 goto reject;
1175         }
1176
1177         /*
1178          * XXX: zswap reclaim does not work with cgroups yet. Without a
1179          * cgroup-aware entry LRU, we will push out entries system-wide based on
1180          * local cgroup limits.
1181          */
1182         objcg = get_obj_cgroup_from_page(page);
1183         if (objcg && !obj_cgroup_may_zswap(objcg)) {
1184                 ret = -ENOMEM;
1185                 goto reject;
1186         }
1187
1188         /* reclaim space if needed */
1189         if (zswap_is_full()) {
1190                 zswap_pool_limit_hit++;
1191                 zswap_pool_reached_full = true;
1192                 goto shrink;
1193         }
1194
1195         if (zswap_pool_reached_full) {
1196                if (!zswap_can_accept()) {
1197                         ret = -ENOMEM;
1198                         goto reject;
1199                 } else
1200                         zswap_pool_reached_full = false;
1201         }
1202
1203         /* allocate entry */
1204         entry = zswap_entry_cache_alloc(GFP_KERNEL);
1205         if (!entry) {
1206                 zswap_reject_kmemcache_fail++;
1207                 ret = -ENOMEM;
1208                 goto reject;
1209         }
1210
1211         if (zswap_same_filled_pages_enabled) {
1212                 src = kmap_atomic(page);
1213                 if (zswap_is_page_same_filled(src, &value)) {
1214                         kunmap_atomic(src);
1215                         entry->offset = offset;
1216                         entry->length = 0;
1217                         entry->value = value;
1218                         atomic_inc(&zswap_same_filled_pages);
1219                         goto insert_entry;
1220                 }
1221                 kunmap_atomic(src);
1222         }
1223
1224         if (!zswap_non_same_filled_pages_enabled) {
1225                 ret = -EINVAL;
1226                 goto freepage;
1227         }
1228
1229         /* if entry is successfully added, it keeps the reference */
1230         entry->pool = zswap_pool_current_get();
1231         if (!entry->pool) {
1232                 ret = -EINVAL;
1233                 goto freepage;
1234         }
1235
1236         /* compress */
1237         acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1238
1239         mutex_lock(acomp_ctx->mutex);
1240
1241         dst = acomp_ctx->dstmem;
1242         sg_init_table(&input, 1);
1243         sg_set_page(&input, page, PAGE_SIZE, 0);
1244
1245         /* zswap_dstmem is of size (PAGE_SIZE * 2). Reflect same in sg_list */
1246         sg_init_one(&output, dst, PAGE_SIZE * 2);
1247         acomp_request_set_params(acomp_ctx->req, &input, &output, PAGE_SIZE, dlen);
1248         /*
1249          * it maybe looks a little bit silly that we send an asynchronous request,
1250          * then wait for its completion synchronously. This makes the process look
1251          * synchronous in fact.
1252          * Theoretically, acomp supports users send multiple acomp requests in one
1253          * acomp instance, then get those requests done simultaneously. but in this
1254          * case, frontswap actually does store and load page by page, there is no
1255          * existing method to send the second page before the first page is done
1256          * in one thread doing frontswap.
1257          * but in different threads running on different cpu, we have different
1258          * acomp instance, so multiple threads can do (de)compression in parallel.
1259          */
1260         ret = crypto_wait_req(crypto_acomp_compress(acomp_ctx->req), &acomp_ctx->wait);
1261         dlen = acomp_ctx->req->dlen;
1262
1263         if (ret) {
1264                 ret = -EINVAL;
1265                 goto put_dstmem;
1266         }
1267
1268         /* store */
1269         hlen = zpool_evictable(entry->pool->zpool) ? sizeof(zhdr) : 0;
1270         gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
1271         if (zpool_malloc_support_movable(entry->pool->zpool))
1272                 gfp |= __GFP_HIGHMEM | __GFP_MOVABLE;
1273         ret = zpool_malloc(entry->pool->zpool, hlen + dlen, gfp, &handle);
1274         if (ret == -ENOSPC) {
1275                 zswap_reject_compress_poor++;
1276                 goto put_dstmem;
1277         }
1278         if (ret) {
1279                 zswap_reject_alloc_fail++;
1280                 goto put_dstmem;
1281         }
1282         buf = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_WO);
1283         memcpy(buf, &zhdr, hlen);
1284         memcpy(buf + hlen, dst, dlen);
1285         zpool_unmap_handle(entry->pool->zpool, handle);
1286         mutex_unlock(acomp_ctx->mutex);
1287
1288         /* populate entry */
1289         entry->offset = offset;
1290         entry->handle = handle;
1291         entry->length = dlen;
1292
1293 insert_entry:
1294         entry->objcg = objcg;
1295         if (objcg) {
1296                 obj_cgroup_charge_zswap(objcg, entry->length);
1297                 /* Account before objcg ref is moved to tree */
1298                 count_objcg_event(objcg, ZSWPOUT);
1299         }
1300
1301         /* map */
1302         spin_lock(&tree->lock);
1303         do {
1304                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1305                 if (ret == -EEXIST) {
1306                         zswap_duplicate_entry++;
1307                         /* remove from rbtree */
1308                         zswap_rb_erase(&tree->rbroot, dupentry);
1309                         zswap_entry_put(tree, dupentry);
1310                 }
1311         } while (ret == -EEXIST);
1312         spin_unlock(&tree->lock);
1313
1314         /* update stats */
1315         atomic_inc(&zswap_stored_pages);
1316         zswap_update_total_size();
1317         count_vm_event(ZSWPOUT);
1318
1319         return 0;
1320
1321 put_dstmem:
1322         mutex_unlock(acomp_ctx->mutex);
1323         zswap_pool_put(entry->pool);
1324 freepage:
1325         zswap_entry_cache_free(entry);
1326 reject:
1327         if (objcg)
1328                 obj_cgroup_put(objcg);
1329         return ret;
1330
1331 shrink:
1332         pool = zswap_pool_last_get();
1333         if (pool)
1334                 queue_work(shrink_wq, &pool->shrink_work);
1335         ret = -ENOMEM;
1336         goto reject;
1337 }
1338
1339 /*
1340  * returns 0 if the page was successfully decompressed
1341  * return -1 on entry not found or error
1342 */
1343 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1344                                 struct page *page)
1345 {
1346         struct zswap_tree *tree = zswap_trees[type];
1347         struct zswap_entry *entry;
1348         struct scatterlist input, output;
1349         struct crypto_acomp_ctx *acomp_ctx;
1350         u8 *src, *dst, *tmp;
1351         unsigned int dlen;
1352         int ret;
1353
1354         /* find */
1355         spin_lock(&tree->lock);
1356         entry = zswap_entry_find_get(&tree->rbroot, offset);
1357         if (!entry) {
1358                 /* entry was written back */
1359                 spin_unlock(&tree->lock);
1360                 return -1;
1361         }
1362         spin_unlock(&tree->lock);
1363
1364         if (!entry->length) {
1365                 dst = kmap_atomic(page);
1366                 zswap_fill_page(dst, entry->value);
1367                 kunmap_atomic(dst);
1368                 ret = 0;
1369                 goto stats;
1370         }
1371
1372         if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1373                 tmp = kmalloc(entry->length, GFP_KERNEL);
1374                 if (!tmp) {
1375                         ret = -ENOMEM;
1376                         goto freeentry;
1377                 }
1378         }
1379
1380         /* decompress */
1381         dlen = PAGE_SIZE;
1382         src = zpool_map_handle(entry->pool->zpool, entry->handle, ZPOOL_MM_RO);
1383         if (zpool_evictable(entry->pool->zpool))
1384                 src += sizeof(struct zswap_header);
1385
1386         if (!zpool_can_sleep_mapped(entry->pool->zpool)) {
1387                 memcpy(tmp, src, entry->length);
1388                 src = tmp;
1389                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1390         }
1391
1392         acomp_ctx = raw_cpu_ptr(entry->pool->acomp_ctx);
1393         mutex_lock(acomp_ctx->mutex);
1394         sg_init_one(&input, src, entry->length);
1395         sg_init_table(&output, 1);
1396         sg_set_page(&output, page, PAGE_SIZE, 0);
1397         acomp_request_set_params(acomp_ctx->req, &input, &output, entry->length, dlen);
1398         ret = crypto_wait_req(crypto_acomp_decompress(acomp_ctx->req), &acomp_ctx->wait);
1399         mutex_unlock(acomp_ctx->mutex);
1400
1401         if (zpool_can_sleep_mapped(entry->pool->zpool))
1402                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
1403         else
1404                 kfree(tmp);
1405
1406         BUG_ON(ret);
1407 stats:
1408         count_vm_event(ZSWPIN);
1409         if (entry->objcg)
1410                 count_objcg_event(entry->objcg, ZSWPIN);
1411 freeentry:
1412         spin_lock(&tree->lock);
1413         zswap_entry_put(tree, entry);
1414         spin_unlock(&tree->lock);
1415
1416         return ret;
1417 }
1418
1419 /* frees an entry in zswap */
1420 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1421 {
1422         struct zswap_tree *tree = zswap_trees[type];
1423         struct zswap_entry *entry;
1424
1425         /* find */
1426         spin_lock(&tree->lock);
1427         entry = zswap_rb_search(&tree->rbroot, offset);
1428         if (!entry) {
1429                 /* entry was written back */
1430                 spin_unlock(&tree->lock);
1431                 return;
1432         }
1433
1434         /* remove from rbtree */
1435         zswap_rb_erase(&tree->rbroot, entry);
1436
1437         /* drop the initial reference from entry creation */
1438         zswap_entry_put(tree, entry);
1439
1440         spin_unlock(&tree->lock);
1441 }
1442
1443 /* frees all zswap entries for the given swap type */
1444 static void zswap_frontswap_invalidate_area(unsigned type)
1445 {
1446         struct zswap_tree *tree = zswap_trees[type];
1447         struct zswap_entry *entry, *n;
1448
1449         if (!tree)
1450                 return;
1451
1452         /* walk the tree and free everything */
1453         spin_lock(&tree->lock);
1454         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1455                 zswap_free_entry(entry);
1456         tree->rbroot = RB_ROOT;
1457         spin_unlock(&tree->lock);
1458         kfree(tree);
1459         zswap_trees[type] = NULL;
1460 }
1461
1462 static void zswap_frontswap_init(unsigned type)
1463 {
1464         struct zswap_tree *tree;
1465
1466         tree = kzalloc(sizeof(*tree), GFP_KERNEL);
1467         if (!tree) {
1468                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1469                 return;
1470         }
1471
1472         tree->rbroot = RB_ROOT;
1473         spin_lock_init(&tree->lock);
1474         zswap_trees[type] = tree;
1475 }
1476
1477 static const struct frontswap_ops zswap_frontswap_ops = {
1478         .store = zswap_frontswap_store,
1479         .load = zswap_frontswap_load,
1480         .invalidate_page = zswap_frontswap_invalidate_page,
1481         .invalidate_area = zswap_frontswap_invalidate_area,
1482         .init = zswap_frontswap_init
1483 };
1484
1485 /*********************************
1486 * debugfs functions
1487 **********************************/
1488 #ifdef CONFIG_DEBUG_FS
1489 #include <linux/debugfs.h>
1490
1491 static struct dentry *zswap_debugfs_root;
1492
1493 static int zswap_debugfs_init(void)
1494 {
1495         if (!debugfs_initialized())
1496                 return -ENODEV;
1497
1498         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1499
1500         debugfs_create_u64("pool_limit_hit", 0444,
1501                            zswap_debugfs_root, &zswap_pool_limit_hit);
1502         debugfs_create_u64("reject_reclaim_fail", 0444,
1503                            zswap_debugfs_root, &zswap_reject_reclaim_fail);
1504         debugfs_create_u64("reject_alloc_fail", 0444,
1505                            zswap_debugfs_root, &zswap_reject_alloc_fail);
1506         debugfs_create_u64("reject_kmemcache_fail", 0444,
1507                            zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1508         debugfs_create_u64("reject_compress_poor", 0444,
1509                            zswap_debugfs_root, &zswap_reject_compress_poor);
1510         debugfs_create_u64("written_back_pages", 0444,
1511                            zswap_debugfs_root, &zswap_written_back_pages);
1512         debugfs_create_u64("duplicate_entry", 0444,
1513                            zswap_debugfs_root, &zswap_duplicate_entry);
1514         debugfs_create_u64("pool_total_size", 0444,
1515                            zswap_debugfs_root, &zswap_pool_total_size);
1516         debugfs_create_atomic_t("stored_pages", 0444,
1517                                 zswap_debugfs_root, &zswap_stored_pages);
1518         debugfs_create_atomic_t("same_filled_pages", 0444,
1519                                 zswap_debugfs_root, &zswap_same_filled_pages);
1520
1521         return 0;
1522 }
1523 #else
1524 static int zswap_debugfs_init(void)
1525 {
1526         return 0;
1527 }
1528 #endif
1529
1530 /*********************************
1531 * module init and exit
1532 **********************************/
1533 static int zswap_setup(void)
1534 {
1535         struct zswap_pool *pool;
1536         int ret;
1537
1538         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
1539         if (!zswap_entry_cache) {
1540                 pr_err("entry cache creation failed\n");
1541                 goto cache_fail;
1542         }
1543
1544         ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1545                                 zswap_dstmem_prepare, zswap_dstmem_dead);
1546         if (ret) {
1547                 pr_err("dstmem alloc failed\n");
1548                 goto dstmem_fail;
1549         }
1550
1551         ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1552                                       "mm/zswap_pool:prepare",
1553                                       zswap_cpu_comp_prepare,
1554                                       zswap_cpu_comp_dead);
1555         if (ret)
1556                 goto hp_fail;
1557
1558         pool = __zswap_pool_create_fallback();
1559         if (pool) {
1560                 pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1561                         zpool_get_type(pool->zpool));
1562                 list_add(&pool->list, &zswap_pools);
1563                 zswap_has_pool = true;
1564         } else {
1565                 pr_err("pool creation failed\n");
1566                 zswap_enabled = false;
1567         }
1568
1569         shrink_wq = create_workqueue("zswap-shrink");
1570         if (!shrink_wq)
1571                 goto fallback_fail;
1572
1573         ret = frontswap_register_ops(&zswap_frontswap_ops);
1574         if (ret)
1575                 goto destroy_wq;
1576         if (zswap_debugfs_init())
1577                 pr_warn("debugfs initialization failed\n");
1578         zswap_init_state = ZSWAP_INIT_SUCCEED;
1579         return 0;
1580
1581 destroy_wq:
1582         destroy_workqueue(shrink_wq);
1583 fallback_fail:
1584         if (pool)
1585                 zswap_pool_destroy(pool);
1586 hp_fail:
1587         cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1588 dstmem_fail:
1589         kmem_cache_destroy(zswap_entry_cache);
1590 cache_fail:
1591         /* if built-in, we aren't unloaded on failure; don't allow use */
1592         zswap_init_state = ZSWAP_INIT_FAILED;
1593         zswap_enabled = false;
1594         return -ENOMEM;
1595 }
1596
1597 static int __init zswap_init(void)
1598 {
1599         if (!zswap_enabled)
1600                 return 0;
1601         return zswap_setup();
1602 }
1603 /* must be late so crypto has time to come up */
1604 late_initcall(zswap_init);
1605
1606 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1607 MODULE_DESCRIPTION("Compressed cache for swap pages");