mm/zswap.c

   1 /*
   2  * zswap.c - zswap driver file
   3  *
   4  * zswap is a backend for frontswap that takes pages that are in the process
   5  * of being swapped out and attempts to compress and store them in a
   6  * RAM-based memory pool.  This can result in a significant I/O reduction on
   7  * the swap device and, in the case where decompressing from RAM is faster
   8  * than reading from the swap device, can also improve workload performance.
   9  *
  10  * Copyright (C) 2012  Seth Jennings <sjenning@linux.vnet.ibm.com>
  11  *
  12  * This program is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU General Public License
  14  * as published by the Free Software Foundation; either version 2
  15  * of the License, or (at your option) any later version.
  16  *
  17  * This program is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  20  * GNU General Public License for more details.
  21 */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/module.h>
  26 #include <linux/cpu.h>
  27 #include <linux/highmem.h>
  28 #include <linux/slab.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/types.h>
  31 #include <linux/atomic.h>
  32 #include <linux/frontswap.h>
  33 #include <linux/rbtree.h>
  34 #include <linux/swap.h>
  35 #include <linux/crypto.h>
  36 #include <linux/mempool.h>
  37 #include <linux/zpool.h>
  38
  39 #include <linux/mm_types.h>
  40 #include <linux/page-flags.h>
  41 #include <linux/swapops.h>
  42 #include <linux/writeback.h>
  43 #include <linux/pagemap.h>
  44
  45 /*********************************
  46 * statistics
  47 **********************************/
  48 /* Total bytes used by the compressed storage */
  49 static u64 zswap_pool_total_size;
  50 /* The number of compressed pages currently stored in zswap */
  51 static atomic_t zswap_stored_pages = ATOMIC_INIT(0);
  52
  53 /*
  54  * The statistics below are not protected from concurrent access for
  55  * performance reasons so they may not be a 100% accurate.  However,
  56  * they do provide useful information on roughly how many times a
  57  * certain event is occurring.
  58 */
  59
  60 /* Pool limit was hit (see zswap_max_pool_percent) */
  61 static u64 zswap_pool_limit_hit;
  62 /* Pages written back when pool limit was reached */
  63 static u64 zswap_written_back_pages;
  64 /* Store failed due to a reclaim failure after pool limit was reached */
  65 static u64 zswap_reject_reclaim_fail;
  66 /* Compressed page was too big for the allocator to (optimally) store */
  67 static u64 zswap_reject_compress_poor;
  68 /* Store failed because underlying allocator could not get memory */
  69 static u64 zswap_reject_alloc_fail;
  70 /* Store failed because the entry metadata could not be allocated (rare) */
  71 static u64 zswap_reject_kmemcache_fail;
  72 /* Duplicate store was encountered (rare) */
  73 static u64 zswap_duplicate_entry;
  74
  75 /*********************************
  76 * tunables
  77 **********************************/
  78
  79 /* Enable/disable zswap (disabled by default) */
  80 static bool zswap_enabled;
  81 module_param_named(enabled, zswap_enabled, bool, 0644);
  82
  83 /* Crypto compressor to use */
  84 #define ZSWAP_COMPRESSOR_DEFAULT "lzo"
  85 static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
  86 static int zswap_compressor_param_set(const char *,
  87                                       const struct kernel_param *);
  88 static struct kernel_param_ops zswap_compressor_param_ops = {
  89         .set =          zswap_compressor_param_set,
  90         .get =          param_get_charp,
  91         .free =         param_free_charp,
  92 };
  93 module_param_cb(compressor, &zswap_compressor_param_ops,
  94                 &zswap_compressor, 0644);
  95
  96 /* Compressed storage zpool to use */
  97 #define ZSWAP_ZPOOL_DEFAULT "zbud"
  98 static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
  99 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 100 static struct kernel_param_ops zswap_zpool_param_ops = {
 101         .set =          zswap_zpool_param_set,
 102         .get =          param_get_charp,
 103         .free =         param_free_charp,
 104 };
 105 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 0644);
 106
 107 /* The maximum percentage of memory that the compressed pool can occupy */
 108 static unsigned int zswap_max_pool_percent = 20;
 109 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
 110
 111 /*********************************
 112 * data structures
 113 **********************************/
 114
 115 struct zswap_pool {
 116         struct zpool *zpool;
 117         struct crypto_comp * __percpu *tfm;
 118         struct kref kref;
 119         struct list_head list;
 120         struct work_struct work;
 121         struct hlist_node node;
 122         char tfm_name[CRYPTO_MAX_ALG_NAME];
 123 };
 124
 125 /*
 126  * struct zswap_entry
 127  *
 128  * This structure contains the metadata for tracking a single compressed
 129  * page within zswap.
 130  *
 131  * rbnode - links the entry into red-black tree for the appropriate swap type
 132  * offset - the swap offset for the entry.  Index into the red-black tree.
 133  * refcount - the number of outstanding reference to the entry. This is needed
 134  *            to protect against premature freeing of the entry by code
 135  *            concurrent calls to load, invalidate, and writeback.  The lock
 136  *            for the zswap_tree structure that contains the entry must
 137  *            be held while changing the refcount.  Since the lock must
 138  *            be held, there is no reason to also make refcount atomic.
 139  * length - the length in bytes of the compressed page data.  Needed during
 140  *          decompression
 141  * pool - the zswap_pool the entry's data is in
 142  * handle - zpool allocation handle that stores the compressed page data
 143  */
 144 struct zswap_entry {
 145         struct rb_node rbnode;
 146         pgoff_t offset;
 147         int refcount;
 148         unsigned int length;
 149         struct zswap_pool *pool;
 150         unsigned long handle;
 151 };
 152
 153 struct zswap_header {
 154         swp_entry_t swpentry;
 155 };
 156
 157 /*
 158  * The tree lock in the zswap_tree struct protects a few things:
 159  * - the rbtree
 160  * - the refcount field of each entry in the tree
 161  */
 162 struct zswap_tree {
 163         struct rb_root rbroot;
 164         spinlock_t lock;
 165 };
 166
 167 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 168
 169 /* RCU-protected iteration */
 170 static LIST_HEAD(zswap_pools);
 171 /* protects zswap_pools list modification */
 172 static DEFINE_SPINLOCK(zswap_pools_lock);
 173 /* pool counter to provide unique names to zpool */
 174 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
 175
 176 /* used by param callback function */
 177 static bool zswap_init_started;
 178
 179 /*********************************
 180 * helpers and fwd declarations
 181 **********************************/
 182
 183 #define zswap_pool_debug(msg, p)                                \
 184         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 185                  zpool_get_type((p)->zpool))
 186
 187 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 188 static int zswap_pool_get(struct zswap_pool *pool);
 189 static void zswap_pool_put(struct zswap_pool *pool);
 190
 191 static const struct zpool_ops zswap_zpool_ops = {
 192         .evict = zswap_writeback_entry
 193 };
 194
 195 static bool zswap_is_full(void)
 196 {
 197         return totalram_pages * zswap_max_pool_percent / 100 <
 198                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 199 }
 200
 201 static void zswap_update_total_size(void)
 202 {
 203         struct zswap_pool *pool;
 204         u64 total = 0;
 205
 206         rcu_read_lock();
 207
 208         list_for_each_entry_rcu(pool, &zswap_pools, list)
 209                 total += zpool_get_total_size(pool->zpool);
 210
 211         rcu_read_unlock();
 212
 213         zswap_pool_total_size = total;
 214 }
 215
 216 /*********************************
 217 * zswap entry functions
 218 **********************************/
 219 static struct kmem_cache *zswap_entry_cache;
 220
 221 static int __init zswap_entry_cache_create(void)
 222 {
 223         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 224         return zswap_entry_cache == NULL;
 225 }
 226
 227 static void __init zswap_entry_cache_destroy(void)
 228 {
 229         kmem_cache_destroy(zswap_entry_cache);
 230 }
 231
 232 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 233 {
 234         struct zswap_entry *entry;
 235         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 236         if (!entry)
 237                 return NULL;
 238         entry->refcount = 1;
 239         RB_CLEAR_NODE(&entry->rbnode);
 240         return entry;
 241 }
 242
 243 static void zswap_entry_cache_free(struct zswap_entry *entry)
 244 {
 245         kmem_cache_free(zswap_entry_cache, entry);
 246 }
 247
 248 /*********************************
 249 * rbtree functions
 250 **********************************/
 251 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 252 {
 253         struct rb_node *node = root->rb_node;
 254         struct zswap_entry *entry;
 255
 256         while (node) {
 257                 entry = rb_entry(node, struct zswap_entry, rbnode);
 258                 if (entry->offset > offset)
 259                         node = node->rb_left;
 260                 else if (entry->offset < offset)
 261                         node = node->rb_right;
 262                 else
 263                         return entry;
 264         }
 265         return NULL;
 266 }
 267
 268 /*
 269  * In the case that a entry with the same offset is found, a pointer to
 270  * the existing entry is stored in dupentry and the function returns -EEXIST
 271  */
 272 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 273                         struct zswap_entry **dupentry)
 274 {
 275         struct rb_node **link = &root->rb_node, *parent = NULL;
 276         struct zswap_entry *myentry;
 277
 278         while (*link) {
 279                 parent = *link;
 280                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 281                 if (myentry->offset > entry->offset)
 282                         link = &(*link)->rb_left;
 283                 else if (myentry->offset < entry->offset)
 284                         link = &(*link)->rb_right;
 285                 else {
 286                         *dupentry = myentry;
 287                         return -EEXIST;
 288                 }
 289         }
 290         rb_link_node(&entry->rbnode, parent, link);
 291         rb_insert_color(&entry->rbnode, root);
 292         return 0;
 293 }
 294
 295 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 296 {
 297         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 298                 rb_erase(&entry->rbnode, root);
 299                 RB_CLEAR_NODE(&entry->rbnode);
 300         }
 301 }
 302
 303 /*
 304  * Carries out the common pattern of freeing and entry's zpool allocation,
 305  * freeing the entry itself, and decrementing the number of stored pages.
 306  */
 307 static void zswap_free_entry(struct zswap_entry *entry)
 308 {
 309         zpool_free(entry->pool->zpool, entry->handle);
 310         zswap_pool_put(entry->pool);
 311         zswap_entry_cache_free(entry);
 312         atomic_dec(&zswap_stored_pages);
 313         zswap_update_total_size();
 314 }
 315
 316 /* caller must hold the tree lock */
 317 static void zswap_entry_get(struct zswap_entry *entry)
 318 {
 319         entry->refcount++;
 320 }
 321
 322 /* caller must hold the tree lock
 323 * remove from the tree and free it, if nobody reference the entry
 324 */
 325 static void zswap_entry_put(struct zswap_tree *tree,
 326                         struct zswap_entry *entry)
 327 {
 328         int refcount = --entry->refcount;
 329
 330         BUG_ON(refcount < 0);
 331         if (refcount == 0) {
 332                 zswap_rb_erase(&tree->rbroot, entry);
 333                 zswap_free_entry(entry);
 334         }
 335 }
 336
 337 /* caller must hold the tree lock */
 338 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 339                                 pgoff_t offset)
 340 {
 341         struct zswap_entry *entry;
 342
 343         entry = zswap_rb_search(root, offset);
 344         if (entry)
 345                 zswap_entry_get(entry);
 346
 347         return entry;
 348 }
 349
 350 /*********************************
 351 * per-cpu code
 352 **********************************/
 353 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 354
 355 static int zswap_dstmem_prepare(unsigned int cpu)
 356 {
 357         u8 *dst;
 358
 359         dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 360         if (!dst) {
 361                 pr_err("can't allocate compressor buffer\n");
 362                 return -ENOMEM;
 363         }
 364         per_cpu(zswap_dstmem, cpu) = dst;
 365         return 0;
 366 }
 367
 368 static int zswap_dstmem_dead(unsigned int cpu)
 369 {
 370         u8 *dst;
 371
 372         dst = per_cpu(zswap_dstmem, cpu);
 373         kfree(dst);
 374         per_cpu(zswap_dstmem, cpu) = NULL;
 375
 376         return 0;
 377 }
 378
 379 static int zswap_cpu_comp_prepare(unsigned int cpu, struct hlist_node *node)
 380 {
 381         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 382         struct crypto_comp *tfm;
 383
 384         if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 385                 return 0;
 386
 387         tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 388         if (IS_ERR_OR_NULL(tfm)) {
 389                 pr_err("could not alloc crypto comp %s : %ld\n",
 390                        pool->tfm_name, PTR_ERR(tfm));
 391                 return -ENOMEM;
 392         }
 393         *per_cpu_ptr(pool->tfm, cpu) = tfm;
 394         return 0;
 395 }
 396
 397 static int zswap_cpu_comp_dead(unsigned int cpu, struct hlist_node *node)
 398 {
 399         struct zswap_pool *pool = hlist_entry(node, struct zswap_pool, node);
 400         struct crypto_comp *tfm;
 401
 402         tfm = *per_cpu_ptr(pool->tfm, cpu);
 403         if (!IS_ERR_OR_NULL(tfm))
 404                 crypto_free_comp(tfm);
 405         *per_cpu_ptr(pool->tfm, cpu) = NULL;
 406         return 0;
 407 }
 408
 409 /*********************************
 410 * pool functions
 411 **********************************/
 412
 413 static struct zswap_pool *__zswap_pool_current(void)
 414 {
 415         struct zswap_pool *pool;
 416
 417         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 418         WARN_ON(!pool);
 419
 420         return pool;
 421 }
 422
 423 static struct zswap_pool *zswap_pool_current(void)
 424 {
 425         assert_spin_locked(&zswap_pools_lock);
 426
 427         return __zswap_pool_current();
 428 }
 429
 430 static struct zswap_pool *zswap_pool_current_get(void)
 431 {
 432         struct zswap_pool *pool;
 433
 434         rcu_read_lock();
 435
 436         pool = __zswap_pool_current();
 437         if (!pool || !zswap_pool_get(pool))
 438                 pool = NULL;
 439
 440         rcu_read_unlock();
 441
 442         return pool;
 443 }
 444
 445 static struct zswap_pool *zswap_pool_last_get(void)
 446 {
 447         struct zswap_pool *pool, *last = NULL;
 448
 449         rcu_read_lock();
 450
 451         list_for_each_entry_rcu(pool, &zswap_pools, list)
 452                 last = pool;
 453         if (!WARN_ON(!last) && !zswap_pool_get(last))
 454                 last = NULL;
 455
 456         rcu_read_unlock();
 457
 458         return last;
 459 }
 460
 461 /* type and compressor must be null-terminated */
 462 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 463 {
 464         struct zswap_pool *pool;
 465
 466         assert_spin_locked(&zswap_pools_lock);
 467
 468         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 469                 if (strcmp(pool->tfm_name, compressor))
 470                         continue;
 471                 if (strcmp(zpool_get_type(pool->zpool), type))
 472                         continue;
 473                 /* if we can't get it, it's about to be destroyed */
 474                 if (!zswap_pool_get(pool))
 475                         continue;
 476                 return pool;
 477         }
 478
 479         return NULL;
 480 }
 481
 482 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 483 {
 484         struct zswap_pool *pool;
 485         char name[38]; /* 'zswap' + 32 char (max) num + \0 */
 486         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
 487         int ret;
 488
 489         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 490         if (!pool) {
 491                 pr_err("pool alloc failed\n");
 492                 return NULL;
 493         }
 494
 495         /* unique name for each pool specifically required by zsmalloc */
 496         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
 497
 498         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
 499         if (!pool->zpool) {
 500                 pr_err("%s zpool not available\n", type);
 501                 goto error;
 502         }
 503         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 504
 505         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 506         pool->tfm = alloc_percpu(struct crypto_comp *);
 507         if (!pool->tfm) {
 508                 pr_err("percpu alloc failed\n");
 509                 goto error;
 510         }
 511
 512         ret = cpuhp_state_add_instance(CPUHP_MM_ZSWP_POOL_PREPARE,
 513                                        &pool->node);
 514         if (ret)
 515                 goto error;
 516         pr_debug("using %s compressor\n", pool->tfm_name);
 517
 518         /* being the current pool takes 1 ref; this func expects the
 519          * caller to always add the new pool as the current pool
 520          */
 521         kref_init(&pool->kref);
 522         INIT_LIST_HEAD(&pool->list);
 523
 524         zswap_pool_debug("created", pool);
 525
 526         return pool;
 527
 528 error:
 529         free_percpu(pool->tfm);
 530         if (pool->zpool)
 531                 zpool_destroy_pool(pool->zpool);
 532         kfree(pool);
 533         return NULL;
 534 }
 535
 536 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
 537 {
 538         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 539                 if (!strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
 540                         pr_err("default compressor %s not available\n",
 541                                zswap_compressor);
 542                         return NULL;
 543                 }
 544                 pr_err("compressor %s not available, using default %s\n",
 545                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 546                 param_free_charp(&zswap_compressor);
 547                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
 548         }
 549         if (!zpool_has_pool(zswap_zpool_type)) {
 550                 if (!strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
 551                         pr_err("default zpool %s not available\n",
 552                                zswap_zpool_type);
 553                         return NULL;
 554                 }
 555                 pr_err("zpool %s not available, using default %s\n",
 556                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 557                 param_free_charp(&zswap_zpool_type);
 558                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 559         }
 560
 561         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 562 }
 563
 564 static void zswap_pool_destroy(struct zswap_pool *pool)
 565 {
 566         zswap_pool_debug("destroying", pool);
 567
 568         cpuhp_state_remove_instance(CPUHP_MM_ZSWP_POOL_PREPARE, &pool->node);
 569         free_percpu(pool->tfm);
 570         zpool_destroy_pool(pool->zpool);
 571         kfree(pool);
 572 }
 573
 574 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 575 {
 576         return kref_get_unless_zero(&pool->kref);
 577 }
 578
 579 static void __zswap_pool_release(struct work_struct *work)
 580 {
 581         struct zswap_pool *pool = container_of(work, typeof(*pool), work);
 582
 583         synchronize_rcu();
 584
 585         /* nobody should have been able to get a kref... */
 586         WARN_ON(kref_get_unless_zero(&pool->kref));
 587
 588         /* pool is now off zswap_pools list and has no references. */
 589         zswap_pool_destroy(pool);
 590 }
 591
 592 static void __zswap_pool_empty(struct kref *kref)
 593 {
 594         struct zswap_pool *pool;
 595
 596         pool = container_of(kref, typeof(*pool), kref);
 597
 598         spin_lock(&zswap_pools_lock);
 599
 600         WARN_ON(pool == zswap_pool_current());
 601
 602         list_del_rcu(&pool->list);
 603
 604         INIT_WORK(&pool->work, __zswap_pool_release);
 605         schedule_work(&pool->work);
 606
 607         spin_unlock(&zswap_pools_lock);
 608 }
 609
 610 static void zswap_pool_put(struct zswap_pool *pool)
 611 {
 612         kref_put(&pool->kref, __zswap_pool_empty);
 613 }
 614
 615 /*********************************
 616 * param callbacks
 617 **********************************/
 618
 619 /* val must be a null-terminated string */
 620 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 621                              char *type, char *compressor)
 622 {
 623         struct zswap_pool *pool, *put_pool = NULL;
 624         char *s = strstrip((char *)val);
 625         int ret;
 626
 627         /* no change required */
 628         if (!strcmp(s, *(char **)kp->arg))
 629                 return 0;
 630
 631         /* if this is load-time (pre-init) param setting,
 632          * don't create a pool; that's done during init.
 633          */
 634         if (!zswap_init_started)
 635                 return param_set_charp(s, kp);
 636
 637         if (!type) {
 638                 if (!zpool_has_pool(s)) {
 639                         pr_err("zpool %s not available\n", s);
 640                         return -ENOENT;
 641                 }
 642                 type = s;
 643         } else if (!compressor) {
 644                 if (!crypto_has_comp(s, 0, 0)) {
 645                         pr_err("compressor %s not available\n", s);
 646                         return -ENOENT;
 647                 }
 648                 compressor = s;
 649         } else {
 650                 WARN_ON(1);
 651                 return -EINVAL;
 652         }
 653
 654         spin_lock(&zswap_pools_lock);
 655
 656         pool = zswap_pool_find_get(type, compressor);
 657         if (pool) {
 658                 zswap_pool_debug("using existing", pool);
 659                 list_del_rcu(&pool->list);
 660         } else {
 661                 spin_unlock(&zswap_pools_lock);
 662                 pool = zswap_pool_create(type, compressor);
 663                 spin_lock(&zswap_pools_lock);
 664         }
 665
 666         if (pool)
 667                 ret = param_set_charp(s, kp);
 668         else
 669                 ret = -EINVAL;
 670
 671         if (!ret) {
 672                 put_pool = zswap_pool_current();
 673                 list_add_rcu(&pool->list, &zswap_pools);
 674         } else if (pool) {
 675                 /* add the possibly pre-existing pool to the end of the pools
 676                  * list; if it's new (and empty) then it'll be removed and
 677                  * destroyed by the put after we drop the lock
 678                  */
 679                 list_add_tail_rcu(&pool->list, &zswap_pools);
 680                 put_pool = pool;
 681         }
 682
 683         spin_unlock(&zswap_pools_lock);
 684
 685         /* drop the ref from either the old current pool,
 686          * or the new pool we failed to add
 687          */
 688         if (put_pool)
 689                 zswap_pool_put(put_pool);
 690
 691         return ret;
 692 }
 693
 694 static int zswap_compressor_param_set(const char *val,
 695                                       const struct kernel_param *kp)
 696 {
 697         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
 698 }
 699
 700 static int zswap_zpool_param_set(const char *val,
 701                                  const struct kernel_param *kp)
 702 {
 703         return __zswap_param_set(val, kp, NULL, zswap_compressor);
 704 }
 705
 706 /*********************************
 707 * writeback code
 708 **********************************/
 709 /* return enum for zswap_get_swap_cache_page */
 710 enum zswap_get_swap_ret {
 711         ZSWAP_SWAPCACHE_NEW,
 712         ZSWAP_SWAPCACHE_EXIST,
 713         ZSWAP_SWAPCACHE_FAIL,
 714 };
 715
 716 /*
 717  * zswap_get_swap_cache_page
 718  *
 719  * This is an adaption of read_swap_cache_async()
 720  *
 721  * This function tries to find a page with the given swap entry
 722  * in the swapper_space address space (the swap cache).  If the page
 723  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 724  * added to the swap cache, and returned in retpage.
 725  *
 726  * If success, the swap cache page is returned in retpage
 727  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 728  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 729  *     the new page is added to swapcache and locked
 730  * Returns ZSWAP_SWAPCACHE_FAIL on error
 731  */
 732 static int zswap_get_swap_cache_page(swp_entry_t entry,
 733                                 struct page **retpage)
 734 {
 735         bool page_was_allocated;
 736
 737         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 738                         NULL, 0, &page_was_allocated);
 739         if (page_was_allocated)
 740                 return ZSWAP_SWAPCACHE_NEW;
 741         if (!*retpage)
 742                 return ZSWAP_SWAPCACHE_FAIL;
 743         return ZSWAP_SWAPCACHE_EXIST;
 744 }
 745
 746 /*
 747  * Attempts to free an entry by adding a page to the swap cache,
 748  * decompressing the entry data into the page, and issuing a
 749  * bio write to write the page back to the swap device.
 750  *
 751  * This can be thought of as a "resumed writeback" of the page
 752  * to the swap device.  We are basically resuming the same swap
 753  * writeback path that was intercepted with the frontswap_store()
 754  * in the first place.  After the page has been decompressed into
 755  * the swap cache, the compressed version stored by zswap can be
 756  * freed.
 757  */
 758 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 759 {
 760         struct zswap_header *zhdr;
 761         swp_entry_t swpentry;
 762         struct zswap_tree *tree;
 763         pgoff_t offset;
 764         struct zswap_entry *entry;
 765         struct page *page;
 766         struct crypto_comp *tfm;
 767         u8 *src, *dst;
 768         unsigned int dlen;
 769         int ret;
 770         struct writeback_control wbc = {
 771                 .sync_mode = WB_SYNC_NONE,
 772         };
 773
 774         /* extract swpentry from data */
 775         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 776         swpentry = zhdr->swpentry; /* here */
 777         zpool_unmap_handle(pool, handle);
 778         tree = zswap_trees[swp_type(swpentry)];
 779         offset = swp_offset(swpentry);
 780
 781         /* find and ref zswap entry */
 782         spin_lock(&tree->lock);
 783         entry = zswap_entry_find_get(&tree->rbroot, offset);
 784         if (!entry) {
 785                 /* entry was invalidated */
 786                 spin_unlock(&tree->lock);
 787                 return 0;
 788         }
 789         spin_unlock(&tree->lock);
 790         BUG_ON(offset != entry->offset);
 791
 792         /* try to allocate swap cache page */
 793         switch (zswap_get_swap_cache_page(swpentry, &page)) {
 794         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
 795                 ret = -ENOMEM;
 796                 goto fail;
 797
 798         case ZSWAP_SWAPCACHE_EXIST:
 799                 /* page is already in the swap cache, ignore for now */
 800                 put_page(page);
 801                 ret = -EEXIST;
 802                 goto fail;
 803
 804         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
 805                 /* decompress */
 806                 dlen = PAGE_SIZE;
 807                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 808                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
 809                 dst = kmap_atomic(page);
 810                 tfm = *get_cpu_ptr(entry->pool->tfm);
 811                 ret = crypto_comp_decompress(tfm, src, entry->length,
 812                                              dst, &dlen);
 813                 put_cpu_ptr(entry->pool->tfm);
 814                 kunmap_atomic(dst);
 815                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
 816                 BUG_ON(ret);
 817                 BUG_ON(dlen != PAGE_SIZE);
 818
 819                 /* page is up to date */
 820                 SetPageUptodate(page);
 821         }
 822
 823         /* move it to the tail of the inactive list after end_writeback */
 824         SetPageReclaim(page);
 825
 826         /* start writeback */
 827         __swap_writepage(page, &wbc, end_swap_bio_write);
 828         put_page(page);
 829         zswap_written_back_pages++;
 830
 831         spin_lock(&tree->lock);
 832         /* drop local reference */
 833         zswap_entry_put(tree, entry);
 834
 835         /*
 836         * There are two possible situations for entry here:
 837         * (1) refcount is 1(normal case),  entry is valid and on the tree
 838         * (2) refcount is 0, entry is freed and not on the tree
 839         *     because invalidate happened during writeback
 840         *  search the tree and free the entry if find entry
 841         */
 842         if (entry == zswap_rb_search(&tree->rbroot, offset))
 843                 zswap_entry_put(tree, entry);
 844         spin_unlock(&tree->lock);
 845
 846         goto end;
 847
 848         /*
 849         * if we get here due to ZSWAP_SWAPCACHE_EXIST
 850         * a load may happening concurrently
 851         * it is safe and okay to not free the entry
 852         * if we free the entry in the following put
 853         * it it either okay to return !0
 854         */
 855 fail:
 856         spin_lock(&tree->lock);
 857         zswap_entry_put(tree, entry);
 858         spin_unlock(&tree->lock);
 859
 860 end:
 861         return ret;
 862 }
 863
 864 static int zswap_shrink(void)
 865 {
 866         struct zswap_pool *pool;
 867         int ret;
 868
 869         pool = zswap_pool_last_get();
 870         if (!pool)
 871                 return -ENOENT;
 872
 873         ret = zpool_shrink(pool->zpool, 1, NULL);
 874
 875         zswap_pool_put(pool);
 876
 877         return ret;
 878 }
 879
 880 /*********************************
 881 * frontswap hooks
 882 **********************************/
 883 /* attempts to compress and store an single page */
 884 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 885                                 struct page *page)
 886 {
 887         struct zswap_tree *tree = zswap_trees[type];
 888         struct zswap_entry *entry, *dupentry;
 889         struct crypto_comp *tfm;
 890         int ret;
 891         unsigned int dlen = PAGE_SIZE, len;
 892         unsigned long handle;
 893         char *buf;
 894         u8 *src, *dst;
 895         struct zswap_header *zhdr;
 896
 897         if (!zswap_enabled || !tree) {
 898                 ret = -ENODEV;
 899                 goto reject;
 900         }
 901
 902         /* reclaim space if needed */
 903         if (zswap_is_full()) {
 904                 zswap_pool_limit_hit++;
 905                 if (zswap_shrink()) {
 906                         zswap_reject_reclaim_fail++;
 907                         ret = -ENOMEM;
 908                         goto reject;
 909                 }
 910         }
 911
 912         /* allocate entry */
 913         entry = zswap_entry_cache_alloc(GFP_KERNEL);
 914         if (!entry) {
 915                 zswap_reject_kmemcache_fail++;
 916                 ret = -ENOMEM;
 917                 goto reject;
 918         }
 919
 920         /* if entry is successfully added, it keeps the reference */
 921         entry->pool = zswap_pool_current_get();
 922         if (!entry->pool) {
 923                 ret = -EINVAL;
 924                 goto freepage;
 925         }
 926
 927         /* compress */
 928         dst = get_cpu_var(zswap_dstmem);
 929         tfm = *get_cpu_ptr(entry->pool->tfm);
 930         src = kmap_atomic(page);
 931         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
 932         kunmap_atomic(src);
 933         put_cpu_ptr(entry->pool->tfm);
 934         if (ret) {
 935                 ret = -EINVAL;
 936                 goto put_dstmem;
 937         }
 938
 939         /* store */
 940         len = dlen + sizeof(struct zswap_header);
 941         ret = zpool_malloc(entry->pool->zpool, len,
 942                            __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
 943                            &handle);
 944         if (ret == -ENOSPC) {
 945                 zswap_reject_compress_poor++;
 946                 goto put_dstmem;
 947         }
 948         if (ret) {
 949                 zswap_reject_alloc_fail++;
 950                 goto put_dstmem;
 951         }
 952         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
 953         zhdr->swpentry = swp_entry(type, offset);
 954         buf = (u8 *)(zhdr + 1);
 955         memcpy(buf, dst, dlen);
 956         zpool_unmap_handle(entry->pool->zpool, handle);
 957         put_cpu_var(zswap_dstmem);
 958
 959         /* populate entry */
 960         entry->offset = offset;
 961         entry->handle = handle;
 962         entry->length = dlen;
 963
 964         /* map */
 965         spin_lock(&tree->lock);
 966         do {
 967                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
 968                 if (ret == -EEXIST) {
 969                         zswap_duplicate_entry++;
 970                         /* remove from rbtree */
 971                         zswap_rb_erase(&tree->rbroot, dupentry);
 972                         zswap_entry_put(tree, dupentry);
 973                 }
 974         } while (ret == -EEXIST);
 975         spin_unlock(&tree->lock);
 976
 977         /* update stats */
 978         atomic_inc(&zswap_stored_pages);
 979         zswap_update_total_size();
 980
 981         return 0;
 982
 983 put_dstmem:
 984         put_cpu_var(zswap_dstmem);
 985         zswap_pool_put(entry->pool);
 986 freepage:
 987         zswap_entry_cache_free(entry);
 988 reject:
 989         return ret;
 990 }
 991
 992 /*
 993  * returns 0 if the page was successfully decompressed
 994  * return -1 on entry not found or error
 995 */
 996 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
 997                                 struct page *page)
 998 {
 999         struct zswap_tree *tree = zswap_trees[type];
1000         struct zswap_entry *entry;
1001         struct crypto_comp *tfm;
1002         u8 *src, *dst;
1003         unsigned int dlen;
1004         int ret;
1005
1006         /* find */
1007         spin_lock(&tree->lock);
1008         entry = zswap_entry_find_get(&tree->rbroot, offset);
1009         if (!entry) {
1010                 /* entry was written back */
1011                 spin_unlock(&tree->lock);
1012                 return -1;
1013         }
1014         spin_unlock(&tree->lock);
1015
1016         /* decompress */
1017         dlen = PAGE_SIZE;
1018         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1019                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1020         dst = kmap_atomic(page);
1021         tfm = *get_cpu_ptr(entry->pool->tfm);
1022         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1023         put_cpu_ptr(entry->pool->tfm);
1024         kunmap_atomic(dst);
1025         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1026         BUG_ON(ret);
1027
1028         spin_lock(&tree->lock);
1029         zswap_entry_put(tree, entry);
1030         spin_unlock(&tree->lock);
1031
1032         return 0;
1033 }
1034
1035 /* frees an entry in zswap */
1036 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1037 {
1038         struct zswap_tree *tree = zswap_trees[type];
1039         struct zswap_entry *entry;
1040
1041         /* find */
1042         spin_lock(&tree->lock);
1043         entry = zswap_rb_search(&tree->rbroot, offset);
1044         if (!entry) {
1045                 /* entry was written back */
1046                 spin_unlock(&tree->lock);
1047                 return;
1048         }
1049
1050         /* remove from rbtree */
1051         zswap_rb_erase(&tree->rbroot, entry);
1052
1053         /* drop the initial reference from entry creation */
1054         zswap_entry_put(tree, entry);
1055
1056         spin_unlock(&tree->lock);
1057 }
1058
1059 /* frees all zswap entries for the given swap type */
1060 static void zswap_frontswap_invalidate_area(unsigned type)
1061 {
1062         struct zswap_tree *tree = zswap_trees[type];
1063         struct zswap_entry *entry, *n;
1064
1065         if (!tree)
1066                 return;
1067
1068         /* walk the tree and free everything */
1069         spin_lock(&tree->lock);
1070         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1071                 zswap_free_entry(entry);
1072         tree->rbroot = RB_ROOT;
1073         spin_unlock(&tree->lock);
1074         kfree(tree);
1075         zswap_trees[type] = NULL;
1076 }
1077
1078 static void zswap_frontswap_init(unsigned type)
1079 {
1080         struct zswap_tree *tree;
1081
1082         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1083         if (!tree) {
1084                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1085                 return;
1086         }
1087
1088         tree->rbroot = RB_ROOT;
1089         spin_lock_init(&tree->lock);
1090         zswap_trees[type] = tree;
1091 }
1092
1093 static struct frontswap_ops zswap_frontswap_ops = {
1094         .store = zswap_frontswap_store,
1095         .load = zswap_frontswap_load,
1096         .invalidate_page = zswap_frontswap_invalidate_page,
1097         .invalidate_area = zswap_frontswap_invalidate_area,
1098         .init = zswap_frontswap_init
1099 };
1100
1101 /*********************************
1102 * debugfs functions
1103 **********************************/
1104 #ifdef CONFIG_DEBUG_FS
1105 #include <linux/debugfs.h>
1106
1107 static struct dentry *zswap_debugfs_root;
1108
1109 static int __init zswap_debugfs_init(void)
1110 {
1111         if (!debugfs_initialized())
1112                 return -ENODEV;
1113
1114         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1115         if (!zswap_debugfs_root)
1116                 return -ENOMEM;
1117
1118         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1119                         zswap_debugfs_root, &zswap_pool_limit_hit);
1120         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1121                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1122         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1123                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1124         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1125                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1126         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1127                         zswap_debugfs_root, &zswap_reject_compress_poor);
1128         debugfs_create_u64("written_back_pages", S_IRUGO,
1129                         zswap_debugfs_root, &zswap_written_back_pages);
1130         debugfs_create_u64("duplicate_entry", S_IRUGO,
1131                         zswap_debugfs_root, &zswap_duplicate_entry);
1132         debugfs_create_u64("pool_total_size", S_IRUGO,
1133                         zswap_debugfs_root, &zswap_pool_total_size);
1134         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1135                         zswap_debugfs_root, &zswap_stored_pages);
1136
1137         return 0;
1138 }
1139
1140 static void __exit zswap_debugfs_exit(void)
1141 {
1142         debugfs_remove_recursive(zswap_debugfs_root);
1143 }
1144 #else
1145 static int __init zswap_debugfs_init(void)
1146 {
1147         return 0;
1148 }
1149
1150 static void __exit zswap_debugfs_exit(void) { }
1151 #endif
1152
1153 /*********************************
1154 * module init and exit
1155 **********************************/
1156 static int __init init_zswap(void)
1157 {
1158         struct zswap_pool *pool;
1159         int ret;
1160
1161         zswap_init_started = true;
1162
1163         if (zswap_entry_cache_create()) {
1164                 pr_err("entry cache creation failed\n");
1165                 goto cache_fail;
1166         }
1167
1168         ret = cpuhp_setup_state(CPUHP_MM_ZSWP_MEM_PREPARE, "mm/zswap:prepare",
1169                                 zswap_dstmem_prepare, zswap_dstmem_dead);
1170         if (ret) {
1171                 pr_err("dstmem alloc failed\n");
1172                 goto dstmem_fail;
1173         }
1174
1175         ret = cpuhp_setup_state_multi(CPUHP_MM_ZSWP_POOL_PREPARE,
1176                                       "mm/zswap_pool:prepare",
1177                                       zswap_cpu_comp_prepare,
1178                                       zswap_cpu_comp_dead);
1179         if (ret)
1180                 goto hp_fail;
1181
1182         pool = __zswap_pool_create_fallback();
1183         if (!pool) {
1184                 pr_err("pool creation failed\n");
1185                 goto pool_fail;
1186         }
1187         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1188                 zpool_get_type(pool->zpool));
1189
1190         list_add(&pool->list, &zswap_pools);
1191
1192         frontswap_register_ops(&zswap_frontswap_ops);
1193         if (zswap_debugfs_init())
1194                 pr_warn("debugfs initialization failed\n");
1195         return 0;
1196
1197 pool_fail:
1198         cpuhp_remove_state_nocalls(CPUHP_MM_ZSWP_POOL_PREPARE);
1199 hp_fail:
1200         cpuhp_remove_state(CPUHP_MM_ZSWP_MEM_PREPARE);
1201 dstmem_fail:
1202         zswap_entry_cache_destroy();
1203 cache_fail:
1204         return -ENOMEM;
1205 }
1206 /* must be late so crypto has time to come up */
1207 late_initcall(init_zswap);
1208
1209 MODULE_LICENSE("GPL");
1210 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1211 MODULE_DESCRIPTION("Compressed cache for swap pages");