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[CRYPTO_MAX_ALG_NAME] = ZSWAP_COMPRESSOR_DEFAULT;
  86 static struct kparam_string zswap_compressor_kparam = {
  87         .string =       zswap_compressor,
  88         .maxlen =       sizeof(zswap_compressor),
  89 };
  90 static int zswap_compressor_param_set(const char *,
  91                                       const struct kernel_param *);
  92 static struct kernel_param_ops zswap_compressor_param_ops = {
  93         .set =          zswap_compressor_param_set,
  94         .get =          param_get_string,
  95 };
  96 module_param_cb(compressor, &zswap_compressor_param_ops,
  97                 &zswap_compressor_kparam, 0644);
  98
  99 /* Compressed storage zpool to use */
 100 #define ZSWAP_ZPOOL_DEFAULT "zbud"
 101 static char zswap_zpool_type[32 /* arbitrary */] = ZSWAP_ZPOOL_DEFAULT;
 102 static struct kparam_string zswap_zpool_kparam = {
 103         .string =       zswap_zpool_type,
 104         .maxlen =       sizeof(zswap_zpool_type),
 105 };
 106 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 107 static struct kernel_param_ops zswap_zpool_param_ops = {
 108         .set =  zswap_zpool_param_set,
 109         .get =  param_get_string,
 110 };
 111 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_kparam, 0644);
 112
 113 /* The maximum percentage of memory that the compressed pool can occupy */
 114 static unsigned int zswap_max_pool_percent = 20;
 115 module_param_named(max_pool_percent, zswap_max_pool_percent, uint, 0644);
 116
 117 /*********************************
 118 * data structures
 119 **********************************/
 120
 121 struct zswap_pool {
 122         struct zpool *zpool;
 123         struct crypto_comp * __percpu *tfm;
 124         struct kref kref;
 125         struct list_head list;
 126         struct rcu_head rcu_head;
 127         struct notifier_block notifier;
 128         char tfm_name[CRYPTO_MAX_ALG_NAME];
 129 };
 130
 131 /*
 132  * struct zswap_entry
 133  *
 134  * This structure contains the metadata for tracking a single compressed
 135  * page within zswap.
 136  *
 137  * rbnode - links the entry into red-black tree for the appropriate swap type
 138  * offset - the swap offset for the entry.  Index into the red-black tree.
 139  * refcount - the number of outstanding reference to the entry. This is needed
 140  *            to protect against premature freeing of the entry by code
 141  *            concurrent calls to load, invalidate, and writeback.  The lock
 142  *            for the zswap_tree structure that contains the entry must
 143  *            be held while changing the refcount.  Since the lock must
 144  *            be held, there is no reason to also make refcount atomic.
 145  * length - the length in bytes of the compressed page data.  Needed during
 146  *          decompression
 147  * pool - the zswap_pool the entry's data is in
 148  * handle - zpool allocation handle that stores the compressed page data
 149  */
 150 struct zswap_entry {
 151         struct rb_node rbnode;
 152         pgoff_t offset;
 153         int refcount;
 154         unsigned int length;
 155         struct zswap_pool *pool;
 156         unsigned long handle;
 157 };
 158
 159 struct zswap_header {
 160         swp_entry_t swpentry;
 161 };
 162
 163 /*
 164  * The tree lock in the zswap_tree struct protects a few things:
 165  * - the rbtree
 166  * - the refcount field of each entry in the tree
 167  */
 168 struct zswap_tree {
 169         struct rb_root rbroot;
 170         spinlock_t lock;
 171 };
 172
 173 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 174
 175 /* RCU-protected iteration */
 176 static LIST_HEAD(zswap_pools);
 177 /* protects zswap_pools list modification */
 178 static DEFINE_SPINLOCK(zswap_pools_lock);
 179
 180 /* used by param callback function */
 181 static bool zswap_init_started;
 182
 183 /*********************************
 184 * helpers and fwd declarations
 185 **********************************/
 186
 187 #define zswap_pool_debug(msg, p)                                \
 188         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 189                  zpool_get_type((p)->zpool))
 190
 191 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 192 static int zswap_pool_get(struct zswap_pool *pool);
 193 static void zswap_pool_put(struct zswap_pool *pool);
 194
 195 static const struct zpool_ops zswap_zpool_ops = {
 196         .evict = zswap_writeback_entry
 197 };
 198
 199 static bool zswap_is_full(void)
 200 {
 201         return totalram_pages * zswap_max_pool_percent / 100 <
 202                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 203 }
 204
 205 static void zswap_update_total_size(void)
 206 {
 207         struct zswap_pool *pool;
 208         u64 total = 0;
 209
 210         rcu_read_lock();
 211
 212         list_for_each_entry_rcu(pool, &zswap_pools, list)
 213                 total += zpool_get_total_size(pool->zpool);
 214
 215         rcu_read_unlock();
 216
 217         zswap_pool_total_size = total;
 218 }
 219
 220 /*********************************
 221 * zswap entry functions
 222 **********************************/
 223 static struct kmem_cache *zswap_entry_cache;
 224
 225 static int __init zswap_entry_cache_create(void)
 226 {
 227         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 228         return zswap_entry_cache == NULL;
 229 }
 230
 231 static void __init zswap_entry_cache_destroy(void)
 232 {
 233         kmem_cache_destroy(zswap_entry_cache);
 234 }
 235
 236 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 237 {
 238         struct zswap_entry *entry;
 239         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 240         if (!entry)
 241                 return NULL;
 242         entry->refcount = 1;
 243         RB_CLEAR_NODE(&entry->rbnode);
 244         return entry;
 245 }
 246
 247 static void zswap_entry_cache_free(struct zswap_entry *entry)
 248 {
 249         kmem_cache_free(zswap_entry_cache, entry);
 250 }
 251
 252 /*********************************
 253 * rbtree functions
 254 **********************************/
 255 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 256 {
 257         struct rb_node *node = root->rb_node;
 258         struct zswap_entry *entry;
 259
 260         while (node) {
 261                 entry = rb_entry(node, struct zswap_entry, rbnode);
 262                 if (entry->offset > offset)
 263                         node = node->rb_left;
 264                 else if (entry->offset < offset)
 265                         node = node->rb_right;
 266                 else
 267                         return entry;
 268         }
 269         return NULL;
 270 }
 271
 272 /*
 273  * In the case that a entry with the same offset is found, a pointer to
 274  * the existing entry is stored in dupentry and the function returns -EEXIST
 275  */
 276 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 277                         struct zswap_entry **dupentry)
 278 {
 279         struct rb_node **link = &root->rb_node, *parent = NULL;
 280         struct zswap_entry *myentry;
 281
 282         while (*link) {
 283                 parent = *link;
 284                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 285                 if (myentry->offset > entry->offset)
 286                         link = &(*link)->rb_left;
 287                 else if (myentry->offset < entry->offset)
 288                         link = &(*link)->rb_right;
 289                 else {
 290                         *dupentry = myentry;
 291                         return -EEXIST;
 292                 }
 293         }
 294         rb_link_node(&entry->rbnode, parent, link);
 295         rb_insert_color(&entry->rbnode, root);
 296         return 0;
 297 }
 298
 299 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 300 {
 301         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 302                 rb_erase(&entry->rbnode, root);
 303                 RB_CLEAR_NODE(&entry->rbnode);
 304         }
 305 }
 306
 307 /*
 308  * Carries out the common pattern of freeing and entry's zpool allocation,
 309  * freeing the entry itself, and decrementing the number of stored pages.
 310  */
 311 static void zswap_free_entry(struct zswap_entry *entry)
 312 {
 313         zpool_free(entry->pool->zpool, entry->handle);
 314         zswap_pool_put(entry->pool);
 315         zswap_entry_cache_free(entry);
 316         atomic_dec(&zswap_stored_pages);
 317         zswap_update_total_size();
 318 }
 319
 320 /* caller must hold the tree lock */
 321 static void zswap_entry_get(struct zswap_entry *entry)
 322 {
 323         entry->refcount++;
 324 }
 325
 326 /* caller must hold the tree lock
 327 * remove from the tree and free it, if nobody reference the entry
 328 */
 329 static void zswap_entry_put(struct zswap_tree *tree,
 330                         struct zswap_entry *entry)
 331 {
 332         int refcount = --entry->refcount;
 333
 334         BUG_ON(refcount < 0);
 335         if (refcount == 0) {
 336                 zswap_rb_erase(&tree->rbroot, entry);
 337                 zswap_free_entry(entry);
 338         }
 339 }
 340
 341 /* caller must hold the tree lock */
 342 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 343                                 pgoff_t offset)
 344 {
 345         struct zswap_entry *entry = NULL;
 346
 347         entry = zswap_rb_search(root, offset);
 348         if (entry)
 349                 zswap_entry_get(entry);
 350
 351         return entry;
 352 }
 353
 354 /*********************************
 355 * per-cpu code
 356 **********************************/
 357 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 358
 359 static int __zswap_cpu_dstmem_notifier(unsigned long action, unsigned long cpu)
 360 {
 361         u8 *dst;
 362
 363         switch (action) {
 364         case CPU_UP_PREPARE:
 365                 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 366                 if (!dst) {
 367                         pr_err("can't allocate compressor buffer\n");
 368                         return NOTIFY_BAD;
 369                 }
 370                 per_cpu(zswap_dstmem, cpu) = dst;
 371                 break;
 372         case CPU_DEAD:
 373         case CPU_UP_CANCELED:
 374                 dst = per_cpu(zswap_dstmem, cpu);
 375                 kfree(dst);
 376                 per_cpu(zswap_dstmem, cpu) = NULL;
 377                 break;
 378         default:
 379                 break;
 380         }
 381         return NOTIFY_OK;
 382 }
 383
 384 static int zswap_cpu_dstmem_notifier(struct notifier_block *nb,
 385                                      unsigned long action, void *pcpu)
 386 {
 387         return __zswap_cpu_dstmem_notifier(action, (unsigned long)pcpu);
 388 }
 389
 390 static struct notifier_block zswap_dstmem_notifier = {
 391         .notifier_call =        zswap_cpu_dstmem_notifier,
 392 };
 393
 394 static int __init zswap_cpu_dstmem_init(void)
 395 {
 396         unsigned long cpu;
 397
 398         cpu_notifier_register_begin();
 399         for_each_online_cpu(cpu)
 400                 if (__zswap_cpu_dstmem_notifier(CPU_UP_PREPARE, cpu) ==
 401                     NOTIFY_BAD)
 402                         goto cleanup;
 403         __register_cpu_notifier(&zswap_dstmem_notifier);
 404         cpu_notifier_register_done();
 405         return 0;
 406
 407 cleanup:
 408         for_each_online_cpu(cpu)
 409                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 410         cpu_notifier_register_done();
 411         return -ENOMEM;
 412 }
 413
 414 static void zswap_cpu_dstmem_destroy(void)
 415 {
 416         unsigned long cpu;
 417
 418         cpu_notifier_register_begin();
 419         for_each_online_cpu(cpu)
 420                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 421         __unregister_cpu_notifier(&zswap_dstmem_notifier);
 422         cpu_notifier_register_done();
 423 }
 424
 425 static int __zswap_cpu_comp_notifier(struct zswap_pool *pool,
 426                                      unsigned long action, unsigned long cpu)
 427 {
 428         struct crypto_comp *tfm;
 429
 430         switch (action) {
 431         case CPU_UP_PREPARE:
 432                 if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 433                         break;
 434                 tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 435                 if (IS_ERR_OR_NULL(tfm)) {
 436                         pr_err("could not alloc crypto comp %s : %ld\n",
 437                                pool->tfm_name, PTR_ERR(tfm));
 438                         return NOTIFY_BAD;
 439                 }
 440                 *per_cpu_ptr(pool->tfm, cpu) = tfm;
 441                 break;
 442         case CPU_DEAD:
 443         case CPU_UP_CANCELED:
 444                 tfm = *per_cpu_ptr(pool->tfm, cpu);
 445                 if (!IS_ERR_OR_NULL(tfm))
 446                         crypto_free_comp(tfm);
 447                 *per_cpu_ptr(pool->tfm, cpu) = NULL;
 448                 break;
 449         default:
 450                 break;
 451         }
 452         return NOTIFY_OK;
 453 }
 454
 455 static int zswap_cpu_comp_notifier(struct notifier_block *nb,
 456                                    unsigned long action, void *pcpu)
 457 {
 458         unsigned long cpu = (unsigned long)pcpu;
 459         struct zswap_pool *pool = container_of(nb, typeof(*pool), notifier);
 460
 461         return __zswap_cpu_comp_notifier(pool, action, cpu);
 462 }
 463
 464 static int zswap_cpu_comp_init(struct zswap_pool *pool)
 465 {
 466         unsigned long cpu;
 467
 468         memset(&pool->notifier, 0, sizeof(pool->notifier));
 469         pool->notifier.notifier_call = zswap_cpu_comp_notifier;
 470
 471         cpu_notifier_register_begin();
 472         for_each_online_cpu(cpu)
 473                 if (__zswap_cpu_comp_notifier(pool, CPU_UP_PREPARE, cpu) ==
 474                     NOTIFY_BAD)
 475                         goto cleanup;
 476         __register_cpu_notifier(&pool->notifier);
 477         cpu_notifier_register_done();
 478         return 0;
 479
 480 cleanup:
 481         for_each_online_cpu(cpu)
 482                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 483         cpu_notifier_register_done();
 484         return -ENOMEM;
 485 }
 486
 487 static void zswap_cpu_comp_destroy(struct zswap_pool *pool)
 488 {
 489         unsigned long cpu;
 490
 491         cpu_notifier_register_begin();
 492         for_each_online_cpu(cpu)
 493                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 494         __unregister_cpu_notifier(&pool->notifier);
 495         cpu_notifier_register_done();
 496 }
 497
 498 /*********************************
 499 * pool functions
 500 **********************************/
 501
 502 static struct zswap_pool *__zswap_pool_current(void)
 503 {
 504         struct zswap_pool *pool;
 505
 506         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 507         WARN_ON(!pool);
 508
 509         return pool;
 510 }
 511
 512 static struct zswap_pool *zswap_pool_current(void)
 513 {
 514         assert_spin_locked(&zswap_pools_lock);
 515
 516         return __zswap_pool_current();
 517 }
 518
 519 static struct zswap_pool *zswap_pool_current_get(void)
 520 {
 521         struct zswap_pool *pool;
 522
 523         rcu_read_lock();
 524
 525         pool = __zswap_pool_current();
 526         if (!pool || !zswap_pool_get(pool))
 527                 pool = NULL;
 528
 529         rcu_read_unlock();
 530
 531         return pool;
 532 }
 533
 534 static struct zswap_pool *zswap_pool_last_get(void)
 535 {
 536         struct zswap_pool *pool, *last = NULL;
 537
 538         rcu_read_lock();
 539
 540         list_for_each_entry_rcu(pool, &zswap_pools, list)
 541                 last = pool;
 542         if (!WARN_ON(!last) && !zswap_pool_get(last))
 543                 last = NULL;
 544
 545         rcu_read_unlock();
 546
 547         return last;
 548 }
 549
 550 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 551 {
 552         struct zswap_pool *pool;
 553
 554         assert_spin_locked(&zswap_pools_lock);
 555
 556         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 557                 if (strncmp(pool->tfm_name, compressor, sizeof(pool->tfm_name)))
 558                         continue;
 559                 if (strncmp(zpool_get_type(pool->zpool), type,
 560                             sizeof(zswap_zpool_type)))
 561                         continue;
 562                 /* if we can't get it, it's about to be destroyed */
 563                 if (!zswap_pool_get(pool))
 564                         continue;
 565                 return pool;
 566         }
 567
 568         return NULL;
 569 }
 570
 571 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 572 {
 573         struct zswap_pool *pool;
 574         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN;
 575
 576         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 577         if (!pool) {
 578                 pr_err("pool alloc failed\n");
 579                 return NULL;
 580         }
 581
 582         pool->zpool = zpool_create_pool(type, "zswap", gfp, &zswap_zpool_ops);
 583         if (!pool->zpool) {
 584                 pr_err("%s zpool not available\n", type);
 585                 goto error;
 586         }
 587         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 588
 589         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 590         pool->tfm = alloc_percpu(struct crypto_comp *);
 591         if (!pool->tfm) {
 592                 pr_err("percpu alloc failed\n");
 593                 goto error;
 594         }
 595
 596         if (zswap_cpu_comp_init(pool))
 597                 goto error;
 598         pr_debug("using %s compressor\n", pool->tfm_name);
 599
 600         /* being the current pool takes 1 ref; this func expects the
 601          * caller to always add the new pool as the current pool
 602          */
 603         kref_init(&pool->kref);
 604         INIT_LIST_HEAD(&pool->list);
 605
 606         zswap_pool_debug("created", pool);
 607
 608         return pool;
 609
 610 error:
 611         free_percpu(pool->tfm);
 612         if (pool->zpool)
 613                 zpool_destroy_pool(pool->zpool);
 614         kfree(pool);
 615         return NULL;
 616 }
 617
 618 static struct zswap_pool *__zswap_pool_create_fallback(void)
 619 {
 620         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 621                 pr_err("compressor %s not available, using default %s\n",
 622                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 623                 strncpy(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT,
 624                         sizeof(zswap_compressor));
 625         }
 626         if (!zpool_has_pool(zswap_zpool_type)) {
 627                 pr_err("zpool %s not available, using default %s\n",
 628                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 629                 strncpy(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT,
 630                         sizeof(zswap_zpool_type));
 631         }
 632
 633         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 634 }
 635
 636 static void zswap_pool_destroy(struct zswap_pool *pool)
 637 {
 638         zswap_pool_debug("destroying", pool);
 639
 640         zswap_cpu_comp_destroy(pool);
 641         free_percpu(pool->tfm);
 642         zpool_destroy_pool(pool->zpool);
 643         kfree(pool);
 644 }
 645
 646 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 647 {
 648         return kref_get_unless_zero(&pool->kref);
 649 }
 650
 651 static void __zswap_pool_release(struct rcu_head *head)
 652 {
 653         struct zswap_pool *pool = container_of(head, typeof(*pool), rcu_head);
 654
 655         /* nobody should have been able to get a kref... */
 656         WARN_ON(kref_get_unless_zero(&pool->kref));
 657
 658         /* pool is now off zswap_pools list and has no references. */
 659         zswap_pool_destroy(pool);
 660 }
 661
 662 static void __zswap_pool_empty(struct kref *kref)
 663 {
 664         struct zswap_pool *pool;
 665
 666         pool = container_of(kref, typeof(*pool), kref);
 667
 668         spin_lock(&zswap_pools_lock);
 669
 670         WARN_ON(pool == zswap_pool_current());
 671
 672         list_del_rcu(&pool->list);
 673         call_rcu(&pool->rcu_head, __zswap_pool_release);
 674
 675         spin_unlock(&zswap_pools_lock);
 676 }
 677
 678 static void zswap_pool_put(struct zswap_pool *pool)
 679 {
 680         kref_put(&pool->kref, __zswap_pool_empty);
 681 }
 682
 683 /*********************************
 684 * param callbacks
 685 **********************************/
 686
 687 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 688                              char *type, char *compressor)
 689 {
 690         struct zswap_pool *pool, *put_pool = NULL;
 691         char str[kp->str->maxlen], *s;
 692         int ret;
 693
 694         /*
 695          * kp is either zswap_zpool_kparam or zswap_compressor_kparam, defined
 696          * at the top of this file, so maxlen is CRYPTO_MAX_ALG_NAME (64) or
 697          * 32 (arbitrary).
 698          */
 699         strlcpy(str, val, kp->str->maxlen);
 700         s = strim(str);
 701
 702         /* if this is load-time (pre-init) param setting,
 703          * don't create a pool; that's done during init.
 704          */
 705         if (!zswap_init_started)
 706                 return param_set_copystring(s, kp);
 707
 708         /* no change required */
 709         if (!strncmp(kp->str->string, s, kp->str->maxlen))
 710                 return 0;
 711
 712         if (!type) {
 713                 type = s;
 714                 if (!zpool_has_pool(type)) {
 715                         pr_err("zpool %s not available\n", type);
 716                         return -ENOENT;
 717                 }
 718         } else if (!compressor) {
 719                 compressor = s;
 720                 if (!crypto_has_comp(compressor, 0, 0)) {
 721                         pr_err("compressor %s not available\n", compressor);
 722                         return -ENOENT;
 723                 }
 724         }
 725
 726         spin_lock(&zswap_pools_lock);
 727
 728         pool = zswap_pool_find_get(type, compressor);
 729         if (pool) {
 730                 zswap_pool_debug("using existing", pool);
 731                 list_del_rcu(&pool->list);
 732         } else {
 733                 spin_unlock(&zswap_pools_lock);
 734                 pool = zswap_pool_create(type, compressor);
 735                 spin_lock(&zswap_pools_lock);
 736         }
 737
 738         if (pool)
 739                 ret = param_set_copystring(s, kp);
 740         else
 741                 ret = -EINVAL;
 742
 743         if (!ret) {
 744                 put_pool = zswap_pool_current();
 745                 list_add_rcu(&pool->list, &zswap_pools);
 746         } else if (pool) {
 747                 /* add the possibly pre-existing pool to the end of the pools
 748                  * list; if it's new (and empty) then it'll be removed and
 749                  * destroyed by the put after we drop the lock
 750                  */
 751                 list_add_tail_rcu(&pool->list, &zswap_pools);
 752                 put_pool = pool;
 753         }
 754
 755         spin_unlock(&zswap_pools_lock);
 756
 757         /* drop the ref from either the old current pool,
 758          * or the new pool we failed to add
 759          */
 760         if (put_pool)
 761                 zswap_pool_put(put_pool);
 762
 763         return ret;
 764 }
 765
 766 static int zswap_compressor_param_set(const char *val,
 767                                       const struct kernel_param *kp)
 768 {
 769         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
 770 }
 771
 772 static int zswap_zpool_param_set(const char *val,
 773                                  const struct kernel_param *kp)
 774 {
 775         return __zswap_param_set(val, kp, NULL, zswap_compressor);
 776 }
 777
 778 /*********************************
 779 * writeback code
 780 **********************************/
 781 /* return enum for zswap_get_swap_cache_page */
 782 enum zswap_get_swap_ret {
 783         ZSWAP_SWAPCACHE_NEW,
 784         ZSWAP_SWAPCACHE_EXIST,
 785         ZSWAP_SWAPCACHE_FAIL,
 786 };
 787
 788 /*
 789  * zswap_get_swap_cache_page
 790  *
 791  * This is an adaption of read_swap_cache_async()
 792  *
 793  * This function tries to find a page with the given swap entry
 794  * in the swapper_space address space (the swap cache).  If the page
 795  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 796  * added to the swap cache, and returned in retpage.
 797  *
 798  * If success, the swap cache page is returned in retpage
 799  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 800  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 801  *     the new page is added to swapcache and locked
 802  * Returns ZSWAP_SWAPCACHE_FAIL on error
 803  */
 804 static int zswap_get_swap_cache_page(swp_entry_t entry,
 805                                 struct page **retpage)
 806 {
 807         bool page_was_allocated;
 808
 809         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 810                         NULL, 0, &page_was_allocated);
 811         if (page_was_allocated)
 812                 return ZSWAP_SWAPCACHE_NEW;
 813         if (!*retpage)
 814                 return ZSWAP_SWAPCACHE_FAIL;
 815         return ZSWAP_SWAPCACHE_EXIST;
 816 }
 817
 818 /*
 819  * Attempts to free an entry by adding a page to the swap cache,
 820  * decompressing the entry data into the page, and issuing a
 821  * bio write to write the page back to the swap device.
 822  *
 823  * This can be thought of as a "resumed writeback" of the page
 824  * to the swap device.  We are basically resuming the same swap
 825  * writeback path that was intercepted with the frontswap_store()
 826  * in the first place.  After the page has been decompressed into
 827  * the swap cache, the compressed version stored by zswap can be
 828  * freed.
 829  */
 830 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 831 {
 832         struct zswap_header *zhdr;
 833         swp_entry_t swpentry;
 834         struct zswap_tree *tree;
 835         pgoff_t offset;
 836         struct zswap_entry *entry;
 837         struct page *page;
 838         struct crypto_comp *tfm;
 839         u8 *src, *dst;
 840         unsigned int dlen;
 841         int ret;
 842         struct writeback_control wbc = {
 843                 .sync_mode = WB_SYNC_NONE,
 844         };
 845
 846         /* extract swpentry from data */
 847         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 848         swpentry = zhdr->swpentry; /* here */
 849         zpool_unmap_handle(pool, handle);
 850         tree = zswap_trees[swp_type(swpentry)];
 851         offset = swp_offset(swpentry);
 852
 853         /* find and ref zswap entry */
 854         spin_lock(&tree->lock);
 855         entry = zswap_entry_find_get(&tree->rbroot, offset);
 856         if (!entry) {
 857                 /* entry was invalidated */
 858                 spin_unlock(&tree->lock);
 859                 return 0;
 860         }
 861         spin_unlock(&tree->lock);
 862         BUG_ON(offset != entry->offset);
 863
 864         /* try to allocate swap cache page */
 865         switch (zswap_get_swap_cache_page(swpentry, &page)) {
 866         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
 867                 ret = -ENOMEM;
 868                 goto fail;
 869
 870         case ZSWAP_SWAPCACHE_EXIST:
 871                 /* page is already in the swap cache, ignore for now */
 872                 page_cache_release(page);
 873                 ret = -EEXIST;
 874                 goto fail;
 875
 876         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
 877                 /* decompress */
 878                 dlen = PAGE_SIZE;
 879                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 880                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
 881                 dst = kmap_atomic(page);
 882                 tfm = *get_cpu_ptr(entry->pool->tfm);
 883                 ret = crypto_comp_decompress(tfm, src, entry->length,
 884                                              dst, &dlen);
 885                 put_cpu_ptr(entry->pool->tfm);
 886                 kunmap_atomic(dst);
 887                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
 888                 BUG_ON(ret);
 889                 BUG_ON(dlen != PAGE_SIZE);
 890
 891                 /* page is up to date */
 892                 SetPageUptodate(page);
 893         }
 894
 895         /* move it to the tail of the inactive list after end_writeback */
 896         SetPageReclaim(page);
 897
 898         /* start writeback */
 899         __swap_writepage(page, &wbc, end_swap_bio_write);
 900         page_cache_release(page);
 901         zswap_written_back_pages++;
 902
 903         spin_lock(&tree->lock);
 904         /* drop local reference */
 905         zswap_entry_put(tree, entry);
 906
 907         /*
 908         * There are two possible situations for entry here:
 909         * (1) refcount is 1(normal case),  entry is valid and on the tree
 910         * (2) refcount is 0, entry is freed and not on the tree
 911         *     because invalidate happened during writeback
 912         *  search the tree and free the entry if find entry
 913         */
 914         if (entry == zswap_rb_search(&tree->rbroot, offset))
 915                 zswap_entry_put(tree, entry);
 916         spin_unlock(&tree->lock);
 917
 918         goto end;
 919
 920         /*
 921         * if we get here due to ZSWAP_SWAPCACHE_EXIST
 922         * a load may happening concurrently
 923         * it is safe and okay to not free the entry
 924         * if we free the entry in the following put
 925         * it it either okay to return !0
 926         */
 927 fail:
 928         spin_lock(&tree->lock);
 929         zswap_entry_put(tree, entry);
 930         spin_unlock(&tree->lock);
 931
 932 end:
 933         return ret;
 934 }
 935
 936 static int zswap_shrink(void)
 937 {
 938         struct zswap_pool *pool;
 939         int ret;
 940
 941         pool = zswap_pool_last_get();
 942         if (!pool)
 943                 return -ENOENT;
 944
 945         ret = zpool_shrink(pool->zpool, 1, NULL);
 946
 947         zswap_pool_put(pool);
 948
 949         return ret;
 950 }
 951
 952 /*********************************
 953 * frontswap hooks
 954 **********************************/
 955 /* attempts to compress and store an single page */
 956 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 957                                 struct page *page)
 958 {
 959         struct zswap_tree *tree = zswap_trees[type];
 960         struct zswap_entry *entry, *dupentry;
 961         struct crypto_comp *tfm;
 962         int ret;
 963         unsigned int dlen = PAGE_SIZE, len;
 964         unsigned long handle;
 965         char *buf;
 966         u8 *src, *dst;
 967         struct zswap_header *zhdr;
 968
 969         if (!zswap_enabled || !tree) {
 970                 ret = -ENODEV;
 971                 goto reject;
 972         }
 973
 974         /* reclaim space if needed */
 975         if (zswap_is_full()) {
 976                 zswap_pool_limit_hit++;
 977                 if (zswap_shrink()) {
 978                         zswap_reject_reclaim_fail++;
 979                         ret = -ENOMEM;
 980                         goto reject;
 981                 }
 982         }
 983
 984         /* allocate entry */
 985         entry = zswap_entry_cache_alloc(GFP_KERNEL);
 986         if (!entry) {
 987                 zswap_reject_kmemcache_fail++;
 988                 ret = -ENOMEM;
 989                 goto reject;
 990         }
 991
 992         /* if entry is successfully added, it keeps the reference */
 993         entry->pool = zswap_pool_current_get();
 994         if (!entry->pool) {
 995                 ret = -EINVAL;
 996                 goto freepage;
 997         }
 998
 999         /* compress */
1000         dst = get_cpu_var(zswap_dstmem);
1001         tfm = *get_cpu_ptr(entry->pool->tfm);
1002         src = kmap_atomic(page);
1003         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
1004         kunmap_atomic(src);
1005         put_cpu_ptr(entry->pool->tfm);
1006         if (ret) {
1007                 ret = -EINVAL;
1008                 goto put_dstmem;
1009         }
1010
1011         /* store */
1012         len = dlen + sizeof(struct zswap_header);
1013         ret = zpool_malloc(entry->pool->zpool, len,
1014                            __GFP_NORETRY | __GFP_NOWARN, &handle);
1015         if (ret == -ENOSPC) {
1016                 zswap_reject_compress_poor++;
1017                 goto put_dstmem;
1018         }
1019         if (ret) {
1020                 zswap_reject_alloc_fail++;
1021                 goto put_dstmem;
1022         }
1023         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1024         zhdr->swpentry = swp_entry(type, offset);
1025         buf = (u8 *)(zhdr + 1);
1026         memcpy(buf, dst, dlen);
1027         zpool_unmap_handle(entry->pool->zpool, handle);
1028         put_cpu_var(zswap_dstmem);
1029
1030         /* populate entry */
1031         entry->offset = offset;
1032         entry->handle = handle;
1033         entry->length = dlen;
1034
1035         /* map */
1036         spin_lock(&tree->lock);
1037         do {
1038                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1039                 if (ret == -EEXIST) {
1040                         zswap_duplicate_entry++;
1041                         /* remove from rbtree */
1042                         zswap_rb_erase(&tree->rbroot, dupentry);
1043                         zswap_entry_put(tree, dupentry);
1044                 }
1045         } while (ret == -EEXIST);
1046         spin_unlock(&tree->lock);
1047
1048         /* update stats */
1049         atomic_inc(&zswap_stored_pages);
1050         zswap_update_total_size();
1051
1052         return 0;
1053
1054 put_dstmem:
1055         put_cpu_var(zswap_dstmem);
1056         zswap_pool_put(entry->pool);
1057 freepage:
1058         zswap_entry_cache_free(entry);
1059 reject:
1060         return ret;
1061 }
1062
1063 /*
1064  * returns 0 if the page was successfully decompressed
1065  * return -1 on entry not found or error
1066 */
1067 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1068                                 struct page *page)
1069 {
1070         struct zswap_tree *tree = zswap_trees[type];
1071         struct zswap_entry *entry;
1072         struct crypto_comp *tfm;
1073         u8 *src, *dst;
1074         unsigned int dlen;
1075         int ret;
1076
1077         /* find */
1078         spin_lock(&tree->lock);
1079         entry = zswap_entry_find_get(&tree->rbroot, offset);
1080         if (!entry) {
1081                 /* entry was written back */
1082                 spin_unlock(&tree->lock);
1083                 return -1;
1084         }
1085         spin_unlock(&tree->lock);
1086
1087         /* decompress */
1088         dlen = PAGE_SIZE;
1089         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1090                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1091         dst = kmap_atomic(page);
1092         tfm = *get_cpu_ptr(entry->pool->tfm);
1093         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1094         put_cpu_ptr(entry->pool->tfm);
1095         kunmap_atomic(dst);
1096         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1097         BUG_ON(ret);
1098
1099         spin_lock(&tree->lock);
1100         zswap_entry_put(tree, entry);
1101         spin_unlock(&tree->lock);
1102
1103         return 0;
1104 }
1105
1106 /* frees an entry in zswap */
1107 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1108 {
1109         struct zswap_tree *tree = zswap_trees[type];
1110         struct zswap_entry *entry;
1111
1112         /* find */
1113         spin_lock(&tree->lock);
1114         entry = zswap_rb_search(&tree->rbroot, offset);
1115         if (!entry) {
1116                 /* entry was written back */
1117                 spin_unlock(&tree->lock);
1118                 return;
1119         }
1120
1121         /* remove from rbtree */
1122         zswap_rb_erase(&tree->rbroot, entry);
1123
1124         /* drop the initial reference from entry creation */
1125         zswap_entry_put(tree, entry);
1126
1127         spin_unlock(&tree->lock);
1128 }
1129
1130 /* frees all zswap entries for the given swap type */
1131 static void zswap_frontswap_invalidate_area(unsigned type)
1132 {
1133         struct zswap_tree *tree = zswap_trees[type];
1134         struct zswap_entry *entry, *n;
1135
1136         if (!tree)
1137                 return;
1138
1139         /* walk the tree and free everything */
1140         spin_lock(&tree->lock);
1141         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1142                 zswap_free_entry(entry);
1143         tree->rbroot = RB_ROOT;
1144         spin_unlock(&tree->lock);
1145         kfree(tree);
1146         zswap_trees[type] = NULL;
1147 }
1148
1149 static void zswap_frontswap_init(unsigned type)
1150 {
1151         struct zswap_tree *tree;
1152
1153         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1154         if (!tree) {
1155                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1156                 return;
1157         }
1158
1159         tree->rbroot = RB_ROOT;
1160         spin_lock_init(&tree->lock);
1161         zswap_trees[type] = tree;
1162 }
1163
1164 static struct frontswap_ops zswap_frontswap_ops = {
1165         .store = zswap_frontswap_store,
1166         .load = zswap_frontswap_load,
1167         .invalidate_page = zswap_frontswap_invalidate_page,
1168         .invalidate_area = zswap_frontswap_invalidate_area,
1169         .init = zswap_frontswap_init
1170 };
1171
1172 /*********************************
1173 * debugfs functions
1174 **********************************/
1175 #ifdef CONFIG_DEBUG_FS
1176 #include <linux/debugfs.h>
1177
1178 static struct dentry *zswap_debugfs_root;
1179
1180 static int __init zswap_debugfs_init(void)
1181 {
1182         if (!debugfs_initialized())
1183                 return -ENODEV;
1184
1185         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1186         if (!zswap_debugfs_root)
1187                 return -ENOMEM;
1188
1189         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1190                         zswap_debugfs_root, &zswap_pool_limit_hit);
1191         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1192                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1193         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1194                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1195         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1196                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1197         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1198                         zswap_debugfs_root, &zswap_reject_compress_poor);
1199         debugfs_create_u64("written_back_pages", S_IRUGO,
1200                         zswap_debugfs_root, &zswap_written_back_pages);
1201         debugfs_create_u64("duplicate_entry", S_IRUGO,
1202                         zswap_debugfs_root, &zswap_duplicate_entry);
1203         debugfs_create_u64("pool_total_size", S_IRUGO,
1204                         zswap_debugfs_root, &zswap_pool_total_size);
1205         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1206                         zswap_debugfs_root, &zswap_stored_pages);
1207
1208         return 0;
1209 }
1210
1211 static void __exit zswap_debugfs_exit(void)
1212 {
1213         debugfs_remove_recursive(zswap_debugfs_root);
1214 }
1215 #else
1216 static int __init zswap_debugfs_init(void)
1217 {
1218         return 0;
1219 }
1220
1221 static void __exit zswap_debugfs_exit(void) { }
1222 #endif
1223
1224 /*********************************
1225 * module init and exit
1226 **********************************/
1227 static int __init init_zswap(void)
1228 {
1229         struct zswap_pool *pool;
1230
1231         zswap_init_started = true;
1232
1233         if (zswap_entry_cache_create()) {
1234                 pr_err("entry cache creation failed\n");
1235                 goto cache_fail;
1236         }
1237
1238         if (zswap_cpu_dstmem_init()) {
1239                 pr_err("dstmem alloc failed\n");
1240                 goto dstmem_fail;
1241         }
1242
1243         pool = __zswap_pool_create_fallback();
1244         if (!pool) {
1245                 pr_err("pool creation failed\n");
1246                 goto pool_fail;
1247         }
1248         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1249                 zpool_get_type(pool->zpool));
1250
1251         list_add(&pool->list, &zswap_pools);
1252
1253         frontswap_register_ops(&zswap_frontswap_ops);
1254         if (zswap_debugfs_init())
1255                 pr_warn("debugfs initialization failed\n");
1256         return 0;
1257
1258 pool_fail:
1259         zswap_cpu_dstmem_destroy();
1260 dstmem_fail:
1261         zswap_entry_cache_destroy();
1262 cache_fail:
1263         return -ENOMEM;
1264 }
1265 /* must be late so crypto has time to come up */
1266 late_initcall(init_zswap);
1267
1268 MODULE_LICENSE("GPL");
1269 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1270 MODULE_DESCRIPTION("Compressed cache for swap pages");