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 rcu_head rcu_head;
 121         struct notifier_block notifier;
 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
 174 /* used by param callback function */
 175 static bool zswap_init_started;
 176
 177 /*********************************
 178 * helpers and fwd declarations
 179 **********************************/
 180
 181 #define zswap_pool_debug(msg, p)                                \
 182         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 183                  zpool_get_type((p)->zpool))
 184
 185 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 186 static int zswap_pool_get(struct zswap_pool *pool);
 187 static void zswap_pool_put(struct zswap_pool *pool);
 188
 189 static const struct zpool_ops zswap_zpool_ops = {
 190         .evict = zswap_writeback_entry
 191 };
 192
 193 static bool zswap_is_full(void)
 194 {
 195         return totalram_pages * zswap_max_pool_percent / 100 <
 196                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 197 }
 198
 199 static void zswap_update_total_size(void)
 200 {
 201         struct zswap_pool *pool;
 202         u64 total = 0;
 203
 204         rcu_read_lock();
 205
 206         list_for_each_entry_rcu(pool, &zswap_pools, list)
 207                 total += zpool_get_total_size(pool->zpool);
 208
 209         rcu_read_unlock();
 210
 211         zswap_pool_total_size = total;
 212 }
 213
 214 /*********************************
 215 * zswap entry functions
 216 **********************************/
 217 static struct kmem_cache *zswap_entry_cache;
 218
 219 static int __init zswap_entry_cache_create(void)
 220 {
 221         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 222         return zswap_entry_cache == NULL;
 223 }
 224
 225 static void __init zswap_entry_cache_destroy(void)
 226 {
 227         kmem_cache_destroy(zswap_entry_cache);
 228 }
 229
 230 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 231 {
 232         struct zswap_entry *entry;
 233         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 234         if (!entry)
 235                 return NULL;
 236         entry->refcount = 1;
 237         RB_CLEAR_NODE(&entry->rbnode);
 238         return entry;
 239 }
 240
 241 static void zswap_entry_cache_free(struct zswap_entry *entry)
 242 {
 243         kmem_cache_free(zswap_entry_cache, entry);
 244 }
 245
 246 /*********************************
 247 * rbtree functions
 248 **********************************/
 249 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 250 {
 251         struct rb_node *node = root->rb_node;
 252         struct zswap_entry *entry;
 253
 254         while (node) {
 255                 entry = rb_entry(node, struct zswap_entry, rbnode);
 256                 if (entry->offset > offset)
 257                         node = node->rb_left;
 258                 else if (entry->offset < offset)
 259                         node = node->rb_right;
 260                 else
 261                         return entry;
 262         }
 263         return NULL;
 264 }
 265
 266 /*
 267  * In the case that a entry with the same offset is found, a pointer to
 268  * the existing entry is stored in dupentry and the function returns -EEXIST
 269  */
 270 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 271                         struct zswap_entry **dupentry)
 272 {
 273         struct rb_node **link = &root->rb_node, *parent = NULL;
 274         struct zswap_entry *myentry;
 275
 276         while (*link) {
 277                 parent = *link;
 278                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 279                 if (myentry->offset > entry->offset)
 280                         link = &(*link)->rb_left;
 281                 else if (myentry->offset < entry->offset)
 282                         link = &(*link)->rb_right;
 283                 else {
 284                         *dupentry = myentry;
 285                         return -EEXIST;
 286                 }
 287         }
 288         rb_link_node(&entry->rbnode, parent, link);
 289         rb_insert_color(&entry->rbnode, root);
 290         return 0;
 291 }
 292
 293 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 294 {
 295         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 296                 rb_erase(&entry->rbnode, root);
 297                 RB_CLEAR_NODE(&entry->rbnode);
 298         }
 299 }
 300
 301 /*
 302  * Carries out the common pattern of freeing and entry's zpool allocation,
 303  * freeing the entry itself, and decrementing the number of stored pages.
 304  */
 305 static void zswap_free_entry(struct zswap_entry *entry)
 306 {
 307         zpool_free(entry->pool->zpool, entry->handle);
 308         zswap_pool_put(entry->pool);
 309         zswap_entry_cache_free(entry);
 310         atomic_dec(&zswap_stored_pages);
 311         zswap_update_total_size();
 312 }
 313
 314 /* caller must hold the tree lock */
 315 static void zswap_entry_get(struct zswap_entry *entry)
 316 {
 317         entry->refcount++;
 318 }
 319
 320 /* caller must hold the tree lock
 321 * remove from the tree and free it, if nobody reference the entry
 322 */
 323 static void zswap_entry_put(struct zswap_tree *tree,
 324                         struct zswap_entry *entry)
 325 {
 326         int refcount = --entry->refcount;
 327
 328         BUG_ON(refcount < 0);
 329         if (refcount == 0) {
 330                 zswap_rb_erase(&tree->rbroot, entry);
 331                 zswap_free_entry(entry);
 332         }
 333 }
 334
 335 /* caller must hold the tree lock */
 336 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 337                                 pgoff_t offset)
 338 {
 339         struct zswap_entry *entry;
 340
 341         entry = zswap_rb_search(root, offset);
 342         if (entry)
 343                 zswap_entry_get(entry);
 344
 345         return entry;
 346 }
 347
 348 /*********************************
 349 * per-cpu code
 350 **********************************/
 351 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 352
 353 static int __zswap_cpu_dstmem_notifier(unsigned long action, unsigned long cpu)
 354 {
 355         u8 *dst;
 356
 357         switch (action) {
 358         case CPU_UP_PREPARE:
 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 NOTIFY_BAD;
 363                 }
 364                 per_cpu(zswap_dstmem, cpu) = dst;
 365                 break;
 366         case CPU_DEAD:
 367         case CPU_UP_CANCELED:
 368                 dst = per_cpu(zswap_dstmem, cpu);
 369                 kfree(dst);
 370                 per_cpu(zswap_dstmem, cpu) = NULL;
 371                 break;
 372         default:
 373                 break;
 374         }
 375         return NOTIFY_OK;
 376 }
 377
 378 static int zswap_cpu_dstmem_notifier(struct notifier_block *nb,
 379                                      unsigned long action, void *pcpu)
 380 {
 381         return __zswap_cpu_dstmem_notifier(action, (unsigned long)pcpu);
 382 }
 383
 384 static struct notifier_block zswap_dstmem_notifier = {
 385         .notifier_call =        zswap_cpu_dstmem_notifier,
 386 };
 387
 388 static int __init zswap_cpu_dstmem_init(void)
 389 {
 390         unsigned long cpu;
 391
 392         cpu_notifier_register_begin();
 393         for_each_online_cpu(cpu)
 394                 if (__zswap_cpu_dstmem_notifier(CPU_UP_PREPARE, cpu) ==
 395                     NOTIFY_BAD)
 396                         goto cleanup;
 397         __register_cpu_notifier(&zswap_dstmem_notifier);
 398         cpu_notifier_register_done();
 399         return 0;
 400
 401 cleanup:
 402         for_each_online_cpu(cpu)
 403                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 404         cpu_notifier_register_done();
 405         return -ENOMEM;
 406 }
 407
 408 static void zswap_cpu_dstmem_destroy(void)
 409 {
 410         unsigned long cpu;
 411
 412         cpu_notifier_register_begin();
 413         for_each_online_cpu(cpu)
 414                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 415         __unregister_cpu_notifier(&zswap_dstmem_notifier);
 416         cpu_notifier_register_done();
 417 }
 418
 419 static int __zswap_cpu_comp_notifier(struct zswap_pool *pool,
 420                                      unsigned long action, unsigned long cpu)
 421 {
 422         struct crypto_comp *tfm;
 423
 424         switch (action) {
 425         case CPU_UP_PREPARE:
 426                 if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 427                         break;
 428                 tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 429                 if (IS_ERR_OR_NULL(tfm)) {
 430                         pr_err("could not alloc crypto comp %s : %ld\n",
 431                                pool->tfm_name, PTR_ERR(tfm));
 432                         return NOTIFY_BAD;
 433                 }
 434                 *per_cpu_ptr(pool->tfm, cpu) = tfm;
 435                 break;
 436         case CPU_DEAD:
 437         case CPU_UP_CANCELED:
 438                 tfm = *per_cpu_ptr(pool->tfm, cpu);
 439                 if (!IS_ERR_OR_NULL(tfm))
 440                         crypto_free_comp(tfm);
 441                 *per_cpu_ptr(pool->tfm, cpu) = NULL;
 442                 break;
 443         default:
 444                 break;
 445         }
 446         return NOTIFY_OK;
 447 }
 448
 449 static int zswap_cpu_comp_notifier(struct notifier_block *nb,
 450                                    unsigned long action, void *pcpu)
 451 {
 452         unsigned long cpu = (unsigned long)pcpu;
 453         struct zswap_pool *pool = container_of(nb, typeof(*pool), notifier);
 454
 455         return __zswap_cpu_comp_notifier(pool, action, cpu);
 456 }
 457
 458 static int zswap_cpu_comp_init(struct zswap_pool *pool)
 459 {
 460         unsigned long cpu;
 461
 462         memset(&pool->notifier, 0, sizeof(pool->notifier));
 463         pool->notifier.notifier_call = zswap_cpu_comp_notifier;
 464
 465         cpu_notifier_register_begin();
 466         for_each_online_cpu(cpu)
 467                 if (__zswap_cpu_comp_notifier(pool, CPU_UP_PREPARE, cpu) ==
 468                     NOTIFY_BAD)
 469                         goto cleanup;
 470         __register_cpu_notifier(&pool->notifier);
 471         cpu_notifier_register_done();
 472         return 0;
 473
 474 cleanup:
 475         for_each_online_cpu(cpu)
 476                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 477         cpu_notifier_register_done();
 478         return -ENOMEM;
 479 }
 480
 481 static void zswap_cpu_comp_destroy(struct zswap_pool *pool)
 482 {
 483         unsigned long cpu;
 484
 485         cpu_notifier_register_begin();
 486         for_each_online_cpu(cpu)
 487                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 488         __unregister_cpu_notifier(&pool->notifier);
 489         cpu_notifier_register_done();
 490 }
 491
 492 /*********************************
 493 * pool functions
 494 **********************************/
 495
 496 static struct zswap_pool *__zswap_pool_current(void)
 497 {
 498         struct zswap_pool *pool;
 499
 500         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 501         WARN_ON(!pool);
 502
 503         return pool;
 504 }
 505
 506 static struct zswap_pool *zswap_pool_current(void)
 507 {
 508         assert_spin_locked(&zswap_pools_lock);
 509
 510         return __zswap_pool_current();
 511 }
 512
 513 static struct zswap_pool *zswap_pool_current_get(void)
 514 {
 515         struct zswap_pool *pool;
 516
 517         rcu_read_lock();
 518
 519         pool = __zswap_pool_current();
 520         if (!pool || !zswap_pool_get(pool))
 521                 pool = NULL;
 522
 523         rcu_read_unlock();
 524
 525         return pool;
 526 }
 527
 528 static struct zswap_pool *zswap_pool_last_get(void)
 529 {
 530         struct zswap_pool *pool, *last = NULL;
 531
 532         rcu_read_lock();
 533
 534         list_for_each_entry_rcu(pool, &zswap_pools, list)
 535                 last = pool;
 536         if (!WARN_ON(!last) && !zswap_pool_get(last))
 537                 last = NULL;
 538
 539         rcu_read_unlock();
 540
 541         return last;
 542 }
 543
 544 /* type and compressor must be null-terminated */
 545 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 546 {
 547         struct zswap_pool *pool;
 548
 549         assert_spin_locked(&zswap_pools_lock);
 550
 551         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 552                 if (strcmp(pool->tfm_name, compressor))
 553                         continue;
 554                 if (strcmp(zpool_get_type(pool->zpool), type))
 555                         continue;
 556                 /* if we can't get it, it's about to be destroyed */
 557                 if (!zswap_pool_get(pool))
 558                         continue;
 559                 return pool;
 560         }
 561
 562         return NULL;
 563 }
 564
 565 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 566 {
 567         struct zswap_pool *pool;
 568         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
 569
 570         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 571         if (!pool) {
 572                 pr_err("pool alloc failed\n");
 573                 return NULL;
 574         }
 575
 576         pool->zpool = zpool_create_pool(type, "zswap", gfp, &zswap_zpool_ops);
 577         if (!pool->zpool) {
 578                 pr_err("%s zpool not available\n", type);
 579                 goto error;
 580         }
 581         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 582
 583         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 584         pool->tfm = alloc_percpu(struct crypto_comp *);
 585         if (!pool->tfm) {
 586                 pr_err("percpu alloc failed\n");
 587                 goto error;
 588         }
 589
 590         if (zswap_cpu_comp_init(pool))
 591                 goto error;
 592         pr_debug("using %s compressor\n", pool->tfm_name);
 593
 594         /* being the current pool takes 1 ref; this func expects the
 595          * caller to always add the new pool as the current pool
 596          */
 597         kref_init(&pool->kref);
 598         INIT_LIST_HEAD(&pool->list);
 599
 600         zswap_pool_debug("created", pool);
 601
 602         return pool;
 603
 604 error:
 605         free_percpu(pool->tfm);
 606         if (pool->zpool)
 607                 zpool_destroy_pool(pool->zpool);
 608         kfree(pool);
 609         return NULL;
 610 }
 611
 612 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
 613 {
 614         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 615                 if (!strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
 616                         pr_err("default compressor %s not available\n",
 617                                zswap_compressor);
 618                         return NULL;
 619                 }
 620                 pr_err("compressor %s not available, using default %s\n",
 621                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 622                 param_free_charp(&zswap_compressor);
 623                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
 624         }
 625         if (!zpool_has_pool(zswap_zpool_type)) {
 626                 if (!strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
 627                         pr_err("default zpool %s not available\n",
 628                                zswap_zpool_type);
 629                         return NULL;
 630                 }
 631                 pr_err("zpool %s not available, using default %s\n",
 632                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 633                 param_free_charp(&zswap_zpool_type);
 634                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 635         }
 636
 637         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 638 }
 639
 640 static void zswap_pool_destroy(struct zswap_pool *pool)
 641 {
 642         zswap_pool_debug("destroying", pool);
 643
 644         zswap_cpu_comp_destroy(pool);
 645         free_percpu(pool->tfm);
 646         zpool_destroy_pool(pool->zpool);
 647         kfree(pool);
 648 }
 649
 650 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 651 {
 652         return kref_get_unless_zero(&pool->kref);
 653 }
 654
 655 static void __zswap_pool_release(struct rcu_head *head)
 656 {
 657         struct zswap_pool *pool = container_of(head, typeof(*pool), rcu_head);
 658
 659         /* nobody should have been able to get a kref... */
 660         WARN_ON(kref_get_unless_zero(&pool->kref));
 661
 662         /* pool is now off zswap_pools list and has no references. */
 663         zswap_pool_destroy(pool);
 664 }
 665
 666 static void __zswap_pool_empty(struct kref *kref)
 667 {
 668         struct zswap_pool *pool;
 669
 670         pool = container_of(kref, typeof(*pool), kref);
 671
 672         spin_lock(&zswap_pools_lock);
 673
 674         WARN_ON(pool == zswap_pool_current());
 675
 676         list_del_rcu(&pool->list);
 677         call_rcu(&pool->rcu_head, __zswap_pool_release);
 678
 679         spin_unlock(&zswap_pools_lock);
 680 }
 681
 682 static void zswap_pool_put(struct zswap_pool *pool)
 683 {
 684         kref_put(&pool->kref, __zswap_pool_empty);
 685 }
 686
 687 /*********************************
 688 * param callbacks
 689 **********************************/
 690
 691 /* val must be a null-terminated string */
 692 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 693                              char *type, char *compressor)
 694 {
 695         struct zswap_pool *pool, *put_pool = NULL;
 696         char *s = strstrip((char *)val);
 697         int ret;
 698
 699         /* no change required */
 700         if (!strcmp(s, *(char **)kp->arg))
 701                 return 0;
 702
 703         /* if this is load-time (pre-init) param setting,
 704          * don't create a pool; that's done during init.
 705          */
 706         if (!zswap_init_started)
 707                 return param_set_charp(s, kp);
 708
 709         if (!type) {
 710                 if (!zpool_has_pool(s)) {
 711                         pr_err("zpool %s not available\n", s);
 712                         return -ENOENT;
 713                 }
 714                 type = s;
 715         } else if (!compressor) {
 716                 if (!crypto_has_comp(s, 0, 0)) {
 717                         pr_err("compressor %s not available\n", s);
 718                         return -ENOENT;
 719                 }
 720                 compressor = s;
 721         } else {
 722                 WARN_ON(1);
 723                 return -EINVAL;
 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_charp(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 | __GFP_KSWAPD_RECLAIM,
1015                            &handle);
1016         if (ret == -ENOSPC) {
1017                 zswap_reject_compress_poor++;
1018                 goto put_dstmem;
1019         }
1020         if (ret) {
1021                 zswap_reject_alloc_fail++;
1022                 goto put_dstmem;
1023         }
1024         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1025         zhdr->swpentry = swp_entry(type, offset);
1026         buf = (u8 *)(zhdr + 1);
1027         memcpy(buf, dst, dlen);
1028         zpool_unmap_handle(entry->pool->zpool, handle);
1029         put_cpu_var(zswap_dstmem);
1030
1031         /* populate entry */
1032         entry->offset = offset;
1033         entry->handle = handle;
1034         entry->length = dlen;
1035
1036         /* map */
1037         spin_lock(&tree->lock);
1038         do {
1039                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1040                 if (ret == -EEXIST) {
1041                         zswap_duplicate_entry++;
1042                         /* remove from rbtree */
1043                         zswap_rb_erase(&tree->rbroot, dupentry);
1044                         zswap_entry_put(tree, dupentry);
1045                 }
1046         } while (ret == -EEXIST);
1047         spin_unlock(&tree->lock);
1048
1049         /* update stats */
1050         atomic_inc(&zswap_stored_pages);
1051         zswap_update_total_size();
1052
1053         return 0;
1054
1055 put_dstmem:
1056         put_cpu_var(zswap_dstmem);
1057         zswap_pool_put(entry->pool);
1058 freepage:
1059         zswap_entry_cache_free(entry);
1060 reject:
1061         return ret;
1062 }
1063
1064 /*
1065  * returns 0 if the page was successfully decompressed
1066  * return -1 on entry not found or error
1067 */
1068 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1069                                 struct page *page)
1070 {
1071         struct zswap_tree *tree = zswap_trees[type];
1072         struct zswap_entry *entry;
1073         struct crypto_comp *tfm;
1074         u8 *src, *dst;
1075         unsigned int dlen;
1076         int ret;
1077
1078         /* find */
1079         spin_lock(&tree->lock);
1080         entry = zswap_entry_find_get(&tree->rbroot, offset);
1081         if (!entry) {
1082                 /* entry was written back */
1083                 spin_unlock(&tree->lock);
1084                 return -1;
1085         }
1086         spin_unlock(&tree->lock);
1087
1088         /* decompress */
1089         dlen = PAGE_SIZE;
1090         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1091                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1092         dst = kmap_atomic(page);
1093         tfm = *get_cpu_ptr(entry->pool->tfm);
1094         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1095         put_cpu_ptr(entry->pool->tfm);
1096         kunmap_atomic(dst);
1097         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1098         BUG_ON(ret);
1099
1100         spin_lock(&tree->lock);
1101         zswap_entry_put(tree, entry);
1102         spin_unlock(&tree->lock);
1103
1104         return 0;
1105 }
1106
1107 /* frees an entry in zswap */
1108 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1109 {
1110         struct zswap_tree *tree = zswap_trees[type];
1111         struct zswap_entry *entry;
1112
1113         /* find */
1114         spin_lock(&tree->lock);
1115         entry = zswap_rb_search(&tree->rbroot, offset);
1116         if (!entry) {
1117                 /* entry was written back */
1118                 spin_unlock(&tree->lock);
1119                 return;
1120         }
1121
1122         /* remove from rbtree */
1123         zswap_rb_erase(&tree->rbroot, entry);
1124
1125         /* drop the initial reference from entry creation */
1126         zswap_entry_put(tree, entry);
1127
1128         spin_unlock(&tree->lock);
1129 }
1130
1131 /* frees all zswap entries for the given swap type */
1132 static void zswap_frontswap_invalidate_area(unsigned type)
1133 {
1134         struct zswap_tree *tree = zswap_trees[type];
1135         struct zswap_entry *entry, *n;
1136
1137         if (!tree)
1138                 return;
1139
1140         /* walk the tree and free everything */
1141         spin_lock(&tree->lock);
1142         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1143                 zswap_free_entry(entry);
1144         tree->rbroot = RB_ROOT;
1145         spin_unlock(&tree->lock);
1146         kfree(tree);
1147         zswap_trees[type] = NULL;
1148 }
1149
1150 static void zswap_frontswap_init(unsigned type)
1151 {
1152         struct zswap_tree *tree;
1153
1154         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1155         if (!tree) {
1156                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1157                 return;
1158         }
1159
1160         tree->rbroot = RB_ROOT;
1161         spin_lock_init(&tree->lock);
1162         zswap_trees[type] = tree;
1163 }
1164
1165 static struct frontswap_ops zswap_frontswap_ops = {
1166         .store = zswap_frontswap_store,
1167         .load = zswap_frontswap_load,
1168         .invalidate_page = zswap_frontswap_invalidate_page,
1169         .invalidate_area = zswap_frontswap_invalidate_area,
1170         .init = zswap_frontswap_init
1171 };
1172
1173 /*********************************
1174 * debugfs functions
1175 **********************************/
1176 #ifdef CONFIG_DEBUG_FS
1177 #include <linux/debugfs.h>
1178
1179 static struct dentry *zswap_debugfs_root;
1180
1181 static int __init zswap_debugfs_init(void)
1182 {
1183         if (!debugfs_initialized())
1184                 return -ENODEV;
1185
1186         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1187         if (!zswap_debugfs_root)
1188                 return -ENOMEM;
1189
1190         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1191                         zswap_debugfs_root, &zswap_pool_limit_hit);
1192         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1193                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1194         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1195                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1196         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1197                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1198         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1199                         zswap_debugfs_root, &zswap_reject_compress_poor);
1200         debugfs_create_u64("written_back_pages", S_IRUGO,
1201                         zswap_debugfs_root, &zswap_written_back_pages);
1202         debugfs_create_u64("duplicate_entry", S_IRUGO,
1203                         zswap_debugfs_root, &zswap_duplicate_entry);
1204         debugfs_create_u64("pool_total_size", S_IRUGO,
1205                         zswap_debugfs_root, &zswap_pool_total_size);
1206         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1207                         zswap_debugfs_root, &zswap_stored_pages);
1208
1209         return 0;
1210 }
1211
1212 static void __exit zswap_debugfs_exit(void)
1213 {
1214         debugfs_remove_recursive(zswap_debugfs_root);
1215 }
1216 #else
1217 static int __init zswap_debugfs_init(void)
1218 {
1219         return 0;
1220 }
1221
1222 static void __exit zswap_debugfs_exit(void) { }
1223 #endif
1224
1225 /*********************************
1226 * module init and exit
1227 **********************************/
1228 static int __init init_zswap(void)
1229 {
1230         struct zswap_pool *pool;
1231
1232         zswap_init_started = true;
1233
1234         if (zswap_entry_cache_create()) {
1235                 pr_err("entry cache creation failed\n");
1236                 goto cache_fail;
1237         }
1238
1239         if (zswap_cpu_dstmem_init()) {
1240                 pr_err("dstmem alloc failed\n");
1241                 goto dstmem_fail;
1242         }
1243
1244         pool = __zswap_pool_create_fallback();
1245         if (!pool) {
1246                 pr_err("pool creation failed\n");
1247                 goto pool_fail;
1248         }
1249         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1250                 zpool_get_type(pool->zpool));
1251
1252         list_add(&pool->list, &zswap_pools);
1253
1254         frontswap_register_ops(&zswap_frontswap_ops);
1255         if (zswap_debugfs_init())
1256                 pr_warn("debugfs initialization failed\n");
1257         return 0;
1258
1259 pool_fail:
1260         zswap_cpu_dstmem_destroy();
1261 dstmem_fail:
1262         zswap_entry_cache_destroy();
1263 cache_fail:
1264         return -ENOMEM;
1265 }
1266 /* must be late so crypto has time to come up */
1267 late_initcall(init_zswap);
1268
1269 MODULE_LICENSE("GPL");
1270 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1271 MODULE_DESCRIPTION("Compressed cache for swap pages");