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 static int zswap_enabled_param_set(const char *,
  82                                    const struct kernel_param *);
  83 static struct kernel_param_ops zswap_enabled_param_ops = {
  84         .set =          zswap_enabled_param_set,
  85         .get =          param_get_bool,
  86 };
  87 module_param_cb(enabled, &zswap_enabled_param_ops, &zswap_enabled, 0644);
  88
  89 /* Crypto compressor to use */
  90 #define ZSWAP_COMPRESSOR_DEFAULT "lzo"
  91 static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
  92 static int zswap_compressor_param_set(const char *,
  93                                       const struct kernel_param *);
  94 static struct kernel_param_ops zswap_compressor_param_ops = {
  95         .set =          zswap_compressor_param_set,
  96         .get =          param_get_charp,
  97         .free =         param_free_charp,
  98 };
  99 module_param_cb(compressor, &zswap_compressor_param_ops,
 100                 &zswap_compressor, 0644);
 101
 102 /* Compressed storage zpool to use */
 103 #define ZSWAP_ZPOOL_DEFAULT "zbud"
 104 static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 105 static int zswap_zpool_param_set(const char *, const struct kernel_param *);
 106 static struct kernel_param_ops zswap_zpool_param_ops = {
 107         .set =          zswap_zpool_param_set,
 108         .get =          param_get_charp,
 109         .free =         param_free_charp,
 110 };
 111 module_param_cb(zpool, &zswap_zpool_param_ops, &zswap_zpool_type, 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 work_struct work;
 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 /* pool counter to provide unique names to zpool */
 180 static atomic_t zswap_pools_count = ATOMIC_INIT(0);
 181
 182 /* used by param callback function */
 183 static bool zswap_init_started;
 184
 185 /* fatal error during init */
 186 static bool zswap_init_failed;
 187
 188 /*********************************
 189 * helpers and fwd declarations
 190 **********************************/
 191
 192 #define zswap_pool_debug(msg, p)                                \
 193         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 194                  zpool_get_type((p)->zpool))
 195
 196 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 197 static int zswap_pool_get(struct zswap_pool *pool);
 198 static void zswap_pool_put(struct zswap_pool *pool);
 199
 200 static const struct zpool_ops zswap_zpool_ops = {
 201         .evict = zswap_writeback_entry
 202 };
 203
 204 static bool zswap_is_full(void)
 205 {
 206         return totalram_pages * zswap_max_pool_percent / 100 <
 207                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 208 }
 209
 210 static void zswap_update_total_size(void)
 211 {
 212         struct zswap_pool *pool;
 213         u64 total = 0;
 214
 215         rcu_read_lock();
 216
 217         list_for_each_entry_rcu(pool, &zswap_pools, list)
 218                 total += zpool_get_total_size(pool->zpool);
 219
 220         rcu_read_unlock();
 221
 222         zswap_pool_total_size = total;
 223 }
 224
 225 /*********************************
 226 * zswap entry functions
 227 **********************************/
 228 static struct kmem_cache *zswap_entry_cache;
 229
 230 static int __init zswap_entry_cache_create(void)
 231 {
 232         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 233         return zswap_entry_cache == NULL;
 234 }
 235
 236 static void __init zswap_entry_cache_destroy(void)
 237 {
 238         kmem_cache_destroy(zswap_entry_cache);
 239 }
 240
 241 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 242 {
 243         struct zswap_entry *entry;
 244         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 245         if (!entry)
 246                 return NULL;
 247         entry->refcount = 1;
 248         RB_CLEAR_NODE(&entry->rbnode);
 249         return entry;
 250 }
 251
 252 static void zswap_entry_cache_free(struct zswap_entry *entry)
 253 {
 254         kmem_cache_free(zswap_entry_cache, entry);
 255 }
 256
 257 /*********************************
 258 * rbtree functions
 259 **********************************/
 260 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 261 {
 262         struct rb_node *node = root->rb_node;
 263         struct zswap_entry *entry;
 264
 265         while (node) {
 266                 entry = rb_entry(node, struct zswap_entry, rbnode);
 267                 if (entry->offset > offset)
 268                         node = node->rb_left;
 269                 else if (entry->offset < offset)
 270                         node = node->rb_right;
 271                 else
 272                         return entry;
 273         }
 274         return NULL;
 275 }
 276
 277 /*
 278  * In the case that a entry with the same offset is found, a pointer to
 279  * the existing entry is stored in dupentry and the function returns -EEXIST
 280  */
 281 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 282                         struct zswap_entry **dupentry)
 283 {
 284         struct rb_node **link = &root->rb_node, *parent = NULL;
 285         struct zswap_entry *myentry;
 286
 287         while (*link) {
 288                 parent = *link;
 289                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 290                 if (myentry->offset > entry->offset)
 291                         link = &(*link)->rb_left;
 292                 else if (myentry->offset < entry->offset)
 293                         link = &(*link)->rb_right;
 294                 else {
 295                         *dupentry = myentry;
 296                         return -EEXIST;
 297                 }
 298         }
 299         rb_link_node(&entry->rbnode, parent, link);
 300         rb_insert_color(&entry->rbnode, root);
 301         return 0;
 302 }
 303
 304 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 305 {
 306         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 307                 rb_erase(&entry->rbnode, root);
 308                 RB_CLEAR_NODE(&entry->rbnode);
 309         }
 310 }
 311
 312 /*
 313  * Carries out the common pattern of freeing and entry's zpool allocation,
 314  * freeing the entry itself, and decrementing the number of stored pages.
 315  */
 316 static void zswap_free_entry(struct zswap_entry *entry)
 317 {
 318         zpool_free(entry->pool->zpool, entry->handle);
 319         zswap_pool_put(entry->pool);
 320         zswap_entry_cache_free(entry);
 321         atomic_dec(&zswap_stored_pages);
 322         zswap_update_total_size();
 323 }
 324
 325 /* caller must hold the tree lock */
 326 static void zswap_entry_get(struct zswap_entry *entry)
 327 {
 328         entry->refcount++;
 329 }
 330
 331 /* caller must hold the tree lock
 332 * remove from the tree and free it, if nobody reference the entry
 333 */
 334 static void zswap_entry_put(struct zswap_tree *tree,
 335                         struct zswap_entry *entry)
 336 {
 337         int refcount = --entry->refcount;
 338
 339         BUG_ON(refcount < 0);
 340         if (refcount == 0) {
 341                 zswap_rb_erase(&tree->rbroot, entry);
 342                 zswap_free_entry(entry);
 343         }
 344 }
 345
 346 /* caller must hold the tree lock */
 347 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 348                                 pgoff_t offset)
 349 {
 350         struct zswap_entry *entry;
 351
 352         entry = zswap_rb_search(root, offset);
 353         if (entry)
 354                 zswap_entry_get(entry);
 355
 356         return entry;
 357 }
 358
 359 /*********************************
 360 * per-cpu code
 361 **********************************/
 362 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 363
 364 static int __zswap_cpu_dstmem_notifier(unsigned long action, unsigned long cpu)
 365 {
 366         u8 *dst;
 367
 368         switch (action) {
 369         case CPU_UP_PREPARE:
 370                 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 371                 if (!dst) {
 372                         pr_err("can't allocate compressor buffer\n");
 373                         return NOTIFY_BAD;
 374                 }
 375                 per_cpu(zswap_dstmem, cpu) = dst;
 376                 break;
 377         case CPU_DEAD:
 378         case CPU_UP_CANCELED:
 379                 dst = per_cpu(zswap_dstmem, cpu);
 380                 kfree(dst);
 381                 per_cpu(zswap_dstmem, cpu) = NULL;
 382                 break;
 383         default:
 384                 break;
 385         }
 386         return NOTIFY_OK;
 387 }
 388
 389 static int zswap_cpu_dstmem_notifier(struct notifier_block *nb,
 390                                      unsigned long action, void *pcpu)
 391 {
 392         return __zswap_cpu_dstmem_notifier(action, (unsigned long)pcpu);
 393 }
 394
 395 static struct notifier_block zswap_dstmem_notifier = {
 396         .notifier_call =        zswap_cpu_dstmem_notifier,
 397 };
 398
 399 static int __init zswap_cpu_dstmem_init(void)
 400 {
 401         unsigned long cpu;
 402
 403         cpu_notifier_register_begin();
 404         for_each_online_cpu(cpu)
 405                 if (__zswap_cpu_dstmem_notifier(CPU_UP_PREPARE, cpu) ==
 406                     NOTIFY_BAD)
 407                         goto cleanup;
 408         __register_cpu_notifier(&zswap_dstmem_notifier);
 409         cpu_notifier_register_done();
 410         return 0;
 411
 412 cleanup:
 413         for_each_online_cpu(cpu)
 414                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 415         cpu_notifier_register_done();
 416         return -ENOMEM;
 417 }
 418
 419 static void zswap_cpu_dstmem_destroy(void)
 420 {
 421         unsigned long cpu;
 422
 423         cpu_notifier_register_begin();
 424         for_each_online_cpu(cpu)
 425                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 426         __unregister_cpu_notifier(&zswap_dstmem_notifier);
 427         cpu_notifier_register_done();
 428 }
 429
 430 static int __zswap_cpu_comp_notifier(struct zswap_pool *pool,
 431                                      unsigned long action, unsigned long cpu)
 432 {
 433         struct crypto_comp *tfm;
 434
 435         switch (action) {
 436         case CPU_UP_PREPARE:
 437                 if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 438                         break;
 439                 tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 440                 if (IS_ERR_OR_NULL(tfm)) {
 441                         pr_err("could not alloc crypto comp %s : %ld\n",
 442                                pool->tfm_name, PTR_ERR(tfm));
 443                         return NOTIFY_BAD;
 444                 }
 445                 *per_cpu_ptr(pool->tfm, cpu) = tfm;
 446                 break;
 447         case CPU_DEAD:
 448         case CPU_UP_CANCELED:
 449                 tfm = *per_cpu_ptr(pool->tfm, cpu);
 450                 if (!IS_ERR_OR_NULL(tfm))
 451                         crypto_free_comp(tfm);
 452                 *per_cpu_ptr(pool->tfm, cpu) = NULL;
 453                 break;
 454         default:
 455                 break;
 456         }
 457         return NOTIFY_OK;
 458 }
 459
 460 static int zswap_cpu_comp_notifier(struct notifier_block *nb,
 461                                    unsigned long action, void *pcpu)
 462 {
 463         unsigned long cpu = (unsigned long)pcpu;
 464         struct zswap_pool *pool = container_of(nb, typeof(*pool), notifier);
 465
 466         return __zswap_cpu_comp_notifier(pool, action, cpu);
 467 }
 468
 469 static int zswap_cpu_comp_init(struct zswap_pool *pool)
 470 {
 471         unsigned long cpu;
 472
 473         memset(&pool->notifier, 0, sizeof(pool->notifier));
 474         pool->notifier.notifier_call = zswap_cpu_comp_notifier;
 475
 476         cpu_notifier_register_begin();
 477         for_each_online_cpu(cpu)
 478                 if (__zswap_cpu_comp_notifier(pool, CPU_UP_PREPARE, cpu) ==
 479                     NOTIFY_BAD)
 480                         goto cleanup;
 481         __register_cpu_notifier(&pool->notifier);
 482         cpu_notifier_register_done();
 483         return 0;
 484
 485 cleanup:
 486         for_each_online_cpu(cpu)
 487                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 488         cpu_notifier_register_done();
 489         return -ENOMEM;
 490 }
 491
 492 static void zswap_cpu_comp_destroy(struct zswap_pool *pool)
 493 {
 494         unsigned long cpu;
 495
 496         cpu_notifier_register_begin();
 497         for_each_online_cpu(cpu)
 498                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 499         __unregister_cpu_notifier(&pool->notifier);
 500         cpu_notifier_register_done();
 501 }
 502
 503 /*********************************
 504 * pool functions
 505 **********************************/
 506
 507 static struct zswap_pool *__zswap_pool_current(void)
 508 {
 509         struct zswap_pool *pool;
 510
 511         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 512         WARN_ON(!pool);
 513
 514         return pool;
 515 }
 516
 517 static struct zswap_pool *zswap_pool_current(void)
 518 {
 519         assert_spin_locked(&zswap_pools_lock);
 520
 521         return __zswap_pool_current();
 522 }
 523
 524 static struct zswap_pool *zswap_pool_current_get(void)
 525 {
 526         struct zswap_pool *pool;
 527
 528         rcu_read_lock();
 529
 530         pool = __zswap_pool_current();
 531         if (!pool || !zswap_pool_get(pool))
 532                 pool = NULL;
 533
 534         rcu_read_unlock();
 535
 536         return pool;
 537 }
 538
 539 static struct zswap_pool *zswap_pool_last_get(void)
 540 {
 541         struct zswap_pool *pool, *last = NULL;
 542
 543         rcu_read_lock();
 544
 545         list_for_each_entry_rcu(pool, &zswap_pools, list)
 546                 last = pool;
 547         if (!WARN_ON(!last) && !zswap_pool_get(last))
 548                 last = NULL;
 549
 550         rcu_read_unlock();
 551
 552         return last;
 553 }
 554
 555 /* type and compressor must be null-terminated */
 556 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 557 {
 558         struct zswap_pool *pool;
 559
 560         assert_spin_locked(&zswap_pools_lock);
 561
 562         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 563                 if (strcmp(pool->tfm_name, compressor))
 564                         continue;
 565                 if (strcmp(zpool_get_type(pool->zpool), type))
 566                         continue;
 567                 /* if we can't get it, it's about to be destroyed */
 568                 if (!zswap_pool_get(pool))
 569                         continue;
 570                 return pool;
 571         }
 572
 573         return NULL;
 574 }
 575
 576 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 577 {
 578         struct zswap_pool *pool;
 579         char name[38]; /* 'zswap' + 32 char (max) num + \0 */
 580         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM;
 581
 582         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 583         if (!pool) {
 584                 pr_err("pool alloc failed\n");
 585                 return NULL;
 586         }
 587
 588         /* unique name for each pool specifically required by zsmalloc */
 589         snprintf(name, 38, "zswap%x", atomic_inc_return(&zswap_pools_count));
 590
 591         pool->zpool = zpool_create_pool(type, name, gfp, &zswap_zpool_ops);
 592         if (!pool->zpool) {
 593                 pr_err("%s zpool not available\n", type);
 594                 goto error;
 595         }
 596         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 597
 598         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 599         pool->tfm = alloc_percpu(struct crypto_comp *);
 600         if (!pool->tfm) {
 601                 pr_err("percpu alloc failed\n");
 602                 goto error;
 603         }
 604
 605         if (zswap_cpu_comp_init(pool))
 606                 goto error;
 607         pr_debug("using %s compressor\n", pool->tfm_name);
 608
 609         /* being the current pool takes 1 ref; this func expects the
 610          * caller to always add the new pool as the current pool
 611          */
 612         kref_init(&pool->kref);
 613         INIT_LIST_HEAD(&pool->list);
 614
 615         zswap_pool_debug("created", pool);
 616
 617         return pool;
 618
 619 error:
 620         free_percpu(pool->tfm);
 621         if (pool->zpool)
 622                 zpool_destroy_pool(pool->zpool);
 623         kfree(pool);
 624         return NULL;
 625 }
 626
 627 static __init struct zswap_pool *__zswap_pool_create_fallback(void)
 628 {
 629         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 630                 if (!strcmp(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT)) {
 631                         pr_err("default compressor %s not available\n",
 632                                zswap_compressor);
 633                         return NULL;
 634                 }
 635                 pr_err("compressor %s not available, using default %s\n",
 636                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 637                 param_free_charp(&zswap_compressor);
 638                 zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
 639         }
 640         if (!zpool_has_pool(zswap_zpool_type)) {
 641                 if (!strcmp(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT)) {
 642                         pr_err("default zpool %s not available\n",
 643                                zswap_zpool_type);
 644                         return NULL;
 645                 }
 646                 pr_err("zpool %s not available, using default %s\n",
 647                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 648                 param_free_charp(&zswap_zpool_type);
 649                 zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
 650         }
 651
 652         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 653 }
 654
 655 static void zswap_pool_destroy(struct zswap_pool *pool)
 656 {
 657         zswap_pool_debug("destroying", pool);
 658
 659         zswap_cpu_comp_destroy(pool);
 660         free_percpu(pool->tfm);
 661         zpool_destroy_pool(pool->zpool);
 662         kfree(pool);
 663 }
 664
 665 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 666 {
 667         return kref_get_unless_zero(&pool->kref);
 668 }
 669
 670 static void __zswap_pool_release(struct work_struct *work)
 671 {
 672         struct zswap_pool *pool = container_of(work, typeof(*pool), work);
 673
 674         synchronize_rcu();
 675
 676         /* nobody should have been able to get a kref... */
 677         WARN_ON(kref_get_unless_zero(&pool->kref));
 678
 679         /* pool is now off zswap_pools list and has no references. */
 680         zswap_pool_destroy(pool);
 681 }
 682
 683 static void __zswap_pool_empty(struct kref *kref)
 684 {
 685         struct zswap_pool *pool;
 686
 687         pool = container_of(kref, typeof(*pool), kref);
 688
 689         spin_lock(&zswap_pools_lock);
 690
 691         WARN_ON(pool == zswap_pool_current());
 692
 693         list_del_rcu(&pool->list);
 694
 695         INIT_WORK(&pool->work, __zswap_pool_release);
 696         schedule_work(&pool->work);
 697
 698         spin_unlock(&zswap_pools_lock);
 699 }
 700
 701 static void zswap_pool_put(struct zswap_pool *pool)
 702 {
 703         kref_put(&pool->kref, __zswap_pool_empty);
 704 }
 705
 706 /*********************************
 707 * param callbacks
 708 **********************************/
 709
 710 /* val must be a null-terminated string */
 711 static int __zswap_param_set(const char *val, const struct kernel_param *kp,
 712                              char *type, char *compressor)
 713 {
 714         struct zswap_pool *pool, *put_pool = NULL;
 715         char *s = strstrip((char *)val);
 716         int ret;
 717
 718         if (zswap_init_failed) {
 719                 pr_err("can't set param, initialization failed\n");
 720                 return -ENODEV;
 721         }
 722
 723         /* no change required */
 724         if (!strcmp(s, *(char **)kp->arg))
 725                 return 0;
 726
 727         /* if this is load-time (pre-init) param setting,
 728          * don't create a pool; that's done during init.
 729          */
 730         if (!zswap_init_started)
 731                 return param_set_charp(s, kp);
 732
 733         if (!type) {
 734                 if (!zpool_has_pool(s)) {
 735                         pr_err("zpool %s not available\n", s);
 736                         return -ENOENT;
 737                 }
 738                 type = s;
 739         } else if (!compressor) {
 740                 if (!crypto_has_comp(s, 0, 0)) {
 741                         pr_err("compressor %s not available\n", s);
 742                         return -ENOENT;
 743                 }
 744                 compressor = s;
 745         } else {
 746                 WARN_ON(1);
 747                 return -EINVAL;
 748         }
 749
 750         spin_lock(&zswap_pools_lock);
 751
 752         pool = zswap_pool_find_get(type, compressor);
 753         if (pool) {
 754                 zswap_pool_debug("using existing", pool);
 755                 list_del_rcu(&pool->list);
 756         } else {
 757                 spin_unlock(&zswap_pools_lock);
 758                 pool = zswap_pool_create(type, compressor);
 759                 spin_lock(&zswap_pools_lock);
 760         }
 761
 762         if (pool)
 763                 ret = param_set_charp(s, kp);
 764         else
 765                 ret = -EINVAL;
 766
 767         if (!ret) {
 768                 put_pool = zswap_pool_current();
 769                 list_add_rcu(&pool->list, &zswap_pools);
 770         } else if (pool) {
 771                 /* add the possibly pre-existing pool to the end of the pools
 772                  * list; if it's new (and empty) then it'll be removed and
 773                  * destroyed by the put after we drop the lock
 774                  */
 775                 list_add_tail_rcu(&pool->list, &zswap_pools);
 776                 put_pool = pool;
 777         }
 778
 779         spin_unlock(&zswap_pools_lock);
 780
 781         /* drop the ref from either the old current pool,
 782          * or the new pool we failed to add
 783          */
 784         if (put_pool)
 785                 zswap_pool_put(put_pool);
 786
 787         return ret;
 788 }
 789
 790 static int zswap_compressor_param_set(const char *val,
 791                                       const struct kernel_param *kp)
 792 {
 793         return __zswap_param_set(val, kp, zswap_zpool_type, NULL);
 794 }
 795
 796 static int zswap_zpool_param_set(const char *val,
 797                                  const struct kernel_param *kp)
 798 {
 799         return __zswap_param_set(val, kp, NULL, zswap_compressor);
 800 }
 801
 802 static int zswap_enabled_param_set(const char *val,
 803                                    const struct kernel_param *kp)
 804 {
 805         if (zswap_init_failed) {
 806                 pr_err("can't enable, initialization failed\n");
 807                 return -ENODEV;
 808         }
 809
 810         return param_set_bool(val, kp);
 811 }
 812
 813 /*********************************
 814 * writeback code
 815 **********************************/
 816 /* return enum for zswap_get_swap_cache_page */
 817 enum zswap_get_swap_ret {
 818         ZSWAP_SWAPCACHE_NEW,
 819         ZSWAP_SWAPCACHE_EXIST,
 820         ZSWAP_SWAPCACHE_FAIL,
 821 };
 822
 823 /*
 824  * zswap_get_swap_cache_page
 825  *
 826  * This is an adaption of read_swap_cache_async()
 827  *
 828  * This function tries to find a page with the given swap entry
 829  * in the swapper_space address space (the swap cache).  If the page
 830  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 831  * added to the swap cache, and returned in retpage.
 832  *
 833  * If success, the swap cache page is returned in retpage
 834  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 835  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 836  *     the new page is added to swapcache and locked
 837  * Returns ZSWAP_SWAPCACHE_FAIL on error
 838  */
 839 static int zswap_get_swap_cache_page(swp_entry_t entry,
 840                                 struct page **retpage)
 841 {
 842         bool page_was_allocated;
 843
 844         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 845                         NULL, 0, &page_was_allocated);
 846         if (page_was_allocated)
 847                 return ZSWAP_SWAPCACHE_NEW;
 848         if (!*retpage)
 849                 return ZSWAP_SWAPCACHE_FAIL;
 850         return ZSWAP_SWAPCACHE_EXIST;
 851 }
 852
 853 /*
 854  * Attempts to free an entry by adding a page to the swap cache,
 855  * decompressing the entry data into the page, and issuing a
 856  * bio write to write the page back to the swap device.
 857  *
 858  * This can be thought of as a "resumed writeback" of the page
 859  * to the swap device.  We are basically resuming the same swap
 860  * writeback path that was intercepted with the frontswap_store()
 861  * in the first place.  After the page has been decompressed into
 862  * the swap cache, the compressed version stored by zswap can be
 863  * freed.
 864  */
 865 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 866 {
 867         struct zswap_header *zhdr;
 868         swp_entry_t swpentry;
 869         struct zswap_tree *tree;
 870         pgoff_t offset;
 871         struct zswap_entry *entry;
 872         struct page *page;
 873         struct crypto_comp *tfm;
 874         u8 *src, *dst;
 875         unsigned int dlen;
 876         int ret;
 877         struct writeback_control wbc = {
 878                 .sync_mode = WB_SYNC_NONE,
 879         };
 880
 881         /* extract swpentry from data */
 882         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 883         swpentry = zhdr->swpentry; /* here */
 884         zpool_unmap_handle(pool, handle);
 885         tree = zswap_trees[swp_type(swpentry)];
 886         offset = swp_offset(swpentry);
 887
 888         /* find and ref zswap entry */
 889         spin_lock(&tree->lock);
 890         entry = zswap_entry_find_get(&tree->rbroot, offset);
 891         if (!entry) {
 892                 /* entry was invalidated */
 893                 spin_unlock(&tree->lock);
 894                 return 0;
 895         }
 896         spin_unlock(&tree->lock);
 897         BUG_ON(offset != entry->offset);
 898
 899         /* try to allocate swap cache page */
 900         switch (zswap_get_swap_cache_page(swpentry, &page)) {
 901         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
 902                 ret = -ENOMEM;
 903                 goto fail;
 904
 905         case ZSWAP_SWAPCACHE_EXIST:
 906                 /* page is already in the swap cache, ignore for now */
 907                 put_page(page);
 908                 ret = -EEXIST;
 909                 goto fail;
 910
 911         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
 912                 /* decompress */
 913                 dlen = PAGE_SIZE;
 914                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 915                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
 916                 dst = kmap_atomic(page);
 917                 tfm = *get_cpu_ptr(entry->pool->tfm);
 918                 ret = crypto_comp_decompress(tfm, src, entry->length,
 919                                              dst, &dlen);
 920                 put_cpu_ptr(entry->pool->tfm);
 921                 kunmap_atomic(dst);
 922                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
 923                 BUG_ON(ret);
 924                 BUG_ON(dlen != PAGE_SIZE);
 925
 926                 /* page is up to date */
 927                 SetPageUptodate(page);
 928         }
 929
 930         /* move it to the tail of the inactive list after end_writeback */
 931         SetPageReclaim(page);
 932
 933         /* start writeback */
 934         __swap_writepage(page, &wbc, end_swap_bio_write);
 935         put_page(page);
 936         zswap_written_back_pages++;
 937
 938         spin_lock(&tree->lock);
 939         /* drop local reference */
 940         zswap_entry_put(tree, entry);
 941
 942         /*
 943         * There are two possible situations for entry here:
 944         * (1) refcount is 1(normal case),  entry is valid and on the tree
 945         * (2) refcount is 0, entry is freed and not on the tree
 946         *     because invalidate happened during writeback
 947         *  search the tree and free the entry if find entry
 948         */
 949         if (entry == zswap_rb_search(&tree->rbroot, offset))
 950                 zswap_entry_put(tree, entry);
 951         spin_unlock(&tree->lock);
 952
 953         goto end;
 954
 955         /*
 956         * if we get here due to ZSWAP_SWAPCACHE_EXIST
 957         * a load may happening concurrently
 958         * it is safe and okay to not free the entry
 959         * if we free the entry in the following put
 960         * it it either okay to return !0
 961         */
 962 fail:
 963         spin_lock(&tree->lock);
 964         zswap_entry_put(tree, entry);
 965         spin_unlock(&tree->lock);
 966
 967 end:
 968         return ret;
 969 }
 970
 971 static int zswap_shrink(void)
 972 {
 973         struct zswap_pool *pool;
 974         int ret;
 975
 976         pool = zswap_pool_last_get();
 977         if (!pool)
 978                 return -ENOENT;
 979
 980         ret = zpool_shrink(pool->zpool, 1, NULL);
 981
 982         zswap_pool_put(pool);
 983
 984         return ret;
 985 }
 986
 987 /*********************************
 988 * frontswap hooks
 989 **********************************/
 990 /* attempts to compress and store an single page */
 991 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 992                                 struct page *page)
 993 {
 994         struct zswap_tree *tree = zswap_trees[type];
 995         struct zswap_entry *entry, *dupentry;
 996         struct crypto_comp *tfm;
 997         int ret;
 998         unsigned int dlen = PAGE_SIZE, len;
 999         unsigned long handle;
1000         char *buf;
1001         u8 *src, *dst;
1002         struct zswap_header *zhdr;
1003
1004         if (!zswap_enabled || !tree) {
1005                 ret = -ENODEV;
1006                 goto reject;
1007         }
1008
1009         /* reclaim space if needed */
1010         if (zswap_is_full()) {
1011                 zswap_pool_limit_hit++;
1012                 if (zswap_shrink()) {
1013                         zswap_reject_reclaim_fail++;
1014                         ret = -ENOMEM;
1015                         goto reject;
1016                 }
1017         }
1018
1019         /* allocate entry */
1020         entry = zswap_entry_cache_alloc(GFP_KERNEL);
1021         if (!entry) {
1022                 zswap_reject_kmemcache_fail++;
1023                 ret = -ENOMEM;
1024                 goto reject;
1025         }
1026
1027         /* if entry is successfully added, it keeps the reference */
1028         entry->pool = zswap_pool_current_get();
1029         if (!entry->pool) {
1030                 ret = -EINVAL;
1031                 goto freepage;
1032         }
1033
1034         /* compress */
1035         dst = get_cpu_var(zswap_dstmem);
1036         tfm = *get_cpu_ptr(entry->pool->tfm);
1037         src = kmap_atomic(page);
1038         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
1039         kunmap_atomic(src);
1040         put_cpu_ptr(entry->pool->tfm);
1041         if (ret) {
1042                 ret = -EINVAL;
1043                 goto put_dstmem;
1044         }
1045
1046         /* store */
1047         len = dlen + sizeof(struct zswap_header);
1048         ret = zpool_malloc(entry->pool->zpool, len,
1049                            __GFP_NORETRY | __GFP_NOWARN | __GFP_KSWAPD_RECLAIM,
1050                            &handle);
1051         if (ret == -ENOSPC) {
1052                 zswap_reject_compress_poor++;
1053                 goto put_dstmem;
1054         }
1055         if (ret) {
1056                 zswap_reject_alloc_fail++;
1057                 goto put_dstmem;
1058         }
1059         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
1060         zhdr->swpentry = swp_entry(type, offset);
1061         buf = (u8 *)(zhdr + 1);
1062         memcpy(buf, dst, dlen);
1063         zpool_unmap_handle(entry->pool->zpool, handle);
1064         put_cpu_var(zswap_dstmem);
1065
1066         /* populate entry */
1067         entry->offset = offset;
1068         entry->handle = handle;
1069         entry->length = dlen;
1070
1071         /* map */
1072         spin_lock(&tree->lock);
1073         do {
1074                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
1075                 if (ret == -EEXIST) {
1076                         zswap_duplicate_entry++;
1077                         /* remove from rbtree */
1078                         zswap_rb_erase(&tree->rbroot, dupentry);
1079                         zswap_entry_put(tree, dupentry);
1080                 }
1081         } while (ret == -EEXIST);
1082         spin_unlock(&tree->lock);
1083
1084         /* update stats */
1085         atomic_inc(&zswap_stored_pages);
1086         zswap_update_total_size();
1087
1088         return 0;
1089
1090 put_dstmem:
1091         put_cpu_var(zswap_dstmem);
1092         zswap_pool_put(entry->pool);
1093 freepage:
1094         zswap_entry_cache_free(entry);
1095 reject:
1096         return ret;
1097 }
1098
1099 /*
1100  * returns 0 if the page was successfully decompressed
1101  * return -1 on entry not found or error
1102 */
1103 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
1104                                 struct page *page)
1105 {
1106         struct zswap_tree *tree = zswap_trees[type];
1107         struct zswap_entry *entry;
1108         struct crypto_comp *tfm;
1109         u8 *src, *dst;
1110         unsigned int dlen;
1111         int ret;
1112
1113         /* find */
1114         spin_lock(&tree->lock);
1115         entry = zswap_entry_find_get(&tree->rbroot, offset);
1116         if (!entry) {
1117                 /* entry was written back */
1118                 spin_unlock(&tree->lock);
1119                 return -1;
1120         }
1121         spin_unlock(&tree->lock);
1122
1123         /* decompress */
1124         dlen = PAGE_SIZE;
1125         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
1126                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
1127         dst = kmap_atomic(page);
1128         tfm = *get_cpu_ptr(entry->pool->tfm);
1129         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
1130         put_cpu_ptr(entry->pool->tfm);
1131         kunmap_atomic(dst);
1132         zpool_unmap_handle(entry->pool->zpool, entry->handle);
1133         BUG_ON(ret);
1134
1135         spin_lock(&tree->lock);
1136         zswap_entry_put(tree, entry);
1137         spin_unlock(&tree->lock);
1138
1139         return 0;
1140 }
1141
1142 /* frees an entry in zswap */
1143 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
1144 {
1145         struct zswap_tree *tree = zswap_trees[type];
1146         struct zswap_entry *entry;
1147
1148         /* find */
1149         spin_lock(&tree->lock);
1150         entry = zswap_rb_search(&tree->rbroot, offset);
1151         if (!entry) {
1152                 /* entry was written back */
1153                 spin_unlock(&tree->lock);
1154                 return;
1155         }
1156
1157         /* remove from rbtree */
1158         zswap_rb_erase(&tree->rbroot, entry);
1159
1160         /* drop the initial reference from entry creation */
1161         zswap_entry_put(tree, entry);
1162
1163         spin_unlock(&tree->lock);
1164 }
1165
1166 /* frees all zswap entries for the given swap type */
1167 static void zswap_frontswap_invalidate_area(unsigned type)
1168 {
1169         struct zswap_tree *tree = zswap_trees[type];
1170         struct zswap_entry *entry, *n;
1171
1172         if (!tree)
1173                 return;
1174
1175         /* walk the tree and free everything */
1176         spin_lock(&tree->lock);
1177         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1178                 zswap_free_entry(entry);
1179         tree->rbroot = RB_ROOT;
1180         spin_unlock(&tree->lock);
1181         kfree(tree);
1182         zswap_trees[type] = NULL;
1183 }
1184
1185 static void zswap_frontswap_init(unsigned type)
1186 {
1187         struct zswap_tree *tree;
1188
1189         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1190         if (!tree) {
1191                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1192                 return;
1193         }
1194
1195         tree->rbroot = RB_ROOT;
1196         spin_lock_init(&tree->lock);
1197         zswap_trees[type] = tree;
1198 }
1199
1200 static struct frontswap_ops zswap_frontswap_ops = {
1201         .store = zswap_frontswap_store,
1202         .load = zswap_frontswap_load,
1203         .invalidate_page = zswap_frontswap_invalidate_page,
1204         .invalidate_area = zswap_frontswap_invalidate_area,
1205         .init = zswap_frontswap_init
1206 };
1207
1208 /*********************************
1209 * debugfs functions
1210 **********************************/
1211 #ifdef CONFIG_DEBUG_FS
1212 #include <linux/debugfs.h>
1213
1214 static struct dentry *zswap_debugfs_root;
1215
1216 static int __init zswap_debugfs_init(void)
1217 {
1218         if (!debugfs_initialized())
1219                 return -ENODEV;
1220
1221         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1222         if (!zswap_debugfs_root)
1223                 return -ENOMEM;
1224
1225         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1226                         zswap_debugfs_root, &zswap_pool_limit_hit);
1227         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1228                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1229         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1230                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1231         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1232                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1233         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1234                         zswap_debugfs_root, &zswap_reject_compress_poor);
1235         debugfs_create_u64("written_back_pages", S_IRUGO,
1236                         zswap_debugfs_root, &zswap_written_back_pages);
1237         debugfs_create_u64("duplicate_entry", S_IRUGO,
1238                         zswap_debugfs_root, &zswap_duplicate_entry);
1239         debugfs_create_u64("pool_total_size", S_IRUGO,
1240                         zswap_debugfs_root, &zswap_pool_total_size);
1241         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1242                         zswap_debugfs_root, &zswap_stored_pages);
1243
1244         return 0;
1245 }
1246
1247 static void __exit zswap_debugfs_exit(void)
1248 {
1249         debugfs_remove_recursive(zswap_debugfs_root);
1250 }
1251 #else
1252 static int __init zswap_debugfs_init(void)
1253 {
1254         return 0;
1255 }
1256
1257 static void __exit zswap_debugfs_exit(void) { }
1258 #endif
1259
1260 /*********************************
1261 * module init and exit
1262 **********************************/
1263 static int __init init_zswap(void)
1264 {
1265         struct zswap_pool *pool;
1266
1267         zswap_init_started = true;
1268
1269         if (zswap_entry_cache_create()) {
1270                 pr_err("entry cache creation failed\n");
1271                 goto cache_fail;
1272         }
1273
1274         if (zswap_cpu_dstmem_init()) {
1275                 pr_err("dstmem alloc failed\n");
1276                 goto dstmem_fail;
1277         }
1278
1279         pool = __zswap_pool_create_fallback();
1280         if (!pool) {
1281                 pr_err("pool creation failed\n");
1282                 goto pool_fail;
1283         }
1284         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1285                 zpool_get_type(pool->zpool));
1286
1287         list_add(&pool->list, &zswap_pools);
1288
1289         frontswap_register_ops(&zswap_frontswap_ops);
1290         if (zswap_debugfs_init())
1291                 pr_warn("debugfs initialization failed\n");
1292         return 0;
1293
1294 pool_fail:
1295         zswap_cpu_dstmem_destroy();
1296 dstmem_fail:
1297         zswap_entry_cache_destroy();
1298 cache_fail:
1299         /* if built-in, we aren't unloaded on failure; don't allow use */
1300         zswap_init_failed = true;
1301         zswap_enabled = false;
1302         return -ENOMEM;
1303 }
1304 /* must be late so crypto has time to come up */
1305 late_initcall(init_zswap);
1306
1307 MODULE_LICENSE("GPL");
1308 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1309 MODULE_DESCRIPTION("Compressed cache for swap pages");