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 /* Compressor to be used by zswap (fixed at boot for now) */
  84 #define ZSWAP_COMPRESSOR_DEFAULT "lzo"
  85 static char *zswap_compressor = ZSWAP_COMPRESSOR_DEFAULT;
  86 module_param_named(compressor, zswap_compressor, charp, 0444);
  87
  88 /* The maximum percentage of memory that the compressed pool can occupy */
  89 static unsigned int zswap_max_pool_percent = 20;
  90 module_param_named(max_pool_percent,
  91                         zswap_max_pool_percent, uint, 0644);
  92
  93 /* Compressed storage to use */
  94 #define ZSWAP_ZPOOL_DEFAULT "zbud"
  95 static char *zswap_zpool_type = ZSWAP_ZPOOL_DEFAULT;
  96 module_param_named(zpool, zswap_zpool_type, charp, 0444);
  97
  98 /* zpool is shared by all of zswap backend  */
  99 static struct zpool *zswap_pool;
 100
 101 /*********************************
 102 * data structures
 103 **********************************/
 104
 105 struct zswap_pool {
 106         struct zpool *zpool;
 107         struct crypto_comp * __percpu *tfm;
 108         struct kref kref;
 109         struct list_head list;
 110         struct rcu_head rcu_head;
 111         struct notifier_block notifier;
 112         char tfm_name[CRYPTO_MAX_ALG_NAME];
 113 };
 114
 115 /*
 116  * struct zswap_entry
 117  *
 118  * This structure contains the metadata for tracking a single compressed
 119  * page within zswap.
 120  *
 121  * rbnode - links the entry into red-black tree for the appropriate swap type
 122  * offset - the swap offset for the entry.  Index into the red-black tree.
 123  * refcount - the number of outstanding reference to the entry. This is needed
 124  *            to protect against premature freeing of the entry by code
 125  *            concurrent calls to load, invalidate, and writeback.  The lock
 126  *            for the zswap_tree structure that contains the entry must
 127  *            be held while changing the refcount.  Since the lock must
 128  *            be held, there is no reason to also make refcount atomic.
 129  * length - the length in bytes of the compressed page data.  Needed during
 130  *          decompression
 131  * pool - the zswap_pool the entry's data is in
 132  * handle - zpool allocation handle that stores the compressed page data
 133  */
 134 struct zswap_entry {
 135         struct rb_node rbnode;
 136         pgoff_t offset;
 137         int refcount;
 138         unsigned int length;
 139         struct zswap_pool *pool;
 140         unsigned long handle;
 141 };
 142
 143 struct zswap_header {
 144         swp_entry_t swpentry;
 145 };
 146
 147 /*
 148  * The tree lock in the zswap_tree struct protects a few things:
 149  * - the rbtree
 150  * - the refcount field of each entry in the tree
 151  */
 152 struct zswap_tree {
 153         struct rb_root rbroot;
 154         spinlock_t lock;
 155 };
 156
 157 static struct zswap_tree *zswap_trees[MAX_SWAPFILES];
 158
 159 /* RCU-protected iteration */
 160 static LIST_HEAD(zswap_pools);
 161 /* protects zswap_pools list modification */
 162 static DEFINE_SPINLOCK(zswap_pools_lock);
 163
 164 /*********************************
 165 * helpers and fwd declarations
 166 **********************************/
 167
 168 #define zswap_pool_debug(msg, p)                                \
 169         pr_debug("%s pool %s/%s\n", msg, (p)->tfm_name,         \
 170                  zpool_get_type((p)->zpool))
 171
 172 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle);
 173 static int zswap_pool_get(struct zswap_pool *pool);
 174 static void zswap_pool_put(struct zswap_pool *pool);
 175
 176 static const struct zpool_ops zswap_zpool_ops = {
 177         .evict = zswap_writeback_entry
 178 };
 179
 180 static bool zswap_is_full(void)
 181 {
 182         return totalram_pages * zswap_max_pool_percent / 100 <
 183                 DIV_ROUND_UP(zswap_pool_total_size, PAGE_SIZE);
 184 }
 185
 186 static void zswap_update_total_size(void)
 187 {
 188         struct zswap_pool *pool;
 189         u64 total = 0;
 190
 191         rcu_read_lock();
 192
 193         list_for_each_entry_rcu(pool, &zswap_pools, list)
 194                 total += zpool_get_total_size(pool->zpool);
 195
 196         rcu_read_unlock();
 197
 198         zswap_pool_total_size = total;
 199 }
 200
 201 /*********************************
 202 * zswap entry functions
 203 **********************************/
 204 static struct kmem_cache *zswap_entry_cache;
 205
 206 static int __init zswap_entry_cache_create(void)
 207 {
 208         zswap_entry_cache = KMEM_CACHE(zswap_entry, 0);
 209         return zswap_entry_cache == NULL;
 210 }
 211
 212 static void __init zswap_entry_cache_destroy(void)
 213 {
 214         kmem_cache_destroy(zswap_entry_cache);
 215 }
 216
 217 static struct zswap_entry *zswap_entry_cache_alloc(gfp_t gfp)
 218 {
 219         struct zswap_entry *entry;
 220         entry = kmem_cache_alloc(zswap_entry_cache, gfp);
 221         if (!entry)
 222                 return NULL;
 223         entry->refcount = 1;
 224         RB_CLEAR_NODE(&entry->rbnode);
 225         return entry;
 226 }
 227
 228 static void zswap_entry_cache_free(struct zswap_entry *entry)
 229 {
 230         kmem_cache_free(zswap_entry_cache, entry);
 231 }
 232
 233 /*********************************
 234 * rbtree functions
 235 **********************************/
 236 static struct zswap_entry *zswap_rb_search(struct rb_root *root, pgoff_t offset)
 237 {
 238         struct rb_node *node = root->rb_node;
 239         struct zswap_entry *entry;
 240
 241         while (node) {
 242                 entry = rb_entry(node, struct zswap_entry, rbnode);
 243                 if (entry->offset > offset)
 244                         node = node->rb_left;
 245                 else if (entry->offset < offset)
 246                         node = node->rb_right;
 247                 else
 248                         return entry;
 249         }
 250         return NULL;
 251 }
 252
 253 /*
 254  * In the case that a entry with the same offset is found, a pointer to
 255  * the existing entry is stored in dupentry and the function returns -EEXIST
 256  */
 257 static int zswap_rb_insert(struct rb_root *root, struct zswap_entry *entry,
 258                         struct zswap_entry **dupentry)
 259 {
 260         struct rb_node **link = &root->rb_node, *parent = NULL;
 261         struct zswap_entry *myentry;
 262
 263         while (*link) {
 264                 parent = *link;
 265                 myentry = rb_entry(parent, struct zswap_entry, rbnode);
 266                 if (myentry->offset > entry->offset)
 267                         link = &(*link)->rb_left;
 268                 else if (myentry->offset < entry->offset)
 269                         link = &(*link)->rb_right;
 270                 else {
 271                         *dupentry = myentry;
 272                         return -EEXIST;
 273                 }
 274         }
 275         rb_link_node(&entry->rbnode, parent, link);
 276         rb_insert_color(&entry->rbnode, root);
 277         return 0;
 278 }
 279
 280 static void zswap_rb_erase(struct rb_root *root, struct zswap_entry *entry)
 281 {
 282         if (!RB_EMPTY_NODE(&entry->rbnode)) {
 283                 rb_erase(&entry->rbnode, root);
 284                 RB_CLEAR_NODE(&entry->rbnode);
 285         }
 286 }
 287
 288 /*
 289  * Carries out the common pattern of freeing and entry's zpool allocation,
 290  * freeing the entry itself, and decrementing the number of stored pages.
 291  */
 292 static void zswap_free_entry(struct zswap_entry *entry)
 293 {
 294         zpool_free(entry->pool->zpool, entry->handle);
 295         zswap_pool_put(entry->pool);
 296         zswap_entry_cache_free(entry);
 297         atomic_dec(&zswap_stored_pages);
 298         zswap_update_total_size();
 299 }
 300
 301 /* caller must hold the tree lock */
 302 static void zswap_entry_get(struct zswap_entry *entry)
 303 {
 304         entry->refcount++;
 305 }
 306
 307 /* caller must hold the tree lock
 308 * remove from the tree and free it, if nobody reference the entry
 309 */
 310 static void zswap_entry_put(struct zswap_tree *tree,
 311                         struct zswap_entry *entry)
 312 {
 313         int refcount = --entry->refcount;
 314
 315         BUG_ON(refcount < 0);
 316         if (refcount == 0) {
 317                 zswap_rb_erase(&tree->rbroot, entry);
 318                 zswap_free_entry(entry);
 319         }
 320 }
 321
 322 /* caller must hold the tree lock */
 323 static struct zswap_entry *zswap_entry_find_get(struct rb_root *root,
 324                                 pgoff_t offset)
 325 {
 326         struct zswap_entry *entry = NULL;
 327
 328         entry = zswap_rb_search(root, offset);
 329         if (entry)
 330                 zswap_entry_get(entry);
 331
 332         return entry;
 333 }
 334
 335 /*********************************
 336 * per-cpu code
 337 **********************************/
 338 static DEFINE_PER_CPU(u8 *, zswap_dstmem);
 339
 340 static int __zswap_cpu_dstmem_notifier(unsigned long action, unsigned long cpu)
 341 {
 342         u8 *dst;
 343
 344         switch (action) {
 345         case CPU_UP_PREPARE:
 346                 dst = kmalloc_node(PAGE_SIZE * 2, GFP_KERNEL, cpu_to_node(cpu));
 347                 if (!dst) {
 348                         pr_err("can't allocate compressor buffer\n");
 349                         return NOTIFY_BAD;
 350                 }
 351                 per_cpu(zswap_dstmem, cpu) = dst;
 352                 break;
 353         case CPU_DEAD:
 354         case CPU_UP_CANCELED:
 355                 dst = per_cpu(zswap_dstmem, cpu);
 356                 kfree(dst);
 357                 per_cpu(zswap_dstmem, cpu) = NULL;
 358                 break;
 359         default:
 360                 break;
 361         }
 362         return NOTIFY_OK;
 363 }
 364
 365 static int zswap_cpu_dstmem_notifier(struct notifier_block *nb,
 366                                      unsigned long action, void *pcpu)
 367 {
 368         return __zswap_cpu_dstmem_notifier(action, (unsigned long)pcpu);
 369 }
 370
 371 static struct notifier_block zswap_dstmem_notifier = {
 372         .notifier_call =        zswap_cpu_dstmem_notifier,
 373 };
 374
 375 static int __init zswap_cpu_dstmem_init(void)
 376 {
 377         unsigned long cpu;
 378
 379         cpu_notifier_register_begin();
 380         for_each_online_cpu(cpu)
 381                 if (__zswap_cpu_dstmem_notifier(CPU_UP_PREPARE, cpu) ==
 382                     NOTIFY_BAD)
 383                         goto cleanup;
 384         __register_cpu_notifier(&zswap_dstmem_notifier);
 385         cpu_notifier_register_done();
 386         return 0;
 387
 388 cleanup:
 389         for_each_online_cpu(cpu)
 390                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 391         cpu_notifier_register_done();
 392         return -ENOMEM;
 393 }
 394
 395 static void zswap_cpu_dstmem_destroy(void)
 396 {
 397         unsigned long cpu;
 398
 399         cpu_notifier_register_begin();
 400         for_each_online_cpu(cpu)
 401                 __zswap_cpu_dstmem_notifier(CPU_UP_CANCELED, cpu);
 402         __unregister_cpu_notifier(&zswap_dstmem_notifier);
 403         cpu_notifier_register_done();
 404 }
 405
 406 static int __zswap_cpu_comp_notifier(struct zswap_pool *pool,
 407                                      unsigned long action, unsigned long cpu)
 408 {
 409         struct crypto_comp *tfm;
 410
 411         switch (action) {
 412         case CPU_UP_PREPARE:
 413                 if (WARN_ON(*per_cpu_ptr(pool->tfm, cpu)))
 414                         break;
 415                 tfm = crypto_alloc_comp(pool->tfm_name, 0, 0);
 416                 if (IS_ERR_OR_NULL(tfm)) {
 417                         pr_err("could not alloc crypto comp %s : %ld\n",
 418                                pool->tfm_name, PTR_ERR(tfm));
 419                         return NOTIFY_BAD;
 420                 }
 421                 *per_cpu_ptr(pool->tfm, cpu) = tfm;
 422                 break;
 423         case CPU_DEAD:
 424         case CPU_UP_CANCELED:
 425                 tfm = *per_cpu_ptr(pool->tfm, cpu);
 426                 if (!IS_ERR_OR_NULL(tfm))
 427                         crypto_free_comp(tfm);
 428                 *per_cpu_ptr(pool->tfm, cpu) = NULL;
 429                 break;
 430         default:
 431                 break;
 432         }
 433         return NOTIFY_OK;
 434 }
 435
 436 static int zswap_cpu_comp_notifier(struct notifier_block *nb,
 437                                    unsigned long action, void *pcpu)
 438 {
 439         unsigned long cpu = (unsigned long)pcpu;
 440         struct zswap_pool *pool = container_of(nb, typeof(*pool), notifier);
 441
 442         return __zswap_cpu_comp_notifier(pool, action, cpu);
 443 }
 444
 445 static int zswap_cpu_comp_init(struct zswap_pool *pool)
 446 {
 447         unsigned long cpu;
 448
 449         memset(&pool->notifier, 0, sizeof(pool->notifier));
 450         pool->notifier.notifier_call = zswap_cpu_comp_notifier;
 451
 452         cpu_notifier_register_begin();
 453         for_each_online_cpu(cpu)
 454                 if (__zswap_cpu_comp_notifier(pool, CPU_UP_PREPARE, cpu) ==
 455                     NOTIFY_BAD)
 456                         goto cleanup;
 457         __register_cpu_notifier(&pool->notifier);
 458         cpu_notifier_register_done();
 459         return 0;
 460
 461 cleanup:
 462         for_each_online_cpu(cpu)
 463                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 464         cpu_notifier_register_done();
 465         return -ENOMEM;
 466 }
 467
 468 static void zswap_cpu_comp_destroy(struct zswap_pool *pool)
 469 {
 470         unsigned long cpu;
 471
 472         cpu_notifier_register_begin();
 473         for_each_online_cpu(cpu)
 474                 __zswap_cpu_comp_notifier(pool, CPU_UP_CANCELED, cpu);
 475         __unregister_cpu_notifier(&pool->notifier);
 476         cpu_notifier_register_done();
 477 }
 478
 479 /*********************************
 480 * pool functions
 481 **********************************/
 482
 483 static struct zswap_pool *__zswap_pool_current(void)
 484 {
 485         struct zswap_pool *pool;
 486
 487         pool = list_first_or_null_rcu(&zswap_pools, typeof(*pool), list);
 488         WARN_ON(!pool);
 489
 490         return pool;
 491 }
 492
 493 static struct zswap_pool *zswap_pool_current(void)
 494 {
 495         assert_spin_locked(&zswap_pools_lock);
 496
 497         return __zswap_pool_current();
 498 }
 499
 500 static struct zswap_pool *zswap_pool_current_get(void)
 501 {
 502         struct zswap_pool *pool;
 503
 504         rcu_read_lock();
 505
 506         pool = __zswap_pool_current();
 507         if (!pool || !zswap_pool_get(pool))
 508                 pool = NULL;
 509
 510         rcu_read_unlock();
 511
 512         return pool;
 513 }
 514
 515 static struct zswap_pool *zswap_pool_last_get(void)
 516 {
 517         struct zswap_pool *pool, *last = NULL;
 518
 519         rcu_read_lock();
 520
 521         list_for_each_entry_rcu(pool, &zswap_pools, list)
 522                 last = pool;
 523         if (!WARN_ON(!last) && !zswap_pool_get(last))
 524                 last = NULL;
 525
 526         rcu_read_unlock();
 527
 528         return last;
 529 }
 530
 531 static struct zswap_pool *zswap_pool_find_get(char *type, char *compressor)
 532 {
 533         struct zswap_pool *pool;
 534
 535         assert_spin_locked(&zswap_pools_lock);
 536
 537         list_for_each_entry_rcu(pool, &zswap_pools, list) {
 538                 if (strncmp(pool->tfm_name, compressor, sizeof(pool->tfm_name)))
 539                         continue;
 540                 if (strncmp(zpool_get_type(pool->zpool), type,
 541                             sizeof(zswap_zpool_type)))
 542                         continue;
 543                 /* if we can't get it, it's about to be destroyed */
 544                 if (!zswap_pool_get(pool))
 545                         continue;
 546                 return pool;
 547         }
 548
 549         return NULL;
 550 }
 551
 552 static struct zswap_pool *zswap_pool_create(char *type, char *compressor)
 553 {
 554         struct zswap_pool *pool;
 555         gfp_t gfp = __GFP_NORETRY | __GFP_NOWARN;
 556
 557         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
 558         if (!pool) {
 559                 pr_err("pool alloc failed\n");
 560                 return NULL;
 561         }
 562
 563         pool->zpool = zpool_create_pool(type, "zswap", gfp, &zswap_zpool_ops);
 564         if (!pool->zpool) {
 565                 pr_err("%s zpool not available\n", type);
 566                 goto error;
 567         }
 568         pr_debug("using %s zpool\n", zpool_get_type(pool->zpool));
 569
 570         strlcpy(pool->tfm_name, compressor, sizeof(pool->tfm_name));
 571         pool->tfm = alloc_percpu(struct crypto_comp *);
 572         if (!pool->tfm) {
 573                 pr_err("percpu alloc failed\n");
 574                 goto error;
 575         }
 576
 577         if (zswap_cpu_comp_init(pool))
 578                 goto error;
 579         pr_debug("using %s compressor\n", pool->tfm_name);
 580
 581         /* being the current pool takes 1 ref; this func expects the
 582          * caller to always add the new pool as the current pool
 583          */
 584         kref_init(&pool->kref);
 585         INIT_LIST_HEAD(&pool->list);
 586
 587         zswap_pool_debug("created", pool);
 588
 589         return pool;
 590
 591 error:
 592         free_percpu(pool->tfm);
 593         if (pool->zpool)
 594                 zpool_destroy_pool(pool->zpool);
 595         kfree(pool);
 596         return NULL;
 597 }
 598
 599 static struct zswap_pool *__zswap_pool_create_fallback(void)
 600 {
 601         if (!crypto_has_comp(zswap_compressor, 0, 0)) {
 602                 pr_err("compressor %s not available, using default %s\n",
 603                        zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT);
 604                 strncpy(zswap_compressor, ZSWAP_COMPRESSOR_DEFAULT,
 605                         sizeof(zswap_compressor));
 606         }
 607         if (!zpool_has_pool(zswap_zpool_type)) {
 608                 pr_err("zpool %s not available, using default %s\n",
 609                        zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT);
 610                 strncpy(zswap_zpool_type, ZSWAP_ZPOOL_DEFAULT,
 611                         sizeof(zswap_zpool_type));
 612         }
 613
 614         return zswap_pool_create(zswap_zpool_type, zswap_compressor);
 615 }
 616
 617 static void zswap_pool_destroy(struct zswap_pool *pool)
 618 {
 619         zswap_pool_debug("destroying", pool);
 620
 621         zswap_cpu_comp_destroy(pool);
 622         free_percpu(pool->tfm);
 623         zpool_destroy_pool(pool->zpool);
 624         kfree(pool);
 625 }
 626
 627 static int __must_check zswap_pool_get(struct zswap_pool *pool)
 628 {
 629         return kref_get_unless_zero(&pool->kref);
 630 }
 631
 632 static void __zswap_pool_release(struct rcu_head *head)
 633 {
 634         struct zswap_pool *pool = container_of(head, typeof(*pool), rcu_head);
 635
 636         /* nobody should have been able to get a kref... */
 637         WARN_ON(kref_get_unless_zero(&pool->kref));
 638
 639         /* pool is now off zswap_pools list and has no references. */
 640         zswap_pool_destroy(pool);
 641 }
 642
 643 static void __zswap_pool_empty(struct kref *kref)
 644 {
 645         struct zswap_pool *pool;
 646
 647         pool = container_of(kref, typeof(*pool), kref);
 648
 649         spin_lock(&zswap_pools_lock);
 650
 651         WARN_ON(pool == zswap_pool_current());
 652
 653         list_del_rcu(&pool->list);
 654         call_rcu(&pool->rcu_head, __zswap_pool_release);
 655
 656         spin_unlock(&zswap_pools_lock);
 657 }
 658
 659 static void zswap_pool_put(struct zswap_pool *pool)
 660 {
 661         kref_put(&pool->kref, __zswap_pool_empty);
 662 }
 663
 664 /*********************************
 665 * writeback code
 666 **********************************/
 667 /* return enum for zswap_get_swap_cache_page */
 668 enum zswap_get_swap_ret {
 669         ZSWAP_SWAPCACHE_NEW,
 670         ZSWAP_SWAPCACHE_EXIST,
 671         ZSWAP_SWAPCACHE_FAIL,
 672 };
 673
 674 /*
 675  * zswap_get_swap_cache_page
 676  *
 677  * This is an adaption of read_swap_cache_async()
 678  *
 679  * This function tries to find a page with the given swap entry
 680  * in the swapper_space address space (the swap cache).  If the page
 681  * is found, it is returned in retpage.  Otherwise, a page is allocated,
 682  * added to the swap cache, and returned in retpage.
 683  *
 684  * If success, the swap cache page is returned in retpage
 685  * Returns ZSWAP_SWAPCACHE_EXIST if page was already in the swap cache
 686  * Returns ZSWAP_SWAPCACHE_NEW if the new page needs to be populated,
 687  *     the new page is added to swapcache and locked
 688  * Returns ZSWAP_SWAPCACHE_FAIL on error
 689  */
 690 static int zswap_get_swap_cache_page(swp_entry_t entry,
 691                                 struct page **retpage)
 692 {
 693         bool page_was_allocated;
 694
 695         *retpage = __read_swap_cache_async(entry, GFP_KERNEL,
 696                         NULL, 0, &page_was_allocated);
 697         if (page_was_allocated)
 698                 return ZSWAP_SWAPCACHE_NEW;
 699         if (!*retpage)
 700                 return ZSWAP_SWAPCACHE_FAIL;
 701         return ZSWAP_SWAPCACHE_EXIST;
 702 }
 703
 704 /*
 705  * Attempts to free an entry by adding a page to the swap cache,
 706  * decompressing the entry data into the page, and issuing a
 707  * bio write to write the page back to the swap device.
 708  *
 709  * This can be thought of as a "resumed writeback" of the page
 710  * to the swap device.  We are basically resuming the same swap
 711  * writeback path that was intercepted with the frontswap_store()
 712  * in the first place.  After the page has been decompressed into
 713  * the swap cache, the compressed version stored by zswap can be
 714  * freed.
 715  */
 716 static int zswap_writeback_entry(struct zpool *pool, unsigned long handle)
 717 {
 718         struct zswap_header *zhdr;
 719         swp_entry_t swpentry;
 720         struct zswap_tree *tree;
 721         pgoff_t offset;
 722         struct zswap_entry *entry;
 723         struct page *page;
 724         struct crypto_comp *tfm;
 725         u8 *src, *dst;
 726         unsigned int dlen;
 727         int ret;
 728         struct writeback_control wbc = {
 729                 .sync_mode = WB_SYNC_NONE,
 730         };
 731
 732         /* extract swpentry from data */
 733         zhdr = zpool_map_handle(pool, handle, ZPOOL_MM_RO);
 734         swpentry = zhdr->swpentry; /* here */
 735         zpool_unmap_handle(pool, handle);
 736         tree = zswap_trees[swp_type(swpentry)];
 737         offset = swp_offset(swpentry);
 738
 739         /* find and ref zswap entry */
 740         spin_lock(&tree->lock);
 741         entry = zswap_entry_find_get(&tree->rbroot, offset);
 742         if (!entry) {
 743                 /* entry was invalidated */
 744                 spin_unlock(&tree->lock);
 745                 return 0;
 746         }
 747         spin_unlock(&tree->lock);
 748         BUG_ON(offset != entry->offset);
 749
 750         /* try to allocate swap cache page */
 751         switch (zswap_get_swap_cache_page(swpentry, &page)) {
 752         case ZSWAP_SWAPCACHE_FAIL: /* no memory or invalidate happened */
 753                 ret = -ENOMEM;
 754                 goto fail;
 755
 756         case ZSWAP_SWAPCACHE_EXIST:
 757                 /* page is already in the swap cache, ignore for now */
 758                 page_cache_release(page);
 759                 ret = -EEXIST;
 760                 goto fail;
 761
 762         case ZSWAP_SWAPCACHE_NEW: /* page is locked */
 763                 /* decompress */
 764                 dlen = PAGE_SIZE;
 765                 src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 766                                 ZPOOL_MM_RO) + sizeof(struct zswap_header);
 767                 dst = kmap_atomic(page);
 768                 tfm = *get_cpu_ptr(entry->pool->tfm);
 769                 ret = crypto_comp_decompress(tfm, src, entry->length,
 770                                              dst, &dlen);
 771                 put_cpu_ptr(entry->pool->tfm);
 772                 kunmap_atomic(dst);
 773                 zpool_unmap_handle(entry->pool->zpool, entry->handle);
 774                 BUG_ON(ret);
 775                 BUG_ON(dlen != PAGE_SIZE);
 776
 777                 /* page is up to date */
 778                 SetPageUptodate(page);
 779         }
 780
 781         /* move it to the tail of the inactive list after end_writeback */
 782         SetPageReclaim(page);
 783
 784         /* start writeback */
 785         __swap_writepage(page, &wbc, end_swap_bio_write);
 786         page_cache_release(page);
 787         zswap_written_back_pages++;
 788
 789         spin_lock(&tree->lock);
 790         /* drop local reference */
 791         zswap_entry_put(tree, entry);
 792
 793         /*
 794         * There are two possible situations for entry here:
 795         * (1) refcount is 1(normal case),  entry is valid and on the tree
 796         * (2) refcount is 0, entry is freed and not on the tree
 797         *     because invalidate happened during writeback
 798         *  search the tree and free the entry if find entry
 799         */
 800         if (entry == zswap_rb_search(&tree->rbroot, offset))
 801                 zswap_entry_put(tree, entry);
 802         spin_unlock(&tree->lock);
 803
 804         goto end;
 805
 806         /*
 807         * if we get here due to ZSWAP_SWAPCACHE_EXIST
 808         * a load may happening concurrently
 809         * it is safe and okay to not free the entry
 810         * if we free the entry in the following put
 811         * it it either okay to return !0
 812         */
 813 fail:
 814         spin_lock(&tree->lock);
 815         zswap_entry_put(tree, entry);
 816         spin_unlock(&tree->lock);
 817
 818 end:
 819         return ret;
 820 }
 821
 822 static int zswap_shrink(void)
 823 {
 824         struct zswap_pool *pool;
 825         int ret;
 826
 827         pool = zswap_pool_last_get();
 828         if (!pool)
 829                 return -ENOENT;
 830
 831         ret = zpool_shrink(pool->zpool, 1, NULL);
 832
 833         zswap_pool_put(pool);
 834
 835         return ret;
 836 }
 837
 838 /*********************************
 839 * frontswap hooks
 840 **********************************/
 841 /* attempts to compress and store an single page */
 842 static int zswap_frontswap_store(unsigned type, pgoff_t offset,
 843                                 struct page *page)
 844 {
 845         struct zswap_tree *tree = zswap_trees[type];
 846         struct zswap_entry *entry, *dupentry;
 847         struct crypto_comp *tfm;
 848         int ret;
 849         unsigned int dlen = PAGE_SIZE, len;
 850         unsigned long handle;
 851         char *buf;
 852         u8 *src, *dst;
 853         struct zswap_header *zhdr;
 854
 855         if (!zswap_enabled || !tree) {
 856                 ret = -ENODEV;
 857                 goto reject;
 858         }
 859
 860         /* reclaim space if needed */
 861         if (zswap_is_full()) {
 862                 zswap_pool_limit_hit++;
 863                 if (zswap_shrink()) {
 864                         zswap_reject_reclaim_fail++;
 865                         ret = -ENOMEM;
 866                         goto reject;
 867                 }
 868         }
 869
 870         /* allocate entry */
 871         entry = zswap_entry_cache_alloc(GFP_KERNEL);
 872         if (!entry) {
 873                 zswap_reject_kmemcache_fail++;
 874                 ret = -ENOMEM;
 875                 goto reject;
 876         }
 877
 878         /* if entry is successfully added, it keeps the reference */
 879         entry->pool = zswap_pool_current_get();
 880         if (!entry->pool) {
 881                 ret = -EINVAL;
 882                 goto freepage;
 883         }
 884
 885         /* compress */
 886         dst = get_cpu_var(zswap_dstmem);
 887         tfm = *get_cpu_ptr(entry->pool->tfm);
 888         src = kmap_atomic(page);
 889         ret = crypto_comp_compress(tfm, src, PAGE_SIZE, dst, &dlen);
 890         kunmap_atomic(src);
 891         put_cpu_ptr(entry->pool->tfm);
 892         if (ret) {
 893                 ret = -EINVAL;
 894                 goto put_dstmem;
 895         }
 896
 897         /* store */
 898         len = dlen + sizeof(struct zswap_header);
 899         ret = zpool_malloc(entry->pool->zpool, len,
 900                            __GFP_NORETRY | __GFP_NOWARN, &handle);
 901         if (ret == -ENOSPC) {
 902                 zswap_reject_compress_poor++;
 903                 goto put_dstmem;
 904         }
 905         if (ret) {
 906                 zswap_reject_alloc_fail++;
 907                 goto put_dstmem;
 908         }
 909         zhdr = zpool_map_handle(entry->pool->zpool, handle, ZPOOL_MM_RW);
 910         zhdr->swpentry = swp_entry(type, offset);
 911         buf = (u8 *)(zhdr + 1);
 912         memcpy(buf, dst, dlen);
 913         zpool_unmap_handle(entry->pool->zpool, handle);
 914         put_cpu_var(zswap_dstmem);
 915
 916         /* populate entry */
 917         entry->offset = offset;
 918         entry->handle = handle;
 919         entry->length = dlen;
 920
 921         /* map */
 922         spin_lock(&tree->lock);
 923         do {
 924                 ret = zswap_rb_insert(&tree->rbroot, entry, &dupentry);
 925                 if (ret == -EEXIST) {
 926                         zswap_duplicate_entry++;
 927                         /* remove from rbtree */
 928                         zswap_rb_erase(&tree->rbroot, dupentry);
 929                         zswap_entry_put(tree, dupentry);
 930                 }
 931         } while (ret == -EEXIST);
 932         spin_unlock(&tree->lock);
 933
 934         /* update stats */
 935         atomic_inc(&zswap_stored_pages);
 936         zswap_update_total_size();
 937
 938         return 0;
 939
 940 put_dstmem:
 941         put_cpu_var(zswap_dstmem);
 942         zswap_pool_put(entry->pool);
 943 freepage:
 944         zswap_entry_cache_free(entry);
 945 reject:
 946         return ret;
 947 }
 948
 949 /*
 950  * returns 0 if the page was successfully decompressed
 951  * return -1 on entry not found or error
 952 */
 953 static int zswap_frontswap_load(unsigned type, pgoff_t offset,
 954                                 struct page *page)
 955 {
 956         struct zswap_tree *tree = zswap_trees[type];
 957         struct zswap_entry *entry;
 958         struct crypto_comp *tfm;
 959         u8 *src, *dst;
 960         unsigned int dlen;
 961         int ret;
 962
 963         /* find */
 964         spin_lock(&tree->lock);
 965         entry = zswap_entry_find_get(&tree->rbroot, offset);
 966         if (!entry) {
 967                 /* entry was written back */
 968                 spin_unlock(&tree->lock);
 969                 return -1;
 970         }
 971         spin_unlock(&tree->lock);
 972
 973         /* decompress */
 974         dlen = PAGE_SIZE;
 975         src = (u8 *)zpool_map_handle(entry->pool->zpool, entry->handle,
 976                         ZPOOL_MM_RO) + sizeof(struct zswap_header);
 977         dst = kmap_atomic(page);
 978         tfm = *get_cpu_ptr(entry->pool->tfm);
 979         ret = crypto_comp_decompress(tfm, src, entry->length, dst, &dlen);
 980         put_cpu_ptr(entry->pool->tfm);
 981         kunmap_atomic(dst);
 982         zpool_unmap_handle(entry->pool->zpool, entry->handle);
 983         BUG_ON(ret);
 984
 985         spin_lock(&tree->lock);
 986         zswap_entry_put(tree, entry);
 987         spin_unlock(&tree->lock);
 988
 989         return 0;
 990 }
 991
 992 /* frees an entry in zswap */
 993 static void zswap_frontswap_invalidate_page(unsigned type, pgoff_t offset)
 994 {
 995         struct zswap_tree *tree = zswap_trees[type];
 996         struct zswap_entry *entry;
 997
 998         /* find */
 999         spin_lock(&tree->lock);
1000         entry = zswap_rb_search(&tree->rbroot, offset);
1001         if (!entry) {
1002                 /* entry was written back */
1003                 spin_unlock(&tree->lock);
1004                 return;
1005         }
1006
1007         /* remove from rbtree */
1008         zswap_rb_erase(&tree->rbroot, entry);
1009
1010         /* drop the initial reference from entry creation */
1011         zswap_entry_put(tree, entry);
1012
1013         spin_unlock(&tree->lock);
1014 }
1015
1016 /* frees all zswap entries for the given swap type */
1017 static void zswap_frontswap_invalidate_area(unsigned type)
1018 {
1019         struct zswap_tree *tree = zswap_trees[type];
1020         struct zswap_entry *entry, *n;
1021
1022         if (!tree)
1023                 return;
1024
1025         /* walk the tree and free everything */
1026         spin_lock(&tree->lock);
1027         rbtree_postorder_for_each_entry_safe(entry, n, &tree->rbroot, rbnode)
1028                 zswap_free_entry(entry);
1029         tree->rbroot = RB_ROOT;
1030         spin_unlock(&tree->lock);
1031         kfree(tree);
1032         zswap_trees[type] = NULL;
1033 }
1034
1035 static void zswap_frontswap_init(unsigned type)
1036 {
1037         struct zswap_tree *tree;
1038
1039         tree = kzalloc(sizeof(struct zswap_tree), GFP_KERNEL);
1040         if (!tree) {
1041                 pr_err("alloc failed, zswap disabled for swap type %d\n", type);
1042                 return;
1043         }
1044
1045         tree->rbroot = RB_ROOT;
1046         spin_lock_init(&tree->lock);
1047         zswap_trees[type] = tree;
1048 }
1049
1050 static struct frontswap_ops zswap_frontswap_ops = {
1051         .store = zswap_frontswap_store,
1052         .load = zswap_frontswap_load,
1053         .invalidate_page = zswap_frontswap_invalidate_page,
1054         .invalidate_area = zswap_frontswap_invalidate_area,
1055         .init = zswap_frontswap_init
1056 };
1057
1058 /*********************************
1059 * debugfs functions
1060 **********************************/
1061 #ifdef CONFIG_DEBUG_FS
1062 #include <linux/debugfs.h>
1063
1064 static struct dentry *zswap_debugfs_root;
1065
1066 static int __init zswap_debugfs_init(void)
1067 {
1068         if (!debugfs_initialized())
1069                 return -ENODEV;
1070
1071         zswap_debugfs_root = debugfs_create_dir("zswap", NULL);
1072         if (!zswap_debugfs_root)
1073                 return -ENOMEM;
1074
1075         debugfs_create_u64("pool_limit_hit", S_IRUGO,
1076                         zswap_debugfs_root, &zswap_pool_limit_hit);
1077         debugfs_create_u64("reject_reclaim_fail", S_IRUGO,
1078                         zswap_debugfs_root, &zswap_reject_reclaim_fail);
1079         debugfs_create_u64("reject_alloc_fail", S_IRUGO,
1080                         zswap_debugfs_root, &zswap_reject_alloc_fail);
1081         debugfs_create_u64("reject_kmemcache_fail", S_IRUGO,
1082                         zswap_debugfs_root, &zswap_reject_kmemcache_fail);
1083         debugfs_create_u64("reject_compress_poor", S_IRUGO,
1084                         zswap_debugfs_root, &zswap_reject_compress_poor);
1085         debugfs_create_u64("written_back_pages", S_IRUGO,
1086                         zswap_debugfs_root, &zswap_written_back_pages);
1087         debugfs_create_u64("duplicate_entry", S_IRUGO,
1088                         zswap_debugfs_root, &zswap_duplicate_entry);
1089         debugfs_create_u64("pool_total_size", S_IRUGO,
1090                         zswap_debugfs_root, &zswap_pool_total_size);
1091         debugfs_create_atomic_t("stored_pages", S_IRUGO,
1092                         zswap_debugfs_root, &zswap_stored_pages);
1093
1094         return 0;
1095 }
1096
1097 static void __exit zswap_debugfs_exit(void)
1098 {
1099         debugfs_remove_recursive(zswap_debugfs_root);
1100 }
1101 #else
1102 static int __init zswap_debugfs_init(void)
1103 {
1104         return 0;
1105 }
1106
1107 static void __exit zswap_debugfs_exit(void) { }
1108 #endif
1109
1110 /*********************************
1111 * module init and exit
1112 **********************************/
1113 static int __init init_zswap(void)
1114 {
1115         struct zswap_pool *pool;
1116
1117         if (zswap_entry_cache_create()) {
1118                 pr_err("entry cache creation failed\n");
1119                 goto cache_fail;
1120         }
1121
1122         if (zswap_cpu_dstmem_init()) {
1123                 pr_err("dstmem alloc failed\n");
1124                 goto dstmem_fail;
1125         }
1126
1127         pool = __zswap_pool_create_fallback();
1128         if (!pool) {
1129                 pr_err("pool creation failed\n");
1130                 goto pool_fail;
1131         }
1132         pr_info("loaded using pool %s/%s\n", pool->tfm_name,
1133                 zpool_get_type(pool->zpool));
1134
1135         list_add(&pool->list, &zswap_pools);
1136
1137         frontswap_register_ops(&zswap_frontswap_ops);
1138         if (zswap_debugfs_init())
1139                 pr_warn("debugfs initialization failed\n");
1140         return 0;
1141
1142 pool_fail:
1143         zswap_cpu_dstmem_destroy();
1144 dstmem_fail:
1145         zswap_entry_cache_destroy();
1146 cache_fail:
1147         return -ENOMEM;
1148 }
1149 /* must be late so crypto has time to come up */
1150 late_initcall(init_zswap);
1151
1152 MODULE_LICENSE("GPL");
1153 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
1154 MODULE_DESCRIPTION("Compressed cache for swap pages");