mm/z3fold.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * z3fold.c
   4  *
   5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   6  * Copyright (C) 2016, Sony Mobile Communications Inc.
   7  *
   8  * This implementation is based on zbud written by Seth Jennings.
   9  *
  10  * z3fold is an special purpose allocator for storing compressed pages. It
  11  * can store up to three compressed pages per page which improves the
  12  * compression ratio of zbud while retaining its main concepts (e. g. always
  13  * storing an integral number of objects per page) and simplicity.
  14  * It still has simple and deterministic reclaim properties that make it
  15  * preferable to a higher density approach (with no requirement on integral
  16  * number of object per page) when reclaim is used.
  17  *
  18  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  19  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  20  *
  21  * z3fold doesn't export any API and is meant to be used via zpool API.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/atomic.h>
  27 #include <linux/sched.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/mount.h>
  38 #include <linux/pseudo_fs.h>
  39 #include <linux/fs.h>
  40 #include <linux/preempt.h>
  41 #include <linux/workqueue.h>
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/zpool.h>
  45 #include <linux/magic.h>
  46
  47 /*
  48  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  49  * adjusting internal fragmentation.  It also determines the number of
  50  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  51  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  52  * in the beginning of an allocated page are occupied by z3fold header, so
  53  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  54  * which shows the max number of free chunks in z3fold page, also there will
  55  * be 63, or 62, respectively, freelists per pool.
  56  */
  57 #define NCHUNKS_ORDER   6
  58
  59 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  60 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  61 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  62 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  63 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  64 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
  65
  66 #define BUDDY_MASK      (0x3)
  67 #define BUDDY_SHIFT     2
  68 #define SLOTS_ALIGN     (0x40)
  69
  70 /*****************
  71  * Structures
  72 *****************/
  73 struct z3fold_pool;
  74 struct z3fold_ops {
  75         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  76 };
  77
  78 enum buddy {
  79         HEADLESS = 0,
  80         FIRST,
  81         MIDDLE,
  82         LAST,
  83         BUDDIES_MAX = LAST
  84 };
  85
  86 struct z3fold_buddy_slots {
  87         /*
  88          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  89          * be enough slots to hold all possible variants
  90          */
  91         unsigned long slot[BUDDY_MASK + 1];
  92         unsigned long pool; /* back link + flags */
  93         rwlock_t lock;
  94 };
  95 #define HANDLE_FLAG_MASK        (0x03)
  96
  97 /*
  98  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  99  *                      z3fold page, except for HEADLESS pages
 100  * @buddy:              links the z3fold page into the relevant list in the
 101  *                      pool
 102  * @page_lock:          per-page lock
 103  * @refcount:           reference count for the z3fold page
 104  * @work:               work_struct for page layout optimization
 105  * @slots:              pointer to the structure holding buddy slots
 106  * @pool:               pointer to the containing pool
 107  * @cpu:                CPU which this page "belongs" to
 108  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 109  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 110  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 111  * @first_num:          the starting number (for the first handle)
 112  * @mapped_count:       the number of objects currently mapped
 113  */
 114 struct z3fold_header {
 115         struct list_head buddy;
 116         spinlock_t page_lock;
 117         struct kref refcount;
 118         struct work_struct work;
 119         struct z3fold_buddy_slots *slots;
 120         struct z3fold_pool *pool;
 121         short cpu;
 122         unsigned short first_chunks;
 123         unsigned short middle_chunks;
 124         unsigned short last_chunks;
 125         unsigned short start_middle;
 126         unsigned short first_num:2;
 127         unsigned short mapped_count:2;
 128         unsigned short foreign_handles:2;
 129 };
 130
 131 /**
 132  * struct z3fold_pool - stores metadata for each z3fold pool
 133  * @name:       pool name
 134  * @lock:       protects pool unbuddied/lru lists
 135  * @stale_lock: protects pool stale page list
 136  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 137  *              buddies; the list each z3fold page is added to depends on
 138  *              the size of its free region.
 139  * @lru:        list tracking the z3fold pages in LRU order by most recently
 140  *              added buddy.
 141  * @stale:      list of pages marked for freeing
 142  * @pages_nr:   number of z3fold pages in the pool.
 143  * @c_handle:   cache for z3fold_buddy_slots allocation
 144  * @ops:        pointer to a structure of user defined operations specified at
 145  *              pool creation time.
 146  * @compact_wq: workqueue for page layout background optimization
 147  * @release_wq: workqueue for safe page release
 148  * @work:       work_struct for safe page release
 149  * @inode:      inode for z3fold pseudo filesystem
 150  *
 151  * This structure is allocated at pool creation time and maintains metadata
 152  * pertaining to a particular z3fold pool.
 153  */
 154 struct z3fold_pool {
 155         const char *name;
 156         spinlock_t lock;
 157         spinlock_t stale_lock;
 158         struct list_head *unbuddied;
 159         struct list_head lru;
 160         struct list_head stale;
 161         atomic64_t pages_nr;
 162         struct kmem_cache *c_handle;
 163         const struct z3fold_ops *ops;
 164         struct zpool *zpool;
 165         const struct zpool_ops *zpool_ops;
 166         struct workqueue_struct *compact_wq;
 167         struct workqueue_struct *release_wq;
 168         struct work_struct work;
 169         struct inode *inode;
 170 };
 171
 172 /*
 173  * Internal z3fold page flags
 174  */
 175 enum z3fold_page_flags {
 176         PAGE_HEADLESS = 0,
 177         MIDDLE_CHUNK_MAPPED,
 178         NEEDS_COMPACTING,
 179         PAGE_STALE,
 180         PAGE_CLAIMED, /* by either reclaim or free */
 181 };
 182
 183 /*
 184  * handle flags, go under HANDLE_FLAG_MASK
 185  */
 186 enum z3fold_handle_flags {
 187         HANDLES_ORPHANED = 0,
 188 };
 189
 190 /*
 191  * Forward declarations
 192  */
 193 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 194 static void compact_page_work(struct work_struct *w);
 195
 196 /*****************
 197  * Helpers
 198 *****************/
 199
 200 /* Converts an allocation size in bytes to size in z3fold chunks */
 201 static int size_to_chunks(size_t size)
 202 {
 203         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 204 }
 205
 206 #define for_each_unbuddied_list(_iter, _begin) \
 207         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 208
 209 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 210                                                         gfp_t gfp)
 211 {
 212         struct z3fold_buddy_slots *slots;
 213
 214         slots = kmem_cache_alloc(pool->c_handle,
 215                                  (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
 216
 217         if (slots) {
 218                 memset(slots->slot, 0, sizeof(slots->slot));
 219                 slots->pool = (unsigned long)pool;
 220                 rwlock_init(&slots->lock);
 221         }
 222
 223         return slots;
 224 }
 225
 226 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 227 {
 228         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 229 }
 230
 231 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 232 {
 233         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 234 }
 235
 236 /* Lock a z3fold page */
 237 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 238 {
 239         spin_lock(&zhdr->page_lock);
 240 }
 241
 242 /* Try to lock a z3fold page */
 243 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 244 {
 245         return spin_trylock(&zhdr->page_lock);
 246 }
 247
 248 /* Unlock a z3fold page */
 249 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 250 {
 251         spin_unlock(&zhdr->page_lock);
 252 }
 253
 254
 255 static inline struct z3fold_header *__get_z3fold_header(unsigned long handle,
 256                                                         bool lock)
 257 {
 258         struct z3fold_buddy_slots *slots;
 259         struct z3fold_header *zhdr;
 260         int locked = 0;
 261
 262         if (!(handle & (1 << PAGE_HEADLESS))) {
 263                 slots = handle_to_slots(handle);
 264                 do {
 265                         unsigned long addr;
 266
 267                         read_lock(&slots->lock);
 268                         addr = *(unsigned long *)handle;
 269                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 270                         if (lock)
 271                                 locked = z3fold_page_trylock(zhdr);
 272                         read_unlock(&slots->lock);
 273                         if (locked)
 274                                 break;
 275                         cpu_relax();
 276                 } while (lock);
 277         } else {
 278                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 279         }
 280
 281         return zhdr;
 282 }
 283
 284 /* Returns the z3fold page where a given handle is stored */
 285 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
 286 {
 287         return __get_z3fold_header(h, false);
 288 }
 289
 290 /* return locked z3fold page if it's not headless */
 291 static inline struct z3fold_header *get_z3fold_header(unsigned long h)
 292 {
 293         return __get_z3fold_header(h, true);
 294 }
 295
 296 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 297 {
 298         struct page *page = virt_to_page(zhdr);
 299
 300         if (!test_bit(PAGE_HEADLESS, &page->private))
 301                 z3fold_page_unlock(zhdr);
 302 }
 303
 304 static inline void free_handle(unsigned long handle)
 305 {
 306         struct z3fold_buddy_slots *slots;
 307         struct z3fold_header *zhdr;
 308         int i;
 309         bool is_free;
 310
 311         if (handle & (1 << PAGE_HEADLESS))
 312                 return;
 313
 314         if (WARN_ON(*(unsigned long *)handle == 0))
 315                 return;
 316
 317         zhdr = handle_to_z3fold_header(handle);
 318         slots = handle_to_slots(handle);
 319         write_lock(&slots->lock);
 320         *(unsigned long *)handle = 0;
 321         write_unlock(&slots->lock);
 322         if (zhdr->slots == slots)
 323                 return; /* simple case, nothing else to do */
 324
 325         /* we are freeing a foreign handle if we are here */
 326         zhdr->foreign_handles--;
 327         is_free = true;
 328         read_lock(&slots->lock);
 329         if (!test_bit(HANDLES_ORPHANED, &slots->pool)) {
 330                 read_unlock(&slots->lock);
 331                 return;
 332         }
 333         for (i = 0; i <= BUDDY_MASK; i++) {
 334                 if (slots->slot[i]) {
 335                         is_free = false;
 336                         break;
 337                 }
 338         }
 339         read_unlock(&slots->lock);
 340
 341         if (is_free) {
 342                 struct z3fold_pool *pool = slots_to_pool(slots);
 343
 344                 kmem_cache_free(pool->c_handle, slots);
 345         }
 346 }
 347
 348 static int z3fold_init_fs_context(struct fs_context *fc)
 349 {
 350         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 351 }
 352
 353 static struct file_system_type z3fold_fs = {
 354         .name           = "z3fold",
 355         .init_fs_context = z3fold_init_fs_context,
 356         .kill_sb        = kill_anon_super,
 357 };
 358
 359 static struct vfsmount *z3fold_mnt;
 360 static int z3fold_mount(void)
 361 {
 362         int ret = 0;
 363
 364         z3fold_mnt = kern_mount(&z3fold_fs);
 365         if (IS_ERR(z3fold_mnt))
 366                 ret = PTR_ERR(z3fold_mnt);
 367
 368         return ret;
 369 }
 370
 371 static void z3fold_unmount(void)
 372 {
 373         kern_unmount(z3fold_mnt);
 374 }
 375
 376 static const struct address_space_operations z3fold_aops;
 377 static int z3fold_register_migration(struct z3fold_pool *pool)
 378 {
 379         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 380         if (IS_ERR(pool->inode)) {
 381                 pool->inode = NULL;
 382                 return 1;
 383         }
 384
 385         pool->inode->i_mapping->private_data = pool;
 386         pool->inode->i_mapping->a_ops = &z3fold_aops;
 387         return 0;
 388 }
 389
 390 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 391 {
 392         if (pool->inode)
 393                 iput(pool->inode);
 394  }
 395
 396 /* Initializes the z3fold header of a newly allocated z3fold page */
 397 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 398                                         struct z3fold_pool *pool, gfp_t gfp)
 399 {
 400         struct z3fold_header *zhdr = page_address(page);
 401         struct z3fold_buddy_slots *slots;
 402
 403         INIT_LIST_HEAD(&page->lru);
 404         clear_bit(PAGE_HEADLESS, &page->private);
 405         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 406         clear_bit(NEEDS_COMPACTING, &page->private);
 407         clear_bit(PAGE_STALE, &page->private);
 408         clear_bit(PAGE_CLAIMED, &page->private);
 409         if (headless)
 410                 return zhdr;
 411
 412         slots = alloc_slots(pool, gfp);
 413         if (!slots)
 414                 return NULL;
 415
 416         spin_lock_init(&zhdr->page_lock);
 417         kref_init(&zhdr->refcount);
 418         zhdr->first_chunks = 0;
 419         zhdr->middle_chunks = 0;
 420         zhdr->last_chunks = 0;
 421         zhdr->first_num = 0;
 422         zhdr->start_middle = 0;
 423         zhdr->cpu = -1;
 424         zhdr->foreign_handles = 0;
 425         zhdr->slots = slots;
 426         zhdr->pool = pool;
 427         INIT_LIST_HEAD(&zhdr->buddy);
 428         INIT_WORK(&zhdr->work, compact_page_work);
 429         return zhdr;
 430 }
 431
 432 /* Resets the struct page fields and frees the page */
 433 static void free_z3fold_page(struct page *page, bool headless)
 434 {
 435         if (!headless) {
 436                 lock_page(page);
 437                 __ClearPageMovable(page);
 438                 unlock_page(page);
 439         }
 440         ClearPagePrivate(page);
 441         __free_page(page);
 442 }
 443
 444 /* Helper function to build the index */
 445 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 446 {
 447         return (bud + zhdr->first_num) & BUDDY_MASK;
 448 }
 449
 450 /*
 451  * Encodes the handle of a particular buddy within a z3fold page
 452  * Pool lock should be held as this function accesses first_num
 453  */
 454 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 455                                 struct z3fold_buddy_slots *slots,
 456                                 enum buddy bud)
 457 {
 458         unsigned long h = (unsigned long)zhdr;
 459         int idx = 0;
 460
 461         /*
 462          * For a headless page, its handle is its pointer with the extra
 463          * PAGE_HEADLESS bit set
 464          */
 465         if (bud == HEADLESS)
 466                 return h | (1 << PAGE_HEADLESS);
 467
 468         /* otherwise, return pointer to encoded handle */
 469         idx = __idx(zhdr, bud);
 470         h += idx;
 471         if (bud == LAST)
 472                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 473
 474         write_lock(&slots->lock);
 475         slots->slot[idx] = h;
 476         write_unlock(&slots->lock);
 477         return (unsigned long)&slots->slot[idx];
 478 }
 479
 480 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 481 {
 482         return __encode_handle(zhdr, zhdr->slots, bud);
 483 }
 484
 485 /* only for LAST bud, returns zero otherwise */
 486 static unsigned short handle_to_chunks(unsigned long handle)
 487 {
 488         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 489         unsigned long addr;
 490
 491         read_lock(&slots->lock);
 492         addr = *(unsigned long *)handle;
 493         read_unlock(&slots->lock);
 494         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 495 }
 496
 497 /*
 498  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 499  *  but that doesn't matter. because the masking will result in the
 500  *  correct buddy number.
 501  */
 502 static enum buddy handle_to_buddy(unsigned long handle)
 503 {
 504         struct z3fold_header *zhdr;
 505         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 506         unsigned long addr;
 507
 508         read_lock(&slots->lock);
 509         WARN_ON(handle & (1 << PAGE_HEADLESS));
 510         addr = *(unsigned long *)handle;
 511         read_unlock(&slots->lock);
 512         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 513         return (addr - zhdr->first_num) & BUDDY_MASK;
 514 }
 515
 516 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 517 {
 518         return zhdr->pool;
 519 }
 520
 521 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 522 {
 523         struct page *page = virt_to_page(zhdr);
 524         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 525         bool is_free = true;
 526         int i;
 527
 528         WARN_ON(!list_empty(&zhdr->buddy));
 529         set_bit(PAGE_STALE, &page->private);
 530         clear_bit(NEEDS_COMPACTING, &page->private);
 531         spin_lock(&pool->lock);
 532         if (!list_empty(&page->lru))
 533                 list_del_init(&page->lru);
 534         spin_unlock(&pool->lock);
 535
 536         /* If there are no foreign handles, free the handles array */
 537         read_lock(&zhdr->slots->lock);
 538         for (i = 0; i <= BUDDY_MASK; i++) {
 539                 if (zhdr->slots->slot[i]) {
 540                         is_free = false;
 541                         break;
 542                 }
 543         }
 544         if (!is_free)
 545                 set_bit(HANDLES_ORPHANED, &zhdr->slots->pool);
 546         read_unlock(&zhdr->slots->lock);
 547
 548         if (is_free)
 549                 kmem_cache_free(pool->c_handle, zhdr->slots);
 550
 551         if (locked)
 552                 z3fold_page_unlock(zhdr);
 553
 554         spin_lock(&pool->stale_lock);
 555         list_add(&zhdr->buddy, &pool->stale);
 556         queue_work(pool->release_wq, &pool->work);
 557         spin_unlock(&pool->stale_lock);
 558 }
 559
 560 static void __attribute__((__unused__))
 561                         release_z3fold_page(struct kref *ref)
 562 {
 563         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 564                                                 refcount);
 565         __release_z3fold_page(zhdr, false);
 566 }
 567
 568 static void release_z3fold_page_locked(struct kref *ref)
 569 {
 570         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 571                                                 refcount);
 572         WARN_ON(z3fold_page_trylock(zhdr));
 573         __release_z3fold_page(zhdr, true);
 574 }
 575
 576 static void release_z3fold_page_locked_list(struct kref *ref)
 577 {
 578         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 579                                                refcount);
 580         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 581
 582         spin_lock(&pool->lock);
 583         list_del_init(&zhdr->buddy);
 584         spin_unlock(&pool->lock);
 585
 586         WARN_ON(z3fold_page_trylock(zhdr));
 587         __release_z3fold_page(zhdr, true);
 588 }
 589
 590 static void free_pages_work(struct work_struct *w)
 591 {
 592         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 593
 594         spin_lock(&pool->stale_lock);
 595         while (!list_empty(&pool->stale)) {
 596                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 597                                                 struct z3fold_header, buddy);
 598                 struct page *page = virt_to_page(zhdr);
 599
 600                 list_del(&zhdr->buddy);
 601                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 602                         continue;
 603                 spin_unlock(&pool->stale_lock);
 604                 cancel_work_sync(&zhdr->work);
 605                 free_z3fold_page(page, false);
 606                 cond_resched();
 607                 spin_lock(&pool->stale_lock);
 608         }
 609         spin_unlock(&pool->stale_lock);
 610 }
 611
 612 /*
 613  * Returns the number of free chunks in a z3fold page.
 614  * NB: can't be used with HEADLESS pages.
 615  */
 616 static int num_free_chunks(struct z3fold_header *zhdr)
 617 {
 618         int nfree;
 619         /*
 620          * If there is a middle object, pick up the bigger free space
 621          * either before or after it. Otherwise just subtract the number
 622          * of chunks occupied by the first and the last objects.
 623          */
 624         if (zhdr->middle_chunks != 0) {
 625                 int nfree_before = zhdr->first_chunks ?
 626                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 627                 int nfree_after = zhdr->last_chunks ?
 628                         0 : TOTAL_CHUNKS -
 629                                 (zhdr->start_middle + zhdr->middle_chunks);
 630                 nfree = max(nfree_before, nfree_after);
 631         } else
 632                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 633         return nfree;
 634 }
 635
 636 /* Add to the appropriate unbuddied list */
 637 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 638                                 struct z3fold_header *zhdr)
 639 {
 640         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 641                         zhdr->middle_chunks == 0) {
 642                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 643
 644                 int freechunks = num_free_chunks(zhdr);
 645                 spin_lock(&pool->lock);
 646                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 647                 spin_unlock(&pool->lock);
 648                 zhdr->cpu = smp_processor_id();
 649                 put_cpu_ptr(pool->unbuddied);
 650         }
 651 }
 652
 653 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 654                                 unsigned short dst_chunk)
 655 {
 656         void *beg = zhdr;
 657         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 658                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 659                        zhdr->middle_chunks << CHUNK_SHIFT);
 660 }
 661
 662 static inline bool buddy_single(struct z3fold_header *zhdr)
 663 {
 664         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 665                         (zhdr->first_chunks && zhdr->last_chunks) ||
 666                         (zhdr->middle_chunks && zhdr->last_chunks));
 667 }
 668
 669 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 670 {
 671         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 672         void *p = zhdr;
 673         unsigned long old_handle = 0;
 674         size_t sz = 0;
 675         struct z3fold_header *new_zhdr = NULL;
 676         int first_idx = __idx(zhdr, FIRST);
 677         int middle_idx = __idx(zhdr, MIDDLE);
 678         int last_idx = __idx(zhdr, LAST);
 679         unsigned short *moved_chunks = NULL;
 680
 681         /*
 682          * No need to protect slots here -- all the slots are "local" and
 683          * the page lock is already taken
 684          */
 685         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 686                 p += ZHDR_SIZE_ALIGNED;
 687                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 688                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 689                 moved_chunks = &zhdr->first_chunks;
 690         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 691                 p += zhdr->start_middle << CHUNK_SHIFT;
 692                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 693                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 694                 moved_chunks = &zhdr->middle_chunks;
 695         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 696                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 697                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 698                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 699                 moved_chunks = &zhdr->last_chunks;
 700         }
 701
 702         if (sz > 0) {
 703                 enum buddy new_bud = HEADLESS;
 704                 short chunks = size_to_chunks(sz);
 705                 void *q;
 706
 707                 new_zhdr = __z3fold_alloc(pool, sz, false);
 708                 if (!new_zhdr)
 709                         return NULL;
 710
 711                 if (WARN_ON(new_zhdr == zhdr))
 712                         goto out_fail;
 713
 714                 if (new_zhdr->first_chunks == 0) {
 715                         if (new_zhdr->middle_chunks != 0 &&
 716                                         chunks >= new_zhdr->start_middle) {
 717                                 new_bud = LAST;
 718                         } else {
 719                                 new_bud = FIRST;
 720                         }
 721                 } else if (new_zhdr->last_chunks == 0) {
 722                         new_bud = LAST;
 723                 } else if (new_zhdr->middle_chunks == 0) {
 724                         new_bud = MIDDLE;
 725                 }
 726                 q = new_zhdr;
 727                 switch (new_bud) {
 728                 case FIRST:
 729                         new_zhdr->first_chunks = chunks;
 730                         q += ZHDR_SIZE_ALIGNED;
 731                         break;
 732                 case MIDDLE:
 733                         new_zhdr->middle_chunks = chunks;
 734                         new_zhdr->start_middle =
 735                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 736                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 737                         break;
 738                 case LAST:
 739                         new_zhdr->last_chunks = chunks;
 740                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 741                         break;
 742                 default:
 743                         goto out_fail;
 744                 }
 745                 new_zhdr->foreign_handles++;
 746                 memcpy(q, p, sz);
 747                 write_lock(&zhdr->slots->lock);
 748                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 749                         __idx(new_zhdr, new_bud);
 750                 if (new_bud == LAST)
 751                         *(unsigned long *)old_handle |=
 752                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 753                 write_unlock(&zhdr->slots->lock);
 754                 add_to_unbuddied(pool, new_zhdr);
 755                 z3fold_page_unlock(new_zhdr);
 756
 757                 *moved_chunks = 0;
 758         }
 759
 760         return new_zhdr;
 761
 762 out_fail:
 763         if (new_zhdr) {
 764                 if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
 765                         atomic64_dec(&pool->pages_nr);
 766                 else {
 767                         add_to_unbuddied(pool, new_zhdr);
 768                         z3fold_page_unlock(new_zhdr);
 769                 }
 770         }
 771         return NULL;
 772
 773 }
 774
 775 #define BIG_CHUNK_GAP   3
 776 /* Has to be called with lock held */
 777 static int z3fold_compact_page(struct z3fold_header *zhdr)
 778 {
 779         struct page *page = virt_to_page(zhdr);
 780
 781         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 782                 return 0; /* can't move middle chunk, it's used */
 783
 784         if (unlikely(PageIsolated(page)))
 785                 return 0;
 786
 787         if (zhdr->middle_chunks == 0)
 788                 return 0; /* nothing to compact */
 789
 790         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 791                 /* move to the beginning */
 792                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 793                 zhdr->first_chunks = zhdr->middle_chunks;
 794                 zhdr->middle_chunks = 0;
 795                 zhdr->start_middle = 0;
 796                 zhdr->first_num++;
 797                 return 1;
 798         }
 799
 800         /*
 801          * moving data is expensive, so let's only do that if
 802          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 803          */
 804         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 805             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 806                         BIG_CHUNK_GAP) {
 807                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 808                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 809                 return 1;
 810         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 811                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 812                                         + zhdr->middle_chunks) >=
 813                         BIG_CHUNK_GAP) {
 814                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 815                         zhdr->middle_chunks;
 816                 mchunk_memmove(zhdr, new_start);
 817                 zhdr->start_middle = new_start;
 818                 return 1;
 819         }
 820
 821         return 0;
 822 }
 823
 824 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 825 {
 826         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 827         struct page *page;
 828
 829         page = virt_to_page(zhdr);
 830         if (locked)
 831                 WARN_ON(z3fold_page_trylock(zhdr));
 832         else
 833                 z3fold_page_lock(zhdr);
 834         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 835                 z3fold_page_unlock(zhdr);
 836                 return;
 837         }
 838         spin_lock(&pool->lock);
 839         list_del_init(&zhdr->buddy);
 840         spin_unlock(&pool->lock);
 841
 842         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 843                 atomic64_dec(&pool->pages_nr);
 844                 return;
 845         }
 846
 847         if (unlikely(PageIsolated(page) ||
 848                      test_bit(PAGE_CLAIMED, &page->private) ||
 849                      test_bit(PAGE_STALE, &page->private))) {
 850                 z3fold_page_unlock(zhdr);
 851                 return;
 852         }
 853
 854         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 855             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 856                 if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 857                         atomic64_dec(&pool->pages_nr);
 858                 else
 859                         z3fold_page_unlock(zhdr);
 860                 return;
 861         }
 862
 863         z3fold_compact_page(zhdr);
 864         add_to_unbuddied(pool, zhdr);
 865         z3fold_page_unlock(zhdr);
 866 }
 867
 868 static void compact_page_work(struct work_struct *w)
 869 {
 870         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 871                                                 work);
 872
 873         do_compact_page(zhdr, false);
 874 }
 875
 876 /* returns _locked_ z3fold page header or NULL */
 877 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 878                                                 size_t size, bool can_sleep)
 879 {
 880         struct z3fold_header *zhdr = NULL;
 881         struct page *page;
 882         struct list_head *unbuddied;
 883         int chunks = size_to_chunks(size), i;
 884
 885 lookup:
 886         /* First, try to find an unbuddied z3fold page. */
 887         unbuddied = get_cpu_ptr(pool->unbuddied);
 888         for_each_unbuddied_list(i, chunks) {
 889                 struct list_head *l = &unbuddied[i];
 890
 891                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 892                                         struct z3fold_header, buddy);
 893
 894                 if (!zhdr)
 895                         continue;
 896
 897                 /* Re-check under lock. */
 898                 spin_lock(&pool->lock);
 899                 l = &unbuddied[i];
 900                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 901                                                 struct z3fold_header, buddy)) ||
 902                     !z3fold_page_trylock(zhdr)) {
 903                         spin_unlock(&pool->lock);
 904                         zhdr = NULL;
 905                         put_cpu_ptr(pool->unbuddied);
 906                         if (can_sleep)
 907                                 cond_resched();
 908                         goto lookup;
 909                 }
 910                 list_del_init(&zhdr->buddy);
 911                 zhdr->cpu = -1;
 912                 spin_unlock(&pool->lock);
 913
 914                 page = virt_to_page(zhdr);
 915                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 916                     test_bit(PAGE_CLAIMED, &page->private)) {
 917                         z3fold_page_unlock(zhdr);
 918                         zhdr = NULL;
 919                         put_cpu_ptr(pool->unbuddied);
 920                         if (can_sleep)
 921                                 cond_resched();
 922                         goto lookup;
 923                 }
 924
 925                 /*
 926                  * this page could not be removed from its unbuddied
 927                  * list while pool lock was held, and then we've taken
 928                  * page lock so kref_put could not be called before
 929                  * we got here, so it's safe to just call kref_get()
 930                  */
 931                 kref_get(&zhdr->refcount);
 932                 break;
 933         }
 934         put_cpu_ptr(pool->unbuddied);
 935
 936         if (!zhdr) {
 937                 int cpu;
 938
 939                 /* look for _exact_ match on other cpus' lists */
 940                 for_each_online_cpu(cpu) {
 941                         struct list_head *l;
 942
 943                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 944                         spin_lock(&pool->lock);
 945                         l = &unbuddied[chunks];
 946
 947                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 948                                                 struct z3fold_header, buddy);
 949
 950                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 951                                 spin_unlock(&pool->lock);
 952                                 zhdr = NULL;
 953                                 continue;
 954                         }
 955                         list_del_init(&zhdr->buddy);
 956                         zhdr->cpu = -1;
 957                         spin_unlock(&pool->lock);
 958
 959                         page = virt_to_page(zhdr);
 960                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 961                             test_bit(PAGE_CLAIMED, &page->private)) {
 962                                 z3fold_page_unlock(zhdr);
 963                                 zhdr = NULL;
 964                                 if (can_sleep)
 965                                         cond_resched();
 966                                 continue;
 967                         }
 968                         kref_get(&zhdr->refcount);
 969                         break;
 970                 }
 971         }
 972
 973         return zhdr;
 974 }
 975
 976 /*
 977  * API Functions
 978  */
 979
 980 /**
 981  * z3fold_create_pool() - create a new z3fold pool
 982  * @name:       pool name
 983  * @gfp:        gfp flags when allocating the z3fold pool structure
 984  * @ops:        user-defined operations for the z3fold pool
 985  *
 986  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 987  * failed.
 988  */
 989 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 990                 const struct z3fold_ops *ops)
 991 {
 992         struct z3fold_pool *pool = NULL;
 993         int i, cpu;
 994
 995         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 996         if (!pool)
 997                 goto out;
 998         pool->c_handle = kmem_cache_create("z3fold_handle",
 999                                 sizeof(struct z3fold_buddy_slots),
1000                                 SLOTS_ALIGN, 0, NULL);
1001         if (!pool->c_handle)
1002                 goto out_c;
1003         spin_lock_init(&pool->lock);
1004         spin_lock_init(&pool->stale_lock);
1005         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
1006         if (!pool->unbuddied)
1007                 goto out_pool;
1008         for_each_possible_cpu(cpu) {
1009                 struct list_head *unbuddied =
1010                                 per_cpu_ptr(pool->unbuddied, cpu);
1011                 for_each_unbuddied_list(i, 0)
1012                         INIT_LIST_HEAD(&unbuddied[i]);
1013         }
1014         INIT_LIST_HEAD(&pool->lru);
1015         INIT_LIST_HEAD(&pool->stale);
1016         atomic64_set(&pool->pages_nr, 0);
1017         pool->name = name;
1018         pool->compact_wq = create_singlethread_workqueue(pool->name);
1019         if (!pool->compact_wq)
1020                 goto out_unbuddied;
1021         pool->release_wq = create_singlethread_workqueue(pool->name);
1022         if (!pool->release_wq)
1023                 goto out_wq;
1024         if (z3fold_register_migration(pool))
1025                 goto out_rwq;
1026         INIT_WORK(&pool->work, free_pages_work);
1027         pool->ops = ops;
1028         return pool;
1029
1030 out_rwq:
1031         destroy_workqueue(pool->release_wq);
1032 out_wq:
1033         destroy_workqueue(pool->compact_wq);
1034 out_unbuddied:
1035         free_percpu(pool->unbuddied);
1036 out_pool:
1037         kmem_cache_destroy(pool->c_handle);
1038 out_c:
1039         kfree(pool);
1040 out:
1041         return NULL;
1042 }
1043
1044 /**
1045  * z3fold_destroy_pool() - destroys an existing z3fold pool
1046  * @pool:       the z3fold pool to be destroyed
1047  *
1048  * The pool should be emptied before this function is called.
1049  */
1050 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1051 {
1052         kmem_cache_destroy(pool->c_handle);
1053
1054         /*
1055          * We need to destroy pool->compact_wq before pool->release_wq,
1056          * as any pending work on pool->compact_wq will call
1057          * queue_work(pool->release_wq, &pool->work).
1058          *
1059          * There are still outstanding pages until both workqueues are drained,
1060          * so we cannot unregister migration until then.
1061          */
1062
1063         destroy_workqueue(pool->compact_wq);
1064         destroy_workqueue(pool->release_wq);
1065         z3fold_unregister_migration(pool);
1066         kfree(pool);
1067 }
1068
1069 /**
1070  * z3fold_alloc() - allocates a region of a given size
1071  * @pool:       z3fold pool from which to allocate
1072  * @size:       size in bytes of the desired allocation
1073  * @gfp:        gfp flags used if the pool needs to grow
1074  * @handle:     handle of the new allocation
1075  *
1076  * This function will attempt to find a free region in the pool large enough to
1077  * satisfy the allocation request.  A search of the unbuddied lists is
1078  * performed first. If no suitable free region is found, then a new page is
1079  * allocated and added to the pool to satisfy the request.
1080  *
1081  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
1082  * as z3fold pool pages.
1083  *
1084  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1085  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1086  * a new page.
1087  */
1088 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1089                         unsigned long *handle)
1090 {
1091         int chunks = size_to_chunks(size);
1092         struct z3fold_header *zhdr = NULL;
1093         struct page *page = NULL;
1094         enum buddy bud;
1095         bool can_sleep = gfpflags_allow_blocking(gfp);
1096
1097         if (!size)
1098                 return -EINVAL;
1099
1100         if (size > PAGE_SIZE)
1101                 return -ENOSPC;
1102
1103         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1104                 bud = HEADLESS;
1105         else {
1106 retry:
1107                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1108                 if (zhdr) {
1109                         if (zhdr->first_chunks == 0) {
1110                                 if (zhdr->middle_chunks != 0 &&
1111                                     chunks >= zhdr->start_middle)
1112                                         bud = LAST;
1113                                 else
1114                                         bud = FIRST;
1115                         } else if (zhdr->last_chunks == 0)
1116                                 bud = LAST;
1117                         else if (zhdr->middle_chunks == 0)
1118                                 bud = MIDDLE;
1119                         else {
1120                                 if (kref_put(&zhdr->refcount,
1121                                              release_z3fold_page_locked))
1122                                         atomic64_dec(&pool->pages_nr);
1123                                 else
1124                                         z3fold_page_unlock(zhdr);
1125                                 pr_err("No free chunks in unbuddied\n");
1126                                 WARN_ON(1);
1127                                 goto retry;
1128                         }
1129                         page = virt_to_page(zhdr);
1130                         goto found;
1131                 }
1132                 bud = FIRST;
1133         }
1134
1135         page = NULL;
1136         if (can_sleep) {
1137                 spin_lock(&pool->stale_lock);
1138                 zhdr = list_first_entry_or_null(&pool->stale,
1139                                                 struct z3fold_header, buddy);
1140                 /*
1141                  * Before allocating a page, let's see if we can take one from
1142                  * the stale pages list. cancel_work_sync() can sleep so we
1143                  * limit this case to the contexts where we can sleep
1144                  */
1145                 if (zhdr) {
1146                         list_del(&zhdr->buddy);
1147                         spin_unlock(&pool->stale_lock);
1148                         cancel_work_sync(&zhdr->work);
1149                         page = virt_to_page(zhdr);
1150                 } else {
1151                         spin_unlock(&pool->stale_lock);
1152                 }
1153         }
1154         if (!page)
1155                 page = alloc_page(gfp);
1156
1157         if (!page)
1158                 return -ENOMEM;
1159
1160         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1161         if (!zhdr) {
1162                 __free_page(page);
1163                 return -ENOMEM;
1164         }
1165         atomic64_inc(&pool->pages_nr);
1166
1167         if (bud == HEADLESS) {
1168                 set_bit(PAGE_HEADLESS, &page->private);
1169                 goto headless;
1170         }
1171         if (can_sleep) {
1172                 lock_page(page);
1173                 __SetPageMovable(page, pool->inode->i_mapping);
1174                 unlock_page(page);
1175         } else {
1176                 if (trylock_page(page)) {
1177                         __SetPageMovable(page, pool->inode->i_mapping);
1178                         unlock_page(page);
1179                 }
1180         }
1181         z3fold_page_lock(zhdr);
1182
1183 found:
1184         if (bud == FIRST)
1185                 zhdr->first_chunks = chunks;
1186         else if (bud == LAST)
1187                 zhdr->last_chunks = chunks;
1188         else {
1189                 zhdr->middle_chunks = chunks;
1190                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1191         }
1192         add_to_unbuddied(pool, zhdr);
1193
1194 headless:
1195         spin_lock(&pool->lock);
1196         /* Add/move z3fold page to beginning of LRU */
1197         if (!list_empty(&page->lru))
1198                 list_del(&page->lru);
1199
1200         list_add(&page->lru, &pool->lru);
1201
1202         *handle = encode_handle(zhdr, bud);
1203         spin_unlock(&pool->lock);
1204         if (bud != HEADLESS)
1205                 z3fold_page_unlock(zhdr);
1206
1207         return 0;
1208 }
1209
1210 /**
1211  * z3fold_free() - frees the allocation associated with the given handle
1212  * @pool:       pool in which the allocation resided
1213  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1214  *
1215  * In the case that the z3fold page in which the allocation resides is under
1216  * reclaim, as indicated by the PG_reclaim flag being set, this function
1217  * only sets the first|last_chunks to 0.  The page is actually freed
1218  * once both buddies are evicted (see z3fold_reclaim_page() below).
1219  */
1220 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1221 {
1222         struct z3fold_header *zhdr;
1223         struct page *page;
1224         enum buddy bud;
1225         bool page_claimed;
1226
1227         zhdr = get_z3fold_header(handle);
1228         page = virt_to_page(zhdr);
1229         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1230
1231         if (test_bit(PAGE_HEADLESS, &page->private)) {
1232                 /* if a headless page is under reclaim, just leave.
1233                  * NB: we use test_and_set_bit for a reason: if the bit
1234                  * has not been set before, we release this page
1235                  * immediately so we don't care about its value any more.
1236                  */
1237                 if (!page_claimed) {
1238                         spin_lock(&pool->lock);
1239                         list_del(&page->lru);
1240                         spin_unlock(&pool->lock);
1241                         put_z3fold_header(zhdr);
1242                         free_z3fold_page(page, true);
1243                         atomic64_dec(&pool->pages_nr);
1244                 }
1245                 return;
1246         }
1247
1248         /* Non-headless case */
1249         bud = handle_to_buddy(handle);
1250
1251         switch (bud) {
1252         case FIRST:
1253                 zhdr->first_chunks = 0;
1254                 break;
1255         case MIDDLE:
1256                 zhdr->middle_chunks = 0;
1257                 break;
1258         case LAST:
1259                 zhdr->last_chunks = 0;
1260                 break;
1261         default:
1262                 pr_err("%s: unknown bud %d\n", __func__, bud);
1263                 WARN_ON(1);
1264                 put_z3fold_header(zhdr);
1265                 clear_bit(PAGE_CLAIMED, &page->private);
1266                 return;
1267         }
1268
1269         if (!page_claimed)
1270                 free_handle(handle);
1271         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1272                 atomic64_dec(&pool->pages_nr);
1273                 return;
1274         }
1275         if (page_claimed) {
1276                 /* the page has not been claimed by us */
1277                 z3fold_page_unlock(zhdr);
1278                 return;
1279         }
1280         if (unlikely(PageIsolated(page)) ||
1281             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1282                 put_z3fold_header(zhdr);
1283                 clear_bit(PAGE_CLAIMED, &page->private);
1284                 return;
1285         }
1286         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1287                 spin_lock(&pool->lock);
1288                 list_del_init(&zhdr->buddy);
1289                 spin_unlock(&pool->lock);
1290                 zhdr->cpu = -1;
1291                 kref_get(&zhdr->refcount);
1292                 clear_bit(PAGE_CLAIMED, &page->private);
1293                 do_compact_page(zhdr, true);
1294                 return;
1295         }
1296         kref_get(&zhdr->refcount);
1297         clear_bit(PAGE_CLAIMED, &page->private);
1298         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1299         put_z3fold_header(zhdr);
1300 }
1301
1302 /**
1303  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1304  * @pool:       pool from which a page will attempt to be evicted
1305  * @retries:    number of pages on the LRU list for which eviction will
1306  *              be attempted before failing
1307  *
1308  * z3fold reclaim is different from normal system reclaim in that it is done
1309  * from the bottom, up. This is because only the bottom layer, z3fold, has
1310  * information on how the allocations are organized within each z3fold page.
1311  * This has the potential to create interesting locking situations between
1312  * z3fold and the user, however.
1313  *
1314  * To avoid these, this is how z3fold_reclaim_page() should be called:
1315  *
1316  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1317  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1318  * call the user-defined eviction handler with the pool and handle as
1319  * arguments.
1320  *
1321  * If the handle can not be evicted, the eviction handler should return
1322  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1323  * appropriate list and try the next z3fold page on the LRU up to
1324  * a user defined number of retries.
1325  *
1326  * If the handle is successfully evicted, the eviction handler should
1327  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1328  * contains logic to delay freeing the page if the page is under reclaim,
1329  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1330  *
1331  * If all buddies in the z3fold page are successfully evicted, then the
1332  * z3fold page can be freed.
1333  *
1334  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1335  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1336  * the retry limit was hit.
1337  */
1338 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1339 {
1340         int i, ret = -1;
1341         struct z3fold_header *zhdr = NULL;
1342         struct page *page = NULL;
1343         struct list_head *pos;
1344         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1345
1346         spin_lock(&pool->lock);
1347         if (!pool->ops || !pool->ops->evict || retries == 0) {
1348                 spin_unlock(&pool->lock);
1349                 return -EINVAL;
1350         }
1351         for (i = 0; i < retries; i++) {
1352                 if (list_empty(&pool->lru)) {
1353                         spin_unlock(&pool->lock);
1354                         return -EINVAL;
1355                 }
1356                 list_for_each_prev(pos, &pool->lru) {
1357                         page = list_entry(pos, struct page, lru);
1358
1359                         /* this bit could have been set by free, in which case
1360                          * we pass over to the next page in the pool.
1361                          */
1362                         if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1363                                 page = NULL;
1364                                 continue;
1365                         }
1366
1367                         if (unlikely(PageIsolated(page))) {
1368                                 clear_bit(PAGE_CLAIMED, &page->private);
1369                                 page = NULL;
1370                                 continue;
1371                         }
1372                         zhdr = page_address(page);
1373                         if (test_bit(PAGE_HEADLESS, &page->private))
1374                                 break;
1375
1376                         if (!z3fold_page_trylock(zhdr)) {
1377                                 clear_bit(PAGE_CLAIMED, &page->private);
1378                                 zhdr = NULL;
1379                                 continue; /* can't evict at this point */
1380                         }
1381                         if (zhdr->foreign_handles) {
1382                                 clear_bit(PAGE_CLAIMED, &page->private);
1383                                 z3fold_page_unlock(zhdr);
1384                                 zhdr = NULL;
1385                                 continue; /* can't evict such page */
1386                         }
1387                         kref_get(&zhdr->refcount);
1388                         list_del_init(&zhdr->buddy);
1389                         zhdr->cpu = -1;
1390                         break;
1391                 }
1392
1393                 if (!zhdr)
1394                         break;
1395
1396                 list_del_init(&page->lru);
1397                 spin_unlock(&pool->lock);
1398
1399                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1400                         /*
1401                          * We need encode the handles before unlocking, and
1402                          * use our local slots structure because z3fold_free
1403                          * can zero out zhdr->slots and we can't do much
1404                          * about that
1405                          */
1406                         first_handle = 0;
1407                         last_handle = 0;
1408                         middle_handle = 0;
1409                         if (zhdr->first_chunks)
1410                                 first_handle = encode_handle(zhdr, FIRST);
1411                         if (zhdr->middle_chunks)
1412                                 middle_handle = encode_handle(zhdr, MIDDLE);
1413                         if (zhdr->last_chunks)
1414                                 last_handle = encode_handle(zhdr, LAST);
1415                         /*
1416                          * it's safe to unlock here because we hold a
1417                          * reference to this page
1418                          */
1419                         z3fold_page_unlock(zhdr);
1420                 } else {
1421                         first_handle = encode_handle(zhdr, HEADLESS);
1422                         last_handle = middle_handle = 0;
1423                 }
1424                 /* Issue the eviction callback(s) */
1425                 if (middle_handle) {
1426                         ret = pool->ops->evict(pool, middle_handle);
1427                         if (ret)
1428                                 goto next;
1429                         free_handle(middle_handle);
1430                 }
1431                 if (first_handle) {
1432                         ret = pool->ops->evict(pool, first_handle);
1433                         if (ret)
1434                                 goto next;
1435                         free_handle(first_handle);
1436                 }
1437                 if (last_handle) {
1438                         ret = pool->ops->evict(pool, last_handle);
1439                         if (ret)
1440                                 goto next;
1441                         free_handle(last_handle);
1442                 }
1443 next:
1444                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1445                         if (ret == 0) {
1446                                 free_z3fold_page(page, true);
1447                                 atomic64_dec(&pool->pages_nr);
1448                                 return 0;
1449                         }
1450                         spin_lock(&pool->lock);
1451                         list_add(&page->lru, &pool->lru);
1452                         spin_unlock(&pool->lock);
1453                         clear_bit(PAGE_CLAIMED, &page->private);
1454                 } else {
1455                         z3fold_page_lock(zhdr);
1456                         if (kref_put(&zhdr->refcount,
1457                                         release_z3fold_page_locked)) {
1458                                 atomic64_dec(&pool->pages_nr);
1459                                 return 0;
1460                         }
1461                         /*
1462                          * if we are here, the page is still not completely
1463                          * free. Take the global pool lock then to be able
1464                          * to add it back to the lru list
1465                          */
1466                         spin_lock(&pool->lock);
1467                         list_add(&page->lru, &pool->lru);
1468                         spin_unlock(&pool->lock);
1469                         z3fold_page_unlock(zhdr);
1470                         clear_bit(PAGE_CLAIMED, &page->private);
1471                 }
1472
1473                 /* We started off locked to we need to lock the pool back */
1474                 spin_lock(&pool->lock);
1475         }
1476         spin_unlock(&pool->lock);
1477         return -EAGAIN;
1478 }
1479
1480 /**
1481  * z3fold_map() - maps the allocation associated with the given handle
1482  * @pool:       pool in which the allocation resides
1483  * @handle:     handle associated with the allocation to be mapped
1484  *
1485  * Extracts the buddy number from handle and constructs the pointer to the
1486  * correct starting chunk within the page.
1487  *
1488  * Returns: a pointer to the mapped allocation
1489  */
1490 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1491 {
1492         struct z3fold_header *zhdr;
1493         struct page *page;
1494         void *addr;
1495         enum buddy buddy;
1496
1497         zhdr = get_z3fold_header(handle);
1498         addr = zhdr;
1499         page = virt_to_page(zhdr);
1500
1501         if (test_bit(PAGE_HEADLESS, &page->private))
1502                 goto out;
1503
1504         buddy = handle_to_buddy(handle);
1505         switch (buddy) {
1506         case FIRST:
1507                 addr += ZHDR_SIZE_ALIGNED;
1508                 break;
1509         case MIDDLE:
1510                 addr += zhdr->start_middle << CHUNK_SHIFT;
1511                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1512                 break;
1513         case LAST:
1514                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1515                 break;
1516         default:
1517                 pr_err("unknown buddy id %d\n", buddy);
1518                 WARN_ON(1);
1519                 addr = NULL;
1520                 break;
1521         }
1522
1523         if (addr)
1524                 zhdr->mapped_count++;
1525 out:
1526         put_z3fold_header(zhdr);
1527         return addr;
1528 }
1529
1530 /**
1531  * z3fold_unmap() - unmaps the allocation associated with the given handle
1532  * @pool:       pool in which the allocation resides
1533  * @handle:     handle associated with the allocation to be unmapped
1534  */
1535 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1536 {
1537         struct z3fold_header *zhdr;
1538         struct page *page;
1539         enum buddy buddy;
1540
1541         zhdr = get_z3fold_header(handle);
1542         page = virt_to_page(zhdr);
1543
1544         if (test_bit(PAGE_HEADLESS, &page->private))
1545                 return;
1546
1547         buddy = handle_to_buddy(handle);
1548         if (buddy == MIDDLE)
1549                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1550         zhdr->mapped_count--;
1551         put_z3fold_header(zhdr);
1552 }
1553
1554 /**
1555  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1556  * @pool:       pool whose size is being queried
1557  *
1558  * Returns: size in pages of the given pool.
1559  */
1560 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1561 {
1562         return atomic64_read(&pool->pages_nr);
1563 }
1564
1565 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1566 {
1567         struct z3fold_header *zhdr;
1568         struct z3fold_pool *pool;
1569
1570         VM_BUG_ON_PAGE(!PageMovable(page), page);
1571         VM_BUG_ON_PAGE(PageIsolated(page), page);
1572
1573         if (test_bit(PAGE_HEADLESS, &page->private) ||
1574             test_bit(PAGE_CLAIMED, &page->private))
1575                 return false;
1576
1577         zhdr = page_address(page);
1578         z3fold_page_lock(zhdr);
1579         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1580             test_bit(PAGE_STALE, &page->private))
1581                 goto out;
1582
1583         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1584                 goto out;
1585
1586         pool = zhdr_to_pool(zhdr);
1587         spin_lock(&pool->lock);
1588         if (!list_empty(&zhdr->buddy))
1589                 list_del_init(&zhdr->buddy);
1590         if (!list_empty(&page->lru))
1591                 list_del_init(&page->lru);
1592         spin_unlock(&pool->lock);
1593
1594         kref_get(&zhdr->refcount);
1595         z3fold_page_unlock(zhdr);
1596         return true;
1597
1598 out:
1599         z3fold_page_unlock(zhdr);
1600         return false;
1601 }
1602
1603 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1604                                struct page *page, enum migrate_mode mode)
1605 {
1606         struct z3fold_header *zhdr, *new_zhdr;
1607         struct z3fold_pool *pool;
1608         struct address_space *new_mapping;
1609
1610         VM_BUG_ON_PAGE(!PageMovable(page), page);
1611         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1612         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1613
1614         zhdr = page_address(page);
1615         pool = zhdr_to_pool(zhdr);
1616
1617         if (!z3fold_page_trylock(zhdr)) {
1618                 return -EAGAIN;
1619         }
1620         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1621                 z3fold_page_unlock(zhdr);
1622                 return -EBUSY;
1623         }
1624         if (work_pending(&zhdr->work)) {
1625                 z3fold_page_unlock(zhdr);
1626                 return -EAGAIN;
1627         }
1628         new_zhdr = page_address(newpage);
1629         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1630         newpage->private = page->private;
1631         page->private = 0;
1632         z3fold_page_unlock(zhdr);
1633         spin_lock_init(&new_zhdr->page_lock);
1634         INIT_WORK(&new_zhdr->work, compact_page_work);
1635         /*
1636          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1637          * so we only have to reinitialize it.
1638          */
1639         INIT_LIST_HEAD(&new_zhdr->buddy);
1640         new_mapping = page_mapping(page);
1641         __ClearPageMovable(page);
1642         ClearPagePrivate(page);
1643
1644         get_page(newpage);
1645         z3fold_page_lock(new_zhdr);
1646         if (new_zhdr->first_chunks)
1647                 encode_handle(new_zhdr, FIRST);
1648         if (new_zhdr->last_chunks)
1649                 encode_handle(new_zhdr, LAST);
1650         if (new_zhdr->middle_chunks)
1651                 encode_handle(new_zhdr, MIDDLE);
1652         set_bit(NEEDS_COMPACTING, &newpage->private);
1653         new_zhdr->cpu = smp_processor_id();
1654         spin_lock(&pool->lock);
1655         list_add(&newpage->lru, &pool->lru);
1656         spin_unlock(&pool->lock);
1657         __SetPageMovable(newpage, new_mapping);
1658         z3fold_page_unlock(new_zhdr);
1659
1660         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1661
1662         page_mapcount_reset(page);
1663         put_page(page);
1664         return 0;
1665 }
1666
1667 static void z3fold_page_putback(struct page *page)
1668 {
1669         struct z3fold_header *zhdr;
1670         struct z3fold_pool *pool;
1671
1672         zhdr = page_address(page);
1673         pool = zhdr_to_pool(zhdr);
1674
1675         z3fold_page_lock(zhdr);
1676         if (!list_empty(&zhdr->buddy))
1677                 list_del_init(&zhdr->buddy);
1678         INIT_LIST_HEAD(&page->lru);
1679         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1680                 atomic64_dec(&pool->pages_nr);
1681                 return;
1682         }
1683         spin_lock(&pool->lock);
1684         list_add(&page->lru, &pool->lru);
1685         spin_unlock(&pool->lock);
1686         z3fold_page_unlock(zhdr);
1687 }
1688
1689 static const struct address_space_operations z3fold_aops = {
1690         .isolate_page = z3fold_page_isolate,
1691         .migratepage = z3fold_page_migrate,
1692         .putback_page = z3fold_page_putback,
1693 };
1694
1695 /*****************
1696  * zpool
1697  ****************/
1698
1699 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1700 {
1701         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1702                 return pool->zpool_ops->evict(pool->zpool, handle);
1703         else
1704                 return -ENOENT;
1705 }
1706
1707 static const struct z3fold_ops z3fold_zpool_ops = {
1708         .evict =        z3fold_zpool_evict
1709 };
1710
1711 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1712                                const struct zpool_ops *zpool_ops,
1713                                struct zpool *zpool)
1714 {
1715         struct z3fold_pool *pool;
1716
1717         pool = z3fold_create_pool(name, gfp,
1718                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1719         if (pool) {
1720                 pool->zpool = zpool;
1721                 pool->zpool_ops = zpool_ops;
1722         }
1723         return pool;
1724 }
1725
1726 static void z3fold_zpool_destroy(void *pool)
1727 {
1728         z3fold_destroy_pool(pool);
1729 }
1730
1731 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1732                         unsigned long *handle)
1733 {
1734         return z3fold_alloc(pool, size, gfp, handle);
1735 }
1736 static void z3fold_zpool_free(void *pool, unsigned long handle)
1737 {
1738         z3fold_free(pool, handle);
1739 }
1740
1741 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1742                         unsigned int *reclaimed)
1743 {
1744         unsigned int total = 0;
1745         int ret = -EINVAL;
1746
1747         while (total < pages) {
1748                 ret = z3fold_reclaim_page(pool, 8);
1749                 if (ret < 0)
1750                         break;
1751                 total++;
1752         }
1753
1754         if (reclaimed)
1755                 *reclaimed = total;
1756
1757         return ret;
1758 }
1759
1760 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1761                         enum zpool_mapmode mm)
1762 {
1763         return z3fold_map(pool, handle);
1764 }
1765 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1766 {
1767         z3fold_unmap(pool, handle);
1768 }
1769
1770 static u64 z3fold_zpool_total_size(void *pool)
1771 {
1772         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1773 }
1774
1775 static struct zpool_driver z3fold_zpool_driver = {
1776         .type =         "z3fold",
1777         .owner =        THIS_MODULE,
1778         .create =       z3fold_zpool_create,
1779         .destroy =      z3fold_zpool_destroy,
1780         .malloc =       z3fold_zpool_malloc,
1781         .free =         z3fold_zpool_free,
1782         .shrink =       z3fold_zpool_shrink,
1783         .map =          z3fold_zpool_map,
1784         .unmap =        z3fold_zpool_unmap,
1785         .total_size =   z3fold_zpool_total_size,
1786 };
1787
1788 MODULE_ALIAS("zpool-z3fold");
1789
1790 static int __init init_z3fold(void)
1791 {
1792         int ret;
1793
1794         /* Make sure the z3fold header is not larger than the page size */
1795         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1796         ret = z3fold_mount();
1797         if (ret)
1798                 return ret;
1799
1800         zpool_register_driver(&z3fold_zpool_driver);
1801
1802         return 0;
1803 }
1804
1805 static void __exit exit_z3fold(void)
1806 {
1807         z3fold_unmount();
1808         zpool_unregister_driver(&z3fold_zpool_driver);
1809 }
1810
1811 module_init(init_z3fold);
1812 module_exit(exit_z3fold);
1813
1814 MODULE_LICENSE("GPL");
1815 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1816 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");