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/preempt.h>
  38 #include <linux/workqueue.h>
  39 #include <linux/slab.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/zpool.h>
  42 #include <linux/kmemleak.h>
  43
  44 /*
  45  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  46  * adjusting internal fragmentation.  It also determines the number of
  47  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  48  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  49  * in the beginning of an allocated page are occupied by z3fold header, so
  50  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  51  * which shows the max number of free chunks in z3fold page, also there will
  52  * be 63, or 62, respectively, freelists per pool.
  53  */
  54 #define NCHUNKS_ORDER   6
  55
  56 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  57 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  58 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  59 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  60 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  61 #define NCHUNKS         (TOTAL_CHUNKS - ZHDR_CHUNKS)
  62
  63 #define BUDDY_MASK      (0x3)
  64 #define BUDDY_SHIFT     2
  65 #define SLOTS_ALIGN     (0x40)
  66
  67 /*****************
  68  * Structures
  69 *****************/
  70 struct z3fold_pool;
  71 struct z3fold_ops {
  72         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  73 };
  74
  75 enum buddy {
  76         HEADLESS = 0,
  77         FIRST,
  78         MIDDLE,
  79         LAST,
  80         BUDDIES_MAX = LAST
  81 };
  82
  83 struct z3fold_buddy_slots {
  84         /*
  85          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  86          * be enough slots to hold all possible variants
  87          */
  88         unsigned long slot[BUDDY_MASK + 1];
  89         unsigned long pool; /* back link */
  90         rwlock_t lock;
  91 };
  92 #define HANDLE_FLAG_MASK        (0x03)
  93
  94 /*
  95  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  96  *                      z3fold page, except for HEADLESS pages
  97  * @buddy:              links the z3fold page into the relevant list in the
  98  *                      pool
  99  * @page_lock:          per-page lock
 100  * @refcount:           reference count for the z3fold page
 101  * @work:               work_struct for page layout optimization
 102  * @slots:              pointer to the structure holding buddy slots
 103  * @pool:               pointer to the containing pool
 104  * @cpu:                CPU which this page "belongs" to
 105  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 106  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 107  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 108  * @first_num:          the starting number (for the first handle)
 109  * @mapped_count:       the number of objects currently mapped
 110  */
 111 struct z3fold_header {
 112         struct list_head buddy;
 113         spinlock_t page_lock;
 114         struct kref refcount;
 115         struct work_struct work;
 116         struct z3fold_buddy_slots *slots;
 117         struct z3fold_pool *pool;
 118         short cpu;
 119         unsigned short first_chunks;
 120         unsigned short middle_chunks;
 121         unsigned short last_chunks;
 122         unsigned short start_middle;
 123         unsigned short first_num:2;
 124         unsigned short mapped_count:2;
 125         unsigned short foreign_handles:2;
 126 };
 127
 128 /**
 129  * struct z3fold_pool - stores metadata for each z3fold pool
 130  * @name:       pool name
 131  * @lock:       protects pool unbuddied/lru lists
 132  * @stale_lock: protects pool stale page list
 133  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 134  *              buddies; the list each z3fold page is added to depends on
 135  *              the size of its free region.
 136  * @lru:        list tracking the z3fold pages in LRU order by most recently
 137  *              added buddy.
 138  * @stale:      list of pages marked for freeing
 139  * @pages_nr:   number of z3fold pages in the pool.
 140  * @c_handle:   cache for z3fold_buddy_slots allocation
 141  * @ops:        pointer to a structure of user defined operations specified at
 142  *              pool creation time.
 143  * @zpool:      zpool driver
 144  * @zpool_ops:  zpool operations structure with an evict callback
 145  * @compact_wq: workqueue for page layout background optimization
 146  * @release_wq: workqueue for safe page release
 147  * @work:       work_struct for safe page release
 148  *
 149  * This structure is allocated at pool creation time and maintains metadata
 150  * pertaining to a particular z3fold pool.
 151  */
 152 struct z3fold_pool {
 153         const char *name;
 154         spinlock_t lock;
 155         spinlock_t stale_lock;
 156         struct list_head *unbuddied;
 157         struct list_head lru;
 158         struct list_head stale;
 159         atomic64_t pages_nr;
 160         struct kmem_cache *c_handle;
 161         const struct z3fold_ops *ops;
 162         struct zpool *zpool;
 163         const struct zpool_ops *zpool_ops;
 164         struct workqueue_struct *compact_wq;
 165         struct workqueue_struct *release_wq;
 166         struct work_struct work;
 167 };
 168
 169 /*
 170  * Internal z3fold page flags
 171  */
 172 enum z3fold_page_flags {
 173         PAGE_HEADLESS = 0,
 174         MIDDLE_CHUNK_MAPPED,
 175         NEEDS_COMPACTING,
 176         PAGE_STALE,
 177         PAGE_CLAIMED, /* by either reclaim or free */
 178         PAGE_MIGRATED, /* page is migrated and soon to be released */
 179 };
 180
 181 /*
 182  * handle flags, go under HANDLE_FLAG_MASK
 183  */
 184 enum z3fold_handle_flags {
 185         HANDLES_NOFREE = 0,
 186 };
 187
 188 /*
 189  * Forward declarations
 190  */
 191 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 192 static void compact_page_work(struct work_struct *w);
 193
 194 /*****************
 195  * Helpers
 196 *****************/
 197
 198 /* Converts an allocation size in bytes to size in z3fold chunks */
 199 static int size_to_chunks(size_t size)
 200 {
 201         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 202 }
 203
 204 #define for_each_unbuddied_list(_iter, _begin) \
 205         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 206
 207 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 208                                                         gfp_t gfp)
 209 {
 210         struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
 211                                                              gfp);
 212
 213         if (slots) {
 214                 /* It will be freed separately in free_handle(). */
 215                 kmemleak_not_leak(slots);
 216                 slots->pool = (unsigned long)pool;
 217                 rwlock_init(&slots->lock);
 218         }
 219
 220         return slots;
 221 }
 222
 223 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 224 {
 225         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 226 }
 227
 228 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 229 {
 230         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 231 }
 232
 233 /* Lock a z3fold page */
 234 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 235 {
 236         spin_lock(&zhdr->page_lock);
 237 }
 238
 239 /* Try to lock a z3fold page */
 240 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 241 {
 242         return spin_trylock(&zhdr->page_lock);
 243 }
 244
 245 /* Unlock a z3fold page */
 246 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 247 {
 248         spin_unlock(&zhdr->page_lock);
 249 }
 250
 251 /* return locked z3fold page if it's not headless */
 252 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 253 {
 254         struct z3fold_buddy_slots *slots;
 255         struct z3fold_header *zhdr;
 256         int locked = 0;
 257
 258         if (!(handle & (1 << PAGE_HEADLESS))) {
 259                 slots = handle_to_slots(handle);
 260                 do {
 261                         unsigned long addr;
 262
 263                         read_lock(&slots->lock);
 264                         addr = *(unsigned long *)handle;
 265                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 266                         locked = z3fold_page_trylock(zhdr);
 267                         read_unlock(&slots->lock);
 268                         if (locked) {
 269                                 struct page *page = virt_to_page(zhdr);
 270
 271                                 if (!test_bit(PAGE_MIGRATED, &page->private))
 272                                         break;
 273                                 z3fold_page_unlock(zhdr);
 274                         }
 275                         cpu_relax();
 276                 } while (true);
 277         } else {
 278                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 279         }
 280
 281         return zhdr;
 282 }
 283
 284 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 285 {
 286         struct page *page = virt_to_page(zhdr);
 287
 288         if (!test_bit(PAGE_HEADLESS, &page->private))
 289                 z3fold_page_unlock(zhdr);
 290 }
 291
 292 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 293 {
 294         struct z3fold_buddy_slots *slots;
 295         int i;
 296         bool is_free;
 297
 298         if (WARN_ON(*(unsigned long *)handle == 0))
 299                 return;
 300
 301         slots = handle_to_slots(handle);
 302         write_lock(&slots->lock);
 303         *(unsigned long *)handle = 0;
 304
 305         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 306                 write_unlock(&slots->lock);
 307                 return; /* simple case, nothing else to do */
 308         }
 309
 310         if (zhdr->slots != slots)
 311                 zhdr->foreign_handles--;
 312
 313         is_free = true;
 314         for (i = 0; i <= BUDDY_MASK; i++) {
 315                 if (slots->slot[i]) {
 316                         is_free = false;
 317                         break;
 318                 }
 319         }
 320         write_unlock(&slots->lock);
 321
 322         if (is_free) {
 323                 struct z3fold_pool *pool = slots_to_pool(slots);
 324
 325                 if (zhdr->slots == slots)
 326                         zhdr->slots = NULL;
 327                 kmem_cache_free(pool->c_handle, slots);
 328         }
 329 }
 330
 331 /* Initializes the z3fold header of a newly allocated z3fold page */
 332 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 333                                         struct z3fold_pool *pool, gfp_t gfp)
 334 {
 335         struct z3fold_header *zhdr = page_address(page);
 336         struct z3fold_buddy_slots *slots;
 337
 338         INIT_LIST_HEAD(&page->lru);
 339         clear_bit(PAGE_HEADLESS, &page->private);
 340         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 341         clear_bit(NEEDS_COMPACTING, &page->private);
 342         clear_bit(PAGE_STALE, &page->private);
 343         clear_bit(PAGE_CLAIMED, &page->private);
 344         clear_bit(PAGE_MIGRATED, &page->private);
 345         if (headless)
 346                 return zhdr;
 347
 348         slots = alloc_slots(pool, gfp);
 349         if (!slots)
 350                 return NULL;
 351
 352         memset(zhdr, 0, sizeof(*zhdr));
 353         spin_lock_init(&zhdr->page_lock);
 354         kref_init(&zhdr->refcount);
 355         zhdr->cpu = -1;
 356         zhdr->slots = slots;
 357         zhdr->pool = pool;
 358         INIT_LIST_HEAD(&zhdr->buddy);
 359         INIT_WORK(&zhdr->work, compact_page_work);
 360         return zhdr;
 361 }
 362
 363 /* Resets the struct page fields and frees the page */
 364 static void free_z3fold_page(struct page *page, bool headless)
 365 {
 366         if (!headless) {
 367                 lock_page(page);
 368                 __ClearPageMovable(page);
 369                 unlock_page(page);
 370         }
 371         __free_page(page);
 372 }
 373
 374 /* Helper function to build the index */
 375 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 376 {
 377         return (bud + zhdr->first_num) & BUDDY_MASK;
 378 }
 379
 380 /*
 381  * Encodes the handle of a particular buddy within a z3fold page
 382  * Pool lock should be held as this function accesses first_num
 383  */
 384 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 385                                 struct z3fold_buddy_slots *slots,
 386                                 enum buddy bud)
 387 {
 388         unsigned long h = (unsigned long)zhdr;
 389         int idx = 0;
 390
 391         /*
 392          * For a headless page, its handle is its pointer with the extra
 393          * PAGE_HEADLESS bit set
 394          */
 395         if (bud == HEADLESS)
 396                 return h | (1 << PAGE_HEADLESS);
 397
 398         /* otherwise, return pointer to encoded handle */
 399         idx = __idx(zhdr, bud);
 400         h += idx;
 401         if (bud == LAST)
 402                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 403
 404         write_lock(&slots->lock);
 405         slots->slot[idx] = h;
 406         write_unlock(&slots->lock);
 407         return (unsigned long)&slots->slot[idx];
 408 }
 409
 410 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 411 {
 412         return __encode_handle(zhdr, zhdr->slots, bud);
 413 }
 414
 415 /* only for LAST bud, returns zero otherwise */
 416 static unsigned short handle_to_chunks(unsigned long handle)
 417 {
 418         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 419         unsigned long addr;
 420
 421         read_lock(&slots->lock);
 422         addr = *(unsigned long *)handle;
 423         read_unlock(&slots->lock);
 424         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 425 }
 426
 427 /*
 428  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 429  *  but that doesn't matter. because the masking will result in the
 430  *  correct buddy number.
 431  */
 432 static enum buddy handle_to_buddy(unsigned long handle)
 433 {
 434         struct z3fold_header *zhdr;
 435         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 436         unsigned long addr;
 437
 438         read_lock(&slots->lock);
 439         WARN_ON(handle & (1 << PAGE_HEADLESS));
 440         addr = *(unsigned long *)handle;
 441         read_unlock(&slots->lock);
 442         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 443         return (addr - zhdr->first_num) & BUDDY_MASK;
 444 }
 445
 446 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 447 {
 448         return zhdr->pool;
 449 }
 450
 451 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 452 {
 453         struct page *page = virt_to_page(zhdr);
 454         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 455
 456         WARN_ON(!list_empty(&zhdr->buddy));
 457         set_bit(PAGE_STALE, &page->private);
 458         clear_bit(NEEDS_COMPACTING, &page->private);
 459         spin_lock(&pool->lock);
 460         if (!list_empty(&page->lru))
 461                 list_del_init(&page->lru);
 462         spin_unlock(&pool->lock);
 463
 464         if (locked)
 465                 z3fold_page_unlock(zhdr);
 466
 467         spin_lock(&pool->stale_lock);
 468         list_add(&zhdr->buddy, &pool->stale);
 469         queue_work(pool->release_wq, &pool->work);
 470         spin_unlock(&pool->stale_lock);
 471
 472         atomic64_dec(&pool->pages_nr);
 473 }
 474
 475 static void release_z3fold_page_locked(struct kref *ref)
 476 {
 477         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 478                                                 refcount);
 479         WARN_ON(z3fold_page_trylock(zhdr));
 480         __release_z3fold_page(zhdr, true);
 481 }
 482
 483 static void release_z3fold_page_locked_list(struct kref *ref)
 484 {
 485         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 486                                                refcount);
 487         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 488
 489         spin_lock(&pool->lock);
 490         list_del_init(&zhdr->buddy);
 491         spin_unlock(&pool->lock);
 492
 493         WARN_ON(z3fold_page_trylock(zhdr));
 494         __release_z3fold_page(zhdr, true);
 495 }
 496
 497 static void free_pages_work(struct work_struct *w)
 498 {
 499         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 500
 501         spin_lock(&pool->stale_lock);
 502         while (!list_empty(&pool->stale)) {
 503                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 504                                                 struct z3fold_header, buddy);
 505                 struct page *page = virt_to_page(zhdr);
 506
 507                 list_del(&zhdr->buddy);
 508                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 509                         continue;
 510                 spin_unlock(&pool->stale_lock);
 511                 cancel_work_sync(&zhdr->work);
 512                 free_z3fold_page(page, false);
 513                 cond_resched();
 514                 spin_lock(&pool->stale_lock);
 515         }
 516         spin_unlock(&pool->stale_lock);
 517 }
 518
 519 /*
 520  * Returns the number of free chunks in a z3fold page.
 521  * NB: can't be used with HEADLESS pages.
 522  */
 523 static int num_free_chunks(struct z3fold_header *zhdr)
 524 {
 525         int nfree;
 526         /*
 527          * If there is a middle object, pick up the bigger free space
 528          * either before or after it. Otherwise just subtract the number
 529          * of chunks occupied by the first and the last objects.
 530          */
 531         if (zhdr->middle_chunks != 0) {
 532                 int nfree_before = zhdr->first_chunks ?
 533                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 534                 int nfree_after = zhdr->last_chunks ?
 535                         0 : TOTAL_CHUNKS -
 536                                 (zhdr->start_middle + zhdr->middle_chunks);
 537                 nfree = max(nfree_before, nfree_after);
 538         } else
 539                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 540         return nfree;
 541 }
 542
 543 /* Add to the appropriate unbuddied list */
 544 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 545                                 struct z3fold_header *zhdr)
 546 {
 547         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 548                         zhdr->middle_chunks == 0) {
 549                 struct list_head *unbuddied;
 550                 int freechunks = num_free_chunks(zhdr);
 551
 552                 migrate_disable();
 553                 unbuddied = this_cpu_ptr(pool->unbuddied);
 554                 spin_lock(&pool->lock);
 555                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 556                 spin_unlock(&pool->lock);
 557                 zhdr->cpu = smp_processor_id();
 558                 migrate_enable();
 559         }
 560 }
 561
 562 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 563 {
 564         enum buddy bud = HEADLESS;
 565
 566         if (zhdr->middle_chunks) {
 567                 if (!zhdr->first_chunks &&
 568                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 569                         bud = FIRST;
 570                 else if (!zhdr->last_chunks)
 571                         bud = LAST;
 572         } else {
 573                 if (!zhdr->first_chunks)
 574                         bud = FIRST;
 575                 else if (!zhdr->last_chunks)
 576                         bud = LAST;
 577                 else
 578                         bud = MIDDLE;
 579         }
 580
 581         return bud;
 582 }
 583
 584 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 585                                 unsigned short dst_chunk)
 586 {
 587         void *beg = zhdr;
 588         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 589                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 590                        zhdr->middle_chunks << CHUNK_SHIFT);
 591 }
 592
 593 static inline bool buddy_single(struct z3fold_header *zhdr)
 594 {
 595         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 596                         (zhdr->first_chunks && zhdr->last_chunks) ||
 597                         (zhdr->middle_chunks && zhdr->last_chunks));
 598 }
 599
 600 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 601 {
 602         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 603         void *p = zhdr;
 604         unsigned long old_handle = 0;
 605         size_t sz = 0;
 606         struct z3fold_header *new_zhdr = NULL;
 607         int first_idx = __idx(zhdr, FIRST);
 608         int middle_idx = __idx(zhdr, MIDDLE);
 609         int last_idx = __idx(zhdr, LAST);
 610         unsigned short *moved_chunks = NULL;
 611
 612         /*
 613          * No need to protect slots here -- all the slots are "local" and
 614          * the page lock is already taken
 615          */
 616         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 617                 p += ZHDR_SIZE_ALIGNED;
 618                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 619                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 620                 moved_chunks = &zhdr->first_chunks;
 621         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 622                 p += zhdr->start_middle << CHUNK_SHIFT;
 623                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 624                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 625                 moved_chunks = &zhdr->middle_chunks;
 626         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 627                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 628                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 629                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 630                 moved_chunks = &zhdr->last_chunks;
 631         }
 632
 633         if (sz > 0) {
 634                 enum buddy new_bud = HEADLESS;
 635                 short chunks = size_to_chunks(sz);
 636                 void *q;
 637
 638                 new_zhdr = __z3fold_alloc(pool, sz, false);
 639                 if (!new_zhdr)
 640                         return NULL;
 641
 642                 if (WARN_ON(new_zhdr == zhdr))
 643                         goto out_fail;
 644
 645                 new_bud = get_free_buddy(new_zhdr, chunks);
 646                 q = new_zhdr;
 647                 switch (new_bud) {
 648                 case FIRST:
 649                         new_zhdr->first_chunks = chunks;
 650                         q += ZHDR_SIZE_ALIGNED;
 651                         break;
 652                 case MIDDLE:
 653                         new_zhdr->middle_chunks = chunks;
 654                         new_zhdr->start_middle =
 655                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 656                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 657                         break;
 658                 case LAST:
 659                         new_zhdr->last_chunks = chunks;
 660                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 661                         break;
 662                 default:
 663                         goto out_fail;
 664                 }
 665                 new_zhdr->foreign_handles++;
 666                 memcpy(q, p, sz);
 667                 write_lock(&zhdr->slots->lock);
 668                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 669                         __idx(new_zhdr, new_bud);
 670                 if (new_bud == LAST)
 671                         *(unsigned long *)old_handle |=
 672                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 673                 write_unlock(&zhdr->slots->lock);
 674                 add_to_unbuddied(pool, new_zhdr);
 675                 z3fold_page_unlock(new_zhdr);
 676
 677                 *moved_chunks = 0;
 678         }
 679
 680         return new_zhdr;
 681
 682 out_fail:
 683         if (new_zhdr && !kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) {
 684                 add_to_unbuddied(pool, new_zhdr);
 685                 z3fold_page_unlock(new_zhdr);
 686         }
 687         return NULL;
 688
 689 }
 690
 691 #define BIG_CHUNK_GAP   3
 692 /* Has to be called with lock held */
 693 static int z3fold_compact_page(struct z3fold_header *zhdr)
 694 {
 695         struct page *page = virt_to_page(zhdr);
 696
 697         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 698                 return 0; /* can't move middle chunk, it's used */
 699
 700         if (unlikely(PageIsolated(page)))
 701                 return 0;
 702
 703         if (zhdr->middle_chunks == 0)
 704                 return 0; /* nothing to compact */
 705
 706         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 707                 /* move to the beginning */
 708                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 709                 zhdr->first_chunks = zhdr->middle_chunks;
 710                 zhdr->middle_chunks = 0;
 711                 zhdr->start_middle = 0;
 712                 zhdr->first_num++;
 713                 return 1;
 714         }
 715
 716         /*
 717          * moving data is expensive, so let's only do that if
 718          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 719          */
 720         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 721             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 722                         BIG_CHUNK_GAP) {
 723                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 724                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 725                 return 1;
 726         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 727                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 728                                         + zhdr->middle_chunks) >=
 729                         BIG_CHUNK_GAP) {
 730                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 731                         zhdr->middle_chunks;
 732                 mchunk_memmove(zhdr, new_start);
 733                 zhdr->start_middle = new_start;
 734                 return 1;
 735         }
 736
 737         return 0;
 738 }
 739
 740 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 741 {
 742         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 743         struct page *page;
 744
 745         page = virt_to_page(zhdr);
 746         if (locked)
 747                 WARN_ON(z3fold_page_trylock(zhdr));
 748         else
 749                 z3fold_page_lock(zhdr);
 750         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 751                 z3fold_page_unlock(zhdr);
 752                 return;
 753         }
 754         spin_lock(&pool->lock);
 755         list_del_init(&zhdr->buddy);
 756         spin_unlock(&pool->lock);
 757
 758         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 759                 return;
 760
 761         if (test_bit(PAGE_STALE, &page->private) ||
 762             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 763                 z3fold_page_unlock(zhdr);
 764                 return;
 765         }
 766
 767         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 768             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 769                 if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 770                         clear_bit(PAGE_CLAIMED, &page->private);
 771                         z3fold_page_unlock(zhdr);
 772                 }
 773                 return;
 774         }
 775
 776         z3fold_compact_page(zhdr);
 777         add_to_unbuddied(pool, zhdr);
 778         clear_bit(PAGE_CLAIMED, &page->private);
 779         z3fold_page_unlock(zhdr);
 780 }
 781
 782 static void compact_page_work(struct work_struct *w)
 783 {
 784         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 785                                                 work);
 786
 787         do_compact_page(zhdr, false);
 788 }
 789
 790 /* returns _locked_ z3fold page header or NULL */
 791 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 792                                                 size_t size, bool can_sleep)
 793 {
 794         struct z3fold_header *zhdr = NULL;
 795         struct page *page;
 796         struct list_head *unbuddied;
 797         int chunks = size_to_chunks(size), i;
 798
 799 lookup:
 800         migrate_disable();
 801         /* First, try to find an unbuddied z3fold page. */
 802         unbuddied = this_cpu_ptr(pool->unbuddied);
 803         for_each_unbuddied_list(i, chunks) {
 804                 struct list_head *l = &unbuddied[i];
 805
 806                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 807                                         struct z3fold_header, buddy);
 808
 809                 if (!zhdr)
 810                         continue;
 811
 812                 /* Re-check under lock. */
 813                 spin_lock(&pool->lock);
 814                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 815                                                 struct z3fold_header, buddy)) ||
 816                     !z3fold_page_trylock(zhdr)) {
 817                         spin_unlock(&pool->lock);
 818                         zhdr = NULL;
 819                         migrate_enable();
 820                         if (can_sleep)
 821                                 cond_resched();
 822                         goto lookup;
 823                 }
 824                 list_del_init(&zhdr->buddy);
 825                 zhdr->cpu = -1;
 826                 spin_unlock(&pool->lock);
 827
 828                 page = virt_to_page(zhdr);
 829                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 830                     test_bit(PAGE_CLAIMED, &page->private)) {
 831                         z3fold_page_unlock(zhdr);
 832                         zhdr = NULL;
 833                         migrate_enable();
 834                         if (can_sleep)
 835                                 cond_resched();
 836                         goto lookup;
 837                 }
 838
 839                 /*
 840                  * this page could not be removed from its unbuddied
 841                  * list while pool lock was held, and then we've taken
 842                  * page lock so kref_put could not be called before
 843                  * we got here, so it's safe to just call kref_get()
 844                  */
 845                 kref_get(&zhdr->refcount);
 846                 break;
 847         }
 848         migrate_enable();
 849
 850         if (!zhdr) {
 851                 int cpu;
 852
 853                 /* look for _exact_ match on other cpus' lists */
 854                 for_each_online_cpu(cpu) {
 855                         struct list_head *l;
 856
 857                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 858                         spin_lock(&pool->lock);
 859                         l = &unbuddied[chunks];
 860
 861                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 862                                                 struct z3fold_header, buddy);
 863
 864                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 865                                 spin_unlock(&pool->lock);
 866                                 zhdr = NULL;
 867                                 continue;
 868                         }
 869                         list_del_init(&zhdr->buddy);
 870                         zhdr->cpu = -1;
 871                         spin_unlock(&pool->lock);
 872
 873                         page = virt_to_page(zhdr);
 874                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 875                             test_bit(PAGE_CLAIMED, &page->private)) {
 876                                 z3fold_page_unlock(zhdr);
 877                                 zhdr = NULL;
 878                                 if (can_sleep)
 879                                         cond_resched();
 880                                 continue;
 881                         }
 882                         kref_get(&zhdr->refcount);
 883                         break;
 884                 }
 885         }
 886
 887         if (zhdr && !zhdr->slots) {
 888                 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
 889                 if (!zhdr->slots)
 890                         goto out_fail;
 891         }
 892         return zhdr;
 893
 894 out_fail:
 895         if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 896                 add_to_unbuddied(pool, zhdr);
 897                 z3fold_page_unlock(zhdr);
 898         }
 899         return NULL;
 900 }
 901
 902 /*
 903  * API Functions
 904  */
 905
 906 /**
 907  * z3fold_create_pool() - create a new z3fold pool
 908  * @name:       pool name
 909  * @gfp:        gfp flags when allocating the z3fold pool structure
 910  * @ops:        user-defined operations for the z3fold pool
 911  *
 912  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 913  * failed.
 914  */
 915 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 916                 const struct z3fold_ops *ops)
 917 {
 918         struct z3fold_pool *pool = NULL;
 919         int i, cpu;
 920
 921         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 922         if (!pool)
 923                 goto out;
 924         pool->c_handle = kmem_cache_create("z3fold_handle",
 925                                 sizeof(struct z3fold_buddy_slots),
 926                                 SLOTS_ALIGN, 0, NULL);
 927         if (!pool->c_handle)
 928                 goto out_c;
 929         spin_lock_init(&pool->lock);
 930         spin_lock_init(&pool->stale_lock);
 931         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 932                                          __alignof__(struct list_head));
 933         if (!pool->unbuddied)
 934                 goto out_pool;
 935         for_each_possible_cpu(cpu) {
 936                 struct list_head *unbuddied =
 937                                 per_cpu_ptr(pool->unbuddied, cpu);
 938                 for_each_unbuddied_list(i, 0)
 939                         INIT_LIST_HEAD(&unbuddied[i]);
 940         }
 941         INIT_LIST_HEAD(&pool->lru);
 942         INIT_LIST_HEAD(&pool->stale);
 943         atomic64_set(&pool->pages_nr, 0);
 944         pool->name = name;
 945         pool->compact_wq = create_singlethread_workqueue(pool->name);
 946         if (!pool->compact_wq)
 947                 goto out_unbuddied;
 948         pool->release_wq = create_singlethread_workqueue(pool->name);
 949         if (!pool->release_wq)
 950                 goto out_wq;
 951         INIT_WORK(&pool->work, free_pages_work);
 952         pool->ops = ops;
 953         return pool;
 954
 955 out_wq:
 956         destroy_workqueue(pool->compact_wq);
 957 out_unbuddied:
 958         free_percpu(pool->unbuddied);
 959 out_pool:
 960         kmem_cache_destroy(pool->c_handle);
 961 out_c:
 962         kfree(pool);
 963 out:
 964         return NULL;
 965 }
 966
 967 /**
 968  * z3fold_destroy_pool() - destroys an existing z3fold pool
 969  * @pool:       the z3fold pool to be destroyed
 970  *
 971  * The pool should be emptied before this function is called.
 972  */
 973 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 974 {
 975         kmem_cache_destroy(pool->c_handle);
 976
 977         /*
 978          * We need to destroy pool->compact_wq before pool->release_wq,
 979          * as any pending work on pool->compact_wq will call
 980          * queue_work(pool->release_wq, &pool->work).
 981          *
 982          * There are still outstanding pages until both workqueues are drained,
 983          * so we cannot unregister migration until then.
 984          */
 985
 986         destroy_workqueue(pool->compact_wq);
 987         destroy_workqueue(pool->release_wq);
 988         free_percpu(pool->unbuddied);
 989         kfree(pool);
 990 }
 991
 992 static const struct movable_operations z3fold_mops;
 993
 994 /**
 995  * z3fold_alloc() - allocates a region of a given size
 996  * @pool:       z3fold pool from which to allocate
 997  * @size:       size in bytes of the desired allocation
 998  * @gfp:        gfp flags used if the pool needs to grow
 999  * @handle:     handle of the new allocation
1000  *
1001  * This function will attempt to find a free region in the pool large enough to
1002  * satisfy the allocation request.  A search of the unbuddied lists is
1003  * performed first. If no suitable free region is found, then a new page is
1004  * allocated and added to the pool to satisfy the request.
1005  *
1006  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1007  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1008  * a new page.
1009  */
1010 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1011                         unsigned long *handle)
1012 {
1013         int chunks = size_to_chunks(size);
1014         struct z3fold_header *zhdr = NULL;
1015         struct page *page = NULL;
1016         enum buddy bud;
1017         bool can_sleep = gfpflags_allow_blocking(gfp);
1018
1019         if (!size || (gfp & __GFP_HIGHMEM))
1020                 return -EINVAL;
1021
1022         if (size > PAGE_SIZE)
1023                 return -ENOSPC;
1024
1025         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1026                 bud = HEADLESS;
1027         else {
1028 retry:
1029                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1030                 if (zhdr) {
1031                         bud = get_free_buddy(zhdr, chunks);
1032                         if (bud == HEADLESS) {
1033                                 if (!kref_put(&zhdr->refcount,
1034                                              release_z3fold_page_locked))
1035                                         z3fold_page_unlock(zhdr);
1036                                 pr_err("No free chunks in unbuddied\n");
1037                                 WARN_ON(1);
1038                                 goto retry;
1039                         }
1040                         page = virt_to_page(zhdr);
1041                         goto found;
1042                 }
1043                 bud = FIRST;
1044         }
1045
1046         page = alloc_page(gfp);
1047         if (!page)
1048                 return -ENOMEM;
1049
1050         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1051         if (!zhdr) {
1052                 __free_page(page);
1053                 return -ENOMEM;
1054         }
1055         atomic64_inc(&pool->pages_nr);
1056
1057         if (bud == HEADLESS) {
1058                 set_bit(PAGE_HEADLESS, &page->private);
1059                 goto headless;
1060         }
1061         if (can_sleep) {
1062                 lock_page(page);
1063                 __SetPageMovable(page, &z3fold_mops);
1064                 unlock_page(page);
1065         } else {
1066                 WARN_ON(!trylock_page(page));
1067                 __SetPageMovable(page, &z3fold_mops);
1068                 unlock_page(page);
1069         }
1070         z3fold_page_lock(zhdr);
1071
1072 found:
1073         if (bud == FIRST)
1074                 zhdr->first_chunks = chunks;
1075         else if (bud == LAST)
1076                 zhdr->last_chunks = chunks;
1077         else {
1078                 zhdr->middle_chunks = chunks;
1079                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1080         }
1081         add_to_unbuddied(pool, zhdr);
1082
1083 headless:
1084         spin_lock(&pool->lock);
1085         /* Add/move z3fold page to beginning of LRU */
1086         if (!list_empty(&page->lru))
1087                 list_del(&page->lru);
1088
1089         list_add(&page->lru, &pool->lru);
1090
1091         *handle = encode_handle(zhdr, bud);
1092         spin_unlock(&pool->lock);
1093         if (bud != HEADLESS)
1094                 z3fold_page_unlock(zhdr);
1095
1096         return 0;
1097 }
1098
1099 /**
1100  * z3fold_free() - frees the allocation associated with the given handle
1101  * @pool:       pool in which the allocation resided
1102  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1103  *
1104  * In the case that the z3fold page in which the allocation resides is under
1105  * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1106  * only sets the first|middle|last_chunks to 0.  The page is actually freed
1107  * once all buddies are evicted (see z3fold_reclaim_page() below).
1108  */
1109 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1110 {
1111         struct z3fold_header *zhdr;
1112         struct page *page;
1113         enum buddy bud;
1114         bool page_claimed;
1115
1116         zhdr = get_z3fold_header(handle);
1117         page = virt_to_page(zhdr);
1118         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1119
1120         if (test_bit(PAGE_HEADLESS, &page->private)) {
1121                 /* if a headless page is under reclaim, just leave.
1122                  * NB: we use test_and_set_bit for a reason: if the bit
1123                  * has not been set before, we release this page
1124                  * immediately so we don't care about its value any more.
1125                  */
1126                 if (!page_claimed) {
1127                         spin_lock(&pool->lock);
1128                         list_del(&page->lru);
1129                         spin_unlock(&pool->lock);
1130                         put_z3fold_header(zhdr);
1131                         free_z3fold_page(page, true);
1132                         atomic64_dec(&pool->pages_nr);
1133                 }
1134                 return;
1135         }
1136
1137         /* Non-headless case */
1138         bud = handle_to_buddy(handle);
1139
1140         switch (bud) {
1141         case FIRST:
1142                 zhdr->first_chunks = 0;
1143                 break;
1144         case MIDDLE:
1145                 zhdr->middle_chunks = 0;
1146                 break;
1147         case LAST:
1148                 zhdr->last_chunks = 0;
1149                 break;
1150         default:
1151                 pr_err("%s: unknown bud %d\n", __func__, bud);
1152                 WARN_ON(1);
1153                 put_z3fold_header(zhdr);
1154                 return;
1155         }
1156
1157         if (!page_claimed)
1158                 free_handle(handle, zhdr);
1159         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list))
1160                 return;
1161         if (page_claimed) {
1162                 /* the page has not been claimed by us */
1163                 put_z3fold_header(zhdr);
1164                 return;
1165         }
1166         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1167                 clear_bit(PAGE_CLAIMED, &page->private);
1168                 put_z3fold_header(zhdr);
1169                 return;
1170         }
1171         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1172                 zhdr->cpu = -1;
1173                 kref_get(&zhdr->refcount);
1174                 clear_bit(PAGE_CLAIMED, &page->private);
1175                 do_compact_page(zhdr, true);
1176                 return;
1177         }
1178         kref_get(&zhdr->refcount);
1179         clear_bit(PAGE_CLAIMED, &page->private);
1180         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1181         put_z3fold_header(zhdr);
1182 }
1183
1184 /**
1185  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1186  * @pool:       pool from which a page will attempt to be evicted
1187  * @retries:    number of pages on the LRU list for which eviction will
1188  *              be attempted before failing
1189  *
1190  * z3fold reclaim is different from normal system reclaim in that it is done
1191  * from the bottom, up. This is because only the bottom layer, z3fold, has
1192  * information on how the allocations are organized within each z3fold page.
1193  * This has the potential to create interesting locking situations between
1194  * z3fold and the user, however.
1195  *
1196  * To avoid these, this is how z3fold_reclaim_page() should be called:
1197  *
1198  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1199  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1200  * call the user-defined eviction handler with the pool and handle as
1201  * arguments.
1202  *
1203  * If the handle can not be evicted, the eviction handler should return
1204  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1205  * appropriate list and try the next z3fold page on the LRU up to
1206  * a user defined number of retries.
1207  *
1208  * If the handle is successfully evicted, the eviction handler should
1209  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1210  * contains logic to delay freeing the page if the page is under reclaim,
1211  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1212  *
1213  * If all buddies in the z3fold page are successfully evicted, then the
1214  * z3fold page can be freed.
1215  *
1216  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1217  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1218  * the retry limit was hit.
1219  */
1220 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1221 {
1222         int i, ret = -1;
1223         struct z3fold_header *zhdr = NULL;
1224         struct page *page = NULL;
1225         struct list_head *pos;
1226         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1227         struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN)));
1228
1229         rwlock_init(&slots.lock);
1230         slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE);
1231
1232         spin_lock(&pool->lock);
1233         if (!pool->ops || !pool->ops->evict || retries == 0) {
1234                 spin_unlock(&pool->lock);
1235                 return -EINVAL;
1236         }
1237         for (i = 0; i < retries; i++) {
1238                 if (list_empty(&pool->lru)) {
1239                         spin_unlock(&pool->lock);
1240                         return -EINVAL;
1241                 }
1242                 list_for_each_prev(pos, &pool->lru) {
1243                         page = list_entry(pos, struct page, lru);
1244
1245                         zhdr = page_address(page);
1246                         if (test_bit(PAGE_HEADLESS, &page->private)) {
1247                                 /*
1248                                  * For non-headless pages, we wait to do this
1249                                  * until we have the page lock to avoid racing
1250                                  * with __z3fold_alloc(). Headless pages don't
1251                                  * have a lock (and __z3fold_alloc() will never
1252                                  * see them), but we still need to test and set
1253                                  * PAGE_CLAIMED to avoid racing with
1254                                  * z3fold_free(), so just do it now before
1255                                  * leaving the loop.
1256                                  */
1257                                 if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1258                                         continue;
1259
1260                                 break;
1261                         }
1262
1263                         if (!z3fold_page_trylock(zhdr)) {
1264                                 zhdr = NULL;
1265                                 continue; /* can't evict at this point */
1266                         }
1267
1268                         /* test_and_set_bit is of course atomic, but we still
1269                          * need to do it under page lock, otherwise checking
1270                          * that bit in __z3fold_alloc wouldn't make sense
1271                          */
1272                         if (zhdr->foreign_handles ||
1273                             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1274                                 z3fold_page_unlock(zhdr);
1275                                 zhdr = NULL;
1276                                 continue; /* can't evict such page */
1277                         }
1278                         list_del_init(&zhdr->buddy);
1279                         zhdr->cpu = -1;
1280                         /* See comment in __z3fold_alloc. */
1281                         kref_get(&zhdr->refcount);
1282                         break;
1283                 }
1284
1285                 if (!zhdr)
1286                         break;
1287
1288                 list_del_init(&page->lru);
1289                 spin_unlock(&pool->lock);
1290
1291                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1292                         /*
1293                          * We need encode the handles before unlocking, and
1294                          * use our local slots structure because z3fold_free
1295                          * can zero out zhdr->slots and we can't do much
1296                          * about that
1297                          */
1298                         first_handle = 0;
1299                         last_handle = 0;
1300                         middle_handle = 0;
1301                         memset(slots.slot, 0, sizeof(slots.slot));
1302                         if (zhdr->first_chunks)
1303                                 first_handle = __encode_handle(zhdr, &slots,
1304                                                                 FIRST);
1305                         if (zhdr->middle_chunks)
1306                                 middle_handle = __encode_handle(zhdr, &slots,
1307                                                                 MIDDLE);
1308                         if (zhdr->last_chunks)
1309                                 last_handle = __encode_handle(zhdr, &slots,
1310                                                                 LAST);
1311                         /*
1312                          * it's safe to unlock here because we hold a
1313                          * reference to this page
1314                          */
1315                         z3fold_page_unlock(zhdr);
1316                 } else {
1317                         first_handle = encode_handle(zhdr, HEADLESS);
1318                         last_handle = middle_handle = 0;
1319                 }
1320                 /* Issue the eviction callback(s) */
1321                 if (middle_handle) {
1322                         ret = pool->ops->evict(pool, middle_handle);
1323                         if (ret)
1324                                 goto next;
1325                 }
1326                 if (first_handle) {
1327                         ret = pool->ops->evict(pool, first_handle);
1328                         if (ret)
1329                                 goto next;
1330                 }
1331                 if (last_handle) {
1332                         ret = pool->ops->evict(pool, last_handle);
1333                         if (ret)
1334                                 goto next;
1335                 }
1336 next:
1337                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1338                         if (ret == 0) {
1339                                 free_z3fold_page(page, true);
1340                                 atomic64_dec(&pool->pages_nr);
1341                                 return 0;
1342                         }
1343                         spin_lock(&pool->lock);
1344                         list_add(&page->lru, &pool->lru);
1345                         spin_unlock(&pool->lock);
1346                         clear_bit(PAGE_CLAIMED, &page->private);
1347                 } else {
1348                         struct z3fold_buddy_slots *slots = zhdr->slots;
1349                         z3fold_page_lock(zhdr);
1350                         if (kref_put(&zhdr->refcount,
1351                                         release_z3fold_page_locked)) {
1352                                 kmem_cache_free(pool->c_handle, slots);
1353                                 return 0;
1354                         }
1355                         /*
1356                          * if we are here, the page is still not completely
1357                          * free. Take the global pool lock then to be able
1358                          * to add it back to the lru list
1359                          */
1360                         spin_lock(&pool->lock);
1361                         list_add(&page->lru, &pool->lru);
1362                         spin_unlock(&pool->lock);
1363                         if (list_empty(&zhdr->buddy))
1364                                 add_to_unbuddied(pool, zhdr);
1365                         clear_bit(PAGE_CLAIMED, &page->private);
1366                         z3fold_page_unlock(zhdr);
1367                 }
1368
1369                 /* We started off locked to we need to lock the pool back */
1370                 spin_lock(&pool->lock);
1371         }
1372         spin_unlock(&pool->lock);
1373         return -EAGAIN;
1374 }
1375
1376 /**
1377  * z3fold_map() - maps the allocation associated with the given handle
1378  * @pool:       pool in which the allocation resides
1379  * @handle:     handle associated with the allocation to be mapped
1380  *
1381  * Extracts the buddy number from handle and constructs the pointer to the
1382  * correct starting chunk within the page.
1383  *
1384  * Returns: a pointer to the mapped allocation
1385  */
1386 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1387 {
1388         struct z3fold_header *zhdr;
1389         struct page *page;
1390         void *addr;
1391         enum buddy buddy;
1392
1393         zhdr = get_z3fold_header(handle);
1394         addr = zhdr;
1395         page = virt_to_page(zhdr);
1396
1397         if (test_bit(PAGE_HEADLESS, &page->private))
1398                 goto out;
1399
1400         buddy = handle_to_buddy(handle);
1401         switch (buddy) {
1402         case FIRST:
1403                 addr += ZHDR_SIZE_ALIGNED;
1404                 break;
1405         case MIDDLE:
1406                 addr += zhdr->start_middle << CHUNK_SHIFT;
1407                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1408                 break;
1409         case LAST:
1410                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1411                 break;
1412         default:
1413                 pr_err("unknown buddy id %d\n", buddy);
1414                 WARN_ON(1);
1415                 addr = NULL;
1416                 break;
1417         }
1418
1419         if (addr)
1420                 zhdr->mapped_count++;
1421 out:
1422         put_z3fold_header(zhdr);
1423         return addr;
1424 }
1425
1426 /**
1427  * z3fold_unmap() - unmaps the allocation associated with the given handle
1428  * @pool:       pool in which the allocation resides
1429  * @handle:     handle associated with the allocation to be unmapped
1430  */
1431 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1432 {
1433         struct z3fold_header *zhdr;
1434         struct page *page;
1435         enum buddy buddy;
1436
1437         zhdr = get_z3fold_header(handle);
1438         page = virt_to_page(zhdr);
1439
1440         if (test_bit(PAGE_HEADLESS, &page->private))
1441                 return;
1442
1443         buddy = handle_to_buddy(handle);
1444         if (buddy == MIDDLE)
1445                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1446         zhdr->mapped_count--;
1447         put_z3fold_header(zhdr);
1448 }
1449
1450 /**
1451  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1452  * @pool:       pool whose size is being queried
1453  *
1454  * Returns: size in pages of the given pool.
1455  */
1456 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1457 {
1458         return atomic64_read(&pool->pages_nr);
1459 }
1460
1461 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1462 {
1463         struct z3fold_header *zhdr;
1464         struct z3fold_pool *pool;
1465
1466         VM_BUG_ON_PAGE(!PageMovable(page), page);
1467         VM_BUG_ON_PAGE(PageIsolated(page), page);
1468
1469         if (test_bit(PAGE_HEADLESS, &page->private))
1470                 return false;
1471
1472         zhdr = page_address(page);
1473         z3fold_page_lock(zhdr);
1474         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1475             test_bit(PAGE_STALE, &page->private))
1476                 goto out;
1477
1478         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1479                 goto out;
1480
1481         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1482                 goto out;
1483         pool = zhdr_to_pool(zhdr);
1484         spin_lock(&pool->lock);
1485         if (!list_empty(&zhdr->buddy))
1486                 list_del_init(&zhdr->buddy);
1487         if (!list_empty(&page->lru))
1488                 list_del_init(&page->lru);
1489         spin_unlock(&pool->lock);
1490
1491         kref_get(&zhdr->refcount);
1492         z3fold_page_unlock(zhdr);
1493         return true;
1494
1495 out:
1496         z3fold_page_unlock(zhdr);
1497         return false;
1498 }
1499
1500 static int z3fold_page_migrate(struct page *newpage, struct page *page,
1501                 enum migrate_mode mode)
1502 {
1503         struct z3fold_header *zhdr, *new_zhdr;
1504         struct z3fold_pool *pool;
1505
1506         VM_BUG_ON_PAGE(!PageMovable(page), page);
1507         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1508         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1509         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1510
1511         zhdr = page_address(page);
1512         pool = zhdr_to_pool(zhdr);
1513
1514         if (!z3fold_page_trylock(zhdr))
1515                 return -EAGAIN;
1516         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1517                 clear_bit(PAGE_CLAIMED, &page->private);
1518                 z3fold_page_unlock(zhdr);
1519                 return -EBUSY;
1520         }
1521         if (work_pending(&zhdr->work)) {
1522                 z3fold_page_unlock(zhdr);
1523                 return -EAGAIN;
1524         }
1525         new_zhdr = page_address(newpage);
1526         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1527         newpage->private = page->private;
1528         set_bit(PAGE_MIGRATED, &page->private);
1529         z3fold_page_unlock(zhdr);
1530         spin_lock_init(&new_zhdr->page_lock);
1531         INIT_WORK(&new_zhdr->work, compact_page_work);
1532         /*
1533          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1534          * so we only have to reinitialize it.
1535          */
1536         INIT_LIST_HEAD(&new_zhdr->buddy);
1537         __ClearPageMovable(page);
1538
1539         get_page(newpage);
1540         z3fold_page_lock(new_zhdr);
1541         if (new_zhdr->first_chunks)
1542                 encode_handle(new_zhdr, FIRST);
1543         if (new_zhdr->last_chunks)
1544                 encode_handle(new_zhdr, LAST);
1545         if (new_zhdr->middle_chunks)
1546                 encode_handle(new_zhdr, MIDDLE);
1547         set_bit(NEEDS_COMPACTING, &newpage->private);
1548         new_zhdr->cpu = smp_processor_id();
1549         spin_lock(&pool->lock);
1550         list_add(&newpage->lru, &pool->lru);
1551         spin_unlock(&pool->lock);
1552         __SetPageMovable(newpage, &z3fold_mops);
1553         z3fold_page_unlock(new_zhdr);
1554
1555         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1556
1557         /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1558         page->private = 0;
1559         put_page(page);
1560         return 0;
1561 }
1562
1563 static void z3fold_page_putback(struct page *page)
1564 {
1565         struct z3fold_header *zhdr;
1566         struct z3fold_pool *pool;
1567
1568         zhdr = page_address(page);
1569         pool = zhdr_to_pool(zhdr);
1570
1571         z3fold_page_lock(zhdr);
1572         if (!list_empty(&zhdr->buddy))
1573                 list_del_init(&zhdr->buddy);
1574         INIT_LIST_HEAD(&page->lru);
1575         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
1576                 return;
1577         spin_lock(&pool->lock);
1578         list_add(&page->lru, &pool->lru);
1579         spin_unlock(&pool->lock);
1580         if (list_empty(&zhdr->buddy))
1581                 add_to_unbuddied(pool, zhdr);
1582         clear_bit(PAGE_CLAIMED, &page->private);
1583         z3fold_page_unlock(zhdr);
1584 }
1585
1586 static const struct movable_operations z3fold_mops = {
1587         .isolate_page = z3fold_page_isolate,
1588         .migrate_page = z3fold_page_migrate,
1589         .putback_page = z3fold_page_putback,
1590 };
1591
1592 /*****************
1593  * zpool
1594  ****************/
1595
1596 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1597 {
1598         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1599                 return pool->zpool_ops->evict(pool->zpool, handle);
1600         else
1601                 return -ENOENT;
1602 }
1603
1604 static const struct z3fold_ops z3fold_zpool_ops = {
1605         .evict =        z3fold_zpool_evict
1606 };
1607
1608 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1609                                const struct zpool_ops *zpool_ops,
1610                                struct zpool *zpool)
1611 {
1612         struct z3fold_pool *pool;
1613
1614         pool = z3fold_create_pool(name, gfp,
1615                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1616         if (pool) {
1617                 pool->zpool = zpool;
1618                 pool->zpool_ops = zpool_ops;
1619         }
1620         return pool;
1621 }
1622
1623 static void z3fold_zpool_destroy(void *pool)
1624 {
1625         z3fold_destroy_pool(pool);
1626 }
1627
1628 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1629                         unsigned long *handle)
1630 {
1631         return z3fold_alloc(pool, size, gfp, handle);
1632 }
1633 static void z3fold_zpool_free(void *pool, unsigned long handle)
1634 {
1635         z3fold_free(pool, handle);
1636 }
1637
1638 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1639                         unsigned int *reclaimed)
1640 {
1641         unsigned int total = 0;
1642         int ret = -EINVAL;
1643
1644         while (total < pages) {
1645                 ret = z3fold_reclaim_page(pool, 8);
1646                 if (ret < 0)
1647                         break;
1648                 total++;
1649         }
1650
1651         if (reclaimed)
1652                 *reclaimed = total;
1653
1654         return ret;
1655 }
1656
1657 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1658                         enum zpool_mapmode mm)
1659 {
1660         return z3fold_map(pool, handle);
1661 }
1662 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1663 {
1664         z3fold_unmap(pool, handle);
1665 }
1666
1667 static u64 z3fold_zpool_total_size(void *pool)
1668 {
1669         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1670 }
1671
1672 static struct zpool_driver z3fold_zpool_driver = {
1673         .type =         "z3fold",
1674         .sleep_mapped = true,
1675         .owner =        THIS_MODULE,
1676         .create =       z3fold_zpool_create,
1677         .destroy =      z3fold_zpool_destroy,
1678         .malloc =       z3fold_zpool_malloc,
1679         .free =         z3fold_zpool_free,
1680         .shrink =       z3fold_zpool_shrink,
1681         .map =          z3fold_zpool_map,
1682         .unmap =        z3fold_zpool_unmap,
1683         .total_size =   z3fold_zpool_total_size,
1684 };
1685
1686 MODULE_ALIAS("zpool-z3fold");
1687
1688 static int __init init_z3fold(void)
1689 {
1690         /*
1691          * Make sure the z3fold header is not larger than the page size and
1692          * there has remaining spaces for its buddy.
1693          */
1694         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1695         zpool_register_driver(&z3fold_zpool_driver);
1696
1697         return 0;
1698 }
1699
1700 static void __exit exit_z3fold(void)
1701 {
1702         zpool_unregister_driver(&z3fold_zpool_driver);
1703 }
1704
1705 module_init(init_z3fold);
1706 module_exit(exit_z3fold);
1707
1708 MODULE_LICENSE("GPL");
1709 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1710 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");