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