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/rwlock.h>
  45 #include <linux/zpool.h>
  46 #include <linux/magic.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 + flags */
  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_ORPHANED = 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_alloc(pool->c_handle,
 216                                  (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
 217
 218         if (slots) {
 219                 memset(slots->slot, 0, sizeof(slots->slot));
 220                 slots->pool = (unsigned long)pool;
 221                 rwlock_init(&slots->lock);
 222         }
 223
 224         return slots;
 225 }
 226
 227 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 228 {
 229         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 230 }
 231
 232 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 233 {
 234         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 235 }
 236
 237 /* Lock a z3fold page */
 238 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 239 {
 240         spin_lock(&zhdr->page_lock);
 241 }
 242
 243 /* Try to lock a z3fold page */
 244 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 245 {
 246         return spin_trylock(&zhdr->page_lock);
 247 }
 248
 249 /* Unlock a z3fold page */
 250 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 251 {
 252         spin_unlock(&zhdr->page_lock);
 253 }
 254
 255
 256 static inline struct z3fold_header *__get_z3fold_header(unsigned long handle,
 257                                                         bool lock)
 258 {
 259         struct z3fold_buddy_slots *slots;
 260         struct z3fold_header *zhdr;
 261         int locked = 0;
 262
 263         if (!(handle & (1 << PAGE_HEADLESS))) {
 264                 slots = handle_to_slots(handle);
 265                 do {
 266                         unsigned long addr;
 267
 268                         read_lock(&slots->lock);
 269                         addr = *(unsigned long *)handle;
 270                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 271                         if (lock)
 272                                 locked = z3fold_page_trylock(zhdr);
 273                         read_unlock(&slots->lock);
 274                         if (locked)
 275                                 break;
 276                         cpu_relax();
 277                 } while (lock);
 278         } else {
 279                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 280         }
 281
 282         return zhdr;
 283 }
 284
 285 /* Returns the z3fold page where a given handle is stored */
 286 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
 287 {
 288         return __get_z3fold_header(h, false);
 289 }
 290
 291 /* return locked z3fold page if it's not headless */
 292 static inline struct z3fold_header *get_z3fold_header(unsigned long h)
 293 {
 294         return __get_z3fold_header(h, true);
 295 }
 296
 297 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 298 {
 299         struct page *page = virt_to_page(zhdr);
 300
 301         if (!test_bit(PAGE_HEADLESS, &page->private))
 302                 z3fold_page_unlock(zhdr);
 303 }
 304
 305 static inline void free_handle(unsigned long handle)
 306 {
 307         struct z3fold_buddy_slots *slots;
 308         struct z3fold_header *zhdr;
 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         zhdr = handle_to_z3fold_header(handle);
 319         slots = handle_to_slots(handle);
 320         write_lock(&slots->lock);
 321         *(unsigned long *)handle = 0;
 322         write_unlock(&slots->lock);
 323         if (zhdr->slots == slots)
 324                 return; /* simple case, nothing else to do */
 325
 326         /* we are freeing a foreign handle if we are here */
 327         zhdr->foreign_handles--;
 328         is_free = true;
 329         read_lock(&slots->lock);
 330         if (!test_bit(HANDLES_ORPHANED, &slots->pool)) {
 331                 read_unlock(&slots->lock);
 332                 return;
 333         }
 334         for (i = 0; i <= BUDDY_MASK; i++) {
 335                 if (slots->slot[i]) {
 336                         is_free = false;
 337                         break;
 338                 }
 339         }
 340         read_unlock(&slots->lock);
 341
 342         if (is_free) {
 343                 struct z3fold_pool *pool = slots_to_pool(slots);
 344
 345                 kmem_cache_free(pool->c_handle, slots);
 346         }
 347 }
 348
 349 static int z3fold_init_fs_context(struct fs_context *fc)
 350 {
 351         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 352 }
 353
 354 static struct file_system_type z3fold_fs = {
 355         .name           = "z3fold",
 356         .init_fs_context = z3fold_init_fs_context,
 357         .kill_sb        = kill_anon_super,
 358 };
 359
 360 static struct vfsmount *z3fold_mnt;
 361 static int z3fold_mount(void)
 362 {
 363         int ret = 0;
 364
 365         z3fold_mnt = kern_mount(&z3fold_fs);
 366         if (IS_ERR(z3fold_mnt))
 367                 ret = PTR_ERR(z3fold_mnt);
 368
 369         return ret;
 370 }
 371
 372 static void z3fold_unmount(void)
 373 {
 374         kern_unmount(z3fold_mnt);
 375 }
 376
 377 static const struct address_space_operations z3fold_aops;
 378 static int z3fold_register_migration(struct z3fold_pool *pool)
 379 {
 380         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 381         if (IS_ERR(pool->inode)) {
 382                 pool->inode = NULL;
 383                 return 1;
 384         }
 385
 386         pool->inode->i_mapping->private_data = pool;
 387         pool->inode->i_mapping->a_ops = &z3fold_aops;
 388         return 0;
 389 }
 390
 391 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 392 {
 393         if (pool->inode)
 394                 iput(pool->inode);
 395  }
 396
 397 /* Initializes the z3fold header of a newly allocated z3fold page */
 398 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 399                                         struct z3fold_pool *pool, gfp_t gfp)
 400 {
 401         struct z3fold_header *zhdr = page_address(page);
 402         struct z3fold_buddy_slots *slots;
 403
 404         INIT_LIST_HEAD(&page->lru);
 405         clear_bit(PAGE_HEADLESS, &page->private);
 406         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 407         clear_bit(NEEDS_COMPACTING, &page->private);
 408         clear_bit(PAGE_STALE, &page->private);
 409         clear_bit(PAGE_CLAIMED, &page->private);
 410         if (headless)
 411                 return zhdr;
 412
 413         slots = alloc_slots(pool, gfp);
 414         if (!slots)
 415                 return NULL;
 416
 417         spin_lock_init(&zhdr->page_lock);
 418         kref_init(&zhdr->refcount);
 419         zhdr->first_chunks = 0;
 420         zhdr->middle_chunks = 0;
 421         zhdr->last_chunks = 0;
 422         zhdr->first_num = 0;
 423         zhdr->start_middle = 0;
 424         zhdr->cpu = -1;
 425         zhdr->foreign_handles = 0;
 426         zhdr->slots = slots;
 427         zhdr->pool = pool;
 428         INIT_LIST_HEAD(&zhdr->buddy);
 429         INIT_WORK(&zhdr->work, compact_page_work);
 430         return zhdr;
 431 }
 432
 433 /* Resets the struct page fields and frees the page */
 434 static void free_z3fold_page(struct page *page, bool headless)
 435 {
 436         if (!headless) {
 437                 lock_page(page);
 438                 __ClearPageMovable(page);
 439                 unlock_page(page);
 440         }
 441         ClearPagePrivate(page);
 442         __free_page(page);
 443 }
 444
 445 /* Helper function to build the index */
 446 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 447 {
 448         return (bud + zhdr->first_num) & BUDDY_MASK;
 449 }
 450
 451 /*
 452  * Encodes the handle of a particular buddy within a z3fold page
 453  * Pool lock should be held as this function accesses first_num
 454  */
 455 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 456                                 struct z3fold_buddy_slots *slots,
 457                                 enum buddy bud)
 458 {
 459         unsigned long h = (unsigned long)zhdr;
 460         int idx = 0;
 461
 462         /*
 463          * For a headless page, its handle is its pointer with the extra
 464          * PAGE_HEADLESS bit set
 465          */
 466         if (bud == HEADLESS)
 467                 return h | (1 << PAGE_HEADLESS);
 468
 469         /* otherwise, return pointer to encoded handle */
 470         idx = __idx(zhdr, bud);
 471         h += idx;
 472         if (bud == LAST)
 473                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 474
 475         write_lock(&slots->lock);
 476         slots->slot[idx] = h;
 477         write_unlock(&slots->lock);
 478         return (unsigned long)&slots->slot[idx];
 479 }
 480
 481 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 482 {
 483         return __encode_handle(zhdr, zhdr->slots, bud);
 484 }
 485
 486 /* only for LAST bud, returns zero otherwise */
 487 static unsigned short handle_to_chunks(unsigned long handle)
 488 {
 489         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 490         unsigned long addr;
 491
 492         read_lock(&slots->lock);
 493         addr = *(unsigned long *)handle;
 494         read_unlock(&slots->lock);
 495         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 496 }
 497
 498 /*
 499  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 500  *  but that doesn't matter. because the masking will result in the
 501  *  correct buddy number.
 502  */
 503 static enum buddy handle_to_buddy(unsigned long handle)
 504 {
 505         struct z3fold_header *zhdr;
 506         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 507         unsigned long addr;
 508
 509         read_lock(&slots->lock);
 510         WARN_ON(handle & (1 << PAGE_HEADLESS));
 511         addr = *(unsigned long *)handle;
 512         read_unlock(&slots->lock);
 513         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 514         return (addr - zhdr->first_num) & BUDDY_MASK;
 515 }
 516
 517 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 518 {
 519         return zhdr->pool;
 520 }
 521
 522 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 523 {
 524         struct page *page = virt_to_page(zhdr);
 525         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 526         bool is_free = true;
 527         int i;
 528
 529         WARN_ON(!list_empty(&zhdr->buddy));
 530         set_bit(PAGE_STALE, &page->private);
 531         clear_bit(NEEDS_COMPACTING, &page->private);
 532         spin_lock(&pool->lock);
 533         if (!list_empty(&page->lru))
 534                 list_del_init(&page->lru);
 535         spin_unlock(&pool->lock);
 536
 537         /* If there are no foreign handles, free the handles array */
 538         read_lock(&zhdr->slots->lock);
 539         for (i = 0; i <= BUDDY_MASK; i++) {
 540                 if (zhdr->slots->slot[i]) {
 541                         is_free = false;
 542                         break;
 543                 }
 544         }
 545         if (!is_free)
 546                 set_bit(HANDLES_ORPHANED, &zhdr->slots->pool);
 547         read_unlock(&zhdr->slots->lock);
 548
 549         if (is_free)
 550                 kmem_cache_free(pool->c_handle, zhdr->slots);
 551
 552         if (locked)
 553                 z3fold_page_unlock(zhdr);
 554
 555         spin_lock(&pool->stale_lock);
 556         list_add(&zhdr->buddy, &pool->stale);
 557         queue_work(pool->release_wq, &pool->work);
 558         spin_unlock(&pool->stale_lock);
 559 }
 560
 561 static void __attribute__((__unused__))
 562                         release_z3fold_page(struct kref *ref)
 563 {
 564         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 565                                                 refcount);
 566         __release_z3fold_page(zhdr, false);
 567 }
 568
 569 static void release_z3fold_page_locked(struct kref *ref)
 570 {
 571         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 572                                                 refcount);
 573         WARN_ON(z3fold_page_trylock(zhdr));
 574         __release_z3fold_page(zhdr, true);
 575 }
 576
 577 static void release_z3fold_page_locked_list(struct kref *ref)
 578 {
 579         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 580                                                refcount);
 581         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 582
 583         spin_lock(&pool->lock);
 584         list_del_init(&zhdr->buddy);
 585         spin_unlock(&pool->lock);
 586
 587         WARN_ON(z3fold_page_trylock(zhdr));
 588         __release_z3fold_page(zhdr, true);
 589 }
 590
 591 static void free_pages_work(struct work_struct *w)
 592 {
 593         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 594
 595         spin_lock(&pool->stale_lock);
 596         while (!list_empty(&pool->stale)) {
 597                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 598                                                 struct z3fold_header, buddy);
 599                 struct page *page = virt_to_page(zhdr);
 600
 601                 list_del(&zhdr->buddy);
 602                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 603                         continue;
 604                 spin_unlock(&pool->stale_lock);
 605                 cancel_work_sync(&zhdr->work);
 606                 free_z3fold_page(page, false);
 607                 cond_resched();
 608                 spin_lock(&pool->stale_lock);
 609         }
 610         spin_unlock(&pool->stale_lock);
 611 }
 612
 613 /*
 614  * Returns the number of free chunks in a z3fold page.
 615  * NB: can't be used with HEADLESS pages.
 616  */
 617 static int num_free_chunks(struct z3fold_header *zhdr)
 618 {
 619         int nfree;
 620         /*
 621          * If there is a middle object, pick up the bigger free space
 622          * either before or after it. Otherwise just subtract the number
 623          * of chunks occupied by the first and the last objects.
 624          */
 625         if (zhdr->middle_chunks != 0) {
 626                 int nfree_before = zhdr->first_chunks ?
 627                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 628                 int nfree_after = zhdr->last_chunks ?
 629                         0 : TOTAL_CHUNKS -
 630                                 (zhdr->start_middle + zhdr->middle_chunks);
 631                 nfree = max(nfree_before, nfree_after);
 632         } else
 633                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 634         return nfree;
 635 }
 636
 637 /* Add to the appropriate unbuddied list */
 638 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 639                                 struct z3fold_header *zhdr)
 640 {
 641         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 642                         zhdr->middle_chunks == 0) {
 643                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 644
 645                 int freechunks = num_free_chunks(zhdr);
 646                 spin_lock(&pool->lock);
 647                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 648                 spin_unlock(&pool->lock);
 649                 zhdr->cpu = smp_processor_id();
 650                 put_cpu_ptr(pool->unbuddied);
 651         }
 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                 if (new_zhdr->first_chunks == 0) {
 716                         if (new_zhdr->middle_chunks != 0 &&
 717                                         chunks >= new_zhdr->start_middle) {
 718                                 new_bud = LAST;
 719                         } else {
 720                                 new_bud = FIRST;
 721                         }
 722                 } else if (new_zhdr->last_chunks == 0) {
 723                         new_bud = LAST;
 724                 } else if (new_zhdr->middle_chunks == 0) {
 725                         new_bud = MIDDLE;
 726                 }
 727                 q = new_zhdr;
 728                 switch (new_bud) {
 729                 case FIRST:
 730                         new_zhdr->first_chunks = chunks;
 731                         q += ZHDR_SIZE_ALIGNED;
 732                         break;
 733                 case MIDDLE:
 734                         new_zhdr->middle_chunks = chunks;
 735                         new_zhdr->start_middle =
 736                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 737                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 738                         break;
 739                 case LAST:
 740                         new_zhdr->last_chunks = chunks;
 741                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 742                         break;
 743                 default:
 744                         goto out_fail;
 745                 }
 746                 new_zhdr->foreign_handles++;
 747                 memcpy(q, p, sz);
 748                 write_lock(&zhdr->slots->lock);
 749                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 750                         __idx(new_zhdr, new_bud);
 751                 if (new_bud == LAST)
 752                         *(unsigned long *)old_handle |=
 753                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 754                 write_unlock(&zhdr->slots->lock);
 755                 add_to_unbuddied(pool, new_zhdr);
 756                 z3fold_page_unlock(new_zhdr);
 757
 758                 *moved_chunks = 0;
 759         }
 760
 761         return new_zhdr;
 762
 763 out_fail:
 764         if (new_zhdr) {
 765                 if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
 766                         atomic64_dec(&pool->pages_nr);
 767                 else {
 768                         add_to_unbuddied(pool, new_zhdr);
 769                         z3fold_page_unlock(new_zhdr);
 770                 }
 771         }
 772         return NULL;
 773
 774 }
 775
 776 #define BIG_CHUNK_GAP   3
 777 /* Has to be called with lock held */
 778 static int z3fold_compact_page(struct z3fold_header *zhdr)
 779 {
 780         struct page *page = virt_to_page(zhdr);
 781
 782         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 783                 return 0; /* can't move middle chunk, it's used */
 784
 785         if (unlikely(PageIsolated(page)))
 786                 return 0;
 787
 788         if (zhdr->middle_chunks == 0)
 789                 return 0; /* nothing to compact */
 790
 791         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 792                 /* move to the beginning */
 793                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 794                 zhdr->first_chunks = zhdr->middle_chunks;
 795                 zhdr->middle_chunks = 0;
 796                 zhdr->start_middle = 0;
 797                 zhdr->first_num++;
 798                 return 1;
 799         }
 800
 801         /*
 802          * moving data is expensive, so let's only do that if
 803          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 804          */
 805         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 806             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 807                         BIG_CHUNK_GAP) {
 808                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 809                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 810                 return 1;
 811         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 812                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 813                                         + zhdr->middle_chunks) >=
 814                         BIG_CHUNK_GAP) {
 815                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 816                         zhdr->middle_chunks;
 817                 mchunk_memmove(zhdr, new_start);
 818                 zhdr->start_middle = new_start;
 819                 return 1;
 820         }
 821
 822         return 0;
 823 }
 824
 825 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 826 {
 827         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 828         struct page *page;
 829
 830         page = virt_to_page(zhdr);
 831         if (locked)
 832                 WARN_ON(z3fold_page_trylock(zhdr));
 833         else
 834                 z3fold_page_lock(zhdr);
 835         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 836                 z3fold_page_unlock(zhdr);
 837                 return;
 838         }
 839         spin_lock(&pool->lock);
 840         list_del_init(&zhdr->buddy);
 841         spin_unlock(&pool->lock);
 842
 843         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 844                 atomic64_dec(&pool->pages_nr);
 845                 return;
 846         }
 847
 848         if (unlikely(PageIsolated(page) ||
 849                      test_bit(PAGE_CLAIMED, &page->private) ||
 850                      test_bit(PAGE_STALE, &page->private))) {
 851                 z3fold_page_unlock(zhdr);
 852                 return;
 853         }
 854
 855         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 856             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 857                 if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 858                         atomic64_dec(&pool->pages_nr);
 859                 else
 860                         z3fold_page_unlock(zhdr);
 861                 return;
 862         }
 863
 864         z3fold_compact_page(zhdr);
 865         add_to_unbuddied(pool, zhdr);
 866         z3fold_page_unlock(zhdr);
 867 }
 868
 869 static void compact_page_work(struct work_struct *w)
 870 {
 871         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 872                                                 work);
 873
 874         do_compact_page(zhdr, false);
 875 }
 876
 877 /* returns _locked_ z3fold page header or NULL */
 878 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 879                                                 size_t size, bool can_sleep)
 880 {
 881         struct z3fold_header *zhdr = NULL;
 882         struct page *page;
 883         struct list_head *unbuddied;
 884         int chunks = size_to_chunks(size), i;
 885
 886 lookup:
 887         /* First, try to find an unbuddied z3fold page. */
 888         unbuddied = get_cpu_ptr(pool->unbuddied);
 889         for_each_unbuddied_list(i, chunks) {
 890                 struct list_head *l = &unbuddied[i];
 891
 892                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 893                                         struct z3fold_header, buddy);
 894
 895                 if (!zhdr)
 896                         continue;
 897
 898                 /* Re-check under lock. */
 899                 spin_lock(&pool->lock);
 900                 l = &unbuddied[i];
 901                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 902                                                 struct z3fold_header, buddy)) ||
 903                     !z3fold_page_trylock(zhdr)) {
 904                         spin_unlock(&pool->lock);
 905                         zhdr = NULL;
 906                         put_cpu_ptr(pool->unbuddied);
 907                         if (can_sleep)
 908                                 cond_resched();
 909                         goto lookup;
 910                 }
 911                 list_del_init(&zhdr->buddy);
 912                 zhdr->cpu = -1;
 913                 spin_unlock(&pool->lock);
 914
 915                 page = virt_to_page(zhdr);
 916                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 917                     test_bit(PAGE_CLAIMED, &page->private)) {
 918                         z3fold_page_unlock(zhdr);
 919                         zhdr = NULL;
 920                         put_cpu_ptr(pool->unbuddied);
 921                         if (can_sleep)
 922                                 cond_resched();
 923                         goto lookup;
 924                 }
 925
 926                 /*
 927                  * this page could not be removed from its unbuddied
 928                  * list while pool lock was held, and then we've taken
 929                  * page lock so kref_put could not be called before
 930                  * we got here, so it's safe to just call kref_get()
 931                  */
 932                 kref_get(&zhdr->refcount);
 933                 break;
 934         }
 935         put_cpu_ptr(pool->unbuddied);
 936
 937         if (!zhdr) {
 938                 int cpu;
 939
 940                 /* look for _exact_ match on other cpus' lists */
 941                 for_each_online_cpu(cpu) {
 942                         struct list_head *l;
 943
 944                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 945                         spin_lock(&pool->lock);
 946                         l = &unbuddied[chunks];
 947
 948                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 949                                                 struct z3fold_header, buddy);
 950
 951                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 952                                 spin_unlock(&pool->lock);
 953                                 zhdr = NULL;
 954                                 continue;
 955                         }
 956                         list_del_init(&zhdr->buddy);
 957                         zhdr->cpu = -1;
 958                         spin_unlock(&pool->lock);
 959
 960                         page = virt_to_page(zhdr);
 961                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 962                             test_bit(PAGE_CLAIMED, &page->private)) {
 963                                 z3fold_page_unlock(zhdr);
 964                                 zhdr = NULL;
 965                                 if (can_sleep)
 966                                         cond_resched();
 967                                 continue;
 968                         }
 969                         kref_get(&zhdr->refcount);
 970                         break;
 971                 }
 972         }
 973
 974         return zhdr;
 975 }
 976
 977 /*
 978  * API Functions
 979  */
 980
 981 /**
 982  * z3fold_create_pool() - create a new z3fold pool
 983  * @name:       pool name
 984  * @gfp:        gfp flags when allocating the z3fold pool structure
 985  * @ops:        user-defined operations for the z3fold pool
 986  *
 987  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 988  * failed.
 989  */
 990 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 991                 const struct z3fold_ops *ops)
 992 {
 993         struct z3fold_pool *pool = NULL;
 994         int i, cpu;
 995
 996         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 997         if (!pool)
 998                 goto out;
 999         pool->c_handle = kmem_cache_create("z3fold_handle",
1000                                 sizeof(struct z3fold_buddy_slots),
1001                                 SLOTS_ALIGN, 0, NULL);
1002         if (!pool->c_handle)
1003                 goto out_c;
1004         spin_lock_init(&pool->lock);
1005         spin_lock_init(&pool->stale_lock);
1006         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
1007         if (!pool->unbuddied)
1008                 goto out_pool;
1009         for_each_possible_cpu(cpu) {
1010                 struct list_head *unbuddied =
1011                                 per_cpu_ptr(pool->unbuddied, cpu);
1012                 for_each_unbuddied_list(i, 0)
1013                         INIT_LIST_HEAD(&unbuddied[i]);
1014         }
1015         INIT_LIST_HEAD(&pool->lru);
1016         INIT_LIST_HEAD(&pool->stale);
1017         atomic64_set(&pool->pages_nr, 0);
1018         pool->name = name;
1019         pool->compact_wq = create_singlethread_workqueue(pool->name);
1020         if (!pool->compact_wq)
1021                 goto out_unbuddied;
1022         pool->release_wq = create_singlethread_workqueue(pool->name);
1023         if (!pool->release_wq)
1024                 goto out_wq;
1025         if (z3fold_register_migration(pool))
1026                 goto out_rwq;
1027         INIT_WORK(&pool->work, free_pages_work);
1028         pool->ops = ops;
1029         return pool;
1030
1031 out_rwq:
1032         destroy_workqueue(pool->release_wq);
1033 out_wq:
1034         destroy_workqueue(pool->compact_wq);
1035 out_unbuddied:
1036         free_percpu(pool->unbuddied);
1037 out_pool:
1038         kmem_cache_destroy(pool->c_handle);
1039 out_c:
1040         kfree(pool);
1041 out:
1042         return NULL;
1043 }
1044
1045 /**
1046  * z3fold_destroy_pool() - destroys an existing z3fold pool
1047  * @pool:       the z3fold pool to be destroyed
1048  *
1049  * The pool should be emptied before this function is called.
1050  */
1051 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1052 {
1053         kmem_cache_destroy(pool->c_handle);
1054
1055         /*
1056          * We need to destroy pool->compact_wq before pool->release_wq,
1057          * as any pending work on pool->compact_wq will call
1058          * queue_work(pool->release_wq, &pool->work).
1059          *
1060          * There are still outstanding pages until both workqueues are drained,
1061          * so we cannot unregister migration until then.
1062          */
1063
1064         destroy_workqueue(pool->compact_wq);
1065         destroy_workqueue(pool->release_wq);
1066         z3fold_unregister_migration(pool);
1067         kfree(pool);
1068 }
1069
1070 /**
1071  * z3fold_alloc() - allocates a region of a given size
1072  * @pool:       z3fold pool from which to allocate
1073  * @size:       size in bytes of the desired allocation
1074  * @gfp:        gfp flags used if the pool needs to grow
1075  * @handle:     handle of the new allocation
1076  *
1077  * This function will attempt to find a free region in the pool large enough to
1078  * satisfy the allocation request.  A search of the unbuddied lists is
1079  * performed first. If no suitable free region is found, then a new page is
1080  * allocated and added to the pool to satisfy the request.
1081  *
1082  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
1083  * as z3fold pool pages.
1084  *
1085  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1086  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1087  * a new page.
1088  */
1089 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1090                         unsigned long *handle)
1091 {
1092         int chunks = size_to_chunks(size);
1093         struct z3fold_header *zhdr = NULL;
1094         struct page *page = NULL;
1095         enum buddy bud;
1096         bool can_sleep = gfpflags_allow_blocking(gfp);
1097
1098         if (!size)
1099                 return -EINVAL;
1100
1101         if (size > PAGE_SIZE)
1102                 return -ENOSPC;
1103
1104         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1105                 bud = HEADLESS;
1106         else {
1107 retry:
1108                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1109                 if (zhdr) {
1110                         if (zhdr->first_chunks == 0) {
1111                                 if (zhdr->middle_chunks != 0 &&
1112                                     chunks >= zhdr->start_middle)
1113                                         bud = LAST;
1114                                 else
1115                                         bud = FIRST;
1116                         } else if (zhdr->last_chunks == 0)
1117                                 bud = LAST;
1118                         else if (zhdr->middle_chunks == 0)
1119                                 bud = MIDDLE;
1120                         else {
1121                                 if (kref_put(&zhdr->refcount,
1122                                              release_z3fold_page_locked))
1123                                         atomic64_dec(&pool->pages_nr);
1124                                 else
1125                                         z3fold_page_unlock(zhdr);
1126                                 pr_err("No free chunks in unbuddied\n");
1127                                 WARN_ON(1);
1128                                 goto retry;
1129                         }
1130                         page = virt_to_page(zhdr);
1131                         goto found;
1132                 }
1133                 bud = FIRST;
1134         }
1135
1136         page = NULL;
1137         if (can_sleep) {
1138                 spin_lock(&pool->stale_lock);
1139                 zhdr = list_first_entry_or_null(&pool->stale,
1140                                                 struct z3fold_header, buddy);
1141                 /*
1142                  * Before allocating a page, let's see if we can take one from
1143                  * the stale pages list. cancel_work_sync() can sleep so we
1144                  * limit this case to the contexts where we can sleep
1145                  */
1146                 if (zhdr) {
1147                         list_del(&zhdr->buddy);
1148                         spin_unlock(&pool->stale_lock);
1149                         cancel_work_sync(&zhdr->work);
1150                         page = virt_to_page(zhdr);
1151                 } else {
1152                         spin_unlock(&pool->stale_lock);
1153                 }
1154         }
1155         if (!page)
1156                 page = alloc_page(gfp);
1157
1158         if (!page)
1159                 return -ENOMEM;
1160
1161         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1162         if (!zhdr) {
1163                 __free_page(page);
1164                 return -ENOMEM;
1165         }
1166         atomic64_inc(&pool->pages_nr);
1167
1168         if (bud == HEADLESS) {
1169                 set_bit(PAGE_HEADLESS, &page->private);
1170                 goto headless;
1171         }
1172         if (can_sleep) {
1173                 lock_page(page);
1174                 __SetPageMovable(page, pool->inode->i_mapping);
1175                 unlock_page(page);
1176         } else {
1177                 if (trylock_page(page)) {
1178                         __SetPageMovable(page, pool->inode->i_mapping);
1179                         unlock_page(page);
1180                 }
1181         }
1182         z3fold_page_lock(zhdr);
1183
1184 found:
1185         if (bud == FIRST)
1186                 zhdr->first_chunks = chunks;
1187         else if (bud == LAST)
1188                 zhdr->last_chunks = chunks;
1189         else {
1190                 zhdr->middle_chunks = chunks;
1191                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1192         }
1193         add_to_unbuddied(pool, zhdr);
1194
1195 headless:
1196         spin_lock(&pool->lock);
1197         /* Add/move z3fold page to beginning of LRU */
1198         if (!list_empty(&page->lru))
1199                 list_del(&page->lru);
1200
1201         list_add(&page->lru, &pool->lru);
1202
1203         *handle = encode_handle(zhdr, bud);
1204         spin_unlock(&pool->lock);
1205         if (bud != HEADLESS)
1206                 z3fold_page_unlock(zhdr);
1207
1208         return 0;
1209 }
1210
1211 /**
1212  * z3fold_free() - frees the allocation associated with the given handle
1213  * @pool:       pool in which the allocation resided
1214  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1215  *
1216  * In the case that the z3fold page in which the allocation resides is under
1217  * reclaim, as indicated by the PG_reclaim flag being set, this function
1218  * only sets the first|last_chunks to 0.  The page is actually freed
1219  * once both buddies are evicted (see z3fold_reclaim_page() below).
1220  */
1221 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1222 {
1223         struct z3fold_header *zhdr;
1224         struct page *page;
1225         enum buddy bud;
1226         bool page_claimed;
1227
1228         zhdr = get_z3fold_header(handle);
1229         page = virt_to_page(zhdr);
1230         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1231
1232         if (test_bit(PAGE_HEADLESS, &page->private)) {
1233                 /* if a headless page is under reclaim, just leave.
1234                  * NB: we use test_and_set_bit for a reason: if the bit
1235                  * has not been set before, we release this page
1236                  * immediately so we don't care about its value any more.
1237                  */
1238                 if (!page_claimed) {
1239                         spin_lock(&pool->lock);
1240                         list_del(&page->lru);
1241                         spin_unlock(&pool->lock);
1242                         put_z3fold_header(zhdr);
1243                         free_z3fold_page(page, true);
1244                         atomic64_dec(&pool->pages_nr);
1245                 }
1246                 return;
1247         }
1248
1249         /* Non-headless case */
1250         bud = handle_to_buddy(handle);
1251
1252         switch (bud) {
1253         case FIRST:
1254                 zhdr->first_chunks = 0;
1255                 break;
1256         case MIDDLE:
1257                 zhdr->middle_chunks = 0;
1258                 break;
1259         case LAST:
1260                 zhdr->last_chunks = 0;
1261                 break;
1262         default:
1263                 pr_err("%s: unknown bud %d\n", __func__, bud);
1264                 WARN_ON(1);
1265                 put_z3fold_header(zhdr);
1266                 clear_bit(PAGE_CLAIMED, &page->private);
1267                 return;
1268         }
1269
1270         if (!page_claimed)
1271                 free_handle(handle);
1272         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1273                 atomic64_dec(&pool->pages_nr);
1274                 return;
1275         }
1276         if (page_claimed) {
1277                 /* the page has not been claimed by us */
1278                 z3fold_page_unlock(zhdr);
1279                 return;
1280         }
1281         if (unlikely(PageIsolated(page)) ||
1282             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1283                 put_z3fold_header(zhdr);
1284                 clear_bit(PAGE_CLAIMED, &page->private);
1285                 return;
1286         }
1287         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1288                 spin_lock(&pool->lock);
1289                 list_del_init(&zhdr->buddy);
1290                 spin_unlock(&pool->lock);
1291                 zhdr->cpu = -1;
1292                 kref_get(&zhdr->refcount);
1293                 clear_bit(PAGE_CLAIMED, &page->private);
1294                 do_compact_page(zhdr, true);
1295                 return;
1296         }
1297         kref_get(&zhdr->refcount);
1298         clear_bit(PAGE_CLAIMED, &page->private);
1299         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1300         put_z3fold_header(zhdr);
1301 }
1302
1303 /**
1304  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1305  * @pool:       pool from which a page will attempt to be evicted
1306  * @retries:    number of pages on the LRU list for which eviction will
1307  *              be attempted before failing
1308  *
1309  * z3fold reclaim is different from normal system reclaim in that it is done
1310  * from the bottom, up. This is because only the bottom layer, z3fold, has
1311  * information on how the allocations are organized within each z3fold page.
1312  * This has the potential to create interesting locking situations between
1313  * z3fold and the user, however.
1314  *
1315  * To avoid these, this is how z3fold_reclaim_page() should be called:
1316  *
1317  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1318  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1319  * call the user-defined eviction handler with the pool and handle as
1320  * arguments.
1321  *
1322  * If the handle can not be evicted, the eviction handler should return
1323  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1324  * appropriate list and try the next z3fold page on the LRU up to
1325  * a user defined number of retries.
1326  *
1327  * If the handle is successfully evicted, the eviction handler should
1328  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1329  * contains logic to delay freeing the page if the page is under reclaim,
1330  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1331  *
1332  * If all buddies in the z3fold page are successfully evicted, then the
1333  * z3fold page can be freed.
1334  *
1335  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1336  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1337  * the retry limit was hit.
1338  */
1339 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1340 {
1341         int i, ret = -1;
1342         struct z3fold_header *zhdr = NULL;
1343         struct page *page = NULL;
1344         struct list_head *pos;
1345         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1346
1347         spin_lock(&pool->lock);
1348         if (!pool->ops || !pool->ops->evict || retries == 0) {
1349                 spin_unlock(&pool->lock);
1350                 return -EINVAL;
1351         }
1352         for (i = 0; i < retries; i++) {
1353                 if (list_empty(&pool->lru)) {
1354                         spin_unlock(&pool->lock);
1355                         return -EINVAL;
1356                 }
1357                 list_for_each_prev(pos, &pool->lru) {
1358                         page = list_entry(pos, struct page, lru);
1359
1360                         /* this bit could have been set by free, in which case
1361                          * we pass over to the next page in the pool.
1362                          */
1363                         if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1364                                 page = NULL;
1365                                 continue;
1366                         }
1367
1368                         if (unlikely(PageIsolated(page))) {
1369                                 clear_bit(PAGE_CLAIMED, &page->private);
1370                                 page = NULL;
1371                                 continue;
1372                         }
1373                         zhdr = page_address(page);
1374                         if (test_bit(PAGE_HEADLESS, &page->private))
1375                                 break;
1376
1377                         if (!z3fold_page_trylock(zhdr)) {
1378                                 clear_bit(PAGE_CLAIMED, &page->private);
1379                                 zhdr = NULL;
1380                                 continue; /* can't evict at this point */
1381                         }
1382                         if (zhdr->foreign_handles) {
1383                                 clear_bit(PAGE_CLAIMED, &page->private);
1384                                 z3fold_page_unlock(zhdr);
1385                                 zhdr = NULL;
1386                                 continue; /* can't evict such page */
1387                         }
1388                         kref_get(&zhdr->refcount);
1389                         list_del_init(&zhdr->buddy);
1390                         zhdr->cpu = -1;
1391                         break;
1392                 }
1393
1394                 if (!zhdr)
1395                         break;
1396
1397                 list_del_init(&page->lru);
1398                 spin_unlock(&pool->lock);
1399
1400                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1401                         /*
1402                          * We need encode the handles before unlocking, and
1403                          * use our local slots structure because z3fold_free
1404                          * can zero out zhdr->slots and we can't do much
1405                          * about that
1406                          */
1407                         first_handle = 0;
1408                         last_handle = 0;
1409                         middle_handle = 0;
1410                         if (zhdr->first_chunks)
1411                                 first_handle = encode_handle(zhdr, FIRST);
1412                         if (zhdr->middle_chunks)
1413                                 middle_handle = encode_handle(zhdr, MIDDLE);
1414                         if (zhdr->last_chunks)
1415                                 last_handle = encode_handle(zhdr, LAST);
1416                         /*
1417                          * it's safe to unlock here because we hold a
1418                          * reference to this page
1419                          */
1420                         z3fold_page_unlock(zhdr);
1421                 } else {
1422                         first_handle = encode_handle(zhdr, HEADLESS);
1423                         last_handle = middle_handle = 0;
1424                 }
1425                 /* Issue the eviction callback(s) */
1426                 if (middle_handle) {
1427                         ret = pool->ops->evict(pool, middle_handle);
1428                         if (ret)
1429                                 goto next;
1430                         free_handle(middle_handle);
1431                 }
1432                 if (first_handle) {
1433                         ret = pool->ops->evict(pool, first_handle);
1434                         if (ret)
1435                                 goto next;
1436                         free_handle(first_handle);
1437                 }
1438                 if (last_handle) {
1439                         ret = pool->ops->evict(pool, last_handle);
1440                         if (ret)
1441                                 goto next;
1442                         free_handle(last_handle);
1443                 }
1444 next:
1445                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1446                         if (ret == 0) {
1447                                 free_z3fold_page(page, true);
1448                                 atomic64_dec(&pool->pages_nr);
1449                                 return 0;
1450                         }
1451                         spin_lock(&pool->lock);
1452                         list_add(&page->lru, &pool->lru);
1453                         spin_unlock(&pool->lock);
1454                         clear_bit(PAGE_CLAIMED, &page->private);
1455                 } else {
1456                         z3fold_page_lock(zhdr);
1457                         if (kref_put(&zhdr->refcount,
1458                                         release_z3fold_page_locked)) {
1459                                 atomic64_dec(&pool->pages_nr);
1460                                 return 0;
1461                         }
1462                         /*
1463                          * if we are here, the page is still not completely
1464                          * free. Take the global pool lock then to be able
1465                          * to add it back to the lru list
1466                          */
1467                         spin_lock(&pool->lock);
1468                         list_add(&page->lru, &pool->lru);
1469                         spin_unlock(&pool->lock);
1470                         z3fold_page_unlock(zhdr);
1471                         clear_bit(PAGE_CLAIMED, &page->private);
1472                 }
1473
1474                 /* We started off locked to we need to lock the pool back */
1475                 spin_lock(&pool->lock);
1476         }
1477         spin_unlock(&pool->lock);
1478         return -EAGAIN;
1479 }
1480
1481 /**
1482  * z3fold_map() - maps the allocation associated with the given handle
1483  * @pool:       pool in which the allocation resides
1484  * @handle:     handle associated with the allocation to be mapped
1485  *
1486  * Extracts the buddy number from handle and constructs the pointer to the
1487  * correct starting chunk within the page.
1488  *
1489  * Returns: a pointer to the mapped allocation
1490  */
1491 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1492 {
1493         struct z3fold_header *zhdr;
1494         struct page *page;
1495         void *addr;
1496         enum buddy buddy;
1497
1498         zhdr = get_z3fold_header(handle);
1499         addr = zhdr;
1500         page = virt_to_page(zhdr);
1501
1502         if (test_bit(PAGE_HEADLESS, &page->private))
1503                 goto out;
1504
1505         buddy = handle_to_buddy(handle);
1506         switch (buddy) {
1507         case FIRST:
1508                 addr += ZHDR_SIZE_ALIGNED;
1509                 break;
1510         case MIDDLE:
1511                 addr += zhdr->start_middle << CHUNK_SHIFT;
1512                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1513                 break;
1514         case LAST:
1515                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1516                 break;
1517         default:
1518                 pr_err("unknown buddy id %d\n", buddy);
1519                 WARN_ON(1);
1520                 addr = NULL;
1521                 break;
1522         }
1523
1524         if (addr)
1525                 zhdr->mapped_count++;
1526 out:
1527         put_z3fold_header(zhdr);
1528         return addr;
1529 }
1530
1531 /**
1532  * z3fold_unmap() - unmaps the allocation associated with the given handle
1533  * @pool:       pool in which the allocation resides
1534  * @handle:     handle associated with the allocation to be unmapped
1535  */
1536 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1537 {
1538         struct z3fold_header *zhdr;
1539         struct page *page;
1540         enum buddy buddy;
1541
1542         zhdr = get_z3fold_header(handle);
1543         page = virt_to_page(zhdr);
1544
1545         if (test_bit(PAGE_HEADLESS, &page->private))
1546                 return;
1547
1548         buddy = handle_to_buddy(handle);
1549         if (buddy == MIDDLE)
1550                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1551         zhdr->mapped_count--;
1552         put_z3fold_header(zhdr);
1553 }
1554
1555 /**
1556  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1557  * @pool:       pool whose size is being queried
1558  *
1559  * Returns: size in pages of the given pool.
1560  */
1561 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1562 {
1563         return atomic64_read(&pool->pages_nr);
1564 }
1565
1566 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1567 {
1568         struct z3fold_header *zhdr;
1569         struct z3fold_pool *pool;
1570
1571         VM_BUG_ON_PAGE(!PageMovable(page), page);
1572         VM_BUG_ON_PAGE(PageIsolated(page), page);
1573
1574         if (test_bit(PAGE_HEADLESS, &page->private) ||
1575             test_bit(PAGE_CLAIMED, &page->private))
1576                 return false;
1577
1578         zhdr = page_address(page);
1579         z3fold_page_lock(zhdr);
1580         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1581             test_bit(PAGE_STALE, &page->private))
1582                 goto out;
1583
1584         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1585                 goto out;
1586
1587         pool = zhdr_to_pool(zhdr);
1588         spin_lock(&pool->lock);
1589         if (!list_empty(&zhdr->buddy))
1590                 list_del_init(&zhdr->buddy);
1591         if (!list_empty(&page->lru))
1592                 list_del_init(&page->lru);
1593         spin_unlock(&pool->lock);
1594
1595         kref_get(&zhdr->refcount);
1596         z3fold_page_unlock(zhdr);
1597         return true;
1598
1599 out:
1600         z3fold_page_unlock(zhdr);
1601         return false;
1602 }
1603
1604 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1605                                struct page *page, enum migrate_mode mode)
1606 {
1607         struct z3fold_header *zhdr, *new_zhdr;
1608         struct z3fold_pool *pool;
1609         struct address_space *new_mapping;
1610
1611         VM_BUG_ON_PAGE(!PageMovable(page), page);
1612         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1613         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1614
1615         zhdr = page_address(page);
1616         pool = zhdr_to_pool(zhdr);
1617
1618         if (!z3fold_page_trylock(zhdr)) {
1619                 return -EAGAIN;
1620         }
1621         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1622                 z3fold_page_unlock(zhdr);
1623                 return -EBUSY;
1624         }
1625         if (work_pending(&zhdr->work)) {
1626                 z3fold_page_unlock(zhdr);
1627                 return -EAGAIN;
1628         }
1629         new_zhdr = page_address(newpage);
1630         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1631         newpage->private = page->private;
1632         page->private = 0;
1633         z3fold_page_unlock(zhdr);
1634         spin_lock_init(&new_zhdr->page_lock);
1635         INIT_WORK(&new_zhdr->work, compact_page_work);
1636         /*
1637          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1638          * so we only have to reinitialize it.
1639          */
1640         INIT_LIST_HEAD(&new_zhdr->buddy);
1641         new_mapping = page_mapping(page);
1642         __ClearPageMovable(page);
1643         ClearPagePrivate(page);
1644
1645         get_page(newpage);
1646         z3fold_page_lock(new_zhdr);
1647         if (new_zhdr->first_chunks)
1648                 encode_handle(new_zhdr, FIRST);
1649         if (new_zhdr->last_chunks)
1650                 encode_handle(new_zhdr, LAST);
1651         if (new_zhdr->middle_chunks)
1652                 encode_handle(new_zhdr, MIDDLE);
1653         set_bit(NEEDS_COMPACTING, &newpage->private);
1654         new_zhdr->cpu = smp_processor_id();
1655         spin_lock(&pool->lock);
1656         list_add(&newpage->lru, &pool->lru);
1657         spin_unlock(&pool->lock);
1658         __SetPageMovable(newpage, new_mapping);
1659         z3fold_page_unlock(new_zhdr);
1660
1661         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1662
1663         page_mapcount_reset(page);
1664         put_page(page);
1665         return 0;
1666 }
1667
1668 static void z3fold_page_putback(struct page *page)
1669 {
1670         struct z3fold_header *zhdr;
1671         struct z3fold_pool *pool;
1672
1673         zhdr = page_address(page);
1674         pool = zhdr_to_pool(zhdr);
1675
1676         z3fold_page_lock(zhdr);
1677         if (!list_empty(&zhdr->buddy))
1678                 list_del_init(&zhdr->buddy);
1679         INIT_LIST_HEAD(&page->lru);
1680         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1681                 atomic64_dec(&pool->pages_nr);
1682                 return;
1683         }
1684         spin_lock(&pool->lock);
1685         list_add(&page->lru, &pool->lru);
1686         spin_unlock(&pool->lock);
1687         z3fold_page_unlock(zhdr);
1688 }
1689
1690 static const struct address_space_operations z3fold_aops = {
1691         .isolate_page = z3fold_page_isolate,
1692         .migratepage = z3fold_page_migrate,
1693         .putback_page = z3fold_page_putback,
1694 };
1695
1696 /*****************
1697  * zpool
1698  ****************/
1699
1700 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1701 {
1702         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1703                 return pool->zpool_ops->evict(pool->zpool, handle);
1704         else
1705                 return -ENOENT;
1706 }
1707
1708 static const struct z3fold_ops z3fold_zpool_ops = {
1709         .evict =        z3fold_zpool_evict
1710 };
1711
1712 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1713                                const struct zpool_ops *zpool_ops,
1714                                struct zpool *zpool)
1715 {
1716         struct z3fold_pool *pool;
1717
1718         pool = z3fold_create_pool(name, gfp,
1719                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1720         if (pool) {
1721                 pool->zpool = zpool;
1722                 pool->zpool_ops = zpool_ops;
1723         }
1724         return pool;
1725 }
1726
1727 static void z3fold_zpool_destroy(void *pool)
1728 {
1729         z3fold_destroy_pool(pool);
1730 }
1731
1732 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1733                         unsigned long *handle)
1734 {
1735         return z3fold_alloc(pool, size, gfp, handle);
1736 }
1737 static void z3fold_zpool_free(void *pool, unsigned long handle)
1738 {
1739         z3fold_free(pool, handle);
1740 }
1741
1742 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1743                         unsigned int *reclaimed)
1744 {
1745         unsigned int total = 0;
1746         int ret = -EINVAL;
1747
1748         while (total < pages) {
1749                 ret = z3fold_reclaim_page(pool, 8);
1750                 if (ret < 0)
1751                         break;
1752                 total++;
1753         }
1754
1755         if (reclaimed)
1756                 *reclaimed = total;
1757
1758         return ret;
1759 }
1760
1761 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1762                         enum zpool_mapmode mm)
1763 {
1764         return z3fold_map(pool, handle);
1765 }
1766 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1767 {
1768         z3fold_unmap(pool, handle);
1769 }
1770
1771 static u64 z3fold_zpool_total_size(void *pool)
1772 {
1773         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1774 }
1775
1776 static struct zpool_driver z3fold_zpool_driver = {
1777         .type =         "z3fold",
1778         .owner =        THIS_MODULE,
1779         .create =       z3fold_zpool_create,
1780         .destroy =      z3fold_zpool_destroy,
1781         .malloc =       z3fold_zpool_malloc,
1782         .free =         z3fold_zpool_free,
1783         .shrink =       z3fold_zpool_shrink,
1784         .map =          z3fold_zpool_map,
1785         .unmap =        z3fold_zpool_unmap,
1786         .total_size =   z3fold_zpool_total_size,
1787 };
1788
1789 MODULE_ALIAS("zpool-z3fold");
1790
1791 static int __init init_z3fold(void)
1792 {
1793         int ret;
1794
1795         /* Make sure the z3fold header is not larger than the page size */
1796         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1797         ret = z3fold_mount();
1798         if (ret)
1799                 return ret;
1800
1801         zpool_register_driver(&z3fold_zpool_driver);
1802
1803         return 0;
1804 }
1805
1806 static void __exit exit_z3fold(void)
1807 {
1808         z3fold_unmount();
1809         zpool_unregister_driver(&z3fold_zpool_driver);
1810 }
1811
1812 module_init(init_z3fold);
1813 module_exit(exit_z3fold);
1814
1815 MODULE_LICENSE("GPL");
1816 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1817 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");