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 + 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_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)
 307 {
 308         struct z3fold_buddy_slots *slots;
 309         struct z3fold_header *zhdr;
 310         int i;
 311         bool is_free;
 312
 313         if (handle & (1 << PAGE_HEADLESS))
 314                 return;
 315
 316         if (WARN_ON(*(unsigned long *)handle == 0))
 317                 return;
 318
 319         zhdr = handle_to_z3fold_header(handle);
 320         slots = handle_to_slots(handle);
 321         write_lock(&slots->lock);
 322         *(unsigned long *)handle = 0;
 323         if (zhdr->slots == slots) {
 324                 write_unlock(&slots->lock);
 325                 return; /* simple case, nothing else to do */
 326         }
 327
 328         /* we are freeing a foreign handle if we are here */
 329         zhdr->foreign_handles--;
 330         is_free = true;
 331         if (!test_bit(HANDLES_ORPHANED, &slots->pool)) {
 332                 write_unlock(&slots->lock);
 333                 return;
 334         }
 335         for (i = 0; i <= BUDDY_MASK; i++) {
 336                 if (slots->slot[i]) {
 337                         is_free = false;
 338                         break;
 339                 }
 340         }
 341         write_unlock(&slots->lock);
 342
 343         if (is_free) {
 344                 struct z3fold_pool *pool = slots_to_pool(slots);
 345
 346                 kmem_cache_free(pool->c_handle, slots);
 347         }
 348 }
 349
 350 static int z3fold_init_fs_context(struct fs_context *fc)
 351 {
 352         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 353 }
 354
 355 static struct file_system_type z3fold_fs = {
 356         .name           = "z3fold",
 357         .init_fs_context = z3fold_init_fs_context,
 358         .kill_sb        = kill_anon_super,
 359 };
 360
 361 static struct vfsmount *z3fold_mnt;
 362 static int z3fold_mount(void)
 363 {
 364         int ret = 0;
 365
 366         z3fold_mnt = kern_mount(&z3fold_fs);
 367         if (IS_ERR(z3fold_mnt))
 368                 ret = PTR_ERR(z3fold_mnt);
 369
 370         return ret;
 371 }
 372
 373 static void z3fold_unmount(void)
 374 {
 375         kern_unmount(z3fold_mnt);
 376 }
 377
 378 static const struct address_space_operations z3fold_aops;
 379 static int z3fold_register_migration(struct z3fold_pool *pool)
 380 {
 381         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 382         if (IS_ERR(pool->inode)) {
 383                 pool->inode = NULL;
 384                 return 1;
 385         }
 386
 387         pool->inode->i_mapping->private_data = pool;
 388         pool->inode->i_mapping->a_ops = &z3fold_aops;
 389         return 0;
 390 }
 391
 392 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 393 {
 394         if (pool->inode)
 395                 iput(pool->inode);
 396  }
 397
 398 /* Initializes the z3fold header of a newly allocated z3fold page */
 399 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 400                                         struct z3fold_pool *pool, gfp_t gfp)
 401 {
 402         struct z3fold_header *zhdr = page_address(page);
 403         struct z3fold_buddy_slots *slots;
 404
 405         INIT_LIST_HEAD(&page->lru);
 406         clear_bit(PAGE_HEADLESS, &page->private);
 407         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 408         clear_bit(NEEDS_COMPACTING, &page->private);
 409         clear_bit(PAGE_STALE, &page->private);
 410         clear_bit(PAGE_CLAIMED, &page->private);
 411         if (headless)
 412                 return zhdr;
 413
 414         slots = alloc_slots(pool, gfp);
 415         if (!slots)
 416                 return NULL;
 417
 418         spin_lock_init(&zhdr->page_lock);
 419         kref_init(&zhdr->refcount);
 420         zhdr->first_chunks = 0;
 421         zhdr->middle_chunks = 0;
 422         zhdr->last_chunks = 0;
 423         zhdr->first_num = 0;
 424         zhdr->start_middle = 0;
 425         zhdr->cpu = -1;
 426         zhdr->foreign_handles = 0;
 427         zhdr->mapped_count = 0;
 428         zhdr->slots = slots;
 429         zhdr->pool = pool;
 430         INIT_LIST_HEAD(&zhdr->buddy);
 431         INIT_WORK(&zhdr->work, compact_page_work);
 432         return zhdr;
 433 }
 434
 435 /* Resets the struct page fields and frees the page */
 436 static void free_z3fold_page(struct page *page, bool headless)
 437 {
 438         if (!headless) {
 439                 lock_page(page);
 440                 __ClearPageMovable(page);
 441                 unlock_page(page);
 442         }
 443         ClearPagePrivate(page);
 444         __free_page(page);
 445 }
 446
 447 /* Helper function to build the index */
 448 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 449 {
 450         return (bud + zhdr->first_num) & BUDDY_MASK;
 451 }
 452
 453 /*
 454  * Encodes the handle of a particular buddy within a z3fold page
 455  * Pool lock should be held as this function accesses first_num
 456  */
 457 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 458                                 struct z3fold_buddy_slots *slots,
 459                                 enum buddy bud)
 460 {
 461         unsigned long h = (unsigned long)zhdr;
 462         int idx = 0;
 463
 464         /*
 465          * For a headless page, its handle is its pointer with the extra
 466          * PAGE_HEADLESS bit set
 467          */
 468         if (bud == HEADLESS)
 469                 return h | (1 << PAGE_HEADLESS);
 470
 471         /* otherwise, return pointer to encoded handle */
 472         idx = __idx(zhdr, bud);
 473         h += idx;
 474         if (bud == LAST)
 475                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 476
 477         write_lock(&slots->lock);
 478         slots->slot[idx] = h;
 479         write_unlock(&slots->lock);
 480         return (unsigned long)&slots->slot[idx];
 481 }
 482
 483 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 484 {
 485         return __encode_handle(zhdr, zhdr->slots, bud);
 486 }
 487
 488 /* only for LAST bud, returns zero otherwise */
 489 static unsigned short handle_to_chunks(unsigned long handle)
 490 {
 491         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 492         unsigned long addr;
 493
 494         read_lock(&slots->lock);
 495         addr = *(unsigned long *)handle;
 496         read_unlock(&slots->lock);
 497         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 498 }
 499
 500 /*
 501  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 502  *  but that doesn't matter. because the masking will result in the
 503  *  correct buddy number.
 504  */
 505 static enum buddy handle_to_buddy(unsigned long handle)
 506 {
 507         struct z3fold_header *zhdr;
 508         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 509         unsigned long addr;
 510
 511         read_lock(&slots->lock);
 512         WARN_ON(handle & (1 << PAGE_HEADLESS));
 513         addr = *(unsigned long *)handle;
 514         read_unlock(&slots->lock);
 515         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 516         return (addr - zhdr->first_num) & BUDDY_MASK;
 517 }
 518
 519 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 520 {
 521         return zhdr->pool;
 522 }
 523
 524 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 525 {
 526         struct page *page = virt_to_page(zhdr);
 527         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 528         bool is_free = true;
 529         int i;
 530
 531         WARN_ON(!list_empty(&zhdr->buddy));
 532         set_bit(PAGE_STALE, &page->private);
 533         clear_bit(NEEDS_COMPACTING, &page->private);
 534         spin_lock(&pool->lock);
 535         if (!list_empty(&page->lru))
 536                 list_del_init(&page->lru);
 537         spin_unlock(&pool->lock);
 538
 539         /* If there are no foreign handles, free the handles array */
 540         read_lock(&zhdr->slots->lock);
 541         for (i = 0; i <= BUDDY_MASK; i++) {
 542                 if (zhdr->slots->slot[i]) {
 543                         is_free = false;
 544                         break;
 545                 }
 546         }
 547         if (!is_free)
 548                 set_bit(HANDLES_ORPHANED, &zhdr->slots->pool);
 549         read_unlock(&zhdr->slots->lock);
 550
 551         if (is_free)
 552                 kmem_cache_free(pool->c_handle, zhdr->slots);
 553
 554         if (locked)
 555                 z3fold_page_unlock(zhdr);
 556
 557         spin_lock(&pool->stale_lock);
 558         list_add(&zhdr->buddy, &pool->stale);
 559         queue_work(pool->release_wq, &pool->work);
 560         spin_unlock(&pool->stale_lock);
 561 }
 562
 563 static void __attribute__((__unused__))
 564                         release_z3fold_page(struct kref *ref)
 565 {
 566         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 567                                                 refcount);
 568         __release_z3fold_page(zhdr, false);
 569 }
 570
 571 static void release_z3fold_page_locked(struct kref *ref)
 572 {
 573         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 574                                                 refcount);
 575         WARN_ON(z3fold_page_trylock(zhdr));
 576         __release_z3fold_page(zhdr, true);
 577 }
 578
 579 static void release_z3fold_page_locked_list(struct kref *ref)
 580 {
 581         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 582                                                refcount);
 583         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 584
 585         spin_lock(&pool->lock);
 586         list_del_init(&zhdr->buddy);
 587         spin_unlock(&pool->lock);
 588
 589         WARN_ON(z3fold_page_trylock(zhdr));
 590         __release_z3fold_page(zhdr, true);
 591 }
 592
 593 static void free_pages_work(struct work_struct *w)
 594 {
 595         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 596
 597         spin_lock(&pool->stale_lock);
 598         while (!list_empty(&pool->stale)) {
 599                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 600                                                 struct z3fold_header, buddy);
 601                 struct page *page = virt_to_page(zhdr);
 602
 603                 list_del(&zhdr->buddy);
 604                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 605                         continue;
 606                 spin_unlock(&pool->stale_lock);
 607                 cancel_work_sync(&zhdr->work);
 608                 free_z3fold_page(page, false);
 609                 cond_resched();
 610                 spin_lock(&pool->stale_lock);
 611         }
 612         spin_unlock(&pool->stale_lock);
 613 }
 614
 615 /*
 616  * Returns the number of free chunks in a z3fold page.
 617  * NB: can't be used with HEADLESS pages.
 618  */
 619 static int num_free_chunks(struct z3fold_header *zhdr)
 620 {
 621         int nfree;
 622         /*
 623          * If there is a middle object, pick up the bigger free space
 624          * either before or after it. Otherwise just subtract the number
 625          * of chunks occupied by the first and the last objects.
 626          */
 627         if (zhdr->middle_chunks != 0) {
 628                 int nfree_before = zhdr->first_chunks ?
 629                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 630                 int nfree_after = zhdr->last_chunks ?
 631                         0 : TOTAL_CHUNKS -
 632                                 (zhdr->start_middle + zhdr->middle_chunks);
 633                 nfree = max(nfree_before, nfree_after);
 634         } else
 635                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 636         return nfree;
 637 }
 638
 639 /* Add to the appropriate unbuddied list */
 640 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 641                                 struct z3fold_header *zhdr)
 642 {
 643         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 644                         zhdr->middle_chunks == 0) {
 645                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 646
 647                 int freechunks = num_free_chunks(zhdr);
 648                 spin_lock(&pool->lock);
 649                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 650                 spin_unlock(&pool->lock);
 651                 zhdr->cpu = smp_processor_id();
 652                 put_cpu_ptr(pool->unbuddied);
 653         }
 654 }
 655
 656 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 657                                 unsigned short dst_chunk)
 658 {
 659         void *beg = zhdr;
 660         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 661                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 662                        zhdr->middle_chunks << CHUNK_SHIFT);
 663 }
 664
 665 static inline bool buddy_single(struct z3fold_header *zhdr)
 666 {
 667         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 668                         (zhdr->first_chunks && zhdr->last_chunks) ||
 669                         (zhdr->middle_chunks && zhdr->last_chunks));
 670 }
 671
 672 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 673 {
 674         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 675         void *p = zhdr;
 676         unsigned long old_handle = 0;
 677         size_t sz = 0;
 678         struct z3fold_header *new_zhdr = NULL;
 679         int first_idx = __idx(zhdr, FIRST);
 680         int middle_idx = __idx(zhdr, MIDDLE);
 681         int last_idx = __idx(zhdr, LAST);
 682         unsigned short *moved_chunks = NULL;
 683
 684         /*
 685          * No need to protect slots here -- all the slots are "local" and
 686          * the page lock is already taken
 687          */
 688         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 689                 p += ZHDR_SIZE_ALIGNED;
 690                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 691                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 692                 moved_chunks = &zhdr->first_chunks;
 693         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 694                 p += zhdr->start_middle << CHUNK_SHIFT;
 695                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 696                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 697                 moved_chunks = &zhdr->middle_chunks;
 698         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 699                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 700                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 701                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 702                 moved_chunks = &zhdr->last_chunks;
 703         }
 704
 705         if (sz > 0) {
 706                 enum buddy new_bud = HEADLESS;
 707                 short chunks = size_to_chunks(sz);
 708                 void *q;
 709
 710                 new_zhdr = __z3fold_alloc(pool, sz, false);
 711                 if (!new_zhdr)
 712                         return NULL;
 713
 714                 if (WARN_ON(new_zhdr == zhdr))
 715                         goto out_fail;
 716
 717                 if (new_zhdr->first_chunks == 0) {
 718                         if (new_zhdr->middle_chunks != 0 &&
 719                                         chunks >= new_zhdr->start_middle) {
 720                                 new_bud = LAST;
 721                         } else {
 722                                 new_bud = FIRST;
 723                         }
 724                 } else if (new_zhdr->last_chunks == 0) {
 725                         new_bud = LAST;
 726                 } else if (new_zhdr->middle_chunks == 0) {
 727                         new_bud = MIDDLE;
 728                 }
 729                 q = new_zhdr;
 730                 switch (new_bud) {
 731                 case FIRST:
 732                         new_zhdr->first_chunks = chunks;
 733                         q += ZHDR_SIZE_ALIGNED;
 734                         break;
 735                 case MIDDLE:
 736                         new_zhdr->middle_chunks = chunks;
 737                         new_zhdr->start_middle =
 738                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 739                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 740                         break;
 741                 case LAST:
 742                         new_zhdr->last_chunks = chunks;
 743                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 744                         break;
 745                 default:
 746                         goto out_fail;
 747                 }
 748                 new_zhdr->foreign_handles++;
 749                 memcpy(q, p, sz);
 750                 write_lock(&zhdr->slots->lock);
 751                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 752                         __idx(new_zhdr, new_bud);
 753                 if (new_bud == LAST)
 754                         *(unsigned long *)old_handle |=
 755                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 756                 write_unlock(&zhdr->slots->lock);
 757                 add_to_unbuddied(pool, new_zhdr);
 758                 z3fold_page_unlock(new_zhdr);
 759
 760                 *moved_chunks = 0;
 761         }
 762
 763         return new_zhdr;
 764
 765 out_fail:
 766         if (new_zhdr) {
 767                 if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
 768                         atomic64_dec(&pool->pages_nr);
 769                 else {
 770                         add_to_unbuddied(pool, new_zhdr);
 771                         z3fold_page_unlock(new_zhdr);
 772                 }
 773         }
 774         return NULL;
 775
 776 }
 777
 778 #define BIG_CHUNK_GAP   3
 779 /* Has to be called with lock held */
 780 static int z3fold_compact_page(struct z3fold_header *zhdr)
 781 {
 782         struct page *page = virt_to_page(zhdr);
 783
 784         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 785                 return 0; /* can't move middle chunk, it's used */
 786
 787         if (unlikely(PageIsolated(page)))
 788                 return 0;
 789
 790         if (zhdr->middle_chunks == 0)
 791                 return 0; /* nothing to compact */
 792
 793         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 794                 /* move to the beginning */
 795                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 796                 zhdr->first_chunks = zhdr->middle_chunks;
 797                 zhdr->middle_chunks = 0;
 798                 zhdr->start_middle = 0;
 799                 zhdr->first_num++;
 800                 return 1;
 801         }
 802
 803         /*
 804          * moving data is expensive, so let's only do that if
 805          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 806          */
 807         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 808             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 809                         BIG_CHUNK_GAP) {
 810                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 811                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 812                 return 1;
 813         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 814                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 815                                         + zhdr->middle_chunks) >=
 816                         BIG_CHUNK_GAP) {
 817                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 818                         zhdr->middle_chunks;
 819                 mchunk_memmove(zhdr, new_start);
 820                 zhdr->start_middle = new_start;
 821                 return 1;
 822         }
 823
 824         return 0;
 825 }
 826
 827 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 828 {
 829         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 830         struct page *page;
 831
 832         page = virt_to_page(zhdr);
 833         if (locked)
 834                 WARN_ON(z3fold_page_trylock(zhdr));
 835         else
 836                 z3fold_page_lock(zhdr);
 837         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 838                 z3fold_page_unlock(zhdr);
 839                 return;
 840         }
 841         spin_lock(&pool->lock);
 842         list_del_init(&zhdr->buddy);
 843         spin_unlock(&pool->lock);
 844
 845         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 846                 atomic64_dec(&pool->pages_nr);
 847                 return;
 848         }
 849
 850         if (unlikely(PageIsolated(page) ||
 851                      test_bit(PAGE_CLAIMED, &page->private) ||
 852                      test_bit(PAGE_STALE, &page->private))) {
 853                 z3fold_page_unlock(zhdr);
 854                 return;
 855         }
 856
 857         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 858             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 859                 if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 860                         atomic64_dec(&pool->pages_nr);
 861                 else
 862                         z3fold_page_unlock(zhdr);
 863                 return;
 864         }
 865
 866         z3fold_compact_page(zhdr);
 867         add_to_unbuddied(pool, zhdr);
 868         z3fold_page_unlock(zhdr);
 869 }
 870
 871 static void compact_page_work(struct work_struct *w)
 872 {
 873         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 874                                                 work);
 875
 876         do_compact_page(zhdr, false);
 877 }
 878
 879 /* returns _locked_ z3fold page header or NULL */
 880 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 881                                                 size_t size, bool can_sleep)
 882 {
 883         struct z3fold_header *zhdr = NULL;
 884         struct page *page;
 885         struct list_head *unbuddied;
 886         int chunks = size_to_chunks(size), i;
 887
 888 lookup:
 889         /* First, try to find an unbuddied z3fold page. */
 890         unbuddied = get_cpu_ptr(pool->unbuddied);
 891         for_each_unbuddied_list(i, chunks) {
 892                 struct list_head *l = &unbuddied[i];
 893
 894                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 895                                         struct z3fold_header, buddy);
 896
 897                 if (!zhdr)
 898                         continue;
 899
 900                 /* Re-check under lock. */
 901                 spin_lock(&pool->lock);
 902                 l = &unbuddied[i];
 903                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 904                                                 struct z3fold_header, buddy)) ||
 905                     !z3fold_page_trylock(zhdr)) {
 906                         spin_unlock(&pool->lock);
 907                         zhdr = NULL;
 908                         put_cpu_ptr(pool->unbuddied);
 909                         if (can_sleep)
 910                                 cond_resched();
 911                         goto lookup;
 912                 }
 913                 list_del_init(&zhdr->buddy);
 914                 zhdr->cpu = -1;
 915                 spin_unlock(&pool->lock);
 916
 917                 page = virt_to_page(zhdr);
 918                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 919                     test_bit(PAGE_CLAIMED, &page->private)) {
 920                         z3fold_page_unlock(zhdr);
 921                         zhdr = NULL;
 922                         put_cpu_ptr(pool->unbuddied);
 923                         if (can_sleep)
 924                                 cond_resched();
 925                         goto lookup;
 926                 }
 927
 928                 /*
 929                  * this page could not be removed from its unbuddied
 930                  * list while pool lock was held, and then we've taken
 931                  * page lock so kref_put could not be called before
 932                  * we got here, so it's safe to just call kref_get()
 933                  */
 934                 kref_get(&zhdr->refcount);
 935                 break;
 936         }
 937         put_cpu_ptr(pool->unbuddied);
 938
 939         if (!zhdr) {
 940                 int cpu;
 941
 942                 /* look for _exact_ match on other cpus' lists */
 943                 for_each_online_cpu(cpu) {
 944                         struct list_head *l;
 945
 946                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 947                         spin_lock(&pool->lock);
 948                         l = &unbuddied[chunks];
 949
 950                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 951                                                 struct z3fold_header, buddy);
 952
 953                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 954                                 spin_unlock(&pool->lock);
 955                                 zhdr = NULL;
 956                                 continue;
 957                         }
 958                         list_del_init(&zhdr->buddy);
 959                         zhdr->cpu = -1;
 960                         spin_unlock(&pool->lock);
 961
 962                         page = virt_to_page(zhdr);
 963                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 964                             test_bit(PAGE_CLAIMED, &page->private)) {
 965                                 z3fold_page_unlock(zhdr);
 966                                 zhdr = NULL;
 967                                 if (can_sleep)
 968                                         cond_resched();
 969                                 continue;
 970                         }
 971                         kref_get(&zhdr->refcount);
 972                         break;
 973                 }
 974         }
 975
 976         return zhdr;
 977 }
 978
 979 /*
 980  * API Functions
 981  */
 982
 983 /**
 984  * z3fold_create_pool() - create a new z3fold pool
 985  * @name:       pool name
 986  * @gfp:        gfp flags when allocating the z3fold pool structure
 987  * @ops:        user-defined operations for the z3fold pool
 988  *
 989  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 990  * failed.
 991  */
 992 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 993                 const struct z3fold_ops *ops)
 994 {
 995         struct z3fold_pool *pool = NULL;
 996         int i, cpu;
 997
 998         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 999         if (!pool)
1000                 goto out;
1001         pool->c_handle = kmem_cache_create("z3fold_handle",
1002                                 sizeof(struct z3fold_buddy_slots),
1003                                 SLOTS_ALIGN, 0, NULL);
1004         if (!pool->c_handle)
1005                 goto out_c;
1006         spin_lock_init(&pool->lock);
1007         spin_lock_init(&pool->stale_lock);
1008         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
1009         if (!pool->unbuddied)
1010                 goto out_pool;
1011         for_each_possible_cpu(cpu) {
1012                 struct list_head *unbuddied =
1013                                 per_cpu_ptr(pool->unbuddied, cpu);
1014                 for_each_unbuddied_list(i, 0)
1015                         INIT_LIST_HEAD(&unbuddied[i]);
1016         }
1017         INIT_LIST_HEAD(&pool->lru);
1018         INIT_LIST_HEAD(&pool->stale);
1019         atomic64_set(&pool->pages_nr, 0);
1020         pool->name = name;
1021         pool->compact_wq = create_singlethread_workqueue(pool->name);
1022         if (!pool->compact_wq)
1023                 goto out_unbuddied;
1024         pool->release_wq = create_singlethread_workqueue(pool->name);
1025         if (!pool->release_wq)
1026                 goto out_wq;
1027         if (z3fold_register_migration(pool))
1028                 goto out_rwq;
1029         INIT_WORK(&pool->work, free_pages_work);
1030         pool->ops = ops;
1031         return pool;
1032
1033 out_rwq:
1034         destroy_workqueue(pool->release_wq);
1035 out_wq:
1036         destroy_workqueue(pool->compact_wq);
1037 out_unbuddied:
1038         free_percpu(pool->unbuddied);
1039 out_pool:
1040         kmem_cache_destroy(pool->c_handle);
1041 out_c:
1042         kfree(pool);
1043 out:
1044         return NULL;
1045 }
1046
1047 /**
1048  * z3fold_destroy_pool() - destroys an existing z3fold pool
1049  * @pool:       the z3fold pool to be destroyed
1050  *
1051  * The pool should be emptied before this function is called.
1052  */
1053 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1054 {
1055         kmem_cache_destroy(pool->c_handle);
1056
1057         /*
1058          * We need to destroy pool->compact_wq before pool->release_wq,
1059          * as any pending work on pool->compact_wq will call
1060          * queue_work(pool->release_wq, &pool->work).
1061          *
1062          * There are still outstanding pages until both workqueues are drained,
1063          * so we cannot unregister migration until then.
1064          */
1065
1066         destroy_workqueue(pool->compact_wq);
1067         destroy_workqueue(pool->release_wq);
1068         z3fold_unregister_migration(pool);
1069         kfree(pool);
1070 }
1071
1072 /**
1073  * z3fold_alloc() - allocates a region of a given size
1074  * @pool:       z3fold pool from which to allocate
1075  * @size:       size in bytes of the desired allocation
1076  * @gfp:        gfp flags used if the pool needs to grow
1077  * @handle:     handle of the new allocation
1078  *
1079  * This function will attempt to find a free region in the pool large enough to
1080  * satisfy the allocation request.  A search of the unbuddied lists is
1081  * performed first. If no suitable free region is found, then a new page is
1082  * allocated and added to the pool to satisfy the request.
1083  *
1084  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
1085  * as z3fold pool pages.
1086  *
1087  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1088  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1089  * a new page.
1090  */
1091 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1092                         unsigned long *handle)
1093 {
1094         int chunks = size_to_chunks(size);
1095         struct z3fold_header *zhdr = NULL;
1096         struct page *page = NULL;
1097         enum buddy bud;
1098         bool can_sleep = gfpflags_allow_blocking(gfp);
1099
1100         if (!size)
1101                 return -EINVAL;
1102
1103         if (size > PAGE_SIZE)
1104                 return -ENOSPC;
1105
1106         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1107                 bud = HEADLESS;
1108         else {
1109 retry:
1110                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1111                 if (zhdr) {
1112                         if (zhdr->first_chunks == 0) {
1113                                 if (zhdr->middle_chunks != 0 &&
1114                                     chunks >= zhdr->start_middle)
1115                                         bud = LAST;
1116                                 else
1117                                         bud = FIRST;
1118                         } else if (zhdr->last_chunks == 0)
1119                                 bud = LAST;
1120                         else if (zhdr->middle_chunks == 0)
1121                                 bud = MIDDLE;
1122                         else {
1123                                 if (kref_put(&zhdr->refcount,
1124                                              release_z3fold_page_locked))
1125                                         atomic64_dec(&pool->pages_nr);
1126                                 else
1127                                         z3fold_page_unlock(zhdr);
1128                                 pr_err("No free chunks in unbuddied\n");
1129                                 WARN_ON(1);
1130                                 goto retry;
1131                         }
1132                         page = virt_to_page(zhdr);
1133                         goto found;
1134                 }
1135                 bud = FIRST;
1136         }
1137
1138         page = NULL;
1139         if (can_sleep) {
1140                 spin_lock(&pool->stale_lock);
1141                 zhdr = list_first_entry_or_null(&pool->stale,
1142                                                 struct z3fold_header, buddy);
1143                 /*
1144                  * Before allocating a page, let's see if we can take one from
1145                  * the stale pages list. cancel_work_sync() can sleep so we
1146                  * limit this case to the contexts where we can sleep
1147                  */
1148                 if (zhdr) {
1149                         list_del(&zhdr->buddy);
1150                         spin_unlock(&pool->stale_lock);
1151                         cancel_work_sync(&zhdr->work);
1152                         page = virt_to_page(zhdr);
1153                 } else {
1154                         spin_unlock(&pool->stale_lock);
1155                 }
1156         }
1157         if (!page)
1158                 page = alloc_page(gfp);
1159
1160         if (!page)
1161                 return -ENOMEM;
1162
1163         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1164         if (!zhdr) {
1165                 __free_page(page);
1166                 return -ENOMEM;
1167         }
1168         atomic64_inc(&pool->pages_nr);
1169
1170         if (bud == HEADLESS) {
1171                 set_bit(PAGE_HEADLESS, &page->private);
1172                 goto headless;
1173         }
1174         if (can_sleep) {
1175                 lock_page(page);
1176                 __SetPageMovable(page, pool->inode->i_mapping);
1177                 unlock_page(page);
1178         } else {
1179                 if (trylock_page(page)) {
1180                         __SetPageMovable(page, pool->inode->i_mapping);
1181                         unlock_page(page);
1182                 }
1183         }
1184         z3fold_page_lock(zhdr);
1185
1186 found:
1187         if (bud == FIRST)
1188                 zhdr->first_chunks = chunks;
1189         else if (bud == LAST)
1190                 zhdr->last_chunks = chunks;
1191         else {
1192                 zhdr->middle_chunks = chunks;
1193                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1194         }
1195         add_to_unbuddied(pool, zhdr);
1196
1197 headless:
1198         spin_lock(&pool->lock);
1199         /* Add/move z3fold page to beginning of LRU */
1200         if (!list_empty(&page->lru))
1201                 list_del(&page->lru);
1202
1203         list_add(&page->lru, &pool->lru);
1204
1205         *handle = encode_handle(zhdr, bud);
1206         spin_unlock(&pool->lock);
1207         if (bud != HEADLESS)
1208                 z3fold_page_unlock(zhdr);
1209
1210         return 0;
1211 }
1212
1213 /**
1214  * z3fold_free() - frees the allocation associated with the given handle
1215  * @pool:       pool in which the allocation resided
1216  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1217  *
1218  * In the case that the z3fold page in which the allocation resides is under
1219  * reclaim, as indicated by the PG_reclaim flag being set, this function
1220  * only sets the first|last_chunks to 0.  The page is actually freed
1221  * once both buddies are evicted (see z3fold_reclaim_page() below).
1222  */
1223 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1224 {
1225         struct z3fold_header *zhdr;
1226         struct page *page;
1227         enum buddy bud;
1228         bool page_claimed;
1229
1230         zhdr = get_z3fold_header(handle);
1231         page = virt_to_page(zhdr);
1232         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1233
1234         if (test_bit(PAGE_HEADLESS, &page->private)) {
1235                 /* if a headless page is under reclaim, just leave.
1236                  * NB: we use test_and_set_bit for a reason: if the bit
1237                  * has not been set before, we release this page
1238                  * immediately so we don't care about its value any more.
1239                  */
1240                 if (!page_claimed) {
1241                         spin_lock(&pool->lock);
1242                         list_del(&page->lru);
1243                         spin_unlock(&pool->lock);
1244                         put_z3fold_header(zhdr);
1245                         free_z3fold_page(page, true);
1246                         atomic64_dec(&pool->pages_nr);
1247                 }
1248                 return;
1249         }
1250
1251         /* Non-headless case */
1252         bud = handle_to_buddy(handle);
1253
1254         switch (bud) {
1255         case FIRST:
1256                 zhdr->first_chunks = 0;
1257                 break;
1258         case MIDDLE:
1259                 zhdr->middle_chunks = 0;
1260                 break;
1261         case LAST:
1262                 zhdr->last_chunks = 0;
1263                 break;
1264         default:
1265                 pr_err("%s: unknown bud %d\n", __func__, bud);
1266                 WARN_ON(1);
1267                 put_z3fold_header(zhdr);
1268                 clear_bit(PAGE_CLAIMED, &page->private);
1269                 return;
1270         }
1271
1272         if (!page_claimed)
1273                 free_handle(handle);
1274         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1275                 atomic64_dec(&pool->pages_nr);
1276                 return;
1277         }
1278         if (page_claimed) {
1279                 /* the page has not been claimed by us */
1280                 z3fold_page_unlock(zhdr);
1281                 return;
1282         }
1283         if (unlikely(PageIsolated(page)) ||
1284             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1285                 put_z3fold_header(zhdr);
1286                 clear_bit(PAGE_CLAIMED, &page->private);
1287                 return;
1288         }
1289         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1290                 spin_lock(&pool->lock);
1291                 list_del_init(&zhdr->buddy);
1292                 spin_unlock(&pool->lock);
1293                 zhdr->cpu = -1;
1294                 kref_get(&zhdr->refcount);
1295                 clear_bit(PAGE_CLAIMED, &page->private);
1296                 do_compact_page(zhdr, true);
1297                 return;
1298         }
1299         kref_get(&zhdr->refcount);
1300         clear_bit(PAGE_CLAIMED, &page->private);
1301         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1302         put_z3fold_header(zhdr);
1303 }
1304
1305 /**
1306  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1307  * @pool:       pool from which a page will attempt to be evicted
1308  * @retries:    number of pages on the LRU list for which eviction will
1309  *              be attempted before failing
1310  *
1311  * z3fold reclaim is different from normal system reclaim in that it is done
1312  * from the bottom, up. This is because only the bottom layer, z3fold, has
1313  * information on how the allocations are organized within each z3fold page.
1314  * This has the potential to create interesting locking situations between
1315  * z3fold and the user, however.
1316  *
1317  * To avoid these, this is how z3fold_reclaim_page() should be called:
1318  *
1319  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1320  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1321  * call the user-defined eviction handler with the pool and handle as
1322  * arguments.
1323  *
1324  * If the handle can not be evicted, the eviction handler should return
1325  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1326  * appropriate list and try the next z3fold page on the LRU up to
1327  * a user defined number of retries.
1328  *
1329  * If the handle is successfully evicted, the eviction handler should
1330  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1331  * contains logic to delay freeing the page if the page is under reclaim,
1332  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1333  *
1334  * If all buddies in the z3fold page are successfully evicted, then the
1335  * z3fold page can be freed.
1336  *
1337  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1338  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1339  * the retry limit was hit.
1340  */
1341 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1342 {
1343         int i, ret = -1;
1344         struct z3fold_header *zhdr = NULL;
1345         struct page *page = NULL;
1346         struct list_head *pos;
1347         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1348
1349         spin_lock(&pool->lock);
1350         if (!pool->ops || !pool->ops->evict || retries == 0) {
1351                 spin_unlock(&pool->lock);
1352                 return -EINVAL;
1353         }
1354         for (i = 0; i < retries; i++) {
1355                 if (list_empty(&pool->lru)) {
1356                         spin_unlock(&pool->lock);
1357                         return -EINVAL;
1358                 }
1359                 list_for_each_prev(pos, &pool->lru) {
1360                         page = list_entry(pos, struct page, lru);
1361
1362                         /* this bit could have been set by free, in which case
1363                          * we pass over to the next page in the pool.
1364                          */
1365                         if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1366                                 page = NULL;
1367                                 continue;
1368                         }
1369
1370                         if (unlikely(PageIsolated(page))) {
1371                                 clear_bit(PAGE_CLAIMED, &page->private);
1372                                 page = NULL;
1373                                 continue;
1374                         }
1375                         zhdr = page_address(page);
1376                         if (test_bit(PAGE_HEADLESS, &page->private))
1377                                 break;
1378
1379                         if (!z3fold_page_trylock(zhdr)) {
1380                                 clear_bit(PAGE_CLAIMED, &page->private);
1381                                 zhdr = NULL;
1382                                 continue; /* can't evict at this point */
1383                         }
1384                         if (zhdr->foreign_handles) {
1385                                 clear_bit(PAGE_CLAIMED, &page->private);
1386                                 z3fold_page_unlock(zhdr);
1387                                 zhdr = NULL;
1388                                 continue; /* can't evict such page */
1389                         }
1390                         kref_get(&zhdr->refcount);
1391                         list_del_init(&zhdr->buddy);
1392                         zhdr->cpu = -1;
1393                         break;
1394                 }
1395
1396                 if (!zhdr)
1397                         break;
1398
1399                 list_del_init(&page->lru);
1400                 spin_unlock(&pool->lock);
1401
1402                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1403                         /*
1404                          * We need encode the handles before unlocking, and
1405                          * use our local slots structure because z3fold_free
1406                          * can zero out zhdr->slots and we can't do much
1407                          * about that
1408                          */
1409                         first_handle = 0;
1410                         last_handle = 0;
1411                         middle_handle = 0;
1412                         if (zhdr->first_chunks)
1413                                 first_handle = encode_handle(zhdr, FIRST);
1414                         if (zhdr->middle_chunks)
1415                                 middle_handle = encode_handle(zhdr, MIDDLE);
1416                         if (zhdr->last_chunks)
1417                                 last_handle = encode_handle(zhdr, LAST);
1418                         /*
1419                          * it's safe to unlock here because we hold a
1420                          * reference to this page
1421                          */
1422                         z3fold_page_unlock(zhdr);
1423                 } else {
1424                         first_handle = encode_handle(zhdr, HEADLESS);
1425                         last_handle = middle_handle = 0;
1426                 }
1427                 /* Issue the eviction callback(s) */
1428                 if (middle_handle) {
1429                         ret = pool->ops->evict(pool, middle_handle);
1430                         if (ret)
1431                                 goto next;
1432                         free_handle(middle_handle);
1433                 }
1434                 if (first_handle) {
1435                         ret = pool->ops->evict(pool, first_handle);
1436                         if (ret)
1437                                 goto next;
1438                         free_handle(first_handle);
1439                 }
1440                 if (last_handle) {
1441                         ret = pool->ops->evict(pool, last_handle);
1442                         if (ret)
1443                                 goto next;
1444                         free_handle(last_handle);
1445                 }
1446 next:
1447                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1448                         if (ret == 0) {
1449                                 free_z3fold_page(page, true);
1450                                 atomic64_dec(&pool->pages_nr);
1451                                 return 0;
1452                         }
1453                         spin_lock(&pool->lock);
1454                         list_add(&page->lru, &pool->lru);
1455                         spin_unlock(&pool->lock);
1456                         clear_bit(PAGE_CLAIMED, &page->private);
1457                 } else {
1458                         z3fold_page_lock(zhdr);
1459                         if (kref_put(&zhdr->refcount,
1460                                         release_z3fold_page_locked)) {
1461                                 atomic64_dec(&pool->pages_nr);
1462                                 return 0;
1463                         }
1464                         /*
1465                          * if we are here, the page is still not completely
1466                          * free. Take the global pool lock then to be able
1467                          * to add it back to the lru list
1468                          */
1469                         spin_lock(&pool->lock);
1470                         list_add(&page->lru, &pool->lru);
1471                         spin_unlock(&pool->lock);
1472                         z3fold_page_unlock(zhdr);
1473                         clear_bit(PAGE_CLAIMED, &page->private);
1474                 }
1475
1476                 /* We started off locked to we need to lock the pool back */
1477                 spin_lock(&pool->lock);
1478         }
1479         spin_unlock(&pool->lock);
1480         return -EAGAIN;
1481 }
1482
1483 /**
1484  * z3fold_map() - maps the allocation associated with the given handle
1485  * @pool:       pool in which the allocation resides
1486  * @handle:     handle associated with the allocation to be mapped
1487  *
1488  * Extracts the buddy number from handle and constructs the pointer to the
1489  * correct starting chunk within the page.
1490  *
1491  * Returns: a pointer to the mapped allocation
1492  */
1493 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1494 {
1495         struct z3fold_header *zhdr;
1496         struct page *page;
1497         void *addr;
1498         enum buddy buddy;
1499
1500         zhdr = get_z3fold_header(handle);
1501         addr = zhdr;
1502         page = virt_to_page(zhdr);
1503
1504         if (test_bit(PAGE_HEADLESS, &page->private))
1505                 goto out;
1506
1507         buddy = handle_to_buddy(handle);
1508         switch (buddy) {
1509         case FIRST:
1510                 addr += ZHDR_SIZE_ALIGNED;
1511                 break;
1512         case MIDDLE:
1513                 addr += zhdr->start_middle << CHUNK_SHIFT;
1514                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1515                 break;
1516         case LAST:
1517                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1518                 break;
1519         default:
1520                 pr_err("unknown buddy id %d\n", buddy);
1521                 WARN_ON(1);
1522                 addr = NULL;
1523                 break;
1524         }
1525
1526         if (addr)
1527                 zhdr->mapped_count++;
1528 out:
1529         put_z3fold_header(zhdr);
1530         return addr;
1531 }
1532
1533 /**
1534  * z3fold_unmap() - unmaps the allocation associated with the given handle
1535  * @pool:       pool in which the allocation resides
1536  * @handle:     handle associated with the allocation to be unmapped
1537  */
1538 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1539 {
1540         struct z3fold_header *zhdr;
1541         struct page *page;
1542         enum buddy buddy;
1543
1544         zhdr = get_z3fold_header(handle);
1545         page = virt_to_page(zhdr);
1546
1547         if (test_bit(PAGE_HEADLESS, &page->private))
1548                 return;
1549
1550         buddy = handle_to_buddy(handle);
1551         if (buddy == MIDDLE)
1552                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1553         zhdr->mapped_count--;
1554         put_z3fold_header(zhdr);
1555 }
1556
1557 /**
1558  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1559  * @pool:       pool whose size is being queried
1560  *
1561  * Returns: size in pages of the given pool.
1562  */
1563 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1564 {
1565         return atomic64_read(&pool->pages_nr);
1566 }
1567
1568 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1569 {
1570         struct z3fold_header *zhdr;
1571         struct z3fold_pool *pool;
1572
1573         VM_BUG_ON_PAGE(!PageMovable(page), page);
1574         VM_BUG_ON_PAGE(PageIsolated(page), page);
1575
1576         if (test_bit(PAGE_HEADLESS, &page->private) ||
1577             test_bit(PAGE_CLAIMED, &page->private))
1578                 return false;
1579
1580         zhdr = page_address(page);
1581         z3fold_page_lock(zhdr);
1582         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1583             test_bit(PAGE_STALE, &page->private))
1584                 goto out;
1585
1586         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1587                 goto out;
1588
1589         pool = zhdr_to_pool(zhdr);
1590         spin_lock(&pool->lock);
1591         if (!list_empty(&zhdr->buddy))
1592                 list_del_init(&zhdr->buddy);
1593         if (!list_empty(&page->lru))
1594                 list_del_init(&page->lru);
1595         spin_unlock(&pool->lock);
1596
1597         kref_get(&zhdr->refcount);
1598         z3fold_page_unlock(zhdr);
1599         return true;
1600
1601 out:
1602         z3fold_page_unlock(zhdr);
1603         return false;
1604 }
1605
1606 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1607                                struct page *page, enum migrate_mode mode)
1608 {
1609         struct z3fold_header *zhdr, *new_zhdr;
1610         struct z3fold_pool *pool;
1611         struct address_space *new_mapping;
1612
1613         VM_BUG_ON_PAGE(!PageMovable(page), page);
1614         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1615         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1616
1617         zhdr = page_address(page);
1618         pool = zhdr_to_pool(zhdr);
1619
1620         if (!z3fold_page_trylock(zhdr)) {
1621                 return -EAGAIN;
1622         }
1623         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1624                 z3fold_page_unlock(zhdr);
1625                 return -EBUSY;
1626         }
1627         if (work_pending(&zhdr->work)) {
1628                 z3fold_page_unlock(zhdr);
1629                 return -EAGAIN;
1630         }
1631         new_zhdr = page_address(newpage);
1632         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1633         newpage->private = page->private;
1634         page->private = 0;
1635         z3fold_page_unlock(zhdr);
1636         spin_lock_init(&new_zhdr->page_lock);
1637         INIT_WORK(&new_zhdr->work, compact_page_work);
1638         /*
1639          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1640          * so we only have to reinitialize it.
1641          */
1642         INIT_LIST_HEAD(&new_zhdr->buddy);
1643         new_mapping = page_mapping(page);
1644         __ClearPageMovable(page);
1645         ClearPagePrivate(page);
1646
1647         get_page(newpage);
1648         z3fold_page_lock(new_zhdr);
1649         if (new_zhdr->first_chunks)
1650                 encode_handle(new_zhdr, FIRST);
1651         if (new_zhdr->last_chunks)
1652                 encode_handle(new_zhdr, LAST);
1653         if (new_zhdr->middle_chunks)
1654                 encode_handle(new_zhdr, MIDDLE);
1655         set_bit(NEEDS_COMPACTING, &newpage->private);
1656         new_zhdr->cpu = smp_processor_id();
1657         spin_lock(&pool->lock);
1658         list_add(&newpage->lru, &pool->lru);
1659         spin_unlock(&pool->lock);
1660         __SetPageMovable(newpage, new_mapping);
1661         z3fold_page_unlock(new_zhdr);
1662
1663         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1664
1665         page_mapcount_reset(page);
1666         put_page(page);
1667         return 0;
1668 }
1669
1670 static void z3fold_page_putback(struct page *page)
1671 {
1672         struct z3fold_header *zhdr;
1673         struct z3fold_pool *pool;
1674
1675         zhdr = page_address(page);
1676         pool = zhdr_to_pool(zhdr);
1677
1678         z3fold_page_lock(zhdr);
1679         if (!list_empty(&zhdr->buddy))
1680                 list_del_init(&zhdr->buddy);
1681         INIT_LIST_HEAD(&page->lru);
1682         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1683                 atomic64_dec(&pool->pages_nr);
1684                 return;
1685         }
1686         spin_lock(&pool->lock);
1687         list_add(&page->lru, &pool->lru);
1688         spin_unlock(&pool->lock);
1689         z3fold_page_unlock(zhdr);
1690 }
1691
1692 static const struct address_space_operations z3fold_aops = {
1693         .isolate_page = z3fold_page_isolate,
1694         .migratepage = z3fold_page_migrate,
1695         .putback_page = z3fold_page_putback,
1696 };
1697
1698 /*****************
1699  * zpool
1700  ****************/
1701
1702 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1703 {
1704         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1705                 return pool->zpool_ops->evict(pool->zpool, handle);
1706         else
1707                 return -ENOENT;
1708 }
1709
1710 static const struct z3fold_ops z3fold_zpool_ops = {
1711         .evict =        z3fold_zpool_evict
1712 };
1713
1714 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1715                                const struct zpool_ops *zpool_ops,
1716                                struct zpool *zpool)
1717 {
1718         struct z3fold_pool *pool;
1719
1720         pool = z3fold_create_pool(name, gfp,
1721                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1722         if (pool) {
1723                 pool->zpool = zpool;
1724                 pool->zpool_ops = zpool_ops;
1725         }
1726         return pool;
1727 }
1728
1729 static void z3fold_zpool_destroy(void *pool)
1730 {
1731         z3fold_destroy_pool(pool);
1732 }
1733
1734 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1735                         unsigned long *handle)
1736 {
1737         return z3fold_alloc(pool, size, gfp, handle);
1738 }
1739 static void z3fold_zpool_free(void *pool, unsigned long handle)
1740 {
1741         z3fold_free(pool, handle);
1742 }
1743
1744 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1745                         unsigned int *reclaimed)
1746 {
1747         unsigned int total = 0;
1748         int ret = -EINVAL;
1749
1750         while (total < pages) {
1751                 ret = z3fold_reclaim_page(pool, 8);
1752                 if (ret < 0)
1753                         break;
1754                 total++;
1755         }
1756
1757         if (reclaimed)
1758                 *reclaimed = total;
1759
1760         return ret;
1761 }
1762
1763 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1764                         enum zpool_mapmode mm)
1765 {
1766         return z3fold_map(pool, handle);
1767 }
1768 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1769 {
1770         z3fold_unmap(pool, handle);
1771 }
1772
1773 static u64 z3fold_zpool_total_size(void *pool)
1774 {
1775         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1776 }
1777
1778 static struct zpool_driver z3fold_zpool_driver = {
1779         .type =         "z3fold",
1780         .owner =        THIS_MODULE,
1781         .create =       z3fold_zpool_create,
1782         .destroy =      z3fold_zpool_destroy,
1783         .malloc =       z3fold_zpool_malloc,
1784         .free =         z3fold_zpool_free,
1785         .shrink =       z3fold_zpool_shrink,
1786         .map =          z3fold_zpool_map,
1787         .unmap =        z3fold_zpool_unmap,
1788         .total_size =   z3fold_zpool_total_size,
1789 };
1790
1791 MODULE_ALIAS("zpool-z3fold");
1792
1793 static int __init init_z3fold(void)
1794 {
1795         int ret;
1796
1797         /* Make sure the z3fold header is not larger than the page size */
1798         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1799         ret = z3fold_mount();
1800         if (ret)
1801                 return ret;
1802
1803         zpool_register_driver(&z3fold_zpool_driver);
1804
1805         return 0;
1806 }
1807
1808 static void __exit exit_z3fold(void)
1809 {
1810         z3fold_unmount();
1811         zpool_unregister_driver(&z3fold_zpool_driver);
1812 }
1813
1814 module_init(init_z3fold);
1815 module_exit(exit_z3fold);
1816
1817 MODULE_LICENSE("GPL");
1818 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1819 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");