mm/z3fold.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * z3fold.c
   4  *
   5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   6  * Copyright (C) 2016, Sony Mobile Communications Inc.
   7  *
   8  * This implementation is based on zbud written by Seth Jennings.
   9  *
  10  * z3fold is an special purpose allocator for storing compressed pages. It
  11  * can store up to three compressed pages per page which improves the
  12  * compression ratio of zbud while retaining its main concepts (e. g. always
  13  * storing an integral number of objects per page) and simplicity.
  14  * It still has simple and deterministic reclaim properties that make it
  15  * preferable to a higher density approach (with no requirement on integral
  16  * number of object per page) when reclaim is used.
  17  *
  18  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  19  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  20  *
  21  * z3fold doesn't export any API and is meant to be used via zpool API.
  22  */
  23
  24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  25
  26 #include <linux/atomic.h>
  27 #include <linux/sched.h>
  28 #include <linux/cpumask.h>
  29 #include <linux/list.h>
  30 #include <linux/mm.h>
  31 #include <linux/module.h>
  32 #include <linux/page-flags.h>
  33 #include <linux/migrate.h>
  34 #include <linux/node.h>
  35 #include <linux/compaction.h>
  36 #include <linux/percpu.h>
  37 #include <linux/mount.h>
  38 #include <linux/pseudo_fs.h>
  39 #include <linux/fs.h>
  40 #include <linux/preempt.h>
  41 #include <linux/workqueue.h>
  42 #include <linux/slab.h>
  43 #include <linux/spinlock.h>
  44 #include <linux/zpool.h>
  45 #include <linux/magic.h>
  46 #include <linux/kmemleak.h>
  47
  48 /*
  49  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  50  * adjusting internal fragmentation.  It also determines the number of
  51  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  52  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  53  * in the beginning of an allocated page are occupied by z3fold header, so
  54  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  55  * which shows the max number of free chunks in z3fold page, also there will
  56  * be 63, or 62, respectively, freelists per pool.
  57  */
  58 #define NCHUNKS_ORDER   6
  59
  60 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  61 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  62 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  63 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  64 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  65 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
  66
  67 #define BUDDY_MASK      (0x3)
  68 #define BUDDY_SHIFT     2
  69 #define SLOTS_ALIGN     (0x40)
  70
  71 /*****************
  72  * Structures
  73 *****************/
  74 struct z3fold_pool;
  75 struct z3fold_ops {
  76         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  77 };
  78
  79 enum buddy {
  80         HEADLESS = 0,
  81         FIRST,
  82         MIDDLE,
  83         LAST,
  84         BUDDIES_MAX = LAST
  85 };
  86
  87 struct z3fold_buddy_slots {
  88         /*
  89          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  90          * be enough slots to hold all possible variants
  91          */
  92         unsigned long slot[BUDDY_MASK + 1];
  93         unsigned long pool; /* back link */
  94         rwlock_t lock;
  95 };
  96 #define HANDLE_FLAG_MASK        (0x03)
  97
  98 /*
  99  * struct z3fold_header - z3fold page metadata occupying first chunks of each
 100  *                      z3fold page, except for HEADLESS pages
 101  * @buddy:              links the z3fold page into the relevant list in the
 102  *                      pool
 103  * @page_lock:          per-page lock
 104  * @refcount:           reference count for the z3fold page
 105  * @work:               work_struct for page layout optimization
 106  * @slots:              pointer to the structure holding buddy slots
 107  * @pool:               pointer to the containing pool
 108  * @cpu:                CPU which this page "belongs" to
 109  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 110  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 111  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 112  * @first_num:          the starting number (for the first handle)
 113  * @mapped_count:       the number of objects currently mapped
 114  */
 115 struct z3fold_header {
 116         struct list_head buddy;
 117         spinlock_t page_lock;
 118         struct kref refcount;
 119         struct work_struct work;
 120         struct z3fold_buddy_slots *slots;
 121         struct z3fold_pool *pool;
 122         short cpu;
 123         unsigned short first_chunks;
 124         unsigned short middle_chunks;
 125         unsigned short last_chunks;
 126         unsigned short start_middle;
 127         unsigned short first_num:2;
 128         unsigned short mapped_count:2;
 129         unsigned short foreign_handles:2;
 130 };
 131
 132 /**
 133  * struct z3fold_pool - stores metadata for each z3fold pool
 134  * @name:       pool name
 135  * @lock:       protects pool unbuddied/lru lists
 136  * @stale_lock: protects pool stale page list
 137  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 138  *              buddies; the list each z3fold page is added to depends on
 139  *              the size of its free region.
 140  * @lru:        list tracking the z3fold pages in LRU order by most recently
 141  *              added buddy.
 142  * @stale:      list of pages marked for freeing
 143  * @pages_nr:   number of z3fold pages in the pool.
 144  * @c_handle:   cache for z3fold_buddy_slots allocation
 145  * @ops:        pointer to a structure of user defined operations specified at
 146  *              pool creation time.
 147  * @compact_wq: workqueue for page layout background optimization
 148  * @release_wq: workqueue for safe page release
 149  * @work:       work_struct for safe page release
 150  * @inode:      inode for z3fold pseudo filesystem
 151  *
 152  * This structure is allocated at pool creation time and maintains metadata
 153  * pertaining to a particular z3fold pool.
 154  */
 155 struct z3fold_pool {
 156         const char *name;
 157         spinlock_t lock;
 158         spinlock_t stale_lock;
 159         struct list_head *unbuddied;
 160         struct list_head lru;
 161         struct list_head stale;
 162         atomic64_t pages_nr;
 163         struct kmem_cache *c_handle;
 164         const struct z3fold_ops *ops;
 165         struct zpool *zpool;
 166         const struct zpool_ops *zpool_ops;
 167         struct workqueue_struct *compact_wq;
 168         struct workqueue_struct *release_wq;
 169         struct work_struct work;
 170         struct inode *inode;
 171 };
 172
 173 /*
 174  * Internal z3fold page flags
 175  */
 176 enum z3fold_page_flags {
 177         PAGE_HEADLESS = 0,
 178         MIDDLE_CHUNK_MAPPED,
 179         NEEDS_COMPACTING,
 180         PAGE_STALE,
 181         PAGE_CLAIMED, /* by either reclaim or free */
 182 };
 183
 184 /*
 185  * handle flags, go under HANDLE_FLAG_MASK
 186  */
 187 enum z3fold_handle_flags {
 188         HANDLES_NOFREE = 0,
 189 };
 190
 191 /*
 192  * Forward declarations
 193  */
 194 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 195 static void compact_page_work(struct work_struct *w);
 196
 197 /*****************
 198  * Helpers
 199 *****************/
 200
 201 /* Converts an allocation size in bytes to size in z3fold chunks */
 202 static int size_to_chunks(size_t size)
 203 {
 204         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 205 }
 206
 207 #define for_each_unbuddied_list(_iter, _begin) \
 208         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 209
 210 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 211                                                         gfp_t gfp)
 212 {
 213         struct z3fold_buddy_slots *slots;
 214
 215         slots = kmem_cache_zalloc(pool->c_handle,
 216                                  (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
 217
 218         if (slots) {
 219                 /* It will be freed separately in free_handle(). */
 220                 kmemleak_not_leak(slots);
 221                 slots->pool = (unsigned long)pool;
 222                 rwlock_init(&slots->lock);
 223         }
 224
 225         return slots;
 226 }
 227
 228 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 229 {
 230         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 231 }
 232
 233 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 234 {
 235         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 236 }
 237
 238 /* Lock a z3fold page */
 239 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 240 {
 241         spin_lock(&zhdr->page_lock);
 242 }
 243
 244 /* Try to lock a z3fold page */
 245 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 246 {
 247         return spin_trylock(&zhdr->page_lock);
 248 }
 249
 250 /* Unlock a z3fold page */
 251 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 252 {
 253         spin_unlock(&zhdr->page_lock);
 254 }
 255
 256
 257 static inline struct z3fold_header *__get_z3fold_header(unsigned long handle,
 258                                                         bool lock)
 259 {
 260         struct z3fold_buddy_slots *slots;
 261         struct z3fold_header *zhdr;
 262         int locked = 0;
 263
 264         if (!(handle & (1 << PAGE_HEADLESS))) {
 265                 slots = handle_to_slots(handle);
 266                 do {
 267                         unsigned long addr;
 268
 269                         read_lock(&slots->lock);
 270                         addr = *(unsigned long *)handle;
 271                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 272                         if (lock)
 273                                 locked = z3fold_page_trylock(zhdr);
 274                         read_unlock(&slots->lock);
 275                         if (locked)
 276                                 break;
 277                         cpu_relax();
 278                 } while (lock);
 279         } else {
 280                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 281         }
 282
 283         return zhdr;
 284 }
 285
 286 /* Returns the z3fold page where a given handle is stored */
 287 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
 288 {
 289         return __get_z3fold_header(h, false);
 290 }
 291
 292 /* return locked z3fold page if it's not headless */
 293 static inline struct z3fold_header *get_z3fold_header(unsigned long h)
 294 {
 295         return __get_z3fold_header(h, true);
 296 }
 297
 298 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 299 {
 300         struct page *page = virt_to_page(zhdr);
 301
 302         if (!test_bit(PAGE_HEADLESS, &page->private))
 303                 z3fold_page_unlock(zhdr);
 304 }
 305
 306 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 307 {
 308         struct z3fold_buddy_slots *slots;
 309         int i;
 310         bool is_free;
 311
 312         if (handle & (1 << PAGE_HEADLESS))
 313                 return;
 314
 315         if (WARN_ON(*(unsigned long *)handle == 0))
 316                 return;
 317
 318         slots = handle_to_slots(handle);
 319         write_lock(&slots->lock);
 320         *(unsigned long *)handle = 0;
 321
 322         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 323                 write_unlock(&slots->lock);
 324                 return; /* simple case, nothing else to do */
 325         }
 326
 327         if (zhdr->slots != slots)
 328                 zhdr->foreign_handles--;
 329
 330         is_free = true;
 331         for (i = 0; i <= BUDDY_MASK; i++) {
 332                 if (slots->slot[i]) {
 333                         is_free = false;
 334                         break;
 335                 }
 336         }
 337         write_unlock(&slots->lock);
 338
 339         if (is_free) {
 340                 struct z3fold_pool *pool = slots_to_pool(slots);
 341
 342                 if (zhdr->slots == slots)
 343                         zhdr->slots = NULL;
 344                 kmem_cache_free(pool->c_handle, slots);
 345         }
 346 }
 347
 348 static int z3fold_init_fs_context(struct fs_context *fc)
 349 {
 350         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
 351 }
 352
 353 static struct file_system_type z3fold_fs = {
 354         .name           = "z3fold",
 355         .init_fs_context = z3fold_init_fs_context,
 356         .kill_sb        = kill_anon_super,
 357 };
 358
 359 static struct vfsmount *z3fold_mnt;
 360 static int z3fold_mount(void)
 361 {
 362         int ret = 0;
 363
 364         z3fold_mnt = kern_mount(&z3fold_fs);
 365         if (IS_ERR(z3fold_mnt))
 366                 ret = PTR_ERR(z3fold_mnt);
 367
 368         return ret;
 369 }
 370
 371 static void z3fold_unmount(void)
 372 {
 373         kern_unmount(z3fold_mnt);
 374 }
 375
 376 static const struct address_space_operations z3fold_aops;
 377 static int z3fold_register_migration(struct z3fold_pool *pool)
 378 {
 379         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
 380         if (IS_ERR(pool->inode)) {
 381                 pool->inode = NULL;
 382                 return 1;
 383         }
 384
 385         pool->inode->i_mapping->private_data = pool;
 386         pool->inode->i_mapping->a_ops = &z3fold_aops;
 387         return 0;
 388 }
 389
 390 static void z3fold_unregister_migration(struct z3fold_pool *pool)
 391 {
 392         if (pool->inode)
 393                 iput(pool->inode);
 394  }
 395
 396 /* Initializes the z3fold header of a newly allocated z3fold page */
 397 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 398                                         struct z3fold_pool *pool, gfp_t gfp)
 399 {
 400         struct z3fold_header *zhdr = page_address(page);
 401         struct z3fold_buddy_slots *slots;
 402
 403         INIT_LIST_HEAD(&page->lru);
 404         clear_bit(PAGE_HEADLESS, &page->private);
 405         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 406         clear_bit(NEEDS_COMPACTING, &page->private);
 407         clear_bit(PAGE_STALE, &page->private);
 408         clear_bit(PAGE_CLAIMED, &page->private);
 409         if (headless)
 410                 return zhdr;
 411
 412         slots = alloc_slots(pool, gfp);
 413         if (!slots)
 414                 return NULL;
 415
 416         spin_lock_init(&zhdr->page_lock);
 417         kref_init(&zhdr->refcount);
 418         zhdr->first_chunks = 0;
 419         zhdr->middle_chunks = 0;
 420         zhdr->last_chunks = 0;
 421         zhdr->first_num = 0;
 422         zhdr->start_middle = 0;
 423         zhdr->cpu = -1;
 424         zhdr->foreign_handles = 0;
 425         zhdr->mapped_count = 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
 527         WARN_ON(!list_empty(&zhdr->buddy));
 528         set_bit(PAGE_STALE, &page->private);
 529         clear_bit(NEEDS_COMPACTING, &page->private);
 530         spin_lock(&pool->lock);
 531         if (!list_empty(&page->lru))
 532                 list_del_init(&page->lru);
 533         spin_unlock(&pool->lock);
 534
 535         if (locked)
 536                 z3fold_page_unlock(zhdr);
 537
 538         spin_lock(&pool->stale_lock);
 539         list_add(&zhdr->buddy, &pool->stale);
 540         queue_work(pool->release_wq, &pool->work);
 541         spin_unlock(&pool->stale_lock);
 542 }
 543
 544 static void __attribute__((__unused__))
 545                         release_z3fold_page(struct kref *ref)
 546 {
 547         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 548                                                 refcount);
 549         __release_z3fold_page(zhdr, false);
 550 }
 551
 552 static void release_z3fold_page_locked(struct kref *ref)
 553 {
 554         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 555                                                 refcount);
 556         WARN_ON(z3fold_page_trylock(zhdr));
 557         __release_z3fold_page(zhdr, true);
 558 }
 559
 560 static void release_z3fold_page_locked_list(struct kref *ref)
 561 {
 562         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 563                                                refcount);
 564         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 565
 566         spin_lock(&pool->lock);
 567         list_del_init(&zhdr->buddy);
 568         spin_unlock(&pool->lock);
 569
 570         WARN_ON(z3fold_page_trylock(zhdr));
 571         __release_z3fold_page(zhdr, true);
 572 }
 573
 574 static void free_pages_work(struct work_struct *w)
 575 {
 576         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 577
 578         spin_lock(&pool->stale_lock);
 579         while (!list_empty(&pool->stale)) {
 580                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 581                                                 struct z3fold_header, buddy);
 582                 struct page *page = virt_to_page(zhdr);
 583
 584                 list_del(&zhdr->buddy);
 585                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 586                         continue;
 587                 spin_unlock(&pool->stale_lock);
 588                 cancel_work_sync(&zhdr->work);
 589                 free_z3fold_page(page, false);
 590                 cond_resched();
 591                 spin_lock(&pool->stale_lock);
 592         }
 593         spin_unlock(&pool->stale_lock);
 594 }
 595
 596 /*
 597  * Returns the number of free chunks in a z3fold page.
 598  * NB: can't be used with HEADLESS pages.
 599  */
 600 static int num_free_chunks(struct z3fold_header *zhdr)
 601 {
 602         int nfree;
 603         /*
 604          * If there is a middle object, pick up the bigger free space
 605          * either before or after it. Otherwise just subtract the number
 606          * of chunks occupied by the first and the last objects.
 607          */
 608         if (zhdr->middle_chunks != 0) {
 609                 int nfree_before = zhdr->first_chunks ?
 610                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 611                 int nfree_after = zhdr->last_chunks ?
 612                         0 : TOTAL_CHUNKS -
 613                                 (zhdr->start_middle + zhdr->middle_chunks);
 614                 nfree = max(nfree_before, nfree_after);
 615         } else
 616                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 617         return nfree;
 618 }
 619
 620 /* Add to the appropriate unbuddied list */
 621 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 622                                 struct z3fold_header *zhdr)
 623 {
 624         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 625                         zhdr->middle_chunks == 0) {
 626                 struct list_head *unbuddied;
 627                 int freechunks = num_free_chunks(zhdr);
 628
 629                 migrate_disable();
 630                 unbuddied = this_cpu_ptr(pool->unbuddied);
 631                 spin_lock(&pool->lock);
 632                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 633                 spin_unlock(&pool->lock);
 634                 zhdr->cpu = smp_processor_id();
 635                 migrate_enable();
 636         }
 637 }
 638
 639 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 640 {
 641         enum buddy bud = HEADLESS;
 642
 643         if (zhdr->middle_chunks) {
 644                 if (!zhdr->first_chunks &&
 645                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 646                         bud = FIRST;
 647                 else if (!zhdr->last_chunks)
 648                         bud = LAST;
 649         } else {
 650                 if (!zhdr->first_chunks)
 651                         bud = FIRST;
 652                 else if (!zhdr->last_chunks)
 653                         bud = LAST;
 654                 else
 655                         bud = MIDDLE;
 656         }
 657
 658         return bud;
 659 }
 660
 661 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 662                                 unsigned short dst_chunk)
 663 {
 664         void *beg = zhdr;
 665         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 666                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 667                        zhdr->middle_chunks << CHUNK_SHIFT);
 668 }
 669
 670 static inline bool buddy_single(struct z3fold_header *zhdr)
 671 {
 672         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 673                         (zhdr->first_chunks && zhdr->last_chunks) ||
 674                         (zhdr->middle_chunks && zhdr->last_chunks));
 675 }
 676
 677 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 678 {
 679         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 680         void *p = zhdr;
 681         unsigned long old_handle = 0;
 682         size_t sz = 0;
 683         struct z3fold_header *new_zhdr = NULL;
 684         int first_idx = __idx(zhdr, FIRST);
 685         int middle_idx = __idx(zhdr, MIDDLE);
 686         int last_idx = __idx(zhdr, LAST);
 687         unsigned short *moved_chunks = NULL;
 688
 689         /*
 690          * No need to protect slots here -- all the slots are "local" and
 691          * the page lock is already taken
 692          */
 693         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 694                 p += ZHDR_SIZE_ALIGNED;
 695                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 696                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 697                 moved_chunks = &zhdr->first_chunks;
 698         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 699                 p += zhdr->start_middle << CHUNK_SHIFT;
 700                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 701                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 702                 moved_chunks = &zhdr->middle_chunks;
 703         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 704                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 705                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 706                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 707                 moved_chunks = &zhdr->last_chunks;
 708         }
 709
 710         if (sz > 0) {
 711                 enum buddy new_bud = HEADLESS;
 712                 short chunks = size_to_chunks(sz);
 713                 void *q;
 714
 715                 new_zhdr = __z3fold_alloc(pool, sz, false);
 716                 if (!new_zhdr)
 717                         return NULL;
 718
 719                 if (WARN_ON(new_zhdr == zhdr))
 720                         goto out_fail;
 721
 722                 new_bud = get_free_buddy(new_zhdr, chunks);
 723                 q = new_zhdr;
 724                 switch (new_bud) {
 725                 case FIRST:
 726                         new_zhdr->first_chunks = chunks;
 727                         q += ZHDR_SIZE_ALIGNED;
 728                         break;
 729                 case MIDDLE:
 730                         new_zhdr->middle_chunks = chunks;
 731                         new_zhdr->start_middle =
 732                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 733                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 734                         break;
 735                 case LAST:
 736                         new_zhdr->last_chunks = chunks;
 737                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 738                         break;
 739                 default:
 740                         goto out_fail;
 741                 }
 742                 new_zhdr->foreign_handles++;
 743                 memcpy(q, p, sz);
 744                 write_lock(&zhdr->slots->lock);
 745                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 746                         __idx(new_zhdr, new_bud);
 747                 if (new_bud == LAST)
 748                         *(unsigned long *)old_handle |=
 749                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 750                 write_unlock(&zhdr->slots->lock);
 751                 add_to_unbuddied(pool, new_zhdr);
 752                 z3fold_page_unlock(new_zhdr);
 753
 754                 *moved_chunks = 0;
 755         }
 756
 757         return new_zhdr;
 758
 759 out_fail:
 760         if (new_zhdr) {
 761                 if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
 762                         atomic64_dec(&pool->pages_nr);
 763                 else {
 764                         add_to_unbuddied(pool, new_zhdr);
 765                         z3fold_page_unlock(new_zhdr);
 766                 }
 767         }
 768         return NULL;
 769
 770 }
 771
 772 #define BIG_CHUNK_GAP   3
 773 /* Has to be called with lock held */
 774 static int z3fold_compact_page(struct z3fold_header *zhdr)
 775 {
 776         struct page *page = virt_to_page(zhdr);
 777
 778         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 779                 return 0; /* can't move middle chunk, it's used */
 780
 781         if (unlikely(PageIsolated(page)))
 782                 return 0;
 783
 784         if (zhdr->middle_chunks == 0)
 785                 return 0; /* nothing to compact */
 786
 787         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 788                 /* move to the beginning */
 789                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 790                 zhdr->first_chunks = zhdr->middle_chunks;
 791                 zhdr->middle_chunks = 0;
 792                 zhdr->start_middle = 0;
 793                 zhdr->first_num++;
 794                 return 1;
 795         }
 796
 797         /*
 798          * moving data is expensive, so let's only do that if
 799          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 800          */
 801         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 802             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 803                         BIG_CHUNK_GAP) {
 804                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 805                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 806                 return 1;
 807         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 808                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 809                                         + zhdr->middle_chunks) >=
 810                         BIG_CHUNK_GAP) {
 811                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 812                         zhdr->middle_chunks;
 813                 mchunk_memmove(zhdr, new_start);
 814                 zhdr->start_middle = new_start;
 815                 return 1;
 816         }
 817
 818         return 0;
 819 }
 820
 821 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 822 {
 823         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 824         struct page *page;
 825
 826         page = virt_to_page(zhdr);
 827         if (locked)
 828                 WARN_ON(z3fold_page_trylock(zhdr));
 829         else
 830                 z3fold_page_lock(zhdr);
 831         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 832                 z3fold_page_unlock(zhdr);
 833                 return;
 834         }
 835         spin_lock(&pool->lock);
 836         list_del_init(&zhdr->buddy);
 837         spin_unlock(&pool->lock);
 838
 839         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 840                 atomic64_dec(&pool->pages_nr);
 841                 return;
 842         }
 843
 844         if (test_bit(PAGE_STALE, &page->private) ||
 845             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 846                 z3fold_page_unlock(zhdr);
 847                 return;
 848         }
 849
 850         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 851             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 852                 if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 853                         atomic64_dec(&pool->pages_nr);
 854                 else {
 855                         clear_bit(PAGE_CLAIMED, &page->private);
 856                         z3fold_page_unlock(zhdr);
 857                 }
 858                 return;
 859         }
 860
 861         z3fold_compact_page(zhdr);
 862         add_to_unbuddied(pool, zhdr);
 863         clear_bit(PAGE_CLAIMED, &page->private);
 864         z3fold_page_unlock(zhdr);
 865 }
 866
 867 static void compact_page_work(struct work_struct *w)
 868 {
 869         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 870                                                 work);
 871
 872         do_compact_page(zhdr, false);
 873 }
 874
 875 /* returns _locked_ z3fold page header or NULL */
 876 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 877                                                 size_t size, bool can_sleep)
 878 {
 879         struct z3fold_header *zhdr = NULL;
 880         struct page *page;
 881         struct list_head *unbuddied;
 882         int chunks = size_to_chunks(size), i;
 883
 884 lookup:
 885         migrate_disable();
 886         /* First, try to find an unbuddied z3fold page. */
 887         unbuddied = this_cpu_ptr(pool->unbuddied);
 888         for_each_unbuddied_list(i, chunks) {
 889                 struct list_head *l = &unbuddied[i];
 890
 891                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 892                                         struct z3fold_header, buddy);
 893
 894                 if (!zhdr)
 895                         continue;
 896
 897                 /* Re-check under lock. */
 898                 spin_lock(&pool->lock);
 899                 l = &unbuddied[i];
 900                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 901                                                 struct z3fold_header, buddy)) ||
 902                     !z3fold_page_trylock(zhdr)) {
 903                         spin_unlock(&pool->lock);
 904                         zhdr = NULL;
 905                         migrate_enable();
 906                         if (can_sleep)
 907                                 cond_resched();
 908                         goto lookup;
 909                 }
 910                 list_del_init(&zhdr->buddy);
 911                 zhdr->cpu = -1;
 912                 spin_unlock(&pool->lock);
 913
 914                 page = virt_to_page(zhdr);
 915                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 916                     test_bit(PAGE_CLAIMED, &page->private)) {
 917                         z3fold_page_unlock(zhdr);
 918                         zhdr = NULL;
 919                         migrate_enable();
 920                         if (can_sleep)
 921                                 cond_resched();
 922                         goto lookup;
 923                 }
 924
 925                 /*
 926                  * this page could not be removed from its unbuddied
 927                  * list while pool lock was held, and then we've taken
 928                  * page lock so kref_put could not be called before
 929                  * we got here, so it's safe to just call kref_get()
 930                  */
 931                 kref_get(&zhdr->refcount);
 932                 break;
 933         }
 934         migrate_enable();
 935
 936         if (!zhdr) {
 937                 int cpu;
 938
 939                 /* look for _exact_ match on other cpus' lists */
 940                 for_each_online_cpu(cpu) {
 941                         struct list_head *l;
 942
 943                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 944                         spin_lock(&pool->lock);
 945                         l = &unbuddied[chunks];
 946
 947                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 948                                                 struct z3fold_header, buddy);
 949
 950                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 951                                 spin_unlock(&pool->lock);
 952                                 zhdr = NULL;
 953                                 continue;
 954                         }
 955                         list_del_init(&zhdr->buddy);
 956                         zhdr->cpu = -1;
 957                         spin_unlock(&pool->lock);
 958
 959                         page = virt_to_page(zhdr);
 960                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 961                             test_bit(PAGE_CLAIMED, &page->private)) {
 962                                 z3fold_page_unlock(zhdr);
 963                                 zhdr = NULL;
 964                                 if (can_sleep)
 965                                         cond_resched();
 966                                 continue;
 967                         }
 968                         kref_get(&zhdr->refcount);
 969                         break;
 970                 }
 971         }
 972
 973         if (zhdr && !zhdr->slots)
 974                 zhdr->slots = alloc_slots(pool,
 975                                         can_sleep ? GFP_NOIO : GFP_ATOMIC);
 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                         bud = get_free_buddy(zhdr, chunks);
1113                         if (bud == HEADLESS) {
1114                                 if (kref_put(&zhdr->refcount,
1115                                              release_z3fold_page_locked))
1116                                         atomic64_dec(&pool->pages_nr);
1117                                 else
1118                                         z3fold_page_unlock(zhdr);
1119                                 pr_err("No free chunks in unbuddied\n");
1120                                 WARN_ON(1);
1121                                 goto retry;
1122                         }
1123                         page = virt_to_page(zhdr);
1124                         goto found;
1125                 }
1126                 bud = FIRST;
1127         }
1128
1129         page = NULL;
1130         if (can_sleep) {
1131                 spin_lock(&pool->stale_lock);
1132                 zhdr = list_first_entry_or_null(&pool->stale,
1133                                                 struct z3fold_header, buddy);
1134                 /*
1135                  * Before allocating a page, let's see if we can take one from
1136                  * the stale pages list. cancel_work_sync() can sleep so we
1137                  * limit this case to the contexts where we can sleep
1138                  */
1139                 if (zhdr) {
1140                         list_del(&zhdr->buddy);
1141                         spin_unlock(&pool->stale_lock);
1142                         cancel_work_sync(&zhdr->work);
1143                         page = virt_to_page(zhdr);
1144                 } else {
1145                         spin_unlock(&pool->stale_lock);
1146                 }
1147         }
1148         if (!page)
1149                 page = alloc_page(gfp);
1150
1151         if (!page)
1152                 return -ENOMEM;
1153
1154         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1155         if (!zhdr) {
1156                 __free_page(page);
1157                 return -ENOMEM;
1158         }
1159         atomic64_inc(&pool->pages_nr);
1160
1161         if (bud == HEADLESS) {
1162                 set_bit(PAGE_HEADLESS, &page->private);
1163                 goto headless;
1164         }
1165         if (can_sleep) {
1166                 lock_page(page);
1167                 __SetPageMovable(page, pool->inode->i_mapping);
1168                 unlock_page(page);
1169         } else {
1170                 if (trylock_page(page)) {
1171                         __SetPageMovable(page, pool->inode->i_mapping);
1172                         unlock_page(page);
1173                 }
1174         }
1175         z3fold_page_lock(zhdr);
1176
1177 found:
1178         if (bud == FIRST)
1179                 zhdr->first_chunks = chunks;
1180         else if (bud == LAST)
1181                 zhdr->last_chunks = chunks;
1182         else {
1183                 zhdr->middle_chunks = chunks;
1184                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1185         }
1186         add_to_unbuddied(pool, zhdr);
1187
1188 headless:
1189         spin_lock(&pool->lock);
1190         /* Add/move z3fold page to beginning of LRU */
1191         if (!list_empty(&page->lru))
1192                 list_del(&page->lru);
1193
1194         list_add(&page->lru, &pool->lru);
1195
1196         *handle = encode_handle(zhdr, bud);
1197         spin_unlock(&pool->lock);
1198         if (bud != HEADLESS)
1199                 z3fold_page_unlock(zhdr);
1200
1201         return 0;
1202 }
1203
1204 /**
1205  * z3fold_free() - frees the allocation associated with the given handle
1206  * @pool:       pool in which the allocation resided
1207  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1208  *
1209  * In the case that the z3fold page in which the allocation resides is under
1210  * reclaim, as indicated by the PG_reclaim flag being set, this function
1211  * only sets the first|last_chunks to 0.  The page is actually freed
1212  * once both buddies are evicted (see z3fold_reclaim_page() below).
1213  */
1214 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1215 {
1216         struct z3fold_header *zhdr;
1217         struct page *page;
1218         enum buddy bud;
1219         bool page_claimed;
1220
1221         zhdr = get_z3fold_header(handle);
1222         page = virt_to_page(zhdr);
1223         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1224
1225         if (test_bit(PAGE_HEADLESS, &page->private)) {
1226                 /* if a headless page is under reclaim, just leave.
1227                  * NB: we use test_and_set_bit for a reason: if the bit
1228                  * has not been set before, we release this page
1229                  * immediately so we don't care about its value any more.
1230                  */
1231                 if (!page_claimed) {
1232                         spin_lock(&pool->lock);
1233                         list_del(&page->lru);
1234                         spin_unlock(&pool->lock);
1235                         put_z3fold_header(zhdr);
1236                         free_z3fold_page(page, true);
1237                         atomic64_dec(&pool->pages_nr);
1238                 }
1239                 return;
1240         }
1241
1242         /* Non-headless case */
1243         bud = handle_to_buddy(handle);
1244
1245         switch (bud) {
1246         case FIRST:
1247                 zhdr->first_chunks = 0;
1248                 break;
1249         case MIDDLE:
1250                 zhdr->middle_chunks = 0;
1251                 break;
1252         case LAST:
1253                 zhdr->last_chunks = 0;
1254                 break;
1255         default:
1256                 pr_err("%s: unknown bud %d\n", __func__, bud);
1257                 WARN_ON(1);
1258                 put_z3fold_header(zhdr);
1259                 return;
1260         }
1261
1262         if (!page_claimed)
1263                 free_handle(handle, zhdr);
1264         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1265                 atomic64_dec(&pool->pages_nr);
1266                 return;
1267         }
1268         if (page_claimed) {
1269                 /* the page has not been claimed by us */
1270                 z3fold_page_unlock(zhdr);
1271                 return;
1272         }
1273         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1274                 put_z3fold_header(zhdr);
1275                 clear_bit(PAGE_CLAIMED, &page->private);
1276                 return;
1277         }
1278         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1279                 spin_lock(&pool->lock);
1280                 list_del_init(&zhdr->buddy);
1281                 spin_unlock(&pool->lock);
1282                 zhdr->cpu = -1;
1283                 kref_get(&zhdr->refcount);
1284                 clear_bit(PAGE_CLAIMED, &page->private);
1285                 do_compact_page(zhdr, true);
1286                 return;
1287         }
1288         kref_get(&zhdr->refcount);
1289         clear_bit(PAGE_CLAIMED, &page->private);
1290         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1291         put_z3fold_header(zhdr);
1292 }
1293
1294 /**
1295  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1296  * @pool:       pool from which a page will attempt to be evicted
1297  * @retries:    number of pages on the LRU list for which eviction will
1298  *              be attempted before failing
1299  *
1300  * z3fold reclaim is different from normal system reclaim in that it is done
1301  * from the bottom, up. This is because only the bottom layer, z3fold, has
1302  * information on how the allocations are organized within each z3fold page.
1303  * This has the potential to create interesting locking situations between
1304  * z3fold and the user, however.
1305  *
1306  * To avoid these, this is how z3fold_reclaim_page() should be called:
1307  *
1308  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1309  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1310  * call the user-defined eviction handler with the pool and handle as
1311  * arguments.
1312  *
1313  * If the handle can not be evicted, the eviction handler should return
1314  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1315  * appropriate list and try the next z3fold page on the LRU up to
1316  * a user defined number of retries.
1317  *
1318  * If the handle is successfully evicted, the eviction handler should
1319  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1320  * contains logic to delay freeing the page if the page is under reclaim,
1321  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1322  *
1323  * If all buddies in the z3fold page are successfully evicted, then the
1324  * z3fold page can be freed.
1325  *
1326  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1327  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1328  * the retry limit was hit.
1329  */
1330 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1331 {
1332         int i, ret = -1;
1333         struct z3fold_header *zhdr = NULL;
1334         struct page *page = NULL;
1335         struct list_head *pos;
1336         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1337         struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN)));
1338
1339         rwlock_init(&slots.lock);
1340         slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE);
1341
1342         spin_lock(&pool->lock);
1343         if (!pool->ops || !pool->ops->evict || retries == 0) {
1344                 spin_unlock(&pool->lock);
1345                 return -EINVAL;
1346         }
1347         for (i = 0; i < retries; i++) {
1348                 if (list_empty(&pool->lru)) {
1349                         spin_unlock(&pool->lock);
1350                         return -EINVAL;
1351                 }
1352                 list_for_each_prev(pos, &pool->lru) {
1353                         page = list_entry(pos, struct page, lru);
1354
1355                         zhdr = page_address(page);
1356                         if (test_bit(PAGE_HEADLESS, &page->private))
1357                                 break;
1358
1359                         if (kref_get_unless_zero(&zhdr->refcount) == 0) {
1360                                 zhdr = NULL;
1361                                 break;
1362                         }
1363                         if (!z3fold_page_trylock(zhdr)) {
1364                                 if (kref_put(&zhdr->refcount,
1365                                                 release_z3fold_page))
1366                                         atomic64_dec(&pool->pages_nr);
1367                                 zhdr = NULL;
1368                                 continue; /* can't evict at this point */
1369                         }
1370
1371                         /* test_and_set_bit is of course atomic, but we still
1372                          * need to do it under page lock, otherwise checking
1373                          * that bit in __z3fold_alloc wouldn't make sense
1374                          */
1375                         if (zhdr->foreign_handles ||
1376                             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1377                                 if (kref_put(&zhdr->refcount,
1378                                                 release_z3fold_page))
1379                                         atomic64_dec(&pool->pages_nr);
1380                                 else
1381                                         z3fold_page_unlock(zhdr);
1382                                 zhdr = NULL;
1383                                 continue; /* can't evict such page */
1384                         }
1385                         list_del_init(&zhdr->buddy);
1386                         zhdr->cpu = -1;
1387                         break;
1388                 }
1389
1390                 if (!zhdr)
1391                         break;
1392
1393                 list_del_init(&page->lru);
1394                 spin_unlock(&pool->lock);
1395
1396                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1397                         /*
1398                          * We need encode the handles before unlocking, and
1399                          * use our local slots structure because z3fold_free
1400                          * can zero out zhdr->slots and we can't do much
1401                          * about that
1402                          */
1403                         first_handle = 0;
1404                         last_handle = 0;
1405                         middle_handle = 0;
1406                         memset(slots.slot, 0, sizeof(slots.slot));
1407                         if (zhdr->first_chunks)
1408                                 first_handle = __encode_handle(zhdr, &slots,
1409                                                                 FIRST);
1410                         if (zhdr->middle_chunks)
1411                                 middle_handle = __encode_handle(zhdr, &slots,
1412                                                                 MIDDLE);
1413                         if (zhdr->last_chunks)
1414                                 last_handle = __encode_handle(zhdr, &slots,
1415                                                                 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                 }
1431                 if (first_handle) {
1432                         ret = pool->ops->evict(pool, first_handle);
1433                         if (ret)
1434                                 goto next;
1435                 }
1436                 if (last_handle) {
1437                         ret = pool->ops->evict(pool, last_handle);
1438                         if (ret)
1439                                 goto next;
1440                 }
1441 next:
1442                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1443                         if (ret == 0) {
1444                                 free_z3fold_page(page, true);
1445                                 atomic64_dec(&pool->pages_nr);
1446                                 return 0;
1447                         }
1448                         spin_lock(&pool->lock);
1449                         list_add(&page->lru, &pool->lru);
1450                         spin_unlock(&pool->lock);
1451                         clear_bit(PAGE_CLAIMED, &page->private);
1452                 } else {
1453                         struct z3fold_buddy_slots *slots = zhdr->slots;
1454                         z3fold_page_lock(zhdr);
1455                         if (kref_put(&zhdr->refcount,
1456                                         release_z3fold_page_locked)) {
1457                                 kmem_cache_free(pool->c_handle, slots);
1458                                 atomic64_dec(&pool->pages_nr);
1459                                 return 0;
1460                         }
1461                         /*
1462                          * if we are here, the page is still not completely
1463                          * free. Take the global pool lock then to be able
1464                          * to add it back to the lru list
1465                          */
1466                         spin_lock(&pool->lock);
1467                         list_add(&page->lru, &pool->lru);
1468                         spin_unlock(&pool->lock);
1469                         z3fold_page_unlock(zhdr);
1470                         clear_bit(PAGE_CLAIMED, &page->private);
1471                 }
1472
1473                 /* We started off locked to we need to lock the pool back */
1474                 spin_lock(&pool->lock);
1475         }
1476         spin_unlock(&pool->lock);
1477         return -EAGAIN;
1478 }
1479
1480 /**
1481  * z3fold_map() - maps the allocation associated with the given handle
1482  * @pool:       pool in which the allocation resides
1483  * @handle:     handle associated with the allocation to be mapped
1484  *
1485  * Extracts the buddy number from handle and constructs the pointer to the
1486  * correct starting chunk within the page.
1487  *
1488  * Returns: a pointer to the mapped allocation
1489  */
1490 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1491 {
1492         struct z3fold_header *zhdr;
1493         struct page *page;
1494         void *addr;
1495         enum buddy buddy;
1496
1497         zhdr = get_z3fold_header(handle);
1498         addr = zhdr;
1499         page = virt_to_page(zhdr);
1500
1501         if (test_bit(PAGE_HEADLESS, &page->private))
1502                 goto out;
1503
1504         buddy = handle_to_buddy(handle);
1505         switch (buddy) {
1506         case FIRST:
1507                 addr += ZHDR_SIZE_ALIGNED;
1508                 break;
1509         case MIDDLE:
1510                 addr += zhdr->start_middle << CHUNK_SHIFT;
1511                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1512                 break;
1513         case LAST:
1514                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1515                 break;
1516         default:
1517                 pr_err("unknown buddy id %d\n", buddy);
1518                 WARN_ON(1);
1519                 addr = NULL;
1520                 break;
1521         }
1522
1523         if (addr)
1524                 zhdr->mapped_count++;
1525 out:
1526         put_z3fold_header(zhdr);
1527         return addr;
1528 }
1529
1530 /**
1531  * z3fold_unmap() - unmaps the allocation associated with the given handle
1532  * @pool:       pool in which the allocation resides
1533  * @handle:     handle associated with the allocation to be unmapped
1534  */
1535 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1536 {
1537         struct z3fold_header *zhdr;
1538         struct page *page;
1539         enum buddy buddy;
1540
1541         zhdr = get_z3fold_header(handle);
1542         page = virt_to_page(zhdr);
1543
1544         if (test_bit(PAGE_HEADLESS, &page->private))
1545                 return;
1546
1547         buddy = handle_to_buddy(handle);
1548         if (buddy == MIDDLE)
1549                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1550         zhdr->mapped_count--;
1551         put_z3fold_header(zhdr);
1552 }
1553
1554 /**
1555  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1556  * @pool:       pool whose size is being queried
1557  *
1558  * Returns: size in pages of the given pool.
1559  */
1560 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1561 {
1562         return atomic64_read(&pool->pages_nr);
1563 }
1564
1565 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1566 {
1567         struct z3fold_header *zhdr;
1568         struct z3fold_pool *pool;
1569
1570         VM_BUG_ON_PAGE(!PageMovable(page), page);
1571         VM_BUG_ON_PAGE(PageIsolated(page), page);
1572
1573         if (test_bit(PAGE_HEADLESS, &page->private))
1574                 return false;
1575
1576         zhdr = page_address(page);
1577         z3fold_page_lock(zhdr);
1578         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1579             test_bit(PAGE_STALE, &page->private))
1580                 goto out;
1581
1582         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1583                 goto out;
1584
1585         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1586                 goto out;
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(!test_bit(PAGE_CLAIMED, &page->private), page);
1614         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1615
1616         zhdr = page_address(page);
1617         pool = zhdr_to_pool(zhdr);
1618
1619         if (!z3fold_page_trylock(zhdr))
1620                 return -EAGAIN;
1621         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1622                 z3fold_page_unlock(zhdr);
1623                 clear_bit(PAGE_CLAIMED, &page->private);
1624                 return -EBUSY;
1625         }
1626         if (work_pending(&zhdr->work)) {
1627                 z3fold_page_unlock(zhdr);
1628                 return -EAGAIN;
1629         }
1630         new_zhdr = page_address(newpage);
1631         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1632         newpage->private = page->private;
1633         page->private = 0;
1634         z3fold_page_unlock(zhdr);
1635         spin_lock_init(&new_zhdr->page_lock);
1636         INIT_WORK(&new_zhdr->work, compact_page_work);
1637         /*
1638          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1639          * so we only have to reinitialize it.
1640          */
1641         INIT_LIST_HEAD(&new_zhdr->buddy);
1642         new_mapping = page_mapping(page);
1643         __ClearPageMovable(page);
1644         ClearPagePrivate(page);
1645
1646         get_page(newpage);
1647         z3fold_page_lock(new_zhdr);
1648         if (new_zhdr->first_chunks)
1649                 encode_handle(new_zhdr, FIRST);
1650         if (new_zhdr->last_chunks)
1651                 encode_handle(new_zhdr, LAST);
1652         if (new_zhdr->middle_chunks)
1653                 encode_handle(new_zhdr, MIDDLE);
1654         set_bit(NEEDS_COMPACTING, &newpage->private);
1655         new_zhdr->cpu = smp_processor_id();
1656         spin_lock(&pool->lock);
1657         list_add(&newpage->lru, &pool->lru);
1658         spin_unlock(&pool->lock);
1659         __SetPageMovable(newpage, new_mapping);
1660         z3fold_page_unlock(new_zhdr);
1661
1662         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1663
1664         page_mapcount_reset(page);
1665         clear_bit(PAGE_CLAIMED, &page->private);
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         clear_bit(PAGE_CLAIMED, &page->private);
1690         z3fold_page_unlock(zhdr);
1691 }
1692
1693 static const struct address_space_operations z3fold_aops = {
1694         .isolate_page = z3fold_page_isolate,
1695         .migratepage = z3fold_page_migrate,
1696         .putback_page = z3fold_page_putback,
1697 };
1698
1699 /*****************
1700  * zpool
1701  ****************/
1702
1703 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1704 {
1705         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1706                 return pool->zpool_ops->evict(pool->zpool, handle);
1707         else
1708                 return -ENOENT;
1709 }
1710
1711 static const struct z3fold_ops z3fold_zpool_ops = {
1712         .evict =        z3fold_zpool_evict
1713 };
1714
1715 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1716                                const struct zpool_ops *zpool_ops,
1717                                struct zpool *zpool)
1718 {
1719         struct z3fold_pool *pool;
1720
1721         pool = z3fold_create_pool(name, gfp,
1722                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1723         if (pool) {
1724                 pool->zpool = zpool;
1725                 pool->zpool_ops = zpool_ops;
1726         }
1727         return pool;
1728 }
1729
1730 static void z3fold_zpool_destroy(void *pool)
1731 {
1732         z3fold_destroy_pool(pool);
1733 }
1734
1735 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1736                         unsigned long *handle)
1737 {
1738         return z3fold_alloc(pool, size, gfp, handle);
1739 }
1740 static void z3fold_zpool_free(void *pool, unsigned long handle)
1741 {
1742         z3fold_free(pool, handle);
1743 }
1744
1745 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1746                         unsigned int *reclaimed)
1747 {
1748         unsigned int total = 0;
1749         int ret = -EINVAL;
1750
1751         while (total < pages) {
1752                 ret = z3fold_reclaim_page(pool, 8);
1753                 if (ret < 0)
1754                         break;
1755                 total++;
1756         }
1757
1758         if (reclaimed)
1759                 *reclaimed = total;
1760
1761         return ret;
1762 }
1763
1764 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1765                         enum zpool_mapmode mm)
1766 {
1767         return z3fold_map(pool, handle);
1768 }
1769 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1770 {
1771         z3fold_unmap(pool, handle);
1772 }
1773
1774 static u64 z3fold_zpool_total_size(void *pool)
1775 {
1776         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1777 }
1778
1779 static struct zpool_driver z3fold_zpool_driver = {
1780         .type =         "z3fold",
1781         .owner =        THIS_MODULE,
1782         .create =       z3fold_zpool_create,
1783         .destroy =      z3fold_zpool_destroy,
1784         .malloc =       z3fold_zpool_malloc,
1785         .free =         z3fold_zpool_free,
1786         .shrink =       z3fold_zpool_shrink,
1787         .map =          z3fold_zpool_map,
1788         .unmap =        z3fold_zpool_unmap,
1789         .total_size =   z3fold_zpool_total_size,
1790 };
1791
1792 MODULE_ALIAS("zpool-z3fold");
1793
1794 static int __init init_z3fold(void)
1795 {
1796         int ret;
1797
1798         /* Make sure the z3fold header is not larger than the page size */
1799         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1800         ret = z3fold_mount();
1801         if (ret)
1802                 return ret;
1803
1804         zpool_register_driver(&z3fold_zpool_driver);
1805
1806         return 0;
1807 }
1808
1809 static void __exit exit_z3fold(void)
1810 {
1811         z3fold_unmount();
1812         zpool_unregister_driver(&z3fold_zpool_driver);
1813 }
1814
1815 module_init(init_z3fold);
1816 module_exit(exit_z3fold);
1817
1818 MODULE_LICENSE("GPL");
1819 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1820 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");