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