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/preempt.h>
  38 #include <linux/workqueue.h>
  39 #include <linux/slab.h>
  40 #include <linux/spinlock.h>
  41 #include <linux/zpool.h>
  42 #include <linux/kmemleak.h>
  43
  44 /*
  45  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  46  * adjusting internal fragmentation.  It also determines the number of
  47  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  48  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  49  * in the beginning of an allocated page are occupied by z3fold header, so
  50  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  51  * which shows the max number of free chunks in z3fold page, also there will
  52  * be 63, or 62, respectively, freelists per pool.
  53  */
  54 #define NCHUNKS_ORDER   6
  55
  56 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  57 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  58 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  59 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  60 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  61 #define NCHUNKS         (TOTAL_CHUNKS - ZHDR_CHUNKS)
  62
  63 #define BUDDY_MASK      (0x3)
  64 #define BUDDY_SHIFT     2
  65 #define SLOTS_ALIGN     (0x40)
  66
  67 /*****************
  68  * Structures
  69 *****************/
  70 struct z3fold_pool;
  71
  72 enum buddy {
  73         HEADLESS = 0,
  74         FIRST,
  75         MIDDLE,
  76         LAST,
  77         BUDDIES_MAX = LAST
  78 };
  79
  80 struct z3fold_buddy_slots {
  81         /*
  82          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
  83          * be enough slots to hold all possible variants
  84          */
  85         unsigned long slot[BUDDY_MASK + 1];
  86         unsigned long pool; /* back link */
  87         rwlock_t lock;
  88 };
  89 #define HANDLE_FLAG_MASK        (0x03)
  90
  91 /*
  92  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  93  *                      z3fold page, except for HEADLESS pages
  94  * @buddy:              links the z3fold page into the relevant list in the
  95  *                      pool
  96  * @page_lock:          per-page lock
  97  * @refcount:           reference count for the z3fold page
  98  * @work:               work_struct for page layout optimization
  99  * @slots:              pointer to the structure holding buddy slots
 100  * @pool:               pointer to the containing pool
 101  * @cpu:                CPU which this page "belongs" to
 102  * @first_chunks:       the size of the first buddy in chunks, 0 if free
 103  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
 104  * @last_chunks:        the size of the last buddy in chunks, 0 if free
 105  * @first_num:          the starting number (for the first handle)
 106  * @mapped_count:       the number of objects currently mapped
 107  */
 108 struct z3fold_header {
 109         struct list_head buddy;
 110         spinlock_t page_lock;
 111         struct kref refcount;
 112         struct work_struct work;
 113         struct z3fold_buddy_slots *slots;
 114         struct z3fold_pool *pool;
 115         short cpu;
 116         unsigned short first_chunks;
 117         unsigned short middle_chunks;
 118         unsigned short last_chunks;
 119         unsigned short start_middle;
 120         unsigned short first_num:2;
 121         unsigned short mapped_count:2;
 122         unsigned short foreign_handles:2;
 123 };
 124
 125 /**
 126  * struct z3fold_pool - stores metadata for each z3fold pool
 127  * @name:       pool name
 128  * @lock:       protects pool unbuddied lists
 129  * @stale_lock: protects pool stale page list
 130  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 131  *              buddies; the list each z3fold page is added to depends on
 132  *              the size of its free region.
 133  * @stale:      list of pages marked for freeing
 134  * @pages_nr:   number of z3fold pages in the pool.
 135  * @c_handle:   cache for z3fold_buddy_slots allocation
 136  * @zpool:      zpool driver
 137  * @zpool_ops:  zpool operations structure with an evict callback
 138  * @compact_wq: workqueue for page layout background optimization
 139  * @release_wq: workqueue for safe page release
 140  * @work:       work_struct for safe page release
 141  *
 142  * This structure is allocated at pool creation time and maintains metadata
 143  * pertaining to a particular z3fold pool.
 144  */
 145 struct z3fold_pool {
 146         const char *name;
 147         spinlock_t lock;
 148         spinlock_t stale_lock;
 149         struct list_head *unbuddied;
 150         struct list_head stale;
 151         atomic64_t pages_nr;
 152         struct kmem_cache *c_handle;
 153         struct workqueue_struct *compact_wq;
 154         struct workqueue_struct *release_wq;
 155         struct work_struct work;
 156 };
 157
 158 /*
 159  * Internal z3fold page flags
 160  */
 161 enum z3fold_page_flags {
 162         PAGE_HEADLESS = 0,
 163         MIDDLE_CHUNK_MAPPED,
 164         NEEDS_COMPACTING,
 165         PAGE_STALE,
 166         PAGE_CLAIMED, /* by either reclaim or free */
 167         PAGE_MIGRATED, /* page is migrated and soon to be released */
 168 };
 169
 170 /*
 171  * handle flags, go under HANDLE_FLAG_MASK
 172  */
 173 enum z3fold_handle_flags {
 174         HANDLES_NOFREE = 0,
 175 };
 176
 177 /*
 178  * Forward declarations
 179  */
 180 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 181 static void compact_page_work(struct work_struct *w);
 182
 183 /*****************
 184  * Helpers
 185 *****************/
 186
 187 /* Converts an allocation size in bytes to size in z3fold chunks */
 188 static int size_to_chunks(size_t size)
 189 {
 190         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 191 }
 192
 193 #define for_each_unbuddied_list(_iter, _begin) \
 194         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 195
 196 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 197                                                         gfp_t gfp)
 198 {
 199         struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
 200                                                              gfp);
 201
 202         if (slots) {
 203                 /* It will be freed separately in free_handle(). */
 204                 kmemleak_not_leak(slots);
 205                 slots->pool = (unsigned long)pool;
 206                 rwlock_init(&slots->lock);
 207         }
 208
 209         return slots;
 210 }
 211
 212 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 213 {
 214         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 215 }
 216
 217 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 218 {
 219         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 220 }
 221
 222 /* Lock a z3fold page */
 223 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 224 {
 225         spin_lock(&zhdr->page_lock);
 226 }
 227
 228 /* Try to lock a z3fold page */
 229 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 230 {
 231         return spin_trylock(&zhdr->page_lock);
 232 }
 233
 234 /* Unlock a z3fold page */
 235 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 236 {
 237         spin_unlock(&zhdr->page_lock);
 238 }
 239
 240 /* return locked z3fold page if it's not headless */
 241 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 242 {
 243         struct z3fold_buddy_slots *slots;
 244         struct z3fold_header *zhdr;
 245         int locked = 0;
 246
 247         if (!(handle & (1 << PAGE_HEADLESS))) {
 248                 slots = handle_to_slots(handle);
 249                 do {
 250                         unsigned long addr;
 251
 252                         read_lock(&slots->lock);
 253                         addr = *(unsigned long *)handle;
 254                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 255                         locked = z3fold_page_trylock(zhdr);
 256                         read_unlock(&slots->lock);
 257                         if (locked) {
 258                                 struct page *page = virt_to_page(zhdr);
 259
 260                                 if (!test_bit(PAGE_MIGRATED, &page->private))
 261                                         break;
 262                                 z3fold_page_unlock(zhdr);
 263                         }
 264                         cpu_relax();
 265                 } while (true);
 266         } else {
 267                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 268         }
 269
 270         return zhdr;
 271 }
 272
 273 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 274 {
 275         struct page *page = virt_to_page(zhdr);
 276
 277         if (!test_bit(PAGE_HEADLESS, &page->private))
 278                 z3fold_page_unlock(zhdr);
 279 }
 280
 281 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 282 {
 283         struct z3fold_buddy_slots *slots;
 284         int i;
 285         bool is_free;
 286
 287         if (WARN_ON(*(unsigned long *)handle == 0))
 288                 return;
 289
 290         slots = handle_to_slots(handle);
 291         write_lock(&slots->lock);
 292         *(unsigned long *)handle = 0;
 293
 294         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 295                 write_unlock(&slots->lock);
 296                 return; /* simple case, nothing else to do */
 297         }
 298
 299         if (zhdr->slots != slots)
 300                 zhdr->foreign_handles--;
 301
 302         is_free = true;
 303         for (i = 0; i <= BUDDY_MASK; i++) {
 304                 if (slots->slot[i]) {
 305                         is_free = false;
 306                         break;
 307                 }
 308         }
 309         write_unlock(&slots->lock);
 310
 311         if (is_free) {
 312                 struct z3fold_pool *pool = slots_to_pool(slots);
 313
 314                 if (zhdr->slots == slots)
 315                         zhdr->slots = NULL;
 316                 kmem_cache_free(pool->c_handle, slots);
 317         }
 318 }
 319
 320 /* Initializes the z3fold header of a newly allocated z3fold page */
 321 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 322                                         struct z3fold_pool *pool, gfp_t gfp)
 323 {
 324         struct z3fold_header *zhdr = page_address(page);
 325         struct z3fold_buddy_slots *slots;
 326
 327         clear_bit(PAGE_HEADLESS, &page->private);
 328         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 329         clear_bit(NEEDS_COMPACTING, &page->private);
 330         clear_bit(PAGE_STALE, &page->private);
 331         clear_bit(PAGE_CLAIMED, &page->private);
 332         clear_bit(PAGE_MIGRATED, &page->private);
 333         if (headless)
 334                 return zhdr;
 335
 336         slots = alloc_slots(pool, gfp);
 337         if (!slots)
 338                 return NULL;
 339
 340         memset(zhdr, 0, sizeof(*zhdr));
 341         spin_lock_init(&zhdr->page_lock);
 342         kref_init(&zhdr->refcount);
 343         zhdr->cpu = -1;
 344         zhdr->slots = slots;
 345         zhdr->pool = pool;
 346         INIT_LIST_HEAD(&zhdr->buddy);
 347         INIT_WORK(&zhdr->work, compact_page_work);
 348         return zhdr;
 349 }
 350
 351 /* Resets the struct page fields and frees the page */
 352 static void free_z3fold_page(struct page *page, bool headless)
 353 {
 354         if (!headless) {
 355                 lock_page(page);
 356                 __ClearPageMovable(page);
 357                 unlock_page(page);
 358         }
 359         __free_page(page);
 360 }
 361
 362 /* Helper function to build the index */
 363 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 364 {
 365         return (bud + zhdr->first_num) & BUDDY_MASK;
 366 }
 367
 368 /*
 369  * Encodes the handle of a particular buddy within a z3fold page
 370  * Pool lock should be held as this function accesses first_num
 371  */
 372 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 373                                 struct z3fold_buddy_slots *slots,
 374                                 enum buddy bud)
 375 {
 376         unsigned long h = (unsigned long)zhdr;
 377         int idx = 0;
 378
 379         /*
 380          * For a headless page, its handle is its pointer with the extra
 381          * PAGE_HEADLESS bit set
 382          */
 383         if (bud == HEADLESS)
 384                 return h | (1 << PAGE_HEADLESS);
 385
 386         /* otherwise, return pointer to encoded handle */
 387         idx = __idx(zhdr, bud);
 388         h += idx;
 389         if (bud == LAST)
 390                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 391
 392         write_lock(&slots->lock);
 393         slots->slot[idx] = h;
 394         write_unlock(&slots->lock);
 395         return (unsigned long)&slots->slot[idx];
 396 }
 397
 398 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 399 {
 400         return __encode_handle(zhdr, zhdr->slots, bud);
 401 }
 402
 403 /* only for LAST bud, returns zero otherwise */
 404 static unsigned short handle_to_chunks(unsigned long handle)
 405 {
 406         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 407         unsigned long addr;
 408
 409         read_lock(&slots->lock);
 410         addr = *(unsigned long *)handle;
 411         read_unlock(&slots->lock);
 412         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 413 }
 414
 415 /*
 416  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 417  *  but that doesn't matter. because the masking will result in the
 418  *  correct buddy number.
 419  */
 420 static enum buddy handle_to_buddy(unsigned long handle)
 421 {
 422         struct z3fold_header *zhdr;
 423         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 424         unsigned long addr;
 425
 426         read_lock(&slots->lock);
 427         WARN_ON(handle & (1 << PAGE_HEADLESS));
 428         addr = *(unsigned long *)handle;
 429         read_unlock(&slots->lock);
 430         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 431         return (addr - zhdr->first_num) & BUDDY_MASK;
 432 }
 433
 434 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 435 {
 436         return zhdr->pool;
 437 }
 438
 439 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 440 {
 441         struct page *page = virt_to_page(zhdr);
 442         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 443
 444         WARN_ON(!list_empty(&zhdr->buddy));
 445         set_bit(PAGE_STALE, &page->private);
 446         clear_bit(NEEDS_COMPACTING, &page->private);
 447         spin_lock(&pool->lock);
 448         spin_unlock(&pool->lock);
 449
 450         if (locked)
 451                 z3fold_page_unlock(zhdr);
 452
 453         spin_lock(&pool->stale_lock);
 454         list_add(&zhdr->buddy, &pool->stale);
 455         queue_work(pool->release_wq, &pool->work);
 456         spin_unlock(&pool->stale_lock);
 457
 458         atomic64_dec(&pool->pages_nr);
 459 }
 460
 461 static void release_z3fold_page_locked(struct kref *ref)
 462 {
 463         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 464                                                 refcount);
 465         WARN_ON(z3fold_page_trylock(zhdr));
 466         __release_z3fold_page(zhdr, true);
 467 }
 468
 469 static void release_z3fold_page_locked_list(struct kref *ref)
 470 {
 471         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 472                                                refcount);
 473         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 474
 475         spin_lock(&pool->lock);
 476         list_del_init(&zhdr->buddy);
 477         spin_unlock(&pool->lock);
 478
 479         WARN_ON(z3fold_page_trylock(zhdr));
 480         __release_z3fold_page(zhdr, true);
 481 }
 482
 483 static void free_pages_work(struct work_struct *w)
 484 {
 485         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 486
 487         spin_lock(&pool->stale_lock);
 488         while (!list_empty(&pool->stale)) {
 489                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 490                                                 struct z3fold_header, buddy);
 491                 struct page *page = virt_to_page(zhdr);
 492
 493                 list_del(&zhdr->buddy);
 494                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 495                         continue;
 496                 spin_unlock(&pool->stale_lock);
 497                 cancel_work_sync(&zhdr->work);
 498                 free_z3fold_page(page, false);
 499                 cond_resched();
 500                 spin_lock(&pool->stale_lock);
 501         }
 502         spin_unlock(&pool->stale_lock);
 503 }
 504
 505 /*
 506  * Returns the number of free chunks in a z3fold page.
 507  * NB: can't be used with HEADLESS pages.
 508  */
 509 static int num_free_chunks(struct z3fold_header *zhdr)
 510 {
 511         int nfree;
 512         /*
 513          * If there is a middle object, pick up the bigger free space
 514          * either before or after it. Otherwise just subtract the number
 515          * of chunks occupied by the first and the last objects.
 516          */
 517         if (zhdr->middle_chunks != 0) {
 518                 int nfree_before = zhdr->first_chunks ?
 519                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 520                 int nfree_after = zhdr->last_chunks ?
 521                         0 : TOTAL_CHUNKS -
 522                                 (zhdr->start_middle + zhdr->middle_chunks);
 523                 nfree = max(nfree_before, nfree_after);
 524         } else
 525                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 526         return nfree;
 527 }
 528
 529 /* Add to the appropriate unbuddied list */
 530 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 531                                 struct z3fold_header *zhdr)
 532 {
 533         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 534                         zhdr->middle_chunks == 0) {
 535                 struct list_head *unbuddied;
 536                 int freechunks = num_free_chunks(zhdr);
 537
 538                 migrate_disable();
 539                 unbuddied = this_cpu_ptr(pool->unbuddied);
 540                 spin_lock(&pool->lock);
 541                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 542                 spin_unlock(&pool->lock);
 543                 zhdr->cpu = smp_processor_id();
 544                 migrate_enable();
 545         }
 546 }
 547
 548 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 549 {
 550         enum buddy bud = HEADLESS;
 551
 552         if (zhdr->middle_chunks) {
 553                 if (!zhdr->first_chunks &&
 554                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 555                         bud = FIRST;
 556                 else if (!zhdr->last_chunks)
 557                         bud = LAST;
 558         } else {
 559                 if (!zhdr->first_chunks)
 560                         bud = FIRST;
 561                 else if (!zhdr->last_chunks)
 562                         bud = LAST;
 563                 else
 564                         bud = MIDDLE;
 565         }
 566
 567         return bud;
 568 }
 569
 570 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 571                                 unsigned short dst_chunk)
 572 {
 573         void *beg = zhdr;
 574         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 575                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 576                        zhdr->middle_chunks << CHUNK_SHIFT);
 577 }
 578
 579 static inline bool buddy_single(struct z3fold_header *zhdr)
 580 {
 581         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 582                         (zhdr->first_chunks && zhdr->last_chunks) ||
 583                         (zhdr->middle_chunks && zhdr->last_chunks));
 584 }
 585
 586 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 587 {
 588         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 589         void *p = zhdr;
 590         unsigned long old_handle = 0;
 591         size_t sz = 0;
 592         struct z3fold_header *new_zhdr = NULL;
 593         int first_idx = __idx(zhdr, FIRST);
 594         int middle_idx = __idx(zhdr, MIDDLE);
 595         int last_idx = __idx(zhdr, LAST);
 596         unsigned short *moved_chunks = NULL;
 597
 598         /*
 599          * No need to protect slots here -- all the slots are "local" and
 600          * the page lock is already taken
 601          */
 602         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 603                 p += ZHDR_SIZE_ALIGNED;
 604                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 605                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 606                 moved_chunks = &zhdr->first_chunks;
 607         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 608                 p += zhdr->start_middle << CHUNK_SHIFT;
 609                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 610                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 611                 moved_chunks = &zhdr->middle_chunks;
 612         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 613                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 614                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 615                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 616                 moved_chunks = &zhdr->last_chunks;
 617         }
 618
 619         if (sz > 0) {
 620                 enum buddy new_bud = HEADLESS;
 621                 short chunks = size_to_chunks(sz);
 622                 void *q;
 623
 624                 new_zhdr = __z3fold_alloc(pool, sz, false);
 625                 if (!new_zhdr)
 626                         return NULL;
 627
 628                 if (WARN_ON(new_zhdr == zhdr))
 629                         goto out_fail;
 630
 631                 new_bud = get_free_buddy(new_zhdr, chunks);
 632                 q = new_zhdr;
 633                 switch (new_bud) {
 634                 case FIRST:
 635                         new_zhdr->first_chunks = chunks;
 636                         q += ZHDR_SIZE_ALIGNED;
 637                         break;
 638                 case MIDDLE:
 639                         new_zhdr->middle_chunks = chunks;
 640                         new_zhdr->start_middle =
 641                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 642                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 643                         break;
 644                 case LAST:
 645                         new_zhdr->last_chunks = chunks;
 646                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 647                         break;
 648                 default:
 649                         goto out_fail;
 650                 }
 651                 new_zhdr->foreign_handles++;
 652                 memcpy(q, p, sz);
 653                 write_lock(&zhdr->slots->lock);
 654                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 655                         __idx(new_zhdr, new_bud);
 656                 if (new_bud == LAST)
 657                         *(unsigned long *)old_handle |=
 658                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 659                 write_unlock(&zhdr->slots->lock);
 660                 add_to_unbuddied(pool, new_zhdr);
 661                 z3fold_page_unlock(new_zhdr);
 662
 663                 *moved_chunks = 0;
 664         }
 665
 666         return new_zhdr;
 667
 668 out_fail:
 669         if (new_zhdr && !kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) {
 670                 add_to_unbuddied(pool, new_zhdr);
 671                 z3fold_page_unlock(new_zhdr);
 672         }
 673         return NULL;
 674
 675 }
 676
 677 #define BIG_CHUNK_GAP   3
 678 /* Has to be called with lock held */
 679 static int z3fold_compact_page(struct z3fold_header *zhdr)
 680 {
 681         struct page *page = virt_to_page(zhdr);
 682
 683         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 684                 return 0; /* can't move middle chunk, it's used */
 685
 686         if (unlikely(PageIsolated(page)))
 687                 return 0;
 688
 689         if (zhdr->middle_chunks == 0)
 690                 return 0; /* nothing to compact */
 691
 692         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 693                 /* move to the beginning */
 694                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 695                 zhdr->first_chunks = zhdr->middle_chunks;
 696                 zhdr->middle_chunks = 0;
 697                 zhdr->start_middle = 0;
 698                 zhdr->first_num++;
 699                 return 1;
 700         }
 701
 702         /*
 703          * moving data is expensive, so let's only do that if
 704          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 705          */
 706         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 707             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 708                         BIG_CHUNK_GAP) {
 709                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 710                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 711                 return 1;
 712         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 713                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 714                                         + zhdr->middle_chunks) >=
 715                         BIG_CHUNK_GAP) {
 716                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 717                         zhdr->middle_chunks;
 718                 mchunk_memmove(zhdr, new_start);
 719                 zhdr->start_middle = new_start;
 720                 return 1;
 721         }
 722
 723         return 0;
 724 }
 725
 726 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 727 {
 728         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 729         struct page *page;
 730
 731         page = virt_to_page(zhdr);
 732         if (locked)
 733                 WARN_ON(z3fold_page_trylock(zhdr));
 734         else
 735                 z3fold_page_lock(zhdr);
 736         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 737                 z3fold_page_unlock(zhdr);
 738                 return;
 739         }
 740         spin_lock(&pool->lock);
 741         list_del_init(&zhdr->buddy);
 742         spin_unlock(&pool->lock);
 743
 744         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 745                 return;
 746
 747         if (test_bit(PAGE_STALE, &page->private) ||
 748             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 749                 z3fold_page_unlock(zhdr);
 750                 return;
 751         }
 752
 753         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 754             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 755                 if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 756                         clear_bit(PAGE_CLAIMED, &page->private);
 757                         z3fold_page_unlock(zhdr);
 758                 }
 759                 return;
 760         }
 761
 762         z3fold_compact_page(zhdr);
 763         add_to_unbuddied(pool, zhdr);
 764         clear_bit(PAGE_CLAIMED, &page->private);
 765         z3fold_page_unlock(zhdr);
 766 }
 767
 768 static void compact_page_work(struct work_struct *w)
 769 {
 770         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 771                                                 work);
 772
 773         do_compact_page(zhdr, false);
 774 }
 775
 776 /* returns _locked_ z3fold page header or NULL */
 777 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 778                                                 size_t size, bool can_sleep)
 779 {
 780         struct z3fold_header *zhdr = NULL;
 781         struct page *page;
 782         struct list_head *unbuddied;
 783         int chunks = size_to_chunks(size), i;
 784
 785 lookup:
 786         migrate_disable();
 787         /* First, try to find an unbuddied z3fold page. */
 788         unbuddied = this_cpu_ptr(pool->unbuddied);
 789         for_each_unbuddied_list(i, chunks) {
 790                 struct list_head *l = &unbuddied[i];
 791
 792                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 793                                         struct z3fold_header, buddy);
 794
 795                 if (!zhdr)
 796                         continue;
 797
 798                 /* Re-check under lock. */
 799                 spin_lock(&pool->lock);
 800                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 801                                                 struct z3fold_header, buddy)) ||
 802                     !z3fold_page_trylock(zhdr)) {
 803                         spin_unlock(&pool->lock);
 804                         zhdr = NULL;
 805                         migrate_enable();
 806                         if (can_sleep)
 807                                 cond_resched();
 808                         goto lookup;
 809                 }
 810                 list_del_init(&zhdr->buddy);
 811                 zhdr->cpu = -1;
 812                 spin_unlock(&pool->lock);
 813
 814                 page = virt_to_page(zhdr);
 815                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 816                     test_bit(PAGE_CLAIMED, &page->private)) {
 817                         z3fold_page_unlock(zhdr);
 818                         zhdr = NULL;
 819                         migrate_enable();
 820                         if (can_sleep)
 821                                 cond_resched();
 822                         goto lookup;
 823                 }
 824
 825                 /*
 826                  * this page could not be removed from its unbuddied
 827                  * list while pool lock was held, and then we've taken
 828                  * page lock so kref_put could not be called before
 829                  * we got here, so it's safe to just call kref_get()
 830                  */
 831                 kref_get(&zhdr->refcount);
 832                 break;
 833         }
 834         migrate_enable();
 835
 836         if (!zhdr) {
 837                 int cpu;
 838
 839                 /* look for _exact_ match on other cpus' lists */
 840                 for_each_online_cpu(cpu) {
 841                         struct list_head *l;
 842
 843                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 844                         spin_lock(&pool->lock);
 845                         l = &unbuddied[chunks];
 846
 847                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 848                                                 struct z3fold_header, buddy);
 849
 850                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 851                                 spin_unlock(&pool->lock);
 852                                 zhdr = NULL;
 853                                 continue;
 854                         }
 855                         list_del_init(&zhdr->buddy);
 856                         zhdr->cpu = -1;
 857                         spin_unlock(&pool->lock);
 858
 859                         page = virt_to_page(zhdr);
 860                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 861                             test_bit(PAGE_CLAIMED, &page->private)) {
 862                                 z3fold_page_unlock(zhdr);
 863                                 zhdr = NULL;
 864                                 if (can_sleep)
 865                                         cond_resched();
 866                                 continue;
 867                         }
 868                         kref_get(&zhdr->refcount);
 869                         break;
 870                 }
 871         }
 872
 873         if (zhdr && !zhdr->slots) {
 874                 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
 875                 if (!zhdr->slots)
 876                         goto out_fail;
 877         }
 878         return zhdr;
 879
 880 out_fail:
 881         if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 882                 add_to_unbuddied(pool, zhdr);
 883                 z3fold_page_unlock(zhdr);
 884         }
 885         return NULL;
 886 }
 887
 888 /*
 889  * API Functions
 890  */
 891
 892 /**
 893  * z3fold_create_pool() - create a new z3fold pool
 894  * @name:       pool name
 895  * @gfp:        gfp flags when allocating the z3fold pool structure
 896  *
 897  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 898  * failed.
 899  */
 900 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp)
 901 {
 902         struct z3fold_pool *pool = NULL;
 903         int i, cpu;
 904
 905         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 906         if (!pool)
 907                 goto out;
 908         pool->c_handle = kmem_cache_create("z3fold_handle",
 909                                 sizeof(struct z3fold_buddy_slots),
 910                                 SLOTS_ALIGN, 0, NULL);
 911         if (!pool->c_handle)
 912                 goto out_c;
 913         spin_lock_init(&pool->lock);
 914         spin_lock_init(&pool->stale_lock);
 915         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 916                                          __alignof__(struct list_head));
 917         if (!pool->unbuddied)
 918                 goto out_pool;
 919         for_each_possible_cpu(cpu) {
 920                 struct list_head *unbuddied =
 921                                 per_cpu_ptr(pool->unbuddied, cpu);
 922                 for_each_unbuddied_list(i, 0)
 923                         INIT_LIST_HEAD(&unbuddied[i]);
 924         }
 925         INIT_LIST_HEAD(&pool->stale);
 926         atomic64_set(&pool->pages_nr, 0);
 927         pool->name = name;
 928         pool->compact_wq = create_singlethread_workqueue(pool->name);
 929         if (!pool->compact_wq)
 930                 goto out_unbuddied;
 931         pool->release_wq = create_singlethread_workqueue(pool->name);
 932         if (!pool->release_wq)
 933                 goto out_wq;
 934         INIT_WORK(&pool->work, free_pages_work);
 935         return pool;
 936
 937 out_wq:
 938         destroy_workqueue(pool->compact_wq);
 939 out_unbuddied:
 940         free_percpu(pool->unbuddied);
 941 out_pool:
 942         kmem_cache_destroy(pool->c_handle);
 943 out_c:
 944         kfree(pool);
 945 out:
 946         return NULL;
 947 }
 948
 949 /**
 950  * z3fold_destroy_pool() - destroys an existing z3fold pool
 951  * @pool:       the z3fold pool to be destroyed
 952  *
 953  * The pool should be emptied before this function is called.
 954  */
 955 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 956 {
 957         kmem_cache_destroy(pool->c_handle);
 958
 959         /*
 960          * We need to destroy pool->compact_wq before pool->release_wq,
 961          * as any pending work on pool->compact_wq will call
 962          * queue_work(pool->release_wq, &pool->work).
 963          *
 964          * There are still outstanding pages until both workqueues are drained,
 965          * so we cannot unregister migration until then.
 966          */
 967
 968         destroy_workqueue(pool->compact_wq);
 969         destroy_workqueue(pool->release_wq);
 970         free_percpu(pool->unbuddied);
 971         kfree(pool);
 972 }
 973
 974 static const struct movable_operations z3fold_mops;
 975
 976 /**
 977  * z3fold_alloc() - allocates a region of a given size
 978  * @pool:       z3fold pool from which to allocate
 979  * @size:       size in bytes of the desired allocation
 980  * @gfp:        gfp flags used if the pool needs to grow
 981  * @handle:     handle of the new allocation
 982  *
 983  * This function will attempt to find a free region in the pool large enough to
 984  * satisfy the allocation request.  A search of the unbuddied lists is
 985  * performed first. If no suitable free region is found, then a new page is
 986  * allocated and added to the pool to satisfy the request.
 987  *
 988  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 989  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 990  * a new page.
 991  */
 992 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 993                         unsigned long *handle)
 994 {
 995         int chunks = size_to_chunks(size);
 996         struct z3fold_header *zhdr = NULL;
 997         struct page *page = NULL;
 998         enum buddy bud;
 999         bool can_sleep = gfpflags_allow_blocking(gfp);
1000
1001         if (!size || (gfp & __GFP_HIGHMEM))
1002                 return -EINVAL;
1003
1004         if (size > PAGE_SIZE)
1005                 return -ENOSPC;
1006
1007         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1008                 bud = HEADLESS;
1009         else {
1010 retry:
1011                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1012                 if (zhdr) {
1013                         bud = get_free_buddy(zhdr, chunks);
1014                         if (bud == HEADLESS) {
1015                                 if (!kref_put(&zhdr->refcount,
1016                                              release_z3fold_page_locked))
1017                                         z3fold_page_unlock(zhdr);
1018                                 pr_err("No free chunks in unbuddied\n");
1019                                 WARN_ON(1);
1020                                 goto retry;
1021                         }
1022                         page = virt_to_page(zhdr);
1023                         goto found;
1024                 }
1025                 bud = FIRST;
1026         }
1027
1028         page = alloc_page(gfp);
1029         if (!page)
1030                 return -ENOMEM;
1031
1032         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1033         if (!zhdr) {
1034                 __free_page(page);
1035                 return -ENOMEM;
1036         }
1037         atomic64_inc(&pool->pages_nr);
1038
1039         if (bud == HEADLESS) {
1040                 set_bit(PAGE_HEADLESS, &page->private);
1041                 goto headless;
1042         }
1043         if (can_sleep) {
1044                 lock_page(page);
1045                 __SetPageMovable(page, &z3fold_mops);
1046                 unlock_page(page);
1047         } else {
1048                 WARN_ON(!trylock_page(page));
1049                 __SetPageMovable(page, &z3fold_mops);
1050                 unlock_page(page);
1051         }
1052         z3fold_page_lock(zhdr);
1053
1054 found:
1055         if (bud == FIRST)
1056                 zhdr->first_chunks = chunks;
1057         else if (bud == LAST)
1058                 zhdr->last_chunks = chunks;
1059         else {
1060                 zhdr->middle_chunks = chunks;
1061                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1062         }
1063         add_to_unbuddied(pool, zhdr);
1064
1065 headless:
1066         spin_lock(&pool->lock);
1067         *handle = encode_handle(zhdr, bud);
1068         spin_unlock(&pool->lock);
1069         if (bud != HEADLESS)
1070                 z3fold_page_unlock(zhdr);
1071
1072         return 0;
1073 }
1074
1075 /**
1076  * z3fold_free() - frees the allocation associated with the given handle
1077  * @pool:       pool in which the allocation resided
1078  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1079  *
1080  * In the case that the z3fold page in which the allocation resides is under
1081  * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1082  * only sets the first|middle|last_chunks to 0.  The page is actually freed
1083  * once all buddies are evicted (see z3fold_reclaim_page() below).
1084  */
1085 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1086 {
1087         struct z3fold_header *zhdr;
1088         struct page *page;
1089         enum buddy bud;
1090         bool page_claimed;
1091
1092         zhdr = get_z3fold_header(handle);
1093         page = virt_to_page(zhdr);
1094         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1095
1096         if (test_bit(PAGE_HEADLESS, &page->private)) {
1097                 /* if a headless page is under reclaim, just leave.
1098                  * NB: we use test_and_set_bit for a reason: if the bit
1099                  * has not been set before, we release this page
1100                  * immediately so we don't care about its value any more.
1101                  */
1102                 if (!page_claimed) {
1103                         put_z3fold_header(zhdr);
1104                         free_z3fold_page(page, true);
1105                         atomic64_dec(&pool->pages_nr);
1106                 }
1107                 return;
1108         }
1109
1110         /* Non-headless case */
1111         bud = handle_to_buddy(handle);
1112
1113         switch (bud) {
1114         case FIRST:
1115                 zhdr->first_chunks = 0;
1116                 break;
1117         case MIDDLE:
1118                 zhdr->middle_chunks = 0;
1119                 break;
1120         case LAST:
1121                 zhdr->last_chunks = 0;
1122                 break;
1123         default:
1124                 pr_err("%s: unknown bud %d\n", __func__, bud);
1125                 WARN_ON(1);
1126                 put_z3fold_header(zhdr);
1127                 return;
1128         }
1129
1130         if (!page_claimed)
1131                 free_handle(handle, zhdr);
1132         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list))
1133                 return;
1134         if (page_claimed) {
1135                 /* the page has not been claimed by us */
1136                 put_z3fold_header(zhdr);
1137                 return;
1138         }
1139         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1140                 clear_bit(PAGE_CLAIMED, &page->private);
1141                 put_z3fold_header(zhdr);
1142                 return;
1143         }
1144         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1145                 zhdr->cpu = -1;
1146                 kref_get(&zhdr->refcount);
1147                 clear_bit(PAGE_CLAIMED, &page->private);
1148                 do_compact_page(zhdr, true);
1149                 return;
1150         }
1151         kref_get(&zhdr->refcount);
1152         clear_bit(PAGE_CLAIMED, &page->private);
1153         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1154         put_z3fold_header(zhdr);
1155 }
1156
1157 /**
1158  * z3fold_map() - maps the allocation associated with the given handle
1159  * @pool:       pool in which the allocation resides
1160  * @handle:     handle associated with the allocation to be mapped
1161  *
1162  * Extracts the buddy number from handle and constructs the pointer to the
1163  * correct starting chunk within the page.
1164  *
1165  * Returns: a pointer to the mapped allocation
1166  */
1167 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1168 {
1169         struct z3fold_header *zhdr;
1170         struct page *page;
1171         void *addr;
1172         enum buddy buddy;
1173
1174         zhdr = get_z3fold_header(handle);
1175         addr = zhdr;
1176         page = virt_to_page(zhdr);
1177
1178         if (test_bit(PAGE_HEADLESS, &page->private))
1179                 goto out;
1180
1181         buddy = handle_to_buddy(handle);
1182         switch (buddy) {
1183         case FIRST:
1184                 addr += ZHDR_SIZE_ALIGNED;
1185                 break;
1186         case MIDDLE:
1187                 addr += zhdr->start_middle << CHUNK_SHIFT;
1188                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1189                 break;
1190         case LAST:
1191                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1192                 break;
1193         default:
1194                 pr_err("unknown buddy id %d\n", buddy);
1195                 WARN_ON(1);
1196                 addr = NULL;
1197                 break;
1198         }
1199
1200         if (addr)
1201                 zhdr->mapped_count++;
1202 out:
1203         put_z3fold_header(zhdr);
1204         return addr;
1205 }
1206
1207 /**
1208  * z3fold_unmap() - unmaps the allocation associated with the given handle
1209  * @pool:       pool in which the allocation resides
1210  * @handle:     handle associated with the allocation to be unmapped
1211  */
1212 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1213 {
1214         struct z3fold_header *zhdr;
1215         struct page *page;
1216         enum buddy buddy;
1217
1218         zhdr = get_z3fold_header(handle);
1219         page = virt_to_page(zhdr);
1220
1221         if (test_bit(PAGE_HEADLESS, &page->private))
1222                 return;
1223
1224         buddy = handle_to_buddy(handle);
1225         if (buddy == MIDDLE)
1226                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1227         zhdr->mapped_count--;
1228         put_z3fold_header(zhdr);
1229 }
1230
1231 /**
1232  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1233  * @pool:       pool whose size is being queried
1234  *
1235  * Returns: size in pages of the given pool.
1236  */
1237 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1238 {
1239         return atomic64_read(&pool->pages_nr);
1240 }
1241
1242 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1243 {
1244         struct z3fold_header *zhdr;
1245         struct z3fold_pool *pool;
1246
1247         VM_BUG_ON_PAGE(PageIsolated(page), page);
1248
1249         if (test_bit(PAGE_HEADLESS, &page->private))
1250                 return false;
1251
1252         zhdr = page_address(page);
1253         z3fold_page_lock(zhdr);
1254         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1255             test_bit(PAGE_STALE, &page->private))
1256                 goto out;
1257
1258         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1259                 goto out;
1260
1261         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1262                 goto out;
1263         pool = zhdr_to_pool(zhdr);
1264         spin_lock(&pool->lock);
1265         if (!list_empty(&zhdr->buddy))
1266                 list_del_init(&zhdr->buddy);
1267         spin_unlock(&pool->lock);
1268
1269         kref_get(&zhdr->refcount);
1270         z3fold_page_unlock(zhdr);
1271         return true;
1272
1273 out:
1274         z3fold_page_unlock(zhdr);
1275         return false;
1276 }
1277
1278 static int z3fold_page_migrate(struct page *newpage, struct page *page,
1279                 enum migrate_mode mode)
1280 {
1281         struct z3fold_header *zhdr, *new_zhdr;
1282         struct z3fold_pool *pool;
1283
1284         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1285         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1286         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1287
1288         zhdr = page_address(page);
1289         pool = zhdr_to_pool(zhdr);
1290
1291         if (!z3fold_page_trylock(zhdr))
1292                 return -EAGAIN;
1293         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1294                 clear_bit(PAGE_CLAIMED, &page->private);
1295                 z3fold_page_unlock(zhdr);
1296                 return -EBUSY;
1297         }
1298         if (work_pending(&zhdr->work)) {
1299                 z3fold_page_unlock(zhdr);
1300                 return -EAGAIN;
1301         }
1302         new_zhdr = page_address(newpage);
1303         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1304         newpage->private = page->private;
1305         set_bit(PAGE_MIGRATED, &page->private);
1306         z3fold_page_unlock(zhdr);
1307         spin_lock_init(&new_zhdr->page_lock);
1308         INIT_WORK(&new_zhdr->work, compact_page_work);
1309         /*
1310          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1311          * so we only have to reinitialize it.
1312          */
1313         INIT_LIST_HEAD(&new_zhdr->buddy);
1314         __ClearPageMovable(page);
1315
1316         get_page(newpage);
1317         z3fold_page_lock(new_zhdr);
1318         if (new_zhdr->first_chunks)
1319                 encode_handle(new_zhdr, FIRST);
1320         if (new_zhdr->last_chunks)
1321                 encode_handle(new_zhdr, LAST);
1322         if (new_zhdr->middle_chunks)
1323                 encode_handle(new_zhdr, MIDDLE);
1324         set_bit(NEEDS_COMPACTING, &newpage->private);
1325         new_zhdr->cpu = smp_processor_id();
1326         __SetPageMovable(newpage, &z3fold_mops);
1327         z3fold_page_unlock(new_zhdr);
1328
1329         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1330
1331         /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1332         page->private = 0;
1333         put_page(page);
1334         return 0;
1335 }
1336
1337 static void z3fold_page_putback(struct page *page)
1338 {
1339         struct z3fold_header *zhdr;
1340         struct z3fold_pool *pool;
1341
1342         zhdr = page_address(page);
1343         pool = zhdr_to_pool(zhdr);
1344
1345         z3fold_page_lock(zhdr);
1346         if (!list_empty(&zhdr->buddy))
1347                 list_del_init(&zhdr->buddy);
1348         INIT_LIST_HEAD(&page->lru);
1349         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
1350                 return;
1351         if (list_empty(&zhdr->buddy))
1352                 add_to_unbuddied(pool, zhdr);
1353         clear_bit(PAGE_CLAIMED, &page->private);
1354         z3fold_page_unlock(zhdr);
1355 }
1356
1357 static const struct movable_operations z3fold_mops = {
1358         .isolate_page = z3fold_page_isolate,
1359         .migrate_page = z3fold_page_migrate,
1360         .putback_page = z3fold_page_putback,
1361 };
1362
1363 /*****************
1364  * zpool
1365  ****************/
1366
1367 static void *z3fold_zpool_create(const char *name, gfp_t gfp)
1368 {
1369         return z3fold_create_pool(name, gfp);
1370 }
1371
1372 static void z3fold_zpool_destroy(void *pool)
1373 {
1374         z3fold_destroy_pool(pool);
1375 }
1376
1377 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1378                         unsigned long *handle)
1379 {
1380         return z3fold_alloc(pool, size, gfp, handle);
1381 }
1382 static void z3fold_zpool_free(void *pool, unsigned long handle)
1383 {
1384         z3fold_free(pool, handle);
1385 }
1386
1387 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1388                         enum zpool_mapmode mm)
1389 {
1390         return z3fold_map(pool, handle);
1391 }
1392 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1393 {
1394         z3fold_unmap(pool, handle);
1395 }
1396
1397 static u64 z3fold_zpool_total_size(void *pool)
1398 {
1399         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1400 }
1401
1402 static struct zpool_driver z3fold_zpool_driver = {
1403         .type =         "z3fold",
1404         .sleep_mapped = true,
1405         .owner =        THIS_MODULE,
1406         .create =       z3fold_zpool_create,
1407         .destroy =      z3fold_zpool_destroy,
1408         .malloc =       z3fold_zpool_malloc,
1409         .free =         z3fold_zpool_free,
1410         .map =          z3fold_zpool_map,
1411         .unmap =        z3fold_zpool_unmap,
1412         .total_size =   z3fold_zpool_total_size,
1413 };
1414
1415 MODULE_ALIAS("zpool-z3fold");
1416
1417 static int __init init_z3fold(void)
1418 {
1419         /*
1420          * Make sure the z3fold header is not larger than the page size and
1421          * there has remaining spaces for its buddy.
1422          */
1423         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1424         zpool_register_driver(&z3fold_zpool_driver);
1425
1426         return 0;
1427 }
1428
1429 static void __exit exit_z3fold(void)
1430 {
1431         zpool_unregister_driver(&z3fold_zpool_driver);
1432 }
1433
1434 module_init(init_z3fold);
1435 module_exit(exit_z3fold);
1436
1437 MODULE_LICENSE("GPL");
1438 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1439 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");