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/lru 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  * @lru:        list tracking the z3fold pages in LRU order by most recently
 134  *              added buddy.
 135  * @stale:      list of pages marked for freeing
 136  * @pages_nr:   number of z3fold pages in the pool.
 137  * @c_handle:   cache for z3fold_buddy_slots allocation
 138  * @zpool:      zpool driver
 139  * @zpool_ops:  zpool operations structure with an evict callback
 140  * @compact_wq: workqueue for page layout background optimization
 141  * @release_wq: workqueue for safe page release
 142  * @work:       work_struct for safe page release
 143  *
 144  * This structure is allocated at pool creation time and maintains metadata
 145  * pertaining to a particular z3fold pool.
 146  */
 147 struct z3fold_pool {
 148         const char *name;
 149         spinlock_t lock;
 150         spinlock_t stale_lock;
 151         struct list_head *unbuddied;
 152         struct list_head lru;
 153         struct list_head stale;
 154         atomic64_t pages_nr;
 155         struct kmem_cache *c_handle;
 156         struct zpool *zpool;
 157         const struct zpool_ops *zpool_ops;
 158         struct workqueue_struct *compact_wq;
 159         struct workqueue_struct *release_wq;
 160         struct work_struct work;
 161 };
 162
 163 /*
 164  * Internal z3fold page flags
 165  */
 166 enum z3fold_page_flags {
 167         PAGE_HEADLESS = 0,
 168         MIDDLE_CHUNK_MAPPED,
 169         NEEDS_COMPACTING,
 170         PAGE_STALE,
 171         PAGE_CLAIMED, /* by either reclaim or free */
 172         PAGE_MIGRATED, /* page is migrated and soon to be released */
 173 };
 174
 175 /*
 176  * handle flags, go under HANDLE_FLAG_MASK
 177  */
 178 enum z3fold_handle_flags {
 179         HANDLES_NOFREE = 0,
 180 };
 181
 182 /*
 183  * Forward declarations
 184  */
 185 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
 186 static void compact_page_work(struct work_struct *w);
 187
 188 /*****************
 189  * Helpers
 190 *****************/
 191
 192 /* Converts an allocation size in bytes to size in z3fold chunks */
 193 static int size_to_chunks(size_t size)
 194 {
 195         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 196 }
 197
 198 #define for_each_unbuddied_list(_iter, _begin) \
 199         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 200
 201 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
 202                                                         gfp_t gfp)
 203 {
 204         struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
 205                                                              gfp);
 206
 207         if (slots) {
 208                 /* It will be freed separately in free_handle(). */
 209                 kmemleak_not_leak(slots);
 210                 slots->pool = (unsigned long)pool;
 211                 rwlock_init(&slots->lock);
 212         }
 213
 214         return slots;
 215 }
 216
 217 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
 218 {
 219         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
 220 }
 221
 222 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
 223 {
 224         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
 225 }
 226
 227 /* Lock a z3fold page */
 228 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 229 {
 230         spin_lock(&zhdr->page_lock);
 231 }
 232
 233 /* Try to lock a z3fold page */
 234 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 235 {
 236         return spin_trylock(&zhdr->page_lock);
 237 }
 238
 239 /* Unlock a z3fold page */
 240 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 241 {
 242         spin_unlock(&zhdr->page_lock);
 243 }
 244
 245 /* return locked z3fold page if it's not headless */
 246 static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
 247 {
 248         struct z3fold_buddy_slots *slots;
 249         struct z3fold_header *zhdr;
 250         int locked = 0;
 251
 252         if (!(handle & (1 << PAGE_HEADLESS))) {
 253                 slots = handle_to_slots(handle);
 254                 do {
 255                         unsigned long addr;
 256
 257                         read_lock(&slots->lock);
 258                         addr = *(unsigned long *)handle;
 259                         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 260                         locked = z3fold_page_trylock(zhdr);
 261                         read_unlock(&slots->lock);
 262                         if (locked) {
 263                                 struct page *page = virt_to_page(zhdr);
 264
 265                                 if (!test_bit(PAGE_MIGRATED, &page->private))
 266                                         break;
 267                                 z3fold_page_unlock(zhdr);
 268                         }
 269                         cpu_relax();
 270                 } while (true);
 271         } else {
 272                 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
 273         }
 274
 275         return zhdr;
 276 }
 277
 278 static inline void put_z3fold_header(struct z3fold_header *zhdr)
 279 {
 280         struct page *page = virt_to_page(zhdr);
 281
 282         if (!test_bit(PAGE_HEADLESS, &page->private))
 283                 z3fold_page_unlock(zhdr);
 284 }
 285
 286 static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
 287 {
 288         struct z3fold_buddy_slots *slots;
 289         int i;
 290         bool is_free;
 291
 292         if (WARN_ON(*(unsigned long *)handle == 0))
 293                 return;
 294
 295         slots = handle_to_slots(handle);
 296         write_lock(&slots->lock);
 297         *(unsigned long *)handle = 0;
 298
 299         if (test_bit(HANDLES_NOFREE, &slots->pool)) {
 300                 write_unlock(&slots->lock);
 301                 return; /* simple case, nothing else to do */
 302         }
 303
 304         if (zhdr->slots != slots)
 305                 zhdr->foreign_handles--;
 306
 307         is_free = true;
 308         for (i = 0; i <= BUDDY_MASK; i++) {
 309                 if (slots->slot[i]) {
 310                         is_free = false;
 311                         break;
 312                 }
 313         }
 314         write_unlock(&slots->lock);
 315
 316         if (is_free) {
 317                 struct z3fold_pool *pool = slots_to_pool(slots);
 318
 319                 if (zhdr->slots == slots)
 320                         zhdr->slots = NULL;
 321                 kmem_cache_free(pool->c_handle, slots);
 322         }
 323 }
 324
 325 /* Initializes the z3fold header of a newly allocated z3fold page */
 326 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
 327                                         struct z3fold_pool *pool, gfp_t gfp)
 328 {
 329         struct z3fold_header *zhdr = page_address(page);
 330         struct z3fold_buddy_slots *slots;
 331
 332         INIT_LIST_HEAD(&page->lru);
 333         clear_bit(PAGE_HEADLESS, &page->private);
 334         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 335         clear_bit(NEEDS_COMPACTING, &page->private);
 336         clear_bit(PAGE_STALE, &page->private);
 337         clear_bit(PAGE_CLAIMED, &page->private);
 338         clear_bit(PAGE_MIGRATED, &page->private);
 339         if (headless)
 340                 return zhdr;
 341
 342         slots = alloc_slots(pool, gfp);
 343         if (!slots)
 344                 return NULL;
 345
 346         memset(zhdr, 0, sizeof(*zhdr));
 347         spin_lock_init(&zhdr->page_lock);
 348         kref_init(&zhdr->refcount);
 349         zhdr->cpu = -1;
 350         zhdr->slots = slots;
 351         zhdr->pool = pool;
 352         INIT_LIST_HEAD(&zhdr->buddy);
 353         INIT_WORK(&zhdr->work, compact_page_work);
 354         return zhdr;
 355 }
 356
 357 /* Resets the struct page fields and frees the page */
 358 static void free_z3fold_page(struct page *page, bool headless)
 359 {
 360         if (!headless) {
 361                 lock_page(page);
 362                 __ClearPageMovable(page);
 363                 unlock_page(page);
 364         }
 365         __free_page(page);
 366 }
 367
 368 /* Helper function to build the index */
 369 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
 370 {
 371         return (bud + zhdr->first_num) & BUDDY_MASK;
 372 }
 373
 374 /*
 375  * Encodes the handle of a particular buddy within a z3fold page
 376  * Pool lock should be held as this function accesses first_num
 377  */
 378 static unsigned long __encode_handle(struct z3fold_header *zhdr,
 379                                 struct z3fold_buddy_slots *slots,
 380                                 enum buddy bud)
 381 {
 382         unsigned long h = (unsigned long)zhdr;
 383         int idx = 0;
 384
 385         /*
 386          * For a headless page, its handle is its pointer with the extra
 387          * PAGE_HEADLESS bit set
 388          */
 389         if (bud == HEADLESS)
 390                 return h | (1 << PAGE_HEADLESS);
 391
 392         /* otherwise, return pointer to encoded handle */
 393         idx = __idx(zhdr, bud);
 394         h += idx;
 395         if (bud == LAST)
 396                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
 397
 398         write_lock(&slots->lock);
 399         slots->slot[idx] = h;
 400         write_unlock(&slots->lock);
 401         return (unsigned long)&slots->slot[idx];
 402 }
 403
 404 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 405 {
 406         return __encode_handle(zhdr, zhdr->slots, bud);
 407 }
 408
 409 /* only for LAST bud, returns zero otherwise */
 410 static unsigned short handle_to_chunks(unsigned long handle)
 411 {
 412         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 413         unsigned long addr;
 414
 415         read_lock(&slots->lock);
 416         addr = *(unsigned long *)handle;
 417         read_unlock(&slots->lock);
 418         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
 419 }
 420
 421 /*
 422  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 423  *  but that doesn't matter. because the masking will result in the
 424  *  correct buddy number.
 425  */
 426 static enum buddy handle_to_buddy(unsigned long handle)
 427 {
 428         struct z3fold_header *zhdr;
 429         struct z3fold_buddy_slots *slots = handle_to_slots(handle);
 430         unsigned long addr;
 431
 432         read_lock(&slots->lock);
 433         WARN_ON(handle & (1 << PAGE_HEADLESS));
 434         addr = *(unsigned long *)handle;
 435         read_unlock(&slots->lock);
 436         zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
 437         return (addr - zhdr->first_num) & BUDDY_MASK;
 438 }
 439
 440 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 441 {
 442         return zhdr->pool;
 443 }
 444
 445 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 446 {
 447         struct page *page = virt_to_page(zhdr);
 448         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 449
 450         WARN_ON(!list_empty(&zhdr->buddy));
 451         set_bit(PAGE_STALE, &page->private);
 452         clear_bit(NEEDS_COMPACTING, &page->private);
 453         spin_lock(&pool->lock);
 454         if (!list_empty(&page->lru))
 455                 list_del_init(&page->lru);
 456         spin_unlock(&pool->lock);
 457
 458         if (locked)
 459                 z3fold_page_unlock(zhdr);
 460
 461         spin_lock(&pool->stale_lock);
 462         list_add(&zhdr->buddy, &pool->stale);
 463         queue_work(pool->release_wq, &pool->work);
 464         spin_unlock(&pool->stale_lock);
 465
 466         atomic64_dec(&pool->pages_nr);
 467 }
 468
 469 static void release_z3fold_page_locked(struct kref *ref)
 470 {
 471         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 472                                                 refcount);
 473         WARN_ON(z3fold_page_trylock(zhdr));
 474         __release_z3fold_page(zhdr, true);
 475 }
 476
 477 static void release_z3fold_page_locked_list(struct kref *ref)
 478 {
 479         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 480                                                refcount);
 481         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 482
 483         spin_lock(&pool->lock);
 484         list_del_init(&zhdr->buddy);
 485         spin_unlock(&pool->lock);
 486
 487         WARN_ON(z3fold_page_trylock(zhdr));
 488         __release_z3fold_page(zhdr, true);
 489 }
 490
 491 static void free_pages_work(struct work_struct *w)
 492 {
 493         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 494
 495         spin_lock(&pool->stale_lock);
 496         while (!list_empty(&pool->stale)) {
 497                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 498                                                 struct z3fold_header, buddy);
 499                 struct page *page = virt_to_page(zhdr);
 500
 501                 list_del(&zhdr->buddy);
 502                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 503                         continue;
 504                 spin_unlock(&pool->stale_lock);
 505                 cancel_work_sync(&zhdr->work);
 506                 free_z3fold_page(page, false);
 507                 cond_resched();
 508                 spin_lock(&pool->stale_lock);
 509         }
 510         spin_unlock(&pool->stale_lock);
 511 }
 512
 513 /*
 514  * Returns the number of free chunks in a z3fold page.
 515  * NB: can't be used with HEADLESS pages.
 516  */
 517 static int num_free_chunks(struct z3fold_header *zhdr)
 518 {
 519         int nfree;
 520         /*
 521          * If there is a middle object, pick up the bigger free space
 522          * either before or after it. Otherwise just subtract the number
 523          * of chunks occupied by the first and the last objects.
 524          */
 525         if (zhdr->middle_chunks != 0) {
 526                 int nfree_before = zhdr->first_chunks ?
 527                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 528                 int nfree_after = zhdr->last_chunks ?
 529                         0 : TOTAL_CHUNKS -
 530                                 (zhdr->start_middle + zhdr->middle_chunks);
 531                 nfree = max(nfree_before, nfree_after);
 532         } else
 533                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 534         return nfree;
 535 }
 536
 537 /* Add to the appropriate unbuddied list */
 538 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 539                                 struct z3fold_header *zhdr)
 540 {
 541         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 542                         zhdr->middle_chunks == 0) {
 543                 struct list_head *unbuddied;
 544                 int freechunks = num_free_chunks(zhdr);
 545
 546                 migrate_disable();
 547                 unbuddied = this_cpu_ptr(pool->unbuddied);
 548                 spin_lock(&pool->lock);
 549                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 550                 spin_unlock(&pool->lock);
 551                 zhdr->cpu = smp_processor_id();
 552                 migrate_enable();
 553         }
 554 }
 555
 556 static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
 557 {
 558         enum buddy bud = HEADLESS;
 559
 560         if (zhdr->middle_chunks) {
 561                 if (!zhdr->first_chunks &&
 562                     chunks <= zhdr->start_middle - ZHDR_CHUNKS)
 563                         bud = FIRST;
 564                 else if (!zhdr->last_chunks)
 565                         bud = LAST;
 566         } else {
 567                 if (!zhdr->first_chunks)
 568                         bud = FIRST;
 569                 else if (!zhdr->last_chunks)
 570                         bud = LAST;
 571                 else
 572                         bud = MIDDLE;
 573         }
 574
 575         return bud;
 576 }
 577
 578 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 579                                 unsigned short dst_chunk)
 580 {
 581         void *beg = zhdr;
 582         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 583                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 584                        zhdr->middle_chunks << CHUNK_SHIFT);
 585 }
 586
 587 static inline bool buddy_single(struct z3fold_header *zhdr)
 588 {
 589         return !((zhdr->first_chunks && zhdr->middle_chunks) ||
 590                         (zhdr->first_chunks && zhdr->last_chunks) ||
 591                         (zhdr->middle_chunks && zhdr->last_chunks));
 592 }
 593
 594 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
 595 {
 596         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 597         void *p = zhdr;
 598         unsigned long old_handle = 0;
 599         size_t sz = 0;
 600         struct z3fold_header *new_zhdr = NULL;
 601         int first_idx = __idx(zhdr, FIRST);
 602         int middle_idx = __idx(zhdr, MIDDLE);
 603         int last_idx = __idx(zhdr, LAST);
 604         unsigned short *moved_chunks = NULL;
 605
 606         /*
 607          * No need to protect slots here -- all the slots are "local" and
 608          * the page lock is already taken
 609          */
 610         if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
 611                 p += ZHDR_SIZE_ALIGNED;
 612                 sz = zhdr->first_chunks << CHUNK_SHIFT;
 613                 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
 614                 moved_chunks = &zhdr->first_chunks;
 615         } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
 616                 p += zhdr->start_middle << CHUNK_SHIFT;
 617                 sz = zhdr->middle_chunks << CHUNK_SHIFT;
 618                 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
 619                 moved_chunks = &zhdr->middle_chunks;
 620         } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
 621                 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 622                 sz = zhdr->last_chunks << CHUNK_SHIFT;
 623                 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
 624                 moved_chunks = &zhdr->last_chunks;
 625         }
 626
 627         if (sz > 0) {
 628                 enum buddy new_bud = HEADLESS;
 629                 short chunks = size_to_chunks(sz);
 630                 void *q;
 631
 632                 new_zhdr = __z3fold_alloc(pool, sz, false);
 633                 if (!new_zhdr)
 634                         return NULL;
 635
 636                 if (WARN_ON(new_zhdr == zhdr))
 637                         goto out_fail;
 638
 639                 new_bud = get_free_buddy(new_zhdr, chunks);
 640                 q = new_zhdr;
 641                 switch (new_bud) {
 642                 case FIRST:
 643                         new_zhdr->first_chunks = chunks;
 644                         q += ZHDR_SIZE_ALIGNED;
 645                         break;
 646                 case MIDDLE:
 647                         new_zhdr->middle_chunks = chunks;
 648                         new_zhdr->start_middle =
 649                                 new_zhdr->first_chunks + ZHDR_CHUNKS;
 650                         q += new_zhdr->start_middle << CHUNK_SHIFT;
 651                         break;
 652                 case LAST:
 653                         new_zhdr->last_chunks = chunks;
 654                         q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
 655                         break;
 656                 default:
 657                         goto out_fail;
 658                 }
 659                 new_zhdr->foreign_handles++;
 660                 memcpy(q, p, sz);
 661                 write_lock(&zhdr->slots->lock);
 662                 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
 663                         __idx(new_zhdr, new_bud);
 664                 if (new_bud == LAST)
 665                         *(unsigned long *)old_handle |=
 666                                         (new_zhdr->last_chunks << BUDDY_SHIFT);
 667                 write_unlock(&zhdr->slots->lock);
 668                 add_to_unbuddied(pool, new_zhdr);
 669                 z3fold_page_unlock(new_zhdr);
 670
 671                 *moved_chunks = 0;
 672         }
 673
 674         return new_zhdr;
 675
 676 out_fail:
 677         if (new_zhdr && !kref_put(&new_zhdr->refcount, release_z3fold_page_locked)) {
 678                 add_to_unbuddied(pool, new_zhdr);
 679                 z3fold_page_unlock(new_zhdr);
 680         }
 681         return NULL;
 682
 683 }
 684
 685 #define BIG_CHUNK_GAP   3
 686 /* Has to be called with lock held */
 687 static int z3fold_compact_page(struct z3fold_header *zhdr)
 688 {
 689         struct page *page = virt_to_page(zhdr);
 690
 691         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 692                 return 0; /* can't move middle chunk, it's used */
 693
 694         if (unlikely(PageIsolated(page)))
 695                 return 0;
 696
 697         if (zhdr->middle_chunks == 0)
 698                 return 0; /* nothing to compact */
 699
 700         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 701                 /* move to the beginning */
 702                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 703                 zhdr->first_chunks = zhdr->middle_chunks;
 704                 zhdr->middle_chunks = 0;
 705                 zhdr->start_middle = 0;
 706                 zhdr->first_num++;
 707                 return 1;
 708         }
 709
 710         /*
 711          * moving data is expensive, so let's only do that if
 712          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 713          */
 714         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 715             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 716                         BIG_CHUNK_GAP) {
 717                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 718                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 719                 return 1;
 720         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 721                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 722                                         + zhdr->middle_chunks) >=
 723                         BIG_CHUNK_GAP) {
 724                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 725                         zhdr->middle_chunks;
 726                 mchunk_memmove(zhdr, new_start);
 727                 zhdr->start_middle = new_start;
 728                 return 1;
 729         }
 730
 731         return 0;
 732 }
 733
 734 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 735 {
 736         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 737         struct page *page;
 738
 739         page = virt_to_page(zhdr);
 740         if (locked)
 741                 WARN_ON(z3fold_page_trylock(zhdr));
 742         else
 743                 z3fold_page_lock(zhdr);
 744         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 745                 z3fold_page_unlock(zhdr);
 746                 return;
 747         }
 748         spin_lock(&pool->lock);
 749         list_del_init(&zhdr->buddy);
 750         spin_unlock(&pool->lock);
 751
 752         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
 753                 return;
 754
 755         if (test_bit(PAGE_STALE, &page->private) ||
 756             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 757                 z3fold_page_unlock(zhdr);
 758                 return;
 759         }
 760
 761         if (!zhdr->foreign_handles && buddy_single(zhdr) &&
 762             zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
 763                 if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 764                         clear_bit(PAGE_CLAIMED, &page->private);
 765                         z3fold_page_unlock(zhdr);
 766                 }
 767                 return;
 768         }
 769
 770         z3fold_compact_page(zhdr);
 771         add_to_unbuddied(pool, zhdr);
 772         clear_bit(PAGE_CLAIMED, &page->private);
 773         z3fold_page_unlock(zhdr);
 774 }
 775
 776 static void compact_page_work(struct work_struct *w)
 777 {
 778         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 779                                                 work);
 780
 781         do_compact_page(zhdr, false);
 782 }
 783
 784 /* returns _locked_ z3fold page header or NULL */
 785 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 786                                                 size_t size, bool can_sleep)
 787 {
 788         struct z3fold_header *zhdr = NULL;
 789         struct page *page;
 790         struct list_head *unbuddied;
 791         int chunks = size_to_chunks(size), i;
 792
 793 lookup:
 794         migrate_disable();
 795         /* First, try to find an unbuddied z3fold page. */
 796         unbuddied = this_cpu_ptr(pool->unbuddied);
 797         for_each_unbuddied_list(i, chunks) {
 798                 struct list_head *l = &unbuddied[i];
 799
 800                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 801                                         struct z3fold_header, buddy);
 802
 803                 if (!zhdr)
 804                         continue;
 805
 806                 /* Re-check under lock. */
 807                 spin_lock(&pool->lock);
 808                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 809                                                 struct z3fold_header, buddy)) ||
 810                     !z3fold_page_trylock(zhdr)) {
 811                         spin_unlock(&pool->lock);
 812                         zhdr = NULL;
 813                         migrate_enable();
 814                         if (can_sleep)
 815                                 cond_resched();
 816                         goto lookup;
 817                 }
 818                 list_del_init(&zhdr->buddy);
 819                 zhdr->cpu = -1;
 820                 spin_unlock(&pool->lock);
 821
 822                 page = virt_to_page(zhdr);
 823                 if (test_bit(NEEDS_COMPACTING, &page->private) ||
 824                     test_bit(PAGE_CLAIMED, &page->private)) {
 825                         z3fold_page_unlock(zhdr);
 826                         zhdr = NULL;
 827                         migrate_enable();
 828                         if (can_sleep)
 829                                 cond_resched();
 830                         goto lookup;
 831                 }
 832
 833                 /*
 834                  * this page could not be removed from its unbuddied
 835                  * list while pool lock was held, and then we've taken
 836                  * page lock so kref_put could not be called before
 837                  * we got here, so it's safe to just call kref_get()
 838                  */
 839                 kref_get(&zhdr->refcount);
 840                 break;
 841         }
 842         migrate_enable();
 843
 844         if (!zhdr) {
 845                 int cpu;
 846
 847                 /* look for _exact_ match on other cpus' lists */
 848                 for_each_online_cpu(cpu) {
 849                         struct list_head *l;
 850
 851                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 852                         spin_lock(&pool->lock);
 853                         l = &unbuddied[chunks];
 854
 855                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 856                                                 struct z3fold_header, buddy);
 857
 858                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 859                                 spin_unlock(&pool->lock);
 860                                 zhdr = NULL;
 861                                 continue;
 862                         }
 863                         list_del_init(&zhdr->buddy);
 864                         zhdr->cpu = -1;
 865                         spin_unlock(&pool->lock);
 866
 867                         page = virt_to_page(zhdr);
 868                         if (test_bit(NEEDS_COMPACTING, &page->private) ||
 869                             test_bit(PAGE_CLAIMED, &page->private)) {
 870                                 z3fold_page_unlock(zhdr);
 871                                 zhdr = NULL;
 872                                 if (can_sleep)
 873                                         cond_resched();
 874                                 continue;
 875                         }
 876                         kref_get(&zhdr->refcount);
 877                         break;
 878                 }
 879         }
 880
 881         if (zhdr && !zhdr->slots) {
 882                 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
 883                 if (!zhdr->slots)
 884                         goto out_fail;
 885         }
 886         return zhdr;
 887
 888 out_fail:
 889         if (!kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 890                 add_to_unbuddied(pool, zhdr);
 891                 z3fold_page_unlock(zhdr);
 892         }
 893         return NULL;
 894 }
 895
 896 /*
 897  * API Functions
 898  */
 899
 900 /**
 901  * z3fold_create_pool() - create a new z3fold pool
 902  * @name:       pool name
 903  * @gfp:        gfp flags when allocating the z3fold pool structure
 904  *
 905  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 906  * failed.
 907  */
 908 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp)
 909 {
 910         struct z3fold_pool *pool = NULL;
 911         int i, cpu;
 912
 913         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 914         if (!pool)
 915                 goto out;
 916         pool->c_handle = kmem_cache_create("z3fold_handle",
 917                                 sizeof(struct z3fold_buddy_slots),
 918                                 SLOTS_ALIGN, 0, NULL);
 919         if (!pool->c_handle)
 920                 goto out_c;
 921         spin_lock_init(&pool->lock);
 922         spin_lock_init(&pool->stale_lock);
 923         pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
 924                                          __alignof__(struct list_head));
 925         if (!pool->unbuddied)
 926                 goto out_pool;
 927         for_each_possible_cpu(cpu) {
 928                 struct list_head *unbuddied =
 929                                 per_cpu_ptr(pool->unbuddied, cpu);
 930                 for_each_unbuddied_list(i, 0)
 931                         INIT_LIST_HEAD(&unbuddied[i]);
 932         }
 933         INIT_LIST_HEAD(&pool->lru);
 934         INIT_LIST_HEAD(&pool->stale);
 935         atomic64_set(&pool->pages_nr, 0);
 936         pool->name = name;
 937         pool->compact_wq = create_singlethread_workqueue(pool->name);
 938         if (!pool->compact_wq)
 939                 goto out_unbuddied;
 940         pool->release_wq = create_singlethread_workqueue(pool->name);
 941         if (!pool->release_wq)
 942                 goto out_wq;
 943         INIT_WORK(&pool->work, free_pages_work);
 944         return pool;
 945
 946 out_wq:
 947         destroy_workqueue(pool->compact_wq);
 948 out_unbuddied:
 949         free_percpu(pool->unbuddied);
 950 out_pool:
 951         kmem_cache_destroy(pool->c_handle);
 952 out_c:
 953         kfree(pool);
 954 out:
 955         return NULL;
 956 }
 957
 958 /**
 959  * z3fold_destroy_pool() - destroys an existing z3fold pool
 960  * @pool:       the z3fold pool to be destroyed
 961  *
 962  * The pool should be emptied before this function is called.
 963  */
 964 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 965 {
 966         kmem_cache_destroy(pool->c_handle);
 967
 968         /*
 969          * We need to destroy pool->compact_wq before pool->release_wq,
 970          * as any pending work on pool->compact_wq will call
 971          * queue_work(pool->release_wq, &pool->work).
 972          *
 973          * There are still outstanding pages until both workqueues are drained,
 974          * so we cannot unregister migration until then.
 975          */
 976
 977         destroy_workqueue(pool->compact_wq);
 978         destroy_workqueue(pool->release_wq);
 979         free_percpu(pool->unbuddied);
 980         kfree(pool);
 981 }
 982
 983 static const struct movable_operations z3fold_mops;
 984
 985 /**
 986  * z3fold_alloc() - allocates a region of a given size
 987  * @pool:       z3fold pool from which to allocate
 988  * @size:       size in bytes of the desired allocation
 989  * @gfp:        gfp flags used if the pool needs to grow
 990  * @handle:     handle of the new allocation
 991  *
 992  * This function will attempt to find a free region in the pool large enough to
 993  * satisfy the allocation request.  A search of the unbuddied lists is
 994  * performed first. If no suitable free region is found, then a new page is
 995  * allocated and added to the pool to satisfy the request.
 996  *
 997  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 998  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 999  * a new page.
1000  */
1001 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1002                         unsigned long *handle)
1003 {
1004         int chunks = size_to_chunks(size);
1005         struct z3fold_header *zhdr = NULL;
1006         struct page *page = NULL;
1007         enum buddy bud;
1008         bool can_sleep = gfpflags_allow_blocking(gfp);
1009
1010         if (!size || (gfp & __GFP_HIGHMEM))
1011                 return -EINVAL;
1012
1013         if (size > PAGE_SIZE)
1014                 return -ENOSPC;
1015
1016         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1017                 bud = HEADLESS;
1018         else {
1019 retry:
1020                 zhdr = __z3fold_alloc(pool, size, can_sleep);
1021                 if (zhdr) {
1022                         bud = get_free_buddy(zhdr, chunks);
1023                         if (bud == HEADLESS) {
1024                                 if (!kref_put(&zhdr->refcount,
1025                                              release_z3fold_page_locked))
1026                                         z3fold_page_unlock(zhdr);
1027                                 pr_err("No free chunks in unbuddied\n");
1028                                 WARN_ON(1);
1029                                 goto retry;
1030                         }
1031                         page = virt_to_page(zhdr);
1032                         goto found;
1033                 }
1034                 bud = FIRST;
1035         }
1036
1037         page = alloc_page(gfp);
1038         if (!page)
1039                 return -ENOMEM;
1040
1041         zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1042         if (!zhdr) {
1043                 __free_page(page);
1044                 return -ENOMEM;
1045         }
1046         atomic64_inc(&pool->pages_nr);
1047
1048         if (bud == HEADLESS) {
1049                 set_bit(PAGE_HEADLESS, &page->private);
1050                 goto headless;
1051         }
1052         if (can_sleep) {
1053                 lock_page(page);
1054                 __SetPageMovable(page, &z3fold_mops);
1055                 unlock_page(page);
1056         } else {
1057                 WARN_ON(!trylock_page(page));
1058                 __SetPageMovable(page, &z3fold_mops);
1059                 unlock_page(page);
1060         }
1061         z3fold_page_lock(zhdr);
1062
1063 found:
1064         if (bud == FIRST)
1065                 zhdr->first_chunks = chunks;
1066         else if (bud == LAST)
1067                 zhdr->last_chunks = chunks;
1068         else {
1069                 zhdr->middle_chunks = chunks;
1070                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1071         }
1072         add_to_unbuddied(pool, zhdr);
1073
1074 headless:
1075         spin_lock(&pool->lock);
1076         /* Add/move z3fold page to beginning of LRU */
1077         if (!list_empty(&page->lru))
1078                 list_del(&page->lru);
1079
1080         list_add(&page->lru, &pool->lru);
1081
1082         *handle = encode_handle(zhdr, bud);
1083         spin_unlock(&pool->lock);
1084         if (bud != HEADLESS)
1085                 z3fold_page_unlock(zhdr);
1086
1087         return 0;
1088 }
1089
1090 /**
1091  * z3fold_free() - frees the allocation associated with the given handle
1092  * @pool:       pool in which the allocation resided
1093  * @handle:     handle associated with the allocation returned by z3fold_alloc()
1094  *
1095  * In the case that the z3fold page in which the allocation resides is under
1096  * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1097  * only sets the first|middle|last_chunks to 0.  The page is actually freed
1098  * once all buddies are evicted (see z3fold_reclaim_page() below).
1099  */
1100 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1101 {
1102         struct z3fold_header *zhdr;
1103         struct page *page;
1104         enum buddy bud;
1105         bool page_claimed;
1106
1107         zhdr = get_z3fold_header(handle);
1108         page = virt_to_page(zhdr);
1109         page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1110
1111         if (test_bit(PAGE_HEADLESS, &page->private)) {
1112                 /* if a headless page is under reclaim, just leave.
1113                  * NB: we use test_and_set_bit for a reason: if the bit
1114                  * has not been set before, we release this page
1115                  * immediately so we don't care about its value any more.
1116                  */
1117                 if (!page_claimed) {
1118                         spin_lock(&pool->lock);
1119                         list_del(&page->lru);
1120                         spin_unlock(&pool->lock);
1121                         put_z3fold_header(zhdr);
1122                         free_z3fold_page(page, true);
1123                         atomic64_dec(&pool->pages_nr);
1124                 }
1125                 return;
1126         }
1127
1128         /* Non-headless case */
1129         bud = handle_to_buddy(handle);
1130
1131         switch (bud) {
1132         case FIRST:
1133                 zhdr->first_chunks = 0;
1134                 break;
1135         case MIDDLE:
1136                 zhdr->middle_chunks = 0;
1137                 break;
1138         case LAST:
1139                 zhdr->last_chunks = 0;
1140                 break;
1141         default:
1142                 pr_err("%s: unknown bud %d\n", __func__, bud);
1143                 WARN_ON(1);
1144                 put_z3fold_header(zhdr);
1145                 return;
1146         }
1147
1148         if (!page_claimed)
1149                 free_handle(handle, zhdr);
1150         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list))
1151                 return;
1152         if (page_claimed) {
1153                 /* the page has not been claimed by us */
1154                 put_z3fold_header(zhdr);
1155                 return;
1156         }
1157         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1158                 clear_bit(PAGE_CLAIMED, &page->private);
1159                 put_z3fold_header(zhdr);
1160                 return;
1161         }
1162         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1163                 zhdr->cpu = -1;
1164                 kref_get(&zhdr->refcount);
1165                 clear_bit(PAGE_CLAIMED, &page->private);
1166                 do_compact_page(zhdr, true);
1167                 return;
1168         }
1169         kref_get(&zhdr->refcount);
1170         clear_bit(PAGE_CLAIMED, &page->private);
1171         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1172         put_z3fold_header(zhdr);
1173 }
1174
1175 /**
1176  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1177  * @pool:       pool from which a page will attempt to be evicted
1178  * @retries:    number of pages on the LRU list for which eviction will
1179  *              be attempted before failing
1180  *
1181  * z3fold reclaim is different from normal system reclaim in that it is done
1182  * from the bottom, up. This is because only the bottom layer, z3fold, has
1183  * information on how the allocations are organized within each z3fold page.
1184  * This has the potential to create interesting locking situations between
1185  * z3fold and the user, however.
1186  *
1187  * To avoid these, this is how z3fold_reclaim_page() should be called:
1188  *
1189  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1190  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1191  * call the user-defined eviction handler with the pool and handle as
1192  * arguments.
1193  *
1194  * If the handle can not be evicted, the eviction handler should return
1195  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1196  * appropriate list and try the next z3fold page on the LRU up to
1197  * a user defined number of retries.
1198  *
1199  * If the handle is successfully evicted, the eviction handler should
1200  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1201  * contains logic to delay freeing the page if the page is under reclaim,
1202  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1203  *
1204  * If all buddies in the z3fold page are successfully evicted, then the
1205  * z3fold page can be freed.
1206  *
1207  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1208  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1209  * the retry limit was hit.
1210  */
1211 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1212 {
1213         int i, ret = -1;
1214         struct z3fold_header *zhdr = NULL;
1215         struct page *page = NULL;
1216         struct list_head *pos;
1217         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1218         struct z3fold_buddy_slots slots __attribute__((aligned(SLOTS_ALIGN)));
1219
1220         rwlock_init(&slots.lock);
1221         slots.pool = (unsigned long)pool | (1 << HANDLES_NOFREE);
1222
1223         spin_lock(&pool->lock);
1224         for (i = 0; i < retries; i++) {
1225                 if (list_empty(&pool->lru)) {
1226                         spin_unlock(&pool->lock);
1227                         return -EINVAL;
1228                 }
1229                 list_for_each_prev(pos, &pool->lru) {
1230                         page = list_entry(pos, struct page, lru);
1231
1232                         zhdr = page_address(page);
1233                         if (test_bit(PAGE_HEADLESS, &page->private)) {
1234                                 /*
1235                                  * For non-headless pages, we wait to do this
1236                                  * until we have the page lock to avoid racing
1237                                  * with __z3fold_alloc(). Headless pages don't
1238                                  * have a lock (and __z3fold_alloc() will never
1239                                  * see them), but we still need to test and set
1240                                  * PAGE_CLAIMED to avoid racing with
1241                                  * z3fold_free(), so just do it now before
1242                                  * leaving the loop.
1243                                  */
1244                                 if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1245                                         continue;
1246
1247                                 break;
1248                         }
1249
1250                         if (!z3fold_page_trylock(zhdr)) {
1251                                 zhdr = NULL;
1252                                 continue; /* can't evict at this point */
1253                         }
1254
1255                         /* test_and_set_bit is of course atomic, but we still
1256                          * need to do it under page lock, otherwise checking
1257                          * that bit in __z3fold_alloc wouldn't make sense
1258                          */
1259                         if (zhdr->foreign_handles ||
1260                             test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1261                                 z3fold_page_unlock(zhdr);
1262                                 zhdr = NULL;
1263                                 continue; /* can't evict such page */
1264                         }
1265                         list_del_init(&zhdr->buddy);
1266                         zhdr->cpu = -1;
1267                         /* See comment in __z3fold_alloc. */
1268                         kref_get(&zhdr->refcount);
1269                         break;
1270                 }
1271
1272                 if (!zhdr)
1273                         break;
1274
1275                 list_del_init(&page->lru);
1276                 spin_unlock(&pool->lock);
1277
1278                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1279                         /*
1280                          * We need encode the handles before unlocking, and
1281                          * use our local slots structure because z3fold_free
1282                          * can zero out zhdr->slots and we can't do much
1283                          * about that
1284                          */
1285                         first_handle = 0;
1286                         last_handle = 0;
1287                         middle_handle = 0;
1288                         memset(slots.slot, 0, sizeof(slots.slot));
1289                         if (zhdr->first_chunks)
1290                                 first_handle = __encode_handle(zhdr, &slots,
1291                                                                 FIRST);
1292                         if (zhdr->middle_chunks)
1293                                 middle_handle = __encode_handle(zhdr, &slots,
1294                                                                 MIDDLE);
1295                         if (zhdr->last_chunks)
1296                                 last_handle = __encode_handle(zhdr, &slots,
1297                                                                 LAST);
1298                         /*
1299                          * it's safe to unlock here because we hold a
1300                          * reference to this page
1301                          */
1302                         z3fold_page_unlock(zhdr);
1303                 } else {
1304                         first_handle = encode_handle(zhdr, HEADLESS);
1305                         last_handle = middle_handle = 0;
1306                 }
1307                 /* Issue the eviction callback(s) */
1308                 if (middle_handle) {
1309                         ret = pool->zpool_ops->evict(pool->zpool, middle_handle);
1310                         if (ret)
1311                                 goto next;
1312                 }
1313                 if (first_handle) {
1314                         ret = pool->zpool_ops->evict(pool->zpool, first_handle);
1315                         if (ret)
1316                                 goto next;
1317                 }
1318                 if (last_handle) {
1319                         ret = pool->zpool_ops->evict(pool->zpool, last_handle);
1320                         if (ret)
1321                                 goto next;
1322                 }
1323 next:
1324                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1325                         if (ret == 0) {
1326                                 free_z3fold_page(page, true);
1327                                 atomic64_dec(&pool->pages_nr);
1328                                 return 0;
1329                         }
1330                         spin_lock(&pool->lock);
1331                         list_add(&page->lru, &pool->lru);
1332                         spin_unlock(&pool->lock);
1333                         clear_bit(PAGE_CLAIMED, &page->private);
1334                 } else {
1335                         struct z3fold_buddy_slots *slots = zhdr->slots;
1336                         z3fold_page_lock(zhdr);
1337                         if (kref_put(&zhdr->refcount,
1338                                         release_z3fold_page_locked)) {
1339                                 kmem_cache_free(pool->c_handle, slots);
1340                                 return 0;
1341                         }
1342                         /*
1343                          * if we are here, the page is still not completely
1344                          * free. Take the global pool lock then to be able
1345                          * to add it back to the lru list
1346                          */
1347                         spin_lock(&pool->lock);
1348                         list_add(&page->lru, &pool->lru);
1349                         spin_unlock(&pool->lock);
1350                         if (list_empty(&zhdr->buddy))
1351                                 add_to_unbuddied(pool, zhdr);
1352                         clear_bit(PAGE_CLAIMED, &page->private);
1353                         z3fold_page_unlock(zhdr);
1354                 }
1355
1356                 /* We started off locked to we need to lock the pool back */
1357                 spin_lock(&pool->lock);
1358         }
1359         spin_unlock(&pool->lock);
1360         return -EAGAIN;
1361 }
1362
1363 /**
1364  * z3fold_map() - maps the allocation associated with the given handle
1365  * @pool:       pool in which the allocation resides
1366  * @handle:     handle associated with the allocation to be mapped
1367  *
1368  * Extracts the buddy number from handle and constructs the pointer to the
1369  * correct starting chunk within the page.
1370  *
1371  * Returns: a pointer to the mapped allocation
1372  */
1373 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1374 {
1375         struct z3fold_header *zhdr;
1376         struct page *page;
1377         void *addr;
1378         enum buddy buddy;
1379
1380         zhdr = get_z3fold_header(handle);
1381         addr = zhdr;
1382         page = virt_to_page(zhdr);
1383
1384         if (test_bit(PAGE_HEADLESS, &page->private))
1385                 goto out;
1386
1387         buddy = handle_to_buddy(handle);
1388         switch (buddy) {
1389         case FIRST:
1390                 addr += ZHDR_SIZE_ALIGNED;
1391                 break;
1392         case MIDDLE:
1393                 addr += zhdr->start_middle << CHUNK_SHIFT;
1394                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1395                 break;
1396         case LAST:
1397                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1398                 break;
1399         default:
1400                 pr_err("unknown buddy id %d\n", buddy);
1401                 WARN_ON(1);
1402                 addr = NULL;
1403                 break;
1404         }
1405
1406         if (addr)
1407                 zhdr->mapped_count++;
1408 out:
1409         put_z3fold_header(zhdr);
1410         return addr;
1411 }
1412
1413 /**
1414  * z3fold_unmap() - unmaps the allocation associated with the given handle
1415  * @pool:       pool in which the allocation resides
1416  * @handle:     handle associated with the allocation to be unmapped
1417  */
1418 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1419 {
1420         struct z3fold_header *zhdr;
1421         struct page *page;
1422         enum buddy buddy;
1423
1424         zhdr = get_z3fold_header(handle);
1425         page = virt_to_page(zhdr);
1426
1427         if (test_bit(PAGE_HEADLESS, &page->private))
1428                 return;
1429
1430         buddy = handle_to_buddy(handle);
1431         if (buddy == MIDDLE)
1432                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1433         zhdr->mapped_count--;
1434         put_z3fold_header(zhdr);
1435 }
1436
1437 /**
1438  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1439  * @pool:       pool whose size is being queried
1440  *
1441  * Returns: size in pages of the given pool.
1442  */
1443 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1444 {
1445         return atomic64_read(&pool->pages_nr);
1446 }
1447
1448 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1449 {
1450         struct z3fold_header *zhdr;
1451         struct z3fold_pool *pool;
1452
1453         VM_BUG_ON_PAGE(!PageMovable(page), page);
1454         VM_BUG_ON_PAGE(PageIsolated(page), page);
1455
1456         if (test_bit(PAGE_HEADLESS, &page->private))
1457                 return false;
1458
1459         zhdr = page_address(page);
1460         z3fold_page_lock(zhdr);
1461         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1462             test_bit(PAGE_STALE, &page->private))
1463                 goto out;
1464
1465         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1466                 goto out;
1467
1468         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1469                 goto out;
1470         pool = zhdr_to_pool(zhdr);
1471         spin_lock(&pool->lock);
1472         if (!list_empty(&zhdr->buddy))
1473                 list_del_init(&zhdr->buddy);
1474         if (!list_empty(&page->lru))
1475                 list_del_init(&page->lru);
1476         spin_unlock(&pool->lock);
1477
1478         kref_get(&zhdr->refcount);
1479         z3fold_page_unlock(zhdr);
1480         return true;
1481
1482 out:
1483         z3fold_page_unlock(zhdr);
1484         return false;
1485 }
1486
1487 static int z3fold_page_migrate(struct page *newpage, struct page *page,
1488                 enum migrate_mode mode)
1489 {
1490         struct z3fold_header *zhdr, *new_zhdr;
1491         struct z3fold_pool *pool;
1492
1493         VM_BUG_ON_PAGE(!PageMovable(page), page);
1494         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1495         VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1496         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1497
1498         zhdr = page_address(page);
1499         pool = zhdr_to_pool(zhdr);
1500
1501         if (!z3fold_page_trylock(zhdr))
1502                 return -EAGAIN;
1503         if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1504                 clear_bit(PAGE_CLAIMED, &page->private);
1505                 z3fold_page_unlock(zhdr);
1506                 return -EBUSY;
1507         }
1508         if (work_pending(&zhdr->work)) {
1509                 z3fold_page_unlock(zhdr);
1510                 return -EAGAIN;
1511         }
1512         new_zhdr = page_address(newpage);
1513         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1514         newpage->private = page->private;
1515         set_bit(PAGE_MIGRATED, &page->private);
1516         z3fold_page_unlock(zhdr);
1517         spin_lock_init(&new_zhdr->page_lock);
1518         INIT_WORK(&new_zhdr->work, compact_page_work);
1519         /*
1520          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1521          * so we only have to reinitialize it.
1522          */
1523         INIT_LIST_HEAD(&new_zhdr->buddy);
1524         __ClearPageMovable(page);
1525
1526         get_page(newpage);
1527         z3fold_page_lock(new_zhdr);
1528         if (new_zhdr->first_chunks)
1529                 encode_handle(new_zhdr, FIRST);
1530         if (new_zhdr->last_chunks)
1531                 encode_handle(new_zhdr, LAST);
1532         if (new_zhdr->middle_chunks)
1533                 encode_handle(new_zhdr, MIDDLE);
1534         set_bit(NEEDS_COMPACTING, &newpage->private);
1535         new_zhdr->cpu = smp_processor_id();
1536         spin_lock(&pool->lock);
1537         list_add(&newpage->lru, &pool->lru);
1538         spin_unlock(&pool->lock);
1539         __SetPageMovable(newpage, &z3fold_mops);
1540         z3fold_page_unlock(new_zhdr);
1541
1542         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1543
1544         /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1545         page->private = 0;
1546         put_page(page);
1547         return 0;
1548 }
1549
1550 static void z3fold_page_putback(struct page *page)
1551 {
1552         struct z3fold_header *zhdr;
1553         struct z3fold_pool *pool;
1554
1555         zhdr = page_address(page);
1556         pool = zhdr_to_pool(zhdr);
1557
1558         z3fold_page_lock(zhdr);
1559         if (!list_empty(&zhdr->buddy))
1560                 list_del_init(&zhdr->buddy);
1561         INIT_LIST_HEAD(&page->lru);
1562         if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
1563                 return;
1564         spin_lock(&pool->lock);
1565         list_add(&page->lru, &pool->lru);
1566         spin_unlock(&pool->lock);
1567         if (list_empty(&zhdr->buddy))
1568                 add_to_unbuddied(pool, zhdr);
1569         clear_bit(PAGE_CLAIMED, &page->private);
1570         z3fold_page_unlock(zhdr);
1571 }
1572
1573 static const struct movable_operations z3fold_mops = {
1574         .isolate_page = z3fold_page_isolate,
1575         .migrate_page = z3fold_page_migrate,
1576         .putback_page = z3fold_page_putback,
1577 };
1578
1579 /*****************
1580  * zpool
1581  ****************/
1582
1583 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1584                                const struct zpool_ops *zpool_ops,
1585                                struct zpool *zpool)
1586 {
1587         struct z3fold_pool *pool;
1588
1589         pool = z3fold_create_pool(name, gfp);
1590         if (pool) {
1591                 pool->zpool = zpool;
1592                 pool->zpool_ops = zpool_ops;
1593         }
1594         return pool;
1595 }
1596
1597 static void z3fold_zpool_destroy(void *pool)
1598 {
1599         z3fold_destroy_pool(pool);
1600 }
1601
1602 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1603                         unsigned long *handle)
1604 {
1605         return z3fold_alloc(pool, size, gfp, handle);
1606 }
1607 static void z3fold_zpool_free(void *pool, unsigned long handle)
1608 {
1609         z3fold_free(pool, handle);
1610 }
1611
1612 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1613                         unsigned int *reclaimed)
1614 {
1615         unsigned int total = 0;
1616         int ret = -EINVAL;
1617
1618         while (total < pages) {
1619                 ret = z3fold_reclaim_page(pool, 8);
1620                 if (ret < 0)
1621                         break;
1622                 total++;
1623         }
1624
1625         if (reclaimed)
1626                 *reclaimed = total;
1627
1628         return ret;
1629 }
1630
1631 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1632                         enum zpool_mapmode mm)
1633 {
1634         return z3fold_map(pool, handle);
1635 }
1636 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1637 {
1638         z3fold_unmap(pool, handle);
1639 }
1640
1641 static u64 z3fold_zpool_total_size(void *pool)
1642 {
1643         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1644 }
1645
1646 static struct zpool_driver z3fold_zpool_driver = {
1647         .type =         "z3fold",
1648         .sleep_mapped = true,
1649         .owner =        THIS_MODULE,
1650         .create =       z3fold_zpool_create,
1651         .destroy =      z3fold_zpool_destroy,
1652         .malloc =       z3fold_zpool_malloc,
1653         .free =         z3fold_zpool_free,
1654         .shrink =       z3fold_zpool_shrink,
1655         .map =          z3fold_zpool_map,
1656         .unmap =        z3fold_zpool_unmap,
1657         .total_size =   z3fold_zpool_total_size,
1658 };
1659
1660 MODULE_ALIAS("zpool-z3fold");
1661
1662 static int __init init_z3fold(void)
1663 {
1664         /*
1665          * Make sure the z3fold header is not larger than the page size and
1666          * there has remaining spaces for its buddy.
1667          */
1668         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1669         zpool_register_driver(&z3fold_zpool_driver);
1670
1671         return 0;
1672 }
1673
1674 static void __exit exit_z3fold(void)
1675 {
1676         zpool_unregister_driver(&z3fold_zpool_driver);
1677 }
1678
1679 module_init(init_z3fold);
1680 module_exit(exit_z3fold);
1681
1682 MODULE_LICENSE("GPL");
1683 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1684 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");