mm/z3fold.c

   1 /*
   2  * z3fold.c
   3  *
   4  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
   5  * Copyright (C) 2016, Sony Mobile Communications Inc.
   6  *
   7  * This implementation is based on zbud written by Seth Jennings.
   8  *
   9  * z3fold is an special purpose allocator for storing compressed pages. It
  10  * can store up to three compressed pages per page which improves the
  11  * compression ratio of zbud while retaining its main concepts (e. g. always
  12  * storing an integral number of objects per page) and simplicity.
  13  * It still has simple and deterministic reclaim properties that make it
  14  * preferable to a higher density approach (with no requirement on integral
  15  * number of object per page) when reclaim is used.
  16  *
  17  * As in zbud, pages are divided into "chunks".  The size of the chunks is
  18  * fixed at compile time and is determined by NCHUNKS_ORDER below.
  19  *
  20  * z3fold doesn't export any API and is meant to be used via zpool API.
  21  */
  22
  23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  24
  25 #include <linux/atomic.h>
  26 #include <linux/sched.h>
  27 #include <linux/list.h>
  28 #include <linux/mm.h>
  29 #include <linux/module.h>
  30 #include <linux/percpu.h>
  31 #include <linux/preempt.h>
  32 #include <linux/workqueue.h>
  33 #include <linux/slab.h>
  34 #include <linux/spinlock.h>
  35 #include <linux/zpool.h>
  36
  37 /*****************
  38  * Structures
  39 *****************/
  40 struct z3fold_pool;
  41 struct z3fold_ops {
  42         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
  43 };
  44
  45 enum buddy {
  46         HEADLESS = 0,
  47         FIRST,
  48         MIDDLE,
  49         LAST,
  50         BUDDIES_MAX
  51 };
  52
  53 /*
  54  * struct z3fold_header - z3fold page metadata occupying first chunks of each
  55  *                      z3fold page, except for HEADLESS pages
  56  * @buddy:              links the z3fold page into the relevant list in the
  57  *                      pool
  58  * @page_lock:          per-page lock
  59  * @refcount:           reference count for the z3fold page
  60  * @work:               work_struct for page layout optimization
  61  * @pool:               pointer to the pool which this page belongs to
  62  * @cpu:                CPU which this page "belongs" to
  63  * @first_chunks:       the size of the first buddy in chunks, 0 if free
  64  * @middle_chunks:      the size of the middle buddy in chunks, 0 if free
  65  * @last_chunks:        the size of the last buddy in chunks, 0 if free
  66  * @first_num:          the starting number (for the first handle)
  67  */
  68 struct z3fold_header {
  69         struct list_head buddy;
  70         spinlock_t page_lock;
  71         struct kref refcount;
  72         struct work_struct work;
  73         struct z3fold_pool *pool;
  74         short cpu;
  75         unsigned short first_chunks;
  76         unsigned short middle_chunks;
  77         unsigned short last_chunks;
  78         unsigned short start_middle;
  79         unsigned short first_num:2;
  80 };
  81
  82 /*
  83  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
  84  * adjusting internal fragmentation.  It also determines the number of
  85  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
  86  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
  87  * in the beginning of an allocated page are occupied by z3fold header, so
  88  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
  89  * which shows the max number of free chunks in z3fold page, also there will
  90  * be 63, or 62, respectively, freelists per pool.
  91  */
  92 #define NCHUNKS_ORDER   6
  93
  94 #define CHUNK_SHIFT     (PAGE_SHIFT - NCHUNKS_ORDER)
  95 #define CHUNK_SIZE      (1 << CHUNK_SHIFT)
  96 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
  97 #define ZHDR_CHUNKS     (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
  98 #define TOTAL_CHUNKS    (PAGE_SIZE >> CHUNK_SHIFT)
  99 #define NCHUNKS         ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
 100
 101 #define BUDDY_MASK      (0x3)
 102
 103 /**
 104  * struct z3fold_pool - stores metadata for each z3fold pool
 105  * @name:       pool name
 106  * @lock:       protects pool unbuddied/lru lists
 107  * @stale_lock: protects pool stale page list
 108  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 109  *              buddies; the list each z3fold page is added to depends on
 110  *              the size of its free region.
 111  * @lru:        list tracking the z3fold pages in LRU order by most recently
 112  *              added buddy.
 113  * @stale:      list of pages marked for freeing
 114  * @pages_nr:   number of z3fold pages in the pool.
 115  * @ops:        pointer to a structure of user defined operations specified at
 116  *              pool creation time.
 117  * @compact_wq: workqueue for page layout background optimization
 118  * @release_wq: workqueue for safe page release
 119  * @work:       work_struct for safe page release
 120  *
 121  * This structure is allocated at pool creation time and maintains metadata
 122  * pertaining to a particular z3fold pool.
 123  */
 124 struct z3fold_pool {
 125         const char *name;
 126         spinlock_t lock;
 127         spinlock_t stale_lock;
 128         struct list_head *unbuddied;
 129         struct list_head lru;
 130         struct list_head stale;
 131         atomic64_t pages_nr;
 132         const struct z3fold_ops *ops;
 133         struct zpool *zpool;
 134         const struct zpool_ops *zpool_ops;
 135         struct workqueue_struct *compact_wq;
 136         struct workqueue_struct *release_wq;
 137         struct work_struct work;
 138 };
 139
 140 /*
 141  * Internal z3fold page flags
 142  */
 143 enum z3fold_page_flags {
 144         PAGE_HEADLESS = 0,
 145         MIDDLE_CHUNK_MAPPED,
 146         NEEDS_COMPACTING,
 147         PAGE_STALE
 148 };
 149
 150 /*****************
 151  * Helpers
 152 *****************/
 153
 154 /* Converts an allocation size in bytes to size in z3fold chunks */
 155 static int size_to_chunks(size_t size)
 156 {
 157         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 158 }
 159
 160 #define for_each_unbuddied_list(_iter, _begin) \
 161         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 162
 163 static void compact_page_work(struct work_struct *w);
 164
 165 /* Initializes the z3fold header of a newly allocated z3fold page */
 166 static struct z3fold_header *init_z3fold_page(struct page *page,
 167                                         struct z3fold_pool *pool)
 168 {
 169         struct z3fold_header *zhdr = page_address(page);
 170
 171         INIT_LIST_HEAD(&page->lru);
 172         clear_bit(PAGE_HEADLESS, &page->private);
 173         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 174         clear_bit(NEEDS_COMPACTING, &page->private);
 175         clear_bit(PAGE_STALE, &page->private);
 176
 177         spin_lock_init(&zhdr->page_lock);
 178         kref_init(&zhdr->refcount);
 179         zhdr->first_chunks = 0;
 180         zhdr->middle_chunks = 0;
 181         zhdr->last_chunks = 0;
 182         zhdr->first_num = 0;
 183         zhdr->start_middle = 0;
 184         zhdr->cpu = -1;
 185         zhdr->pool = pool;
 186         INIT_LIST_HEAD(&zhdr->buddy);
 187         INIT_WORK(&zhdr->work, compact_page_work);
 188         return zhdr;
 189 }
 190
 191 /* Resets the struct page fields and frees the page */
 192 static void free_z3fold_page(struct page *page)
 193 {
 194         __free_page(page);
 195 }
 196
 197 /* Lock a z3fold page */
 198 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 199 {
 200         spin_lock(&zhdr->page_lock);
 201 }
 202
 203 /* Try to lock a z3fold page */
 204 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 205 {
 206         return spin_trylock(&zhdr->page_lock);
 207 }
 208
 209 /* Unlock a z3fold page */
 210 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 211 {
 212         spin_unlock(&zhdr->page_lock);
 213 }
 214
 215 /*
 216  * Encodes the handle of a particular buddy within a z3fold page
 217  * Pool lock should be held as this function accesses first_num
 218  */
 219 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 220 {
 221         unsigned long handle;
 222
 223         handle = (unsigned long)zhdr;
 224         if (bud != HEADLESS)
 225                 handle += (bud + zhdr->first_num) & BUDDY_MASK;
 226         return handle;
 227 }
 228
 229 /* Returns the z3fold page where a given handle is stored */
 230 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
 231 {
 232         return (struct z3fold_header *)(handle & PAGE_MASK);
 233 }
 234
 235 /*
 236  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 237  *  but that doesn't matter. because the masking will result in the
 238  *  correct buddy number.
 239  */
 240 static enum buddy handle_to_buddy(unsigned long handle)
 241 {
 242         struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
 243         return (handle - zhdr->first_num) & BUDDY_MASK;
 244 }
 245
 246 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 247 {
 248         struct page *page = virt_to_page(zhdr);
 249         struct z3fold_pool *pool = zhdr->pool;
 250
 251         WARN_ON(!list_empty(&zhdr->buddy));
 252         set_bit(PAGE_STALE, &page->private);
 253         clear_bit(NEEDS_COMPACTING, &page->private);
 254         spin_lock(&pool->lock);
 255         if (!list_empty(&page->lru))
 256                 list_del(&page->lru);
 257         spin_unlock(&pool->lock);
 258         if (locked)
 259                 z3fold_page_unlock(zhdr);
 260         spin_lock(&pool->stale_lock);
 261         list_add(&zhdr->buddy, &pool->stale);
 262         queue_work(pool->release_wq, &pool->work);
 263         spin_unlock(&pool->stale_lock);
 264 }
 265
 266 static void __attribute__((__unused__))
 267                         release_z3fold_page(struct kref *ref)
 268 {
 269         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 270                                                 refcount);
 271         __release_z3fold_page(zhdr, false);
 272 }
 273
 274 static void release_z3fold_page_locked(struct kref *ref)
 275 {
 276         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 277                                                 refcount);
 278         WARN_ON(z3fold_page_trylock(zhdr));
 279         __release_z3fold_page(zhdr, true);
 280 }
 281
 282 static void release_z3fold_page_locked_list(struct kref *ref)
 283 {
 284         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 285                                                refcount);
 286         spin_lock(&zhdr->pool->lock);
 287         list_del_init(&zhdr->buddy);
 288         spin_unlock(&zhdr->pool->lock);
 289
 290         WARN_ON(z3fold_page_trylock(zhdr));
 291         __release_z3fold_page(zhdr, true);
 292 }
 293
 294 static void free_pages_work(struct work_struct *w)
 295 {
 296         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 297
 298         spin_lock(&pool->stale_lock);
 299         while (!list_empty(&pool->stale)) {
 300                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 301                                                 struct z3fold_header, buddy);
 302                 struct page *page = virt_to_page(zhdr);
 303
 304                 list_del(&zhdr->buddy);
 305                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 306                         continue;
 307                 spin_unlock(&pool->stale_lock);
 308                 cancel_work_sync(&zhdr->work);
 309                 free_z3fold_page(page);
 310                 cond_resched();
 311                 spin_lock(&pool->stale_lock);
 312         }
 313         spin_unlock(&pool->stale_lock);
 314 }
 315
 316 /*
 317  * Returns the number of free chunks in a z3fold page.
 318  * NB: can't be used with HEADLESS pages.
 319  */
 320 static int num_free_chunks(struct z3fold_header *zhdr)
 321 {
 322         int nfree;
 323         /*
 324          * If there is a middle object, pick up the bigger free space
 325          * either before or after it. Otherwise just subtract the number
 326          * of chunks occupied by the first and the last objects.
 327          */
 328         if (zhdr->middle_chunks != 0) {
 329                 int nfree_before = zhdr->first_chunks ?
 330                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 331                 int nfree_after = zhdr->last_chunks ?
 332                         0 : TOTAL_CHUNKS -
 333                                 (zhdr->start_middle + zhdr->middle_chunks);
 334                 nfree = max(nfree_before, nfree_after);
 335         } else
 336                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 337         return nfree;
 338 }
 339
 340 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 341                                 unsigned short dst_chunk)
 342 {
 343         void *beg = zhdr;
 344         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 345                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 346                        zhdr->middle_chunks << CHUNK_SHIFT);
 347 }
 348
 349 #define BIG_CHUNK_GAP   3
 350 /* Has to be called with lock held */
 351 static int z3fold_compact_page(struct z3fold_header *zhdr)
 352 {
 353         struct page *page = virt_to_page(zhdr);
 354
 355         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 356                 return 0; /* can't move middle chunk, it's used */
 357
 358         if (zhdr->middle_chunks == 0)
 359                 return 0; /* nothing to compact */
 360
 361         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 362                 /* move to the beginning */
 363                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 364                 zhdr->first_chunks = zhdr->middle_chunks;
 365                 zhdr->middle_chunks = 0;
 366                 zhdr->start_middle = 0;
 367                 zhdr->first_num++;
 368                 return 1;
 369         }
 370
 371         /*
 372          * moving data is expensive, so let's only do that if
 373          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 374          */
 375         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 376             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 377                         BIG_CHUNK_GAP) {
 378                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 379                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 380                 return 1;
 381         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 382                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 383                                         + zhdr->middle_chunks) >=
 384                         BIG_CHUNK_GAP) {
 385                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 386                         zhdr->middle_chunks;
 387                 mchunk_memmove(zhdr, new_start);
 388                 zhdr->start_middle = new_start;
 389                 return 1;
 390         }
 391
 392         return 0;
 393 }
 394
 395 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 396 {
 397         struct z3fold_pool *pool = zhdr->pool;
 398         struct page *page;
 399         struct list_head *unbuddied;
 400         int fchunks;
 401
 402         page = virt_to_page(zhdr);
 403         if (locked)
 404                 WARN_ON(z3fold_page_trylock(zhdr));
 405         else
 406                 z3fold_page_lock(zhdr);
 407         if (test_bit(PAGE_STALE, &page->private) ||
 408             !test_and_clear_bit(NEEDS_COMPACTING, &page->private)) {
 409                 z3fold_page_unlock(zhdr);
 410                 return;
 411         }
 412         spin_lock(&pool->lock);
 413         list_del_init(&zhdr->buddy);
 414         spin_unlock(&pool->lock);
 415
 416         z3fold_compact_page(zhdr);
 417         unbuddied = get_cpu_ptr(pool->unbuddied);
 418         fchunks = num_free_chunks(zhdr);
 419         if (fchunks < NCHUNKS &&
 420             (!zhdr->first_chunks || !zhdr->middle_chunks ||
 421                         !zhdr->last_chunks)) {
 422                 /* the page's not completely free and it's unbuddied */
 423                 spin_lock(&pool->lock);
 424                 list_add(&zhdr->buddy, &unbuddied[fchunks]);
 425                 spin_unlock(&pool->lock);
 426                 zhdr->cpu = smp_processor_id();
 427         }
 428         put_cpu_ptr(pool->unbuddied);
 429         z3fold_page_unlock(zhdr);
 430 }
 431
 432 static void compact_page_work(struct work_struct *w)
 433 {
 434         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 435                                                 work);
 436
 437         do_compact_page(zhdr, false);
 438 }
 439
 440
 441 /*
 442  * API Functions
 443  */
 444
 445 /**
 446  * z3fold_create_pool() - create a new z3fold pool
 447  * @name:       pool name
 448  * @gfp:        gfp flags when allocating the z3fold pool structure
 449  * @ops:        user-defined operations for the z3fold pool
 450  *
 451  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 452  * failed.
 453  */
 454 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 455                 const struct z3fold_ops *ops)
 456 {
 457         struct z3fold_pool *pool = NULL;
 458         int i, cpu;
 459
 460         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 461         if (!pool)
 462                 goto out;
 463         spin_lock_init(&pool->lock);
 464         spin_lock_init(&pool->stale_lock);
 465         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 466         for_each_possible_cpu(cpu) {
 467                 struct list_head *unbuddied =
 468                                 per_cpu_ptr(pool->unbuddied, cpu);
 469                 for_each_unbuddied_list(i, 0)
 470                         INIT_LIST_HEAD(&unbuddied[i]);
 471         }
 472         INIT_LIST_HEAD(&pool->lru);
 473         INIT_LIST_HEAD(&pool->stale);
 474         atomic64_set(&pool->pages_nr, 0);
 475         pool->name = name;
 476         pool->compact_wq = create_singlethread_workqueue(pool->name);
 477         if (!pool->compact_wq)
 478                 goto out;
 479         pool->release_wq = create_singlethread_workqueue(pool->name);
 480         if (!pool->release_wq)
 481                 goto out_wq;
 482         INIT_WORK(&pool->work, free_pages_work);
 483         pool->ops = ops;
 484         return pool;
 485
 486 out_wq:
 487         destroy_workqueue(pool->compact_wq);
 488 out:
 489         kfree(pool);
 490         return NULL;
 491 }
 492
 493 /**
 494  * z3fold_destroy_pool() - destroys an existing z3fold pool
 495  * @pool:       the z3fold pool to be destroyed
 496  *
 497  * The pool should be emptied before this function is called.
 498  */
 499 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 500 {
 501         destroy_workqueue(pool->release_wq);
 502         destroy_workqueue(pool->compact_wq);
 503         kfree(pool);
 504 }
 505
 506 /**
 507  * z3fold_alloc() - allocates a region of a given size
 508  * @pool:       z3fold pool from which to allocate
 509  * @size:       size in bytes of the desired allocation
 510  * @gfp:        gfp flags used if the pool needs to grow
 511  * @handle:     handle of the new allocation
 512  *
 513  * This function will attempt to find a free region in the pool large enough to
 514  * satisfy the allocation request.  A search of the unbuddied lists is
 515  * performed first. If no suitable free region is found, then a new page is
 516  * allocated and added to the pool to satisfy the request.
 517  *
 518  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 519  * as z3fold pool pages.
 520  *
 521  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 522  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 523  * a new page.
 524  */
 525 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 526                         unsigned long *handle)
 527 {
 528         int chunks = 0, i, freechunks;
 529         struct z3fold_header *zhdr = NULL;
 530         struct page *page = NULL;
 531         enum buddy bud;
 532         bool can_sleep = (gfp & __GFP_RECLAIM) == __GFP_RECLAIM;
 533
 534         if (!size || (gfp & __GFP_HIGHMEM))
 535                 return -EINVAL;
 536
 537         if (size > PAGE_SIZE)
 538                 return -ENOSPC;
 539
 540         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 541                 bud = HEADLESS;
 542         else {
 543                 struct list_head *unbuddied;
 544                 chunks = size_to_chunks(size);
 545
 546 lookup:
 547                 /* First, try to find an unbuddied z3fold page. */
 548                 unbuddied = get_cpu_ptr(pool->unbuddied);
 549                 for_each_unbuddied_list(i, chunks) {
 550                         struct list_head *l = &unbuddied[i];
 551
 552                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 553                                                 struct z3fold_header, buddy);
 554
 555                         if (!zhdr)
 556                                 continue;
 557
 558                         /* Re-check under lock. */
 559                         spin_lock(&pool->lock);
 560                         l = &unbuddied[i];
 561                         if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 562                                         struct z3fold_header, buddy)) ||
 563                             !z3fold_page_trylock(zhdr)) {
 564                                 spin_unlock(&pool->lock);
 565                                 put_cpu_ptr(pool->unbuddied);
 566                                 goto lookup;
 567                         }
 568                         list_del_init(&zhdr->buddy);
 569                         zhdr->cpu = -1;
 570                         spin_unlock(&pool->lock);
 571
 572                         page = virt_to_page(zhdr);
 573                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 574                                 z3fold_page_unlock(zhdr);
 575                                 zhdr = NULL;
 576                                 put_cpu_ptr(pool->unbuddied);
 577                                 if (can_sleep)
 578                                         cond_resched();
 579                                 goto lookup;
 580                         }
 581
 582                         /*
 583                          * this page could not be removed from its unbuddied
 584                          * list while pool lock was held, and then we've taken
 585                          * page lock so kref_put could not be called before
 586                          * we got here, so it's safe to just call kref_get()
 587                          */
 588                         kref_get(&zhdr->refcount);
 589                         break;
 590                 }
 591                 put_cpu_ptr(pool->unbuddied);
 592
 593                 if (zhdr) {
 594                         if (zhdr->first_chunks == 0) {
 595                                 if (zhdr->middle_chunks != 0 &&
 596                                     chunks >= zhdr->start_middle)
 597                                         bud = LAST;
 598                                 else
 599                                         bud = FIRST;
 600                         } else if (zhdr->last_chunks == 0)
 601                                 bud = LAST;
 602                         else if (zhdr->middle_chunks == 0)
 603                                 bud = MIDDLE;
 604                         else {
 605                                 if (kref_put(&zhdr->refcount,
 606                                              release_z3fold_page_locked))
 607                                         atomic64_dec(&pool->pages_nr);
 608                                 else
 609                                         z3fold_page_unlock(zhdr);
 610                                 pr_err("No free chunks in unbuddied\n");
 611                                 WARN_ON(1);
 612                                 goto lookup;
 613                         }
 614                         goto found;
 615                 }
 616                 bud = FIRST;
 617         }
 618
 619         spin_lock(&pool->stale_lock);
 620         zhdr = list_first_entry_or_null(&pool->stale,
 621                                         struct z3fold_header, buddy);
 622         /*
 623          * Before allocating a page, let's see if we can take one from the
 624          * stale pages list. cancel_work_sync() can sleep so we must make
 625          * sure it won't be called in case we're in atomic context.
 626          */
 627         if (zhdr && (can_sleep || !work_pending(&zhdr->work))) {
 628                 list_del(&zhdr->buddy);
 629                 spin_unlock(&pool->stale_lock);
 630                 if (can_sleep)
 631                         cancel_work_sync(&zhdr->work);
 632                 page = virt_to_page(zhdr);
 633         } else {
 634                 spin_unlock(&pool->stale_lock);
 635                 page = alloc_page(gfp);
 636         }
 637
 638         if (!page)
 639                 return -ENOMEM;
 640
 641         atomic64_inc(&pool->pages_nr);
 642         zhdr = init_z3fold_page(page, pool);
 643
 644         if (bud == HEADLESS) {
 645                 set_bit(PAGE_HEADLESS, &page->private);
 646                 goto headless;
 647         }
 648         z3fold_page_lock(zhdr);
 649
 650 found:
 651         if (bud == FIRST)
 652                 zhdr->first_chunks = chunks;
 653         else if (bud == LAST)
 654                 zhdr->last_chunks = chunks;
 655         else {
 656                 zhdr->middle_chunks = chunks;
 657                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 658         }
 659
 660         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 661                         zhdr->middle_chunks == 0) {
 662                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 663
 664                 /* Add to unbuddied list */
 665                 freechunks = num_free_chunks(zhdr);
 666                 spin_lock(&pool->lock);
 667                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 668                 spin_unlock(&pool->lock);
 669                 zhdr->cpu = smp_processor_id();
 670                 put_cpu_ptr(pool->unbuddied);
 671         }
 672
 673 headless:
 674         spin_lock(&pool->lock);
 675         /* Add/move z3fold page to beginning of LRU */
 676         if (!list_empty(&page->lru))
 677                 list_del(&page->lru);
 678
 679         list_add(&page->lru, &pool->lru);
 680
 681         *handle = encode_handle(zhdr, bud);
 682         spin_unlock(&pool->lock);
 683         if (bud != HEADLESS)
 684                 z3fold_page_unlock(zhdr);
 685
 686         return 0;
 687 }
 688
 689 /**
 690  * z3fold_free() - frees the allocation associated with the given handle
 691  * @pool:       pool in which the allocation resided
 692  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 693  *
 694  * In the case that the z3fold page in which the allocation resides is under
 695  * reclaim, as indicated by the PG_reclaim flag being set, this function
 696  * only sets the first|last_chunks to 0.  The page is actually freed
 697  * once both buddies are evicted (see z3fold_reclaim_page() below).
 698  */
 699 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 700 {
 701         struct z3fold_header *zhdr;
 702         struct page *page;
 703         enum buddy bud;
 704
 705         zhdr = handle_to_z3fold_header(handle);
 706         page = virt_to_page(zhdr);
 707
 708         if (test_bit(PAGE_HEADLESS, &page->private)) {
 709                 /* HEADLESS page stored */
 710                 bud = HEADLESS;
 711         } else {
 712                 z3fold_page_lock(zhdr);
 713                 bud = handle_to_buddy(handle);
 714
 715                 switch (bud) {
 716                 case FIRST:
 717                         zhdr->first_chunks = 0;
 718                         break;
 719                 case MIDDLE:
 720                         zhdr->middle_chunks = 0;
 721                         zhdr->start_middle = 0;
 722                         break;
 723                 case LAST:
 724                         zhdr->last_chunks = 0;
 725                         break;
 726                 default:
 727                         pr_err("%s: unknown bud %d\n", __func__, bud);
 728                         WARN_ON(1);
 729                         z3fold_page_unlock(zhdr);
 730                         return;
 731                 }
 732         }
 733
 734         if (bud == HEADLESS) {
 735                 spin_lock(&pool->lock);
 736                 list_del(&page->lru);
 737                 spin_unlock(&pool->lock);
 738                 free_z3fold_page(page);
 739                 atomic64_dec(&pool->pages_nr);
 740                 return;
 741         }
 742
 743         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
 744                 atomic64_dec(&pool->pages_nr);
 745                 return;
 746         }
 747         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
 748                 z3fold_page_unlock(zhdr);
 749                 return;
 750         }
 751         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
 752                 spin_lock(&pool->lock);
 753                 list_del_init(&zhdr->buddy);
 754                 spin_unlock(&pool->lock);
 755                 zhdr->cpu = -1;
 756                 do_compact_page(zhdr, true);
 757                 return;
 758         }
 759         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
 760         z3fold_page_unlock(zhdr);
 761 }
 762
 763 /**
 764  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
 765  * @pool:       pool from which a page will attempt to be evicted
 766  * @retires:    number of pages on the LRU list for which eviction will
 767  *              be attempted before failing
 768  *
 769  * z3fold reclaim is different from normal system reclaim in that it is done
 770  * from the bottom, up. This is because only the bottom layer, z3fold, has
 771  * information on how the allocations are organized within each z3fold page.
 772  * This has the potential to create interesting locking situations between
 773  * z3fold and the user, however.
 774  *
 775  * To avoid these, this is how z3fold_reclaim_page() should be called:
 776
 777  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
 778  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
 779  * call the user-defined eviction handler with the pool and handle as
 780  * arguments.
 781  *
 782  * If the handle can not be evicted, the eviction handler should return
 783  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
 784  * appropriate list and try the next z3fold page on the LRU up to
 785  * a user defined number of retries.
 786  *
 787  * If the handle is successfully evicted, the eviction handler should
 788  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
 789  * contains logic to delay freeing the page if the page is under reclaim,
 790  * as indicated by the setting of the PG_reclaim flag on the underlying page.
 791  *
 792  * If all buddies in the z3fold page are successfully evicted, then the
 793  * z3fold page can be freed.
 794  *
 795  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
 796  * no pages to evict or an eviction handler is not registered, -EAGAIN if
 797  * the retry limit was hit.
 798  */
 799 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 800 {
 801         int i, ret = 0;
 802         struct z3fold_header *zhdr = NULL;
 803         struct page *page = NULL;
 804         struct list_head *pos;
 805         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
 806
 807         spin_lock(&pool->lock);
 808         if (!pool->ops || !pool->ops->evict || retries == 0) {
 809                 spin_unlock(&pool->lock);
 810                 return -EINVAL;
 811         }
 812         for (i = 0; i < retries; i++) {
 813                 if (list_empty(&pool->lru)) {
 814                         spin_unlock(&pool->lock);
 815                         return -EINVAL;
 816                 }
 817                 list_for_each_prev(pos, &pool->lru) {
 818                         page = list_entry(pos, struct page, lru);
 819                         if (test_bit(PAGE_HEADLESS, &page->private))
 820                                 /* candidate found */
 821                                 break;
 822
 823                         zhdr = page_address(page);
 824                         if (!z3fold_page_trylock(zhdr))
 825                                 continue; /* can't evict at this point */
 826                         kref_get(&zhdr->refcount);
 827                         list_del_init(&zhdr->buddy);
 828                         zhdr->cpu = -1;
 829                 }
 830
 831                 list_del_init(&page->lru);
 832                 spin_unlock(&pool->lock);
 833
 834                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
 835                         /*
 836                          * We need encode the handles before unlocking, since
 837                          * we can race with free that will set
 838                          * (first|last)_chunks to 0
 839                          */
 840                         first_handle = 0;
 841                         last_handle = 0;
 842                         middle_handle = 0;
 843                         if (zhdr->first_chunks)
 844                                 first_handle = encode_handle(zhdr, FIRST);
 845                         if (zhdr->middle_chunks)
 846                                 middle_handle = encode_handle(zhdr, MIDDLE);
 847                         if (zhdr->last_chunks)
 848                                 last_handle = encode_handle(zhdr, LAST);
 849                         /*
 850                          * it's safe to unlock here because we hold a
 851                          * reference to this page
 852                          */
 853                         z3fold_page_unlock(zhdr);
 854                 } else {
 855                         first_handle = encode_handle(zhdr, HEADLESS);
 856                         last_handle = middle_handle = 0;
 857                 }
 858
 859                 /* Issue the eviction callback(s) */
 860                 if (middle_handle) {
 861                         ret = pool->ops->evict(pool, middle_handle);
 862                         if (ret)
 863                                 goto next;
 864                 }
 865                 if (first_handle) {
 866                         ret = pool->ops->evict(pool, first_handle);
 867                         if (ret)
 868                                 goto next;
 869                 }
 870                 if (last_handle) {
 871                         ret = pool->ops->evict(pool, last_handle);
 872                         if (ret)
 873                                 goto next;
 874                 }
 875 next:
 876                 spin_lock(&pool->lock);
 877                 if (test_bit(PAGE_HEADLESS, &page->private)) {
 878                         if (ret == 0) {
 879                                 spin_unlock(&pool->lock);
 880                                 free_z3fold_page(page);
 881                                 return 0;
 882                         }
 883                 } else if (kref_put(&zhdr->refcount, release_z3fold_page)) {
 884                         atomic64_dec(&pool->pages_nr);
 885                         spin_unlock(&pool->lock);
 886                         return 0;
 887                 }
 888
 889                 /*
 890                  * Add to the beginning of LRU.
 891                  * Pool lock has to be kept here to ensure the page has
 892                  * not already been released
 893                  */
 894                 list_add(&page->lru, &pool->lru);
 895         }
 896         spin_unlock(&pool->lock);
 897         return -EAGAIN;
 898 }
 899
 900 /**
 901  * z3fold_map() - maps the allocation associated with the given handle
 902  * @pool:       pool in which the allocation resides
 903  * @handle:     handle associated with the allocation to be mapped
 904  *
 905  * Extracts the buddy number from handle and constructs the pointer to the
 906  * correct starting chunk within the page.
 907  *
 908  * Returns: a pointer to the mapped allocation
 909  */
 910 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
 911 {
 912         struct z3fold_header *zhdr;
 913         struct page *page;
 914         void *addr;
 915         enum buddy buddy;
 916
 917         zhdr = handle_to_z3fold_header(handle);
 918         addr = zhdr;
 919         page = virt_to_page(zhdr);
 920
 921         if (test_bit(PAGE_HEADLESS, &page->private))
 922                 goto out;
 923
 924         z3fold_page_lock(zhdr);
 925         buddy = handle_to_buddy(handle);
 926         switch (buddy) {
 927         case FIRST:
 928                 addr += ZHDR_SIZE_ALIGNED;
 929                 break;
 930         case MIDDLE:
 931                 addr += zhdr->start_middle << CHUNK_SHIFT;
 932                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 933                 break;
 934         case LAST:
 935                 addr += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
 936                 break;
 937         default:
 938                 pr_err("unknown buddy id %d\n", buddy);
 939                 WARN_ON(1);
 940                 addr = NULL;
 941                 break;
 942         }
 943
 944         z3fold_page_unlock(zhdr);
 945 out:
 946         return addr;
 947 }
 948
 949 /**
 950  * z3fold_unmap() - unmaps the allocation associated with the given handle
 951  * @pool:       pool in which the allocation resides
 952  * @handle:     handle associated with the allocation to be unmapped
 953  */
 954 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
 955 {
 956         struct z3fold_header *zhdr;
 957         struct page *page;
 958         enum buddy buddy;
 959
 960         zhdr = handle_to_z3fold_header(handle);
 961         page = virt_to_page(zhdr);
 962
 963         if (test_bit(PAGE_HEADLESS, &page->private))
 964                 return;
 965
 966         z3fold_page_lock(zhdr);
 967         buddy = handle_to_buddy(handle);
 968         if (buddy == MIDDLE)
 969                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 970         z3fold_page_unlock(zhdr);
 971 }
 972
 973 /**
 974  * z3fold_get_pool_size() - gets the z3fold pool size in pages
 975  * @pool:       pool whose size is being queried
 976  *
 977  * Returns: size in pages of the given pool.
 978  */
 979 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
 980 {
 981         return atomic64_read(&pool->pages_nr);
 982 }
 983
 984 /*****************
 985  * zpool
 986  ****************/
 987
 988 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
 989 {
 990         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
 991                 return pool->zpool_ops->evict(pool->zpool, handle);
 992         else
 993                 return -ENOENT;
 994 }
 995
 996 static const struct z3fold_ops z3fold_zpool_ops = {
 997         .evict =        z3fold_zpool_evict
 998 };
 999
1000 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1001                                const struct zpool_ops *zpool_ops,
1002                                struct zpool *zpool)
1003 {
1004         struct z3fold_pool *pool;
1005
1006         pool = z3fold_create_pool(name, gfp,
1007                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1008         if (pool) {
1009                 pool->zpool = zpool;
1010                 pool->zpool_ops = zpool_ops;
1011         }
1012         return pool;
1013 }
1014
1015 static void z3fold_zpool_destroy(void *pool)
1016 {
1017         z3fold_destroy_pool(pool);
1018 }
1019
1020 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1021                         unsigned long *handle)
1022 {
1023         return z3fold_alloc(pool, size, gfp, handle);
1024 }
1025 static void z3fold_zpool_free(void *pool, unsigned long handle)
1026 {
1027         z3fold_free(pool, handle);
1028 }
1029
1030 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1031                         unsigned int *reclaimed)
1032 {
1033         unsigned int total = 0;
1034         int ret = -EINVAL;
1035
1036         while (total < pages) {
1037                 ret = z3fold_reclaim_page(pool, 8);
1038                 if (ret < 0)
1039                         break;
1040                 total++;
1041         }
1042
1043         if (reclaimed)
1044                 *reclaimed = total;
1045
1046         return ret;
1047 }
1048
1049 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1050                         enum zpool_mapmode mm)
1051 {
1052         return z3fold_map(pool, handle);
1053 }
1054 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1055 {
1056         z3fold_unmap(pool, handle);
1057 }
1058
1059 static u64 z3fold_zpool_total_size(void *pool)
1060 {
1061         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1062 }
1063
1064 static struct zpool_driver z3fold_zpool_driver = {
1065         .type =         "z3fold",
1066         .owner =        THIS_MODULE,
1067         .create =       z3fold_zpool_create,
1068         .destroy =      z3fold_zpool_destroy,
1069         .malloc =       z3fold_zpool_malloc,
1070         .free =         z3fold_zpool_free,
1071         .shrink =       z3fold_zpool_shrink,
1072         .map =          z3fold_zpool_map,
1073         .unmap =        z3fold_zpool_unmap,
1074         .total_size =   z3fold_zpool_total_size,
1075 };
1076
1077 MODULE_ALIAS("zpool-z3fold");
1078
1079 static int __init init_z3fold(void)
1080 {
1081         /* Make sure the z3fold header is not larger than the page size */
1082         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1083         zpool_register_driver(&z3fold_zpool_driver);
1084
1085         return 0;
1086 }
1087
1088 static void __exit exit_z3fold(void)
1089 {
1090         zpool_unregister_driver(&z3fold_zpool_driver);
1091 }
1092
1093 module_init(init_z3fold);
1094 module_exit(exit_z3fold);
1095
1096 MODULE_LICENSE("GPL");
1097 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1098 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");