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 #define BUDDY_SHIFT     2
 103
 104 /**
 105  * struct z3fold_pool - stores metadata for each z3fold pool
 106  * @name:       pool name
 107  * @lock:       protects pool unbuddied/lru lists
 108  * @stale_lock: protects pool stale page list
 109  * @unbuddied:  per-cpu array of lists tracking z3fold pages that contain 2-
 110  *              buddies; the list each z3fold page is added to depends on
 111  *              the size of its free region.
 112  * @lru:        list tracking the z3fold pages in LRU order by most recently
 113  *              added buddy.
 114  * @stale:      list of pages marked for freeing
 115  * @pages_nr:   number of z3fold pages in the pool.
 116  * @ops:        pointer to a structure of user defined operations specified at
 117  *              pool creation time.
 118  * @compact_wq: workqueue for page layout background optimization
 119  * @release_wq: workqueue for safe page release
 120  * @work:       work_struct for safe page release
 121  *
 122  * This structure is allocated at pool creation time and maintains metadata
 123  * pertaining to a particular z3fold pool.
 124  */
 125 struct z3fold_pool {
 126         const char *name;
 127         spinlock_t lock;
 128         spinlock_t stale_lock;
 129         struct list_head *unbuddied;
 130         struct list_head lru;
 131         struct list_head stale;
 132         atomic64_t pages_nr;
 133         const struct z3fold_ops *ops;
 134         struct zpool *zpool;
 135         const struct zpool_ops *zpool_ops;
 136         struct workqueue_struct *compact_wq;
 137         struct workqueue_struct *release_wq;
 138         struct work_struct work;
 139 };
 140
 141 /*
 142  * Internal z3fold page flags
 143  */
 144 enum z3fold_page_flags {
 145         PAGE_HEADLESS = 0,
 146         MIDDLE_CHUNK_MAPPED,
 147         NEEDS_COMPACTING,
 148         PAGE_STALE,
 149         PAGE_CLAIMED, /* by either reclaim or free */
 150 };
 151
 152 /*****************
 153  * Helpers
 154 *****************/
 155
 156 /* Converts an allocation size in bytes to size in z3fold chunks */
 157 static int size_to_chunks(size_t size)
 158 {
 159         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
 160 }
 161
 162 #define for_each_unbuddied_list(_iter, _begin) \
 163         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
 164
 165 static void compact_page_work(struct work_struct *w);
 166
 167 /* Initializes the z3fold header of a newly allocated z3fold page */
 168 static struct z3fold_header *init_z3fold_page(struct page *page,
 169                                         struct z3fold_pool *pool)
 170 {
 171         struct z3fold_header *zhdr = page_address(page);
 172
 173         INIT_LIST_HEAD(&page->lru);
 174         clear_bit(PAGE_HEADLESS, &page->private);
 175         clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
 176         clear_bit(NEEDS_COMPACTING, &page->private);
 177         clear_bit(PAGE_STALE, &page->private);
 178         clear_bit(PAGE_CLAIMED, &page->private);
 179
 180         spin_lock_init(&zhdr->page_lock);
 181         kref_init(&zhdr->refcount);
 182         zhdr->first_chunks = 0;
 183         zhdr->middle_chunks = 0;
 184         zhdr->last_chunks = 0;
 185         zhdr->first_num = 0;
 186         zhdr->start_middle = 0;
 187         zhdr->cpu = -1;
 188         zhdr->pool = pool;
 189         INIT_LIST_HEAD(&zhdr->buddy);
 190         INIT_WORK(&zhdr->work, compact_page_work);
 191         return zhdr;
 192 }
 193
 194 /* Resets the struct page fields and frees the page */
 195 static void free_z3fold_page(struct page *page)
 196 {
 197         __free_page(page);
 198 }
 199
 200 /* Lock a z3fold page */
 201 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
 202 {
 203         spin_lock(&zhdr->page_lock);
 204 }
 205
 206 /* Try to lock a z3fold page */
 207 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
 208 {
 209         return spin_trylock(&zhdr->page_lock);
 210 }
 211
 212 /* Unlock a z3fold page */
 213 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
 214 {
 215         spin_unlock(&zhdr->page_lock);
 216 }
 217
 218 /*
 219  * Encodes the handle of a particular buddy within a z3fold page
 220  * Pool lock should be held as this function accesses first_num
 221  */
 222 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
 223 {
 224         unsigned long handle;
 225
 226         handle = (unsigned long)zhdr;
 227         if (bud != HEADLESS) {
 228                 handle |= (bud + zhdr->first_num) & BUDDY_MASK;
 229                 if (bud == LAST)
 230                         handle |= (zhdr->last_chunks << BUDDY_SHIFT);
 231         }
 232         return handle;
 233 }
 234
 235 /* Returns the z3fold page where a given handle is stored */
 236 static struct z3fold_header *handle_to_z3fold_header(unsigned long handle)
 237 {
 238         return (struct z3fold_header *)(handle & PAGE_MASK);
 239 }
 240
 241 /* only for LAST bud, returns zero otherwise */
 242 static unsigned short handle_to_chunks(unsigned long handle)
 243 {
 244         return (handle & ~PAGE_MASK) >> BUDDY_SHIFT;
 245 }
 246
 247 /*
 248  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
 249  *  but that doesn't matter. because the masking will result in the
 250  *  correct buddy number.
 251  */
 252 static enum buddy handle_to_buddy(unsigned long handle)
 253 {
 254         struct z3fold_header *zhdr = handle_to_z3fold_header(handle);
 255         return (handle - zhdr->first_num) & BUDDY_MASK;
 256 }
 257
 258 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
 259 {
 260         return zhdr->pool;
 261 }
 262
 263 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
 264 {
 265         struct page *page = virt_to_page(zhdr);
 266         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 267
 268         WARN_ON(!list_empty(&zhdr->buddy));
 269         set_bit(PAGE_STALE, &page->private);
 270         clear_bit(NEEDS_COMPACTING, &page->private);
 271         spin_lock(&pool->lock);
 272         if (!list_empty(&page->lru))
 273                 list_del(&page->lru);
 274         spin_unlock(&pool->lock);
 275         if (locked)
 276                 z3fold_page_unlock(zhdr);
 277         spin_lock(&pool->stale_lock);
 278         list_add(&zhdr->buddy, &pool->stale);
 279         queue_work(pool->release_wq, &pool->work);
 280         spin_unlock(&pool->stale_lock);
 281 }
 282
 283 static void __attribute__((__unused__))
 284                         release_z3fold_page(struct kref *ref)
 285 {
 286         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 287                                                 refcount);
 288         __release_z3fold_page(zhdr, false);
 289 }
 290
 291 static void release_z3fold_page_locked(struct kref *ref)
 292 {
 293         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 294                                                 refcount);
 295         WARN_ON(z3fold_page_trylock(zhdr));
 296         __release_z3fold_page(zhdr, true);
 297 }
 298
 299 static void release_z3fold_page_locked_list(struct kref *ref)
 300 {
 301         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
 302                                                refcount);
 303         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 304         spin_lock(&pool->lock);
 305         list_del_init(&zhdr->buddy);
 306         spin_unlock(&pool->lock);
 307
 308         WARN_ON(z3fold_page_trylock(zhdr));
 309         __release_z3fold_page(zhdr, true);
 310 }
 311
 312 static void free_pages_work(struct work_struct *w)
 313 {
 314         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
 315
 316         spin_lock(&pool->stale_lock);
 317         while (!list_empty(&pool->stale)) {
 318                 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
 319                                                 struct z3fold_header, buddy);
 320                 struct page *page = virt_to_page(zhdr);
 321
 322                 list_del(&zhdr->buddy);
 323                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
 324                         continue;
 325                 spin_unlock(&pool->stale_lock);
 326                 cancel_work_sync(&zhdr->work);
 327                 free_z3fold_page(page);
 328                 cond_resched();
 329                 spin_lock(&pool->stale_lock);
 330         }
 331         spin_unlock(&pool->stale_lock);
 332 }
 333
 334 /*
 335  * Returns the number of free chunks in a z3fold page.
 336  * NB: can't be used with HEADLESS pages.
 337  */
 338 static int num_free_chunks(struct z3fold_header *zhdr)
 339 {
 340         int nfree;
 341         /*
 342          * If there is a middle object, pick up the bigger free space
 343          * either before or after it. Otherwise just subtract the number
 344          * of chunks occupied by the first and the last objects.
 345          */
 346         if (zhdr->middle_chunks != 0) {
 347                 int nfree_before = zhdr->first_chunks ?
 348                         0 : zhdr->start_middle - ZHDR_CHUNKS;
 349                 int nfree_after = zhdr->last_chunks ?
 350                         0 : TOTAL_CHUNKS -
 351                                 (zhdr->start_middle + zhdr->middle_chunks);
 352                 nfree = max(nfree_before, nfree_after);
 353         } else
 354                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
 355         return nfree;
 356 }
 357
 358 /* Add to the appropriate unbuddied list */
 359 static inline void add_to_unbuddied(struct z3fold_pool *pool,
 360                                 struct z3fold_header *zhdr)
 361 {
 362         if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
 363                         zhdr->middle_chunks == 0) {
 364                 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
 365
 366                 int freechunks = num_free_chunks(zhdr);
 367                 spin_lock(&pool->lock);
 368                 list_add(&zhdr->buddy, &unbuddied[freechunks]);
 369                 spin_unlock(&pool->lock);
 370                 zhdr->cpu = smp_processor_id();
 371                 put_cpu_ptr(pool->unbuddied);
 372         }
 373 }
 374
 375 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
 376                                 unsigned short dst_chunk)
 377 {
 378         void *beg = zhdr;
 379         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
 380                        beg + (zhdr->start_middle << CHUNK_SHIFT),
 381                        zhdr->middle_chunks << CHUNK_SHIFT);
 382 }
 383
 384 #define BIG_CHUNK_GAP   3
 385 /* Has to be called with lock held */
 386 static int z3fold_compact_page(struct z3fold_header *zhdr)
 387 {
 388         struct page *page = virt_to_page(zhdr);
 389
 390         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
 391                 return 0; /* can't move middle chunk, it's used */
 392
 393         if (zhdr->middle_chunks == 0)
 394                 return 0; /* nothing to compact */
 395
 396         if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
 397                 /* move to the beginning */
 398                 mchunk_memmove(zhdr, ZHDR_CHUNKS);
 399                 zhdr->first_chunks = zhdr->middle_chunks;
 400                 zhdr->middle_chunks = 0;
 401                 zhdr->start_middle = 0;
 402                 zhdr->first_num++;
 403                 return 1;
 404         }
 405
 406         /*
 407          * moving data is expensive, so let's only do that if
 408          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
 409          */
 410         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
 411             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
 412                         BIG_CHUNK_GAP) {
 413                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
 414                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 415                 return 1;
 416         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
 417                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
 418                                         + zhdr->middle_chunks) >=
 419                         BIG_CHUNK_GAP) {
 420                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
 421                         zhdr->middle_chunks;
 422                 mchunk_memmove(zhdr, new_start);
 423                 zhdr->start_middle = new_start;
 424                 return 1;
 425         }
 426
 427         return 0;
 428 }
 429
 430 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
 431 {
 432         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
 433         struct page *page;
 434
 435         page = virt_to_page(zhdr);
 436         if (locked)
 437                 WARN_ON(z3fold_page_trylock(zhdr));
 438         else
 439                 z3fold_page_lock(zhdr);
 440         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
 441                 z3fold_page_unlock(zhdr);
 442                 return;
 443         }
 444         spin_lock(&pool->lock);
 445         list_del_init(&zhdr->buddy);
 446         spin_unlock(&pool->lock);
 447
 448         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
 449                 atomic64_dec(&pool->pages_nr);
 450                 return;
 451         }
 452
 453         z3fold_compact_page(zhdr);
 454         add_to_unbuddied(pool, zhdr);
 455         z3fold_page_unlock(zhdr);
 456 }
 457
 458 static void compact_page_work(struct work_struct *w)
 459 {
 460         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
 461                                                 work);
 462
 463         do_compact_page(zhdr, false);
 464 }
 465
 466 /* returns _locked_ z3fold page header or NULL */
 467 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
 468                                                 size_t size, bool can_sleep)
 469 {
 470         struct z3fold_header *zhdr = NULL;
 471         struct page *page;
 472         struct list_head *unbuddied;
 473         int chunks = size_to_chunks(size), i;
 474
 475 lookup:
 476         /* First, try to find an unbuddied z3fold page. */
 477         unbuddied = get_cpu_ptr(pool->unbuddied);
 478         for_each_unbuddied_list(i, chunks) {
 479                 struct list_head *l = &unbuddied[i];
 480
 481                 zhdr = list_first_entry_or_null(READ_ONCE(l),
 482                                         struct z3fold_header, buddy);
 483
 484                 if (!zhdr)
 485                         continue;
 486
 487                 /* Re-check under lock. */
 488                 spin_lock(&pool->lock);
 489                 l = &unbuddied[i];
 490                 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
 491                                                 struct z3fold_header, buddy)) ||
 492                     !z3fold_page_trylock(zhdr)) {
 493                         spin_unlock(&pool->lock);
 494                         zhdr = NULL;
 495                         put_cpu_ptr(pool->unbuddied);
 496                         if (can_sleep)
 497                                 cond_resched();
 498                         goto lookup;
 499                 }
 500                 list_del_init(&zhdr->buddy);
 501                 zhdr->cpu = -1;
 502                 spin_unlock(&pool->lock);
 503
 504                 page = virt_to_page(zhdr);
 505                 if (test_bit(NEEDS_COMPACTING, &page->private)) {
 506                         z3fold_page_unlock(zhdr);
 507                         zhdr = NULL;
 508                         put_cpu_ptr(pool->unbuddied);
 509                         if (can_sleep)
 510                                 cond_resched();
 511                         goto lookup;
 512                 }
 513
 514                 /*
 515                  * this page could not be removed from its unbuddied
 516                  * list while pool lock was held, and then we've taken
 517                  * page lock so kref_put could not be called before
 518                  * we got here, so it's safe to just call kref_get()
 519                  */
 520                 kref_get(&zhdr->refcount);
 521                 break;
 522         }
 523         put_cpu_ptr(pool->unbuddied);
 524
 525         if (!zhdr) {
 526                 int cpu;
 527
 528                 /* look for _exact_ match on other cpus' lists */
 529                 for_each_online_cpu(cpu) {
 530                         struct list_head *l;
 531
 532                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
 533                         spin_lock(&pool->lock);
 534                         l = &unbuddied[chunks];
 535
 536                         zhdr = list_first_entry_or_null(READ_ONCE(l),
 537                                                 struct z3fold_header, buddy);
 538
 539                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
 540                                 spin_unlock(&pool->lock);
 541                                 zhdr = NULL;
 542                                 continue;
 543                         }
 544                         list_del_init(&zhdr->buddy);
 545                         zhdr->cpu = -1;
 546                         spin_unlock(&pool->lock);
 547
 548                         page = virt_to_page(zhdr);
 549                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
 550                                 z3fold_page_unlock(zhdr);
 551                                 zhdr = NULL;
 552                                 if (can_sleep)
 553                                         cond_resched();
 554                                 continue;
 555                         }
 556                         kref_get(&zhdr->refcount);
 557                         break;
 558                 }
 559         }
 560
 561         return zhdr;
 562 }
 563
 564 /*
 565  * API Functions
 566  */
 567
 568 /**
 569  * z3fold_create_pool() - create a new z3fold pool
 570  * @name:       pool name
 571  * @gfp:        gfp flags when allocating the z3fold pool structure
 572  * @ops:        user-defined operations for the z3fold pool
 573  *
 574  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
 575  * failed.
 576  */
 577 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
 578                 const struct z3fold_ops *ops)
 579 {
 580         struct z3fold_pool *pool = NULL;
 581         int i, cpu;
 582
 583         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
 584         if (!pool)
 585                 goto out;
 586         spin_lock_init(&pool->lock);
 587         spin_lock_init(&pool->stale_lock);
 588         pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
 589         if (!pool->unbuddied)
 590                 goto out_pool;
 591         for_each_possible_cpu(cpu) {
 592                 struct list_head *unbuddied =
 593                                 per_cpu_ptr(pool->unbuddied, cpu);
 594                 for_each_unbuddied_list(i, 0)
 595                         INIT_LIST_HEAD(&unbuddied[i]);
 596         }
 597         INIT_LIST_HEAD(&pool->lru);
 598         INIT_LIST_HEAD(&pool->stale);
 599         atomic64_set(&pool->pages_nr, 0);
 600         pool->name = name;
 601         pool->compact_wq = create_singlethread_workqueue(pool->name);
 602         if (!pool->compact_wq)
 603                 goto out_unbuddied;
 604         pool->release_wq = create_singlethread_workqueue(pool->name);
 605         if (!pool->release_wq)
 606                 goto out_wq;
 607         INIT_WORK(&pool->work, free_pages_work);
 608         pool->ops = ops;
 609         return pool;
 610
 611 out_wq:
 612         destroy_workqueue(pool->compact_wq);
 613 out_unbuddied:
 614         free_percpu(pool->unbuddied);
 615 out_pool:
 616         kfree(pool);
 617 out:
 618         return NULL;
 619 }
 620
 621 /**
 622  * z3fold_destroy_pool() - destroys an existing z3fold pool
 623  * @pool:       the z3fold pool to be destroyed
 624  *
 625  * The pool should be emptied before this function is called.
 626  */
 627 static void z3fold_destroy_pool(struct z3fold_pool *pool)
 628 {
 629         destroy_workqueue(pool->release_wq);
 630         destroy_workqueue(pool->compact_wq);
 631         kfree(pool);
 632 }
 633
 634 /**
 635  * z3fold_alloc() - allocates a region of a given size
 636  * @pool:       z3fold pool from which to allocate
 637  * @size:       size in bytes of the desired allocation
 638  * @gfp:        gfp flags used if the pool needs to grow
 639  * @handle:     handle of the new allocation
 640  *
 641  * This function will attempt to find a free region in the pool large enough to
 642  * satisfy the allocation request.  A search of the unbuddied lists is
 643  * performed first. If no suitable free region is found, then a new page is
 644  * allocated and added to the pool to satisfy the request.
 645  *
 646  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
 647  * as z3fold pool pages.
 648  *
 649  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
 650  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
 651  * a new page.
 652  */
 653 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
 654                         unsigned long *handle)
 655 {
 656         int chunks = size_to_chunks(size);
 657         struct z3fold_header *zhdr = NULL;
 658         struct page *page = NULL;
 659         enum buddy bud;
 660         bool can_sleep = gfpflags_allow_blocking(gfp);
 661
 662         if (!size || (gfp & __GFP_HIGHMEM))
 663                 return -EINVAL;
 664
 665         if (size > PAGE_SIZE)
 666                 return -ENOSPC;
 667
 668         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
 669                 bud = HEADLESS;
 670         else {
 671 retry:
 672                 zhdr = __z3fold_alloc(pool, size, can_sleep);
 673                 if (zhdr) {
 674                         if (zhdr->first_chunks == 0) {
 675                                 if (zhdr->middle_chunks != 0 &&
 676                                     chunks >= zhdr->start_middle)
 677                                         bud = LAST;
 678                                 else
 679                                         bud = FIRST;
 680                         } else if (zhdr->last_chunks == 0)
 681                                 bud = LAST;
 682                         else if (zhdr->middle_chunks == 0)
 683                                 bud = MIDDLE;
 684                         else {
 685                                 if (kref_put(&zhdr->refcount,
 686                                              release_z3fold_page_locked))
 687                                         atomic64_dec(&pool->pages_nr);
 688                                 else
 689                                         z3fold_page_unlock(zhdr);
 690                                 pr_err("No free chunks in unbuddied\n");
 691                                 WARN_ON(1);
 692                                 goto retry;
 693                         }
 694                         page = virt_to_page(zhdr);
 695                         goto found;
 696                 }
 697                 bud = FIRST;
 698         }
 699
 700         page = NULL;
 701         if (can_sleep) {
 702                 spin_lock(&pool->stale_lock);
 703                 zhdr = list_first_entry_or_null(&pool->stale,
 704                                                 struct z3fold_header, buddy);
 705                 /*
 706                  * Before allocating a page, let's see if we can take one from
 707                  * the stale pages list. cancel_work_sync() can sleep so we
 708                  * limit this case to the contexts where we can sleep
 709                  */
 710                 if (zhdr) {
 711                         list_del(&zhdr->buddy);
 712                         spin_unlock(&pool->stale_lock);
 713                         cancel_work_sync(&zhdr->work);
 714                         page = virt_to_page(zhdr);
 715                 } else {
 716                         spin_unlock(&pool->stale_lock);
 717                 }
 718         }
 719         if (!page)
 720                 page = alloc_page(gfp);
 721
 722         if (!page)
 723                 return -ENOMEM;
 724
 725         zhdr = init_z3fold_page(page, pool);
 726         if (!zhdr) {
 727                 __free_page(page);
 728                 return -ENOMEM;
 729         }
 730         atomic64_inc(&pool->pages_nr);
 731
 732         if (bud == HEADLESS) {
 733                 set_bit(PAGE_HEADLESS, &page->private);
 734                 goto headless;
 735         }
 736         z3fold_page_lock(zhdr);
 737
 738 found:
 739         if (bud == FIRST)
 740                 zhdr->first_chunks = chunks;
 741         else if (bud == LAST)
 742                 zhdr->last_chunks = chunks;
 743         else {
 744                 zhdr->middle_chunks = chunks;
 745                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
 746         }
 747         add_to_unbuddied(pool, zhdr);
 748
 749 headless:
 750         spin_lock(&pool->lock);
 751         /* Add/move z3fold page to beginning of LRU */
 752         if (!list_empty(&page->lru))
 753                 list_del(&page->lru);
 754
 755         list_add(&page->lru, &pool->lru);
 756
 757         *handle = encode_handle(zhdr, bud);
 758         spin_unlock(&pool->lock);
 759         if (bud != HEADLESS)
 760                 z3fold_page_unlock(zhdr);
 761
 762         return 0;
 763 }
 764
 765 /**
 766  * z3fold_free() - frees the allocation associated with the given handle
 767  * @pool:       pool in which the allocation resided
 768  * @handle:     handle associated with the allocation returned by z3fold_alloc()
 769  *
 770  * In the case that the z3fold page in which the allocation resides is under
 771  * reclaim, as indicated by the PG_reclaim flag being set, this function
 772  * only sets the first|last_chunks to 0.  The page is actually freed
 773  * once both buddies are evicted (see z3fold_reclaim_page() below).
 774  */
 775 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
 776 {
 777         struct z3fold_header *zhdr;
 778         struct page *page;
 779         enum buddy bud;
 780
 781         zhdr = handle_to_z3fold_header(handle);
 782         page = virt_to_page(zhdr);
 783
 784         if (test_bit(PAGE_HEADLESS, &page->private)) {
 785                 /* if a headless page is under reclaim, just leave.
 786                  * NB: we use test_and_set_bit for a reason: if the bit
 787                  * has not been set before, we release this page
 788                  * immediately so we don't care about its value any more.
 789                  */
 790                 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
 791                         spin_lock(&pool->lock);
 792                         list_del(&page->lru);
 793                         spin_unlock(&pool->lock);
 794                         free_z3fold_page(page);
 795                         atomic64_dec(&pool->pages_nr);
 796                 }
 797                 return;
 798         }
 799
 800         /* Non-headless case */
 801         z3fold_page_lock(zhdr);
 802         bud = handle_to_buddy(handle);
 803
 804         switch (bud) {
 805         case FIRST:
 806                 zhdr->first_chunks = 0;
 807                 break;
 808         case MIDDLE:
 809                 zhdr->middle_chunks = 0;
 810                 break;
 811         case LAST:
 812                 zhdr->last_chunks = 0;
 813                 break;
 814         default:
 815                 pr_err("%s: unknown bud %d\n", __func__, bud);
 816                 WARN_ON(1);
 817                 z3fold_page_unlock(zhdr);
 818                 return;
 819         }
 820
 821         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
 822                 atomic64_dec(&pool->pages_nr);
 823                 return;
 824         }
 825         if (test_bit(PAGE_CLAIMED, &page->private)) {
 826                 z3fold_page_unlock(zhdr);
 827                 return;
 828         }
 829         if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
 830                 z3fold_page_unlock(zhdr);
 831                 return;
 832         }
 833         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
 834                 spin_lock(&pool->lock);
 835                 list_del_init(&zhdr->buddy);
 836                 spin_unlock(&pool->lock);
 837                 zhdr->cpu = -1;
 838                 kref_get(&zhdr->refcount);
 839                 do_compact_page(zhdr, true);
 840                 return;
 841         }
 842         kref_get(&zhdr->refcount);
 843         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
 844         z3fold_page_unlock(zhdr);
 845 }
 846
 847 /**
 848  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
 849  * @pool:       pool from which a page will attempt to be evicted
 850  * @retries:    number of pages on the LRU list for which eviction will
 851  *              be attempted before failing
 852  *
 853  * z3fold reclaim is different from normal system reclaim in that it is done
 854  * from the bottom, up. This is because only the bottom layer, z3fold, has
 855  * information on how the allocations are organized within each z3fold page.
 856  * This has the potential to create interesting locking situations between
 857  * z3fold and the user, however.
 858  *
 859  * To avoid these, this is how z3fold_reclaim_page() should be called:
 860  *
 861  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
 862  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
 863  * call the user-defined eviction handler with the pool and handle as
 864  * arguments.
 865  *
 866  * If the handle can not be evicted, the eviction handler should return
 867  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
 868  * appropriate list and try the next z3fold page on the LRU up to
 869  * a user defined number of retries.
 870  *
 871  * If the handle is successfully evicted, the eviction handler should
 872  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
 873  * contains logic to delay freeing the page if the page is under reclaim,
 874  * as indicated by the setting of the PG_reclaim flag on the underlying page.
 875  *
 876  * If all buddies in the z3fold page are successfully evicted, then the
 877  * z3fold page can be freed.
 878  *
 879  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
 880  * no pages to evict or an eviction handler is not registered, -EAGAIN if
 881  * the retry limit was hit.
 882  */
 883 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
 884 {
 885         int i, ret = 0;
 886         struct z3fold_header *zhdr = NULL;
 887         struct page *page = NULL;
 888         struct list_head *pos;
 889         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
 890
 891         spin_lock(&pool->lock);
 892         if (!pool->ops || !pool->ops->evict || retries == 0) {
 893                 spin_unlock(&pool->lock);
 894                 return -EINVAL;
 895         }
 896         for (i = 0; i < retries; i++) {
 897                 if (list_empty(&pool->lru)) {
 898                         spin_unlock(&pool->lock);
 899                         return -EINVAL;
 900                 }
 901                 list_for_each_prev(pos, &pool->lru) {
 902                         page = list_entry(pos, struct page, lru);
 903
 904                         /* this bit could have been set by free, in which case
 905                          * we pass over to the next page in the pool.
 906                          */
 907                         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
 908                                 continue;
 909
 910                         zhdr = page_address(page);
 911                         if (test_bit(PAGE_HEADLESS, &page->private))
 912                                 break;
 913
 914                         if (!z3fold_page_trylock(zhdr)) {
 915                                 zhdr = NULL;
 916                                 continue; /* can't evict at this point */
 917                         }
 918                         kref_get(&zhdr->refcount);
 919                         list_del_init(&zhdr->buddy);
 920                         zhdr->cpu = -1;
 921                         break;
 922                 }
 923
 924                 if (!zhdr)
 925                         break;
 926
 927                 list_del_init(&page->lru);
 928                 spin_unlock(&pool->lock);
 929
 930                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
 931                         /*
 932                          * We need encode the handles before unlocking, since
 933                          * we can race with free that will set
 934                          * (first|last)_chunks to 0
 935                          */
 936                         first_handle = 0;
 937                         last_handle = 0;
 938                         middle_handle = 0;
 939                         if (zhdr->first_chunks)
 940                                 first_handle = encode_handle(zhdr, FIRST);
 941                         if (zhdr->middle_chunks)
 942                                 middle_handle = encode_handle(zhdr, MIDDLE);
 943                         if (zhdr->last_chunks)
 944                                 last_handle = encode_handle(zhdr, LAST);
 945                         /*
 946                          * it's safe to unlock here because we hold a
 947                          * reference to this page
 948                          */
 949                         z3fold_page_unlock(zhdr);
 950                 } else {
 951                         first_handle = encode_handle(zhdr, HEADLESS);
 952                         last_handle = middle_handle = 0;
 953                 }
 954
 955                 /* Issue the eviction callback(s) */
 956                 if (middle_handle) {
 957                         ret = pool->ops->evict(pool, middle_handle);
 958                         if (ret)
 959                                 goto next;
 960                 }
 961                 if (first_handle) {
 962                         ret = pool->ops->evict(pool, first_handle);
 963                         if (ret)
 964                                 goto next;
 965                 }
 966                 if (last_handle) {
 967                         ret = pool->ops->evict(pool, last_handle);
 968                         if (ret)
 969                                 goto next;
 970                 }
 971 next:
 972                 if (test_bit(PAGE_HEADLESS, &page->private)) {
 973                         if (ret == 0) {
 974                                 free_z3fold_page(page);
 975                                 atomic64_dec(&pool->pages_nr);
 976                                 return 0;
 977                         }
 978                         spin_lock(&pool->lock);
 979                         list_add(&page->lru, &pool->lru);
 980                         spin_unlock(&pool->lock);
 981                 } else {
 982                         z3fold_page_lock(zhdr);
 983                         clear_bit(PAGE_CLAIMED, &page->private);
 984                         if (kref_put(&zhdr->refcount,
 985                                         release_z3fold_page_locked)) {
 986                                 atomic64_dec(&pool->pages_nr);
 987                                 return 0;
 988                         }
 989                         /*
 990                          * if we are here, the page is still not completely
 991                          * free. Take the global pool lock then to be able
 992                          * to add it back to the lru list
 993                          */
 994                         spin_lock(&pool->lock);
 995                         list_add(&page->lru, &pool->lru);
 996                         spin_unlock(&pool->lock);
 997                         z3fold_page_unlock(zhdr);
 998                 }
 999
1000                 /* We started off locked to we need to lock the pool back */
1001                 spin_lock(&pool->lock);
1002         }
1003         spin_unlock(&pool->lock);
1004         return -EAGAIN;
1005 }
1006
1007 /**
1008  * z3fold_map() - maps the allocation associated with the given handle
1009  * @pool:       pool in which the allocation resides
1010  * @handle:     handle associated with the allocation to be mapped
1011  *
1012  * Extracts the buddy number from handle and constructs the pointer to the
1013  * correct starting chunk within the page.
1014  *
1015  * Returns: a pointer to the mapped allocation
1016  */
1017 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1018 {
1019         struct z3fold_header *zhdr;
1020         struct page *page;
1021         void *addr;
1022         enum buddy buddy;
1023
1024         zhdr = handle_to_z3fold_header(handle);
1025         addr = zhdr;
1026         page = virt_to_page(zhdr);
1027
1028         if (test_bit(PAGE_HEADLESS, &page->private))
1029                 goto out;
1030
1031         z3fold_page_lock(zhdr);
1032         buddy = handle_to_buddy(handle);
1033         switch (buddy) {
1034         case FIRST:
1035                 addr += ZHDR_SIZE_ALIGNED;
1036                 break;
1037         case MIDDLE:
1038                 addr += zhdr->start_middle << CHUNK_SHIFT;
1039                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1040                 break;
1041         case LAST:
1042                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1043                 break;
1044         default:
1045                 pr_err("unknown buddy id %d\n", buddy);
1046                 WARN_ON(1);
1047                 addr = NULL;
1048                 break;
1049         }
1050
1051         z3fold_page_unlock(zhdr);
1052 out:
1053         return addr;
1054 }
1055
1056 /**
1057  * z3fold_unmap() - unmaps the allocation associated with the given handle
1058  * @pool:       pool in which the allocation resides
1059  * @handle:     handle associated with the allocation to be unmapped
1060  */
1061 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1062 {
1063         struct z3fold_header *zhdr;
1064         struct page *page;
1065         enum buddy buddy;
1066
1067         zhdr = handle_to_z3fold_header(handle);
1068         page = virt_to_page(zhdr);
1069
1070         if (test_bit(PAGE_HEADLESS, &page->private))
1071                 return;
1072
1073         z3fold_page_lock(zhdr);
1074         buddy = handle_to_buddy(handle);
1075         if (buddy == MIDDLE)
1076                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1077         z3fold_page_unlock(zhdr);
1078 }
1079
1080 /**
1081  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1082  * @pool:       pool whose size is being queried
1083  *
1084  * Returns: size in pages of the given pool.
1085  */
1086 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1087 {
1088         return atomic64_read(&pool->pages_nr);
1089 }
1090
1091 /*****************
1092  * zpool
1093  ****************/
1094
1095 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1096 {
1097         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1098                 return pool->zpool_ops->evict(pool->zpool, handle);
1099         else
1100                 return -ENOENT;
1101 }
1102
1103 static const struct z3fold_ops z3fold_zpool_ops = {
1104         .evict =        z3fold_zpool_evict
1105 };
1106
1107 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1108                                const struct zpool_ops *zpool_ops,
1109                                struct zpool *zpool)
1110 {
1111         struct z3fold_pool *pool;
1112
1113         pool = z3fold_create_pool(name, gfp,
1114                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1115         if (pool) {
1116                 pool->zpool = zpool;
1117                 pool->zpool_ops = zpool_ops;
1118         }
1119         return pool;
1120 }
1121
1122 static void z3fold_zpool_destroy(void *pool)
1123 {
1124         z3fold_destroy_pool(pool);
1125 }
1126
1127 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1128                         unsigned long *handle)
1129 {
1130         return z3fold_alloc(pool, size, gfp, handle);
1131 }
1132 static void z3fold_zpool_free(void *pool, unsigned long handle)
1133 {
1134         z3fold_free(pool, handle);
1135 }
1136
1137 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1138                         unsigned int *reclaimed)
1139 {
1140         unsigned int total = 0;
1141         int ret = -EINVAL;
1142
1143         while (total < pages) {
1144                 ret = z3fold_reclaim_page(pool, 8);
1145                 if (ret < 0)
1146                         break;
1147                 total++;
1148         }
1149
1150         if (reclaimed)
1151                 *reclaimed = total;
1152
1153         return ret;
1154 }
1155
1156 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1157                         enum zpool_mapmode mm)
1158 {
1159         return z3fold_map(pool, handle);
1160 }
1161 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1162 {
1163         z3fold_unmap(pool, handle);
1164 }
1165
1166 static u64 z3fold_zpool_total_size(void *pool)
1167 {
1168         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1169 }
1170
1171 static struct zpool_driver z3fold_zpool_driver = {
1172         .type =         "z3fold",
1173         .owner =        THIS_MODULE,
1174         .create =       z3fold_zpool_create,
1175         .destroy =      z3fold_zpool_destroy,
1176         .malloc =       z3fold_zpool_malloc,
1177         .free =         z3fold_zpool_free,
1178         .shrink =       z3fold_zpool_shrink,
1179         .map =          z3fold_zpool_map,
1180         .unmap =        z3fold_zpool_unmap,
1181         .total_size =   z3fold_zpool_total_size,
1182 };
1183
1184 MODULE_ALIAS("zpool-z3fold");
1185
1186 static int __init init_z3fold(void)
1187 {
1188         /* Make sure the z3fold header is not larger than the page size */
1189         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1190         zpool_register_driver(&z3fold_zpool_driver);
1191
1192         return 0;
1193 }
1194
1195 static void __exit exit_z3fold(void)
1196 {
1197         zpool_unregister_driver(&z3fold_zpool_driver);
1198 }
1199
1200 module_init(init_z3fold);
1201 module_exit(exit_z3fold);
1202
1203 MODULE_LICENSE("GPL");
1204 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1205 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");