fs/erofs/zdata.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             https://www.huawei.com/
   5  * Created by Gao Xiang <[email protected]>
   6  */
   7 #include "zdata.h"
   8 #include "compress.h"
   9 #include <linux/prefetch.h>
  10
  11 #include <trace/events/erofs.h>
  12
  13 /*
  14  * a compressed_pages[] placeholder in order to avoid
  15  * being filled with file pages for in-place decompression.
  16  */
  17 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
  18
  19 /* how to allocate cached pages for a pcluster */
  20 enum z_erofs_cache_alloctype {
  21         DONTALLOC,      /* don't allocate any cached pages */
  22         DELAYEDALLOC,   /* delayed allocation (at the time of submitting io) */
  23 };
  24
  25 /*
  26  * tagged pointer with 1-bit tag for all compressed pages
  27  * tag 0 - the page is just found with an extra page reference
  28  */
  29 typedef tagptr1_t compressed_page_t;
  30
  31 #define tag_compressed_page_justfound(page) \
  32         tagptr_fold(compressed_page_t, page, 1)
  33
  34 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
  35 static struct kmem_cache *pcluster_cachep __read_mostly;
  36
  37 void z_erofs_exit_zip_subsystem(void)
  38 {
  39         destroy_workqueue(z_erofs_workqueue);
  40         kmem_cache_destroy(pcluster_cachep);
  41 }
  42
  43 static inline int z_erofs_init_workqueue(void)
  44 {
  45         const unsigned int onlinecpus = num_possible_cpus();
  46
  47         /*
  48          * no need to spawn too many threads, limiting threads could minimum
  49          * scheduling overhead, perhaps per-CPU threads should be better?
  50          */
  51         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
  52                                             WQ_UNBOUND | WQ_HIGHPRI,
  53                                             onlinecpus + onlinecpus / 4);
  54         return z_erofs_workqueue ? 0 : -ENOMEM;
  55 }
  56
  57 static void z_erofs_pcluster_init_once(void *ptr)
  58 {
  59         struct z_erofs_pcluster *pcl = ptr;
  60         struct z_erofs_collection *cl = z_erofs_primarycollection(pcl);
  61         unsigned int i;
  62
  63         mutex_init(&cl->lock);
  64         cl->nr_pages = 0;
  65         cl->vcnt = 0;
  66         for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i)
  67                 pcl->compressed_pages[i] = NULL;
  68 }
  69
  70 int __init z_erofs_init_zip_subsystem(void)
  71 {
  72         pcluster_cachep = kmem_cache_create("erofs_compress",
  73                                             Z_EROFS_WORKGROUP_SIZE, 0,
  74                                             SLAB_RECLAIM_ACCOUNT,
  75                                             z_erofs_pcluster_init_once);
  76         if (pcluster_cachep) {
  77                 if (!z_erofs_init_workqueue())
  78                         return 0;
  79
  80                 kmem_cache_destroy(pcluster_cachep);
  81         }
  82         return -ENOMEM;
  83 }
  84
  85 enum z_erofs_collectmode {
  86         COLLECT_SECONDARY,
  87         COLLECT_PRIMARY,
  88         /*
  89          * The current collection was the tail of an exist chain, in addition
  90          * that the previous processed chained collections are all decided to
  91          * be hooked up to it.
  92          * A new chain will be created for the remaining collections which are
  93          * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
  94          * the next collection cannot reuse the whole page safely in
  95          * the following scenario:
  96          *  ________________________________________________________________
  97          * |      tail (partial) page     |       head (partial) page       |
  98          * |   (belongs to the next cl)   |   (belongs to the current cl)   |
  99          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 100          */
 101         COLLECT_PRIMARY_HOOKED,
 102         COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
 103         /*
 104          * The current collection has been linked with the owned chain, and
 105          * could also be linked with the remaining collections, which means
 106          * if the processing page is the tail page of the collection, thus
 107          * the current collection can safely use the whole page (since
 108          * the previous collection is under control) for in-place I/O, as
 109          * illustrated below:
 110          *  ________________________________________________________________
 111          * |  tail (partial) page |          head (partial) page           |
 112          * |  (of the current cl) |      (of the previous collection)      |
 113          * |  PRIMARY_FOLLOWED or |                                        |
 114          * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
 115          *
 116          * [  (*) the above page can be used as inplace I/O.               ]
 117          */
 118         COLLECT_PRIMARY_FOLLOWED,
 119 };
 120
 121 struct z_erofs_collector {
 122         struct z_erofs_pagevec_ctor vector;
 123
 124         struct z_erofs_pcluster *pcl, *tailpcl;
 125         struct z_erofs_collection *cl;
 126         struct page **compressedpages;
 127         z_erofs_next_pcluster_t owned_head;
 128
 129         enum z_erofs_collectmode mode;
 130 };
 131
 132 struct z_erofs_decompress_frontend {
 133         struct inode *const inode;
 134
 135         struct z_erofs_collector clt;
 136         struct erofs_map_blocks map;
 137
 138         bool readahead;
 139         /* used for applying cache strategy on the fly */
 140         bool backmost;
 141         erofs_off_t headoffset;
 142 };
 143
 144 #define COLLECTOR_INIT() { \
 145         .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 146         .mode = COLLECT_PRIMARY_FOLLOWED }
 147
 148 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 149         .inode = __i, .clt = COLLECTOR_INIT(), \
 150         .backmost = true, }
 151
 152 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
 153 static DEFINE_MUTEX(z_pagemap_global_lock);
 154
 155 static void preload_compressed_pages(struct z_erofs_collector *clt,
 156                                      struct address_space *mc,
 157                                      enum z_erofs_cache_alloctype type)
 158 {
 159         const struct z_erofs_pcluster *pcl = clt->pcl;
 160         const unsigned int clusterpages = BIT(pcl->clusterbits);
 161         struct page **pages = clt->compressedpages;
 162         pgoff_t index = pcl->obj.index + (pages - pcl->compressed_pages);
 163         bool standalone = true;
 164
 165         if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
 166                 return;
 167
 168         for (; pages < pcl->compressed_pages + clusterpages; ++pages) {
 169                 struct page *page;
 170                 compressed_page_t t;
 171
 172                 /* the compressed page was loaded before */
 173                 if (READ_ONCE(*pages))
 174                         continue;
 175
 176                 page = find_get_page(mc, index);
 177
 178                 if (page) {
 179                         t = tag_compressed_page_justfound(page);
 180                 } else if (type == DELAYEDALLOC) {
 181                         t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
 182                 } else {        /* DONTALLOC */
 183                         if (standalone)
 184                                 clt->compressedpages = pages;
 185                         standalone = false;
 186                         continue;
 187                 }
 188
 189                 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
 190                         continue;
 191
 192                 if (page)
 193                         put_page(page);
 194         }
 195
 196         if (standalone)         /* downgrade to PRIMARY_FOLLOWED_NOINPLACE */
 197                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 198 }
 199
 200 /* called by erofs_shrinker to get rid of all compressed_pages */
 201 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 202                                        struct erofs_workgroup *grp)
 203 {
 204         struct z_erofs_pcluster *const pcl =
 205                 container_of(grp, struct z_erofs_pcluster, obj);
 206         struct address_space *const mapping = MNGD_MAPPING(sbi);
 207         const unsigned int clusterpages = BIT(pcl->clusterbits);
 208         int i;
 209
 210         /*
 211          * refcount of workgroup is now freezed as 1,
 212          * therefore no need to worry about available decompression users.
 213          */
 214         for (i = 0; i < clusterpages; ++i) {
 215                 struct page *page = pcl->compressed_pages[i];
 216
 217                 if (!page)
 218                         continue;
 219
 220                 /* block other users from reclaiming or migrating the page */
 221                 if (!trylock_page(page))
 222                         return -EBUSY;
 223
 224                 if (page->mapping != mapping)
 225                         continue;
 226
 227                 /* barrier is implied in the following 'unlock_page' */
 228                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 229                 set_page_private(page, 0);
 230                 ClearPagePrivate(page);
 231
 232                 unlock_page(page);
 233                 put_page(page);
 234         }
 235         return 0;
 236 }
 237
 238 int erofs_try_to_free_cached_page(struct address_space *mapping,
 239                                   struct page *page)
 240 {
 241         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 242         const unsigned int clusterpages = BIT(pcl->clusterbits);
 243         int ret = 0;    /* 0 - busy */
 244
 245         if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
 246                 unsigned int i;
 247
 248                 for (i = 0; i < clusterpages; ++i) {
 249                         if (pcl->compressed_pages[i] == page) {
 250                                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 251                                 ret = 1;
 252                                 break;
 253                         }
 254                 }
 255                 erofs_workgroup_unfreeze(&pcl->obj, 1);
 256
 257                 if (ret) {
 258                         ClearPagePrivate(page);
 259                         put_page(page);
 260                 }
 261         }
 262         return ret;
 263 }
 264
 265 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 266 static inline bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
 267                                           struct page *page)
 268 {
 269         struct z_erofs_pcluster *const pcl = clt->pcl;
 270         const unsigned int clusterpages = BIT(pcl->clusterbits);
 271
 272         while (clt->compressedpages < pcl->compressed_pages + clusterpages) {
 273                 if (!cmpxchg(clt->compressedpages++, NULL, page))
 274                         return true;
 275         }
 276         return false;
 277 }
 278
 279 /* callers must be with collection lock held */
 280 static int z_erofs_attach_page(struct z_erofs_collector *clt,
 281                                struct page *page,
 282                                enum z_erofs_page_type type)
 283 {
 284         int ret;
 285         bool occupied;
 286
 287         /* give priority for inplaceio */
 288         if (clt->mode >= COLLECT_PRIMARY &&
 289             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 290             z_erofs_try_inplace_io(clt, page))
 291                 return 0;
 292
 293         ret = z_erofs_pagevec_enqueue(&clt->vector,
 294                                       page, type, &occupied);
 295         clt->cl->vcnt += (unsigned int)ret;
 296
 297         return ret ? 0 : -EAGAIN;
 298 }
 299
 300 static enum z_erofs_collectmode
 301 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
 302                       z_erofs_next_pcluster_t *owned_head)
 303 {
 304         /* let's claim these following types of pclusters */
 305 retry:
 306         if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
 307                 /* type 1, nil pcluster */
 308                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 309                             *owned_head) != Z_EROFS_PCLUSTER_NIL)
 310                         goto retry;
 311
 312                 *owned_head = &pcl->next;
 313                 /* lucky, I am the followee :) */
 314                 return COLLECT_PRIMARY_FOLLOWED;
 315         } else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
 316                 /*
 317                  * type 2, link to the end of a existing open chain,
 318                  * be careful that its submission itself is governed
 319                  * by the original owned chain.
 320                  */
 321                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 322                             *owned_head) != Z_EROFS_PCLUSTER_TAIL)
 323                         goto retry;
 324                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 325                 return COLLECT_PRIMARY_HOOKED;
 326         }
 327         return COLLECT_PRIMARY; /* :( better luck next time */
 328 }
 329
 330 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
 331                                      struct inode *inode,
 332                                      struct erofs_map_blocks *map)
 333 {
 334         struct z_erofs_pcluster *pcl = clt->pcl;
 335         struct z_erofs_collection *cl;
 336         unsigned int length;
 337
 338         /* to avoid unexpected loop formed by corrupted images */
 339         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 340                 DBG_BUGON(1);
 341                 return -EFSCORRUPTED;
 342         }
 343
 344         cl = z_erofs_primarycollection(pcl);
 345         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 346                 DBG_BUGON(1);
 347                 return -EFSCORRUPTED;
 348         }
 349
 350         length = READ_ONCE(pcl->length);
 351         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 352                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 353                         DBG_BUGON(1);
 354                         return -EFSCORRUPTED;
 355                 }
 356         } else {
 357                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 358
 359                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 360                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 361
 362                 while (llen > length &&
 363                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 364                         cpu_relax();
 365                         length = READ_ONCE(pcl->length);
 366                 }
 367         }
 368         mutex_lock(&cl->lock);
 369         /* used to check tail merging loop due to corrupted images */
 370         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 371                 clt->tailpcl = pcl;
 372         clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
 373         /* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
 374         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 375                 clt->tailpcl = NULL;
 376         clt->cl = cl;
 377         return 0;
 378 }
 379
 380 static int z_erofs_register_collection(struct z_erofs_collector *clt,
 381                                        struct inode *inode,
 382                                        struct erofs_map_blocks *map)
 383 {
 384         struct z_erofs_pcluster *pcl;
 385         struct z_erofs_collection *cl;
 386         struct erofs_workgroup *grp;
 387         int err;
 388
 389         /* no available workgroup, let's allocate one */
 390         pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS);
 391         if (!pcl)
 392                 return -ENOMEM;
 393
 394         atomic_set(&pcl->obj.refcount, 1);
 395         pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 396
 397         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 398                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 399                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 400
 401         if (map->m_flags & EROFS_MAP_ZIPPED)
 402                 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
 403         else
 404                 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
 405
 406         pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0];
 407         pcl->clusterbits -= PAGE_SHIFT;
 408
 409         /* new pclusters should be claimed as type 1, primary and followed */
 410         pcl->next = clt->owned_head;
 411         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 412
 413         cl = z_erofs_primarycollection(pcl);
 414
 415         /* must be cleaned before freeing to slab */
 416         DBG_BUGON(cl->nr_pages);
 417         DBG_BUGON(cl->vcnt);
 418
 419         cl->pageofs = map->m_la & ~PAGE_MASK;
 420
 421         /*
 422          * lock all primary followed works before visible to others
 423          * and mutex_trylock *never* fails for a new pcluster.
 424          */
 425         DBG_BUGON(!mutex_trylock(&cl->lock));
 426
 427         grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
 428         if (IS_ERR(grp)) {
 429                 err = PTR_ERR(grp);
 430                 goto err_out;
 431         }
 432
 433         if (grp != &pcl->obj) {
 434                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 435                 err = -EEXIST;
 436                 goto err_out;
 437         }
 438         /* used to check tail merging loop due to corrupted images */
 439         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 440                 clt->tailpcl = pcl;
 441         clt->owned_head = &pcl->next;
 442         clt->pcl = pcl;
 443         clt->cl = cl;
 444         return 0;
 445
 446 err_out:
 447         mutex_unlock(&cl->lock);
 448         kmem_cache_free(pcluster_cachep, pcl);
 449         return err;
 450 }
 451
 452 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
 453                                    struct inode *inode,
 454                                    struct erofs_map_blocks *map)
 455 {
 456         struct erofs_workgroup *grp;
 457         int ret;
 458
 459         DBG_BUGON(clt->cl);
 460
 461         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
 462         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
 463         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 464
 465         if (!PAGE_ALIGNED(map->m_pa)) {
 466                 DBG_BUGON(1);
 467                 return -EINVAL;
 468         }
 469
 470         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
 471         if (grp) {
 472                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 473         } else {
 474                 ret = z_erofs_register_collection(clt, inode, map);
 475
 476                 if (!ret)
 477                         goto out;
 478                 if (ret != -EEXIST)
 479                         return ret;
 480         }
 481
 482         ret = z_erofs_lookup_collection(clt, inode, map);
 483         if (ret) {
 484                 erofs_workgroup_put(&clt->pcl->obj);
 485                 return ret;
 486         }
 487
 488 out:
 489         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 490                                   clt->cl->pagevec, clt->cl->vcnt);
 491
 492         clt->compressedpages = clt->pcl->compressed_pages;
 493         if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
 494                 clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
 495         return 0;
 496 }
 497
 498 /*
 499  * keep in mind that no referenced pclusters will be freed
 500  * only after a RCU grace period.
 501  */
 502 static void z_erofs_rcu_callback(struct rcu_head *head)
 503 {
 504         struct z_erofs_collection *const cl =
 505                 container_of(head, struct z_erofs_collection, rcu);
 506
 507         kmem_cache_free(pcluster_cachep,
 508                         container_of(cl, struct z_erofs_pcluster,
 509                                      primary_collection));
 510 }
 511
 512 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 513 {
 514         struct z_erofs_pcluster *const pcl =
 515                 container_of(grp, struct z_erofs_pcluster, obj);
 516         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 517
 518         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 519 }
 520
 521 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 522 {
 523         struct z_erofs_pcluster *const pcl =
 524                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 525
 526         erofs_workgroup_put(&pcl->obj);
 527 }
 528
 529 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 530 {
 531         struct z_erofs_collection *cl = clt->cl;
 532
 533         if (!cl)
 534                 return false;
 535
 536         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 537         mutex_unlock(&cl->lock);
 538
 539         /*
 540          * if all pending pages are added, don't hold its reference
 541          * any longer if the pcluster isn't hosted by ourselves.
 542          */
 543         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 544                 z_erofs_collection_put(cl);
 545
 546         clt->cl = NULL;
 547         return true;
 548 }
 549
 550 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 551                                        unsigned int cachestrategy,
 552                                        erofs_off_t la)
 553 {
 554         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 555                 return false;
 556
 557         if (fe->backmost)
 558                 return true;
 559
 560         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 561                 la < fe->headoffset;
 562 }
 563
 564 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 565                                 struct page *page)
 566 {
 567         struct inode *const inode = fe->inode;
 568         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 569         struct erofs_map_blocks *const map = &fe->map;
 570         struct z_erofs_collector *const clt = &fe->clt;
 571         const loff_t offset = page_offset(page);
 572         bool tight = true;
 573
 574         enum z_erofs_cache_alloctype cache_strategy;
 575         enum z_erofs_page_type page_type;
 576         unsigned int cur, end, spiltted, index;
 577         int err = 0;
 578
 579         /* register locked file pages as online pages in pack */
 580         z_erofs_onlinepage_init(page);
 581
 582         spiltted = 0;
 583         end = PAGE_SIZE;
 584 repeat:
 585         cur = end - 1;
 586
 587         /* lucky, within the range of the current map_blocks */
 588         if (offset + cur >= map->m_la &&
 589             offset + cur < map->m_la + map->m_llen) {
 590                 /* didn't get a valid collection previously (very rare) */
 591                 if (!clt->cl)
 592                         goto restart_now;
 593                 goto hitted;
 594         }
 595
 596         /* go ahead the next map_blocks */
 597         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 598
 599         if (z_erofs_collector_end(clt))
 600                 fe->backmost = false;
 601
 602         map->m_la = offset + cur;
 603         map->m_llen = 0;
 604         err = z_erofs_map_blocks_iter(inode, map, 0);
 605         if (err)
 606                 goto err_out;
 607
 608 restart_now:
 609         if (!(map->m_flags & EROFS_MAP_MAPPED))
 610                 goto hitted;
 611
 612         err = z_erofs_collector_begin(clt, inode, map);
 613         if (err)
 614                 goto err_out;
 615
 616         /* preload all compressed pages (maybe downgrade role if necessary) */
 617         if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la))
 618                 cache_strategy = DELAYEDALLOC;
 619         else
 620                 cache_strategy = DONTALLOC;
 621
 622         preload_compressed_pages(clt, MNGD_MAPPING(sbi), cache_strategy);
 623
 624 hitted:
 625         /*
 626          * Ensure the current partial page belongs to this submit chain rather
 627          * than other concurrent submit chains or the noio(bypass) chain since
 628          * those chains are handled asynchronously thus the page cannot be used
 629          * for inplace I/O or pagevec (should be processed in strict order.)
 630          */
 631         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 632                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 633
 634         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 635         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 636                 zero_user_segment(page, cur, end);
 637                 goto next_part;
 638         }
 639
 640         /* let's derive page type */
 641         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 642                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 643                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 644                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 645
 646         if (cur)
 647                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 648
 649 retry:
 650         err = z_erofs_attach_page(clt, page, page_type);
 651         /* should allocate an additional staging page for pagevec */
 652         if (err == -EAGAIN) {
 653                 struct page *const newpage =
 654                                 alloc_page(GFP_NOFS | __GFP_NOFAIL);
 655
 656                 newpage->mapping = Z_EROFS_MAPPING_STAGING;
 657                 err = z_erofs_attach_page(clt, newpage,
 658                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 659                 if (!err)
 660                         goto retry;
 661         }
 662
 663         if (err)
 664                 goto err_out;
 665
 666         index = page->index - (map->m_la >> PAGE_SHIFT);
 667
 668         z_erofs_onlinepage_fixup(page, index, true);
 669
 670         /* bump up the number of spiltted parts of a page */
 671         ++spiltted;
 672         /* also update nr_pages */
 673         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 674 next_part:
 675         /* can be used for verification */
 676         map->m_llen = offset + cur - map->m_la;
 677
 678         end = cur;
 679         if (end > 0)
 680                 goto repeat;
 681
 682 out:
 683         z_erofs_onlinepage_endio(page);
 684
 685         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 686                   __func__, page, spiltted, map->m_llen);
 687         return err;
 688
 689         /* if some error occurred while processing this page */
 690 err_out:
 691         SetPageError(page);
 692         goto out;
 693 }
 694
 695 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 696                                        bool sync, int bios)
 697 {
 698         /* wake up the caller thread for sync decompression */
 699         if (sync) {
 700                 unsigned long flags;
 701
 702                 spin_lock_irqsave(&io->u.wait.lock, flags);
 703                 if (!atomic_add_return(bios, &io->pending_bios))
 704                         wake_up_locked(&io->u.wait);
 705                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 706                 return;
 707         }
 708
 709         if (!atomic_add_return(bios, &io->pending_bios))
 710                 queue_work(z_erofs_workqueue, &io->u.work);
 711 }
 712
 713 static void z_erofs_decompressqueue_endio(struct bio *bio)
 714 {
 715         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
 716         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
 717         blk_status_t err = bio->bi_status;
 718         struct bio_vec *bvec;
 719         struct bvec_iter_all iter_all;
 720
 721         bio_for_each_segment_all(bvec, bio, iter_all) {
 722                 struct page *page = bvec->bv_page;
 723
 724                 DBG_BUGON(PageUptodate(page));
 725                 DBG_BUGON(!page->mapping);
 726
 727                 if (err)
 728                         SetPageError(page);
 729
 730                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
 731                         if (!err)
 732                                 SetPageUptodate(page);
 733                         unlock_page(page);
 734                 }
 735         }
 736         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
 737         bio_put(bio);
 738 }
 739
 740 static int z_erofs_decompress_pcluster(struct super_block *sb,
 741                                        struct z_erofs_pcluster *pcl,
 742                                        struct list_head *pagepool)
 743 {
 744         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 745         const unsigned int clusterpages = BIT(pcl->clusterbits);
 746         struct z_erofs_pagevec_ctor ctor;
 747         unsigned int i, outputsize, llen, nr_pages;
 748         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 749         struct page **pages, **compressed_pages, *page;
 750
 751         enum z_erofs_page_type page_type;
 752         bool overlapped, partial;
 753         struct z_erofs_collection *cl;
 754         int err;
 755
 756         might_sleep();
 757         cl = z_erofs_primarycollection(pcl);
 758         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 759
 760         mutex_lock(&cl->lock);
 761         nr_pages = cl->nr_pages;
 762
 763         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 764                 pages = pages_onstack;
 765         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 766                    mutex_trylock(&z_pagemap_global_lock)) {
 767                 pages = z_pagemap_global;
 768         } else {
 769                 gfp_t gfp_flags = GFP_KERNEL;
 770
 771                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 772                         gfp_flags |= __GFP_NOFAIL;
 773
 774                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 775                                        gfp_flags);
 776
 777                 /* fallback to global pagemap for the lowmem scenario */
 778                 if (!pages) {
 779                         mutex_lock(&z_pagemap_global_lock);
 780                         pages = z_pagemap_global;
 781                 }
 782         }
 783
 784         for (i = 0; i < nr_pages; ++i)
 785                 pages[i] = NULL;
 786
 787         err = 0;
 788         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 789                                   cl->pagevec, 0);
 790
 791         for (i = 0; i < cl->vcnt; ++i) {
 792                 unsigned int pagenr;
 793
 794                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 795
 796                 /* all pages in pagevec ought to be valid */
 797                 DBG_BUGON(!page);
 798                 DBG_BUGON(!page->mapping);
 799
 800                 if (z_erofs_put_stagingpage(pagepool, page))
 801                         continue;
 802
 803                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 804                         pagenr = 0;
 805                 else
 806                         pagenr = z_erofs_onlinepage_index(page);
 807
 808                 DBG_BUGON(pagenr >= nr_pages);
 809
 810                 /*
 811                  * currently EROFS doesn't support multiref(dedup),
 812                  * so here erroring out one multiref page.
 813                  */
 814                 if (pages[pagenr]) {
 815                         DBG_BUGON(1);
 816                         SetPageError(pages[pagenr]);
 817                         z_erofs_onlinepage_endio(pages[pagenr]);
 818                         err = -EFSCORRUPTED;
 819                 }
 820                 pages[pagenr] = page;
 821         }
 822         z_erofs_pagevec_ctor_exit(&ctor, true);
 823
 824         overlapped = false;
 825         compressed_pages = pcl->compressed_pages;
 826
 827         for (i = 0; i < clusterpages; ++i) {
 828                 unsigned int pagenr;
 829
 830                 page = compressed_pages[i];
 831
 832                 /* all compressed pages ought to be valid */
 833                 DBG_BUGON(!page);
 834                 DBG_BUGON(!page->mapping);
 835
 836                 if (!z_erofs_page_is_staging(page)) {
 837                         if (erofs_page_is_managed(sbi, page)) {
 838                                 if (!PageUptodate(page))
 839                                         err = -EIO;
 840                                 continue;
 841                         }
 842
 843                         /*
 844                          * only if non-head page can be selected
 845                          * for inplace decompression
 846                          */
 847                         pagenr = z_erofs_onlinepage_index(page);
 848
 849                         DBG_BUGON(pagenr >= nr_pages);
 850                         if (pages[pagenr]) {
 851                                 DBG_BUGON(1);
 852                                 SetPageError(pages[pagenr]);
 853                                 z_erofs_onlinepage_endio(pages[pagenr]);
 854                                 err = -EFSCORRUPTED;
 855                         }
 856                         pages[pagenr] = page;
 857
 858                         overlapped = true;
 859                 }
 860
 861                 /* PG_error needs checking for inplaced and staging pages */
 862                 if (PageError(page)) {
 863                         DBG_BUGON(PageUptodate(page));
 864                         err = -EIO;
 865                 }
 866         }
 867
 868         if (err)
 869                 goto out;
 870
 871         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 872         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 873                 outputsize = llen;
 874                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 875         } else {
 876                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 877                 partial = true;
 878         }
 879
 880         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 881                                         .sb = sb,
 882                                         .in = compressed_pages,
 883                                         .out = pages,
 884                                         .pageofs_out = cl->pageofs,
 885                                         .inputsize = PAGE_SIZE,
 886                                         .outputsize = outputsize,
 887                                         .alg = pcl->algorithmformat,
 888                                         .inplace_io = overlapped,
 889                                         .partial_decoding = partial
 890                                  }, pagepool);
 891
 892 out:
 893         /* must handle all compressed pages before endding pages */
 894         for (i = 0; i < clusterpages; ++i) {
 895                 page = compressed_pages[i];
 896
 897                 if (erofs_page_is_managed(sbi, page))
 898                         continue;
 899
 900                 /* recycle all individual staging pages */
 901                 (void)z_erofs_put_stagingpage(pagepool, page);
 902
 903                 WRITE_ONCE(compressed_pages[i], NULL);
 904         }
 905
 906         for (i = 0; i < nr_pages; ++i) {
 907                 page = pages[i];
 908                 if (!page)
 909                         continue;
 910
 911                 DBG_BUGON(!page->mapping);
 912
 913                 /* recycle all individual staging pages */
 914                 if (z_erofs_put_stagingpage(pagepool, page))
 915                         continue;
 916
 917                 if (err < 0)
 918                         SetPageError(page);
 919
 920                 z_erofs_onlinepage_endio(page);
 921         }
 922
 923         if (pages == z_pagemap_global)
 924                 mutex_unlock(&z_pagemap_global_lock);
 925         else if (pages != pages_onstack)
 926                 kvfree(pages);
 927
 928         cl->nr_pages = 0;
 929         cl->vcnt = 0;
 930
 931         /* all cl locks MUST be taken before the following line */
 932         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
 933
 934         /* all cl locks SHOULD be released right now */
 935         mutex_unlock(&cl->lock);
 936
 937         z_erofs_collection_put(cl);
 938         return err;
 939 }
 940
 941 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
 942                                      struct list_head *pagepool)
 943 {
 944         z_erofs_next_pcluster_t owned = io->head;
 945
 946         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
 947                 struct z_erofs_pcluster *pcl;
 948
 949                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
 950                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
 951
 952                 /* no possible that 'owned' equals NULL */
 953                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
 954
 955                 pcl = container_of(owned, struct z_erofs_pcluster, next);
 956                 owned = READ_ONCE(pcl->next);
 957
 958                 z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
 959         }
 960 }
 961
 962 static void z_erofs_decompressqueue_work(struct work_struct *work)
 963 {
 964         struct z_erofs_decompressqueue *bgq =
 965                 container_of(work, struct z_erofs_decompressqueue, u.work);
 966         LIST_HEAD(pagepool);
 967
 968         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 969         z_erofs_decompress_queue(bgq, &pagepool);
 970
 971         put_pages_list(&pagepool);
 972         kvfree(bgq);
 973 }
 974
 975 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
 976                                                unsigned int nr,
 977                                                struct list_head *pagepool,
 978                                                struct address_space *mc,
 979                                                gfp_t gfp)
 980 {
 981         const pgoff_t index = pcl->obj.index;
 982         bool tocache = false;
 983
 984         struct address_space *mapping;
 985         struct page *oldpage, *page;
 986
 987         compressed_page_t t;
 988         int justfound;
 989
 990 repeat:
 991         page = READ_ONCE(pcl->compressed_pages[nr]);
 992         oldpage = page;
 993
 994         if (!page)
 995                 goto out_allocpage;
 996
 997         /*
 998          * the cached page has not been allocated and
 999          * an placeholder is out there, prepare it now.
1000          */
1001         if (page == PAGE_UNALLOCATED) {
1002                 tocache = true;
1003                 goto out_allocpage;
1004         }
1005
1006         /* process the target tagged pointer */
1007         t = tagptr_init(compressed_page_t, page);
1008         justfound = tagptr_unfold_tags(t);
1009         page = tagptr_unfold_ptr(t);
1010
1011         mapping = READ_ONCE(page->mapping);
1012
1013         /*
1014          * unmanaged (file) pages are all locked solidly,
1015          * therefore it is impossible for `mapping' to be NULL.
1016          */
1017         if (mapping && mapping != mc)
1018                 /* ought to be unmanaged pages */
1019                 goto out;
1020
1021         lock_page(page);
1022
1023         /* only true if page reclaim goes wrong, should never happen */
1024         DBG_BUGON(justfound && PagePrivate(page));
1025
1026         /* the page is still in manage cache */
1027         if (page->mapping == mc) {
1028                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1029
1030                 ClearPageError(page);
1031                 if (!PagePrivate(page)) {
1032                         /*
1033                          * impossible to be !PagePrivate(page) for
1034                          * the current restriction as well if
1035                          * the page is already in compressed_pages[].
1036                          */
1037                         DBG_BUGON(!justfound);
1038
1039                         justfound = 0;
1040                         set_page_private(page, (unsigned long)pcl);
1041                         SetPagePrivate(page);
1042                 }
1043
1044                 /* no need to submit io if it is already up-to-date */
1045                 if (PageUptodate(page)) {
1046                         unlock_page(page);
1047                         page = NULL;
1048                 }
1049                 goto out;
1050         }
1051
1052         /*
1053          * the managed page has been truncated, it's unsafe to
1054          * reuse this one, let's allocate a new cache-managed page.
1055          */
1056         DBG_BUGON(page->mapping);
1057         DBG_BUGON(!justfound);
1058
1059         tocache = true;
1060         unlock_page(page);
1061         put_page(page);
1062 out_allocpage:
1063         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1064         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1065                 /* non-LRU / non-movable temporary page is needed */
1066                 page->mapping = Z_EROFS_MAPPING_STAGING;
1067                 tocache = false;
1068         }
1069
1070         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1071                 if (tocache) {
1072                         /* since it added to managed cache successfully */
1073                         unlock_page(page);
1074                         put_page(page);
1075                 } else {
1076                         list_add(&page->lru, pagepool);
1077                 }
1078                 cond_resched();
1079                 goto repeat;
1080         }
1081         set_page_private(page, (unsigned long)pcl);
1082         SetPagePrivate(page);
1083 out:    /* the only exit (for tracing and debugging) */
1084         return page;
1085 }
1086
1087 static struct z_erofs_decompressqueue *
1088 jobqueue_init(struct super_block *sb,
1089               struct z_erofs_decompressqueue *fgq, bool *fg)
1090 {
1091         struct z_erofs_decompressqueue *q;
1092
1093         if (fg && !*fg) {
1094                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1095                 if (!q) {
1096                         *fg = true;
1097                         goto fg_out;
1098                 }
1099                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1100         } else {
1101 fg_out:
1102                 q = fgq;
1103                 init_waitqueue_head(&fgq->u.wait);
1104                 atomic_set(&fgq->pending_bios, 0);
1105         }
1106         q->sb = sb;
1107         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1108         return q;
1109 }
1110
1111 /* define decompression jobqueue types */
1112 enum {
1113         JQ_BYPASS,
1114         JQ_SUBMIT,
1115         NR_JOBQUEUES,
1116 };
1117
1118 static void *jobqueueset_init(struct super_block *sb,
1119                               struct z_erofs_decompressqueue *q[],
1120                               struct z_erofs_decompressqueue *fgq, bool *fg)
1121 {
1122         /*
1123          * if managed cache is enabled, bypass jobqueue is needed,
1124          * no need to read from device for all pclusters in this queue.
1125          */
1126         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1127         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1128
1129         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1130 }
1131
1132 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1133                                     z_erofs_next_pcluster_t qtail[],
1134                                     z_erofs_next_pcluster_t owned_head)
1135 {
1136         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1137         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1138
1139         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1140         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1141                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1142
1143         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1144
1145         WRITE_ONCE(*submit_qtail, owned_head);
1146         WRITE_ONCE(*bypass_qtail, &pcl->next);
1147
1148         qtail[JQ_BYPASS] = &pcl->next;
1149 }
1150
1151 static void z_erofs_submit_queue(struct super_block *sb,
1152                                  struct z_erofs_decompress_frontend *f,
1153                                  struct list_head *pagepool,
1154                                  struct z_erofs_decompressqueue *fgq,
1155                                  bool *force_fg)
1156 {
1157         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1158         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1159         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1160         void *bi_private;
1161         z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1162         /* since bio will be NULL, no need to initialize last_index */
1163         pgoff_t last_index;
1164         unsigned int nr_bios = 0;
1165         struct bio *bio = NULL;
1166
1167         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1168         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1169         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1170
1171         /* by default, all need io submission */
1172         q[JQ_SUBMIT]->head = owned_head;
1173
1174         do {
1175                 struct z_erofs_pcluster *pcl;
1176                 pgoff_t cur, end;
1177                 unsigned int i = 0;
1178                 bool bypass = true;
1179
1180                 /* no possible 'owned_head' equals the following */
1181                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1182                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1183
1184                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1185
1186                 cur = pcl->obj.index;
1187                 end = cur + BIT(pcl->clusterbits);
1188
1189                 /* close the main owned chain at first */
1190                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1191                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1192
1193                 do {
1194                         struct page *page;
1195
1196                         page = pickup_page_for_submission(pcl, i++, pagepool,
1197                                                           MNGD_MAPPING(sbi),
1198                                                           GFP_NOFS);
1199                         if (!page)
1200                                 continue;
1201
1202                         if (bio && cur != last_index + 1) {
1203 submit_bio_retry:
1204                                 submit_bio(bio);
1205                                 bio = NULL;
1206                         }
1207
1208                         if (!bio) {
1209                                 bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1210
1211                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1212                                 bio_set_dev(bio, sb->s_bdev);
1213                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1214                                         LOG_SECTORS_PER_BLOCK;
1215                                 bio->bi_private = bi_private;
1216                                 bio->bi_opf = REQ_OP_READ;
1217                                 if (f->readahead)
1218                                         bio->bi_opf |= REQ_RAHEAD;
1219                                 ++nr_bios;
1220                         }
1221
1222                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1223                                 goto submit_bio_retry;
1224
1225                         last_index = cur;
1226                         bypass = false;
1227                 } while (++cur < end);
1228
1229                 if (!bypass)
1230                         qtail[JQ_SUBMIT] = &pcl->next;
1231                 else
1232                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1233         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1234
1235         if (bio)
1236                 submit_bio(bio);
1237
1238         /*
1239          * although background is preferred, no one is pending for submission.
1240          * don't issue workqueue for decompression but drop it directly instead.
1241          */
1242         if (!*force_fg && !nr_bios) {
1243                 kvfree(q[JQ_SUBMIT]);
1244                 return;
1245         }
1246         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1247 }
1248
1249 static void z_erofs_runqueue(struct super_block *sb,
1250                              struct z_erofs_decompress_frontend *f,
1251                              struct list_head *pagepool, bool force_fg)
1252 {
1253         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1254
1255         if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1256                 return;
1257         z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1258
1259         /* handle bypass queue (no i/o pclusters) immediately */
1260         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1261
1262         if (!force_fg)
1263                 return;
1264
1265         /* wait until all bios are completed */
1266         io_wait_event(io[JQ_SUBMIT].u.wait,
1267                       !atomic_read(&io[JQ_SUBMIT].pending_bios));
1268
1269         /* handle synchronous decompress queue in the caller context */
1270         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1271 }
1272
1273 static int z_erofs_readpage(struct file *file, struct page *page)
1274 {
1275         struct inode *const inode = page->mapping->host;
1276         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1277         int err;
1278         LIST_HEAD(pagepool);
1279
1280         trace_erofs_readpage(page, false);
1281
1282         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1283
1284         err = z_erofs_do_read_page(&f, page);
1285         (void)z_erofs_collector_end(&f.clt);
1286
1287         /* if some compressed cluster ready, need submit them anyway */
1288         z_erofs_runqueue(inode->i_sb, &f, &pagepool, true);
1289
1290         if (err)
1291                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1292
1293         if (f.map.mpage)
1294                 put_page(f.map.mpage);
1295
1296         /* clean up the remaining free pages */
1297         put_pages_list(&pagepool);
1298         return err;
1299 }
1300
1301 static void z_erofs_readahead(struct readahead_control *rac)
1302 {
1303         struct inode *const inode = rac->mapping->host;
1304         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1305
1306         unsigned int nr_pages = readahead_count(rac);
1307         bool sync = (nr_pages <= sbi->ctx.max_sync_decompress_pages);
1308         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1309         struct page *page, *head = NULL;
1310         LIST_HEAD(pagepool);
1311
1312         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1313
1314         f.readahead = true;
1315         f.headoffset = readahead_pos(rac);
1316
1317         while ((page = readahead_page(rac))) {
1318                 prefetchw(&page->flags);
1319
1320                 /*
1321                  * A pure asynchronous readahead is indicated if
1322                  * a PG_readahead marked page is hitted at first.
1323                  * Let's also do asynchronous decompression for this case.
1324                  */
1325                 sync &= !(PageReadahead(page) && !head);
1326
1327                 set_page_private(page, (unsigned long)head);
1328                 head = page;
1329         }
1330
1331         while (head) {
1332                 struct page *page = head;
1333                 int err;
1334
1335                 /* traversal in reverse order */
1336                 head = (void *)page_private(page);
1337
1338                 err = z_erofs_do_read_page(&f, page);
1339                 if (err)
1340                         erofs_err(inode->i_sb,
1341                                   "readahead error at page %lu @ nid %llu",
1342                                   page->index, EROFS_I(inode)->nid);
1343                 put_page(page);
1344         }
1345
1346         (void)z_erofs_collector_end(&f.clt);
1347
1348         z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync);
1349
1350         if (f.map.mpage)
1351                 put_page(f.map.mpage);
1352
1353         /* clean up the remaining free pages */
1354         put_pages_list(&pagepool);
1355 }
1356
1357 const struct address_space_operations z_erofs_aops = {
1358         .readpage = z_erofs_readpage,
1359         .readahead = z_erofs_readahead,
1360 };
1361