fs/erofs/zdata.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             http://www.huawei.com/
   5  * Created by Gao Xiang <gaoxiang25@huawei.com>
   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         const unsigned int flags = WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE;
  47
  48         /*
  49          * no need to spawn too many threads, limiting threads could minimum
  50          * scheduling overhead, perhaps per-CPU threads should be better?
  51          */
  52         z_erofs_workqueue = alloc_workqueue("erofs_unzipd", flags,
  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 static void z_erofs_pcluster_init_always(struct z_erofs_pcluster *pcl)
  71 {
  72         struct z_erofs_collection *cl = z_erofs_primarycollection(pcl);
  73
  74         atomic_set(&pcl->obj.refcount, 1);
  75
  76         DBG_BUGON(cl->nr_pages);
  77         DBG_BUGON(cl->vcnt);
  78 }
  79
  80 int __init z_erofs_init_zip_subsystem(void)
  81 {
  82         pcluster_cachep = kmem_cache_create("erofs_compress",
  83                                             Z_EROFS_WORKGROUP_SIZE, 0,
  84                                             SLAB_RECLAIM_ACCOUNT,
  85                                             z_erofs_pcluster_init_once);
  86         if (pcluster_cachep) {
  87                 if (!z_erofs_init_workqueue())
  88                         return 0;
  89
  90                 kmem_cache_destroy(pcluster_cachep);
  91         }
  92         return -ENOMEM;
  93 }
  94
  95 enum z_erofs_collectmode {
  96         COLLECT_SECONDARY,
  97         COLLECT_PRIMARY,
  98         /*
  99          * The current collection was the tail of an exist chain, in addition
 100          * that the previous processed chained collections are all decided to
 101          * be hooked up to it.
 102          * A new chain will be created for the remaining collections which are
 103          * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
 104          * the next collection cannot reuse the whole page safely in
 105          * the following scenario:
 106          *  ________________________________________________________________
 107          * |      tail (partial) page     |       head (partial) page       |
 108          * |   (belongs to the next cl)   |   (belongs to the current cl)   |
 109          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 110          */
 111         COLLECT_PRIMARY_HOOKED,
 112         COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
 113         /*
 114          * The current collection has been linked with the owned chain, and
 115          * could also be linked with the remaining collections, which means
 116          * if the processing page is the tail page of the collection, thus
 117          * the current collection can safely use the whole page (since
 118          * the previous collection is under control) for in-place I/O, as
 119          * illustrated below:
 120          *  ________________________________________________________________
 121          * |  tail (partial) page |          head (partial) page           |
 122          * |  (of the current cl) |      (of the previous collection)      |
 123          * |  PRIMARY_FOLLOWED or |                                        |
 124          * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
 125          *
 126          * [  (*) the above page can be used as inplace I/O.               ]
 127          */
 128         COLLECT_PRIMARY_FOLLOWED,
 129 };
 130
 131 struct z_erofs_collector {
 132         struct z_erofs_pagevec_ctor vector;
 133
 134         struct z_erofs_pcluster *pcl, *tailpcl;
 135         struct z_erofs_collection *cl;
 136         struct page **compressedpages;
 137         z_erofs_next_pcluster_t owned_head;
 138
 139         enum z_erofs_collectmode mode;
 140 };
 141
 142 struct z_erofs_decompress_frontend {
 143         struct inode *const inode;
 144
 145         struct z_erofs_collector clt;
 146         struct erofs_map_blocks map;
 147
 148         /* used for applying cache strategy on the fly */
 149         bool backmost;
 150         erofs_off_t headoffset;
 151 };
 152
 153 #define COLLECTOR_INIT() { \
 154         .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 155         .mode = COLLECT_PRIMARY_FOLLOWED }
 156
 157 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 158         .inode = __i, .clt = COLLECTOR_INIT(), \
 159         .backmost = true, }
 160
 161 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
 162 static DEFINE_MUTEX(z_pagemap_global_lock);
 163
 164 static void preload_compressed_pages(struct z_erofs_collector *clt,
 165                                      struct address_space *mc,
 166                                      enum z_erofs_cache_alloctype type,
 167                                      struct list_head *pagepool)
 168 {
 169         const struct z_erofs_pcluster *pcl = clt->pcl;
 170         const unsigned int clusterpages = BIT(pcl->clusterbits);
 171         struct page **pages = clt->compressedpages;
 172         pgoff_t index = pcl->obj.index + (pages - pcl->compressed_pages);
 173         bool standalone = true;
 174
 175         if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
 176                 return;
 177
 178         for (; pages < pcl->compressed_pages + clusterpages; ++pages) {
 179                 struct page *page;
 180                 compressed_page_t t;
 181
 182                 /* the compressed page was loaded before */
 183                 if (READ_ONCE(*pages))
 184                         continue;
 185
 186                 page = find_get_page(mc, index);
 187
 188                 if (page) {
 189                         t = tag_compressed_page_justfound(page);
 190                 } else if (type == DELAYEDALLOC) {
 191                         t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
 192                 } else {        /* DONTALLOC */
 193                         if (standalone)
 194                                 clt->compressedpages = pages;
 195                         standalone = false;
 196                         continue;
 197                 }
 198
 199                 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
 200                         continue;
 201
 202                 if (page)
 203                         put_page(page);
 204         }
 205
 206         if (standalone)         /* downgrade to PRIMARY_FOLLOWED_NOINPLACE */
 207                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 208 }
 209
 210 /* called by erofs_shrinker to get rid of all compressed_pages */
 211 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 212                                        struct erofs_workgroup *grp)
 213 {
 214         struct z_erofs_pcluster *const pcl =
 215                 container_of(grp, struct z_erofs_pcluster, obj);
 216         struct address_space *const mapping = MNGD_MAPPING(sbi);
 217         const unsigned int clusterpages = BIT(pcl->clusterbits);
 218         int i;
 219
 220         /*
 221          * refcount of workgroup is now freezed as 1,
 222          * therefore no need to worry about available decompression users.
 223          */
 224         for (i = 0; i < clusterpages; ++i) {
 225                 struct page *page = pcl->compressed_pages[i];
 226
 227                 if (!page)
 228                         continue;
 229
 230                 /* block other users from reclaiming or migrating the page */
 231                 if (!trylock_page(page))
 232                         return -EBUSY;
 233
 234                 if (page->mapping != mapping)
 235                         continue;
 236
 237                 /* barrier is implied in the following 'unlock_page' */
 238                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 239                 set_page_private(page, 0);
 240                 ClearPagePrivate(page);
 241
 242                 unlock_page(page);
 243                 put_page(page);
 244         }
 245         return 0;
 246 }
 247
 248 int erofs_try_to_free_cached_page(struct address_space *mapping,
 249                                   struct page *page)
 250 {
 251         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 252         const unsigned int clusterpages = BIT(pcl->clusterbits);
 253         int ret = 0;    /* 0 - busy */
 254
 255         if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
 256                 unsigned int i;
 257
 258                 for (i = 0; i < clusterpages; ++i) {
 259                         if (pcl->compressed_pages[i] == page) {
 260                                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 261                                 ret = 1;
 262                                 break;
 263                         }
 264                 }
 265                 erofs_workgroup_unfreeze(&pcl->obj, 1);
 266
 267                 if (ret) {
 268                         ClearPagePrivate(page);
 269                         put_page(page);
 270                 }
 271         }
 272         return ret;
 273 }
 274
 275 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 276 static inline bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
 277                                           struct page *page)
 278 {
 279         struct z_erofs_pcluster *const pcl = clt->pcl;
 280         const unsigned int clusterpages = BIT(pcl->clusterbits);
 281
 282         while (clt->compressedpages < pcl->compressed_pages + clusterpages) {
 283                 if (!cmpxchg(clt->compressedpages++, NULL, page))
 284                         return true;
 285         }
 286         return false;
 287 }
 288
 289 /* callers must be with collection lock held */
 290 static int z_erofs_attach_page(struct z_erofs_collector *clt,
 291                                struct page *page,
 292                                enum z_erofs_page_type type)
 293 {
 294         int ret;
 295         bool occupied;
 296
 297         /* give priority for inplaceio */
 298         if (clt->mode >= COLLECT_PRIMARY &&
 299             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 300             z_erofs_try_inplace_io(clt, page))
 301                 return 0;
 302
 303         ret = z_erofs_pagevec_enqueue(&clt->vector,
 304                                       page, type, &occupied);
 305         clt->cl->vcnt += (unsigned int)ret;
 306
 307         return ret ? 0 : -EAGAIN;
 308 }
 309
 310 static enum z_erofs_collectmode
 311 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
 312                       z_erofs_next_pcluster_t *owned_head)
 313 {
 314         /* let's claim these following types of pclusters */
 315 retry:
 316         if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
 317                 /* type 1, nil pcluster */
 318                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 319                             *owned_head) != Z_EROFS_PCLUSTER_NIL)
 320                         goto retry;
 321
 322                 *owned_head = &pcl->next;
 323                 /* lucky, I am the followee :) */
 324                 return COLLECT_PRIMARY_FOLLOWED;
 325         } else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
 326                 /*
 327                  * type 2, link to the end of a existing open chain,
 328                  * be careful that its submission itself is governed
 329                  * by the original owned chain.
 330                  */
 331                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 332                             *owned_head) != Z_EROFS_PCLUSTER_TAIL)
 333                         goto retry;
 334                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 335                 return COLLECT_PRIMARY_HOOKED;
 336         }
 337         return COLLECT_PRIMARY; /* :( better luck next time */
 338 }
 339
 340 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
 341                                      struct inode *inode,
 342                                      struct erofs_map_blocks *map)
 343 {
 344         struct erofs_workgroup *grp;
 345         struct z_erofs_pcluster *pcl;
 346         struct z_erofs_collection *cl;
 347         unsigned int length;
 348         bool tag;
 349
 350         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT, &tag);
 351         if (!grp)
 352                 return -ENOENT;
 353
 354         pcl = container_of(grp, struct z_erofs_pcluster, obj);
 355         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 356                 DBG_BUGON(1);
 357                 erofs_workgroup_put(grp);
 358                 return -EFSCORRUPTED;
 359         }
 360
 361         cl = z_erofs_primarycollection(pcl);
 362         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 363                 DBG_BUGON(1);
 364                 erofs_workgroup_put(grp);
 365                 return -EFSCORRUPTED;
 366         }
 367
 368         length = READ_ONCE(pcl->length);
 369         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 370                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 371                         DBG_BUGON(1);
 372                         erofs_workgroup_put(grp);
 373                         return -EFSCORRUPTED;
 374                 }
 375         } else {
 376                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 377
 378                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 379                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 380
 381                 while (llen > length &&
 382                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 383                         cpu_relax();
 384                         length = READ_ONCE(pcl->length);
 385                 }
 386         }
 387         mutex_lock(&cl->lock);
 388         /* used to check tail merging loop due to corrupted images */
 389         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 390                 clt->tailpcl = pcl;
 391         clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
 392         /* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
 393         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 394                 clt->tailpcl = NULL;
 395         clt->pcl = pcl;
 396         clt->cl = cl;
 397         return 0;
 398 }
 399
 400 static int z_erofs_register_collection(struct z_erofs_collector *clt,
 401                                        struct inode *inode,
 402                                        struct erofs_map_blocks *map)
 403 {
 404         struct z_erofs_pcluster *pcl;
 405         struct z_erofs_collection *cl;
 406         int err;
 407
 408         /* no available workgroup, let's allocate one */
 409         pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS);
 410         if (!pcl)
 411                 return -ENOMEM;
 412
 413         z_erofs_pcluster_init_always(pcl);
 414         pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 415
 416         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 417                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 418                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 419
 420         if (map->m_flags & EROFS_MAP_ZIPPED)
 421                 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
 422         else
 423                 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
 424
 425         pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0];
 426         pcl->clusterbits -= PAGE_SHIFT;
 427
 428         /* new pclusters should be claimed as type 1, primary and followed */
 429         pcl->next = clt->owned_head;
 430         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 431
 432         cl = z_erofs_primarycollection(pcl);
 433         cl->pageofs = map->m_la & ~PAGE_MASK;
 434
 435         /*
 436          * lock all primary followed works before visible to others
 437          * and mutex_trylock *never* fails for a new pcluster.
 438          */
 439         mutex_trylock(&cl->lock);
 440
 441         err = erofs_register_workgroup(inode->i_sb, &pcl->obj, 0);
 442         if (err) {
 443                 mutex_unlock(&cl->lock);
 444                 kmem_cache_free(pcluster_cachep, pcl);
 445                 return -EAGAIN;
 446         }
 447         /* used to check tail merging loop due to corrupted images */
 448         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 449                 clt->tailpcl = pcl;
 450         clt->owned_head = &pcl->next;
 451         clt->pcl = pcl;
 452         clt->cl = cl;
 453         return 0;
 454 }
 455
 456 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
 457                                    struct inode *inode,
 458                                    struct erofs_map_blocks *map)
 459 {
 460         int ret;
 461
 462         DBG_BUGON(clt->cl);
 463
 464         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
 465         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
 466         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 467
 468         if (!PAGE_ALIGNED(map->m_pa)) {
 469                 DBG_BUGON(1);
 470                 return -EINVAL;
 471         }
 472
 473 repeat:
 474         ret = z_erofs_lookup_collection(clt, inode, map);
 475         if (ret == -ENOENT) {
 476                 ret = z_erofs_register_collection(clt, inode, map);
 477
 478                 /* someone registered at the same time, give another try */
 479                 if (ret == -EAGAIN) {
 480                         cond_resched();
 481                         goto repeat;
 482                 }
 483         }
 484
 485         if (ret)
 486                 return ret;
 487
 488         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 489                                   clt->cl->pagevec, clt->cl->vcnt);
 490
 491         clt->compressedpages = clt->pcl->compressed_pages;
 492         if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
 493                 clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
 494         return 0;
 495 }
 496
 497 /*
 498  * keep in mind that no referenced pclusters will be freed
 499  * only after a RCU grace period.
 500  */
 501 static void z_erofs_rcu_callback(struct rcu_head *head)
 502 {
 503         struct z_erofs_collection *const cl =
 504                 container_of(head, struct z_erofs_collection, rcu);
 505
 506         kmem_cache_free(pcluster_cachep,
 507                         container_of(cl, struct z_erofs_pcluster,
 508                                      primary_collection));
 509 }
 510
 511 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 512 {
 513         struct z_erofs_pcluster *const pcl =
 514                 container_of(grp, struct z_erofs_pcluster, obj);
 515         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 516
 517         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 518 }
 519
 520 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 521 {
 522         struct z_erofs_pcluster *const pcl =
 523                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 524
 525         erofs_workgroup_put(&pcl->obj);
 526 }
 527
 528 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 529 {
 530         struct z_erofs_collection *cl = clt->cl;
 531
 532         if (!cl)
 533                 return false;
 534
 535         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 536         mutex_unlock(&cl->lock);
 537
 538         /*
 539          * if all pending pages are added, don't hold its reference
 540          * any longer if the pcluster isn't hosted by ourselves.
 541          */
 542         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 543                 z_erofs_collection_put(cl);
 544
 545         clt->cl = NULL;
 546         return true;
 547 }
 548
 549 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 550                                        unsigned int cachestrategy,
 551                                        erofs_off_t la)
 552 {
 553         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 554                 return false;
 555
 556         if (fe->backmost)
 557                 return true;
 558
 559         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 560                 la < fe->headoffset;
 561 }
 562
 563 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 564                                 struct page *page,
 565                                 struct list_head *pagepool)
 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->cache_strategy, map->m_la))
 618                 cache_strategy = DELAYEDALLOC;
 619         else
 620                 cache_strategy = DONTALLOC;
 621
 622         preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 623                                  cache_strategy, pagepool);
 624
 625 hitted:
 626         /*
 627          * Ensure the current partial page belongs to this submit chain rather
 628          * than other concurrent submit chains or the noio(bypass) chain since
 629          * those chains are handled asynchronously thus the page cannot be used
 630          * for inplace I/O or pagevec (should be processed in strict order.)
 631          */
 632         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 633                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 634
 635         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 636         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 637                 zero_user_segment(page, cur, end);
 638                 goto next_part;
 639         }
 640
 641         /* let's derive page type */
 642         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 643                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 644                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 645                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 646
 647         if (cur)
 648                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 649
 650 retry:
 651         err = z_erofs_attach_page(clt, page, page_type);
 652         /* should allocate an additional staging page for pagevec */
 653         if (err == -EAGAIN) {
 654                 struct page *const newpage =
 655                         erofs_allocpage(pagepool, GFP_NOFS | __GFP_NOFAIL);
 656
 657                 newpage->mapping = Z_EROFS_MAPPING_STAGING;
 658                 err = z_erofs_attach_page(clt, newpage,
 659                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 660                 if (!err)
 661                         goto retry;
 662         }
 663
 664         if (err)
 665                 goto err_out;
 666
 667         index = page->index - (map->m_la >> PAGE_SHIFT);
 668
 669         z_erofs_onlinepage_fixup(page, index, true);
 670
 671         /* bump up the number of spiltted parts of a page */
 672         ++spiltted;
 673         /* also update nr_pages */
 674         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 675 next_part:
 676         /* can be used for verification */
 677         map->m_llen = offset + cur - map->m_la;
 678
 679         end = cur;
 680         if (end > 0)
 681                 goto repeat;
 682
 683 out:
 684         z_erofs_onlinepage_endio(page);
 685
 686         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 687                   __func__, page, spiltted, map->m_llen);
 688         return err;
 689
 690         /* if some error occurred while processing this page */
 691 err_out:
 692         SetPageError(page);
 693         goto out;
 694 }
 695
 696 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 697                                        bool sync, int bios)
 698 {
 699         /* wake up the caller thread for sync decompression */
 700         if (sync) {
 701                 unsigned long flags;
 702
 703                 spin_lock_irqsave(&io->u.wait.lock, flags);
 704                 if (!atomic_add_return(bios, &io->pending_bios))
 705                         wake_up_locked(&io->u.wait);
 706                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 707                 return;
 708         }
 709
 710         if (!atomic_add_return(bios, &io->pending_bios))
 711                 queue_work(z_erofs_workqueue, &io->u.work);
 712 }
 713
 714 static void z_erofs_decompressqueue_endio(struct bio *bio)
 715 {
 716         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
 717         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
 718         blk_status_t err = bio->bi_status;
 719         struct bio_vec *bvec;
 720         struct bvec_iter_all iter_all;
 721
 722         bio_for_each_segment_all(bvec, bio, iter_all) {
 723                 struct page *page = bvec->bv_page;
 724
 725                 DBG_BUGON(PageUptodate(page));
 726                 DBG_BUGON(!page->mapping);
 727
 728                 if (err)
 729                         SetPageError(page);
 730
 731                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
 732                         if (!err)
 733                                 SetPageUptodate(page);
 734                         unlock_page(page);
 735                 }
 736         }
 737         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
 738         bio_put(bio);
 739 }
 740
 741 static int z_erofs_decompress_pcluster(struct super_block *sb,
 742                                        struct z_erofs_pcluster *pcl,
 743                                        struct list_head *pagepool)
 744 {
 745         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 746         const unsigned int clusterpages = BIT(pcl->clusterbits);
 747         struct z_erofs_pagevec_ctor ctor;
 748         unsigned int i, outputsize, llen, nr_pages;
 749         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 750         struct page **pages, **compressed_pages, *page;
 751
 752         enum z_erofs_page_type page_type;
 753         bool overlapped, partial;
 754         struct z_erofs_collection *cl;
 755         int err;
 756
 757         might_sleep();
 758         cl = z_erofs_primarycollection(pcl);
 759         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 760
 761         mutex_lock(&cl->lock);
 762         nr_pages = cl->nr_pages;
 763
 764         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 765                 pages = pages_onstack;
 766         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 767                    mutex_trylock(&z_pagemap_global_lock)) {
 768                 pages = z_pagemap_global;
 769         } else {
 770                 gfp_t gfp_flags = GFP_KERNEL;
 771
 772                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 773                         gfp_flags |= __GFP_NOFAIL;
 774
 775                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 776                                        gfp_flags);
 777
 778                 /* fallback to global pagemap for the lowmem scenario */
 779                 if (!pages) {
 780                         mutex_lock(&z_pagemap_global_lock);
 781                         pages = z_pagemap_global;
 782                 }
 783         }
 784
 785         for (i = 0; i < nr_pages; ++i)
 786                 pages[i] = NULL;
 787
 788         err = 0;
 789         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 790                                   cl->pagevec, 0);
 791
 792         for (i = 0; i < cl->vcnt; ++i) {
 793                 unsigned int pagenr;
 794
 795                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 796
 797                 /* all pages in pagevec ought to be valid */
 798                 DBG_BUGON(!page);
 799                 DBG_BUGON(!page->mapping);
 800
 801                 if (z_erofs_put_stagingpage(pagepool, page))
 802                         continue;
 803
 804                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 805                         pagenr = 0;
 806                 else
 807                         pagenr = z_erofs_onlinepage_index(page);
 808
 809                 DBG_BUGON(pagenr >= nr_pages);
 810
 811                 /*
 812                  * currently EROFS doesn't support multiref(dedup),
 813                  * so here erroring out one multiref page.
 814                  */
 815                 if (pages[pagenr]) {
 816                         DBG_BUGON(1);
 817                         SetPageError(pages[pagenr]);
 818                         z_erofs_onlinepage_endio(pages[pagenr]);
 819                         err = -EFSCORRUPTED;
 820                 }
 821                 pages[pagenr] = page;
 822         }
 823         z_erofs_pagevec_ctor_exit(&ctor, true);
 824
 825         overlapped = false;
 826         compressed_pages = pcl->compressed_pages;
 827
 828         for (i = 0; i < clusterpages; ++i) {
 829                 unsigned int pagenr;
 830
 831                 page = compressed_pages[i];
 832
 833                 /* all compressed pages ought to be valid */
 834                 DBG_BUGON(!page);
 835                 DBG_BUGON(!page->mapping);
 836
 837                 if (!z_erofs_page_is_staging(page)) {
 838                         if (erofs_page_is_managed(sbi, page)) {
 839                                 if (!PageUptodate(page))
 840                                         err = -EIO;
 841                                 continue;
 842                         }
 843
 844                         /*
 845                          * only if non-head page can be selected
 846                          * for inplace decompression
 847                          */
 848                         pagenr = z_erofs_onlinepage_index(page);
 849
 850                         DBG_BUGON(pagenr >= nr_pages);
 851                         if (pages[pagenr]) {
 852                                 DBG_BUGON(1);
 853                                 SetPageError(pages[pagenr]);
 854                                 z_erofs_onlinepage_endio(pages[pagenr]);
 855                                 err = -EFSCORRUPTED;
 856                         }
 857                         pages[pagenr] = page;
 858
 859                         overlapped = true;
 860                 }
 861
 862                 /* PG_error needs checking for inplaced and staging pages */
 863                 if (PageError(page)) {
 864                         DBG_BUGON(PageUptodate(page));
 865                         err = -EIO;
 866                 }
 867         }
 868
 869         if (err)
 870                 goto out;
 871
 872         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 873         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 874                 outputsize = llen;
 875                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 876         } else {
 877                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 878                 partial = true;
 879         }
 880
 881         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 882                                         .sb = sb,
 883                                         .in = compressed_pages,
 884                                         .out = pages,
 885                                         .pageofs_out = cl->pageofs,
 886                                         .inputsize = PAGE_SIZE,
 887                                         .outputsize = outputsize,
 888                                         .alg = pcl->algorithmformat,
 889                                         .inplace_io = overlapped,
 890                                         .partial_decoding = partial
 891                                  }, pagepool);
 892
 893 out:
 894         /* must handle all compressed pages before endding pages */
 895         for (i = 0; i < clusterpages; ++i) {
 896                 page = compressed_pages[i];
 897
 898                 if (erofs_page_is_managed(sbi, page))
 899                         continue;
 900
 901                 /* recycle all individual staging pages */
 902                 (void)z_erofs_put_stagingpage(pagepool, page);
 903
 904                 WRITE_ONCE(compressed_pages[i], NULL);
 905         }
 906
 907         for (i = 0; i < nr_pages; ++i) {
 908                 page = pages[i];
 909                 if (!page)
 910                         continue;
 911
 912                 DBG_BUGON(!page->mapping);
 913
 914                 /* recycle all individual staging pages */
 915                 if (z_erofs_put_stagingpage(pagepool, page))
 916                         continue;
 917
 918                 if (err < 0)
 919                         SetPageError(page);
 920
 921                 z_erofs_onlinepage_endio(page);
 922         }
 923
 924         if (pages == z_pagemap_global)
 925                 mutex_unlock(&z_pagemap_global_lock);
 926         else if (pages != pages_onstack)
 927                 kvfree(pages);
 928
 929         cl->nr_pages = 0;
 930         cl->vcnt = 0;
 931
 932         /* all cl locks MUST be taken before the following line */
 933         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
 934
 935         /* all cl locks SHOULD be released right now */
 936         mutex_unlock(&cl->lock);
 937
 938         z_erofs_collection_put(cl);
 939         return err;
 940 }
 941
 942 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
 943                                      struct list_head *pagepool)
 944 {
 945         z_erofs_next_pcluster_t owned = io->head;
 946
 947         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
 948                 struct z_erofs_pcluster *pcl;
 949
 950                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
 951                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
 952
 953                 /* no possible that 'owned' equals NULL */
 954                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
 955
 956                 pcl = container_of(owned, struct z_erofs_pcluster, next);
 957                 owned = READ_ONCE(pcl->next);
 958
 959                 z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
 960         }
 961 }
 962
 963 static void z_erofs_decompressqueue_work(struct work_struct *work)
 964 {
 965         struct z_erofs_decompressqueue *bgq =
 966                 container_of(work, struct z_erofs_decompressqueue, u.work);
 967         LIST_HEAD(pagepool);
 968
 969         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 970         z_erofs_decompress_queue(bgq, &pagepool);
 971
 972         put_pages_list(&pagepool);
 973         kvfree(bgq);
 974 }
 975
 976 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
 977                                                unsigned int nr,
 978                                                struct list_head *pagepool,
 979                                                struct address_space *mc,
 980                                                gfp_t gfp)
 981 {
 982         const pgoff_t index = pcl->obj.index;
 983         bool tocache = false;
 984
 985         struct address_space *mapping;
 986         struct page *oldpage, *page;
 987
 988         compressed_page_t t;
 989         int justfound;
 990
 991 repeat:
 992         page = READ_ONCE(pcl->compressed_pages[nr]);
 993         oldpage = page;
 994
 995         if (!page)
 996                 goto out_allocpage;
 997
 998         /*
 999          * the cached page has not been allocated and
1000          * an placeholder is out there, prepare it now.
1001          */
1002         if (page == PAGE_UNALLOCATED) {
1003                 tocache = true;
1004                 goto out_allocpage;
1005         }
1006
1007         /* process the target tagged pointer */
1008         t = tagptr_init(compressed_page_t, page);
1009         justfound = tagptr_unfold_tags(t);
1010         page = tagptr_unfold_ptr(t);
1011
1012         mapping = READ_ONCE(page->mapping);
1013
1014         /*
1015          * unmanaged (file) pages are all locked solidly,
1016          * therefore it is impossible for `mapping' to be NULL.
1017          */
1018         if (mapping && mapping != mc)
1019                 /* ought to be unmanaged pages */
1020                 goto out;
1021
1022         lock_page(page);
1023
1024         /* only true if page reclaim goes wrong, should never happen */
1025         DBG_BUGON(justfound && PagePrivate(page));
1026
1027         /* the page is still in manage cache */
1028         if (page->mapping == mc) {
1029                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1030
1031                 ClearPageError(page);
1032                 if (!PagePrivate(page)) {
1033                         /*
1034                          * impossible to be !PagePrivate(page) for
1035                          * the current restriction as well if
1036                          * the page is already in compressed_pages[].
1037                          */
1038                         DBG_BUGON(!justfound);
1039
1040                         justfound = 0;
1041                         set_page_private(page, (unsigned long)pcl);
1042                         SetPagePrivate(page);
1043                 }
1044
1045                 /* no need to submit io if it is already up-to-date */
1046                 if (PageUptodate(page)) {
1047                         unlock_page(page);
1048                         page = NULL;
1049                 }
1050                 goto out;
1051         }
1052
1053         /*
1054          * the managed page has been truncated, it's unsafe to
1055          * reuse this one, let's allocate a new cache-managed page.
1056          */
1057         DBG_BUGON(page->mapping);
1058         DBG_BUGON(!justfound);
1059
1060         tocache = true;
1061         unlock_page(page);
1062         put_page(page);
1063 out_allocpage:
1064         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1065         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1066                 /* non-LRU / non-movable temporary page is needed */
1067                 page->mapping = Z_EROFS_MAPPING_STAGING;
1068                 tocache = false;
1069         }
1070
1071         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1072                 if (tocache) {
1073                         /* since it added to managed cache successfully */
1074                         unlock_page(page);
1075                         put_page(page);
1076                 } else {
1077                         list_add(&page->lru, pagepool);
1078                 }
1079                 cond_resched();
1080                 goto repeat;
1081         }
1082         set_page_private(page, (unsigned long)pcl);
1083         SetPagePrivate(page);
1084 out:    /* the only exit (for tracing and debugging) */
1085         return page;
1086 }
1087
1088 static struct z_erofs_decompressqueue *
1089 jobqueue_init(struct super_block *sb,
1090               struct z_erofs_decompressqueue *fgq, bool *fg)
1091 {
1092         struct z_erofs_decompressqueue *q;
1093
1094         if (fg && !*fg) {
1095                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1096                 if (!q) {
1097                         *fg = true;
1098                         goto fg_out;
1099                 }
1100                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1101         } else {
1102 fg_out:
1103                 q = fgq;
1104                 init_waitqueue_head(&fgq->u.wait);
1105                 atomic_set(&fgq->pending_bios, 0);
1106         }
1107         q->sb = sb;
1108         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1109         return q;
1110 }
1111
1112 /* define decompression jobqueue types */
1113 enum {
1114         JQ_BYPASS,
1115         JQ_SUBMIT,
1116         NR_JOBQUEUES,
1117 };
1118
1119 static void *jobqueueset_init(struct super_block *sb,
1120                               struct z_erofs_decompressqueue *q[],
1121                               struct z_erofs_decompressqueue *fgq, bool *fg)
1122 {
1123         /*
1124          * if managed cache is enabled, bypass jobqueue is needed,
1125          * no need to read from device for all pclusters in this queue.
1126          */
1127         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1128         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1129
1130         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1131 }
1132
1133 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1134                                     z_erofs_next_pcluster_t qtail[],
1135                                     z_erofs_next_pcluster_t owned_head)
1136 {
1137         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1138         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1139
1140         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1141         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1142                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1143
1144         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1145
1146         WRITE_ONCE(*submit_qtail, owned_head);
1147         WRITE_ONCE(*bypass_qtail, &pcl->next);
1148
1149         qtail[JQ_BYPASS] = &pcl->next;
1150 }
1151
1152 static bool postsubmit_is_all_bypassed(struct z_erofs_decompressqueue *q[],
1153                                        unsigned int nr_bios, bool force_fg)
1154 {
1155         /*
1156          * although background is preferred, no one is pending for submission.
1157          * don't issue workqueue for decompression but drop it directly instead.
1158          */
1159         if (force_fg || nr_bios)
1160                 return false;
1161
1162         kvfree(q[JQ_SUBMIT]);
1163         return true;
1164 }
1165
1166 static bool z_erofs_submit_queue(struct super_block *sb,
1167                                  z_erofs_next_pcluster_t owned_head,
1168                                  struct list_head *pagepool,
1169                                  struct z_erofs_decompressqueue *fgq,
1170                                  bool *force_fg)
1171 {
1172         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1173         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1174         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1175         struct bio *bio;
1176         void *bi_private;
1177         /* since bio will be NULL, no need to initialize last_index */
1178         pgoff_t uninitialized_var(last_index);
1179         bool force_submit = false;
1180         unsigned int nr_bios;
1181
1182         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1183                 return false;
1184
1185         force_submit = false;
1186         bio = NULL;
1187         nr_bios = 0;
1188         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1189         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1190         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1191
1192         /* by default, all need io submission */
1193         q[JQ_SUBMIT]->head = owned_head;
1194
1195         do {
1196                 struct z_erofs_pcluster *pcl;
1197                 unsigned int clusterpages;
1198                 pgoff_t first_index;
1199                 struct page *page;
1200                 unsigned int i = 0, bypass = 0;
1201                 int err;
1202
1203                 /* no possible 'owned_head' equals the following */
1204                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1205                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1206
1207                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1208
1209                 clusterpages = BIT(pcl->clusterbits);
1210
1211                 /* close the main owned chain at first */
1212                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1213                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1214
1215                 first_index = pcl->obj.index;
1216                 force_submit |= (first_index != last_index + 1);
1217
1218 repeat:
1219                 page = pickup_page_for_submission(pcl, i, pagepool,
1220                                                   MNGD_MAPPING(sbi),
1221                                                   GFP_NOFS);
1222                 if (!page) {
1223                         force_submit = true;
1224                         ++bypass;
1225                         goto skippage;
1226                 }
1227
1228                 if (bio && force_submit) {
1229 submit_bio_retry:
1230                         submit_bio(bio);
1231                         bio = NULL;
1232                 }
1233
1234                 if (!bio) {
1235                         bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1236
1237                         bio->bi_end_io = z_erofs_decompressqueue_endio;
1238                         bio_set_dev(bio, sb->s_bdev);
1239                         bio->bi_iter.bi_sector = (sector_t)(first_index + i) <<
1240                                 LOG_SECTORS_PER_BLOCK;
1241                         bio->bi_private = bi_private;
1242                         bio->bi_opf = REQ_OP_READ;
1243
1244                         ++nr_bios;
1245                 }
1246
1247                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1248                 if (err < PAGE_SIZE)
1249                         goto submit_bio_retry;
1250
1251                 force_submit = false;
1252                 last_index = first_index + i;
1253 skippage:
1254                 if (++i < clusterpages)
1255                         goto repeat;
1256
1257                 if (bypass < clusterpages)
1258                         qtail[JQ_SUBMIT] = &pcl->next;
1259                 else
1260                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1261         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1262
1263         if (bio)
1264                 submit_bio(bio);
1265
1266         if (postsubmit_is_all_bypassed(q, nr_bios, *force_fg))
1267                 return true;
1268
1269         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1270         return true;
1271 }
1272
1273 static void z_erofs_runqueue(struct super_block *sb,
1274                              struct z_erofs_collector *clt,
1275                              struct list_head *pagepool, bool force_fg)
1276 {
1277         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1278
1279         if (!z_erofs_submit_queue(sb, clt->owned_head,
1280                                   pagepool, io, &force_fg))
1281                 return;
1282
1283         /* handle bypass queue (no i/o pclusters) immediately */
1284         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1285
1286         if (!force_fg)
1287                 return;
1288
1289         /* wait until all bios are completed */
1290         io_wait_event(io[JQ_SUBMIT].u.wait,
1291                       !atomic_read(&io[JQ_SUBMIT].pending_bios));
1292
1293         /* handle synchronous decompress queue in the caller context */
1294         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1295 }
1296
1297 static int z_erofs_readpage(struct file *file, struct page *page)
1298 {
1299         struct inode *const inode = page->mapping->host;
1300         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1301         int err;
1302         LIST_HEAD(pagepool);
1303
1304         trace_erofs_readpage(page, false);
1305
1306         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1307
1308         err = z_erofs_do_read_page(&f, page, &pagepool);
1309         (void)z_erofs_collector_end(&f.clt);
1310
1311         /* if some compressed cluster ready, need submit them anyway */
1312         z_erofs_runqueue(inode->i_sb, &f.clt, &pagepool, true);
1313
1314         if (err)
1315                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1316
1317         if (f.map.mpage)
1318                 put_page(f.map.mpage);
1319
1320         /* clean up the remaining free pages */
1321         put_pages_list(&pagepool);
1322         return err;
1323 }
1324
1325 static bool should_decompress_synchronously(struct erofs_sb_info *sbi,
1326                                             unsigned int nr)
1327 {
1328         return nr <= sbi->max_sync_decompress_pages;
1329 }
1330
1331 static int z_erofs_readpages(struct file *filp, struct address_space *mapping,
1332                              struct list_head *pages, unsigned int nr_pages)
1333 {
1334         struct inode *const inode = mapping->host;
1335         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1336
1337         bool sync = should_decompress_synchronously(sbi, nr_pages);
1338         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1339         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1340         struct page *head = NULL;
1341         LIST_HEAD(pagepool);
1342
1343         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1344                               nr_pages, false);
1345
1346         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1347
1348         for (; nr_pages; --nr_pages) {
1349                 struct page *page = lru_to_page(pages);
1350
1351                 prefetchw(&page->flags);
1352                 list_del(&page->lru);
1353
1354                 /*
1355                  * A pure asynchronous readahead is indicated if
1356                  * a PG_readahead marked page is hitted at first.
1357                  * Let's also do asynchronous decompression for this case.
1358                  */
1359                 sync &= !(PageReadahead(page) && !head);
1360
1361                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1362                         list_add(&page->lru, &pagepool);
1363                         continue;
1364                 }
1365
1366                 set_page_private(page, (unsigned long)head);
1367                 head = page;
1368         }
1369
1370         while (head) {
1371                 struct page *page = head;
1372                 int err;
1373
1374                 /* traversal in reverse order */
1375                 head = (void *)page_private(page);
1376
1377                 err = z_erofs_do_read_page(&f, page, &pagepool);
1378                 if (err)
1379                         erofs_err(inode->i_sb,
1380                                   "readahead error at page %lu @ nid %llu",
1381                                   page->index, EROFS_I(inode)->nid);
1382                 put_page(page);
1383         }
1384
1385         (void)z_erofs_collector_end(&f.clt);
1386
1387         z_erofs_runqueue(inode->i_sb, &f.clt, &pagepool, sync);
1388
1389         if (f.map.mpage)
1390                 put_page(f.map.mpage);
1391
1392         /* clean up the remaining free pages */
1393         put_pages_list(&pagepool);
1394         return 0;
1395 }
1396
1397 const struct address_space_operations z_erofs_aops = {
1398         .readpage = z_erofs_readpage,
1399         .readpages = z_erofs_readpages,
1400 };
1401