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 struct z_erofs_collection *cllookup(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 NULL;
 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 ERR_PTR(-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 ERR_PTR(-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 ERR_PTR(-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 cl;
 398 }
 399
 400 static struct z_erofs_collection *clregister(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 ERR_PTR(-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 ERR_PTR(-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 cl;
 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         struct z_erofs_collection *cl;
 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         cl = cllookup(clt, inode, map);
 475         if (!cl) {
 476                 cl = clregister(clt, inode, map);
 477
 478                 if (cl == ERR_PTR(-EAGAIN))
 479                         goto repeat;
 480         }
 481
 482         if (IS_ERR(cl))
 483                 return PTR_ERR(cl);
 484
 485         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 486                                   cl->pagevec, cl->vcnt);
 487
 488         clt->compressedpages = clt->pcl->compressed_pages;
 489         if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
 490                 clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
 491         return 0;
 492 }
 493
 494 /*
 495  * keep in mind that no referenced pclusters will be freed
 496  * only after a RCU grace period.
 497  */
 498 static void z_erofs_rcu_callback(struct rcu_head *head)
 499 {
 500         struct z_erofs_collection *const cl =
 501                 container_of(head, struct z_erofs_collection, rcu);
 502
 503         kmem_cache_free(pcluster_cachep,
 504                         container_of(cl, struct z_erofs_pcluster,
 505                                      primary_collection));
 506 }
 507
 508 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 509 {
 510         struct z_erofs_pcluster *const pcl =
 511                 container_of(grp, struct z_erofs_pcluster, obj);
 512         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 513
 514         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 515 }
 516
 517 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 518 {
 519         struct z_erofs_pcluster *const pcl =
 520                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 521
 522         erofs_workgroup_put(&pcl->obj);
 523 }
 524
 525 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 526 {
 527         struct z_erofs_collection *cl = clt->cl;
 528
 529         if (!cl)
 530                 return false;
 531
 532         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 533         mutex_unlock(&cl->lock);
 534
 535         /*
 536          * if all pending pages are added, don't hold its reference
 537          * any longer if the pcluster isn't hosted by ourselves.
 538          */
 539         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 540                 z_erofs_collection_put(cl);
 541
 542         clt->cl = NULL;
 543         return true;
 544 }
 545
 546 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 547                                                gfp_t gfp)
 548 {
 549         struct page *page = erofs_allocpage(pagepool, gfp, true);
 550
 551         page->mapping = Z_EROFS_MAPPING_STAGING;
 552         return page;
 553 }
 554
 555 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 556                                        unsigned int cachestrategy,
 557                                        erofs_off_t la)
 558 {
 559         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 560                 return false;
 561
 562         if (fe->backmost)
 563                 return true;
 564
 565         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 566                 la < fe->headoffset;
 567 }
 568
 569 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 570                                 struct page *page,
 571                                 struct list_head *pagepool)
 572 {
 573         struct inode *const inode = fe->inode;
 574         struct erofs_sb_info *const sbi __maybe_unused = EROFS_I_SB(inode);
 575         struct erofs_map_blocks *const map = &fe->map;
 576         struct z_erofs_collector *const clt = &fe->clt;
 577         const loff_t offset = page_offset(page);
 578         bool tight = (clt->mode >= COLLECT_PRIMARY_HOOKED);
 579
 580         enum z_erofs_cache_alloctype cache_strategy;
 581         enum z_erofs_page_type page_type;
 582         unsigned int cur, end, spiltted, index;
 583         int err = 0;
 584
 585         /* register locked file pages as online pages in pack */
 586         z_erofs_onlinepage_init(page);
 587
 588         spiltted = 0;
 589         end = PAGE_SIZE;
 590 repeat:
 591         cur = end - 1;
 592
 593         /* lucky, within the range of the current map_blocks */
 594         if (offset + cur >= map->m_la &&
 595             offset + cur < map->m_la + map->m_llen) {
 596                 /* didn't get a valid collection previously (very rare) */
 597                 if (!clt->cl)
 598                         goto restart_now;
 599                 goto hitted;
 600         }
 601
 602         /* go ahead the next map_blocks */
 603         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 604
 605         if (z_erofs_collector_end(clt))
 606                 fe->backmost = false;
 607
 608         map->m_la = offset + cur;
 609         map->m_llen = 0;
 610         err = z_erofs_map_blocks_iter(inode, map, 0);
 611         if (err)
 612                 goto err_out;
 613
 614 restart_now:
 615         if (!(map->m_flags & EROFS_MAP_MAPPED))
 616                 goto hitted;
 617
 618         err = z_erofs_collector_begin(clt, inode, map);
 619         if (err)
 620                 goto err_out;
 621
 622         /* preload all compressed pages (maybe downgrade role if necessary) */
 623         if (should_alloc_managed_pages(fe, sbi->cache_strategy, map->m_la))
 624                 cache_strategy = DELAYEDALLOC;
 625         else
 626                 cache_strategy = DONTALLOC;
 627
 628         preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 629                                  cache_strategy, pagepool);
 630
 631         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED);
 632 hitted:
 633         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 634         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 635                 zero_user_segment(page, cur, end);
 636                 goto next_part;
 637         }
 638
 639         /* let's derive page type */
 640         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 641                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 642                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 643                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 644
 645         if (cur)
 646                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 647
 648 retry:
 649         err = z_erofs_attach_page(clt, page, page_type);
 650         /* should allocate an additional staging page for pagevec */
 651         if (err == -EAGAIN) {
 652                 struct page *const newpage =
 653                         __stagingpage_alloc(pagepool, GFP_NOFS);
 654
 655                 err = z_erofs_attach_page(clt, newpage,
 656                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 657                 if (!err)
 658                         goto retry;
 659         }
 660
 661         if (err)
 662                 goto err_out;
 663
 664         index = page->index - (map->m_la >> PAGE_SHIFT);
 665
 666         z_erofs_onlinepage_fixup(page, index, true);
 667
 668         /* bump up the number of spiltted parts of a page */
 669         ++spiltted;
 670         /* also update nr_pages */
 671         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 672 next_part:
 673         /* can be used for verification */
 674         map->m_llen = offset + cur - map->m_la;
 675
 676         end = cur;
 677         if (end > 0)
 678                 goto repeat;
 679
 680 out:
 681         z_erofs_onlinepage_endio(page);
 682
 683         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 684                   __func__, page, spiltted, map->m_llen);
 685         return err;
 686
 687         /* if some error occurred while processing this page */
 688 err_out:
 689         SetPageError(page);
 690         goto out;
 691 }
 692
 693 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 694 {
 695         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 696         struct z_erofs_unzip_io *io = tagptr_unfold_ptr(t);
 697         bool background = tagptr_unfold_tags(t);
 698
 699         if (!background) {
 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 inline void z_erofs_vle_read_endio(struct bio *bio)
 714 {
 715         struct erofs_sb_info *sbi = NULL;
 716         blk_status_t err = bio->bi_status;
 717         struct bio_vec *bvec;
 718         struct bvec_iter_all iter_all;
 719
 720         bio_for_each_segment_all(bvec, bio, iter_all) {
 721                 struct page *page = bvec->bv_page;
 722                 bool cachemngd = false;
 723
 724                 DBG_BUGON(PageUptodate(page));
 725                 DBG_BUGON(!page->mapping);
 726
 727                 if (!sbi && !z_erofs_page_is_staging(page))
 728                         sbi = EROFS_SB(page->mapping->host->i_sb);
 729
 730                 /* sbi should already be gotten if the page is managed */
 731                 if (sbi)
 732                         cachemngd = erofs_page_is_managed(sbi, page);
 733
 734                 if (err)
 735                         SetPageError(page);
 736                 else if (cachemngd)
 737                         SetPageUptodate(page);
 738
 739                 if (cachemngd)
 740                         unlock_page(page);
 741         }
 742
 743         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 744         bio_put(bio);
 745 }
 746
 747 static int z_erofs_decompress_pcluster(struct super_block *sb,
 748                                        struct z_erofs_pcluster *pcl,
 749                                        struct list_head *pagepool)
 750 {
 751         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 752         const unsigned int clusterpages = BIT(pcl->clusterbits);
 753         struct z_erofs_pagevec_ctor ctor;
 754         unsigned int i, outputsize, llen, nr_pages;
 755         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 756         struct page **pages, **compressed_pages, *page;
 757
 758         enum z_erofs_page_type page_type;
 759         bool overlapped, partial;
 760         struct z_erofs_collection *cl;
 761         int err;
 762
 763         might_sleep();
 764         cl = z_erofs_primarycollection(pcl);
 765         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 766
 767         mutex_lock(&cl->lock);
 768         nr_pages = cl->nr_pages;
 769
 770         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 771                 pages = pages_onstack;
 772         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 773                    mutex_trylock(&z_pagemap_global_lock)) {
 774                 pages = z_pagemap_global;
 775         } else {
 776                 gfp_t gfp_flags = GFP_KERNEL;
 777
 778                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 779                         gfp_flags |= __GFP_NOFAIL;
 780
 781                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 782                                        gfp_flags);
 783
 784                 /* fallback to global pagemap for the lowmem scenario */
 785                 if (!pages) {
 786                         mutex_lock(&z_pagemap_global_lock);
 787                         pages = z_pagemap_global;
 788                 }
 789         }
 790
 791         for (i = 0; i < nr_pages; ++i)
 792                 pages[i] = NULL;
 793
 794         err = 0;
 795         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 796                                   cl->pagevec, 0);
 797
 798         for (i = 0; i < cl->vcnt; ++i) {
 799                 unsigned int pagenr;
 800
 801                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 802
 803                 /* all pages in pagevec ought to be valid */
 804                 DBG_BUGON(!page);
 805                 DBG_BUGON(!page->mapping);
 806
 807                 if (z_erofs_put_stagingpage(pagepool, page))
 808                         continue;
 809
 810                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 811                         pagenr = 0;
 812                 else
 813                         pagenr = z_erofs_onlinepage_index(page);
 814
 815                 DBG_BUGON(pagenr >= nr_pages);
 816
 817                 /*
 818                  * currently EROFS doesn't support multiref(dedup),
 819                  * so here erroring out one multiref page.
 820                  */
 821                 if (pages[pagenr]) {
 822                         DBG_BUGON(1);
 823                         SetPageError(pages[pagenr]);
 824                         z_erofs_onlinepage_endio(pages[pagenr]);
 825                         err = -EFSCORRUPTED;
 826                 }
 827                 pages[pagenr] = page;
 828         }
 829         z_erofs_pagevec_ctor_exit(&ctor, true);
 830
 831         overlapped = false;
 832         compressed_pages = pcl->compressed_pages;
 833
 834         for (i = 0; i < clusterpages; ++i) {
 835                 unsigned int pagenr;
 836
 837                 page = compressed_pages[i];
 838
 839                 /* all compressed pages ought to be valid */
 840                 DBG_BUGON(!page);
 841                 DBG_BUGON(!page->mapping);
 842
 843                 if (!z_erofs_page_is_staging(page)) {
 844                         if (erofs_page_is_managed(sbi, page)) {
 845                                 if (!PageUptodate(page))
 846                                         err = -EIO;
 847                                 continue;
 848                         }
 849
 850                         /*
 851                          * only if non-head page can be selected
 852                          * for inplace decompression
 853                          */
 854                         pagenr = z_erofs_onlinepage_index(page);
 855
 856                         DBG_BUGON(pagenr >= nr_pages);
 857                         if (pages[pagenr]) {
 858                                 DBG_BUGON(1);
 859                                 SetPageError(pages[pagenr]);
 860                                 z_erofs_onlinepage_endio(pages[pagenr]);
 861                                 err = -EFSCORRUPTED;
 862                         }
 863                         pages[pagenr] = page;
 864
 865                         overlapped = true;
 866                 }
 867
 868                 /* PG_error needs checking for inplaced and staging pages */
 869                 if (PageError(page)) {
 870                         DBG_BUGON(PageUptodate(page));
 871                         err = -EIO;
 872                 }
 873         }
 874
 875         if (err)
 876                 goto out;
 877
 878         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 879         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 880                 outputsize = llen;
 881                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 882         } else {
 883                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 884                 partial = true;
 885         }
 886
 887         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 888                                         .sb = sb,
 889                                         .in = compressed_pages,
 890                                         .out = pages,
 891                                         .pageofs_out = cl->pageofs,
 892                                         .inputsize = PAGE_SIZE,
 893                                         .outputsize = outputsize,
 894                                         .alg = pcl->algorithmformat,
 895                                         .inplace_io = overlapped,
 896                                         .partial_decoding = partial
 897                                  }, pagepool);
 898
 899 out:
 900         /* must handle all compressed pages before endding pages */
 901         for (i = 0; i < clusterpages; ++i) {
 902                 page = compressed_pages[i];
 903
 904                 if (erofs_page_is_managed(sbi, page))
 905                         continue;
 906
 907                 /* recycle all individual staging pages */
 908                 (void)z_erofs_put_stagingpage(pagepool, page);
 909
 910                 WRITE_ONCE(compressed_pages[i], NULL);
 911         }
 912
 913         for (i = 0; i < nr_pages; ++i) {
 914                 page = pages[i];
 915                 if (!page)
 916                         continue;
 917
 918                 DBG_BUGON(!page->mapping);
 919
 920                 /* recycle all individual staging pages */
 921                 if (z_erofs_put_stagingpage(pagepool, page))
 922                         continue;
 923
 924                 if (err < 0)
 925                         SetPageError(page);
 926
 927                 z_erofs_onlinepage_endio(page);
 928         }
 929
 930         if (pages == z_pagemap_global)
 931                 mutex_unlock(&z_pagemap_global_lock);
 932         else if (pages != pages_onstack)
 933                 kvfree(pages);
 934
 935         cl->nr_pages = 0;
 936         cl->vcnt = 0;
 937
 938         /* all cl locks MUST be taken before the following line */
 939         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
 940
 941         /* all cl locks SHOULD be released right now */
 942         mutex_unlock(&cl->lock);
 943
 944         z_erofs_collection_put(cl);
 945         return err;
 946 }
 947
 948 static void z_erofs_vle_unzip_all(struct super_block *sb,
 949                                   struct z_erofs_unzip_io *io,
 950                                   struct list_head *pagepool)
 951 {
 952         z_erofs_next_pcluster_t owned = io->head;
 953
 954         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
 955                 struct z_erofs_pcluster *pcl;
 956
 957                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
 958                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
 959
 960                 /* no possible that 'owned' equals NULL */
 961                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
 962
 963                 pcl = container_of(owned, struct z_erofs_pcluster, next);
 964                 owned = READ_ONCE(pcl->next);
 965
 966                 z_erofs_decompress_pcluster(sb, pcl, pagepool);
 967         }
 968 }
 969
 970 static void z_erofs_vle_unzip_wq(struct work_struct *work)
 971 {
 972         struct z_erofs_unzip_io_sb *iosb =
 973                 container_of(work, struct z_erofs_unzip_io_sb, io.u.work);
 974         LIST_HEAD(pagepool);
 975
 976         DBG_BUGON(iosb->io.head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 977         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &pagepool);
 978
 979         put_pages_list(&pagepool);
 980         kvfree(iosb);
 981 }
 982
 983 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
 984                                                unsigned int nr,
 985                                                struct list_head *pagepool,
 986                                                struct address_space *mc,
 987                                                gfp_t gfp)
 988 {
 989         /* determined at compile time to avoid too many #ifdefs */
 990         const bool nocache = __builtin_constant_p(mc) ? !mc : false;
 991         const pgoff_t index = pcl->obj.index;
 992         bool tocache = false;
 993
 994         struct address_space *mapping;
 995         struct page *oldpage, *page;
 996
 997         compressed_page_t t;
 998         int justfound;
 999
1000 repeat:
1001         page = READ_ONCE(pcl->compressed_pages[nr]);
1002         oldpage = page;
1003
1004         if (!page)
1005                 goto out_allocpage;
1006
1007         /*
1008          * the cached page has not been allocated and
1009          * an placeholder is out there, prepare it now.
1010          */
1011         if (!nocache && page == PAGE_UNALLOCATED) {
1012                 tocache = true;
1013                 goto out_allocpage;
1014         }
1015
1016         /* process the target tagged pointer */
1017         t = tagptr_init(compressed_page_t, page);
1018         justfound = tagptr_unfold_tags(t);
1019         page = tagptr_unfold_ptr(t);
1020
1021         mapping = READ_ONCE(page->mapping);
1022
1023         /*
1024          * if managed cache is disabled, it's no way to
1025          * get such a cached-like page.
1026          */
1027         if (nocache) {
1028                 /* if managed cache is disabled, it is impossible `justfound' */
1029                 DBG_BUGON(justfound);
1030
1031                 /* and it should be locked, not uptodate, and not truncated */
1032                 DBG_BUGON(!PageLocked(page));
1033                 DBG_BUGON(PageUptodate(page));
1034                 DBG_BUGON(!mapping);
1035                 goto out;
1036         }
1037
1038         /*
1039          * unmanaged (file) pages are all locked solidly,
1040          * therefore it is impossible for `mapping' to be NULL.
1041          */
1042         if (mapping && mapping != mc)
1043                 /* ought to be unmanaged pages */
1044                 goto out;
1045
1046         lock_page(page);
1047
1048         /* only true if page reclaim goes wrong, should never happen */
1049         DBG_BUGON(justfound && PagePrivate(page));
1050
1051         /* the page is still in manage cache */
1052         if (page->mapping == mc) {
1053                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1054
1055                 ClearPageError(page);
1056                 if (!PagePrivate(page)) {
1057                         /*
1058                          * impossible to be !PagePrivate(page) for
1059                          * the current restriction as well if
1060                          * the page is already in compressed_pages[].
1061                          */
1062                         DBG_BUGON(!justfound);
1063
1064                         justfound = 0;
1065                         set_page_private(page, (unsigned long)pcl);
1066                         SetPagePrivate(page);
1067                 }
1068
1069                 /* no need to submit io if it is already up-to-date */
1070                 if (PageUptodate(page)) {
1071                         unlock_page(page);
1072                         page = NULL;
1073                 }
1074                 goto out;
1075         }
1076
1077         /*
1078          * the managed page has been truncated, it's unsafe to
1079          * reuse this one, let's allocate a new cache-managed page.
1080          */
1081         DBG_BUGON(page->mapping);
1082         DBG_BUGON(!justfound);
1083
1084         tocache = true;
1085         unlock_page(page);
1086         put_page(page);
1087 out_allocpage:
1088         page = __stagingpage_alloc(pagepool, gfp);
1089         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1090                 list_add(&page->lru, pagepool);
1091                 cpu_relax();
1092                 goto repeat;
1093         }
1094         if (nocache || !tocache)
1095                 goto out;
1096         if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1097                 page->mapping = Z_EROFS_MAPPING_STAGING;
1098                 goto out;
1099         }
1100
1101         set_page_private(page, (unsigned long)pcl);
1102         SetPagePrivate(page);
1103 out:    /* the only exit (for tracing and debugging) */
1104         return page;
1105 }
1106
1107 static struct z_erofs_unzip_io *jobqueue_init(struct super_block *sb,
1108                                               struct z_erofs_unzip_io *io,
1109                                               bool foreground)
1110 {
1111         struct z_erofs_unzip_io_sb *iosb;
1112
1113         if (foreground) {
1114                 /* waitqueue available for foreground io */
1115                 DBG_BUGON(!io);
1116
1117                 init_waitqueue_head(&io->u.wait);
1118                 atomic_set(&io->pending_bios, 0);
1119                 goto out;
1120         }
1121
1122         iosb = kvzalloc(sizeof(*iosb), GFP_KERNEL | __GFP_NOFAIL);
1123         DBG_BUGON(!iosb);
1124
1125         /* initialize fields in the allocated descriptor */
1126         io = &iosb->io;
1127         iosb->sb = sb;
1128         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1129 out:
1130         io->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1131         return io;
1132 }
1133
1134 /* define decompression jobqueue types */
1135 enum {
1136         JQ_BYPASS,
1137         JQ_SUBMIT,
1138         NR_JOBQUEUES,
1139 };
1140
1141 static void *jobqueueset_init(struct super_block *sb,
1142                               z_erofs_next_pcluster_t qtail[],
1143                               struct z_erofs_unzip_io *q[],
1144                               struct z_erofs_unzip_io *fgq,
1145                               bool forcefg)
1146 {
1147         /*
1148          * if managed cache is enabled, bypass jobqueue is needed,
1149          * no need to read from device for all pclusters in this queue.
1150          */
1151         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1152         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1153
1154         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1155         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1156
1157         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1158 }
1159
1160 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1161                                     z_erofs_next_pcluster_t qtail[],
1162                                     z_erofs_next_pcluster_t owned_head)
1163 {
1164         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1165         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1166
1167         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1168         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1169                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1170
1171         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1172
1173         WRITE_ONCE(*submit_qtail, owned_head);
1174         WRITE_ONCE(*bypass_qtail, &pcl->next);
1175
1176         qtail[JQ_BYPASS] = &pcl->next;
1177 }
1178
1179 static bool postsubmit_is_all_bypassed(struct z_erofs_unzip_io *q[],
1180                                        unsigned int nr_bios,
1181                                        bool force_fg)
1182 {
1183         /*
1184          * although background is preferred, no one is pending for submission.
1185          * don't issue workqueue for decompression but drop it directly instead.
1186          */
1187         if (force_fg || nr_bios)
1188                 return false;
1189
1190         kvfree(container_of(q[JQ_SUBMIT], struct z_erofs_unzip_io_sb, io));
1191         return true;
1192 }
1193
1194 static bool z_erofs_vle_submit_all(struct super_block *sb,
1195                                    z_erofs_next_pcluster_t owned_head,
1196                                    struct list_head *pagepool,
1197                                    struct z_erofs_unzip_io *fgq,
1198                                    bool force_fg)
1199 {
1200         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
1201         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1202         struct z_erofs_unzip_io *q[NR_JOBQUEUES];
1203         struct bio *bio;
1204         void *bi_private;
1205         /* since bio will be NULL, no need to initialize last_index */
1206         pgoff_t uninitialized_var(last_index);
1207         bool force_submit = false;
1208         unsigned int nr_bios;
1209
1210         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1211                 return false;
1212
1213         force_submit = false;
1214         bio = NULL;
1215         nr_bios = 0;
1216         bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1217
1218         /* by default, all need io submission */
1219         q[JQ_SUBMIT]->head = owned_head;
1220
1221         do {
1222                 struct z_erofs_pcluster *pcl;
1223                 unsigned int clusterpages;
1224                 pgoff_t first_index;
1225                 struct page *page;
1226                 unsigned int i = 0, bypass = 0;
1227                 int err;
1228
1229                 /* no possible 'owned_head' equals the following */
1230                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1231                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1232
1233                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1234
1235                 clusterpages = BIT(pcl->clusterbits);
1236
1237                 /* close the main owned chain at first */
1238                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1239                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1240
1241                 first_index = pcl->obj.index;
1242                 force_submit |= (first_index != last_index + 1);
1243
1244 repeat:
1245                 page = pickup_page_for_submission(pcl, i, pagepool,
1246                                                   MNGD_MAPPING(sbi),
1247                                                   GFP_NOFS);
1248                 if (!page) {
1249                         force_submit = true;
1250                         ++bypass;
1251                         goto skippage;
1252                 }
1253
1254                 if (bio && force_submit) {
1255 submit_bio_retry:
1256                         submit_bio(bio);
1257                         bio = NULL;
1258                 }
1259
1260                 if (!bio) {
1261                         bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1262
1263                         bio->bi_end_io = z_erofs_vle_read_endio;
1264                         bio_set_dev(bio, sb->s_bdev);
1265                         bio->bi_iter.bi_sector = (sector_t)(first_index + i) <<
1266                                 LOG_SECTORS_PER_BLOCK;
1267                         bio->bi_private = bi_private;
1268                         bio->bi_opf = REQ_OP_READ;
1269
1270                         ++nr_bios;
1271                 }
1272
1273                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1274                 if (err < PAGE_SIZE)
1275                         goto submit_bio_retry;
1276
1277                 force_submit = false;
1278                 last_index = first_index + i;
1279 skippage:
1280                 if (++i < clusterpages)
1281                         goto repeat;
1282
1283                 if (bypass < clusterpages)
1284                         qtail[JQ_SUBMIT] = &pcl->next;
1285                 else
1286                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1287         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1288
1289         if (bio)
1290                 submit_bio(bio);
1291
1292         if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1293                 return true;
1294
1295         z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1296         return true;
1297 }
1298
1299 static void z_erofs_submit_and_unzip(struct super_block *sb,
1300                                      struct z_erofs_collector *clt,
1301                                      struct list_head *pagepool,
1302                                      bool force_fg)
1303 {
1304         struct z_erofs_unzip_io io[NR_JOBQUEUES];
1305
1306         if (!z_erofs_vle_submit_all(sb, clt->owned_head,
1307                                     pagepool, io, force_fg))
1308                 return;
1309
1310         /* decompress no I/O pclusters immediately */
1311         z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1312
1313         if (!force_fg)
1314                 return;
1315
1316         /* wait until all bios are completed */
1317         wait_event(io[JQ_SUBMIT].u.wait,
1318                    !atomic_read(&io[JQ_SUBMIT].pending_bios));
1319
1320         /* let's synchronous decompression */
1321         z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1322 }
1323
1324 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1325                                              struct page *page)
1326 {
1327         struct inode *const inode = page->mapping->host;
1328         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1329         int err;
1330         LIST_HEAD(pagepool);
1331
1332         trace_erofs_readpage(page, false);
1333
1334         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1335
1336         err = z_erofs_do_read_page(&f, page, &pagepool);
1337         (void)z_erofs_collector_end(&f.clt);
1338
1339         /* if some compressed cluster ready, need submit them anyway */
1340         z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, true);
1341
1342         if (err)
1343                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1344
1345         if (f.map.mpage)
1346                 put_page(f.map.mpage);
1347
1348         /* clean up the remaining free pages */
1349         put_pages_list(&pagepool);
1350         return err;
1351 }
1352
1353 static bool should_decompress_synchronously(struct erofs_sb_info *sbi,
1354                                             unsigned int nr)
1355 {
1356         return nr <= sbi->max_sync_decompress_pages;
1357 }
1358
1359 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1360                                               struct address_space *mapping,
1361                                               struct list_head *pages,
1362                                               unsigned int nr_pages)
1363 {
1364         struct inode *const inode = mapping->host;
1365         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1366
1367         bool sync = should_decompress_synchronously(sbi, nr_pages);
1368         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1369         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1370         struct page *head = NULL;
1371         LIST_HEAD(pagepool);
1372
1373         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1374                               nr_pages, false);
1375
1376         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1377
1378         for (; nr_pages; --nr_pages) {
1379                 struct page *page = lru_to_page(pages);
1380
1381                 prefetchw(&page->flags);
1382                 list_del(&page->lru);
1383
1384                 /*
1385                  * A pure asynchronous readahead is indicated if
1386                  * a PG_readahead marked page is hitted at first.
1387                  * Let's also do asynchronous decompression for this case.
1388                  */
1389                 sync &= !(PageReadahead(page) && !head);
1390
1391                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1392                         list_add(&page->lru, &pagepool);
1393                         continue;
1394                 }
1395
1396                 set_page_private(page, (unsigned long)head);
1397                 head = page;
1398         }
1399
1400         while (head) {
1401                 struct page *page = head;
1402                 int err;
1403
1404                 /* traversal in reverse order */
1405                 head = (void *)page_private(page);
1406
1407                 err = z_erofs_do_read_page(&f, page, &pagepool);
1408                 if (err)
1409                         erofs_err(inode->i_sb,
1410                                   "readahead error at page %lu @ nid %llu",
1411                                   page->index, EROFS_I(inode)->nid);
1412                 put_page(page);
1413         }
1414
1415         (void)z_erofs_collector_end(&f.clt);
1416
1417         z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, sync);
1418
1419         if (f.map.mpage)
1420                 put_page(f.map.mpage);
1421
1422         /* clean up the remaining free pages */
1423         put_pages_list(&pagepool);
1424         return 0;
1425 }
1426
1427 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1428         .readpage = z_erofs_vle_normalaccess_readpage,
1429         .readpages = z_erofs_vle_normalaccess_readpages,
1430 };
1431