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, enum z_erofs_page_type type,
 292                                bool pvec_safereuse)
 293 {
 294         int ret;
 295
 296         /* give priority for inplaceio */
 297         if (clt->mode >= COLLECT_PRIMARY &&
 298             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 299             z_erofs_try_inplace_io(clt, page))
 300                 return 0;
 301
 302         ret = z_erofs_pagevec_enqueue(&clt->vector, page, type,
 303                                       pvec_safereuse);
 304         clt->cl->vcnt += (unsigned int)ret;
 305         return ret ? 0 : -EAGAIN;
 306 }
 307
 308 static enum z_erofs_collectmode
 309 try_to_claim_pcluster(struct z_erofs_pcluster *pcl,
 310                       z_erofs_next_pcluster_t *owned_head)
 311 {
 312         /* let's claim these following types of pclusters */
 313 retry:
 314         if (pcl->next == Z_EROFS_PCLUSTER_NIL) {
 315                 /* type 1, nil pcluster */
 316                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 317                             *owned_head) != Z_EROFS_PCLUSTER_NIL)
 318                         goto retry;
 319
 320                 *owned_head = &pcl->next;
 321                 /* lucky, I am the followee :) */
 322                 return COLLECT_PRIMARY_FOLLOWED;
 323         } else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) {
 324                 /*
 325                  * type 2, link to the end of a existing open chain,
 326                  * be careful that its submission itself is governed
 327                  * by the original owned chain.
 328                  */
 329                 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 330                             *owned_head) != Z_EROFS_PCLUSTER_TAIL)
 331                         goto retry;
 332                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 333                 return COLLECT_PRIMARY_HOOKED;
 334         }
 335         return COLLECT_PRIMARY; /* :( better luck next time */
 336 }
 337
 338 static struct z_erofs_collection *cllookup(struct z_erofs_collector *clt,
 339                                            struct inode *inode,
 340                                            struct erofs_map_blocks *map)
 341 {
 342         struct erofs_workgroup *grp;
 343         struct z_erofs_pcluster *pcl;
 344         struct z_erofs_collection *cl;
 345         unsigned int length;
 346         bool tag;
 347
 348         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT, &tag);
 349         if (!grp)
 350                 return NULL;
 351
 352         pcl = container_of(grp, struct z_erofs_pcluster, obj);
 353         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 354                 DBG_BUGON(1);
 355                 erofs_workgroup_put(grp);
 356                 return ERR_PTR(-EFSCORRUPTED);
 357         }
 358
 359         cl = z_erofs_primarycollection(pcl);
 360         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 361                 DBG_BUGON(1);
 362                 erofs_workgroup_put(grp);
 363                 return ERR_PTR(-EFSCORRUPTED);
 364         }
 365
 366         length = READ_ONCE(pcl->length);
 367         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 368                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 369                         DBG_BUGON(1);
 370                         erofs_workgroup_put(grp);
 371                         return ERR_PTR(-EFSCORRUPTED);
 372                 }
 373         } else {
 374                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 375
 376                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 377                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 378
 379                 while (llen > length &&
 380                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 381                         cpu_relax();
 382                         length = READ_ONCE(pcl->length);
 383                 }
 384         }
 385         mutex_lock(&cl->lock);
 386         /* used to check tail merging loop due to corrupted images */
 387         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 388                 clt->tailpcl = pcl;
 389         clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head);
 390         /* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */
 391         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 392                 clt->tailpcl = NULL;
 393         clt->pcl = pcl;
 394         clt->cl = cl;
 395         return cl;
 396 }
 397
 398 static struct z_erofs_collection *clregister(struct z_erofs_collector *clt,
 399                                              struct inode *inode,
 400                                              struct erofs_map_blocks *map)
 401 {
 402         struct z_erofs_pcluster *pcl;
 403         struct z_erofs_collection *cl;
 404         int err;
 405
 406         /* no available workgroup, let's allocate one */
 407         pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS);
 408         if (!pcl)
 409                 return ERR_PTR(-ENOMEM);
 410
 411         z_erofs_pcluster_init_always(pcl);
 412         pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 413
 414         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 415                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 416                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 417
 418         if (map->m_flags & EROFS_MAP_ZIPPED)
 419                 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
 420         else
 421                 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
 422
 423         pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0];
 424         pcl->clusterbits -= PAGE_SHIFT;
 425
 426         /* new pclusters should be claimed as type 1, primary and followed */
 427         pcl->next = clt->owned_head;
 428         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 429
 430         cl = z_erofs_primarycollection(pcl);
 431         cl->pageofs = map->m_la & ~PAGE_MASK;
 432
 433         /*
 434          * lock all primary followed works before visible to others
 435          * and mutex_trylock *never* fails for a new pcluster.
 436          */
 437         mutex_trylock(&cl->lock);
 438
 439         err = erofs_register_workgroup(inode->i_sb, &pcl->obj, 0);
 440         if (err) {
 441                 mutex_unlock(&cl->lock);
 442                 kmem_cache_free(pcluster_cachep, pcl);
 443                 return ERR_PTR(-EAGAIN);
 444         }
 445         /* used to check tail merging loop due to corrupted images */
 446         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 447                 clt->tailpcl = pcl;
 448         clt->owned_head = &pcl->next;
 449         clt->pcl = pcl;
 450         clt->cl = cl;
 451         return cl;
 452 }
 453
 454 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
 455                                    struct inode *inode,
 456                                    struct erofs_map_blocks *map)
 457 {
 458         struct z_erofs_collection *cl;
 459
 460         DBG_BUGON(clt->cl);
 461
 462         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
 463         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
 464         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 465
 466         if (!PAGE_ALIGNED(map->m_pa)) {
 467                 DBG_BUGON(1);
 468                 return -EINVAL;
 469         }
 470
 471 repeat:
 472         cl = cllookup(clt, inode, map);
 473         if (!cl) {
 474                 cl = clregister(clt, inode, map);
 475
 476                 if (cl == ERR_PTR(-EAGAIN))
 477                         goto repeat;
 478         }
 479
 480         if (IS_ERR(cl))
 481                 return PTR_ERR(cl);
 482
 483         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 484                                   cl->pagevec, cl->vcnt);
 485
 486         clt->compressedpages = clt->pcl->compressed_pages;
 487         if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */
 488                 clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES;
 489         return 0;
 490 }
 491
 492 /*
 493  * keep in mind that no referenced pclusters will be freed
 494  * only after a RCU grace period.
 495  */
 496 static void z_erofs_rcu_callback(struct rcu_head *head)
 497 {
 498         struct z_erofs_collection *const cl =
 499                 container_of(head, struct z_erofs_collection, rcu);
 500
 501         kmem_cache_free(pcluster_cachep,
 502                         container_of(cl, struct z_erofs_pcluster,
 503                                      primary_collection));
 504 }
 505
 506 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 507 {
 508         struct z_erofs_pcluster *const pcl =
 509                 container_of(grp, struct z_erofs_pcluster, obj);
 510         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 511
 512         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 513 }
 514
 515 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 516 {
 517         struct z_erofs_pcluster *const pcl =
 518                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 519
 520         erofs_workgroup_put(&pcl->obj);
 521 }
 522
 523 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 524 {
 525         struct z_erofs_collection *cl = clt->cl;
 526
 527         if (!cl)
 528                 return false;
 529
 530         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 531         mutex_unlock(&cl->lock);
 532
 533         /*
 534          * if all pending pages are added, don't hold its reference
 535          * any longer if the pcluster isn't hosted by ourselves.
 536          */
 537         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 538                 z_erofs_collection_put(cl);
 539
 540         clt->cl = NULL;
 541         return true;
 542 }
 543
 544 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 545                                                gfp_t gfp)
 546 {
 547         struct page *page = erofs_allocpage(pagepool, gfp, true);
 548
 549         page->mapping = Z_EROFS_MAPPING_STAGING;
 550         return page;
 551 }
 552
 553 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 554                                        unsigned int cachestrategy,
 555                                        erofs_off_t la)
 556 {
 557         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 558                 return false;
 559
 560         if (fe->backmost)
 561                 return true;
 562
 563         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 564                 la < fe->headoffset;
 565 }
 566
 567 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 568                                 struct page *page,
 569                                 struct list_head *pagepool)
 570 {
 571         struct inode *const inode = fe->inode;
 572         struct erofs_sb_info *const sbi __maybe_unused = EROFS_I_SB(inode);
 573         struct erofs_map_blocks *const map = &fe->map;
 574         struct z_erofs_collector *const clt = &fe->clt;
 575         const loff_t offset = page_offset(page);
 576         bool tight = true;
 577
 578         enum z_erofs_cache_alloctype cache_strategy;
 579         enum z_erofs_page_type page_type;
 580         unsigned int cur, end, spiltted, index;
 581         int err = 0;
 582
 583         /* register locked file pages as online pages in pack */
 584         z_erofs_onlinepage_init(page);
 585
 586         spiltted = 0;
 587         end = PAGE_SIZE;
 588 repeat:
 589         cur = end - 1;
 590
 591         /* lucky, within the range of the current map_blocks */
 592         if (offset + cur >= map->m_la &&
 593             offset + cur < map->m_la + map->m_llen) {
 594                 /* didn't get a valid collection previously (very rare) */
 595                 if (!clt->cl)
 596                         goto restart_now;
 597                 goto hitted;
 598         }
 599
 600         /* go ahead the next map_blocks */
 601         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 602
 603         if (z_erofs_collector_end(clt))
 604                 fe->backmost = false;
 605
 606         map->m_la = offset + cur;
 607         map->m_llen = 0;
 608         err = z_erofs_map_blocks_iter(inode, map, 0);
 609         if (err)
 610                 goto err_out;
 611
 612 restart_now:
 613         if (!(map->m_flags & EROFS_MAP_MAPPED))
 614                 goto hitted;
 615
 616         err = z_erofs_collector_begin(clt, inode, map);
 617         if (err)
 618                 goto err_out;
 619
 620         /* preload all compressed pages (maybe downgrade role if necessary) */
 621         if (should_alloc_managed_pages(fe, sbi->cache_strategy, map->m_la))
 622                 cache_strategy = DELAYEDALLOC;
 623         else
 624                 cache_strategy = DONTALLOC;
 625
 626         preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 627                                  cache_strategy, pagepool);
 628
 629 hitted:
 630         /*
 631          * Ensure the current partial page belongs to this submit chain rather
 632          * than other concurrent submit chains or the noio(bypass) chain since
 633          * those chains are handled asynchronously thus the page cannot be used
 634          * for inplace I/O or pagevec (should be processed in strict order.)
 635          */
 636         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 637                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 638
 639         cur = end - min_t(erofs_off_t, offset + end - map->m_la, end);
 640         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 641                 zero_user_segment(page, cur, end);
 642                 ++spiltted;
 643                 tight = false;
 644                 goto next_part;
 645         }
 646
 647         /* let's derive page type */
 648         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 649                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 650                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 651                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 652
 653         if (cur)
 654                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 655
 656 retry:
 657         err = z_erofs_attach_page(clt, page, page_type,
 658                                   clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 659         /* should allocate an additional staging page for pagevec */
 660         if (err == -EAGAIN) {
 661                 struct page *const newpage =
 662                         __stagingpage_alloc(pagepool, GFP_NOFS);
 663
 664                 err = z_erofs_attach_page(clt, newpage,
 665                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
 666                 if (!err)
 667                         goto retry;
 668         }
 669
 670         if (err)
 671                 goto err_out;
 672
 673         index = page->index - (map->m_la >> PAGE_SHIFT);
 674
 675         z_erofs_onlinepage_fixup(page, index, true);
 676
 677         /* bump up the number of spiltted parts of a page */
 678         ++spiltted;
 679         /* also update nr_pages */
 680         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 681 next_part:
 682         /* can be used for verification */
 683         map->m_llen = offset + cur - map->m_la;
 684
 685         end = cur;
 686         if (end > 0)
 687                 goto repeat;
 688
 689 out:
 690         z_erofs_onlinepage_endio(page);
 691
 692         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 693                   __func__, page, spiltted, map->m_llen);
 694         return err;
 695
 696         /* if some error occurred while processing this page */
 697 err_out:
 698         SetPageError(page);
 699         goto out;
 700 }
 701
 702 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 703 {
 704         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 705         struct z_erofs_unzip_io *io = tagptr_unfold_ptr(t);
 706         bool background = tagptr_unfold_tags(t);
 707
 708         if (!background) {
 709                 unsigned long flags;
 710
 711                 spin_lock_irqsave(&io->u.wait.lock, flags);
 712                 if (!atomic_add_return(bios, &io->pending_bios))
 713                         wake_up_locked(&io->u.wait);
 714                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 715                 return;
 716         }
 717
 718         if (!atomic_add_return(bios, &io->pending_bios))
 719                 queue_work(z_erofs_workqueue, &io->u.work);
 720 }
 721
 722 static inline void z_erofs_vle_read_endio(struct bio *bio)
 723 {
 724         struct erofs_sb_info *sbi = NULL;
 725         blk_status_t err = bio->bi_status;
 726         struct bio_vec *bvec;
 727         struct bvec_iter_all iter_all;
 728
 729         bio_for_each_segment_all(bvec, bio, iter_all) {
 730                 struct page *page = bvec->bv_page;
 731                 bool cachemngd = false;
 732
 733                 DBG_BUGON(PageUptodate(page));
 734                 DBG_BUGON(!page->mapping);
 735
 736                 if (!sbi && !z_erofs_page_is_staging(page))
 737                         sbi = EROFS_SB(page->mapping->host->i_sb);
 738
 739                 /* sbi should already be gotten if the page is managed */
 740                 if (sbi)
 741                         cachemngd = erofs_page_is_managed(sbi, page);
 742
 743                 if (err)
 744                         SetPageError(page);
 745                 else if (cachemngd)
 746                         SetPageUptodate(page);
 747
 748                 if (cachemngd)
 749                         unlock_page(page);
 750         }
 751
 752         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 753         bio_put(bio);
 754 }
 755
 756 static int z_erofs_decompress_pcluster(struct super_block *sb,
 757                                        struct z_erofs_pcluster *pcl,
 758                                        struct list_head *pagepool)
 759 {
 760         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 761         const unsigned int clusterpages = BIT(pcl->clusterbits);
 762         struct z_erofs_pagevec_ctor ctor;
 763         unsigned int i, outputsize, llen, nr_pages;
 764         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 765         struct page **pages, **compressed_pages, *page;
 766
 767         enum z_erofs_page_type page_type;
 768         bool overlapped, partial;
 769         struct z_erofs_collection *cl;
 770         int err;
 771
 772         might_sleep();
 773         cl = z_erofs_primarycollection(pcl);
 774         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 775
 776         mutex_lock(&cl->lock);
 777         nr_pages = cl->nr_pages;
 778
 779         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 780                 pages = pages_onstack;
 781         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 782                    mutex_trylock(&z_pagemap_global_lock)) {
 783                 pages = z_pagemap_global;
 784         } else {
 785                 gfp_t gfp_flags = GFP_KERNEL;
 786
 787                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 788                         gfp_flags |= __GFP_NOFAIL;
 789
 790                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 791                                        gfp_flags);
 792
 793                 /* fallback to global pagemap for the lowmem scenario */
 794                 if (!pages) {
 795                         mutex_lock(&z_pagemap_global_lock);
 796                         pages = z_pagemap_global;
 797                 }
 798         }
 799
 800         for (i = 0; i < nr_pages; ++i)
 801                 pages[i] = NULL;
 802
 803         err = 0;
 804         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 805                                   cl->pagevec, 0);
 806
 807         for (i = 0; i < cl->vcnt; ++i) {
 808                 unsigned int pagenr;
 809
 810                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 811
 812                 /* all pages in pagevec ought to be valid */
 813                 DBG_BUGON(!page);
 814                 DBG_BUGON(!page->mapping);
 815
 816                 if (z_erofs_put_stagingpage(pagepool, page))
 817                         continue;
 818
 819                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 820                         pagenr = 0;
 821                 else
 822                         pagenr = z_erofs_onlinepage_index(page);
 823
 824                 DBG_BUGON(pagenr >= nr_pages);
 825
 826                 /*
 827                  * currently EROFS doesn't support multiref(dedup),
 828                  * so here erroring out one multiref page.
 829                  */
 830                 if (pages[pagenr]) {
 831                         DBG_BUGON(1);
 832                         SetPageError(pages[pagenr]);
 833                         z_erofs_onlinepage_endio(pages[pagenr]);
 834                         err = -EFSCORRUPTED;
 835                 }
 836                 pages[pagenr] = page;
 837         }
 838         z_erofs_pagevec_ctor_exit(&ctor, true);
 839
 840         overlapped = false;
 841         compressed_pages = pcl->compressed_pages;
 842
 843         for (i = 0; i < clusterpages; ++i) {
 844                 unsigned int pagenr;
 845
 846                 page = compressed_pages[i];
 847
 848                 /* all compressed pages ought to be valid */
 849                 DBG_BUGON(!page);
 850                 DBG_BUGON(!page->mapping);
 851
 852                 if (!z_erofs_page_is_staging(page)) {
 853                         if (erofs_page_is_managed(sbi, page)) {
 854                                 if (!PageUptodate(page))
 855                                         err = -EIO;
 856                                 continue;
 857                         }
 858
 859                         /*
 860                          * only if non-head page can be selected
 861                          * for inplace decompression
 862                          */
 863                         pagenr = z_erofs_onlinepage_index(page);
 864
 865                         DBG_BUGON(pagenr >= nr_pages);
 866                         if (pages[pagenr]) {
 867                                 DBG_BUGON(1);
 868                                 SetPageError(pages[pagenr]);
 869                                 z_erofs_onlinepage_endio(pages[pagenr]);
 870                                 err = -EFSCORRUPTED;
 871                         }
 872                         pages[pagenr] = page;
 873
 874                         overlapped = true;
 875                 }
 876
 877                 /* PG_error needs checking for inplaced and staging pages */
 878                 if (PageError(page)) {
 879                         DBG_BUGON(PageUptodate(page));
 880                         err = -EIO;
 881                 }
 882         }
 883
 884         if (err)
 885                 goto out;
 886
 887         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 888         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 889                 outputsize = llen;
 890                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 891         } else {
 892                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 893                 partial = true;
 894         }
 895
 896         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 897                                         .sb = sb,
 898                                         .in = compressed_pages,
 899                                         .out = pages,
 900                                         .pageofs_out = cl->pageofs,
 901                                         .inputsize = PAGE_SIZE,
 902                                         .outputsize = outputsize,
 903                                         .alg = pcl->algorithmformat,
 904                                         .inplace_io = overlapped,
 905                                         .partial_decoding = partial
 906                                  }, pagepool);
 907
 908 out:
 909         /* must handle all compressed pages before endding pages */
 910         for (i = 0; i < clusterpages; ++i) {
 911                 page = compressed_pages[i];
 912
 913                 if (erofs_page_is_managed(sbi, page))
 914                         continue;
 915
 916                 /* recycle all individual staging pages */
 917                 (void)z_erofs_put_stagingpage(pagepool, page);
 918
 919                 WRITE_ONCE(compressed_pages[i], NULL);
 920         }
 921
 922         for (i = 0; i < nr_pages; ++i) {
 923                 page = pages[i];
 924                 if (!page)
 925                         continue;
 926
 927                 DBG_BUGON(!page->mapping);
 928
 929                 /* recycle all individual staging pages */
 930                 if (z_erofs_put_stagingpage(pagepool, page))
 931                         continue;
 932
 933                 if (err < 0)
 934                         SetPageError(page);
 935
 936                 z_erofs_onlinepage_endio(page);
 937         }
 938
 939         if (pages == z_pagemap_global)
 940                 mutex_unlock(&z_pagemap_global_lock);
 941         else if (pages != pages_onstack)
 942                 kvfree(pages);
 943
 944         cl->nr_pages = 0;
 945         cl->vcnt = 0;
 946
 947         /* all cl locks MUST be taken before the following line */
 948         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
 949
 950         /* all cl locks SHOULD be released right now */
 951         mutex_unlock(&cl->lock);
 952
 953         z_erofs_collection_put(cl);
 954         return err;
 955 }
 956
 957 static void z_erofs_vle_unzip_all(struct super_block *sb,
 958                                   struct z_erofs_unzip_io *io,
 959                                   struct list_head *pagepool)
 960 {
 961         z_erofs_next_pcluster_t owned = io->head;
 962
 963         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
 964                 struct z_erofs_pcluster *pcl;
 965
 966                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
 967                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
 968
 969                 /* no possible that 'owned' equals NULL */
 970                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
 971
 972                 pcl = container_of(owned, struct z_erofs_pcluster, next);
 973                 owned = READ_ONCE(pcl->next);
 974
 975                 z_erofs_decompress_pcluster(sb, pcl, pagepool);
 976         }
 977 }
 978
 979 static void z_erofs_vle_unzip_wq(struct work_struct *work)
 980 {
 981         struct z_erofs_unzip_io_sb *iosb =
 982                 container_of(work, struct z_erofs_unzip_io_sb, io.u.work);
 983         LIST_HEAD(pagepool);
 984
 985         DBG_BUGON(iosb->io.head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 986         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &pagepool);
 987
 988         put_pages_list(&pagepool);
 989         kvfree(iosb);
 990 }
 991
 992 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
 993                                                unsigned int nr,
 994                                                struct list_head *pagepool,
 995                                                struct address_space *mc,
 996                                                gfp_t gfp)
 997 {
 998         /* determined at compile time to avoid too many #ifdefs */
 999         const bool nocache = __builtin_constant_p(mc) ? !mc : false;
1000         const pgoff_t index = pcl->obj.index;
1001         bool tocache = false;
1002
1003         struct address_space *mapping;
1004         struct page *oldpage, *page;
1005
1006         compressed_page_t t;
1007         int justfound;
1008
1009 repeat:
1010         page = READ_ONCE(pcl->compressed_pages[nr]);
1011         oldpage = page;
1012
1013         if (!page)
1014                 goto out_allocpage;
1015
1016         /*
1017          * the cached page has not been allocated and
1018          * an placeholder is out there, prepare it now.
1019          */
1020         if (!nocache && page == PAGE_UNALLOCATED) {
1021                 tocache = true;
1022                 goto out_allocpage;
1023         }
1024
1025         /* process the target tagged pointer */
1026         t = tagptr_init(compressed_page_t, page);
1027         justfound = tagptr_unfold_tags(t);
1028         page = tagptr_unfold_ptr(t);
1029
1030         mapping = READ_ONCE(page->mapping);
1031
1032         /*
1033          * if managed cache is disabled, it's no way to
1034          * get such a cached-like page.
1035          */
1036         if (nocache) {
1037                 /* if managed cache is disabled, it is impossible `justfound' */
1038                 DBG_BUGON(justfound);
1039
1040                 /* and it should be locked, not uptodate, and not truncated */
1041                 DBG_BUGON(!PageLocked(page));
1042                 DBG_BUGON(PageUptodate(page));
1043                 DBG_BUGON(!mapping);
1044                 goto out;
1045         }
1046
1047         /*
1048          * unmanaged (file) pages are all locked solidly,
1049          * therefore it is impossible for `mapping' to be NULL.
1050          */
1051         if (mapping && mapping != mc)
1052                 /* ought to be unmanaged pages */
1053                 goto out;
1054
1055         lock_page(page);
1056
1057         /* only true if page reclaim goes wrong, should never happen */
1058         DBG_BUGON(justfound && PagePrivate(page));
1059
1060         /* the page is still in manage cache */
1061         if (page->mapping == mc) {
1062                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1063
1064                 ClearPageError(page);
1065                 if (!PagePrivate(page)) {
1066                         /*
1067                          * impossible to be !PagePrivate(page) for
1068                          * the current restriction as well if
1069                          * the page is already in compressed_pages[].
1070                          */
1071                         DBG_BUGON(!justfound);
1072
1073                         justfound = 0;
1074                         set_page_private(page, (unsigned long)pcl);
1075                         SetPagePrivate(page);
1076                 }
1077
1078                 /* no need to submit io if it is already up-to-date */
1079                 if (PageUptodate(page)) {
1080                         unlock_page(page);
1081                         page = NULL;
1082                 }
1083                 goto out;
1084         }
1085
1086         /*
1087          * the managed page has been truncated, it's unsafe to
1088          * reuse this one, let's allocate a new cache-managed page.
1089          */
1090         DBG_BUGON(page->mapping);
1091         DBG_BUGON(!justfound);
1092
1093         tocache = true;
1094         unlock_page(page);
1095         put_page(page);
1096 out_allocpage:
1097         page = __stagingpage_alloc(pagepool, gfp);
1098         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1099                 list_add(&page->lru, pagepool);
1100                 cpu_relax();
1101                 goto repeat;
1102         }
1103         if (nocache || !tocache)
1104                 goto out;
1105         if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1106                 page->mapping = Z_EROFS_MAPPING_STAGING;
1107                 goto out;
1108         }
1109
1110         set_page_private(page, (unsigned long)pcl);
1111         SetPagePrivate(page);
1112 out:    /* the only exit (for tracing and debugging) */
1113         return page;
1114 }
1115
1116 static struct z_erofs_unzip_io *jobqueue_init(struct super_block *sb,
1117                                               struct z_erofs_unzip_io *io,
1118                                               bool foreground)
1119 {
1120         struct z_erofs_unzip_io_sb *iosb;
1121
1122         if (foreground) {
1123                 /* waitqueue available for foreground io */
1124                 DBG_BUGON(!io);
1125
1126                 init_waitqueue_head(&io->u.wait);
1127                 atomic_set(&io->pending_bios, 0);
1128                 goto out;
1129         }
1130
1131         iosb = kvzalloc(sizeof(*iosb), GFP_KERNEL | __GFP_NOFAIL);
1132         DBG_BUGON(!iosb);
1133
1134         /* initialize fields in the allocated descriptor */
1135         io = &iosb->io;
1136         iosb->sb = sb;
1137         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1138 out:
1139         io->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1140         return io;
1141 }
1142
1143 /* define decompression jobqueue types */
1144 enum {
1145         JQ_BYPASS,
1146         JQ_SUBMIT,
1147         NR_JOBQUEUES,
1148 };
1149
1150 static void *jobqueueset_init(struct super_block *sb,
1151                               z_erofs_next_pcluster_t qtail[],
1152                               struct z_erofs_unzip_io *q[],
1153                               struct z_erofs_unzip_io *fgq,
1154                               bool forcefg)
1155 {
1156         /*
1157          * if managed cache is enabled, bypass jobqueue is needed,
1158          * no need to read from device for all pclusters in this queue.
1159          */
1160         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1161         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1162
1163         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1164         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1165
1166         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1167 }
1168
1169 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1170                                     z_erofs_next_pcluster_t qtail[],
1171                                     z_erofs_next_pcluster_t owned_head)
1172 {
1173         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1174         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1175
1176         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1177         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1178                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1179
1180         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1181
1182         WRITE_ONCE(*submit_qtail, owned_head);
1183         WRITE_ONCE(*bypass_qtail, &pcl->next);
1184
1185         qtail[JQ_BYPASS] = &pcl->next;
1186 }
1187
1188 static bool postsubmit_is_all_bypassed(struct z_erofs_unzip_io *q[],
1189                                        unsigned int nr_bios,
1190                                        bool force_fg)
1191 {
1192         /*
1193          * although background is preferred, no one is pending for submission.
1194          * don't issue workqueue for decompression but drop it directly instead.
1195          */
1196         if (force_fg || nr_bios)
1197                 return false;
1198
1199         kvfree(container_of(q[JQ_SUBMIT], struct z_erofs_unzip_io_sb, io));
1200         return true;
1201 }
1202
1203 static bool z_erofs_vle_submit_all(struct super_block *sb,
1204                                    z_erofs_next_pcluster_t owned_head,
1205                                    struct list_head *pagepool,
1206                                    struct z_erofs_unzip_io *fgq,
1207                                    bool force_fg)
1208 {
1209         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
1210         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1211         struct z_erofs_unzip_io *q[NR_JOBQUEUES];
1212         struct bio *bio;
1213         void *bi_private;
1214         /* since bio will be NULL, no need to initialize last_index */
1215         pgoff_t uninitialized_var(last_index);
1216         bool force_submit = false;
1217         unsigned int nr_bios;
1218
1219         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1220                 return false;
1221
1222         force_submit = false;
1223         bio = NULL;
1224         nr_bios = 0;
1225         bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1226
1227         /* by default, all need io submission */
1228         q[JQ_SUBMIT]->head = owned_head;
1229
1230         do {
1231                 struct z_erofs_pcluster *pcl;
1232                 unsigned int clusterpages;
1233                 pgoff_t first_index;
1234                 struct page *page;
1235                 unsigned int i = 0, bypass = 0;
1236                 int err;
1237
1238                 /* no possible 'owned_head' equals the following */
1239                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1240                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1241
1242                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1243
1244                 clusterpages = BIT(pcl->clusterbits);
1245
1246                 /* close the main owned chain at first */
1247                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1248                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1249
1250                 first_index = pcl->obj.index;
1251                 force_submit |= (first_index != last_index + 1);
1252
1253 repeat:
1254                 page = pickup_page_for_submission(pcl, i, pagepool,
1255                                                   MNGD_MAPPING(sbi),
1256                                                   GFP_NOFS);
1257                 if (!page) {
1258                         force_submit = true;
1259                         ++bypass;
1260                         goto skippage;
1261                 }
1262
1263                 if (bio && force_submit) {
1264 submit_bio_retry:
1265                         submit_bio(bio);
1266                         bio = NULL;
1267                 }
1268
1269                 if (!bio) {
1270                         bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES);
1271
1272                         bio->bi_end_io = z_erofs_vle_read_endio;
1273                         bio_set_dev(bio, sb->s_bdev);
1274                         bio->bi_iter.bi_sector = (sector_t)(first_index + i) <<
1275                                 LOG_SECTORS_PER_BLOCK;
1276                         bio->bi_private = bi_private;
1277                         bio->bi_opf = REQ_OP_READ;
1278
1279                         ++nr_bios;
1280                 }
1281
1282                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1283                 if (err < PAGE_SIZE)
1284                         goto submit_bio_retry;
1285
1286                 force_submit = false;
1287                 last_index = first_index + i;
1288 skippage:
1289                 if (++i < clusterpages)
1290                         goto repeat;
1291
1292                 if (bypass < clusterpages)
1293                         qtail[JQ_SUBMIT] = &pcl->next;
1294                 else
1295                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1296         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1297
1298         if (bio)
1299                 submit_bio(bio);
1300
1301         if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1302                 return true;
1303
1304         z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1305         return true;
1306 }
1307
1308 static void z_erofs_submit_and_unzip(struct super_block *sb,
1309                                      struct z_erofs_collector *clt,
1310                                      struct list_head *pagepool,
1311                                      bool force_fg)
1312 {
1313         struct z_erofs_unzip_io io[NR_JOBQUEUES];
1314
1315         if (!z_erofs_vle_submit_all(sb, clt->owned_head,
1316                                     pagepool, io, force_fg))
1317                 return;
1318
1319         /* decompress no I/O pclusters immediately */
1320         z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1321
1322         if (!force_fg)
1323                 return;
1324
1325         /* wait until all bios are completed */
1326         wait_event(io[JQ_SUBMIT].u.wait,
1327                    !atomic_read(&io[JQ_SUBMIT].pending_bios));
1328
1329         /* let's synchronous decompression */
1330         z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1331 }
1332
1333 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1334                                              struct page *page)
1335 {
1336         struct inode *const inode = page->mapping->host;
1337         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1338         int err;
1339         LIST_HEAD(pagepool);
1340
1341         trace_erofs_readpage(page, false);
1342
1343         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1344
1345         err = z_erofs_do_read_page(&f, page, &pagepool);
1346         (void)z_erofs_collector_end(&f.clt);
1347
1348         /* if some compressed cluster ready, need submit them anyway */
1349         z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, true);
1350
1351         if (err)
1352                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1353
1354         if (f.map.mpage)
1355                 put_page(f.map.mpage);
1356
1357         /* clean up the remaining free pages */
1358         put_pages_list(&pagepool);
1359         return err;
1360 }
1361
1362 static bool should_decompress_synchronously(struct erofs_sb_info *sbi,
1363                                             unsigned int nr)
1364 {
1365         return nr <= sbi->max_sync_decompress_pages;
1366 }
1367
1368 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1369                                               struct address_space *mapping,
1370                                               struct list_head *pages,
1371                                               unsigned int nr_pages)
1372 {
1373         struct inode *const inode = mapping->host;
1374         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1375
1376         bool sync = should_decompress_synchronously(sbi, nr_pages);
1377         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1378         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1379         struct page *head = NULL;
1380         LIST_HEAD(pagepool);
1381
1382         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1383                               nr_pages, false);
1384
1385         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1386
1387         for (; nr_pages; --nr_pages) {
1388                 struct page *page = lru_to_page(pages);
1389
1390                 prefetchw(&page->flags);
1391                 list_del(&page->lru);
1392
1393                 /*
1394                  * A pure asynchronous readahead is indicated if
1395                  * a PG_readahead marked page is hitted at first.
1396                  * Let's also do asynchronous decompression for this case.
1397                  */
1398                 sync &= !(PageReadahead(page) && !head);
1399
1400                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1401                         list_add(&page->lru, &pagepool);
1402                         continue;
1403                 }
1404
1405                 set_page_private(page, (unsigned long)head);
1406                 head = page;
1407         }
1408
1409         while (head) {
1410                 struct page *page = head;
1411                 int err;
1412
1413                 /* traversal in reverse order */
1414                 head = (void *)page_private(page);
1415
1416                 err = z_erofs_do_read_page(&f, page, &pagepool);
1417                 if (err)
1418                         erofs_err(inode->i_sb,
1419                                   "readahead error at page %lu @ nid %llu",
1420                                   page->index, EROFS_I(inode)->nid);
1421                 put_page(page);
1422         }
1423
1424         (void)z_erofs_collector_end(&f.clt);
1425
1426         z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, sync);
1427
1428         if (f.map.mpage)
1429                 put_page(f.map.mpage);
1430
1431         /* clean up the remaining free pages */
1432         put_pages_list(&pagepool);
1433         return 0;
1434 }
1435
1436 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1437         .readpage = z_erofs_vle_normalaccess_readpage,
1438         .readpages = z_erofs_vle_normalaccess_readpages,
1439 };
1440