fs/erofs/zdata.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             https://www.huawei.com/
   5  */
   6 #include "zdata.h"
   7 #include "compress.h"
   8 #include <linux/prefetch.h>
   9
  10 #include <trace/events/erofs.h>
  11
  12 /*
  13  * since pclustersize is variable for big pcluster feature, introduce slab
  14  * pools implementation for different pcluster sizes.
  15  */
  16 struct z_erofs_pcluster_slab {
  17         struct kmem_cache *slab;
  18         unsigned int maxpages;
  19         char name[48];
  20 };
  21
  22 #define _PCLP(n) { .maxpages = n }
  23
  24 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
  25         _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
  26         _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
  27 };
  28
  29 static void z_erofs_destroy_pcluster_pool(void)
  30 {
  31         int i;
  32
  33         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  34                 if (!pcluster_pool[i].slab)
  35                         continue;
  36                 kmem_cache_destroy(pcluster_pool[i].slab);
  37                 pcluster_pool[i].slab = NULL;
  38         }
  39 }
  40
  41 static int z_erofs_create_pcluster_pool(void)
  42 {
  43         struct z_erofs_pcluster_slab *pcs;
  44         struct z_erofs_pcluster *a;
  45         unsigned int size;
  46
  47         for (pcs = pcluster_pool;
  48              pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
  49                 size = struct_size(a, compressed_pages, pcs->maxpages);
  50
  51                 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
  52                 pcs->slab = kmem_cache_create(pcs->name, size, 0,
  53                                               SLAB_RECLAIM_ACCOUNT, NULL);
  54                 if (pcs->slab)
  55                         continue;
  56
  57                 z_erofs_destroy_pcluster_pool();
  58                 return -ENOMEM;
  59         }
  60         return 0;
  61 }
  62
  63 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
  64 {
  65         int i;
  66
  67         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  68                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
  69                 struct z_erofs_pcluster *pcl;
  70
  71                 if (nrpages > pcs->maxpages)
  72                         continue;
  73
  74                 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
  75                 if (!pcl)
  76                         return ERR_PTR(-ENOMEM);
  77                 pcl->pclusterpages = nrpages;
  78                 return pcl;
  79         }
  80         return ERR_PTR(-EINVAL);
  81 }
  82
  83 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
  84 {
  85         int i;
  86
  87         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  88                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
  89
  90                 if (pcl->pclusterpages > pcs->maxpages)
  91                         continue;
  92
  93                 kmem_cache_free(pcs->slab, pcl);
  94                 return;
  95         }
  96         DBG_BUGON(1);
  97 }
  98
  99 /*
 100  * a compressed_pages[] placeholder in order to avoid
 101  * being filled with file pages for in-place decompression.
 102  */
 103 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
 104
 105 /* how to allocate cached pages for a pcluster */
 106 enum z_erofs_cache_alloctype {
 107         DONTALLOC,      /* don't allocate any cached pages */
 108         DELAYEDALLOC,   /* delayed allocation (at the time of submitting io) */
 109         /*
 110          * try to use cached I/O if page allocation succeeds or fallback
 111          * to in-place I/O instead to avoid any direct reclaim.
 112          */
 113         TRYALLOC,
 114 };
 115
 116 /*
 117  * tagged pointer with 1-bit tag for all compressed pages
 118  * tag 0 - the page is just found with an extra page reference
 119  */
 120 typedef tagptr1_t compressed_page_t;
 121
 122 #define tag_compressed_page_justfound(page) \
 123         tagptr_fold(compressed_page_t, page, 1)
 124
 125 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 126
 127 void z_erofs_exit_zip_subsystem(void)
 128 {
 129         destroy_workqueue(z_erofs_workqueue);
 130         z_erofs_destroy_pcluster_pool();
 131 }
 132
 133 static inline int z_erofs_init_workqueue(void)
 134 {
 135         const unsigned int onlinecpus = num_possible_cpus();
 136
 137         /*
 138          * no need to spawn too many threads, limiting threads could minimum
 139          * scheduling overhead, perhaps per-CPU threads should be better?
 140          */
 141         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
 142                                             WQ_UNBOUND | WQ_HIGHPRI,
 143                                             onlinecpus + onlinecpus / 4);
 144         return z_erofs_workqueue ? 0 : -ENOMEM;
 145 }
 146
 147 int __init z_erofs_init_zip_subsystem(void)
 148 {
 149         int err = z_erofs_create_pcluster_pool();
 150
 151         if (err)
 152                 return err;
 153         err = z_erofs_init_workqueue();
 154         if (err)
 155                 z_erofs_destroy_pcluster_pool();
 156         return err;
 157 }
 158
 159 enum z_erofs_collectmode {
 160         COLLECT_SECONDARY,
 161         COLLECT_PRIMARY,
 162         /*
 163          * The current collection was the tail of an exist chain, in addition
 164          * that the previous processed chained collections are all decided to
 165          * be hooked up to it.
 166          * A new chain will be created for the remaining collections which are
 167          * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
 168          * the next collection cannot reuse the whole page safely in
 169          * the following scenario:
 170          *  ________________________________________________________________
 171          * |      tail (partial) page     |       head (partial) page       |
 172          * |   (belongs to the next cl)   |   (belongs to the current cl)   |
 173          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 174          */
 175         COLLECT_PRIMARY_HOOKED,
 176         /*
 177          * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
 178          * could be dispatched into bypass queue later due to uptodated managed
 179          * pages. All related online pages cannot be reused for inplace I/O (or
 180          * pagevec) since it can be directly decoded without I/O submission.
 181          */
 182         COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
 183         /*
 184          * The current collection has been linked with the owned chain, and
 185          * could also be linked with the remaining collections, which means
 186          * if the processing page is the tail page of the collection, thus
 187          * the current collection can safely use the whole page (since
 188          * the previous collection is under control) for in-place I/O, as
 189          * illustrated below:
 190          *  ________________________________________________________________
 191          * |  tail (partial) page |          head (partial) page           |
 192          * |  (of the current cl) |      (of the previous collection)      |
 193          * |  PRIMARY_FOLLOWED or |                                        |
 194          * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
 195          *
 196          * [  (*) the above page can be used as inplace I/O.               ]
 197          */
 198         COLLECT_PRIMARY_FOLLOWED,
 199 };
 200
 201 struct z_erofs_collector {
 202         struct z_erofs_pagevec_ctor vector;
 203
 204         struct z_erofs_pcluster *pcl, *tailpcl;
 205         struct z_erofs_collection *cl;
 206         /* a pointer used to pick up inplace I/O pages */
 207         struct page **icpage_ptr;
 208         z_erofs_next_pcluster_t owned_head;
 209
 210         enum z_erofs_collectmode mode;
 211 };
 212
 213 struct z_erofs_decompress_frontend {
 214         struct inode *const inode;
 215
 216         struct z_erofs_collector clt;
 217         struct erofs_map_blocks map;
 218
 219         bool readahead;
 220         /* used for applying cache strategy on the fly */
 221         bool backmost;
 222         erofs_off_t headoffset;
 223 };
 224
 225 #define COLLECTOR_INIT() { \
 226         .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 227         .mode = COLLECT_PRIMARY_FOLLOWED }
 228
 229 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 230         .inode = __i, .clt = COLLECTOR_INIT(), \
 231         .backmost = true, }
 232
 233 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
 234 static DEFINE_MUTEX(z_pagemap_global_lock);
 235
 236 static void preload_compressed_pages(struct z_erofs_collector *clt,
 237                                      struct address_space *mc,
 238                                      enum z_erofs_cache_alloctype type,
 239                                      struct list_head *pagepool)
 240 {
 241         struct z_erofs_pcluster *pcl = clt->pcl;
 242         bool standalone = true;
 243         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 244                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 245         struct page **pages;
 246         pgoff_t index;
 247
 248         if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
 249                 return;
 250
 251         pages = pcl->compressed_pages;
 252         index = pcl->obj.index;
 253         for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
 254                 struct page *page;
 255                 compressed_page_t t;
 256                 struct page *newpage = NULL;
 257
 258                 /* the compressed page was loaded before */
 259                 if (READ_ONCE(*pages))
 260                         continue;
 261
 262                 page = find_get_page(mc, index);
 263
 264                 if (page) {
 265                         t = tag_compressed_page_justfound(page);
 266                 } else {
 267                         /* I/O is needed, no possible to decompress directly */
 268                         standalone = false;
 269                         switch (type) {
 270                         case DELAYEDALLOC:
 271                                 t = tagptr_init(compressed_page_t,
 272                                                 PAGE_UNALLOCATED);
 273                                 break;
 274                         case TRYALLOC:
 275                                 newpage = erofs_allocpage(pagepool, gfp);
 276                                 if (!newpage)
 277                                         continue;
 278                                 set_page_private(newpage,
 279                                                  Z_EROFS_PREALLOCATED_PAGE);
 280                                 t = tag_compressed_page_justfound(newpage);
 281                                 break;
 282                         default:        /* DONTALLOC */
 283                                 continue;
 284                         }
 285                 }
 286
 287                 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
 288                         continue;
 289
 290                 if (page) {
 291                         put_page(page);
 292                 } else if (newpage) {
 293                         set_page_private(newpage, 0);
 294                         list_add(&newpage->lru, pagepool);
 295                 }
 296         }
 297
 298         /*
 299          * don't do inplace I/O if all compressed pages are available in
 300          * managed cache since it can be moved to the bypass queue instead.
 301          */
 302         if (standalone)
 303                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 304 }
 305
 306 /* called by erofs_shrinker to get rid of all compressed_pages */
 307 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 308                                        struct erofs_workgroup *grp)
 309 {
 310         struct z_erofs_pcluster *const pcl =
 311                 container_of(grp, struct z_erofs_pcluster, obj);
 312         int i;
 313
 314         /*
 315          * refcount of workgroup is now freezed as 1,
 316          * therefore no need to worry about available decompression users.
 317          */
 318         for (i = 0; i < pcl->pclusterpages; ++i) {
 319                 struct page *page = pcl->compressed_pages[i];
 320
 321                 if (!page)
 322                         continue;
 323
 324                 /* block other users from reclaiming or migrating the page */
 325                 if (!trylock_page(page))
 326                         return -EBUSY;
 327
 328                 if (!erofs_page_is_managed(sbi, page))
 329                         continue;
 330
 331                 /* barrier is implied in the following 'unlock_page' */
 332                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 333                 detach_page_private(page);
 334                 unlock_page(page);
 335         }
 336         return 0;
 337 }
 338
 339 int erofs_try_to_free_cached_page(struct page *page)
 340 {
 341         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 342         int ret = 0;    /* 0 - busy */
 343
 344         if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
 345                 unsigned int i;
 346
 347                 for (i = 0; i < pcl->pclusterpages; ++i) {
 348                         if (pcl->compressed_pages[i] == page) {
 349                                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 350                                 ret = 1;
 351                                 break;
 352                         }
 353                 }
 354                 erofs_workgroup_unfreeze(&pcl->obj, 1);
 355
 356                 if (ret)
 357                         detach_page_private(page);
 358         }
 359         return ret;
 360 }
 361
 362 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 363 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
 364                                    struct page *page)
 365 {
 366         struct z_erofs_pcluster *const pcl = clt->pcl;
 367
 368         while (clt->icpage_ptr > pcl->compressed_pages)
 369                 if (!cmpxchg(--clt->icpage_ptr, NULL, page))
 370                         return true;
 371         return false;
 372 }
 373
 374 /* callers must be with collection lock held */
 375 static int z_erofs_attach_page(struct z_erofs_collector *clt,
 376                                struct page *page,
 377                                enum z_erofs_page_type type)
 378 {
 379         int ret;
 380
 381         /* give priority for inplaceio */
 382         if (clt->mode >= COLLECT_PRIMARY &&
 383             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 384             z_erofs_try_inplace_io(clt, page))
 385                 return 0;
 386
 387         ret = z_erofs_pagevec_enqueue(&clt->vector, page, type);
 388         clt->cl->vcnt += (unsigned int)ret;
 389
 390         return ret ? 0 : -EAGAIN;
 391 }
 392
 393 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt)
 394 {
 395         struct z_erofs_pcluster *pcl = clt->pcl;
 396         z_erofs_next_pcluster_t *owned_head = &clt->owned_head;
 397
 398         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 399         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 400                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 401                 *owned_head = &pcl->next;
 402                 /* so we can attach this pcluster to our submission chain. */
 403                 clt->mode = COLLECT_PRIMARY_FOLLOWED;
 404                 return;
 405         }
 406
 407         /*
 408          * type 2, link to the end of an existing open chain, be careful
 409          * that its submission is controlled by the original attached chain.
 410          */
 411         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 412                     *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
 413                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 414                 clt->mode = COLLECT_PRIMARY_HOOKED;
 415                 clt->tailpcl = NULL;
 416                 return;
 417         }
 418         /* type 3, it belongs to a chain, but it isn't the end of the chain */
 419         clt->mode = COLLECT_PRIMARY;
 420 }
 421
 422 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
 423                                      struct inode *inode,
 424                                      struct erofs_map_blocks *map)
 425 {
 426         struct z_erofs_pcluster *pcl = clt->pcl;
 427         struct z_erofs_collection *cl;
 428         unsigned int length;
 429
 430         /* to avoid unexpected loop formed by corrupted images */
 431         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 432                 DBG_BUGON(1);
 433                 return -EFSCORRUPTED;
 434         }
 435
 436         cl = z_erofs_primarycollection(pcl);
 437         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 438                 DBG_BUGON(1);
 439                 return -EFSCORRUPTED;
 440         }
 441
 442         length = READ_ONCE(pcl->length);
 443         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 444                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 445                         DBG_BUGON(1);
 446                         return -EFSCORRUPTED;
 447                 }
 448         } else {
 449                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 450
 451                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 452                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 453
 454                 while (llen > length &&
 455                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 456                         cpu_relax();
 457                         length = READ_ONCE(pcl->length);
 458                 }
 459         }
 460         mutex_lock(&cl->lock);
 461         /* used to check tail merging loop due to corrupted images */
 462         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 463                 clt->tailpcl = pcl;
 464
 465         z_erofs_try_to_claim_pcluster(clt);
 466         clt->cl = cl;
 467         return 0;
 468 }
 469
 470 static int z_erofs_register_collection(struct z_erofs_collector *clt,
 471                                        struct inode *inode,
 472                                        struct erofs_map_blocks *map)
 473 {
 474         struct z_erofs_pcluster *pcl;
 475         struct z_erofs_collection *cl;
 476         struct erofs_workgroup *grp;
 477         int err;
 478
 479         /* no available pcluster, let's allocate one */
 480         pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT);
 481         if (IS_ERR(pcl))
 482                 return PTR_ERR(pcl);
 483
 484         atomic_set(&pcl->obj.refcount, 1);
 485         pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 486
 487         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 488                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 489                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 490
 491         if (map->m_flags & EROFS_MAP_ZIPPED)
 492                 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4;
 493         else
 494                 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED;
 495
 496         /* new pclusters should be claimed as type 1, primary and followed */
 497         pcl->next = clt->owned_head;
 498         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 499
 500         cl = z_erofs_primarycollection(pcl);
 501         cl->pageofs = map->m_la & ~PAGE_MASK;
 502
 503         /*
 504          * lock all primary followed works before visible to others
 505          * and mutex_trylock *never* fails for a new pcluster.
 506          */
 507         mutex_init(&cl->lock);
 508         DBG_BUGON(!mutex_trylock(&cl->lock));
 509
 510         grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
 511         if (IS_ERR(grp)) {
 512                 err = PTR_ERR(grp);
 513                 goto err_out;
 514         }
 515
 516         if (grp != &pcl->obj) {
 517                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 518                 err = -EEXIST;
 519                 goto err_out;
 520         }
 521         /* used to check tail merging loop due to corrupted images */
 522         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 523                 clt->tailpcl = pcl;
 524         clt->owned_head = &pcl->next;
 525         clt->pcl = pcl;
 526         clt->cl = cl;
 527         return 0;
 528
 529 err_out:
 530         mutex_unlock(&cl->lock);
 531         z_erofs_free_pcluster(pcl);
 532         return err;
 533 }
 534
 535 static int z_erofs_collector_begin(struct z_erofs_collector *clt,
 536                                    struct inode *inode,
 537                                    struct erofs_map_blocks *map)
 538 {
 539         struct erofs_workgroup *grp;
 540         int ret;
 541
 542         DBG_BUGON(clt->cl);
 543
 544         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */
 545         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL);
 546         DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 547
 548         if (!PAGE_ALIGNED(map->m_pa)) {
 549                 DBG_BUGON(1);
 550                 return -EINVAL;
 551         }
 552
 553         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
 554         if (grp) {
 555                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 556         } else {
 557                 ret = z_erofs_register_collection(clt, inode, map);
 558
 559                 if (!ret)
 560                         goto out;
 561                 if (ret != -EEXIST)
 562                         return ret;
 563         }
 564
 565         ret = z_erofs_lookup_collection(clt, inode, map);
 566         if (ret) {
 567                 erofs_workgroup_put(&clt->pcl->obj);
 568                 return ret;
 569         }
 570
 571 out:
 572         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 573                                   clt->cl->pagevec, clt->cl->vcnt);
 574
 575         /* since file-backed online pages are traversed in reverse order */
 576         clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages;
 577         return 0;
 578 }
 579
 580 /*
 581  * keep in mind that no referenced pclusters will be freed
 582  * only after a RCU grace period.
 583  */
 584 static void z_erofs_rcu_callback(struct rcu_head *head)
 585 {
 586         struct z_erofs_collection *const cl =
 587                 container_of(head, struct z_erofs_collection, rcu);
 588
 589         z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
 590                                            primary_collection));
 591 }
 592
 593 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 594 {
 595         struct z_erofs_pcluster *const pcl =
 596                 container_of(grp, struct z_erofs_pcluster, obj);
 597         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 598
 599         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 600 }
 601
 602 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 603 {
 604         struct z_erofs_pcluster *const pcl =
 605                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 606
 607         erofs_workgroup_put(&pcl->obj);
 608 }
 609
 610 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 611 {
 612         struct z_erofs_collection *cl = clt->cl;
 613
 614         if (!cl)
 615                 return false;
 616
 617         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 618         mutex_unlock(&cl->lock);
 619
 620         /*
 621          * if all pending pages are added, don't hold its reference
 622          * any longer if the pcluster isn't hosted by ourselves.
 623          */
 624         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 625                 z_erofs_collection_put(cl);
 626
 627         clt->cl = NULL;
 628         return true;
 629 }
 630
 631 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 632                                        unsigned int cachestrategy,
 633                                        erofs_off_t la)
 634 {
 635         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 636                 return false;
 637
 638         if (fe->backmost)
 639                 return true;
 640
 641         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 642                 la < fe->headoffset;
 643 }
 644
 645 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 646                                 struct page *page, struct list_head *pagepool)
 647 {
 648         struct inode *const inode = fe->inode;
 649         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 650         struct erofs_map_blocks *const map = &fe->map;
 651         struct z_erofs_collector *const clt = &fe->clt;
 652         const loff_t offset = page_offset(page);
 653         bool tight = true;
 654
 655         enum z_erofs_cache_alloctype cache_strategy;
 656         enum z_erofs_page_type page_type;
 657         unsigned int cur, end, spiltted, index;
 658         int err = 0;
 659
 660         /* register locked file pages as online pages in pack */
 661         z_erofs_onlinepage_init(page);
 662
 663         spiltted = 0;
 664         end = PAGE_SIZE;
 665 repeat:
 666         cur = end - 1;
 667
 668         /* lucky, within the range of the current map_blocks */
 669         if (offset + cur >= map->m_la &&
 670             offset + cur < map->m_la + map->m_llen) {
 671                 /* didn't get a valid collection previously (very rare) */
 672                 if (!clt->cl)
 673                         goto restart_now;
 674                 goto hitted;
 675         }
 676
 677         /* go ahead the next map_blocks */
 678         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 679
 680         if (z_erofs_collector_end(clt))
 681                 fe->backmost = false;
 682
 683         map->m_la = offset + cur;
 684         map->m_llen = 0;
 685         err = z_erofs_map_blocks_iter(inode, map, 0);
 686         if (err)
 687                 goto err_out;
 688
 689 restart_now:
 690         if (!(map->m_flags & EROFS_MAP_MAPPED))
 691                 goto hitted;
 692
 693         err = z_erofs_collector_begin(clt, inode, map);
 694         if (err)
 695                 goto err_out;
 696
 697         /* preload all compressed pages (maybe downgrade role if necessary) */
 698         if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la))
 699                 cache_strategy = TRYALLOC;
 700         else
 701                 cache_strategy = DONTALLOC;
 702
 703         preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 704                                  cache_strategy, pagepool);
 705
 706 hitted:
 707         /*
 708          * Ensure the current partial page belongs to this submit chain rather
 709          * than other concurrent submit chains or the noio(bypass) chain since
 710          * those chains are handled asynchronously thus the page cannot be used
 711          * for inplace I/O or pagevec (should be processed in strict order.)
 712          */
 713         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 714                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 715
 716         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 717         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 718                 zero_user_segment(page, cur, end);
 719                 goto next_part;
 720         }
 721
 722         /* let's derive page type */
 723         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 724                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 725                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 726                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 727
 728         if (cur)
 729                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 730
 731 retry:
 732         err = z_erofs_attach_page(clt, page, page_type);
 733         /* should allocate an additional short-lived page for pagevec */
 734         if (err == -EAGAIN) {
 735                 struct page *const newpage =
 736                                 alloc_page(GFP_NOFS | __GFP_NOFAIL);
 737
 738                 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
 739                 err = z_erofs_attach_page(clt, newpage,
 740                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 741                 if (!err)
 742                         goto retry;
 743         }
 744
 745         if (err)
 746                 goto err_out;
 747
 748         index = page->index - (map->m_la >> PAGE_SHIFT);
 749
 750         z_erofs_onlinepage_fixup(page, index, true);
 751
 752         /* bump up the number of spiltted parts of a page */
 753         ++spiltted;
 754         /* also update nr_pages */
 755         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 756 next_part:
 757         /* can be used for verification */
 758         map->m_llen = offset + cur - map->m_la;
 759
 760         end = cur;
 761         if (end > 0)
 762                 goto repeat;
 763
 764 out:
 765         z_erofs_onlinepage_endio(page);
 766
 767         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 768                   __func__, page, spiltted, map->m_llen);
 769         return err;
 770
 771         /* if some error occurred while processing this page */
 772 err_out:
 773         SetPageError(page);
 774         goto out;
 775 }
 776
 777 static void z_erofs_decompressqueue_work(struct work_struct *work);
 778 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
 779                                        bool sync, int bios)
 780 {
 781         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
 782
 783         /* wake up the caller thread for sync decompression */
 784         if (sync) {
 785                 unsigned long flags;
 786
 787                 spin_lock_irqsave(&io->u.wait.lock, flags);
 788                 if (!atomic_add_return(bios, &io->pending_bios))
 789                         wake_up_locked(&io->u.wait);
 790                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 791                 return;
 792         }
 793
 794         if (atomic_add_return(bios, &io->pending_bios))
 795                 return;
 796         /* Use workqueue and sync decompression for atomic contexts only */
 797         if (in_atomic() || irqs_disabled()) {
 798                 queue_work(z_erofs_workqueue, &io->u.work);
 799                 sbi->ctx.readahead_sync_decompress = true;
 800                 return;
 801         }
 802         z_erofs_decompressqueue_work(&io->u.work);
 803 }
 804
 805 static bool z_erofs_page_is_invalidated(struct page *page)
 806 {
 807         return !page->mapping && !z_erofs_is_shortlived_page(page);
 808 }
 809
 810 static void z_erofs_decompressqueue_endio(struct bio *bio)
 811 {
 812         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
 813         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
 814         blk_status_t err = bio->bi_status;
 815         struct bio_vec *bvec;
 816         struct bvec_iter_all iter_all;
 817
 818         bio_for_each_segment_all(bvec, bio, iter_all) {
 819                 struct page *page = bvec->bv_page;
 820
 821                 DBG_BUGON(PageUptodate(page));
 822                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 823
 824                 if (err)
 825                         SetPageError(page);
 826
 827                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
 828                         if (!err)
 829                                 SetPageUptodate(page);
 830                         unlock_page(page);
 831                 }
 832         }
 833         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
 834         bio_put(bio);
 835 }
 836
 837 static int z_erofs_decompress_pcluster(struct super_block *sb,
 838                                        struct z_erofs_pcluster *pcl,
 839                                        struct list_head *pagepool)
 840 {
 841         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 842         struct z_erofs_pagevec_ctor ctor;
 843         unsigned int i, inputsize, outputsize, llen, nr_pages;
 844         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 845         struct page **pages, **compressed_pages, *page;
 846
 847         enum z_erofs_page_type page_type;
 848         bool overlapped, partial;
 849         struct z_erofs_collection *cl;
 850         int err;
 851
 852         might_sleep();
 853         cl = z_erofs_primarycollection(pcl);
 854         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 855
 856         mutex_lock(&cl->lock);
 857         nr_pages = cl->nr_pages;
 858
 859         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 860                 pages = pages_onstack;
 861         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 862                    mutex_trylock(&z_pagemap_global_lock)) {
 863                 pages = z_pagemap_global;
 864         } else {
 865                 gfp_t gfp_flags = GFP_KERNEL;
 866
 867                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 868                         gfp_flags |= __GFP_NOFAIL;
 869
 870                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 871                                        gfp_flags);
 872
 873                 /* fallback to global pagemap for the lowmem scenario */
 874                 if (!pages) {
 875                         mutex_lock(&z_pagemap_global_lock);
 876                         pages = z_pagemap_global;
 877                 }
 878         }
 879
 880         for (i = 0; i < nr_pages; ++i)
 881                 pages[i] = NULL;
 882
 883         err = 0;
 884         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 885                                   cl->pagevec, 0);
 886
 887         for (i = 0; i < cl->vcnt; ++i) {
 888                 unsigned int pagenr;
 889
 890                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 891
 892                 /* all pages in pagevec ought to be valid */
 893                 DBG_BUGON(!page);
 894                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 895
 896                 if (z_erofs_put_shortlivedpage(pagepool, page))
 897                         continue;
 898
 899                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 900                         pagenr = 0;
 901                 else
 902                         pagenr = z_erofs_onlinepage_index(page);
 903
 904                 DBG_BUGON(pagenr >= nr_pages);
 905
 906                 /*
 907                  * currently EROFS doesn't support multiref(dedup),
 908                  * so here erroring out one multiref page.
 909                  */
 910                 if (pages[pagenr]) {
 911                         DBG_BUGON(1);
 912                         SetPageError(pages[pagenr]);
 913                         z_erofs_onlinepage_endio(pages[pagenr]);
 914                         err = -EFSCORRUPTED;
 915                 }
 916                 pages[pagenr] = page;
 917         }
 918         z_erofs_pagevec_ctor_exit(&ctor, true);
 919
 920         overlapped = false;
 921         compressed_pages = pcl->compressed_pages;
 922
 923         for (i = 0; i < pcl->pclusterpages; ++i) {
 924                 unsigned int pagenr;
 925
 926                 page = compressed_pages[i];
 927
 928                 /* all compressed pages ought to be valid */
 929                 DBG_BUGON(!page);
 930                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 931
 932                 if (!z_erofs_is_shortlived_page(page)) {
 933                         if (erofs_page_is_managed(sbi, page)) {
 934                                 if (!PageUptodate(page))
 935                                         err = -EIO;
 936                                 continue;
 937                         }
 938
 939                         /*
 940                          * only if non-head page can be selected
 941                          * for inplace decompression
 942                          */
 943                         pagenr = z_erofs_onlinepage_index(page);
 944
 945                         DBG_BUGON(pagenr >= nr_pages);
 946                         if (pages[pagenr]) {
 947                                 DBG_BUGON(1);
 948                                 SetPageError(pages[pagenr]);
 949                                 z_erofs_onlinepage_endio(pages[pagenr]);
 950                                 err = -EFSCORRUPTED;
 951                         }
 952                         pages[pagenr] = page;
 953
 954                         overlapped = true;
 955                 }
 956
 957                 /* PG_error needs checking for all non-managed pages */
 958                 if (PageError(page)) {
 959                         DBG_BUGON(PageUptodate(page));
 960                         err = -EIO;
 961                 }
 962         }
 963
 964         if (err)
 965                 goto out;
 966
 967         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 968         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 969                 outputsize = llen;
 970                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 971         } else {
 972                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 973                 partial = true;
 974         }
 975
 976         inputsize = pcl->pclusterpages * PAGE_SIZE;
 977         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 978                                         .sb = sb,
 979                                         .in = compressed_pages,
 980                                         .out = pages,
 981                                         .pageofs_out = cl->pageofs,
 982                                         .inputsize = inputsize,
 983                                         .outputsize = outputsize,
 984                                         .alg = pcl->algorithmformat,
 985                                         .inplace_io = overlapped,
 986                                         .partial_decoding = partial
 987                                  }, pagepool);
 988
 989 out:
 990         /* must handle all compressed pages before ending pages */
 991         for (i = 0; i < pcl->pclusterpages; ++i) {
 992                 page = compressed_pages[i];
 993
 994                 if (erofs_page_is_managed(sbi, page))
 995                         continue;
 996
 997                 /* recycle all individual short-lived pages */
 998                 (void)z_erofs_put_shortlivedpage(pagepool, page);
 999
1000                 WRITE_ONCE(compressed_pages[i], NULL);
1001         }
1002
1003         for (i = 0; i < nr_pages; ++i) {
1004                 page = pages[i];
1005                 if (!page)
1006                         continue;
1007
1008                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1009
1010                 /* recycle all individual short-lived pages */
1011                 if (z_erofs_put_shortlivedpage(pagepool, page))
1012                         continue;
1013
1014                 if (err < 0)
1015                         SetPageError(page);
1016
1017                 z_erofs_onlinepage_endio(page);
1018         }
1019
1020         if (pages == z_pagemap_global)
1021                 mutex_unlock(&z_pagemap_global_lock);
1022         else if (pages != pages_onstack)
1023                 kvfree(pages);
1024
1025         cl->nr_pages = 0;
1026         cl->vcnt = 0;
1027
1028         /* all cl locks MUST be taken before the following line */
1029         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1030
1031         /* all cl locks SHOULD be released right now */
1032         mutex_unlock(&cl->lock);
1033
1034         z_erofs_collection_put(cl);
1035         return err;
1036 }
1037
1038 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1039                                      struct list_head *pagepool)
1040 {
1041         z_erofs_next_pcluster_t owned = io->head;
1042
1043         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1044                 struct z_erofs_pcluster *pcl;
1045
1046                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1047                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1048
1049                 /* no possible that 'owned' equals NULL */
1050                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1051
1052                 pcl = container_of(owned, struct z_erofs_pcluster, next);
1053                 owned = READ_ONCE(pcl->next);
1054
1055                 z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1056         }
1057 }
1058
1059 static void z_erofs_decompressqueue_work(struct work_struct *work)
1060 {
1061         struct z_erofs_decompressqueue *bgq =
1062                 container_of(work, struct z_erofs_decompressqueue, u.work);
1063         LIST_HEAD(pagepool);
1064
1065         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1066         z_erofs_decompress_queue(bgq, &pagepool);
1067
1068         put_pages_list(&pagepool);
1069         kvfree(bgq);
1070 }
1071
1072 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1073                                                unsigned int nr,
1074                                                struct list_head *pagepool,
1075                                                struct address_space *mc,
1076                                                gfp_t gfp)
1077 {
1078         const pgoff_t index = pcl->obj.index;
1079         bool tocache = false;
1080
1081         struct address_space *mapping;
1082         struct page *oldpage, *page;
1083
1084         compressed_page_t t;
1085         int justfound;
1086
1087 repeat:
1088         page = READ_ONCE(pcl->compressed_pages[nr]);
1089         oldpage = page;
1090
1091         if (!page)
1092                 goto out_allocpage;
1093
1094         /*
1095          * the cached page has not been allocated and
1096          * an placeholder is out there, prepare it now.
1097          */
1098         if (page == PAGE_UNALLOCATED) {
1099                 tocache = true;
1100                 goto out_allocpage;
1101         }
1102
1103         /* process the target tagged pointer */
1104         t = tagptr_init(compressed_page_t, page);
1105         justfound = tagptr_unfold_tags(t);
1106         page = tagptr_unfold_ptr(t);
1107
1108         /*
1109          * preallocated cached pages, which is used to avoid direct reclaim
1110          * otherwise, it will go inplace I/O path instead.
1111          */
1112         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1113                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1114                 set_page_private(page, 0);
1115                 tocache = true;
1116                 goto out_tocache;
1117         }
1118         mapping = READ_ONCE(page->mapping);
1119
1120         /*
1121          * file-backed online pages in plcuster are all locked steady,
1122          * therefore it is impossible for `mapping' to be NULL.
1123          */
1124         if (mapping && mapping != mc)
1125                 /* ought to be unmanaged pages */
1126                 goto out;
1127
1128         /* directly return for shortlived page as well */
1129         if (z_erofs_is_shortlived_page(page))
1130                 goto out;
1131
1132         lock_page(page);
1133
1134         /* only true if page reclaim goes wrong, should never happen */
1135         DBG_BUGON(justfound && PagePrivate(page));
1136
1137         /* the page is still in manage cache */
1138         if (page->mapping == mc) {
1139                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1140
1141                 ClearPageError(page);
1142                 if (!PagePrivate(page)) {
1143                         /*
1144                          * impossible to be !PagePrivate(page) for
1145                          * the current restriction as well if
1146                          * the page is already in compressed_pages[].
1147                          */
1148                         DBG_BUGON(!justfound);
1149
1150                         justfound = 0;
1151                         set_page_private(page, (unsigned long)pcl);
1152                         SetPagePrivate(page);
1153                 }
1154
1155                 /* no need to submit io if it is already up-to-date */
1156                 if (PageUptodate(page)) {
1157                         unlock_page(page);
1158                         page = NULL;
1159                 }
1160                 goto out;
1161         }
1162
1163         /*
1164          * the managed page has been truncated, it's unsafe to
1165          * reuse this one, let's allocate a new cache-managed page.
1166          */
1167         DBG_BUGON(page->mapping);
1168         DBG_BUGON(!justfound);
1169
1170         tocache = true;
1171         unlock_page(page);
1172         put_page(page);
1173 out_allocpage:
1174         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1175         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1176                 list_add(&page->lru, pagepool);
1177                 cond_resched();
1178                 goto repeat;
1179         }
1180 out_tocache:
1181         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1182                 /* turn into temporary page if fails (1 ref) */
1183                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1184                 goto out;
1185         }
1186         attach_page_private(page, pcl);
1187         /* drop a refcount added by allocpage (then we have 2 refs here) */
1188         put_page(page);
1189
1190 out:    /* the only exit (for tracing and debugging) */
1191         return page;
1192 }
1193
1194 static struct z_erofs_decompressqueue *
1195 jobqueue_init(struct super_block *sb,
1196               struct z_erofs_decompressqueue *fgq, bool *fg)
1197 {
1198         struct z_erofs_decompressqueue *q;
1199
1200         if (fg && !*fg) {
1201                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1202                 if (!q) {
1203                         *fg = true;
1204                         goto fg_out;
1205                 }
1206                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1207         } else {
1208 fg_out:
1209                 q = fgq;
1210                 init_waitqueue_head(&fgq->u.wait);
1211                 atomic_set(&fgq->pending_bios, 0);
1212         }
1213         q->sb = sb;
1214         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1215         return q;
1216 }
1217
1218 /* define decompression jobqueue types */
1219 enum {
1220         JQ_BYPASS,
1221         JQ_SUBMIT,
1222         NR_JOBQUEUES,
1223 };
1224
1225 static void *jobqueueset_init(struct super_block *sb,
1226                               struct z_erofs_decompressqueue *q[],
1227                               struct z_erofs_decompressqueue *fgq, bool *fg)
1228 {
1229         /*
1230          * if managed cache is enabled, bypass jobqueue is needed,
1231          * no need to read from device for all pclusters in this queue.
1232          */
1233         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1234         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1235
1236         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1237 }
1238
1239 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1240                                     z_erofs_next_pcluster_t qtail[],
1241                                     z_erofs_next_pcluster_t owned_head)
1242 {
1243         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1244         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1245
1246         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1247         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1248                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1249
1250         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1251
1252         WRITE_ONCE(*submit_qtail, owned_head);
1253         WRITE_ONCE(*bypass_qtail, &pcl->next);
1254
1255         qtail[JQ_BYPASS] = &pcl->next;
1256 }
1257
1258 static void z_erofs_submit_queue(struct super_block *sb,
1259                                  struct z_erofs_decompress_frontend *f,
1260                                  struct list_head *pagepool,
1261                                  struct z_erofs_decompressqueue *fgq,
1262                                  bool *force_fg)
1263 {
1264         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1265         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1266         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1267         void *bi_private;
1268         z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1269         /* since bio will be NULL, no need to initialize last_index */
1270         pgoff_t last_index;
1271         unsigned int nr_bios = 0;
1272         struct bio *bio = NULL;
1273
1274         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1275         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1276         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1277
1278         /* by default, all need io submission */
1279         q[JQ_SUBMIT]->head = owned_head;
1280
1281         do {
1282                 struct z_erofs_pcluster *pcl;
1283                 pgoff_t cur, end;
1284                 unsigned int i = 0;
1285                 bool bypass = true;
1286
1287                 /* no possible 'owned_head' equals the following */
1288                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1289                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1290
1291                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1292
1293                 cur = pcl->obj.index;
1294                 end = cur + pcl->pclusterpages;
1295
1296                 /* close the main owned chain at first */
1297                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1298                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1299
1300                 do {
1301                         struct page *page;
1302
1303                         page = pickup_page_for_submission(pcl, i++, pagepool,
1304                                                           MNGD_MAPPING(sbi),
1305                                                           GFP_NOFS);
1306                         if (!page)
1307                                 continue;
1308
1309                         if (bio && cur != last_index + 1) {
1310 submit_bio_retry:
1311                                 submit_bio(bio);
1312                                 bio = NULL;
1313                         }
1314
1315                         if (!bio) {
1316                                 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS);
1317
1318                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1319                                 bio_set_dev(bio, sb->s_bdev);
1320                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1321                                         LOG_SECTORS_PER_BLOCK;
1322                                 bio->bi_private = bi_private;
1323                                 bio->bi_opf = REQ_OP_READ;
1324                                 if (f->readahead)
1325                                         bio->bi_opf |= REQ_RAHEAD;
1326                                 ++nr_bios;
1327                         }
1328
1329                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1330                                 goto submit_bio_retry;
1331
1332                         last_index = cur;
1333                         bypass = false;
1334                 } while (++cur < end);
1335
1336                 if (!bypass)
1337                         qtail[JQ_SUBMIT] = &pcl->next;
1338                 else
1339                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1340         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1341
1342         if (bio)
1343                 submit_bio(bio);
1344
1345         /*
1346          * although background is preferred, no one is pending for submission.
1347          * don't issue workqueue for decompression but drop it directly instead.
1348          */
1349         if (!*force_fg && !nr_bios) {
1350                 kvfree(q[JQ_SUBMIT]);
1351                 return;
1352         }
1353         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1354 }
1355
1356 static void z_erofs_runqueue(struct super_block *sb,
1357                              struct z_erofs_decompress_frontend *f,
1358                              struct list_head *pagepool, bool force_fg)
1359 {
1360         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1361
1362         if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1363                 return;
1364         z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1365
1366         /* handle bypass queue (no i/o pclusters) immediately */
1367         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1368
1369         if (!force_fg)
1370                 return;
1371
1372         /* wait until all bios are completed */
1373         io_wait_event(io[JQ_SUBMIT].u.wait,
1374                       !atomic_read(&io[JQ_SUBMIT].pending_bios));
1375
1376         /* handle synchronous decompress queue in the caller context */
1377         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1378 }
1379
1380 static int z_erofs_readpage(struct file *file, struct page *page)
1381 {
1382         struct inode *const inode = page->mapping->host;
1383         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1384         int err;
1385         LIST_HEAD(pagepool);
1386
1387         trace_erofs_readpage(page, false);
1388
1389         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1390
1391         err = z_erofs_do_read_page(&f, page, &pagepool);
1392         (void)z_erofs_collector_end(&f.clt);
1393
1394         /* if some compressed cluster ready, need submit them anyway */
1395         z_erofs_runqueue(inode->i_sb, &f, &pagepool, true);
1396
1397         if (err)
1398                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1399
1400         if (f.map.mpage)
1401                 put_page(f.map.mpage);
1402
1403         /* clean up the remaining free pages */
1404         put_pages_list(&pagepool);
1405         return err;
1406 }
1407
1408 static void z_erofs_readahead(struct readahead_control *rac)
1409 {
1410         struct inode *const inode = rac->mapping->host;
1411         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1412
1413         unsigned int nr_pages = readahead_count(rac);
1414         bool sync = (sbi->ctx.readahead_sync_decompress &&
1415                         nr_pages <= sbi->ctx.max_sync_decompress_pages);
1416         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1417         struct page *page, *head = NULL;
1418         LIST_HEAD(pagepool);
1419
1420         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1421
1422         f.readahead = true;
1423         f.headoffset = readahead_pos(rac);
1424
1425         while ((page = readahead_page(rac))) {
1426                 prefetchw(&page->flags);
1427
1428                 /*
1429                  * A pure asynchronous readahead is indicated if
1430                  * a PG_readahead marked page is hitted at first.
1431                  * Let's also do asynchronous decompression for this case.
1432                  */
1433                 sync &= !(PageReadahead(page) && !head);
1434
1435                 set_page_private(page, (unsigned long)head);
1436                 head = page;
1437         }
1438
1439         while (head) {
1440                 struct page *page = head;
1441                 int err;
1442
1443                 /* traversal in reverse order */
1444                 head = (void *)page_private(page);
1445
1446                 err = z_erofs_do_read_page(&f, page, &pagepool);
1447                 if (err)
1448                         erofs_err(inode->i_sb,
1449                                   "readahead error at page %lu @ nid %llu",
1450                                   page->index, EROFS_I(inode)->nid);
1451                 put_page(page);
1452         }
1453
1454         (void)z_erofs_collector_end(&f.clt);
1455
1456         z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync);
1457
1458         if (f.map.mpage)
1459                 put_page(f.map.mpage);
1460
1461         /* clean up the remaining free pages */
1462         put_pages_list(&pagepool);
1463 }
1464
1465 const struct address_space_operations z_erofs_aops = {
1466         .readpage = z_erofs_readpage,
1467         .readahead = z_erofs_readahead,
1468 };