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         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
  86         int i;
  87
  88         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
  89                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
  90
  91                 if (pclusterpages > pcs->maxpages)
  92                         continue;
  93
  94                 kmem_cache_free(pcs->slab, pcl);
  95                 return;
  96         }
  97         DBG_BUGON(1);
  98 }
  99
 100 /* how to allocate cached pages for a pcluster */
 101 enum z_erofs_cache_alloctype {
 102         DONTALLOC,      /* don't allocate any cached pages */
 103         /*
 104          * try to use cached I/O if page allocation succeeds or fallback
 105          * to in-place I/O instead to avoid any direct reclaim.
 106          */
 107         TRYALLOC,
 108 };
 109
 110 /*
 111  * tagged pointer with 1-bit tag for all compressed pages
 112  * tag 0 - the page is just found with an extra page reference
 113  */
 114 typedef tagptr1_t compressed_page_t;
 115
 116 #define tag_compressed_page_justfound(page) \
 117         tagptr_fold(compressed_page_t, page, 1)
 118
 119 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 120
 121 void z_erofs_exit_zip_subsystem(void)
 122 {
 123         destroy_workqueue(z_erofs_workqueue);
 124         z_erofs_destroy_pcluster_pool();
 125 }
 126
 127 static inline int z_erofs_init_workqueue(void)
 128 {
 129         const unsigned int onlinecpus = num_possible_cpus();
 130
 131         /*
 132          * no need to spawn too many threads, limiting threads could minimum
 133          * scheduling overhead, perhaps per-CPU threads should be better?
 134          */
 135         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
 136                                             WQ_UNBOUND | WQ_HIGHPRI,
 137                                             onlinecpus + onlinecpus / 4);
 138         return z_erofs_workqueue ? 0 : -ENOMEM;
 139 }
 140
 141 int __init z_erofs_init_zip_subsystem(void)
 142 {
 143         int err = z_erofs_create_pcluster_pool();
 144
 145         if (err)
 146                 return err;
 147         err = z_erofs_init_workqueue();
 148         if (err)
 149                 z_erofs_destroy_pcluster_pool();
 150         return err;
 151 }
 152
 153 enum z_erofs_collectmode {
 154         COLLECT_SECONDARY,
 155         COLLECT_PRIMARY,
 156         /*
 157          * The current collection was the tail of an exist chain, in addition
 158          * that the previous processed chained collections are all decided to
 159          * be hooked up to it.
 160          * A new chain will be created for the remaining collections which are
 161          * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED,
 162          * the next collection cannot reuse the whole page safely in
 163          * the following scenario:
 164          *  ________________________________________________________________
 165          * |      tail (partial) page     |       head (partial) page       |
 166          * |   (belongs to the next cl)   |   (belongs to the current cl)   |
 167          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 168          */
 169         COLLECT_PRIMARY_HOOKED,
 170         /*
 171          * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it
 172          * could be dispatched into bypass queue later due to uptodated managed
 173          * pages. All related online pages cannot be reused for inplace I/O (or
 174          * pagevec) since it can be directly decoded without I/O submission.
 175          */
 176         COLLECT_PRIMARY_FOLLOWED_NOINPLACE,
 177         /*
 178          * The current collection has been linked with the owned chain, and
 179          * could also be linked with the remaining collections, which means
 180          * if the processing page is the tail page of the collection, thus
 181          * the current collection can safely use the whole page (since
 182          * the previous collection is under control) for in-place I/O, as
 183          * illustrated below:
 184          *  ________________________________________________________________
 185          * |  tail (partial) page |          head (partial) page           |
 186          * |  (of the current cl) |      (of the previous collection)      |
 187          * |  PRIMARY_FOLLOWED or |                                        |
 188          * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________|
 189          *
 190          * [  (*) the above page can be used as inplace I/O.               ]
 191          */
 192         COLLECT_PRIMARY_FOLLOWED,
 193 };
 194
 195 struct z_erofs_collector {
 196         struct z_erofs_pagevec_ctor vector;
 197
 198         struct z_erofs_pcluster *pcl, *tailpcl;
 199         struct z_erofs_collection *cl;
 200         /* a pointer used to pick up inplace I/O pages */
 201         struct page **icpage_ptr;
 202         z_erofs_next_pcluster_t owned_head;
 203
 204         enum z_erofs_collectmode mode;
 205 };
 206
 207 struct z_erofs_decompress_frontend {
 208         struct inode *const inode;
 209
 210         struct z_erofs_collector clt;
 211         struct erofs_map_blocks map;
 212
 213         bool readahead;
 214         /* used for applying cache strategy on the fly */
 215         bool backmost;
 216         erofs_off_t headoffset;
 217 };
 218
 219 #define COLLECTOR_INIT() { \
 220         .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 221         .mode = COLLECT_PRIMARY_FOLLOWED }
 222
 223 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 224         .inode = __i, .clt = COLLECTOR_INIT(), \
 225         .backmost = true, }
 226
 227 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES];
 228 static DEFINE_MUTEX(z_pagemap_global_lock);
 229
 230 static void preload_compressed_pages(struct z_erofs_collector *clt,
 231                                      struct address_space *mc,
 232                                      enum z_erofs_cache_alloctype type,
 233                                      struct page **pagepool)
 234 {
 235         struct z_erofs_pcluster *pcl = clt->pcl;
 236         bool standalone = true;
 237         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 238                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 239         struct page **pages;
 240         pgoff_t index;
 241
 242         if (clt->mode < COLLECT_PRIMARY_FOLLOWED)
 243                 return;
 244
 245         pages = pcl->compressed_pages;
 246         index = pcl->obj.index;
 247         for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) {
 248                 struct page *page;
 249                 compressed_page_t t;
 250                 struct page *newpage = NULL;
 251
 252                 /* the compressed page was loaded before */
 253                 if (READ_ONCE(*pages))
 254                         continue;
 255
 256                 page = find_get_page(mc, index);
 257
 258                 if (page) {
 259                         t = tag_compressed_page_justfound(page);
 260                 } else {
 261                         /* I/O is needed, no possible to decompress directly */
 262                         standalone = false;
 263                         switch (type) {
 264                         case TRYALLOC:
 265                                 newpage = erofs_allocpage(pagepool, gfp);
 266                                 if (!newpage)
 267                                         continue;
 268                                 set_page_private(newpage,
 269                                                  Z_EROFS_PREALLOCATED_PAGE);
 270                                 t = tag_compressed_page_justfound(newpage);
 271                                 break;
 272                         default:        /* DONTALLOC */
 273                                 continue;
 274                         }
 275                 }
 276
 277                 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t)))
 278                         continue;
 279
 280                 if (page)
 281                         put_page(page);
 282                 else if (newpage)
 283                         erofs_pagepool_add(pagepool, newpage);
 284         }
 285
 286         /*
 287          * don't do inplace I/O if all compressed pages are available in
 288          * managed cache since it can be moved to the bypass queue instead.
 289          */
 290         if (standalone)
 291                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 292 }
 293
 294 /* called by erofs_shrinker to get rid of all compressed_pages */
 295 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 296                                        struct erofs_workgroup *grp)
 297 {
 298         struct z_erofs_pcluster *const pcl =
 299                 container_of(grp, struct z_erofs_pcluster, obj);
 300         int i;
 301
 302         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 303         /*
 304          * refcount of workgroup is now freezed as 1,
 305          * therefore no need to worry about available decompression users.
 306          */
 307         for (i = 0; i < pcl->pclusterpages; ++i) {
 308                 struct page *page = pcl->compressed_pages[i];
 309
 310                 if (!page)
 311                         continue;
 312
 313                 /* block other users from reclaiming or migrating the page */
 314                 if (!trylock_page(page))
 315                         return -EBUSY;
 316
 317                 if (!erofs_page_is_managed(sbi, page))
 318                         continue;
 319
 320                 /* barrier is implied in the following 'unlock_page' */
 321                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 322                 detach_page_private(page);
 323                 unlock_page(page);
 324         }
 325         return 0;
 326 }
 327
 328 int erofs_try_to_free_cached_page(struct page *page)
 329 {
 330         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 331         int ret = 0;    /* 0 - busy */
 332
 333         if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) {
 334                 unsigned int i;
 335
 336                 DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 337                 for (i = 0; i < pcl->pclusterpages; ++i) {
 338                         if (pcl->compressed_pages[i] == page) {
 339                                 WRITE_ONCE(pcl->compressed_pages[i], NULL);
 340                                 ret = 1;
 341                                 break;
 342                         }
 343                 }
 344                 erofs_workgroup_unfreeze(&pcl->obj, 1);
 345
 346                 if (ret)
 347                         detach_page_private(page);
 348         }
 349         return ret;
 350 }
 351
 352 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 353 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt,
 354                                    struct page *page)
 355 {
 356         struct z_erofs_pcluster *const pcl = clt->pcl;
 357
 358         while (clt->icpage_ptr > pcl->compressed_pages)
 359                 if (!cmpxchg(--clt->icpage_ptr, NULL, page))
 360                         return true;
 361         return false;
 362 }
 363
 364 /* callers must be with collection lock held */
 365 static int z_erofs_attach_page(struct z_erofs_collector *clt,
 366                                struct page *page, enum z_erofs_page_type type,
 367                                bool pvec_safereuse)
 368 {
 369         int ret;
 370
 371         /* give priority for inplaceio */
 372         if (clt->mode >= COLLECT_PRIMARY &&
 373             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 374             z_erofs_try_inplace_io(clt, page))
 375                 return 0;
 376
 377         ret = z_erofs_pagevec_enqueue(&clt->vector, page, type,
 378                                       pvec_safereuse);
 379         clt->cl->vcnt += (unsigned int)ret;
 380         return ret ? 0 : -EAGAIN;
 381 }
 382
 383 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt)
 384 {
 385         struct z_erofs_pcluster *pcl = clt->pcl;
 386         z_erofs_next_pcluster_t *owned_head = &clt->owned_head;
 387
 388         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 389         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 390                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 391                 *owned_head = &pcl->next;
 392                 /* so we can attach this pcluster to our submission chain. */
 393                 clt->mode = COLLECT_PRIMARY_FOLLOWED;
 394                 return;
 395         }
 396
 397         /*
 398          * type 2, link to the end of an existing open chain, be careful
 399          * that its submission is controlled by the original attached chain.
 400          */
 401         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 402                     *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
 403                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 404                 clt->mode = COLLECT_PRIMARY_HOOKED;
 405                 clt->tailpcl = NULL;
 406                 return;
 407         }
 408         /* type 3, it belongs to a chain, but it isn't the end of the chain */
 409         clt->mode = COLLECT_PRIMARY;
 410 }
 411
 412 static int z_erofs_lookup_collection(struct z_erofs_collector *clt,
 413                                      struct inode *inode,
 414                                      struct erofs_map_blocks *map)
 415 {
 416         struct z_erofs_pcluster *pcl = clt->pcl;
 417         struct z_erofs_collection *cl;
 418         unsigned int length;
 419
 420         /* to avoid unexpected loop formed by corrupted images */
 421         if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) {
 422                 DBG_BUGON(1);
 423                 return -EFSCORRUPTED;
 424         }
 425
 426         cl = z_erofs_primarycollection(pcl);
 427         if (cl->pageofs != (map->m_la & ~PAGE_MASK)) {
 428                 DBG_BUGON(1);
 429                 return -EFSCORRUPTED;
 430         }
 431
 432         length = READ_ONCE(pcl->length);
 433         if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) {
 434                 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) {
 435                         DBG_BUGON(1);
 436                         return -EFSCORRUPTED;
 437                 }
 438         } else {
 439                 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT;
 440
 441                 if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 442                         llen |= Z_EROFS_PCLUSTER_FULL_LENGTH;
 443
 444                 while (llen > length &&
 445                        length != cmpxchg_relaxed(&pcl->length, length, llen)) {
 446                         cpu_relax();
 447                         length = READ_ONCE(pcl->length);
 448                 }
 449         }
 450         mutex_lock(&cl->lock);
 451         /* used to check tail merging loop due to corrupted images */
 452         if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL)
 453                 clt->tailpcl = pcl;
 454
 455         z_erofs_try_to_claim_pcluster(clt);
 456         clt->cl = cl;
 457         return 0;
 458 }
 459
 460 static int z_erofs_register_collection(struct z_erofs_collector *clt,
 461                                        struct inode *inode,
 462                                        struct erofs_map_blocks *map)
 463 {
 464         bool ztailpacking = map->m_flags & EROFS_MAP_META;
 465         struct z_erofs_pcluster *pcl;
 466         struct z_erofs_collection *cl;
 467         struct erofs_workgroup *grp;
 468         int err;
 469
 470         if (!(map->m_flags & EROFS_MAP_ENCODED)) {
 471                 DBG_BUGON(1);
 472                 return -EFSCORRUPTED;
 473         }
 474
 475         /* no available pcluster, let's allocate one */
 476         pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
 477                                      map->m_plen >> PAGE_SHIFT);
 478         if (IS_ERR(pcl))
 479                 return PTR_ERR(pcl);
 480
 481         atomic_set(&pcl->obj.refcount, 1);
 482         pcl->algorithmformat = map->m_algorithmformat;
 483         pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) |
 484                 (map->m_flags & EROFS_MAP_FULL_MAPPED ?
 485                         Z_EROFS_PCLUSTER_FULL_LENGTH : 0);
 486
 487         /* new pclusters should be claimed as type 1, primary and followed */
 488         pcl->next = clt->owned_head;
 489         clt->mode = COLLECT_PRIMARY_FOLLOWED;
 490
 491         cl = z_erofs_primarycollection(pcl);
 492         cl->pageofs = map->m_la & ~PAGE_MASK;
 493
 494         /*
 495          * lock all primary followed works before visible to others
 496          * and mutex_trylock *never* fails for a new pcluster.
 497          */
 498         mutex_init(&cl->lock);
 499         DBG_BUGON(!mutex_trylock(&cl->lock));
 500
 501         if (ztailpacking) {
 502                 pcl->obj.index = 0;     /* which indicates ztailpacking */
 503                 pcl->pageofs_in = erofs_blkoff(map->m_pa);
 504                 pcl->tailpacking_size = map->m_plen;
 505         } else {
 506                 pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 507
 508                 grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj);
 509                 if (IS_ERR(grp)) {
 510                         err = PTR_ERR(grp);
 511                         goto err_out;
 512                 }
 513
 514                 if (grp != &pcl->obj) {
 515                         clt->pcl = container_of(grp,
 516                                         struct z_erofs_pcluster, obj);
 517                         err = -EEXIST;
 518                         goto err_out;
 519                 }
 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 (map->m_flags & EROFS_MAP_META) {
 549                 if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
 550                         DBG_BUGON(1);
 551                         return -EFSCORRUPTED;
 552                 }
 553                 goto tailpacking;
 554         }
 555
 556         grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT);
 557         if (grp) {
 558                 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 559         } else {
 560 tailpacking:
 561                 ret = z_erofs_register_collection(clt, inode, map);
 562                 if (!ret)
 563                         goto out;
 564                 if (ret != -EEXIST)
 565                         return ret;
 566         }
 567
 568         ret = z_erofs_lookup_collection(clt, inode, map);
 569         if (ret) {
 570                 erofs_workgroup_put(&clt->pcl->obj);
 571                 return ret;
 572         }
 573
 574 out:
 575         z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 576                                   clt->cl->pagevec, clt->cl->vcnt);
 577         /* since file-backed online pages are traversed in reverse order */
 578         clt->icpage_ptr = clt->pcl->compressed_pages +
 579                         z_erofs_pclusterpages(clt->pcl);
 580         return 0;
 581 }
 582
 583 /*
 584  * keep in mind that no referenced pclusters will be freed
 585  * only after a RCU grace period.
 586  */
 587 static void z_erofs_rcu_callback(struct rcu_head *head)
 588 {
 589         struct z_erofs_collection *const cl =
 590                 container_of(head, struct z_erofs_collection, rcu);
 591
 592         z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster,
 593                                            primary_collection));
 594 }
 595
 596 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 597 {
 598         struct z_erofs_pcluster *const pcl =
 599                 container_of(grp, struct z_erofs_pcluster, obj);
 600         struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl);
 601
 602         call_rcu(&cl->rcu, z_erofs_rcu_callback);
 603 }
 604
 605 static void z_erofs_collection_put(struct z_erofs_collection *cl)
 606 {
 607         struct z_erofs_pcluster *const pcl =
 608                 container_of(cl, struct z_erofs_pcluster, primary_collection);
 609
 610         erofs_workgroup_put(&pcl->obj);
 611 }
 612
 613 static bool z_erofs_collector_end(struct z_erofs_collector *clt)
 614 {
 615         struct z_erofs_collection *cl = clt->cl;
 616
 617         if (!cl)
 618                 return false;
 619
 620         z_erofs_pagevec_ctor_exit(&clt->vector, false);
 621         mutex_unlock(&cl->lock);
 622
 623         /*
 624          * if all pending pages are added, don't hold its reference
 625          * any longer if the pcluster isn't hosted by ourselves.
 626          */
 627         if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE)
 628                 z_erofs_collection_put(cl);
 629
 630         clt->cl = NULL;
 631         return true;
 632 }
 633
 634 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 635                                        unsigned int cachestrategy,
 636                                        erofs_off_t la)
 637 {
 638         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 639                 return false;
 640
 641         if (fe->backmost)
 642                 return true;
 643
 644         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 645                 la < fe->headoffset;
 646 }
 647
 648 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 649                                 struct page *page, struct page **pagepool)
 650 {
 651         struct inode *const inode = fe->inode;
 652         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 653         struct erofs_map_blocks *const map = &fe->map;
 654         struct z_erofs_collector *const clt = &fe->clt;
 655         const loff_t offset = page_offset(page);
 656         bool tight = true;
 657
 658         enum z_erofs_cache_alloctype cache_strategy;
 659         enum z_erofs_page_type page_type;
 660         unsigned int cur, end, spiltted, index;
 661         int err = 0;
 662
 663         /* register locked file pages as online pages in pack */
 664         z_erofs_onlinepage_init(page);
 665
 666         spiltted = 0;
 667         end = PAGE_SIZE;
 668 repeat:
 669         cur = end - 1;
 670
 671         /* lucky, within the range of the current map_blocks */
 672         if (offset + cur >= map->m_la &&
 673             offset + cur < map->m_la + map->m_llen) {
 674                 /* didn't get a valid collection previously (very rare) */
 675                 if (!clt->cl)
 676                         goto restart_now;
 677                 goto hitted;
 678         }
 679
 680         /* go ahead the next map_blocks */
 681         erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur);
 682
 683         if (z_erofs_collector_end(clt))
 684                 fe->backmost = false;
 685
 686         map->m_la = offset + cur;
 687         map->m_llen = 0;
 688         err = z_erofs_map_blocks_iter(inode, map, 0);
 689         if (err)
 690                 goto err_out;
 691
 692 restart_now:
 693         if (!(map->m_flags & EROFS_MAP_MAPPED))
 694                 goto hitted;
 695
 696         err = z_erofs_collector_begin(clt, inode, map);
 697         if (err)
 698                 goto err_out;
 699
 700         if (z_erofs_is_inline_pcluster(clt->pcl)) {
 701                 void *mp;
 702
 703                 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
 704                                         erofs_blknr(map->m_pa), EROFS_NO_KMAP);
 705                 if (IS_ERR(mp)) {
 706                         err = PTR_ERR(mp);
 707                         erofs_err(inode->i_sb,
 708                                   "failed to get inline page, err %d", err);
 709                         goto err_out;
 710                 }
 711                 get_page(fe->map.buf.page);
 712                 WRITE_ONCE(clt->pcl->compressed_pages[0], fe->map.buf.page);
 713                 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE;
 714         } else {
 715                 /* preload all compressed pages (can change mode if needed) */
 716                 if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy,
 717                                                map->m_la))
 718                         cache_strategy = TRYALLOC;
 719                 else
 720                         cache_strategy = DONTALLOC;
 721
 722                 preload_compressed_pages(clt, MNGD_MAPPING(sbi),
 723                                          cache_strategy, pagepool);
 724         }
 725
 726 hitted:
 727         /*
 728          * Ensure the current partial page belongs to this submit chain rather
 729          * than other concurrent submit chains or the noio(bypass) chain since
 730          * those chains are handled asynchronously thus the page cannot be used
 731          * for inplace I/O or pagevec (should be processed in strict order.)
 732          */
 733         tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED &&
 734                   clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE);
 735
 736         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 737         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 738                 zero_user_segment(page, cur, end);
 739                 goto next_part;
 740         }
 741
 742         /* let's derive page type */
 743         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 744                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 745                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 746                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 747
 748         if (cur)
 749                 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 750
 751 retry:
 752         err = z_erofs_attach_page(clt, page, page_type,
 753                                   clt->mode >= COLLECT_PRIMARY_FOLLOWED);
 754         /* should allocate an additional short-lived page for pagevec */
 755         if (err == -EAGAIN) {
 756                 struct page *const newpage =
 757                                 alloc_page(GFP_NOFS | __GFP_NOFAIL);
 758
 759                 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE);
 760                 err = z_erofs_attach_page(clt, newpage,
 761                                           Z_EROFS_PAGE_TYPE_EXCLUSIVE, true);
 762                 if (!err)
 763                         goto retry;
 764         }
 765
 766         if (err)
 767                 goto err_out;
 768
 769         index = page->index - (map->m_la >> PAGE_SHIFT);
 770
 771         z_erofs_onlinepage_fixup(page, index, true);
 772
 773         /* bump up the number of spiltted parts of a page */
 774         ++spiltted;
 775         /* also update nr_pages */
 776         clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1);
 777 next_part:
 778         /* can be used for verification */
 779         map->m_llen = offset + cur - map->m_la;
 780
 781         end = cur;
 782         if (end > 0)
 783                 goto repeat;
 784
 785 out:
 786         z_erofs_onlinepage_endio(page);
 787
 788         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 789                   __func__, page, spiltted, map->m_llen);
 790         return err;
 791
 792         /* if some error occurred while processing this page */
 793 err_out:
 794         SetPageError(page);
 795         goto out;
 796 }
 797
 798 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
 799                                        unsigned int readahead_pages)
 800 {
 801         /* auto: enable for readpage, disable for readahead */
 802         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
 803             !readahead_pages)
 804                 return true;
 805
 806         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
 807             (readahead_pages <= sbi->opt.max_sync_decompress_pages))
 808                 return true;
 809
 810         return false;
 811 }
 812
 813 static bool z_erofs_page_is_invalidated(struct page *page)
 814 {
 815         return !page->mapping && !z_erofs_is_shortlived_page(page);
 816 }
 817
 818 static int z_erofs_decompress_pcluster(struct super_block *sb,
 819                                        struct z_erofs_pcluster *pcl,
 820                                        struct page **pagepool)
 821 {
 822         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 823         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 824         struct z_erofs_pagevec_ctor ctor;
 825         unsigned int i, inputsize, outputsize, llen, nr_pages;
 826         struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES];
 827         struct page **pages, **compressed_pages, *page;
 828
 829         enum z_erofs_page_type page_type;
 830         bool overlapped, partial;
 831         struct z_erofs_collection *cl;
 832         int err;
 833
 834         might_sleep();
 835         cl = z_erofs_primarycollection(pcl);
 836         DBG_BUGON(!READ_ONCE(cl->nr_pages));
 837
 838         mutex_lock(&cl->lock);
 839         nr_pages = cl->nr_pages;
 840
 841         if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) {
 842                 pages = pages_onstack;
 843         } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES &&
 844                    mutex_trylock(&z_pagemap_global_lock)) {
 845                 pages = z_pagemap_global;
 846         } else {
 847                 gfp_t gfp_flags = GFP_KERNEL;
 848
 849                 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES)
 850                         gfp_flags |= __GFP_NOFAIL;
 851
 852                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 853                                        gfp_flags);
 854
 855                 /* fallback to global pagemap for the lowmem scenario */
 856                 if (!pages) {
 857                         mutex_lock(&z_pagemap_global_lock);
 858                         pages = z_pagemap_global;
 859                 }
 860         }
 861
 862         for (i = 0; i < nr_pages; ++i)
 863                 pages[i] = NULL;
 864
 865         err = 0;
 866         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 867                                   cl->pagevec, 0);
 868
 869         for (i = 0; i < cl->vcnt; ++i) {
 870                 unsigned int pagenr;
 871
 872                 page = z_erofs_pagevec_dequeue(&ctor, &page_type);
 873
 874                 /* all pages in pagevec ought to be valid */
 875                 DBG_BUGON(!page);
 876                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 877
 878                 if (z_erofs_put_shortlivedpage(pagepool, page))
 879                         continue;
 880
 881                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 882                         pagenr = 0;
 883                 else
 884                         pagenr = z_erofs_onlinepage_index(page);
 885
 886                 DBG_BUGON(pagenr >= nr_pages);
 887
 888                 /*
 889                  * currently EROFS doesn't support multiref(dedup),
 890                  * so here erroring out one multiref page.
 891                  */
 892                 if (pages[pagenr]) {
 893                         DBG_BUGON(1);
 894                         SetPageError(pages[pagenr]);
 895                         z_erofs_onlinepage_endio(pages[pagenr]);
 896                         err = -EFSCORRUPTED;
 897                 }
 898                 pages[pagenr] = page;
 899         }
 900         z_erofs_pagevec_ctor_exit(&ctor, true);
 901
 902         overlapped = false;
 903         compressed_pages = pcl->compressed_pages;
 904
 905         for (i = 0; i < pclusterpages; ++i) {
 906                 unsigned int pagenr;
 907
 908                 page = compressed_pages[i];
 909                 /* all compressed pages ought to be valid */
 910                 DBG_BUGON(!page);
 911
 912                 if (z_erofs_is_inline_pcluster(pcl)) {
 913                         if (!PageUptodate(page))
 914                                 err = -EIO;
 915                         continue;
 916                 }
 917
 918                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 919                 if (!z_erofs_is_shortlived_page(page)) {
 920                         if (erofs_page_is_managed(sbi, page)) {
 921                                 if (!PageUptodate(page))
 922                                         err = -EIO;
 923                                 continue;
 924                         }
 925
 926                         /*
 927                          * only if non-head page can be selected
 928                          * for inplace decompression
 929                          */
 930                         pagenr = z_erofs_onlinepage_index(page);
 931
 932                         DBG_BUGON(pagenr >= nr_pages);
 933                         if (pages[pagenr]) {
 934                                 DBG_BUGON(1);
 935                                 SetPageError(pages[pagenr]);
 936                                 z_erofs_onlinepage_endio(pages[pagenr]);
 937                                 err = -EFSCORRUPTED;
 938                         }
 939                         pages[pagenr] = page;
 940
 941                         overlapped = true;
 942                 }
 943
 944                 /* PG_error needs checking for all non-managed pages */
 945                 if (PageError(page)) {
 946                         DBG_BUGON(PageUptodate(page));
 947                         err = -EIO;
 948                 }
 949         }
 950
 951         if (err)
 952                 goto out;
 953
 954         llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT;
 955         if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) {
 956                 outputsize = llen;
 957                 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH);
 958         } else {
 959                 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs;
 960                 partial = true;
 961         }
 962
 963         if (z_erofs_is_inline_pcluster(pcl))
 964                 inputsize = pcl->tailpacking_size;
 965         else
 966                 inputsize = pclusterpages * PAGE_SIZE;
 967
 968         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
 969                                         .sb = sb,
 970                                         .in = compressed_pages,
 971                                         .out = pages,
 972                                         .pageofs_in = pcl->pageofs_in,
 973                                         .pageofs_out = cl->pageofs,
 974                                         .inputsize = inputsize,
 975                                         .outputsize = outputsize,
 976                                         .alg = pcl->algorithmformat,
 977                                         .inplace_io = overlapped,
 978                                         .partial_decoding = partial
 979                                  }, pagepool);
 980
 981 out:
 982         /* must handle all compressed pages before actual file pages */
 983         if (z_erofs_is_inline_pcluster(pcl)) {
 984                 page = compressed_pages[0];
 985                 WRITE_ONCE(compressed_pages[0], NULL);
 986                 put_page(page);
 987         } else {
 988                 for (i = 0; i < pclusterpages; ++i) {
 989                         page = compressed_pages[i];
 990
 991                         if (erofs_page_is_managed(sbi, page))
 992                                 continue;
 993
 994                         /* recycle all individual short-lived pages */
 995                         (void)z_erofs_put_shortlivedpage(pagepool, page);
 996                         WRITE_ONCE(compressed_pages[i], NULL);
 997                 }
 998         }
 999
1000         for (i = 0; i < nr_pages; ++i) {
1001                 page = pages[i];
1002                 if (!page)
1003                         continue;
1004
1005                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1006
1007                 /* recycle all individual short-lived pages */
1008                 if (z_erofs_put_shortlivedpage(pagepool, page))
1009                         continue;
1010
1011                 if (err < 0)
1012                         SetPageError(page);
1013
1014                 z_erofs_onlinepage_endio(page);
1015         }
1016
1017         if (pages == z_pagemap_global)
1018                 mutex_unlock(&z_pagemap_global_lock);
1019         else if (pages != pages_onstack)
1020                 kvfree(pages);
1021
1022         cl->nr_pages = 0;
1023         cl->vcnt = 0;
1024
1025         /* all cl locks MUST be taken before the following line */
1026         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1027
1028         /* all cl locks SHOULD be released right now */
1029         mutex_unlock(&cl->lock);
1030
1031         z_erofs_collection_put(cl);
1032         return err;
1033 }
1034
1035 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1036                                      struct page **pagepool)
1037 {
1038         z_erofs_next_pcluster_t owned = io->head;
1039
1040         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1041                 struct z_erofs_pcluster *pcl;
1042
1043                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1044                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1045
1046                 /* no possible that 'owned' equals NULL */
1047                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1048
1049                 pcl = container_of(owned, struct z_erofs_pcluster, next);
1050                 owned = READ_ONCE(pcl->next);
1051
1052                 z_erofs_decompress_pcluster(io->sb, pcl, pagepool);
1053         }
1054 }
1055
1056 static void z_erofs_decompressqueue_work(struct work_struct *work)
1057 {
1058         struct z_erofs_decompressqueue *bgq =
1059                 container_of(work, struct z_erofs_decompressqueue, u.work);
1060         struct page *pagepool = NULL;
1061
1062         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1063         z_erofs_decompress_queue(bgq, &pagepool);
1064
1065         erofs_release_pages(&pagepool);
1066         kvfree(bgq);
1067 }
1068
1069 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1070                                        bool sync, int bios)
1071 {
1072         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1073
1074         /* wake up the caller thread for sync decompression */
1075         if (sync) {
1076                 unsigned long flags;
1077
1078                 spin_lock_irqsave(&io->u.wait.lock, flags);
1079                 if (!atomic_add_return(bios, &io->pending_bios))
1080                         wake_up_locked(&io->u.wait);
1081                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
1082                 return;
1083         }
1084
1085         if (atomic_add_return(bios, &io->pending_bios))
1086                 return;
1087         /* Use workqueue and sync decompression for atomic contexts only */
1088         if (in_atomic() || irqs_disabled()) {
1089                 queue_work(z_erofs_workqueue, &io->u.work);
1090                 /* enable sync decompression for readahead */
1091                 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1092                         sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1093                 return;
1094         }
1095         z_erofs_decompressqueue_work(&io->u.work);
1096 }
1097
1098 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1099                                                unsigned int nr,
1100                                                struct page **pagepool,
1101                                                struct address_space *mc,
1102                                                gfp_t gfp)
1103 {
1104         const pgoff_t index = pcl->obj.index;
1105         bool tocache = false;
1106
1107         struct address_space *mapping;
1108         struct page *oldpage, *page;
1109
1110         compressed_page_t t;
1111         int justfound;
1112
1113 repeat:
1114         page = READ_ONCE(pcl->compressed_pages[nr]);
1115         oldpage = page;
1116
1117         if (!page)
1118                 goto out_allocpage;
1119
1120         /* process the target tagged pointer */
1121         t = tagptr_init(compressed_page_t, page);
1122         justfound = tagptr_unfold_tags(t);
1123         page = tagptr_unfold_ptr(t);
1124
1125         /*
1126          * preallocated cached pages, which is used to avoid direct reclaim
1127          * otherwise, it will go inplace I/O path instead.
1128          */
1129         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1130                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1131                 set_page_private(page, 0);
1132                 tocache = true;
1133                 goto out_tocache;
1134         }
1135         mapping = READ_ONCE(page->mapping);
1136
1137         /*
1138          * file-backed online pages in plcuster are all locked steady,
1139          * therefore it is impossible for `mapping' to be NULL.
1140          */
1141         if (mapping && mapping != mc)
1142                 /* ought to be unmanaged pages */
1143                 goto out;
1144
1145         /* directly return for shortlived page as well */
1146         if (z_erofs_is_shortlived_page(page))
1147                 goto out;
1148
1149         lock_page(page);
1150
1151         /* only true if page reclaim goes wrong, should never happen */
1152         DBG_BUGON(justfound && PagePrivate(page));
1153
1154         /* the page is still in manage cache */
1155         if (page->mapping == mc) {
1156                 WRITE_ONCE(pcl->compressed_pages[nr], page);
1157
1158                 ClearPageError(page);
1159                 if (!PagePrivate(page)) {
1160                         /*
1161                          * impossible to be !PagePrivate(page) for
1162                          * the current restriction as well if
1163                          * the page is already in compressed_pages[].
1164                          */
1165                         DBG_BUGON(!justfound);
1166
1167                         justfound = 0;
1168                         set_page_private(page, (unsigned long)pcl);
1169                         SetPagePrivate(page);
1170                 }
1171
1172                 /* no need to submit io if it is already up-to-date */
1173                 if (PageUptodate(page)) {
1174                         unlock_page(page);
1175                         page = NULL;
1176                 }
1177                 goto out;
1178         }
1179
1180         /*
1181          * the managed page has been truncated, it's unsafe to
1182          * reuse this one, let's allocate a new cache-managed page.
1183          */
1184         DBG_BUGON(page->mapping);
1185         DBG_BUGON(!justfound);
1186
1187         tocache = true;
1188         unlock_page(page);
1189         put_page(page);
1190 out_allocpage:
1191         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1192         if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) {
1193                 erofs_pagepool_add(pagepool, page);
1194                 cond_resched();
1195                 goto repeat;
1196         }
1197 out_tocache:
1198         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1199                 /* turn into temporary page if fails (1 ref) */
1200                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1201                 goto out;
1202         }
1203         attach_page_private(page, pcl);
1204         /* drop a refcount added by allocpage (then we have 2 refs here) */
1205         put_page(page);
1206
1207 out:    /* the only exit (for tracing and debugging) */
1208         return page;
1209 }
1210
1211 static struct z_erofs_decompressqueue *
1212 jobqueue_init(struct super_block *sb,
1213               struct z_erofs_decompressqueue *fgq, bool *fg)
1214 {
1215         struct z_erofs_decompressqueue *q;
1216
1217         if (fg && !*fg) {
1218                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1219                 if (!q) {
1220                         *fg = true;
1221                         goto fg_out;
1222                 }
1223                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1224         } else {
1225 fg_out:
1226                 q = fgq;
1227                 init_waitqueue_head(&fgq->u.wait);
1228                 atomic_set(&fgq->pending_bios, 0);
1229         }
1230         q->sb = sb;
1231         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1232         return q;
1233 }
1234
1235 /* define decompression jobqueue types */
1236 enum {
1237         JQ_BYPASS,
1238         JQ_SUBMIT,
1239         NR_JOBQUEUES,
1240 };
1241
1242 static void *jobqueueset_init(struct super_block *sb,
1243                               struct z_erofs_decompressqueue *q[],
1244                               struct z_erofs_decompressqueue *fgq, bool *fg)
1245 {
1246         /*
1247          * if managed cache is enabled, bypass jobqueue is needed,
1248          * no need to read from device for all pclusters in this queue.
1249          */
1250         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1251         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1252
1253         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1254 }
1255
1256 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1257                                     z_erofs_next_pcluster_t qtail[],
1258                                     z_erofs_next_pcluster_t owned_head)
1259 {
1260         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1261         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1262
1263         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1264         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1265                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1266
1267         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1268
1269         WRITE_ONCE(*submit_qtail, owned_head);
1270         WRITE_ONCE(*bypass_qtail, &pcl->next);
1271
1272         qtail[JQ_BYPASS] = &pcl->next;
1273 }
1274
1275 static void z_erofs_decompressqueue_endio(struct bio *bio)
1276 {
1277         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
1278         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
1279         blk_status_t err = bio->bi_status;
1280         struct bio_vec *bvec;
1281         struct bvec_iter_all iter_all;
1282
1283         bio_for_each_segment_all(bvec, bio, iter_all) {
1284                 struct page *page = bvec->bv_page;
1285
1286                 DBG_BUGON(PageUptodate(page));
1287                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1288
1289                 if (err)
1290                         SetPageError(page);
1291
1292                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1293                         if (!err)
1294                                 SetPageUptodate(page);
1295                         unlock_page(page);
1296                 }
1297         }
1298         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
1299         bio_put(bio);
1300 }
1301
1302 static void z_erofs_submit_queue(struct super_block *sb,
1303                                  struct z_erofs_decompress_frontend *f,
1304                                  struct page **pagepool,
1305                                  struct z_erofs_decompressqueue *fgq,
1306                                  bool *force_fg)
1307 {
1308         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1309         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1310         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1311         void *bi_private;
1312         z_erofs_next_pcluster_t owned_head = f->clt.owned_head;
1313         /* bio is NULL initially, so no need to initialize last_{index,bdev} */
1314         pgoff_t last_index;
1315         struct block_device *last_bdev;
1316         unsigned int nr_bios = 0;
1317         struct bio *bio = NULL;
1318
1319         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1320         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1321         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1322
1323         /* by default, all need io submission */
1324         q[JQ_SUBMIT]->head = owned_head;
1325
1326         do {
1327                 struct erofs_map_dev mdev;
1328                 struct z_erofs_pcluster *pcl;
1329                 pgoff_t cur, end;
1330                 unsigned int i = 0;
1331                 bool bypass = true;
1332
1333                 /* no possible 'owned_head' equals the following */
1334                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1335                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1336
1337                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1338
1339                 /* close the main owned chain at first */
1340                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1341                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1342                 if (z_erofs_is_inline_pcluster(pcl)) {
1343                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1344                         continue;
1345                 }
1346
1347                 /* no device id here, thus it will always succeed */
1348                 mdev = (struct erofs_map_dev) {
1349                         .m_pa = blknr_to_addr(pcl->obj.index),
1350                 };
1351                 (void)erofs_map_dev(sb, &mdev);
1352
1353                 cur = erofs_blknr(mdev.m_pa);
1354                 end = cur + pcl->pclusterpages;
1355
1356                 do {
1357                         struct page *page;
1358
1359                         page = pickup_page_for_submission(pcl, i++, pagepool,
1360                                                           MNGD_MAPPING(sbi),
1361                                                           GFP_NOFS);
1362                         if (!page)
1363                                 continue;
1364
1365                         if (bio && (cur != last_index + 1 ||
1366                                     last_bdev != mdev.m_bdev)) {
1367 submit_bio_retry:
1368                                 submit_bio(bio);
1369                                 bio = NULL;
1370                         }
1371
1372                         if (!bio) {
1373                                 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS);
1374                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1375
1376                                 bio_set_dev(bio, mdev.m_bdev);
1377                                 last_bdev = mdev.m_bdev;
1378                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1379                                         LOG_SECTORS_PER_BLOCK;
1380                                 bio->bi_private = bi_private;
1381                                 bio->bi_opf = REQ_OP_READ;
1382                                 if (f->readahead)
1383                                         bio->bi_opf |= REQ_RAHEAD;
1384                                 ++nr_bios;
1385                         }
1386
1387                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1388                                 goto submit_bio_retry;
1389
1390                         last_index = cur;
1391                         bypass = false;
1392                 } while (++cur < end);
1393
1394                 if (!bypass)
1395                         qtail[JQ_SUBMIT] = &pcl->next;
1396                 else
1397                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1398         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1399
1400         if (bio)
1401                 submit_bio(bio);
1402
1403         /*
1404          * although background is preferred, no one is pending for submission.
1405          * don't issue workqueue for decompression but drop it directly instead.
1406          */
1407         if (!*force_fg && !nr_bios) {
1408                 kvfree(q[JQ_SUBMIT]);
1409                 return;
1410         }
1411         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1412 }
1413
1414 static void z_erofs_runqueue(struct super_block *sb,
1415                              struct z_erofs_decompress_frontend *f,
1416                              struct page **pagepool, bool force_fg)
1417 {
1418         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1419
1420         if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL)
1421                 return;
1422         z_erofs_submit_queue(sb, f, pagepool, io, &force_fg);
1423
1424         /* handle bypass queue (no i/o pclusters) immediately */
1425         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1426
1427         if (!force_fg)
1428                 return;
1429
1430         /* wait until all bios are completed */
1431         io_wait_event(io[JQ_SUBMIT].u.wait,
1432                       !atomic_read(&io[JQ_SUBMIT].pending_bios));
1433
1434         /* handle synchronous decompress queue in the caller context */
1435         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1436 }
1437
1438 /*
1439  * Since partial uptodate is still unimplemented for now, we have to use
1440  * approximate readmore strategies as a start.
1441  */
1442 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1443                                       struct readahead_control *rac,
1444                                       erofs_off_t end,
1445                                       struct page **pagepool,
1446                                       bool backmost)
1447 {
1448         struct inode *inode = f->inode;
1449         struct erofs_map_blocks *map = &f->map;
1450         erofs_off_t cur;
1451         int err;
1452
1453         if (backmost) {
1454                 map->m_la = end;
1455                 err = z_erofs_map_blocks_iter(inode, map,
1456                                               EROFS_GET_BLOCKS_READMORE);
1457                 if (err)
1458                         return;
1459
1460                 /* expend ra for the trailing edge if readahead */
1461                 if (rac) {
1462                         loff_t newstart = readahead_pos(rac);
1463
1464                         cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1465                         readahead_expand(rac, newstart, cur - newstart);
1466                         return;
1467                 }
1468                 end = round_up(end, PAGE_SIZE);
1469         } else {
1470                 end = round_up(map->m_la, PAGE_SIZE);
1471
1472                 if (!map->m_llen)
1473                         return;
1474         }
1475
1476         cur = map->m_la + map->m_llen - 1;
1477         while (cur >= end) {
1478                 pgoff_t index = cur >> PAGE_SHIFT;
1479                 struct page *page;
1480
1481                 page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1482                 if (!page)
1483                         goto skip;
1484
1485                 if (PageUptodate(page)) {
1486                         unlock_page(page);
1487                         put_page(page);
1488                         goto skip;
1489                 }
1490
1491                 err = z_erofs_do_read_page(f, page, pagepool);
1492                 if (err)
1493                         erofs_err(inode->i_sb,
1494                                   "readmore error at page %lu @ nid %llu",
1495                                   index, EROFS_I(inode)->nid);
1496                 put_page(page);
1497 skip:
1498                 if (cur < PAGE_SIZE)
1499                         break;
1500                 cur = (index << PAGE_SHIFT) - 1;
1501         }
1502 }
1503
1504 static int z_erofs_readpage(struct file *file, struct page *page)
1505 {
1506         struct inode *const inode = page->mapping->host;
1507         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1508         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1509         struct page *pagepool = NULL;
1510         int err;
1511
1512         trace_erofs_readpage(page, false);
1513         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1514
1515         z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1516                                   &pagepool, true);
1517         err = z_erofs_do_read_page(&f, page, &pagepool);
1518         z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1519
1520         (void)z_erofs_collector_end(&f.clt);
1521
1522         /* if some compressed cluster ready, need submit them anyway */
1523         z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1524                          z_erofs_get_sync_decompress_policy(sbi, 0));
1525
1526         if (err)
1527                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1528
1529         erofs_put_metabuf(&f.map.buf);
1530         erofs_release_pages(&pagepool);
1531         return err;
1532 }
1533
1534 static void z_erofs_readahead(struct readahead_control *rac)
1535 {
1536         struct inode *const inode = rac->mapping->host;
1537         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1538         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1539         struct page *pagepool = NULL, *head = NULL, *page;
1540         unsigned int nr_pages;
1541
1542         f.readahead = true;
1543         f.headoffset = readahead_pos(rac);
1544
1545         z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1546                                   readahead_length(rac) - 1, &pagepool, true);
1547         nr_pages = readahead_count(rac);
1548         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1549
1550         while ((page = readahead_page(rac))) {
1551                 set_page_private(page, (unsigned long)head);
1552                 head = page;
1553         }
1554
1555         while (head) {
1556                 struct page *page = head;
1557                 int err;
1558
1559                 /* traversal in reverse order */
1560                 head = (void *)page_private(page);
1561
1562                 err = z_erofs_do_read_page(&f, page, &pagepool);
1563                 if (err)
1564                         erofs_err(inode->i_sb,
1565                                   "readahead error at page %lu @ nid %llu",
1566                                   page->index, EROFS_I(inode)->nid);
1567                 put_page(page);
1568         }
1569         z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1570         (void)z_erofs_collector_end(&f.clt);
1571
1572         z_erofs_runqueue(inode->i_sb, &f, &pagepool,
1573                          z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1574         erofs_put_metabuf(&f.map.buf);
1575         erofs_release_pages(&pagepool);
1576 }
1577
1578 const struct address_space_operations z_erofs_aops = {
1579         .readpage = z_erofs_readpage,
1580         .readahead = z_erofs_readahead,
1581 };