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  * Copyright (C) 2022 Alibaba Cloud
   6  */
   7 #include "compress.h"
   8 #include <linux/psi.h>
   9 #include <linux/cpuhotplug.h>
  10 #include <trace/events/erofs.h>
  11
  12 #define Z_EROFS_PCLUSTER_MAX_PAGES      (Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE)
  13 #define Z_EROFS_INLINE_BVECS            2
  14
  15 /*
  16  * let's leave a type here in case of introducing
  17  * another tagged pointer later.
  18  */
  19 typedef void *z_erofs_next_pcluster_t;
  20
  21 struct z_erofs_bvec {
  22         struct page *page;
  23         int offset;
  24         unsigned int end;
  25 };
  26
  27 #define __Z_EROFS_BVSET(name, total) \
  28 struct name { \
  29         /* point to the next page which contains the following bvecs */ \
  30         struct page *nextpage; \
  31         struct z_erofs_bvec bvec[total]; \
  32 }
  33 __Z_EROFS_BVSET(z_erofs_bvset,);
  34 __Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS);
  35
  36 /*
  37  * Structure fields follow one of the following exclusion rules.
  38  *
  39  * I: Modifiable by initialization/destruction paths and read-only
  40  *    for everyone else;
  41  *
  42  * L: Field should be protected by the pcluster lock;
  43  *
  44  * A: Field should be accessed / updated in atomic for parallelized code.
  45  */
  46 struct z_erofs_pcluster {
  47         struct erofs_workgroup obj;
  48         struct mutex lock;
  49
  50         /* A: point to next chained pcluster or TAILs */
  51         z_erofs_next_pcluster_t next;
  52
  53         /* L: the maximum decompression size of this round */
  54         unsigned int length;
  55
  56         /* L: total number of bvecs */
  57         unsigned int vcnt;
  58
  59         /* I: page offset of start position of decompression */
  60         unsigned short pageofs_out;
  61
  62         /* I: page offset of inline compressed data */
  63         unsigned short pageofs_in;
  64
  65         union {
  66                 /* L: inline a certain number of bvec for bootstrap */
  67                 struct z_erofs_bvset_inline bvset;
  68
  69                 /* I: can be used to free the pcluster by RCU. */
  70                 struct rcu_head rcu;
  71         };
  72
  73         union {
  74                 /* I: physical cluster size in pages */
  75                 unsigned short pclusterpages;
  76
  77                 /* I: tailpacking inline compressed size */
  78                 unsigned short tailpacking_size;
  79         };
  80
  81         /* I: compression algorithm format */
  82         unsigned char algorithmformat;
  83
  84         /* L: whether partial decompression or not */
  85         bool partial;
  86
  87         /* L: indicate several pageofs_outs or not */
  88         bool multibases;
  89
  90         /* A: compressed bvecs (can be cached or inplaced pages) */
  91         struct z_erofs_bvec compressed_bvecs[];
  92 };
  93
  94 /* the end of a chain of pclusters */
  95 #define Z_EROFS_PCLUSTER_TAIL           ((void *) 0x700 + POISON_POINTER_DELTA)
  96 #define Z_EROFS_PCLUSTER_NIL            (NULL)
  97
  98 struct z_erofs_decompressqueue {
  99         struct super_block *sb;
 100         atomic_t pending_bios;
 101         z_erofs_next_pcluster_t head;
 102
 103         union {
 104                 struct completion done;
 105                 struct work_struct work;
 106                 struct kthread_work kthread_work;
 107         } u;
 108         bool eio, sync;
 109 };
 110
 111 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)
 112 {
 113         return !pcl->obj.index;
 114 }
 115
 116 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)
 117 {
 118         if (z_erofs_is_inline_pcluster(pcl))
 119                 return 1;
 120         return pcl->pclusterpages;
 121 }
 122
 123 /*
 124  * bit 30: I/O error occurred on this page
 125  * bit 0 - 29: remaining parts to complete this page
 126  */
 127 #define Z_EROFS_PAGE_EIO                        (1 << 30)
 128
 129 static inline void z_erofs_onlinepage_init(struct page *page)
 130 {
 131         union {
 132                 atomic_t o;
 133                 unsigned long v;
 134         } u = { .o = ATOMIC_INIT(1) };
 135
 136         set_page_private(page, u.v);
 137         smp_wmb();
 138         SetPagePrivate(page);
 139 }
 140
 141 static inline void z_erofs_onlinepage_split(struct page *page)
 142 {
 143         atomic_inc((atomic_t *)&page->private);
 144 }
 145
 146 static void z_erofs_onlinepage_endio(struct page *page, int err)
 147 {
 148         int orig, v;
 149
 150         DBG_BUGON(!PagePrivate(page));
 151
 152         do {
 153                 orig = atomic_read((atomic_t *)&page->private);
 154                 v = (orig - 1) | (err ? Z_EROFS_PAGE_EIO : 0);
 155         } while (atomic_cmpxchg((atomic_t *)&page->private, orig, v) != orig);
 156
 157         if (!(v & ~Z_EROFS_PAGE_EIO)) {
 158                 set_page_private(page, 0);
 159                 ClearPagePrivate(page);
 160                 if (!(v & Z_EROFS_PAGE_EIO))
 161                         SetPageUptodate(page);
 162                 unlock_page(page);
 163         }
 164 }
 165
 166 #define Z_EROFS_ONSTACK_PAGES           32
 167
 168 /*
 169  * since pclustersize is variable for big pcluster feature, introduce slab
 170  * pools implementation for different pcluster sizes.
 171  */
 172 struct z_erofs_pcluster_slab {
 173         struct kmem_cache *slab;
 174         unsigned int maxpages;
 175         char name[48];
 176 };
 177
 178 #define _PCLP(n) { .maxpages = n }
 179
 180 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
 181         _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
 182         _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
 183 };
 184
 185 struct z_erofs_bvec_iter {
 186         struct page *bvpage;
 187         struct z_erofs_bvset *bvset;
 188         unsigned int nr, cur;
 189 };
 190
 191 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
 192 {
 193         if (iter->bvpage)
 194                 kunmap_local(iter->bvset);
 195         return iter->bvpage;
 196 }
 197
 198 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
 199 {
 200         unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
 201         /* have to access nextpage in advance, otherwise it will be unmapped */
 202         struct page *nextpage = iter->bvset->nextpage;
 203         struct page *oldpage;
 204
 205         DBG_BUGON(!nextpage);
 206         oldpage = z_erofs_bvec_iter_end(iter);
 207         iter->bvpage = nextpage;
 208         iter->bvset = kmap_local_page(nextpage);
 209         iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
 210         iter->cur = 0;
 211         return oldpage;
 212 }
 213
 214 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
 215                                     struct z_erofs_bvset_inline *bvset,
 216                                     unsigned int bootstrap_nr,
 217                                     unsigned int cur)
 218 {
 219         *iter = (struct z_erofs_bvec_iter) {
 220                 .nr = bootstrap_nr,
 221                 .bvset = (struct z_erofs_bvset *)bvset,
 222         };
 223
 224         while (cur > iter->nr) {
 225                 cur -= iter->nr;
 226                 z_erofs_bvset_flip(iter);
 227         }
 228         iter->cur = cur;
 229 }
 230
 231 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
 232                                 struct z_erofs_bvec *bvec,
 233                                 struct page **candidate_bvpage,
 234                                 struct page **pagepool)
 235 {
 236         if (iter->cur >= iter->nr) {
 237                 struct page *nextpage = *candidate_bvpage;
 238
 239                 if (!nextpage) {
 240                         nextpage = erofs_allocpage(pagepool, GFP_NOFS);
 241                         if (!nextpage)
 242                                 return -ENOMEM;
 243                         set_page_private(nextpage, Z_EROFS_SHORTLIVED_PAGE);
 244                 }
 245                 DBG_BUGON(iter->bvset->nextpage);
 246                 iter->bvset->nextpage = nextpage;
 247                 z_erofs_bvset_flip(iter);
 248
 249                 iter->bvset->nextpage = NULL;
 250                 *candidate_bvpage = NULL;
 251         }
 252         iter->bvset->bvec[iter->cur++] = *bvec;
 253         return 0;
 254 }
 255
 256 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,
 257                                  struct z_erofs_bvec *bvec,
 258                                  struct page **old_bvpage)
 259 {
 260         if (iter->cur == iter->nr)
 261                 *old_bvpage = z_erofs_bvset_flip(iter);
 262         else
 263                 *old_bvpage = NULL;
 264         *bvec = iter->bvset->bvec[iter->cur++];
 265 }
 266
 267 static void z_erofs_destroy_pcluster_pool(void)
 268 {
 269         int i;
 270
 271         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 272                 if (!pcluster_pool[i].slab)
 273                         continue;
 274                 kmem_cache_destroy(pcluster_pool[i].slab);
 275                 pcluster_pool[i].slab = NULL;
 276         }
 277 }
 278
 279 static int z_erofs_create_pcluster_pool(void)
 280 {
 281         struct z_erofs_pcluster_slab *pcs;
 282         struct z_erofs_pcluster *a;
 283         unsigned int size;
 284
 285         for (pcs = pcluster_pool;
 286              pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
 287                 size = struct_size(a, compressed_bvecs, pcs->maxpages);
 288
 289                 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
 290                 pcs->slab = kmem_cache_create(pcs->name, size, 0,
 291                                               SLAB_RECLAIM_ACCOUNT, NULL);
 292                 if (pcs->slab)
 293                         continue;
 294
 295                 z_erofs_destroy_pcluster_pool();
 296                 return -ENOMEM;
 297         }
 298         return 0;
 299 }
 300
 301 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
 302 {
 303         int i;
 304
 305         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 306                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 307                 struct z_erofs_pcluster *pcl;
 308
 309                 if (nrpages > pcs->maxpages)
 310                         continue;
 311
 312                 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
 313                 if (!pcl)
 314                         return ERR_PTR(-ENOMEM);
 315                 pcl->pclusterpages = nrpages;
 316                 return pcl;
 317         }
 318         return ERR_PTR(-EINVAL);
 319 }
 320
 321 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
 322 {
 323         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 324         int i;
 325
 326         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 327                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 328
 329                 if (pclusterpages > pcs->maxpages)
 330                         continue;
 331
 332                 kmem_cache_free(pcs->slab, pcl);
 333                 return;
 334         }
 335         DBG_BUGON(1);
 336 }
 337
 338 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 339
 340 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
 341 static struct kthread_worker __rcu **z_erofs_pcpu_workers;
 342
 343 static void erofs_destroy_percpu_workers(void)
 344 {
 345         struct kthread_worker *worker;
 346         unsigned int cpu;
 347
 348         for_each_possible_cpu(cpu) {
 349                 worker = rcu_dereference_protected(
 350                                         z_erofs_pcpu_workers[cpu], 1);
 351                 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
 352                 if (worker)
 353                         kthread_destroy_worker(worker);
 354         }
 355         kfree(z_erofs_pcpu_workers);
 356 }
 357
 358 static struct kthread_worker *erofs_init_percpu_worker(int cpu)
 359 {
 360         struct kthread_worker *worker =
 361                 kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);
 362
 363         if (IS_ERR(worker))
 364                 return worker;
 365         if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))
 366                 sched_set_fifo_low(worker->task);
 367         return worker;
 368 }
 369
 370 static int erofs_init_percpu_workers(void)
 371 {
 372         struct kthread_worker *worker;
 373         unsigned int cpu;
 374
 375         z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),
 376                         sizeof(struct kthread_worker *), GFP_ATOMIC);
 377         if (!z_erofs_pcpu_workers)
 378                 return -ENOMEM;
 379
 380         for_each_online_cpu(cpu) {      /* could miss cpu{off,on}line? */
 381                 worker = erofs_init_percpu_worker(cpu);
 382                 if (!IS_ERR(worker))
 383                         rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
 384         }
 385         return 0;
 386 }
 387 #else
 388 static inline void erofs_destroy_percpu_workers(void) {}
 389 static inline int erofs_init_percpu_workers(void) { return 0; }
 390 #endif
 391
 392 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)
 393 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
 394 static enum cpuhp_state erofs_cpuhp_state;
 395
 396 static int erofs_cpu_online(unsigned int cpu)
 397 {
 398         struct kthread_worker *worker, *old;
 399
 400         worker = erofs_init_percpu_worker(cpu);
 401         if (IS_ERR(worker))
 402                 return PTR_ERR(worker);
 403
 404         spin_lock(&z_erofs_pcpu_worker_lock);
 405         old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
 406                         lockdep_is_held(&z_erofs_pcpu_worker_lock));
 407         if (!old)
 408                 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
 409         spin_unlock(&z_erofs_pcpu_worker_lock);
 410         if (old)
 411                 kthread_destroy_worker(worker);
 412         return 0;
 413 }
 414
 415 static int erofs_cpu_offline(unsigned int cpu)
 416 {
 417         struct kthread_worker *worker;
 418
 419         spin_lock(&z_erofs_pcpu_worker_lock);
 420         worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
 421                         lockdep_is_held(&z_erofs_pcpu_worker_lock));
 422         rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
 423         spin_unlock(&z_erofs_pcpu_worker_lock);
 424
 425         synchronize_rcu();
 426         if (worker)
 427                 kthread_destroy_worker(worker);
 428         return 0;
 429 }
 430
 431 static int erofs_cpu_hotplug_init(void)
 432 {
 433         int state;
 434
 435         state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 436                         "fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);
 437         if (state < 0)
 438                 return state;
 439
 440         erofs_cpuhp_state = state;
 441         return 0;
 442 }
 443
 444 static void erofs_cpu_hotplug_destroy(void)
 445 {
 446         if (erofs_cpuhp_state)
 447                 cpuhp_remove_state_nocalls(erofs_cpuhp_state);
 448 }
 449 #else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */
 450 static inline int erofs_cpu_hotplug_init(void) { return 0; }
 451 static inline void erofs_cpu_hotplug_destroy(void) {}
 452 #endif
 453
 454 void z_erofs_exit_zip_subsystem(void)
 455 {
 456         erofs_cpu_hotplug_destroy();
 457         erofs_destroy_percpu_workers();
 458         destroy_workqueue(z_erofs_workqueue);
 459         z_erofs_destroy_pcluster_pool();
 460 }
 461
 462 int __init z_erofs_init_zip_subsystem(void)
 463 {
 464         int err = z_erofs_create_pcluster_pool();
 465
 466         if (err)
 467                 goto out_error_pcluster_pool;
 468
 469         z_erofs_workqueue = alloc_workqueue("erofs_worker",
 470                         WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());
 471         if (!z_erofs_workqueue) {
 472                 err = -ENOMEM;
 473                 goto out_error_workqueue_init;
 474         }
 475
 476         err = erofs_init_percpu_workers();
 477         if (err)
 478                 goto out_error_pcpu_worker;
 479
 480         err = erofs_cpu_hotplug_init();
 481         if (err < 0)
 482                 goto out_error_cpuhp_init;
 483         return err;
 484
 485 out_error_cpuhp_init:
 486         erofs_destroy_percpu_workers();
 487 out_error_pcpu_worker:
 488         destroy_workqueue(z_erofs_workqueue);
 489 out_error_workqueue_init:
 490         z_erofs_destroy_pcluster_pool();
 491 out_error_pcluster_pool:
 492         return err;
 493 }
 494
 495 enum z_erofs_pclustermode {
 496         Z_EROFS_PCLUSTER_INFLIGHT,
 497         /*
 498          * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
 499          * could be dispatched into bypass queue later due to uptodated managed
 500          * pages. All related online pages cannot be reused for inplace I/O (or
 501          * bvpage) since it can be directly decoded without I/O submission.
 502          */
 503         Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
 504         /*
 505          * The pcluster was just linked to a decompression chain by us.  It can
 506          * also be linked with the remaining pclusters, which means if the
 507          * processing page is the tail page of a pcluster, this pcluster can
 508          * safely use the whole page (since the previous pcluster is within the
 509          * same chain) for in-place I/O, as illustrated below:
 510          *  ___________________________________________________
 511          * |  tail (partial) page  |    head (partial) page    |
 512          * |  (of the current pcl) |   (of the previous pcl)   |
 513          * |___PCLUSTER_FOLLOWED___|_____PCLUSTER_FOLLOWED_____|
 514          *
 515          * [  (*) the page above can be used as inplace I/O.   ]
 516          */
 517         Z_EROFS_PCLUSTER_FOLLOWED,
 518 };
 519
 520 struct z_erofs_decompress_frontend {
 521         struct inode *const inode;
 522         struct erofs_map_blocks map;
 523         struct z_erofs_bvec_iter biter;
 524
 525         struct page *pagepool;
 526         struct page *candidate_bvpage;
 527         struct z_erofs_pcluster *pcl;
 528         z_erofs_next_pcluster_t owned_head;
 529         enum z_erofs_pclustermode mode;
 530
 531         erofs_off_t headoffset;
 532
 533         /* a pointer used to pick up inplace I/O pages */
 534         unsigned int icur;
 535 };
 536
 537 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 538         .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 539         .mode = Z_EROFS_PCLUSTER_FOLLOWED }
 540
 541 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe)
 542 {
 543         unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy;
 544
 545         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 546                 return false;
 547
 548         if (!(fe->map.m_flags & EROFS_MAP_FULL_MAPPED))
 549                 return true;
 550
 551         if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 552             fe->map.m_la < fe->headoffset)
 553                 return true;
 554
 555         return false;
 556 }
 557
 558 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe)
 559 {
 560         struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
 561         struct z_erofs_pcluster *pcl = fe->pcl;
 562         bool shouldalloc = z_erofs_should_alloc_cache(fe);
 563         bool standalone = true;
 564         /*
 565          * optimistic allocation without direct reclaim since inplace I/O
 566          * can be used if low memory otherwise.
 567          */
 568         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 569                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 570         unsigned int i;
 571
 572         if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
 573                 return;
 574
 575         for (i = 0; i < pcl->pclusterpages; ++i) {
 576                 struct page *page;
 577                 void *t;        /* mark pages just found for debugging */
 578                 struct page *newpage = NULL;
 579
 580                 /* the compressed page was loaded before */
 581                 if (READ_ONCE(pcl->compressed_bvecs[i].page))
 582                         continue;
 583
 584                 page = find_get_page(mc, pcl->obj.index + i);
 585
 586                 if (page) {
 587                         t = (void *)((unsigned long)page | 1);
 588                 } else {
 589                         /* I/O is needed, no possible to decompress directly */
 590                         standalone = false;
 591                         if (!shouldalloc)
 592                                 continue;
 593
 594                         /*
 595                          * try to use cached I/O if page allocation
 596                          * succeeds or fallback to in-place I/O instead
 597                          * to avoid any direct reclaim.
 598                          */
 599                         newpage = erofs_allocpage(&fe->pagepool, gfp);
 600                         if (!newpage)
 601                                 continue;
 602                         set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE);
 603                         t = (void *)((unsigned long)newpage | 1);
 604                 }
 605
 606                 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t))
 607                         continue;
 608
 609                 if (page)
 610                         put_page(page);
 611                 else if (newpage)
 612                         erofs_pagepool_add(&fe->pagepool, newpage);
 613         }
 614
 615         /*
 616          * don't do inplace I/O if all compressed pages are available in
 617          * managed cache since it can be moved to the bypass queue instead.
 618          */
 619         if (standalone)
 620                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
 621 }
 622
 623 /* called by erofs_shrinker to get rid of all compressed_pages */
 624 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 625                                        struct erofs_workgroup *grp)
 626 {
 627         struct z_erofs_pcluster *const pcl =
 628                 container_of(grp, struct z_erofs_pcluster, obj);
 629         int i;
 630
 631         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 632         /*
 633          * refcount of workgroup is now freezed as 0,
 634          * therefore no need to worry about available decompression users.
 635          */
 636         for (i = 0; i < pcl->pclusterpages; ++i) {
 637                 struct page *page = pcl->compressed_bvecs[i].page;
 638
 639                 if (!page)
 640                         continue;
 641
 642                 /* block other users from reclaiming or migrating the page */
 643                 if (!trylock_page(page))
 644                         return -EBUSY;
 645
 646                 if (!erofs_page_is_managed(sbi, page))
 647                         continue;
 648
 649                 /* barrier is implied in the following 'unlock_page' */
 650                 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 651                 detach_page_private(page);
 652                 unlock_page(page);
 653         }
 654         return 0;
 655 }
 656
 657 static bool z_erofs_cache_release_folio(struct folio *folio, gfp_t gfp)
 658 {
 659         struct z_erofs_pcluster *pcl = folio_get_private(folio);
 660         bool ret;
 661         int i;
 662
 663         if (!folio_test_private(folio))
 664                 return true;
 665
 666         ret = false;
 667         spin_lock(&pcl->obj.lockref.lock);
 668         if (pcl->obj.lockref.count > 0)
 669                 goto out;
 670
 671         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 672         for (i = 0; i < pcl->pclusterpages; ++i) {
 673                 if (pcl->compressed_bvecs[i].page == &folio->page) {
 674                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 675                         ret = true;
 676                         break;
 677                 }
 678         }
 679         if (ret)
 680                 folio_detach_private(folio);
 681 out:
 682         spin_unlock(&pcl->obj.lockref.lock);
 683         return ret;
 684 }
 685
 686 /*
 687  * It will be called only on inode eviction. In case that there are still some
 688  * decompression requests in progress, wait with rescheduling for a bit here.
 689  * An extra lock could be introduced instead but it seems unnecessary.
 690  */
 691 static void z_erofs_cache_invalidate_folio(struct folio *folio,
 692                                            size_t offset, size_t length)
 693 {
 694         const size_t stop = length + offset;
 695
 696         /* Check for potential overflow in debug mode */
 697         DBG_BUGON(stop > folio_size(folio) || stop < length);
 698
 699         if (offset == 0 && stop == folio_size(folio))
 700                 while (!z_erofs_cache_release_folio(folio, GFP_NOFS))
 701                         cond_resched();
 702 }
 703
 704 static const struct address_space_operations z_erofs_cache_aops = {
 705         .release_folio = z_erofs_cache_release_folio,
 706         .invalidate_folio = z_erofs_cache_invalidate_folio,
 707 };
 708
 709 int erofs_init_managed_cache(struct super_block *sb)
 710 {
 711         struct inode *const inode = new_inode(sb);
 712
 713         if (!inode)
 714                 return -ENOMEM;
 715
 716         set_nlink(inode, 1);
 717         inode->i_size = OFFSET_MAX;
 718         inode->i_mapping->a_ops = &z_erofs_cache_aops;
 719         mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS);
 720         EROFS_SB(sb)->managed_cache = inode;
 721         return 0;
 722 }
 723
 724 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
 725                                    struct z_erofs_bvec *bvec)
 726 {
 727         struct z_erofs_pcluster *const pcl = fe->pcl;
 728
 729         while (fe->icur > 0) {
 730                 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
 731                              NULL, bvec->page)) {
 732                         pcl->compressed_bvecs[fe->icur] = *bvec;
 733                         return true;
 734                 }
 735         }
 736         return false;
 737 }
 738
 739 /* callers must be with pcluster lock held */
 740 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
 741                                struct z_erofs_bvec *bvec, bool exclusive)
 742 {
 743         int ret;
 744
 745         if (exclusive) {
 746                 /* give priority for inplaceio to use file pages first */
 747                 if (z_erofs_try_inplace_io(fe, bvec))
 748                         return 0;
 749                 /* otherwise, check if it can be used as a bvpage */
 750                 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
 751                     !fe->candidate_bvpage)
 752                         fe->candidate_bvpage = bvec->page;
 753         }
 754         ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage,
 755                                    &fe->pagepool);
 756         fe->pcl->vcnt += (ret >= 0);
 757         return ret;
 758 }
 759
 760 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
 761 {
 762         struct z_erofs_pcluster *pcl = f->pcl;
 763         z_erofs_next_pcluster_t *owned_head = &f->owned_head;
 764
 765         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 766         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 767                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 768                 *owned_head = &pcl->next;
 769                 /* so we can attach this pcluster to our submission chain. */
 770                 f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 771                 return;
 772         }
 773
 774         /* type 2, it belongs to an ongoing chain */
 775         f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
 776 }
 777
 778 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
 779 {
 780         struct erofs_map_blocks *map = &fe->map;
 781         bool ztailpacking = map->m_flags & EROFS_MAP_META;
 782         struct z_erofs_pcluster *pcl;
 783         struct erofs_workgroup *grp;
 784         int err;
 785
 786         if (!(map->m_flags & EROFS_MAP_ENCODED) ||
 787             (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) {
 788                 DBG_BUGON(1);
 789                 return -EFSCORRUPTED;
 790         }
 791
 792         /* no available pcluster, let's allocate one */
 793         pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
 794                                      map->m_plen >> PAGE_SHIFT);
 795         if (IS_ERR(pcl))
 796                 return PTR_ERR(pcl);
 797
 798         spin_lock_init(&pcl->obj.lockref.lock);
 799         pcl->obj.lockref.count = 1;     /* one ref for this request */
 800         pcl->algorithmformat = map->m_algorithmformat;
 801         pcl->length = 0;
 802         pcl->partial = true;
 803
 804         /* new pclusters should be claimed as type 1, primary and followed */
 805         pcl->next = fe->owned_head;
 806         pcl->pageofs_out = map->m_la & ~PAGE_MASK;
 807         fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 808
 809         /*
 810          * lock all primary followed works before visible to others
 811          * and mutex_trylock *never* fails for a new pcluster.
 812          */
 813         mutex_init(&pcl->lock);
 814         DBG_BUGON(!mutex_trylock(&pcl->lock));
 815
 816         if (ztailpacking) {
 817                 pcl->obj.index = 0;     /* which indicates ztailpacking */
 818                 pcl->pageofs_in = erofs_blkoff(fe->inode->i_sb, map->m_pa);
 819                 pcl->tailpacking_size = map->m_plen;
 820         } else {
 821                 pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 822
 823                 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
 824                 if (IS_ERR(grp)) {
 825                         err = PTR_ERR(grp);
 826                         goto err_out;
 827                 }
 828
 829                 if (grp != &pcl->obj) {
 830                         fe->pcl = container_of(grp,
 831                                         struct z_erofs_pcluster, obj);
 832                         err = -EEXIST;
 833                         goto err_out;
 834                 }
 835         }
 836         fe->owned_head = &pcl->next;
 837         fe->pcl = pcl;
 838         return 0;
 839
 840 err_out:
 841         mutex_unlock(&pcl->lock);
 842         z_erofs_free_pcluster(pcl);
 843         return err;
 844 }
 845
 846 static int z_erofs_pcluster_begin(struct z_erofs_decompress_frontend *fe)
 847 {
 848         struct erofs_map_blocks *map = &fe->map;
 849         struct super_block *sb = fe->inode->i_sb;
 850         erofs_blk_t blknr = erofs_blknr(sb, map->m_pa);
 851         struct erofs_workgroup *grp = NULL;
 852         int ret;
 853
 854         DBG_BUGON(fe->pcl);
 855
 856         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
 857         DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
 858
 859         if (!(map->m_flags & EROFS_MAP_META)) {
 860                 grp = erofs_find_workgroup(sb, blknr);
 861         } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
 862                 DBG_BUGON(1);
 863                 return -EFSCORRUPTED;
 864         }
 865
 866         if (grp) {
 867                 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 868                 ret = -EEXIST;
 869         } else {
 870                 ret = z_erofs_register_pcluster(fe);
 871         }
 872
 873         if (ret == -EEXIST) {
 874                 mutex_lock(&fe->pcl->lock);
 875                 z_erofs_try_to_claim_pcluster(fe);
 876         } else if (ret) {
 877                 return ret;
 878         }
 879
 880         z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
 881                                 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
 882         if (!z_erofs_is_inline_pcluster(fe->pcl)) {
 883                 /* bind cache first when cached decompression is preferred */
 884                 z_erofs_bind_cache(fe);
 885         } else {
 886                 void *mptr;
 887
 888                 mptr = erofs_read_metabuf(&map->buf, sb, blknr, EROFS_NO_KMAP);
 889                 if (IS_ERR(mptr)) {
 890                         ret = PTR_ERR(mptr);
 891                         erofs_err(sb, "failed to get inline data %d", ret);
 892                         return ret;
 893                 }
 894                 get_page(map->buf.page);
 895                 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, map->buf.page);
 896                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
 897         }
 898         /* file-backed inplace I/O pages are traversed in reverse order */
 899         fe->icur = z_erofs_pclusterpages(fe->pcl);
 900         return 0;
 901 }
 902
 903 /*
 904  * keep in mind that no referenced pclusters will be freed
 905  * only after a RCU grace period.
 906  */
 907 static void z_erofs_rcu_callback(struct rcu_head *head)
 908 {
 909         z_erofs_free_pcluster(container_of(head,
 910                         struct z_erofs_pcluster, rcu));
 911 }
 912
 913 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 914 {
 915         struct z_erofs_pcluster *const pcl =
 916                 container_of(grp, struct z_erofs_pcluster, obj);
 917
 918         call_rcu(&pcl->rcu, z_erofs_rcu_callback);
 919 }
 920
 921 static void z_erofs_pcluster_end(struct z_erofs_decompress_frontend *fe)
 922 {
 923         struct z_erofs_pcluster *pcl = fe->pcl;
 924
 925         if (!pcl)
 926                 return;
 927
 928         z_erofs_bvec_iter_end(&fe->biter);
 929         mutex_unlock(&pcl->lock);
 930
 931         if (fe->candidate_bvpage)
 932                 fe->candidate_bvpage = NULL;
 933
 934         /*
 935          * if all pending pages are added, don't hold its reference
 936          * any longer if the pcluster isn't hosted by ourselves.
 937          */
 938         if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
 939                 erofs_workgroup_put(&pcl->obj);
 940
 941         fe->pcl = NULL;
 942 }
 943
 944 static int z_erofs_read_fragment(struct super_block *sb, struct page *page,
 945                         unsigned int cur, unsigned int end, erofs_off_t pos)
 946 {
 947         struct inode *packed_inode = EROFS_SB(sb)->packed_inode;
 948         struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
 949         unsigned int cnt;
 950         u8 *src;
 951
 952         if (!packed_inode)
 953                 return -EFSCORRUPTED;
 954
 955         buf.inode = packed_inode;
 956         for (; cur < end; cur += cnt, pos += cnt) {
 957                 cnt = min_t(unsigned int, end - cur,
 958                             sb->s_blocksize - erofs_blkoff(sb, pos));
 959                 src = erofs_bread(&buf, erofs_blknr(sb, pos), EROFS_KMAP);
 960                 if (IS_ERR(src)) {
 961                         erofs_put_metabuf(&buf);
 962                         return PTR_ERR(src);
 963                 }
 964                 memcpy_to_page(page, cur, src + erofs_blkoff(sb, pos), cnt);
 965         }
 966         erofs_put_metabuf(&buf);
 967         return 0;
 968 }
 969
 970 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 971                                 struct page *page)
 972 {
 973         struct inode *const inode = fe->inode;
 974         struct erofs_map_blocks *const map = &fe->map;
 975         const loff_t offset = page_offset(page);
 976         bool tight = true, exclusive;
 977         unsigned int cur, end, len, split;
 978         int err = 0;
 979
 980         z_erofs_onlinepage_init(page);
 981
 982         split = 0;
 983         end = PAGE_SIZE;
 984 repeat:
 985         if (offset + end - 1 < map->m_la ||
 986             offset + end - 1 >= map->m_la + map->m_llen) {
 987                 z_erofs_pcluster_end(fe);
 988                 map->m_la = offset + end - 1;
 989                 map->m_llen = 0;
 990                 err = z_erofs_map_blocks_iter(inode, map, 0);
 991                 if (err)
 992                         goto out;
 993         }
 994
 995         cur = offset > map->m_la ? 0 : map->m_la - offset;
 996         /* bump split parts first to avoid several separate cases */
 997         ++split;
 998
 999         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
1000                 zero_user_segment(page, cur, end);
1001                 tight = false;
1002                 goto next_part;
1003         }
1004
1005         if (map->m_flags & EROFS_MAP_FRAGMENT) {
1006                 erofs_off_t fpos = offset + cur - map->m_la;
1007
1008                 len = min_t(unsigned int, map->m_llen - fpos, end - cur);
1009                 err = z_erofs_read_fragment(inode->i_sb, page, cur, cur + len,
1010                                 EROFS_I(inode)->z_fragmentoff + fpos);
1011                 if (err)
1012                         goto out;
1013                 tight = false;
1014                 goto next_part;
1015         }
1016
1017         if (!fe->pcl) {
1018                 err = z_erofs_pcluster_begin(fe);
1019                 if (err)
1020                         goto out;
1021         }
1022
1023         /*
1024          * Ensure the current partial page belongs to this submit chain rather
1025          * than other concurrent submit chains or the noio(bypass) chain since
1026          * those chains are handled asynchronously thus the page cannot be used
1027          * for inplace I/O or bvpage (should be processed in a strict order.)
1028          */
1029         tight &= (fe->mode > Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
1030         exclusive = (!cur && ((split <= 1) || tight));
1031         if (cur)
1032                 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
1033
1034         err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
1035                                         .page = page,
1036                                         .offset = offset - map->m_la,
1037                                         .end = end,
1038                                   }), exclusive);
1039         if (err)
1040                 goto out;
1041
1042         z_erofs_onlinepage_split(page);
1043         if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
1044                 fe->pcl->multibases = true;
1045         if (fe->pcl->length < offset + end - map->m_la) {
1046                 fe->pcl->length = offset + end - map->m_la;
1047                 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
1048         }
1049         if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
1050             !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
1051             fe->pcl->length == map->m_llen)
1052                 fe->pcl->partial = false;
1053 next_part:
1054         /* shorten the remaining extent to update progress */
1055         map->m_llen = offset + cur - map->m_la;
1056         map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
1057
1058         end = cur;
1059         if (end > 0)
1060                 goto repeat;
1061
1062 out:
1063         z_erofs_onlinepage_endio(page, err);
1064         return err;
1065 }
1066
1067 static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi,
1068                                        unsigned int readahead_pages)
1069 {
1070         /* auto: enable for read_folio, disable for readahead */
1071         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
1072             !readahead_pages)
1073                 return true;
1074
1075         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
1076             (readahead_pages <= sbi->opt.max_sync_decompress_pages))
1077                 return true;
1078
1079         return false;
1080 }
1081
1082 static bool z_erofs_page_is_invalidated(struct page *page)
1083 {
1084         return !page->mapping && !z_erofs_is_shortlived_page(page);
1085 }
1086
1087 struct z_erofs_decompress_backend {
1088         struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
1089         struct super_block *sb;
1090         struct z_erofs_pcluster *pcl;
1091
1092         /* pages with the longest decompressed length for deduplication */
1093         struct page **decompressed_pages;
1094         /* pages to keep the compressed data */
1095         struct page **compressed_pages;
1096
1097         struct list_head decompressed_secondary_bvecs;
1098         struct page **pagepool;
1099         unsigned int onstack_used, nr_pages;
1100 };
1101
1102 struct z_erofs_bvec_item {
1103         struct z_erofs_bvec bvec;
1104         struct list_head list;
1105 };
1106
1107 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
1108                                          struct z_erofs_bvec *bvec)
1109 {
1110         struct z_erofs_bvec_item *item;
1111         unsigned int pgnr;
1112
1113         if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK) &&
1114             (bvec->end == PAGE_SIZE ||
1115              bvec->offset + bvec->end == be->pcl->length)) {
1116                 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
1117                 DBG_BUGON(pgnr >= be->nr_pages);
1118                 if (!be->decompressed_pages[pgnr]) {
1119                         be->decompressed_pages[pgnr] = bvec->page;
1120                         return;
1121                 }
1122         }
1123
1124         /* (cold path) one pcluster is requested multiple times */
1125         item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
1126         item->bvec = *bvec;
1127         list_add(&item->list, &be->decompressed_secondary_bvecs);
1128 }
1129
1130 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
1131                                       int err)
1132 {
1133         unsigned int off0 = be->pcl->pageofs_out;
1134         struct list_head *p, *n;
1135
1136         list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
1137                 struct z_erofs_bvec_item *bvi;
1138                 unsigned int end, cur;
1139                 void *dst, *src;
1140
1141                 bvi = container_of(p, struct z_erofs_bvec_item, list);
1142                 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
1143                 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
1144                             bvi->bvec.end);
1145                 dst = kmap_local_page(bvi->bvec.page);
1146                 while (cur < end) {
1147                         unsigned int pgnr, scur, len;
1148
1149                         pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
1150                         DBG_BUGON(pgnr >= be->nr_pages);
1151
1152                         scur = bvi->bvec.offset + cur -
1153                                         ((pgnr << PAGE_SHIFT) - off0);
1154                         len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
1155                         if (!be->decompressed_pages[pgnr]) {
1156                                 err = -EFSCORRUPTED;
1157                                 cur += len;
1158                                 continue;
1159                         }
1160                         src = kmap_local_page(be->decompressed_pages[pgnr]);
1161                         memcpy(dst + cur, src + scur, len);
1162                         kunmap_local(src);
1163                         cur += len;
1164                 }
1165                 kunmap_local(dst);
1166                 z_erofs_onlinepage_endio(bvi->bvec.page, err);
1167                 list_del(p);
1168                 kfree(bvi);
1169         }
1170 }
1171
1172 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
1173 {
1174         struct z_erofs_pcluster *pcl = be->pcl;
1175         struct z_erofs_bvec_iter biter;
1176         struct page *old_bvpage;
1177         int i;
1178
1179         z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
1180         for (i = 0; i < pcl->vcnt; ++i) {
1181                 struct z_erofs_bvec bvec;
1182
1183                 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
1184
1185                 if (old_bvpage)
1186                         z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1187
1188                 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
1189                 z_erofs_do_decompressed_bvec(be, &bvec);
1190         }
1191
1192         old_bvpage = z_erofs_bvec_iter_end(&biter);
1193         if (old_bvpage)
1194                 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1195 }
1196
1197 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
1198                                   bool *overlapped)
1199 {
1200         struct z_erofs_pcluster *pcl = be->pcl;
1201         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1202         int i, err = 0;
1203
1204         *overlapped = false;
1205         for (i = 0; i < pclusterpages; ++i) {
1206                 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
1207                 struct page *page = bvec->page;
1208
1209                 /* compressed pages ought to be present before decompressing */
1210                 if (!page) {
1211                         DBG_BUGON(1);
1212                         continue;
1213                 }
1214                 be->compressed_pages[i] = page;
1215
1216                 if (z_erofs_is_inline_pcluster(pcl)) {
1217                         if (!PageUptodate(page))
1218                                 err = -EIO;
1219                         continue;
1220                 }
1221
1222                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1223                 if (!z_erofs_is_shortlived_page(page)) {
1224                         if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
1225                                 if (!PageUptodate(page))
1226                                         err = -EIO;
1227                                 continue;
1228                         }
1229                         z_erofs_do_decompressed_bvec(be, bvec);
1230                         *overlapped = true;
1231                 }
1232         }
1233
1234         if (err)
1235                 return err;
1236         return 0;
1237 }
1238
1239 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1240                                        int err)
1241 {
1242         struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1243         struct z_erofs_pcluster *pcl = be->pcl;
1244         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1245         const struct z_erofs_decompressor *decompressor =
1246                                 &erofs_decompressors[pcl->algorithmformat];
1247         unsigned int i, inputsize;
1248         int err2;
1249         struct page *page;
1250         bool overlapped;
1251
1252         mutex_lock(&pcl->lock);
1253         be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1254
1255         /* allocate (de)compressed page arrays if cannot be kept on stack */
1256         be->decompressed_pages = NULL;
1257         be->compressed_pages = NULL;
1258         be->onstack_used = 0;
1259         if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1260                 be->decompressed_pages = be->onstack_pages;
1261                 be->onstack_used = be->nr_pages;
1262                 memset(be->decompressed_pages, 0,
1263                        sizeof(struct page *) * be->nr_pages);
1264         }
1265
1266         if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1267                 be->compressed_pages = be->onstack_pages + be->onstack_used;
1268
1269         if (!be->decompressed_pages)
1270                 be->decompressed_pages =
1271                         kvcalloc(be->nr_pages, sizeof(struct page *),
1272                                  GFP_KERNEL | __GFP_NOFAIL);
1273         if (!be->compressed_pages)
1274                 be->compressed_pages =
1275                         kvcalloc(pclusterpages, sizeof(struct page *),
1276                                  GFP_KERNEL | __GFP_NOFAIL);
1277
1278         z_erofs_parse_out_bvecs(be);
1279         err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1280         if (err2)
1281                 err = err2;
1282         if (err)
1283                 goto out;
1284
1285         if (z_erofs_is_inline_pcluster(pcl))
1286                 inputsize = pcl->tailpacking_size;
1287         else
1288                 inputsize = pclusterpages * PAGE_SIZE;
1289
1290         err = decompressor->decompress(&(struct z_erofs_decompress_req) {
1291                                         .sb = be->sb,
1292                                         .in = be->compressed_pages,
1293                                         .out = be->decompressed_pages,
1294                                         .pageofs_in = pcl->pageofs_in,
1295                                         .pageofs_out = pcl->pageofs_out,
1296                                         .inputsize = inputsize,
1297                                         .outputsize = pcl->length,
1298                                         .alg = pcl->algorithmformat,
1299                                         .inplace_io = overlapped,
1300                                         .partial_decoding = pcl->partial,
1301                                         .fillgaps = pcl->multibases,
1302                                  }, be->pagepool);
1303
1304 out:
1305         /* must handle all compressed pages before actual file pages */
1306         if (z_erofs_is_inline_pcluster(pcl)) {
1307                 page = pcl->compressed_bvecs[0].page;
1308                 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1309                 put_page(page);
1310         } else {
1311                 for (i = 0; i < pclusterpages; ++i) {
1312                         page = pcl->compressed_bvecs[i].page;
1313
1314                         if (erofs_page_is_managed(sbi, page))
1315                                 continue;
1316
1317                         /* recycle all individual short-lived pages */
1318                         (void)z_erofs_put_shortlivedpage(be->pagepool, page);
1319                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1320                 }
1321         }
1322         if (be->compressed_pages < be->onstack_pages ||
1323             be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1324                 kvfree(be->compressed_pages);
1325         z_erofs_fill_other_copies(be, err);
1326
1327         for (i = 0; i < be->nr_pages; ++i) {
1328                 page = be->decompressed_pages[i];
1329                 if (!page)
1330                         continue;
1331
1332                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1333
1334                 /* recycle all individual short-lived pages */
1335                 if (z_erofs_put_shortlivedpage(be->pagepool, page))
1336                         continue;
1337                 z_erofs_onlinepage_endio(page, err);
1338         }
1339
1340         if (be->decompressed_pages != be->onstack_pages)
1341                 kvfree(be->decompressed_pages);
1342
1343         pcl->length = 0;
1344         pcl->partial = true;
1345         pcl->multibases = false;
1346         pcl->bvset.nextpage = NULL;
1347         pcl->vcnt = 0;
1348
1349         /* pcluster lock MUST be taken before the following line */
1350         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1351         mutex_unlock(&pcl->lock);
1352         return err;
1353 }
1354
1355 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1356                                      struct page **pagepool)
1357 {
1358         struct z_erofs_decompress_backend be = {
1359                 .sb = io->sb,
1360                 .pagepool = pagepool,
1361                 .decompressed_secondary_bvecs =
1362                         LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1363         };
1364         z_erofs_next_pcluster_t owned = io->head;
1365
1366         while (owned != Z_EROFS_PCLUSTER_TAIL) {
1367                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1368
1369                 be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1370                 owned = READ_ONCE(be.pcl->next);
1371
1372                 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1373                 if (z_erofs_is_inline_pcluster(be.pcl))
1374                         z_erofs_free_pcluster(be.pcl);
1375                 else
1376                         erofs_workgroup_put(&be.pcl->obj);
1377         }
1378 }
1379
1380 static void z_erofs_decompressqueue_work(struct work_struct *work)
1381 {
1382         struct z_erofs_decompressqueue *bgq =
1383                 container_of(work, struct z_erofs_decompressqueue, u.work);
1384         struct page *pagepool = NULL;
1385
1386         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL);
1387         z_erofs_decompress_queue(bgq, &pagepool);
1388         erofs_release_pages(&pagepool);
1389         kvfree(bgq);
1390 }
1391
1392 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1393 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
1394 {
1395         z_erofs_decompressqueue_work((struct work_struct *)work);
1396 }
1397 #endif
1398
1399 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1400                                        int bios)
1401 {
1402         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1403
1404         /* wake up the caller thread for sync decompression */
1405         if (io->sync) {
1406                 if (!atomic_add_return(bios, &io->pending_bios))
1407                         complete(&io->u.done);
1408                 return;
1409         }
1410
1411         if (atomic_add_return(bios, &io->pending_bios))
1412                 return;
1413         /* Use (kthread_)work and sync decompression for atomic contexts only */
1414         if (!in_task() || irqs_disabled() || rcu_read_lock_any_held()) {
1415 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1416                 struct kthread_worker *worker;
1417
1418                 rcu_read_lock();
1419                 worker = rcu_dereference(
1420                                 z_erofs_pcpu_workers[raw_smp_processor_id()]);
1421                 if (!worker) {
1422                         INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
1423                         queue_work(z_erofs_workqueue, &io->u.work);
1424                 } else {
1425                         kthread_queue_work(worker, &io->u.kthread_work);
1426                 }
1427                 rcu_read_unlock();
1428 #else
1429                 queue_work(z_erofs_workqueue, &io->u.work);
1430 #endif
1431                 /* enable sync decompression for readahead */
1432                 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1433                         sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1434                 return;
1435         }
1436         z_erofs_decompressqueue_work(&io->u.work);
1437 }
1438
1439 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1440                                                unsigned int nr,
1441                                                struct page **pagepool,
1442                                                struct address_space *mc)
1443 {
1444         const pgoff_t index = pcl->obj.index;
1445         gfp_t gfp = mapping_gfp_mask(mc);
1446         bool tocache = false;
1447
1448         struct address_space *mapping;
1449         struct page *oldpage, *page;
1450         int justfound;
1451
1452 repeat:
1453         page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1454         oldpage = page;
1455
1456         if (!page)
1457                 goto out_allocpage;
1458
1459         justfound = (unsigned long)page & 1UL;
1460         page = (struct page *)((unsigned long)page & ~1UL);
1461
1462         /*
1463          * preallocated cached pages, which is used to avoid direct reclaim
1464          * otherwise, it will go inplace I/O path instead.
1465          */
1466         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1467                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1468                 set_page_private(page, 0);
1469                 tocache = true;
1470                 goto out_tocache;
1471         }
1472         mapping = READ_ONCE(page->mapping);
1473
1474         /*
1475          * file-backed online pages in plcuster are all locked steady,
1476          * therefore it is impossible for `mapping' to be NULL.
1477          */
1478         if (mapping && mapping != mc)
1479                 /* ought to be unmanaged pages */
1480                 goto out;
1481
1482         /* directly return for shortlived page as well */
1483         if (z_erofs_is_shortlived_page(page))
1484                 goto out;
1485
1486         lock_page(page);
1487
1488         /* only true if page reclaim goes wrong, should never happen */
1489         DBG_BUGON(justfound && PagePrivate(page));
1490
1491         /* the page is still in manage cache */
1492         if (page->mapping == mc) {
1493                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1494
1495                 if (!PagePrivate(page)) {
1496                         /*
1497                          * impossible to be !PagePrivate(page) for
1498                          * the current restriction as well if
1499                          * the page is already in compressed_bvecs[].
1500                          */
1501                         DBG_BUGON(!justfound);
1502
1503                         justfound = 0;
1504                         set_page_private(page, (unsigned long)pcl);
1505                         SetPagePrivate(page);
1506                 }
1507
1508                 /* no need to submit io if it is already up-to-date */
1509                 if (PageUptodate(page)) {
1510                         unlock_page(page);
1511                         page = NULL;
1512                 }
1513                 goto out;
1514         }
1515
1516         /*
1517          * the managed page has been truncated, it's unsafe to
1518          * reuse this one, let's allocate a new cache-managed page.
1519          */
1520         DBG_BUGON(page->mapping);
1521         DBG_BUGON(!justfound);
1522
1523         tocache = true;
1524         unlock_page(page);
1525         put_page(page);
1526 out_allocpage:
1527         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1528         if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1529                                oldpage, page)) {
1530                 erofs_pagepool_add(pagepool, page);
1531                 cond_resched();
1532                 goto repeat;
1533         }
1534 out_tocache:
1535         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1536                 /* turn into temporary page if fails (1 ref) */
1537                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1538                 goto out;
1539         }
1540         attach_page_private(page, pcl);
1541         /* drop a refcount added by allocpage (then we have 2 refs here) */
1542         put_page(page);
1543
1544 out:    /* the only exit (for tracing and debugging) */
1545         return page;
1546 }
1547
1548 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
1549                               struct z_erofs_decompressqueue *fgq, bool *fg)
1550 {
1551         struct z_erofs_decompressqueue *q;
1552
1553         if (fg && !*fg) {
1554                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1555                 if (!q) {
1556                         *fg = true;
1557                         goto fg_out;
1558                 }
1559 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1560                 kthread_init_work(&q->u.kthread_work,
1561                                   z_erofs_decompressqueue_kthread_work);
1562 #else
1563                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1564 #endif
1565         } else {
1566 fg_out:
1567                 q = fgq;
1568                 init_completion(&fgq->u.done);
1569                 atomic_set(&fgq->pending_bios, 0);
1570                 q->eio = false;
1571                 q->sync = true;
1572         }
1573         q->sb = sb;
1574         q->head = Z_EROFS_PCLUSTER_TAIL;
1575         return q;
1576 }
1577
1578 /* define decompression jobqueue types */
1579 enum {
1580         JQ_BYPASS,
1581         JQ_SUBMIT,
1582         NR_JOBQUEUES,
1583 };
1584
1585 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1586                                     z_erofs_next_pcluster_t qtail[],
1587                                     z_erofs_next_pcluster_t owned_head)
1588 {
1589         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1590         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1591
1592         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL);
1593
1594         WRITE_ONCE(*submit_qtail, owned_head);
1595         WRITE_ONCE(*bypass_qtail, &pcl->next);
1596
1597         qtail[JQ_BYPASS] = &pcl->next;
1598 }
1599
1600 static void z_erofs_decompressqueue_endio(struct bio *bio)
1601 {
1602         struct z_erofs_decompressqueue *q = bio->bi_private;
1603         blk_status_t err = bio->bi_status;
1604         struct bio_vec *bvec;
1605         struct bvec_iter_all iter_all;
1606
1607         bio_for_each_segment_all(bvec, bio, iter_all) {
1608                 struct page *page = bvec->bv_page;
1609
1610                 DBG_BUGON(PageUptodate(page));
1611                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1612
1613                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1614                         if (!err)
1615                                 SetPageUptodate(page);
1616                         unlock_page(page);
1617                 }
1618         }
1619         if (err)
1620                 q->eio = true;
1621         z_erofs_decompress_kickoff(q, -1);
1622         bio_put(bio);
1623 }
1624
1625 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1626                                  struct z_erofs_decompressqueue *fgq,
1627                                  bool *force_fg, bool readahead)
1628 {
1629         struct super_block *sb = f->inode->i_sb;
1630         struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1631         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1632         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1633         z_erofs_next_pcluster_t owned_head = f->owned_head;
1634         /* bio is NULL initially, so no need to initialize last_{index,bdev} */
1635         pgoff_t last_index;
1636         struct block_device *last_bdev;
1637         unsigned int nr_bios = 0;
1638         struct bio *bio = NULL;
1639         unsigned long pflags;
1640         int memstall = 0;
1641
1642         /*
1643          * if managed cache is enabled, bypass jobqueue is needed,
1644          * no need to read from device for all pclusters in this queue.
1645          */
1646         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1647         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);
1648
1649         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1650         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1651
1652         /* by default, all need io submission */
1653         q[JQ_SUBMIT]->head = owned_head;
1654
1655         do {
1656                 struct erofs_map_dev mdev;
1657                 struct z_erofs_pcluster *pcl;
1658                 pgoff_t cur, end;
1659                 unsigned int i = 0;
1660                 bool bypass = true;
1661
1662                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1663                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1664                 owned_head = READ_ONCE(pcl->next);
1665
1666                 if (z_erofs_is_inline_pcluster(pcl)) {
1667                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1668                         continue;
1669                 }
1670
1671                 /* no device id here, thus it will always succeed */
1672                 mdev = (struct erofs_map_dev) {
1673                         .m_pa = erofs_pos(sb, pcl->obj.index),
1674                 };
1675                 (void)erofs_map_dev(sb, &mdev);
1676
1677                 cur = erofs_blknr(sb, mdev.m_pa);
1678                 end = cur + pcl->pclusterpages;
1679
1680                 do {
1681                         struct page *page;
1682
1683                         page = pickup_page_for_submission(pcl, i++,
1684                                         &f->pagepool, mc);
1685                         if (!page)
1686                                 continue;
1687
1688                         if (bio && (cur != last_index + 1 ||
1689                                     last_bdev != mdev.m_bdev)) {
1690 submit_bio_retry:
1691                                 submit_bio(bio);
1692                                 if (memstall) {
1693                                         psi_memstall_leave(&pflags);
1694                                         memstall = 0;
1695                                 }
1696                                 bio = NULL;
1697                         }
1698
1699                         if (unlikely(PageWorkingset(page)) && !memstall) {
1700                                 psi_memstall_enter(&pflags);
1701                                 memstall = 1;
1702                         }
1703
1704                         if (!bio) {
1705                                 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1706                                                 REQ_OP_READ, GFP_NOIO);
1707                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1708
1709                                 last_bdev = mdev.m_bdev;
1710                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1711                                         (sb->s_blocksize_bits - 9);
1712                                 bio->bi_private = q[JQ_SUBMIT];
1713                                 if (readahead)
1714                                         bio->bi_opf |= REQ_RAHEAD;
1715                                 ++nr_bios;
1716                         }
1717
1718                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1719                                 goto submit_bio_retry;
1720
1721                         last_index = cur;
1722                         bypass = false;
1723                 } while (++cur < end);
1724
1725                 if (!bypass)
1726                         qtail[JQ_SUBMIT] = &pcl->next;
1727                 else
1728                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1729         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1730
1731         if (bio) {
1732                 submit_bio(bio);
1733                 if (memstall)
1734                         psi_memstall_leave(&pflags);
1735         }
1736
1737         /*
1738          * although background is preferred, no one is pending for submission.
1739          * don't issue decompression but drop it directly instead.
1740          */
1741         if (!*force_fg && !nr_bios) {
1742                 kvfree(q[JQ_SUBMIT]);
1743                 return;
1744         }
1745         z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);
1746 }
1747
1748 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1749                              bool force_fg, bool ra)
1750 {
1751         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1752
1753         if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1754                 return;
1755         z_erofs_submit_queue(f, io, &force_fg, ra);
1756
1757         /* handle bypass queue (no i/o pclusters) immediately */
1758         z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool);
1759
1760         if (!force_fg)
1761                 return;
1762
1763         /* wait until all bios are completed */
1764         wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1765
1766         /* handle synchronous decompress queue in the caller context */
1767         z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool);
1768 }
1769
1770 /*
1771  * Since partial uptodate is still unimplemented for now, we have to use
1772  * approximate readmore strategies as a start.
1773  */
1774 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1775                 struct readahead_control *rac, bool backmost)
1776 {
1777         struct inode *inode = f->inode;
1778         struct erofs_map_blocks *map = &f->map;
1779         erofs_off_t cur, end, headoffset = f->headoffset;
1780         int err;
1781
1782         if (backmost) {
1783                 if (rac)
1784                         end = headoffset + readahead_length(rac) - 1;
1785                 else
1786                         end = headoffset + PAGE_SIZE - 1;
1787                 map->m_la = end;
1788                 err = z_erofs_map_blocks_iter(inode, map,
1789                                               EROFS_GET_BLOCKS_READMORE);
1790                 if (err)
1791                         return;
1792
1793                 /* expand ra for the trailing edge if readahead */
1794                 if (rac) {
1795                         cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1796                         readahead_expand(rac, headoffset, cur - headoffset);
1797                         return;
1798                 }
1799                 end = round_up(end, PAGE_SIZE);
1800         } else {
1801                 end = round_up(map->m_la, PAGE_SIZE);
1802
1803                 if (!map->m_llen)
1804                         return;
1805         }
1806
1807         cur = map->m_la + map->m_llen - 1;
1808         while ((cur >= end) && (cur < i_size_read(inode))) {
1809                 pgoff_t index = cur >> PAGE_SHIFT;
1810                 struct page *page;
1811
1812                 page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1813                 if (page) {
1814                         if (PageUptodate(page))
1815                                 unlock_page(page);
1816                         else
1817                                 (void)z_erofs_do_read_page(f, page);
1818                         put_page(page);
1819                 }
1820
1821                 if (cur < PAGE_SIZE)
1822                         break;
1823                 cur = (index << PAGE_SHIFT) - 1;
1824         }
1825 }
1826
1827 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1828 {
1829         struct inode *const inode = folio->mapping->host;
1830         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1831         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1832         int err;
1833
1834         trace_erofs_read_folio(folio, false);
1835         f.headoffset = (erofs_off_t)folio->index << PAGE_SHIFT;
1836
1837         z_erofs_pcluster_readmore(&f, NULL, true);
1838         err = z_erofs_do_read_page(&f, &folio->page);
1839         z_erofs_pcluster_readmore(&f, NULL, false);
1840         z_erofs_pcluster_end(&f);
1841
1842         /* if some compressed cluster ready, need submit them anyway */
1843         z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, 0), false);
1844
1845         if (err && err != -EINTR)
1846                 erofs_err(inode->i_sb, "read error %d @ %lu of nid %llu",
1847                           err, folio->index, EROFS_I(inode)->nid);
1848
1849         erofs_put_metabuf(&f.map.buf);
1850         erofs_release_pages(&f.pagepool);
1851         return err;
1852 }
1853
1854 static void z_erofs_readahead(struct readahead_control *rac)
1855 {
1856         struct inode *const inode = rac->mapping->host;
1857         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1858         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1859         struct folio *head = NULL, *folio;
1860         unsigned int nr_folios;
1861         int err;
1862
1863         f.headoffset = readahead_pos(rac);
1864
1865         z_erofs_pcluster_readmore(&f, rac, true);
1866         nr_folios = readahead_count(rac);
1867         trace_erofs_readpages(inode, readahead_index(rac), nr_folios, false);
1868
1869         while ((folio = readahead_folio(rac))) {
1870                 folio->private = head;
1871                 head = folio;
1872         }
1873
1874         /* traverse in reverse order for best metadata I/O performance */
1875         while (head) {
1876                 folio = head;
1877                 head = folio_get_private(folio);
1878
1879                 err = z_erofs_do_read_page(&f, &folio->page);
1880                 if (err && err != -EINTR)
1881                         erofs_err(inode->i_sb, "readahead error at folio %lu @ nid %llu",
1882                                   folio->index, EROFS_I(inode)->nid);
1883         }
1884         z_erofs_pcluster_readmore(&f, rac, false);
1885         z_erofs_pcluster_end(&f);
1886
1887         z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, nr_folios), true);
1888         erofs_put_metabuf(&f.map.buf);
1889         erofs_release_pages(&f.pagepool);
1890 }
1891
1892 const struct address_space_operations z_erofs_aops = {
1893         .read_folio = z_erofs_read_folio,
1894         .readahead = z_erofs_readahead,
1895 };