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