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/prefetch.h>
   9 #include <linux/psi.h>
  10 #include <linux/cpuhotplug.h>
  11 #include <trace/events/erofs.h>
  12
  13 #define Z_EROFS_PCLUSTER_MAX_PAGES      (Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE)
  14 #define Z_EROFS_INLINE_BVECS            2
  15
  16 /*
  17  * let's leave a type here in case of introducing
  18  * another tagged pointer later.
  19  */
  20 typedef void *z_erofs_next_pcluster_t;
  21
  22 struct z_erofs_bvec {
  23         struct page *page;
  24         int offset;
  25         unsigned int end;
  26 };
  27
  28 #define __Z_EROFS_BVSET(name, total) \
  29 struct name { \
  30         /* point to the next page which contains the following bvecs */ \
  31         struct page *nextpage; \
  32         struct z_erofs_bvec bvec[total]; \
  33 }
  34 __Z_EROFS_BVSET(z_erofs_bvset,);
  35 __Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS);
  36
  37 /*
  38  * Structure fields follow one of the following exclusion rules.
  39  *
  40  * I: Modifiable by initialization/destruction paths and read-only
  41  *    for everyone else;
  42  *
  43  * L: Field should be protected by the pcluster lock;
  44  *
  45  * A: Field should be accessed / updated in atomic for parallelized code.
  46  */
  47 struct z_erofs_pcluster {
  48         struct erofs_workgroup obj;
  49         struct mutex lock;
  50
  51         /* A: point to next chained pcluster or TAILs */
  52         z_erofs_next_pcluster_t next;
  53
  54         /* L: the maximum decompression size of this round */
  55         unsigned int length;
  56
  57         /* L: total number of bvecs */
  58         unsigned int vcnt;
  59
  60         /* I: page offset of start position of decompression */
  61         unsigned short pageofs_out;
  62
  63         /* I: page offset of inline compressed data */
  64         unsigned short pageofs_in;
  65
  66         union {
  67                 /* L: inline a certain number of bvec for bootstrap */
  68                 struct z_erofs_bvset_inline bvset;
  69
  70                 /* I: can be used to free the pcluster by RCU. */
  71                 struct rcu_head rcu;
  72         };
  73
  74         union {
  75                 /* I: physical cluster size in pages */
  76                 unsigned short pclusterpages;
  77
  78                 /* I: tailpacking inline compressed size */
  79                 unsigned short tailpacking_size;
  80         };
  81
  82         /* I: compression algorithm format */
  83         unsigned char algorithmformat;
  84
  85         /* L: whether partial decompression or not */
  86         bool partial;
  87
  88         /* L: indicate several pageofs_outs or not */
  89         bool multibases;
  90
  91         /* A: compressed bvecs (can be cached or inplaced pages) */
  92         struct z_erofs_bvec compressed_bvecs[];
  93 };
  94
  95 /* let's avoid the valid 32-bit kernel addresses */
  96
  97 /* the chained workgroup has't submitted io (still open) */
  98 #define Z_EROFS_PCLUSTER_TAIL           ((void *)0x5F0ECAFE)
  99 /* the chained workgroup has already submitted io */
 100 #define Z_EROFS_PCLUSTER_TAIL_CLOSED    ((void *)0x5F0EDEAD)
 101
 102 #define Z_EROFS_PCLUSTER_NIL            (NULL)
 103
 104 struct z_erofs_decompressqueue {
 105         struct super_block *sb;
 106         atomic_t pending_bios;
 107         z_erofs_next_pcluster_t head;
 108
 109         union {
 110                 struct completion done;
 111                 struct work_struct work;
 112                 struct kthread_work kthread_work;
 113         } u;
 114         bool eio, sync;
 115 };
 116
 117 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl)
 118 {
 119         return !pcl->obj.index;
 120 }
 121
 122 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl)
 123 {
 124         if (z_erofs_is_inline_pcluster(pcl))
 125                 return 1;
 126         return pcl->pclusterpages;
 127 }
 128
 129 /*
 130  * bit 30: I/O error occurred on this page
 131  * bit 0 - 29: remaining parts to complete this page
 132  */
 133 #define Z_EROFS_PAGE_EIO                        (1 << 30)
 134
 135 static inline void z_erofs_onlinepage_init(struct page *page)
 136 {
 137         union {
 138                 atomic_t o;
 139                 unsigned long v;
 140         } u = { .o = ATOMIC_INIT(1) };
 141
 142         set_page_private(page, u.v);
 143         smp_wmb();
 144         SetPagePrivate(page);
 145 }
 146
 147 static inline void z_erofs_onlinepage_split(struct page *page)
 148 {
 149         atomic_inc((atomic_t *)&page->private);
 150 }
 151
 152 static inline void z_erofs_page_mark_eio(struct page *page)
 153 {
 154         int orig;
 155
 156         do {
 157                 orig = atomic_read((atomic_t *)&page->private);
 158         } while (atomic_cmpxchg((atomic_t *)&page->private, orig,
 159                                 orig | Z_EROFS_PAGE_EIO) != orig);
 160 }
 161
 162 static inline void z_erofs_onlinepage_endio(struct page *page)
 163 {
 164         unsigned int v;
 165
 166         DBG_BUGON(!PagePrivate(page));
 167         v = atomic_dec_return((atomic_t *)&page->private);
 168         if (!(v & ~Z_EROFS_PAGE_EIO)) {
 169                 set_page_private(page, 0);
 170                 ClearPagePrivate(page);
 171                 if (!(v & Z_EROFS_PAGE_EIO))
 172                         SetPageUptodate(page);
 173                 unlock_page(page);
 174         }
 175 }
 176
 177 #define Z_EROFS_ONSTACK_PAGES           32
 178
 179 /*
 180  * since pclustersize is variable for big pcluster feature, introduce slab
 181  * pools implementation for different pcluster sizes.
 182  */
 183 struct z_erofs_pcluster_slab {
 184         struct kmem_cache *slab;
 185         unsigned int maxpages;
 186         char name[48];
 187 };
 188
 189 #define _PCLP(n) { .maxpages = n }
 190
 191 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
 192         _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
 193         _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
 194 };
 195
 196 struct z_erofs_bvec_iter {
 197         struct page *bvpage;
 198         struct z_erofs_bvset *bvset;
 199         unsigned int nr, cur;
 200 };
 201
 202 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
 203 {
 204         if (iter->bvpage)
 205                 kunmap_local(iter->bvset);
 206         return iter->bvpage;
 207 }
 208
 209 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
 210 {
 211         unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
 212         /* have to access nextpage in advance, otherwise it will be unmapped */
 213         struct page *nextpage = iter->bvset->nextpage;
 214         struct page *oldpage;
 215
 216         DBG_BUGON(!nextpage);
 217         oldpage = z_erofs_bvec_iter_end(iter);
 218         iter->bvpage = nextpage;
 219         iter->bvset = kmap_local_page(nextpage);
 220         iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
 221         iter->cur = 0;
 222         return oldpage;
 223 }
 224
 225 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
 226                                     struct z_erofs_bvset_inline *bvset,
 227                                     unsigned int bootstrap_nr,
 228                                     unsigned int cur)
 229 {
 230         *iter = (struct z_erofs_bvec_iter) {
 231                 .nr = bootstrap_nr,
 232                 .bvset = (struct z_erofs_bvset *)bvset,
 233         };
 234
 235         while (cur > iter->nr) {
 236                 cur -= iter->nr;
 237                 z_erofs_bvset_flip(iter);
 238         }
 239         iter->cur = cur;
 240 }
 241
 242 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
 243                                 struct z_erofs_bvec *bvec,
 244                                 struct page **candidate_bvpage)
 245 {
 246         if (iter->cur == iter->nr) {
 247                 if (!*candidate_bvpage)
 248                         return -EAGAIN;
 249
 250                 DBG_BUGON(iter->bvset->nextpage);
 251                 iter->bvset->nextpage = *candidate_bvpage;
 252                 z_erofs_bvset_flip(iter);
 253
 254                 iter->bvset->nextpage = NULL;
 255                 *candidate_bvpage = NULL;
 256         }
 257         iter->bvset->bvec[iter->cur++] = *bvec;
 258         return 0;
 259 }
 260
 261 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,
 262                                  struct z_erofs_bvec *bvec,
 263                                  struct page **old_bvpage)
 264 {
 265         if (iter->cur == iter->nr)
 266                 *old_bvpage = z_erofs_bvset_flip(iter);
 267         else
 268                 *old_bvpage = NULL;
 269         *bvec = iter->bvset->bvec[iter->cur++];
 270 }
 271
 272 static void z_erofs_destroy_pcluster_pool(void)
 273 {
 274         int i;
 275
 276         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 277                 if (!pcluster_pool[i].slab)
 278                         continue;
 279                 kmem_cache_destroy(pcluster_pool[i].slab);
 280                 pcluster_pool[i].slab = NULL;
 281         }
 282 }
 283
 284 static int z_erofs_create_pcluster_pool(void)
 285 {
 286         struct z_erofs_pcluster_slab *pcs;
 287         struct z_erofs_pcluster *a;
 288         unsigned int size;
 289
 290         for (pcs = pcluster_pool;
 291              pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
 292                 size = struct_size(a, compressed_bvecs, pcs->maxpages);
 293
 294                 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
 295                 pcs->slab = kmem_cache_create(pcs->name, size, 0,
 296                                               SLAB_RECLAIM_ACCOUNT, NULL);
 297                 if (pcs->slab)
 298                         continue;
 299
 300                 z_erofs_destroy_pcluster_pool();
 301                 return -ENOMEM;
 302         }
 303         return 0;
 304 }
 305
 306 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
 307 {
 308         int i;
 309
 310         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 311                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 312                 struct z_erofs_pcluster *pcl;
 313
 314                 if (nrpages > pcs->maxpages)
 315                         continue;
 316
 317                 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
 318                 if (!pcl)
 319                         return ERR_PTR(-ENOMEM);
 320                 pcl->pclusterpages = nrpages;
 321                 return pcl;
 322         }
 323         return ERR_PTR(-EINVAL);
 324 }
 325
 326 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
 327 {
 328         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 329         int i;
 330
 331         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 332                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 333
 334                 if (pclusterpages > pcs->maxpages)
 335                         continue;
 336
 337                 kmem_cache_free(pcs->slab, pcl);
 338                 return;
 339         }
 340         DBG_BUGON(1);
 341 }
 342
 343 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 344
 345 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
 346 static struct kthread_worker __rcu **z_erofs_pcpu_workers;
 347
 348 static void erofs_destroy_percpu_workers(void)
 349 {
 350         struct kthread_worker *worker;
 351         unsigned int cpu;
 352
 353         for_each_possible_cpu(cpu) {
 354                 worker = rcu_dereference_protected(
 355                                         z_erofs_pcpu_workers[cpu], 1);
 356                 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
 357                 if (worker)
 358                         kthread_destroy_worker(worker);
 359         }
 360         kfree(z_erofs_pcpu_workers);
 361 }
 362
 363 static struct kthread_worker *erofs_init_percpu_worker(int cpu)
 364 {
 365         struct kthread_worker *worker =
 366                 kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu);
 367
 368         if (IS_ERR(worker))
 369                 return worker;
 370         if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI))
 371                 sched_set_fifo_low(worker->task);
 372         else
 373                 sched_set_normal(worker->task, 0);
 374         return worker;
 375 }
 376
 377 static int erofs_init_percpu_workers(void)
 378 {
 379         struct kthread_worker *worker;
 380         unsigned int cpu;
 381
 382         z_erofs_pcpu_workers = kcalloc(num_possible_cpus(),
 383                         sizeof(struct kthread_worker *), GFP_ATOMIC);
 384         if (!z_erofs_pcpu_workers)
 385                 return -ENOMEM;
 386
 387         for_each_online_cpu(cpu) {      /* could miss cpu{off,on}line? */
 388                 worker = erofs_init_percpu_worker(cpu);
 389                 if (!IS_ERR(worker))
 390                         rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
 391         }
 392         return 0;
 393 }
 394 #else
 395 static inline void erofs_destroy_percpu_workers(void) {}
 396 static inline int erofs_init_percpu_workers(void) { return 0; }
 397 #endif
 398
 399 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD)
 400 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock);
 401 static enum cpuhp_state erofs_cpuhp_state;
 402
 403 static int erofs_cpu_online(unsigned int cpu)
 404 {
 405         struct kthread_worker *worker, *old;
 406
 407         worker = erofs_init_percpu_worker(cpu);
 408         if (IS_ERR(worker))
 409                 return PTR_ERR(worker);
 410
 411         spin_lock(&z_erofs_pcpu_worker_lock);
 412         old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
 413                         lockdep_is_held(&z_erofs_pcpu_worker_lock));
 414         if (!old)
 415                 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker);
 416         spin_unlock(&z_erofs_pcpu_worker_lock);
 417         if (old)
 418                 kthread_destroy_worker(worker);
 419         return 0;
 420 }
 421
 422 static int erofs_cpu_offline(unsigned int cpu)
 423 {
 424         struct kthread_worker *worker;
 425
 426         spin_lock(&z_erofs_pcpu_worker_lock);
 427         worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu],
 428                         lockdep_is_held(&z_erofs_pcpu_worker_lock));
 429         rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL);
 430         spin_unlock(&z_erofs_pcpu_worker_lock);
 431
 432         synchronize_rcu();
 433         if (worker)
 434                 kthread_destroy_worker(worker);
 435         return 0;
 436 }
 437
 438 static int erofs_cpu_hotplug_init(void)
 439 {
 440         int state;
 441
 442         state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
 443                         "fs/erofs:online", erofs_cpu_online, erofs_cpu_offline);
 444         if (state < 0)
 445                 return state;
 446
 447         erofs_cpuhp_state = state;
 448         return 0;
 449 }
 450
 451 static void erofs_cpu_hotplug_destroy(void)
 452 {
 453         if (erofs_cpuhp_state)
 454                 cpuhp_remove_state_nocalls(erofs_cpuhp_state);
 455 }
 456 #else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */
 457 static inline int erofs_cpu_hotplug_init(void) { return 0; }
 458 static inline void erofs_cpu_hotplug_destroy(void) {}
 459 #endif
 460
 461 void z_erofs_exit_zip_subsystem(void)
 462 {
 463         erofs_cpu_hotplug_destroy();
 464         erofs_destroy_percpu_workers();
 465         destroy_workqueue(z_erofs_workqueue);
 466         z_erofs_destroy_pcluster_pool();
 467 }
 468
 469 int __init z_erofs_init_zip_subsystem(void)
 470 {
 471         int err = z_erofs_create_pcluster_pool();
 472
 473         if (err)
 474                 goto out_error_pcluster_pool;
 475
 476         z_erofs_workqueue = alloc_workqueue("erofs_worker",
 477                         WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus());
 478         if (!z_erofs_workqueue) {
 479                 err = -ENOMEM;
 480                 goto out_error_workqueue_init;
 481         }
 482
 483         err = erofs_init_percpu_workers();
 484         if (err)
 485                 goto out_error_pcpu_worker;
 486
 487         err = erofs_cpu_hotplug_init();
 488         if (err < 0)
 489                 goto out_error_cpuhp_init;
 490         return err;
 491
 492 out_error_cpuhp_init:
 493         erofs_destroy_percpu_workers();
 494 out_error_pcpu_worker:
 495         destroy_workqueue(z_erofs_workqueue);
 496 out_error_workqueue_init:
 497         z_erofs_destroy_pcluster_pool();
 498 out_error_pcluster_pool:
 499         return err;
 500 }
 501
 502 enum z_erofs_pclustermode {
 503         Z_EROFS_PCLUSTER_INFLIGHT,
 504         /*
 505          * The current pclusters was the tail of an exist chain, in addition
 506          * that the previous processed chained pclusters are all decided to
 507          * be hooked up to it.
 508          * A new chain will be created for the remaining pclusters which are
 509          * not processed yet, so different from Z_EROFS_PCLUSTER_FOLLOWED,
 510          * the next pcluster cannot reuse the whole page safely for inplace I/O
 511          * in the following scenario:
 512          *  ________________________________________________________________
 513          * |      tail (partial) page     |       head (partial) page       |
 514          * |   (belongs to the next pcl)  |   (belongs to the current pcl)  |
 515          * |_______PCLUSTER_FOLLOWED______|________PCLUSTER_HOOKED__________|
 516          */
 517         Z_EROFS_PCLUSTER_HOOKED,
 518         /*
 519          * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
 520          * could be dispatched into bypass queue later due to uptodated managed
 521          * pages. All related online pages cannot be reused for inplace I/O (or
 522          * bvpage) since it can be directly decoded without I/O submission.
 523          */
 524         Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
 525         /*
 526          * The current collection has been linked with the owned chain, and
 527          * could also be linked with the remaining collections, which means
 528          * if the processing page is the tail page of the collection, thus
 529          * the current collection can safely use the whole page (since
 530          * the previous collection is under control) for in-place I/O, as
 531          * illustrated below:
 532          *  ________________________________________________________________
 533          * |  tail (partial) page |          head (partial) page           |
 534          * |  (of the current cl) |      (of the previous collection)      |
 535          * | PCLUSTER_FOLLOWED or |                                        |
 536          * |_____PCLUSTER_HOOKED__|___________PCLUSTER_FOLLOWED____________|
 537          *
 538          * [  (*) the above page can be used as inplace I/O.               ]
 539          */
 540         Z_EROFS_PCLUSTER_FOLLOWED,
 541 };
 542
 543 struct z_erofs_decompress_frontend {
 544         struct inode *const inode;
 545         struct erofs_map_blocks map;
 546         struct z_erofs_bvec_iter biter;
 547
 548         struct page *candidate_bvpage;
 549         struct z_erofs_pcluster *pcl, *tailpcl;
 550         z_erofs_next_pcluster_t owned_head;
 551         enum z_erofs_pclustermode mode;
 552
 553         bool readahead;
 554         /* used for applying cache strategy on the fly */
 555         bool backmost;
 556         erofs_off_t headoffset;
 557
 558         /* a pointer used to pick up inplace I/O pages */
 559         unsigned int icur;
 560 };
 561
 562 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 563         .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 564         .mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true }
 565
 566 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe)
 567 {
 568         unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy;
 569
 570         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 571                 return false;
 572
 573         if (fe->backmost)
 574                 return true;
 575
 576         if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 577             fe->map.m_la < fe->headoffset)
 578                 return true;
 579
 580         return false;
 581 }
 582
 583 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe,
 584                                struct page **pagepool)
 585 {
 586         struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
 587         struct z_erofs_pcluster *pcl = fe->pcl;
 588         bool shouldalloc = z_erofs_should_alloc_cache(fe);
 589         bool standalone = true;
 590         /*
 591          * optimistic allocation without direct reclaim since inplace I/O
 592          * can be used if low memory otherwise.
 593          */
 594         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 595                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 596         unsigned int i;
 597
 598         if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
 599                 return;
 600
 601         for (i = 0; i < pcl->pclusterpages; ++i) {
 602                 struct page *page;
 603                 void *t;        /* mark pages just found for debugging */
 604                 struct page *newpage = NULL;
 605
 606                 /* the compressed page was loaded before */
 607                 if (READ_ONCE(pcl->compressed_bvecs[i].page))
 608                         continue;
 609
 610                 page = find_get_page(mc, pcl->obj.index + i);
 611
 612                 if (page) {
 613                         t = (void *)((unsigned long)page | 1);
 614                 } else {
 615                         /* I/O is needed, no possible to decompress directly */
 616                         standalone = false;
 617                         if (!shouldalloc)
 618                                 continue;
 619
 620                         /*
 621                          * try to use cached I/O if page allocation
 622                          * succeeds or fallback to in-place I/O instead
 623                          * to avoid any direct reclaim.
 624                          */
 625                         newpage = erofs_allocpage(pagepool, gfp);
 626                         if (!newpage)
 627                                 continue;
 628                         set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE);
 629                         t = (void *)((unsigned long)newpage | 1);
 630                 }
 631
 632                 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t))
 633                         continue;
 634
 635                 if (page)
 636                         put_page(page);
 637                 else if (newpage)
 638                         erofs_pagepool_add(pagepool, newpage);
 639         }
 640
 641         /*
 642          * don't do inplace I/O if all compressed pages are available in
 643          * managed cache since it can be moved to the bypass queue instead.
 644          */
 645         if (standalone)
 646                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
 647 }
 648
 649 /* called by erofs_shrinker to get rid of all compressed_pages */
 650 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 651                                        struct erofs_workgroup *grp)
 652 {
 653         struct z_erofs_pcluster *const pcl =
 654                 container_of(grp, struct z_erofs_pcluster, obj);
 655         int i;
 656
 657         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 658         /*
 659          * refcount of workgroup is now freezed as 1,
 660          * therefore no need to worry about available decompression users.
 661          */
 662         for (i = 0; i < pcl->pclusterpages; ++i) {
 663                 struct page *page = pcl->compressed_bvecs[i].page;
 664
 665                 if (!page)
 666                         continue;
 667
 668                 /* block other users from reclaiming or migrating the page */
 669                 if (!trylock_page(page))
 670                         return -EBUSY;
 671
 672                 if (!erofs_page_is_managed(sbi, page))
 673                         continue;
 674
 675                 /* barrier is implied in the following 'unlock_page' */
 676                 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 677                 detach_page_private(page);
 678                 unlock_page(page);
 679         }
 680         return 0;
 681 }
 682
 683 int erofs_try_to_free_cached_page(struct page *page)
 684 {
 685         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 686         int ret, i;
 687
 688         if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1))
 689                 return 0;
 690
 691         ret = 0;
 692         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 693         for (i = 0; i < pcl->pclusterpages; ++i) {
 694                 if (pcl->compressed_bvecs[i].page == page) {
 695                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 696                         ret = 1;
 697                         break;
 698                 }
 699         }
 700         erofs_workgroup_unfreeze(&pcl->obj, 1);
 701         if (ret)
 702                 detach_page_private(page);
 703         return ret;
 704 }
 705
 706 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
 707                                    struct z_erofs_bvec *bvec)
 708 {
 709         struct z_erofs_pcluster *const pcl = fe->pcl;
 710
 711         while (fe->icur > 0) {
 712                 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
 713                              NULL, bvec->page)) {
 714                         pcl->compressed_bvecs[fe->icur] = *bvec;
 715                         return true;
 716                 }
 717         }
 718         return false;
 719 }
 720
 721 /* callers must be with pcluster lock held */
 722 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
 723                                struct z_erofs_bvec *bvec, bool exclusive)
 724 {
 725         int ret;
 726
 727         if (exclusive) {
 728                 /* give priority for inplaceio to use file pages first */
 729                 if (z_erofs_try_inplace_io(fe, bvec))
 730                         return 0;
 731                 /* otherwise, check if it can be used as a bvpage */
 732                 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
 733                     !fe->candidate_bvpage)
 734                         fe->candidate_bvpage = bvec->page;
 735         }
 736         ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage);
 737         fe->pcl->vcnt += (ret >= 0);
 738         return ret;
 739 }
 740
 741 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
 742 {
 743         struct z_erofs_pcluster *pcl = f->pcl;
 744         z_erofs_next_pcluster_t *owned_head = &f->owned_head;
 745
 746         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 747         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 748                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 749                 *owned_head = &pcl->next;
 750                 /* so we can attach this pcluster to our submission chain. */
 751                 f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 752                 return;
 753         }
 754
 755         /*
 756          * type 2, link to the end of an existing open chain, be careful
 757          * that its submission is controlled by the original attached chain.
 758          */
 759         if (*owned_head != &pcl->next && pcl != f->tailpcl &&
 760             cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 761                     *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
 762                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 763                 f->mode = Z_EROFS_PCLUSTER_HOOKED;
 764                 f->tailpcl = NULL;
 765                 return;
 766         }
 767         /* type 3, it belongs to a chain, but it isn't the end of the chain */
 768         f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
 769 }
 770
 771 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
 772 {
 773         struct erofs_map_blocks *map = &fe->map;
 774         bool ztailpacking = map->m_flags & EROFS_MAP_META;
 775         struct z_erofs_pcluster *pcl;
 776         struct erofs_workgroup *grp;
 777         int err;
 778
 779         if (!(map->m_flags & EROFS_MAP_ENCODED) ||
 780             (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) {
 781                 DBG_BUGON(1);
 782                 return -EFSCORRUPTED;
 783         }
 784
 785         /* no available pcluster, let's allocate one */
 786         pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
 787                                      map->m_plen >> PAGE_SHIFT);
 788         if (IS_ERR(pcl))
 789                 return PTR_ERR(pcl);
 790
 791         atomic_set(&pcl->obj.refcount, 1);
 792         pcl->algorithmformat = map->m_algorithmformat;
 793         pcl->length = 0;
 794         pcl->partial = true;
 795
 796         /* new pclusters should be claimed as type 1, primary and followed */
 797         pcl->next = fe->owned_head;
 798         pcl->pageofs_out = map->m_la & ~PAGE_MASK;
 799         fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 800
 801         /*
 802          * lock all primary followed works before visible to others
 803          * and mutex_trylock *never* fails for a new pcluster.
 804          */
 805         mutex_init(&pcl->lock);
 806         DBG_BUGON(!mutex_trylock(&pcl->lock));
 807
 808         if (ztailpacking) {
 809                 pcl->obj.index = 0;     /* which indicates ztailpacking */
 810                 pcl->pageofs_in = erofs_blkoff(map->m_pa);
 811                 pcl->tailpacking_size = map->m_plen;
 812         } else {
 813                 pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 814
 815                 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
 816                 if (IS_ERR(grp)) {
 817                         err = PTR_ERR(grp);
 818                         goto err_out;
 819                 }
 820
 821                 if (grp != &pcl->obj) {
 822                         fe->pcl = container_of(grp,
 823                                         struct z_erofs_pcluster, obj);
 824                         err = -EEXIST;
 825                         goto err_out;
 826                 }
 827         }
 828         /* used to check tail merging loop due to corrupted images */
 829         if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
 830                 fe->tailpcl = pcl;
 831         fe->owned_head = &pcl->next;
 832         fe->pcl = pcl;
 833         return 0;
 834
 835 err_out:
 836         mutex_unlock(&pcl->lock);
 837         z_erofs_free_pcluster(pcl);
 838         return err;
 839 }
 840
 841 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
 842 {
 843         struct erofs_map_blocks *map = &fe->map;
 844         struct erofs_workgroup *grp = NULL;
 845         int ret;
 846
 847         DBG_BUGON(fe->pcl);
 848
 849         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
 850         DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
 851         DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 852
 853         if (!(map->m_flags & EROFS_MAP_META)) {
 854                 grp = erofs_find_workgroup(fe->inode->i_sb,
 855                                            map->m_pa >> PAGE_SHIFT);
 856         } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
 857                 DBG_BUGON(1);
 858                 return -EFSCORRUPTED;
 859         }
 860
 861         if (grp) {
 862                 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 863                 ret = -EEXIST;
 864         } else {
 865                 ret = z_erofs_register_pcluster(fe);
 866         }
 867
 868         if (ret == -EEXIST) {
 869                 mutex_lock(&fe->pcl->lock);
 870                 /* used to check tail merging loop due to corrupted images */
 871                 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
 872                         fe->tailpcl = fe->pcl;
 873
 874                 z_erofs_try_to_claim_pcluster(fe);
 875         } else if (ret) {
 876                 return ret;
 877         }
 878         z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
 879                                 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
 880         /* since file-backed online pages are traversed in reverse order */
 881         fe->icur = z_erofs_pclusterpages(fe->pcl);
 882         return 0;
 883 }
 884
 885 /*
 886  * keep in mind that no referenced pclusters will be freed
 887  * only after a RCU grace period.
 888  */
 889 static void z_erofs_rcu_callback(struct rcu_head *head)
 890 {
 891         z_erofs_free_pcluster(container_of(head,
 892                         struct z_erofs_pcluster, rcu));
 893 }
 894
 895 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 896 {
 897         struct z_erofs_pcluster *const pcl =
 898                 container_of(grp, struct z_erofs_pcluster, obj);
 899
 900         call_rcu(&pcl->rcu, z_erofs_rcu_callback);
 901 }
 902
 903 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
 904 {
 905         struct z_erofs_pcluster *pcl = fe->pcl;
 906
 907         if (!pcl)
 908                 return false;
 909
 910         z_erofs_bvec_iter_end(&fe->biter);
 911         mutex_unlock(&pcl->lock);
 912
 913         if (fe->candidate_bvpage) {
 914                 DBG_BUGON(z_erofs_is_shortlived_page(fe->candidate_bvpage));
 915                 fe->candidate_bvpage = NULL;
 916         }
 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         return true;
 927 }
 928
 929 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos,
 930                                  struct page *page, unsigned int pageofs,
 931                                  unsigned int len)
 932 {
 933         struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode;
 934         struct erofs_buf buf = __EROFS_BUF_INITIALIZER;
 935         u8 *src, *dst;
 936         unsigned int i, cnt;
 937
 938         if (!packed_inode)
 939                 return -EFSCORRUPTED;
 940
 941         pos += EROFS_I(inode)->z_fragmentoff;
 942         for (i = 0; i < len; i += cnt) {
 943                 cnt = min_t(unsigned int, len - i,
 944                             EROFS_BLKSIZ - erofs_blkoff(pos));
 945                 src = erofs_bread(&buf, packed_inode,
 946                                   erofs_blknr(pos), EROFS_KMAP);
 947                 if (IS_ERR(src)) {
 948                         erofs_put_metabuf(&buf);
 949                         return PTR_ERR(src);
 950                 }
 951
 952                 dst = kmap_local_page(page);
 953                 memcpy(dst + pageofs + i, src + erofs_blkoff(pos), cnt);
 954                 kunmap_local(dst);
 955                 pos += cnt;
 956         }
 957         erofs_put_metabuf(&buf);
 958         return 0;
 959 }
 960
 961 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 962                                 struct page *page, struct page **pagepool)
 963 {
 964         struct inode *const inode = fe->inode;
 965         struct erofs_map_blocks *const map = &fe->map;
 966         const loff_t offset = page_offset(page);
 967         bool tight = true, exclusive;
 968         unsigned int cur, end, spiltted;
 969         int err = 0;
 970
 971         /* register locked file pages as online pages in pack */
 972         z_erofs_onlinepage_init(page);
 973
 974         spiltted = 0;
 975         end = PAGE_SIZE;
 976 repeat:
 977         cur = end - 1;
 978
 979         if (offset + cur < map->m_la ||
 980             offset + cur >= map->m_la + map->m_llen) {
 981                 erofs_dbg("out-of-range map @ pos %llu", offset + cur);
 982
 983                 if (z_erofs_collector_end(fe))
 984                         fe->backmost = false;
 985                 map->m_la = offset + cur;
 986                 map->m_llen = 0;
 987                 err = z_erofs_map_blocks_iter(inode, map, 0);
 988                 if (err)
 989                         goto out;
 990         } else {
 991                 if (fe->pcl)
 992                         goto hitted;
 993                 /* didn't get a valid pcluster previously (very rare) */
 994         }
 995
 996         if (!(map->m_flags & EROFS_MAP_MAPPED) ||
 997             map->m_flags & EROFS_MAP_FRAGMENT)
 998                 goto hitted;
 999
1000         err = z_erofs_collector_begin(fe);
1001         if (err)
1002                 goto out;
1003
1004         if (z_erofs_is_inline_pcluster(fe->pcl)) {
1005                 void *mp;
1006
1007                 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
1008                                         erofs_blknr(map->m_pa), EROFS_NO_KMAP);
1009                 if (IS_ERR(mp)) {
1010                         err = PTR_ERR(mp);
1011                         erofs_err(inode->i_sb,
1012                                   "failed to get inline page, err %d", err);
1013                         goto out;
1014                 }
1015                 get_page(fe->map.buf.page);
1016                 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page,
1017                            fe->map.buf.page);
1018                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
1019         } else {
1020                 /* bind cache first when cached decompression is preferred */
1021                 z_erofs_bind_cache(fe, pagepool);
1022         }
1023 hitted:
1024         /*
1025          * Ensure the current partial page belongs to this submit chain rather
1026          * than other concurrent submit chains or the noio(bypass) chain since
1027          * those chains are handled asynchronously thus the page cannot be used
1028          * for inplace I/O or bvpage (should be processed in a strict order.)
1029          */
1030         tight &= (fe->mode >= Z_EROFS_PCLUSTER_HOOKED &&
1031                   fe->mode != Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
1032
1033         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
1034         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
1035                 zero_user_segment(page, cur, end);
1036                 goto next_part;
1037         }
1038         if (map->m_flags & EROFS_MAP_FRAGMENT) {
1039                 unsigned int pageofs, skip, len;
1040
1041                 if (offset > map->m_la) {
1042                         pageofs = 0;
1043                         skip = offset - map->m_la;
1044                 } else {
1045                         pageofs = map->m_la & ~PAGE_MASK;
1046                         skip = 0;
1047                 }
1048                 len = min_t(unsigned int, map->m_llen - skip, end - cur);
1049                 err = z_erofs_read_fragment(inode, skip, page, pageofs, len);
1050                 if (err)
1051                         goto out;
1052                 ++spiltted;
1053                 tight = false;
1054                 goto next_part;
1055         }
1056
1057         exclusive = (!cur && (!spiltted || tight));
1058         if (cur)
1059                 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
1060
1061 retry:
1062         err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
1063                                         .page = page,
1064                                         .offset = offset - map->m_la,
1065                                         .end = end,
1066                                   }), exclusive);
1067         /* should allocate an additional short-lived page for bvset */
1068         if (err == -EAGAIN && !fe->candidate_bvpage) {
1069                 fe->candidate_bvpage = alloc_page(GFP_NOFS | __GFP_NOFAIL);
1070                 set_page_private(fe->candidate_bvpage,
1071                                  Z_EROFS_SHORTLIVED_PAGE);
1072                 goto retry;
1073         }
1074
1075         if (err) {
1076                 DBG_BUGON(err == -EAGAIN && fe->candidate_bvpage);
1077                 goto out;
1078         }
1079
1080         z_erofs_onlinepage_split(page);
1081         /* bump up the number of spiltted parts of a page */
1082         ++spiltted;
1083         if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
1084                 fe->pcl->multibases = true;
1085         if (fe->pcl->length < offset + end - map->m_la) {
1086                 fe->pcl->length = offset + end - map->m_la;
1087                 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
1088         }
1089         if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
1090             !(map->m_flags & EROFS_MAP_PARTIAL_REF) &&
1091             fe->pcl->length == map->m_llen)
1092                 fe->pcl->partial = false;
1093 next_part:
1094         /* shorten the remaining extent to update progress */
1095         map->m_llen = offset + cur - map->m_la;
1096         map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
1097
1098         end = cur;
1099         if (end > 0)
1100                 goto repeat;
1101
1102 out:
1103         if (err)
1104                 z_erofs_page_mark_eio(page);
1105         z_erofs_onlinepage_endio(page);
1106
1107         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
1108                   __func__, page, spiltted, map->m_llen);
1109         return err;
1110 }
1111
1112 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
1113                                        unsigned int readahead_pages)
1114 {
1115         /* auto: enable for read_folio, disable for readahead */
1116         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
1117             !readahead_pages)
1118                 return true;
1119
1120         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
1121             (readahead_pages <= sbi->opt.max_sync_decompress_pages))
1122                 return true;
1123
1124         return false;
1125 }
1126
1127 static bool z_erofs_page_is_invalidated(struct page *page)
1128 {
1129         return !page->mapping && !z_erofs_is_shortlived_page(page);
1130 }
1131
1132 struct z_erofs_decompress_backend {
1133         struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
1134         struct super_block *sb;
1135         struct z_erofs_pcluster *pcl;
1136
1137         /* pages with the longest decompressed length for deduplication */
1138         struct page **decompressed_pages;
1139         /* pages to keep the compressed data */
1140         struct page **compressed_pages;
1141
1142         struct list_head decompressed_secondary_bvecs;
1143         struct page **pagepool;
1144         unsigned int onstack_used, nr_pages;
1145 };
1146
1147 struct z_erofs_bvec_item {
1148         struct z_erofs_bvec bvec;
1149         struct list_head list;
1150 };
1151
1152 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
1153                                          struct z_erofs_bvec *bvec)
1154 {
1155         struct z_erofs_bvec_item *item;
1156
1157         if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) {
1158                 unsigned int pgnr;
1159
1160                 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
1161                 DBG_BUGON(pgnr >= be->nr_pages);
1162                 if (!be->decompressed_pages[pgnr]) {
1163                         be->decompressed_pages[pgnr] = bvec->page;
1164                         return;
1165                 }
1166         }
1167
1168         /* (cold path) one pcluster is requested multiple times */
1169         item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
1170         item->bvec = *bvec;
1171         list_add(&item->list, &be->decompressed_secondary_bvecs);
1172 }
1173
1174 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
1175                                       int err)
1176 {
1177         unsigned int off0 = be->pcl->pageofs_out;
1178         struct list_head *p, *n;
1179
1180         list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
1181                 struct z_erofs_bvec_item *bvi;
1182                 unsigned int end, cur;
1183                 void *dst, *src;
1184
1185                 bvi = container_of(p, struct z_erofs_bvec_item, list);
1186                 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
1187                 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
1188                             bvi->bvec.end);
1189                 dst = kmap_local_page(bvi->bvec.page);
1190                 while (cur < end) {
1191                         unsigned int pgnr, scur, len;
1192
1193                         pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
1194                         DBG_BUGON(pgnr >= be->nr_pages);
1195
1196                         scur = bvi->bvec.offset + cur -
1197                                         ((pgnr << PAGE_SHIFT) - off0);
1198                         len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
1199                         if (!be->decompressed_pages[pgnr]) {
1200                                 err = -EFSCORRUPTED;
1201                                 cur += len;
1202                                 continue;
1203                         }
1204                         src = kmap_local_page(be->decompressed_pages[pgnr]);
1205                         memcpy(dst + cur, src + scur, len);
1206                         kunmap_local(src);
1207                         cur += len;
1208                 }
1209                 kunmap_local(dst);
1210                 if (err)
1211                         z_erofs_page_mark_eio(bvi->bvec.page);
1212                 z_erofs_onlinepage_endio(bvi->bvec.page);
1213                 list_del(p);
1214                 kfree(bvi);
1215         }
1216 }
1217
1218 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
1219 {
1220         struct z_erofs_pcluster *pcl = be->pcl;
1221         struct z_erofs_bvec_iter biter;
1222         struct page *old_bvpage;
1223         int i;
1224
1225         z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
1226         for (i = 0; i < pcl->vcnt; ++i) {
1227                 struct z_erofs_bvec bvec;
1228
1229                 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
1230
1231                 if (old_bvpage)
1232                         z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1233
1234                 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
1235                 z_erofs_do_decompressed_bvec(be, &bvec);
1236         }
1237
1238         old_bvpage = z_erofs_bvec_iter_end(&biter);
1239         if (old_bvpage)
1240                 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
1241 }
1242
1243 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
1244                                   bool *overlapped)
1245 {
1246         struct z_erofs_pcluster *pcl = be->pcl;
1247         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1248         int i, err = 0;
1249
1250         *overlapped = false;
1251         for (i = 0; i < pclusterpages; ++i) {
1252                 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
1253                 struct page *page = bvec->page;
1254
1255                 /* compressed pages ought to be present before decompressing */
1256                 if (!page) {
1257                         DBG_BUGON(1);
1258                         continue;
1259                 }
1260                 be->compressed_pages[i] = page;
1261
1262                 if (z_erofs_is_inline_pcluster(pcl)) {
1263                         if (!PageUptodate(page))
1264                                 err = -EIO;
1265                         continue;
1266                 }
1267
1268                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1269                 if (!z_erofs_is_shortlived_page(page)) {
1270                         if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
1271                                 if (!PageUptodate(page))
1272                                         err = -EIO;
1273                                 continue;
1274                         }
1275                         z_erofs_do_decompressed_bvec(be, bvec);
1276                         *overlapped = true;
1277                 }
1278         }
1279
1280         if (err)
1281                 return err;
1282         return 0;
1283 }
1284
1285 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
1286                                        int err)
1287 {
1288         struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
1289         struct z_erofs_pcluster *pcl = be->pcl;
1290         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
1291         unsigned int i, inputsize;
1292         int err2;
1293         struct page *page;
1294         bool overlapped;
1295
1296         mutex_lock(&pcl->lock);
1297         be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
1298
1299         /* allocate (de)compressed page arrays if cannot be kept on stack */
1300         be->decompressed_pages = NULL;
1301         be->compressed_pages = NULL;
1302         be->onstack_used = 0;
1303         if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
1304                 be->decompressed_pages = be->onstack_pages;
1305                 be->onstack_used = be->nr_pages;
1306                 memset(be->decompressed_pages, 0,
1307                        sizeof(struct page *) * be->nr_pages);
1308         }
1309
1310         if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
1311                 be->compressed_pages = be->onstack_pages + be->onstack_used;
1312
1313         if (!be->decompressed_pages)
1314                 be->decompressed_pages =
1315                         kcalloc(be->nr_pages, sizeof(struct page *),
1316                                 GFP_KERNEL | __GFP_NOFAIL);
1317         if (!be->compressed_pages)
1318                 be->compressed_pages =
1319                         kcalloc(pclusterpages, sizeof(struct page *),
1320                                 GFP_KERNEL | __GFP_NOFAIL);
1321
1322         z_erofs_parse_out_bvecs(be);
1323         err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1324         if (err2)
1325                 err = err2;
1326         if (err)
1327                 goto out;
1328
1329         if (z_erofs_is_inline_pcluster(pcl))
1330                 inputsize = pcl->tailpacking_size;
1331         else
1332                 inputsize = pclusterpages * PAGE_SIZE;
1333
1334         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1335                                         .sb = be->sb,
1336                                         .in = be->compressed_pages,
1337                                         .out = be->decompressed_pages,
1338                                         .pageofs_in = pcl->pageofs_in,
1339                                         .pageofs_out = pcl->pageofs_out,
1340                                         .inputsize = inputsize,
1341                                         .outputsize = pcl->length,
1342                                         .alg = pcl->algorithmformat,
1343                                         .inplace_io = overlapped,
1344                                         .partial_decoding = pcl->partial,
1345                                         .fillgaps = pcl->multibases,
1346                                  }, be->pagepool);
1347
1348 out:
1349         /* must handle all compressed pages before actual file pages */
1350         if (z_erofs_is_inline_pcluster(pcl)) {
1351                 page = pcl->compressed_bvecs[0].page;
1352                 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1353                 put_page(page);
1354         } else {
1355                 for (i = 0; i < pclusterpages; ++i) {
1356                         page = pcl->compressed_bvecs[i].page;
1357
1358                         if (erofs_page_is_managed(sbi, page))
1359                                 continue;
1360
1361                         /* recycle all individual short-lived pages */
1362                         (void)z_erofs_put_shortlivedpage(be->pagepool, page);
1363                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1364                 }
1365         }
1366         if (be->compressed_pages < be->onstack_pages ||
1367             be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1368                 kfree(be->compressed_pages);
1369         z_erofs_fill_other_copies(be, err);
1370
1371         for (i = 0; i < be->nr_pages; ++i) {
1372                 page = be->decompressed_pages[i];
1373                 if (!page)
1374                         continue;
1375
1376                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1377
1378                 /* recycle all individual short-lived pages */
1379                 if (z_erofs_put_shortlivedpage(be->pagepool, page))
1380                         continue;
1381                 if (err)
1382                         z_erofs_page_mark_eio(page);
1383                 z_erofs_onlinepage_endio(page);
1384         }
1385
1386         if (be->decompressed_pages != be->onstack_pages)
1387                 kfree(be->decompressed_pages);
1388
1389         pcl->length = 0;
1390         pcl->partial = true;
1391         pcl->multibases = false;
1392         pcl->bvset.nextpage = NULL;
1393         pcl->vcnt = 0;
1394
1395         /* pcluster lock MUST be taken before the following line */
1396         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1397         mutex_unlock(&pcl->lock);
1398         return err;
1399 }
1400
1401 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1402                                      struct page **pagepool)
1403 {
1404         struct z_erofs_decompress_backend be = {
1405                 .sb = io->sb,
1406                 .pagepool = pagepool,
1407                 .decompressed_secondary_bvecs =
1408                         LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1409         };
1410         z_erofs_next_pcluster_t owned = io->head;
1411
1412         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1413                 /* impossible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1414                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1415                 /* impossible that 'owned' equals Z_EROFS_PCLUSTER_NIL */
1416                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1417
1418                 be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1419                 owned = READ_ONCE(be.pcl->next);
1420
1421                 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1422                 erofs_workgroup_put(&be.pcl->obj);
1423         }
1424 }
1425
1426 static void z_erofs_decompressqueue_work(struct work_struct *work)
1427 {
1428         struct z_erofs_decompressqueue *bgq =
1429                 container_of(work, struct z_erofs_decompressqueue, u.work);
1430         struct page *pagepool = NULL;
1431
1432         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1433         z_erofs_decompress_queue(bgq, &pagepool);
1434         erofs_release_pages(&pagepool);
1435         kvfree(bgq);
1436 }
1437
1438 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1439 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work)
1440 {
1441         z_erofs_decompressqueue_work((struct work_struct *)work);
1442 }
1443 #endif
1444
1445 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1446                                        int bios)
1447 {
1448         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1449
1450         /* wake up the caller thread for sync decompression */
1451         if (io->sync) {
1452                 if (!atomic_add_return(bios, &io->pending_bios))
1453                         complete(&io->u.done);
1454                 return;
1455         }
1456
1457         if (atomic_add_return(bios, &io->pending_bios))
1458                 return;
1459         /* Use (kthread_)work and sync decompression for atomic contexts only */
1460         if (in_atomic() || irqs_disabled()) {
1461 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1462                 struct kthread_worker *worker;
1463
1464                 rcu_read_lock();
1465                 worker = rcu_dereference(
1466                                 z_erofs_pcpu_workers[raw_smp_processor_id()]);
1467                 if (!worker) {
1468                         INIT_WORK(&io->u.work, z_erofs_decompressqueue_work);
1469                         queue_work(z_erofs_workqueue, &io->u.work);
1470                 } else {
1471                         kthread_queue_work(worker, &io->u.kthread_work);
1472                 }
1473                 rcu_read_unlock();
1474 #else
1475                 queue_work(z_erofs_workqueue, &io->u.work);
1476 #endif
1477                 /* enable sync decompression for readahead */
1478                 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1479                         sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1480                 return;
1481         }
1482         z_erofs_decompressqueue_work(&io->u.work);
1483 }
1484
1485 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1486                                                unsigned int nr,
1487                                                struct page **pagepool,
1488                                                struct address_space *mc)
1489 {
1490         const pgoff_t index = pcl->obj.index;
1491         gfp_t gfp = mapping_gfp_mask(mc);
1492         bool tocache = false;
1493
1494         struct address_space *mapping;
1495         struct page *oldpage, *page;
1496         int justfound;
1497
1498 repeat:
1499         page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1500         oldpage = page;
1501
1502         if (!page)
1503                 goto out_allocpage;
1504
1505         justfound = (unsigned long)page & 1UL;
1506         page = (struct page *)((unsigned long)page & ~1UL);
1507
1508         /*
1509          * preallocated cached pages, which is used to avoid direct reclaim
1510          * otherwise, it will go inplace I/O path instead.
1511          */
1512         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1513                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1514                 set_page_private(page, 0);
1515                 tocache = true;
1516                 goto out_tocache;
1517         }
1518         mapping = READ_ONCE(page->mapping);
1519
1520         /*
1521          * file-backed online pages in plcuster are all locked steady,
1522          * therefore it is impossible for `mapping' to be NULL.
1523          */
1524         if (mapping && mapping != mc)
1525                 /* ought to be unmanaged pages */
1526                 goto out;
1527
1528         /* directly return for shortlived page as well */
1529         if (z_erofs_is_shortlived_page(page))
1530                 goto out;
1531
1532         lock_page(page);
1533
1534         /* only true if page reclaim goes wrong, should never happen */
1535         DBG_BUGON(justfound && PagePrivate(page));
1536
1537         /* the page is still in manage cache */
1538         if (page->mapping == mc) {
1539                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1540
1541                 if (!PagePrivate(page)) {
1542                         /*
1543                          * impossible to be !PagePrivate(page) for
1544                          * the current restriction as well if
1545                          * the page is already in compressed_bvecs[].
1546                          */
1547                         DBG_BUGON(!justfound);
1548
1549                         justfound = 0;
1550                         set_page_private(page, (unsigned long)pcl);
1551                         SetPagePrivate(page);
1552                 }
1553
1554                 /* no need to submit io if it is already up-to-date */
1555                 if (PageUptodate(page)) {
1556                         unlock_page(page);
1557                         page = NULL;
1558                 }
1559                 goto out;
1560         }
1561
1562         /*
1563          * the managed page has been truncated, it's unsafe to
1564          * reuse this one, let's allocate a new cache-managed page.
1565          */
1566         DBG_BUGON(page->mapping);
1567         DBG_BUGON(!justfound);
1568
1569         tocache = true;
1570         unlock_page(page);
1571         put_page(page);
1572 out_allocpage:
1573         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1574         if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1575                                oldpage, page)) {
1576                 erofs_pagepool_add(pagepool, page);
1577                 cond_resched();
1578                 goto repeat;
1579         }
1580 out_tocache:
1581         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1582                 /* turn into temporary page if fails (1 ref) */
1583                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1584                 goto out;
1585         }
1586         attach_page_private(page, pcl);
1587         /* drop a refcount added by allocpage (then we have 2 refs here) */
1588         put_page(page);
1589
1590 out:    /* the only exit (for tracing and debugging) */
1591         return page;
1592 }
1593
1594 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb,
1595                               struct z_erofs_decompressqueue *fgq, bool *fg)
1596 {
1597         struct z_erofs_decompressqueue *q;
1598
1599         if (fg && !*fg) {
1600                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1601                 if (!q) {
1602                         *fg = true;
1603                         goto fg_out;
1604                 }
1605 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD
1606                 kthread_init_work(&q->u.kthread_work,
1607                                   z_erofs_decompressqueue_kthread_work);
1608 #else
1609                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1610 #endif
1611         } else {
1612 fg_out:
1613                 q = fgq;
1614                 init_completion(&fgq->u.done);
1615                 atomic_set(&fgq->pending_bios, 0);
1616                 q->eio = false;
1617                 q->sync = true;
1618         }
1619         q->sb = sb;
1620         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1621         return q;
1622 }
1623
1624 /* define decompression jobqueue types */
1625 enum {
1626         JQ_BYPASS,
1627         JQ_SUBMIT,
1628         NR_JOBQUEUES,
1629 };
1630
1631 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1632                                     z_erofs_next_pcluster_t qtail[],
1633                                     z_erofs_next_pcluster_t owned_head)
1634 {
1635         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1636         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1637
1638         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1639         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1640                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1641
1642         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1643
1644         WRITE_ONCE(*submit_qtail, owned_head);
1645         WRITE_ONCE(*bypass_qtail, &pcl->next);
1646
1647         qtail[JQ_BYPASS] = &pcl->next;
1648 }
1649
1650 static void z_erofs_decompressqueue_endio(struct bio *bio)
1651 {
1652         struct z_erofs_decompressqueue *q = bio->bi_private;
1653         blk_status_t err = bio->bi_status;
1654         struct bio_vec *bvec;
1655         struct bvec_iter_all iter_all;
1656
1657         bio_for_each_segment_all(bvec, bio, iter_all) {
1658                 struct page *page = bvec->bv_page;
1659
1660                 DBG_BUGON(PageUptodate(page));
1661                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1662
1663                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1664                         if (!err)
1665                                 SetPageUptodate(page);
1666                         unlock_page(page);
1667                 }
1668         }
1669         if (err)
1670                 q->eio = true;
1671         z_erofs_decompress_kickoff(q, -1);
1672         bio_put(bio);
1673 }
1674
1675 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1676                                  struct page **pagepool,
1677                                  struct z_erofs_decompressqueue *fgq,
1678                                  bool *force_fg)
1679 {
1680         struct super_block *sb = f->inode->i_sb;
1681         struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1682         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1683         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1684         z_erofs_next_pcluster_t owned_head = f->owned_head;
1685         /* bio is NULL initially, so no need to initialize last_{index,bdev} */
1686         pgoff_t last_index;
1687         struct block_device *last_bdev;
1688         unsigned int nr_bios = 0;
1689         struct bio *bio = NULL;
1690         unsigned long pflags;
1691         int memstall = 0;
1692
1693         /*
1694          * if managed cache is enabled, bypass jobqueue is needed,
1695          * no need to read from device for all pclusters in this queue.
1696          */
1697         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1698         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg);
1699
1700         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1701         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1702
1703         /* by default, all need io submission */
1704         q[JQ_SUBMIT]->head = owned_head;
1705
1706         do {
1707                 struct erofs_map_dev mdev;
1708                 struct z_erofs_pcluster *pcl;
1709                 pgoff_t cur, end;
1710                 unsigned int i = 0;
1711                 bool bypass = true;
1712
1713                 /* no possible 'owned_head' equals the following */
1714                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1715                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1716
1717                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1718
1719                 /* close the main owned chain at first */
1720                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1721                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1722                 if (z_erofs_is_inline_pcluster(pcl)) {
1723                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1724                         continue;
1725                 }
1726
1727                 /* no device id here, thus it will always succeed */
1728                 mdev = (struct erofs_map_dev) {
1729                         .m_pa = blknr_to_addr(pcl->obj.index),
1730                 };
1731                 (void)erofs_map_dev(sb, &mdev);
1732
1733                 cur = erofs_blknr(mdev.m_pa);
1734                 end = cur + pcl->pclusterpages;
1735
1736                 do {
1737                         struct page *page;
1738
1739                         page = pickup_page_for_submission(pcl, i++, pagepool,
1740                                                           mc);
1741                         if (!page)
1742                                 continue;
1743
1744                         if (bio && (cur != last_index + 1 ||
1745                                     last_bdev != mdev.m_bdev)) {
1746 submit_bio_retry:
1747                                 submit_bio(bio);
1748                                 if (memstall) {
1749                                         psi_memstall_leave(&pflags);
1750                                         memstall = 0;
1751                                 }
1752                                 bio = NULL;
1753                         }
1754
1755                         if (unlikely(PageWorkingset(page)) && !memstall) {
1756                                 psi_memstall_enter(&pflags);
1757                                 memstall = 1;
1758                         }
1759
1760                         if (!bio) {
1761                                 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1762                                                 REQ_OP_READ, GFP_NOIO);
1763                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1764
1765                                 last_bdev = mdev.m_bdev;
1766                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1767                                         LOG_SECTORS_PER_BLOCK;
1768                                 bio->bi_private = q[JQ_SUBMIT];
1769                                 if (f->readahead)
1770                                         bio->bi_opf |= REQ_RAHEAD;
1771                                 ++nr_bios;
1772                         }
1773
1774                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1775                                 goto submit_bio_retry;
1776
1777                         last_index = cur;
1778                         bypass = false;
1779                 } while (++cur < end);
1780
1781                 if (!bypass)
1782                         qtail[JQ_SUBMIT] = &pcl->next;
1783                 else
1784                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1785         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1786
1787         if (bio) {
1788                 submit_bio(bio);
1789                 if (memstall)
1790                         psi_memstall_leave(&pflags);
1791         }
1792
1793         /*
1794          * although background is preferred, no one is pending for submission.
1795          * don't issue decompression but drop it directly instead.
1796          */
1797         if (!*force_fg && !nr_bios) {
1798                 kvfree(q[JQ_SUBMIT]);
1799                 return;
1800         }
1801         z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios);
1802 }
1803
1804 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1805                              struct page **pagepool, bool force_fg)
1806 {
1807         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1808
1809         if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1810                 return;
1811         z_erofs_submit_queue(f, pagepool, io, &force_fg);
1812
1813         /* handle bypass queue (no i/o pclusters) immediately */
1814         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1815
1816         if (!force_fg)
1817                 return;
1818
1819         /* wait until all bios are completed */
1820         wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1821
1822         /* handle synchronous decompress queue in the caller context */
1823         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1824 }
1825
1826 /*
1827  * Since partial uptodate is still unimplemented for now, we have to use
1828  * approximate readmore strategies as a start.
1829  */
1830 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1831                                       struct readahead_control *rac,
1832                                       erofs_off_t end,
1833                                       struct page **pagepool,
1834                                       bool backmost)
1835 {
1836         struct inode *inode = f->inode;
1837         struct erofs_map_blocks *map = &f->map;
1838         erofs_off_t cur;
1839         int err;
1840
1841         if (backmost) {
1842                 map->m_la = end;
1843                 err = z_erofs_map_blocks_iter(inode, map,
1844                                               EROFS_GET_BLOCKS_READMORE);
1845                 if (err)
1846                         return;
1847
1848                 /* expend ra for the trailing edge if readahead */
1849                 if (rac) {
1850                         loff_t newstart = readahead_pos(rac);
1851
1852                         cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1853                         readahead_expand(rac, newstart, cur - newstart);
1854                         return;
1855                 }
1856                 end = round_up(end, PAGE_SIZE);
1857         } else {
1858                 end = round_up(map->m_la, PAGE_SIZE);
1859
1860                 if (!map->m_llen)
1861                         return;
1862         }
1863
1864         cur = map->m_la + map->m_llen - 1;
1865         while (cur >= end) {
1866                 pgoff_t index = cur >> PAGE_SHIFT;
1867                 struct page *page;
1868
1869                 page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1870                 if (page) {
1871                         if (PageUptodate(page)) {
1872                                 unlock_page(page);
1873                         } else {
1874                                 err = z_erofs_do_read_page(f, page, pagepool);
1875                                 if (err)
1876                                         erofs_err(inode->i_sb,
1877                                                   "readmore error at page %lu @ nid %llu",
1878                                                   index, EROFS_I(inode)->nid);
1879                         }
1880                         put_page(page);
1881                 }
1882
1883                 if (cur < PAGE_SIZE)
1884                         break;
1885                 cur = (index << PAGE_SHIFT) - 1;
1886         }
1887 }
1888
1889 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1890 {
1891         struct page *page = &folio->page;
1892         struct inode *const inode = page->mapping->host;
1893         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1894         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1895         struct page *pagepool = NULL;
1896         int err;
1897
1898         trace_erofs_readpage(page, false);
1899         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1900
1901         z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1902                                   &pagepool, true);
1903         err = z_erofs_do_read_page(&f, page, &pagepool);
1904         z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1905
1906         (void)z_erofs_collector_end(&f);
1907
1908         /* if some compressed cluster ready, need submit them anyway */
1909         z_erofs_runqueue(&f, &pagepool,
1910                          z_erofs_get_sync_decompress_policy(sbi, 0));
1911
1912         if (err)
1913                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1914
1915         erofs_put_metabuf(&f.map.buf);
1916         erofs_release_pages(&pagepool);
1917         return err;
1918 }
1919
1920 static void z_erofs_readahead(struct readahead_control *rac)
1921 {
1922         struct inode *const inode = rac->mapping->host;
1923         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1924         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1925         struct page *pagepool = NULL, *head = NULL, *page;
1926         unsigned int nr_pages;
1927
1928         f.readahead = true;
1929         f.headoffset = readahead_pos(rac);
1930
1931         z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1932                                   readahead_length(rac) - 1, &pagepool, true);
1933         nr_pages = readahead_count(rac);
1934         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1935
1936         while ((page = readahead_page(rac))) {
1937                 set_page_private(page, (unsigned long)head);
1938                 head = page;
1939         }
1940
1941         while (head) {
1942                 struct page *page = head;
1943                 int err;
1944
1945                 /* traversal in reverse order */
1946                 head = (void *)page_private(page);
1947
1948                 err = z_erofs_do_read_page(&f, page, &pagepool);
1949                 if (err)
1950                         erofs_err(inode->i_sb,
1951                                   "readahead error at page %lu @ nid %llu",
1952                                   page->index, EROFS_I(inode)->nid);
1953                 put_page(page);
1954         }
1955         z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1956         (void)z_erofs_collector_end(&f);
1957
1958         z_erofs_runqueue(&f, &pagepool,
1959                          z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1960         erofs_put_metabuf(&f.map.buf);
1961         erofs_release_pages(&pagepool);
1962 }
1963
1964 const struct address_space_operations z_erofs_aops = {
1965         .read_folio = z_erofs_read_folio,
1966         .readahead = z_erofs_readahead,
1967 };