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 goto out_error_workqueue_init; 480 481 err = erofs_init_percpu_workers(); 482 if (err) 483 goto out_error_pcpu_worker; 484 485 err = erofs_cpu_hotplug_init(); 486 if (err < 0) 487 goto out_error_cpuhp_init; 488 return err; 489 490 out_error_cpuhp_init: 491 erofs_destroy_percpu_workers(); 492 out_error_pcpu_worker: 493 destroy_workqueue(z_erofs_workqueue); 494 out_error_workqueue_init: 495 z_erofs_destroy_pcluster_pool(); 496 out_error_pcluster_pool: 497 return err; 498 } 499 500 enum z_erofs_pclustermode { 501 Z_EROFS_PCLUSTER_INFLIGHT, 502 /* 503 * The current pclusters was the tail of an exist chain, in addition 504 * that the previous processed chained pclusters are all decided to 505 * be hooked up to it. 506 * A new chain will be created for the remaining pclusters which are 507 * not processed yet, so different from Z_EROFS_PCLUSTER_FOLLOWED, 508 * the next pcluster cannot reuse the whole page safely for inplace I/O 509 * in the following scenario: 510 * ________________________________________________________________ 511 * | tail (partial) page | head (partial) page | 512 * | (belongs to the next pcl) | (belongs to the current pcl) | 513 * |_______PCLUSTER_FOLLOWED______|________PCLUSTER_HOOKED__________| 514 */ 515 Z_EROFS_PCLUSTER_HOOKED, 516 /* 517 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it 518 * could be dispatched into bypass queue later due to uptodated managed 519 * pages. All related online pages cannot be reused for inplace I/O (or 520 * bvpage) since it can be directly decoded without I/O submission. 521 */ 522 Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE, 523 /* 524 * The current collection has been linked with the owned chain, and 525 * could also be linked with the remaining collections, which means 526 * if the processing page is the tail page of the collection, thus 527 * the current collection can safely use the whole page (since 528 * the previous collection is under control) for in-place I/O, as 529 * illustrated below: 530 * ________________________________________________________________ 531 * | tail (partial) page | head (partial) page | 532 * | (of the current cl) | (of the previous collection) | 533 * | PCLUSTER_FOLLOWED or | | 534 * |_____PCLUSTER_HOOKED__|___________PCLUSTER_FOLLOWED____________| 535 * 536 * [ (*) the above page can be used as inplace I/O. ] 537 */ 538 Z_EROFS_PCLUSTER_FOLLOWED, 539 }; 540 541 struct z_erofs_decompress_frontend { 542 struct inode *const inode; 543 struct erofs_map_blocks map; 544 struct z_erofs_bvec_iter biter; 545 546 struct page *candidate_bvpage; 547 struct z_erofs_pcluster *pcl, *tailpcl; 548 z_erofs_next_pcluster_t owned_head; 549 enum z_erofs_pclustermode mode; 550 551 bool readahead; 552 /* used for applying cache strategy on the fly */ 553 bool backmost; 554 erofs_off_t headoffset; 555 556 /* a pointer used to pick up inplace I/O pages */ 557 unsigned int icur; 558 }; 559 560 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 561 .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 562 .mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true } 563 564 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe) 565 { 566 unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy; 567 568 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 569 return false; 570 571 if (fe->backmost) 572 return true; 573 574 if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 575 fe->map.m_la < fe->headoffset) 576 return true; 577 578 return false; 579 } 580 581 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe, 582 struct page **pagepool) 583 { 584 struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode)); 585 struct z_erofs_pcluster *pcl = fe->pcl; 586 bool shouldalloc = z_erofs_should_alloc_cache(fe); 587 bool standalone = true; 588 /* 589 * optimistic allocation without direct reclaim since inplace I/O 590 * can be used if low memory otherwise. 591 */ 592 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 593 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 594 unsigned int i; 595 596 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED) 597 return; 598 599 for (i = 0; i < pcl->pclusterpages; ++i) { 600 struct page *page; 601 void *t; /* mark pages just found for debugging */ 602 struct page *newpage = NULL; 603 604 /* the compressed page was loaded before */ 605 if (READ_ONCE(pcl->compressed_bvecs[i].page)) 606 continue; 607 608 page = find_get_page(mc, pcl->obj.index + i); 609 610 if (page) { 611 t = (void *)((unsigned long)page | 1); 612 } else { 613 /* I/O is needed, no possible to decompress directly */ 614 standalone = false; 615 if (!shouldalloc) 616 continue; 617 618 /* 619 * try to use cached I/O if page allocation 620 * succeeds or fallback to in-place I/O instead 621 * to avoid any direct reclaim. 622 */ 623 newpage = erofs_allocpage(pagepool, gfp); 624 if (!newpage) 625 continue; 626 set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE); 627 t = (void *)((unsigned long)newpage | 1); 628 } 629 630 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t)) 631 continue; 632 633 if (page) 634 put_page(page); 635 else if (newpage) 636 erofs_pagepool_add(pagepool, newpage); 637 } 638 639 /* 640 * don't do inplace I/O if all compressed pages are available in 641 * managed cache since it can be moved to the bypass queue instead. 642 */ 643 if (standalone) 644 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 645 } 646 647 /* called by erofs_shrinker to get rid of all compressed_pages */ 648 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 649 struct erofs_workgroup *grp) 650 { 651 struct z_erofs_pcluster *const pcl = 652 container_of(grp, struct z_erofs_pcluster, obj); 653 int i; 654 655 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 656 /* 657 * refcount of workgroup is now freezed as 1, 658 * therefore no need to worry about available decompression users. 659 */ 660 for (i = 0; i < pcl->pclusterpages; ++i) { 661 struct page *page = pcl->compressed_bvecs[i].page; 662 663 if (!page) 664 continue; 665 666 /* block other users from reclaiming or migrating the page */ 667 if (!trylock_page(page)) 668 return -EBUSY; 669 670 if (!erofs_page_is_managed(sbi, page)) 671 continue; 672 673 /* barrier is implied in the following 'unlock_page' */ 674 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 675 detach_page_private(page); 676 unlock_page(page); 677 } 678 return 0; 679 } 680 681 int erofs_try_to_free_cached_page(struct page *page) 682 { 683 struct z_erofs_pcluster *const pcl = (void *)page_private(page); 684 int ret, i; 685 686 if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1)) 687 return 0; 688 689 ret = 0; 690 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 691 for (i = 0; i < pcl->pclusterpages; ++i) { 692 if (pcl->compressed_bvecs[i].page == page) { 693 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 694 ret = 1; 695 break; 696 } 697 } 698 erofs_workgroup_unfreeze(&pcl->obj, 1); 699 if (ret) 700 detach_page_private(page); 701 return ret; 702 } 703 704 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe, 705 struct z_erofs_bvec *bvec) 706 { 707 struct z_erofs_pcluster *const pcl = fe->pcl; 708 709 while (fe->icur > 0) { 710 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page, 711 NULL, bvec->page)) { 712 pcl->compressed_bvecs[fe->icur] = *bvec; 713 return true; 714 } 715 } 716 return false; 717 } 718 719 /* callers must be with pcluster lock held */ 720 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe, 721 struct z_erofs_bvec *bvec, bool exclusive) 722 { 723 int ret; 724 725 if (exclusive) { 726 /* give priority for inplaceio to use file pages first */ 727 if (z_erofs_try_inplace_io(fe, bvec)) 728 return 0; 729 /* otherwise, check if it can be used as a bvpage */ 730 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED && 731 !fe->candidate_bvpage) 732 fe->candidate_bvpage = bvec->page; 733 } 734 ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage); 735 fe->pcl->vcnt += (ret >= 0); 736 return ret; 737 } 738 739 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f) 740 { 741 struct z_erofs_pcluster *pcl = f->pcl; 742 z_erofs_next_pcluster_t *owned_head = &f->owned_head; 743 744 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 745 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 746 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 747 *owned_head = &pcl->next; 748 /* so we can attach this pcluster to our submission chain. */ 749 f->mode = Z_EROFS_PCLUSTER_FOLLOWED; 750 return; 751 } 752 753 /* 754 * type 2, link to the end of an existing open chain, be careful 755 * that its submission is controlled by the original attached chain. 756 */ 757 if (*owned_head != &pcl->next && pcl != f->tailpcl && 758 cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 759 *owned_head) == Z_EROFS_PCLUSTER_TAIL) { 760 *owned_head = Z_EROFS_PCLUSTER_TAIL; 761 f->mode = Z_EROFS_PCLUSTER_HOOKED; 762 f->tailpcl = NULL; 763 return; 764 } 765 /* type 3, it belongs to a chain, but it isn't the end of the chain */ 766 f->mode = Z_EROFS_PCLUSTER_INFLIGHT; 767 } 768 769 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe) 770 { 771 struct erofs_map_blocks *map = &fe->map; 772 bool ztailpacking = map->m_flags & EROFS_MAP_META; 773 struct z_erofs_pcluster *pcl; 774 struct erofs_workgroup *grp; 775 int err; 776 777 if (!(map->m_flags & EROFS_MAP_ENCODED) || 778 (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) { 779 DBG_BUGON(1); 780 return -EFSCORRUPTED; 781 } 782 783 /* no available pcluster, let's allocate one */ 784 pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 : 785 map->m_plen >> PAGE_SHIFT); 786 if (IS_ERR(pcl)) 787 return PTR_ERR(pcl); 788 789 atomic_set(&pcl->obj.refcount, 1); 790 pcl->algorithmformat = map->m_algorithmformat; 791 pcl->length = 0; 792 pcl->partial = true; 793 794 /* new pclusters should be claimed as type 1, primary and followed */ 795 pcl->next = fe->owned_head; 796 pcl->pageofs_out = map->m_la & ~PAGE_MASK; 797 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED; 798 799 /* 800 * lock all primary followed works before visible to others 801 * and mutex_trylock *never* fails for a new pcluster. 802 */ 803 mutex_init(&pcl->lock); 804 DBG_BUGON(!mutex_trylock(&pcl->lock)); 805 806 if (ztailpacking) { 807 pcl->obj.index = 0; /* which indicates ztailpacking */ 808 pcl->pageofs_in = erofs_blkoff(map->m_pa); 809 pcl->tailpacking_size = map->m_plen; 810 } else { 811 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 812 813 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj); 814 if (IS_ERR(grp)) { 815 err = PTR_ERR(grp); 816 goto err_out; 817 } 818 819 if (grp != &pcl->obj) { 820 fe->pcl = container_of(grp, 821 struct z_erofs_pcluster, obj); 822 err = -EEXIST; 823 goto err_out; 824 } 825 } 826 /* used to check tail merging loop due to corrupted images */ 827 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 828 fe->tailpcl = pcl; 829 fe->owned_head = &pcl->next; 830 fe->pcl = pcl; 831 return 0; 832 833 err_out: 834 mutex_unlock(&pcl->lock); 835 z_erofs_free_pcluster(pcl); 836 return err; 837 } 838 839 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe) 840 { 841 struct erofs_map_blocks *map = &fe->map; 842 struct erofs_workgroup *grp = NULL; 843 int ret; 844 845 DBG_BUGON(fe->pcl); 846 847 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */ 848 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL); 849 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 850 851 if (!(map->m_flags & EROFS_MAP_META)) { 852 grp = erofs_find_workgroup(fe->inode->i_sb, 853 map->m_pa >> PAGE_SHIFT); 854 } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) { 855 DBG_BUGON(1); 856 return -EFSCORRUPTED; 857 } 858 859 if (grp) { 860 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj); 861 ret = -EEXIST; 862 } else { 863 ret = z_erofs_register_pcluster(fe); 864 } 865 866 if (ret == -EEXIST) { 867 mutex_lock(&fe->pcl->lock); 868 /* used to check tail merging loop due to corrupted images */ 869 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 870 fe->tailpcl = fe->pcl; 871 872 z_erofs_try_to_claim_pcluster(fe); 873 } else if (ret) { 874 return ret; 875 } 876 z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset, 877 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt); 878 /* since file-backed online pages are traversed in reverse order */ 879 fe->icur = z_erofs_pclusterpages(fe->pcl); 880 return 0; 881 } 882 883 /* 884 * keep in mind that no referenced pclusters will be freed 885 * only after a RCU grace period. 886 */ 887 static void z_erofs_rcu_callback(struct rcu_head *head) 888 { 889 z_erofs_free_pcluster(container_of(head, 890 struct z_erofs_pcluster, rcu)); 891 } 892 893 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 894 { 895 struct z_erofs_pcluster *const pcl = 896 container_of(grp, struct z_erofs_pcluster, obj); 897 898 call_rcu(&pcl->rcu, z_erofs_rcu_callback); 899 } 900 901 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe) 902 { 903 struct z_erofs_pcluster *pcl = fe->pcl; 904 905 if (!pcl) 906 return false; 907 908 z_erofs_bvec_iter_end(&fe->biter); 909 mutex_unlock(&pcl->lock); 910 911 if (fe->candidate_bvpage) { 912 DBG_BUGON(z_erofs_is_shortlived_page(fe->candidate_bvpage)); 913 fe->candidate_bvpage = NULL; 914 } 915 916 /* 917 * if all pending pages are added, don't hold its reference 918 * any longer if the pcluster isn't hosted by ourselves. 919 */ 920 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE) 921 erofs_workgroup_put(&pcl->obj); 922 923 fe->pcl = NULL; 924 return true; 925 } 926 927 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos, 928 struct page *page, unsigned int pageofs, 929 unsigned int len) 930 { 931 struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode; 932 struct erofs_buf buf = __EROFS_BUF_INITIALIZER; 933 u8 *src, *dst; 934 unsigned int i, cnt; 935 936 if (!packed_inode) 937 return -EFSCORRUPTED; 938 939 pos += EROFS_I(inode)->z_fragmentoff; 940 for (i = 0; i < len; i += cnt) { 941 cnt = min_t(unsigned int, len - i, 942 EROFS_BLKSIZ - erofs_blkoff(pos)); 943 src = erofs_bread(&buf, packed_inode, 944 erofs_blknr(pos), EROFS_KMAP); 945 if (IS_ERR(src)) { 946 erofs_put_metabuf(&buf); 947 return PTR_ERR(src); 948 } 949 950 dst = kmap_local_page(page); 951 memcpy(dst + pageofs + i, src + erofs_blkoff(pos), cnt); 952 kunmap_local(dst); 953 pos += cnt; 954 } 955 erofs_put_metabuf(&buf); 956 return 0; 957 } 958 959 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 960 struct page *page, struct page **pagepool) 961 { 962 struct inode *const inode = fe->inode; 963 struct erofs_map_blocks *const map = &fe->map; 964 const loff_t offset = page_offset(page); 965 bool tight = true, exclusive; 966 unsigned int cur, end, spiltted; 967 int err = 0; 968 969 /* register locked file pages as online pages in pack */ 970 z_erofs_onlinepage_init(page); 971 972 spiltted = 0; 973 end = PAGE_SIZE; 974 repeat: 975 cur = end - 1; 976 977 if (offset + cur < map->m_la || 978 offset + cur >= map->m_la + map->m_llen) { 979 erofs_dbg("out-of-range map @ pos %llu", offset + cur); 980 981 if (z_erofs_collector_end(fe)) 982 fe->backmost = false; 983 map->m_la = offset + cur; 984 map->m_llen = 0; 985 err = z_erofs_map_blocks_iter(inode, map, 0); 986 if (err) 987 goto out; 988 } else { 989 if (fe->pcl) 990 goto hitted; 991 /* didn't get a valid pcluster previously (very rare) */ 992 } 993 994 if (!(map->m_flags & EROFS_MAP_MAPPED) || 995 map->m_flags & EROFS_MAP_FRAGMENT) 996 goto hitted; 997 998 err = z_erofs_collector_begin(fe); 999 if (err) 1000 goto out; 1001 1002 if (z_erofs_is_inline_pcluster(fe->pcl)) { 1003 void *mp; 1004 1005 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb, 1006 erofs_blknr(map->m_pa), EROFS_NO_KMAP); 1007 if (IS_ERR(mp)) { 1008 err = PTR_ERR(mp); 1009 erofs_err(inode->i_sb, 1010 "failed to get inline page, err %d", err); 1011 goto out; 1012 } 1013 get_page(fe->map.buf.page); 1014 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, 1015 fe->map.buf.page); 1016 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 1017 } else { 1018 /* bind cache first when cached decompression is preferred */ 1019 z_erofs_bind_cache(fe, pagepool); 1020 } 1021 hitted: 1022 /* 1023 * Ensure the current partial page belongs to this submit chain rather 1024 * than other concurrent submit chains or the noio(bypass) chain since 1025 * those chains are handled asynchronously thus the page cannot be used 1026 * for inplace I/O or bvpage (should be processed in a strict order.) 1027 */ 1028 tight &= (fe->mode >= Z_EROFS_PCLUSTER_HOOKED && 1029 fe->mode != Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE); 1030 1031 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 1032 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 1033 zero_user_segment(page, cur, end); 1034 goto next_part; 1035 } 1036 if (map->m_flags & EROFS_MAP_FRAGMENT) { 1037 unsigned int pageofs, skip, len; 1038 1039 if (offset > map->m_la) { 1040 pageofs = 0; 1041 skip = offset - map->m_la; 1042 } else { 1043 pageofs = map->m_la & ~PAGE_MASK; 1044 skip = 0; 1045 } 1046 len = min_t(unsigned int, map->m_llen - skip, end - cur); 1047 err = z_erofs_read_fragment(inode, skip, page, pageofs, len); 1048 if (err) 1049 goto out; 1050 ++spiltted; 1051 tight = false; 1052 goto next_part; 1053 } 1054 1055 exclusive = (!cur && (!spiltted || tight)); 1056 if (cur) 1057 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED); 1058 1059 retry: 1060 err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) { 1061 .page = page, 1062 .offset = offset - map->m_la, 1063 .end = end, 1064 }), exclusive); 1065 /* should allocate an additional short-lived page for bvset */ 1066 if (err == -EAGAIN && !fe->candidate_bvpage) { 1067 fe->candidate_bvpage = alloc_page(GFP_NOFS | __GFP_NOFAIL); 1068 set_page_private(fe->candidate_bvpage, 1069 Z_EROFS_SHORTLIVED_PAGE); 1070 goto retry; 1071 } 1072 1073 if (err) { 1074 DBG_BUGON(err == -EAGAIN && fe->candidate_bvpage); 1075 goto out; 1076 } 1077 1078 z_erofs_onlinepage_split(page); 1079 /* bump up the number of spiltted parts of a page */ 1080 ++spiltted; 1081 if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK)) 1082 fe->pcl->multibases = true; 1083 if (fe->pcl->length < offset + end - map->m_la) { 1084 fe->pcl->length = offset + end - map->m_la; 1085 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK; 1086 } 1087 if ((map->m_flags & EROFS_MAP_FULL_MAPPED) && 1088 !(map->m_flags & EROFS_MAP_PARTIAL_REF) && 1089 fe->pcl->length == map->m_llen) 1090 fe->pcl->partial = false; 1091 next_part: 1092 /* shorten the remaining extent to update progress */ 1093 map->m_llen = offset + cur - map->m_la; 1094 map->m_flags &= ~EROFS_MAP_FULL_MAPPED; 1095 1096 end = cur; 1097 if (end > 0) 1098 goto repeat; 1099 1100 out: 1101 if (err) 1102 z_erofs_page_mark_eio(page); 1103 z_erofs_onlinepage_endio(page); 1104 1105 erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 1106 __func__, page, spiltted, map->m_llen); 1107 return err; 1108 } 1109 1110 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi, 1111 unsigned int readahead_pages) 1112 { 1113 /* auto: enable for read_folio, disable for readahead */ 1114 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) && 1115 !readahead_pages) 1116 return true; 1117 1118 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) && 1119 (readahead_pages <= sbi->opt.max_sync_decompress_pages)) 1120 return true; 1121 1122 return false; 1123 } 1124 1125 static bool z_erofs_page_is_invalidated(struct page *page) 1126 { 1127 return !page->mapping && !z_erofs_is_shortlived_page(page); 1128 } 1129 1130 struct z_erofs_decompress_backend { 1131 struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES]; 1132 struct super_block *sb; 1133 struct z_erofs_pcluster *pcl; 1134 1135 /* pages with the longest decompressed length for deduplication */ 1136 struct page **decompressed_pages; 1137 /* pages to keep the compressed data */ 1138 struct page **compressed_pages; 1139 1140 struct list_head decompressed_secondary_bvecs; 1141 struct page **pagepool; 1142 unsigned int onstack_used, nr_pages; 1143 }; 1144 1145 struct z_erofs_bvec_item { 1146 struct z_erofs_bvec bvec; 1147 struct list_head list; 1148 }; 1149 1150 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be, 1151 struct z_erofs_bvec *bvec) 1152 { 1153 struct z_erofs_bvec_item *item; 1154 1155 if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) { 1156 unsigned int pgnr; 1157 1158 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT; 1159 DBG_BUGON(pgnr >= be->nr_pages); 1160 if (!be->decompressed_pages[pgnr]) { 1161 be->decompressed_pages[pgnr] = bvec->page; 1162 return; 1163 } 1164 } 1165 1166 /* (cold path) one pcluster is requested multiple times */ 1167 item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL); 1168 item->bvec = *bvec; 1169 list_add(&item->list, &be->decompressed_secondary_bvecs); 1170 } 1171 1172 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be, 1173 int err) 1174 { 1175 unsigned int off0 = be->pcl->pageofs_out; 1176 struct list_head *p, *n; 1177 1178 list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) { 1179 struct z_erofs_bvec_item *bvi; 1180 unsigned int end, cur; 1181 void *dst, *src; 1182 1183 bvi = container_of(p, struct z_erofs_bvec_item, list); 1184 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0; 1185 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset, 1186 bvi->bvec.end); 1187 dst = kmap_local_page(bvi->bvec.page); 1188 while (cur < end) { 1189 unsigned int pgnr, scur, len; 1190 1191 pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT; 1192 DBG_BUGON(pgnr >= be->nr_pages); 1193 1194 scur = bvi->bvec.offset + cur - 1195 ((pgnr << PAGE_SHIFT) - off0); 1196 len = min_t(unsigned int, end - cur, PAGE_SIZE - scur); 1197 if (!be->decompressed_pages[pgnr]) { 1198 err = -EFSCORRUPTED; 1199 cur += len; 1200 continue; 1201 } 1202 src = kmap_local_page(be->decompressed_pages[pgnr]); 1203 memcpy(dst + cur, src + scur, len); 1204 kunmap_local(src); 1205 cur += len; 1206 } 1207 kunmap_local(dst); 1208 if (err) 1209 z_erofs_page_mark_eio(bvi->bvec.page); 1210 z_erofs_onlinepage_endio(bvi->bvec.page); 1211 list_del(p); 1212 kfree(bvi); 1213 } 1214 } 1215 1216 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be) 1217 { 1218 struct z_erofs_pcluster *pcl = be->pcl; 1219 struct z_erofs_bvec_iter biter; 1220 struct page *old_bvpage; 1221 int i; 1222 1223 z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0); 1224 for (i = 0; i < pcl->vcnt; ++i) { 1225 struct z_erofs_bvec bvec; 1226 1227 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage); 1228 1229 if (old_bvpage) 1230 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1231 1232 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page)); 1233 z_erofs_do_decompressed_bvec(be, &bvec); 1234 } 1235 1236 old_bvpage = z_erofs_bvec_iter_end(&biter); 1237 if (old_bvpage) 1238 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1239 } 1240 1241 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be, 1242 bool *overlapped) 1243 { 1244 struct z_erofs_pcluster *pcl = be->pcl; 1245 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1246 int i, err = 0; 1247 1248 *overlapped = false; 1249 for (i = 0; i < pclusterpages; ++i) { 1250 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i]; 1251 struct page *page = bvec->page; 1252 1253 /* compressed pages ought to be present before decompressing */ 1254 if (!page) { 1255 DBG_BUGON(1); 1256 continue; 1257 } 1258 be->compressed_pages[i] = page; 1259 1260 if (z_erofs_is_inline_pcluster(pcl)) { 1261 if (!PageUptodate(page)) 1262 err = -EIO; 1263 continue; 1264 } 1265 1266 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1267 if (!z_erofs_is_shortlived_page(page)) { 1268 if (erofs_page_is_managed(EROFS_SB(be->sb), page)) { 1269 if (!PageUptodate(page)) 1270 err = -EIO; 1271 continue; 1272 } 1273 z_erofs_do_decompressed_bvec(be, bvec); 1274 *overlapped = true; 1275 } 1276 } 1277 1278 if (err) 1279 return err; 1280 return 0; 1281 } 1282 1283 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be, 1284 int err) 1285 { 1286 struct erofs_sb_info *const sbi = EROFS_SB(be->sb); 1287 struct z_erofs_pcluster *pcl = be->pcl; 1288 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1289 unsigned int i, inputsize; 1290 int err2; 1291 struct page *page; 1292 bool overlapped; 1293 1294 mutex_lock(&pcl->lock); 1295 be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT; 1296 1297 /* allocate (de)compressed page arrays if cannot be kept on stack */ 1298 be->decompressed_pages = NULL; 1299 be->compressed_pages = NULL; 1300 be->onstack_used = 0; 1301 if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) { 1302 be->decompressed_pages = be->onstack_pages; 1303 be->onstack_used = be->nr_pages; 1304 memset(be->decompressed_pages, 0, 1305 sizeof(struct page *) * be->nr_pages); 1306 } 1307 1308 if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES) 1309 be->compressed_pages = be->onstack_pages + be->onstack_used; 1310 1311 if (!be->decompressed_pages) 1312 be->decompressed_pages = 1313 kcalloc(be->nr_pages, sizeof(struct page *), 1314 GFP_KERNEL | __GFP_NOFAIL); 1315 if (!be->compressed_pages) 1316 be->compressed_pages = 1317 kcalloc(pclusterpages, sizeof(struct page *), 1318 GFP_KERNEL | __GFP_NOFAIL); 1319 1320 z_erofs_parse_out_bvecs(be); 1321 err2 = z_erofs_parse_in_bvecs(be, &overlapped); 1322 if (err2) 1323 err = err2; 1324 if (err) 1325 goto out; 1326 1327 if (z_erofs_is_inline_pcluster(pcl)) 1328 inputsize = pcl->tailpacking_size; 1329 else 1330 inputsize = pclusterpages * PAGE_SIZE; 1331 1332 err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 1333 .sb = be->sb, 1334 .in = be->compressed_pages, 1335 .out = be->decompressed_pages, 1336 .pageofs_in = pcl->pageofs_in, 1337 .pageofs_out = pcl->pageofs_out, 1338 .inputsize = inputsize, 1339 .outputsize = pcl->length, 1340 .alg = pcl->algorithmformat, 1341 .inplace_io = overlapped, 1342 .partial_decoding = pcl->partial, 1343 .fillgaps = pcl->multibases, 1344 }, be->pagepool); 1345 1346 out: 1347 /* must handle all compressed pages before actual file pages */ 1348 if (z_erofs_is_inline_pcluster(pcl)) { 1349 page = pcl->compressed_bvecs[0].page; 1350 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL); 1351 put_page(page); 1352 } else { 1353 for (i = 0; i < pclusterpages; ++i) { 1354 page = pcl->compressed_bvecs[i].page; 1355 1356 if (erofs_page_is_managed(sbi, page)) 1357 continue; 1358 1359 /* recycle all individual short-lived pages */ 1360 (void)z_erofs_put_shortlivedpage(be->pagepool, page); 1361 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 1362 } 1363 } 1364 if (be->compressed_pages < be->onstack_pages || 1365 be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES) 1366 kfree(be->compressed_pages); 1367 z_erofs_fill_other_copies(be, err); 1368 1369 for (i = 0; i < be->nr_pages; ++i) { 1370 page = be->decompressed_pages[i]; 1371 if (!page) 1372 continue; 1373 1374 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1375 1376 /* recycle all individual short-lived pages */ 1377 if (z_erofs_put_shortlivedpage(be->pagepool, page)) 1378 continue; 1379 if (err) 1380 z_erofs_page_mark_eio(page); 1381 z_erofs_onlinepage_endio(page); 1382 } 1383 1384 if (be->decompressed_pages != be->onstack_pages) 1385 kfree(be->decompressed_pages); 1386 1387 pcl->length = 0; 1388 pcl->partial = true; 1389 pcl->multibases = false; 1390 pcl->bvset.nextpage = NULL; 1391 pcl->vcnt = 0; 1392 1393 /* pcluster lock MUST be taken before the following line */ 1394 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 1395 mutex_unlock(&pcl->lock); 1396 return err; 1397 } 1398 1399 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 1400 struct page **pagepool) 1401 { 1402 struct z_erofs_decompress_backend be = { 1403 .sb = io->sb, 1404 .pagepool = pagepool, 1405 .decompressed_secondary_bvecs = 1406 LIST_HEAD_INIT(be.decompressed_secondary_bvecs), 1407 }; 1408 z_erofs_next_pcluster_t owned = io->head; 1409 1410 while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 1411 /* impossible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 1412 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 1413 /* impossible that 'owned' equals Z_EROFS_PCLUSTER_NIL */ 1414 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1415 1416 be.pcl = container_of(owned, struct z_erofs_pcluster, next); 1417 owned = READ_ONCE(be.pcl->next); 1418 1419 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0); 1420 erofs_workgroup_put(&be.pcl->obj); 1421 } 1422 } 1423 1424 static void z_erofs_decompressqueue_work(struct work_struct *work) 1425 { 1426 struct z_erofs_decompressqueue *bgq = 1427 container_of(work, struct z_erofs_decompressqueue, u.work); 1428 struct page *pagepool = NULL; 1429 1430 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1431 z_erofs_decompress_queue(bgq, &pagepool); 1432 erofs_release_pages(&pagepool); 1433 kvfree(bgq); 1434 } 1435 1436 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1437 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work) 1438 { 1439 z_erofs_decompressqueue_work((struct work_struct *)work); 1440 } 1441 #endif 1442 1443 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 1444 int bios) 1445 { 1446 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 1447 1448 /* wake up the caller thread for sync decompression */ 1449 if (io->sync) { 1450 if (!atomic_add_return(bios, &io->pending_bios)) 1451 complete(&io->u.done); 1452 return; 1453 } 1454 1455 if (atomic_add_return(bios, &io->pending_bios)) 1456 return; 1457 /* Use (kthread_)work and sync decompression for atomic contexts only */ 1458 if (in_atomic() || irqs_disabled()) { 1459 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1460 struct kthread_worker *worker; 1461 1462 rcu_read_lock(); 1463 worker = rcu_dereference( 1464 z_erofs_pcpu_workers[raw_smp_processor_id()]); 1465 if (!worker) { 1466 INIT_WORK(&io->u.work, z_erofs_decompressqueue_work); 1467 queue_work(z_erofs_workqueue, &io->u.work); 1468 } else { 1469 kthread_queue_work(worker, &io->u.kthread_work); 1470 } 1471 rcu_read_unlock(); 1472 #else 1473 queue_work(z_erofs_workqueue, &io->u.work); 1474 #endif 1475 /* enable sync decompression for readahead */ 1476 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) 1477 sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON; 1478 return; 1479 } 1480 z_erofs_decompressqueue_work(&io->u.work); 1481 } 1482 1483 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1484 unsigned int nr, 1485 struct page **pagepool, 1486 struct address_space *mc) 1487 { 1488 const pgoff_t index = pcl->obj.index; 1489 gfp_t gfp = mapping_gfp_mask(mc); 1490 bool tocache = false; 1491 1492 struct address_space *mapping; 1493 struct page *oldpage, *page; 1494 int justfound; 1495 1496 repeat: 1497 page = READ_ONCE(pcl->compressed_bvecs[nr].page); 1498 oldpage = page; 1499 1500 if (!page) 1501 goto out_allocpage; 1502 1503 justfound = (unsigned long)page & 1UL; 1504 page = (struct page *)((unsigned long)page & ~1UL); 1505 1506 /* 1507 * preallocated cached pages, which is used to avoid direct reclaim 1508 * otherwise, it will go inplace I/O path instead. 1509 */ 1510 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1511 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1512 set_page_private(page, 0); 1513 tocache = true; 1514 goto out_tocache; 1515 } 1516 mapping = READ_ONCE(page->mapping); 1517 1518 /* 1519 * file-backed online pages in plcuster are all locked steady, 1520 * therefore it is impossible for `mapping' to be NULL. 1521 */ 1522 if (mapping && mapping != mc) 1523 /* ought to be unmanaged pages */ 1524 goto out; 1525 1526 /* directly return for shortlived page as well */ 1527 if (z_erofs_is_shortlived_page(page)) 1528 goto out; 1529 1530 lock_page(page); 1531 1532 /* only true if page reclaim goes wrong, should never happen */ 1533 DBG_BUGON(justfound && PagePrivate(page)); 1534 1535 /* the page is still in manage cache */ 1536 if (page->mapping == mc) { 1537 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1538 1539 if (!PagePrivate(page)) { 1540 /* 1541 * impossible to be !PagePrivate(page) for 1542 * the current restriction as well if 1543 * the page is already in compressed_bvecs[]. 1544 */ 1545 DBG_BUGON(!justfound); 1546 1547 justfound = 0; 1548 set_page_private(page, (unsigned long)pcl); 1549 SetPagePrivate(page); 1550 } 1551 1552 /* no need to submit io if it is already up-to-date */ 1553 if (PageUptodate(page)) { 1554 unlock_page(page); 1555 page = NULL; 1556 } 1557 goto out; 1558 } 1559 1560 /* 1561 * the managed page has been truncated, it's unsafe to 1562 * reuse this one, let's allocate a new cache-managed page. 1563 */ 1564 DBG_BUGON(page->mapping); 1565 DBG_BUGON(!justfound); 1566 1567 tocache = true; 1568 unlock_page(page); 1569 put_page(page); 1570 out_allocpage: 1571 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1572 if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page, 1573 oldpage, page)) { 1574 erofs_pagepool_add(pagepool, page); 1575 cond_resched(); 1576 goto repeat; 1577 } 1578 out_tocache: 1579 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1580 /* turn into temporary page if fails (1 ref) */ 1581 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1582 goto out; 1583 } 1584 attach_page_private(page, pcl); 1585 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1586 put_page(page); 1587 1588 out: /* the only exit (for tracing and debugging) */ 1589 return page; 1590 } 1591 1592 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb, 1593 struct z_erofs_decompressqueue *fgq, bool *fg) 1594 { 1595 struct z_erofs_decompressqueue *q; 1596 1597 if (fg && !*fg) { 1598 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1599 if (!q) { 1600 *fg = true; 1601 goto fg_out; 1602 } 1603 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1604 kthread_init_work(&q->u.kthread_work, 1605 z_erofs_decompressqueue_kthread_work); 1606 #else 1607 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1608 #endif 1609 } else { 1610 fg_out: 1611 q = fgq; 1612 init_completion(&fgq->u.done); 1613 atomic_set(&fgq->pending_bios, 0); 1614 q->eio = false; 1615 q->sync = true; 1616 } 1617 q->sb = sb; 1618 q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1619 return q; 1620 } 1621 1622 /* define decompression jobqueue types */ 1623 enum { 1624 JQ_BYPASS, 1625 JQ_SUBMIT, 1626 NR_JOBQUEUES, 1627 }; 1628 1629 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1630 z_erofs_next_pcluster_t qtail[], 1631 z_erofs_next_pcluster_t owned_head) 1632 { 1633 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1634 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1635 1636 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1637 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1638 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1639 1640 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 1641 1642 WRITE_ONCE(*submit_qtail, owned_head); 1643 WRITE_ONCE(*bypass_qtail, &pcl->next); 1644 1645 qtail[JQ_BYPASS] = &pcl->next; 1646 } 1647 1648 static void z_erofs_decompressqueue_endio(struct bio *bio) 1649 { 1650 struct z_erofs_decompressqueue *q = bio->bi_private; 1651 blk_status_t err = bio->bi_status; 1652 struct bio_vec *bvec; 1653 struct bvec_iter_all iter_all; 1654 1655 bio_for_each_segment_all(bvec, bio, iter_all) { 1656 struct page *page = bvec->bv_page; 1657 1658 DBG_BUGON(PageUptodate(page)); 1659 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1660 1661 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 1662 if (!err) 1663 SetPageUptodate(page); 1664 unlock_page(page); 1665 } 1666 } 1667 if (err) 1668 q->eio = true; 1669 z_erofs_decompress_kickoff(q, -1); 1670 bio_put(bio); 1671 } 1672 1673 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f, 1674 struct page **pagepool, 1675 struct z_erofs_decompressqueue *fgq, 1676 bool *force_fg) 1677 { 1678 struct super_block *sb = f->inode->i_sb; 1679 struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb)); 1680 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1681 struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 1682 z_erofs_next_pcluster_t owned_head = f->owned_head; 1683 /* bio is NULL initially, so no need to initialize last_{index,bdev} */ 1684 pgoff_t last_index; 1685 struct block_device *last_bdev; 1686 unsigned int nr_bios = 0; 1687 struct bio *bio = NULL; 1688 unsigned long pflags; 1689 int memstall = 0; 1690 1691 /* 1692 * if managed cache is enabled, bypass jobqueue is needed, 1693 * no need to read from device for all pclusters in this queue. 1694 */ 1695 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1696 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg); 1697 1698 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1699 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1700 1701 /* by default, all need io submission */ 1702 q[JQ_SUBMIT]->head = owned_head; 1703 1704 do { 1705 struct erofs_map_dev mdev; 1706 struct z_erofs_pcluster *pcl; 1707 pgoff_t cur, end; 1708 unsigned int i = 0; 1709 bool bypass = true; 1710 1711 /* no possible 'owned_head' equals the following */ 1712 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1713 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1714 1715 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1716 1717 /* close the main owned chain at first */ 1718 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1719 Z_EROFS_PCLUSTER_TAIL_CLOSED); 1720 if (z_erofs_is_inline_pcluster(pcl)) { 1721 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1722 continue; 1723 } 1724 1725 /* no device id here, thus it will always succeed */ 1726 mdev = (struct erofs_map_dev) { 1727 .m_pa = blknr_to_addr(pcl->obj.index), 1728 }; 1729 (void)erofs_map_dev(sb, &mdev); 1730 1731 cur = erofs_blknr(mdev.m_pa); 1732 end = cur + pcl->pclusterpages; 1733 1734 do { 1735 struct page *page; 1736 1737 page = pickup_page_for_submission(pcl, i++, pagepool, 1738 mc); 1739 if (!page) 1740 continue; 1741 1742 if (bio && (cur != last_index + 1 || 1743 last_bdev != mdev.m_bdev)) { 1744 submit_bio_retry: 1745 submit_bio(bio); 1746 if (memstall) { 1747 psi_memstall_leave(&pflags); 1748 memstall = 0; 1749 } 1750 bio = NULL; 1751 } 1752 1753 if (unlikely(PageWorkingset(page)) && !memstall) { 1754 psi_memstall_enter(&pflags); 1755 memstall = 1; 1756 } 1757 1758 if (!bio) { 1759 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS, 1760 REQ_OP_READ, GFP_NOIO); 1761 bio->bi_end_io = z_erofs_decompressqueue_endio; 1762 1763 last_bdev = mdev.m_bdev; 1764 bio->bi_iter.bi_sector = (sector_t)cur << 1765 LOG_SECTORS_PER_BLOCK; 1766 bio->bi_private = q[JQ_SUBMIT]; 1767 if (f->readahead) 1768 bio->bi_opf |= REQ_RAHEAD; 1769 ++nr_bios; 1770 } 1771 1772 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1773 goto submit_bio_retry; 1774 1775 last_index = cur; 1776 bypass = false; 1777 } while (++cur < end); 1778 1779 if (!bypass) 1780 qtail[JQ_SUBMIT] = &pcl->next; 1781 else 1782 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1783 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1784 1785 if (bio) { 1786 submit_bio(bio); 1787 if (memstall) 1788 psi_memstall_leave(&pflags); 1789 } 1790 1791 /* 1792 * although background is preferred, no one is pending for submission. 1793 * don't issue decompression but drop it directly instead. 1794 */ 1795 if (!*force_fg && !nr_bios) { 1796 kvfree(q[JQ_SUBMIT]); 1797 return; 1798 } 1799 z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios); 1800 } 1801 1802 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f, 1803 struct page **pagepool, bool force_fg) 1804 { 1805 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1806 1807 if (f->owned_head == Z_EROFS_PCLUSTER_TAIL) 1808 return; 1809 z_erofs_submit_queue(f, pagepool, io, &force_fg); 1810 1811 /* handle bypass queue (no i/o pclusters) immediately */ 1812 z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool); 1813 1814 if (!force_fg) 1815 return; 1816 1817 /* wait until all bios are completed */ 1818 wait_for_completion_io(&io[JQ_SUBMIT].u.done); 1819 1820 /* handle synchronous decompress queue in the caller context */ 1821 z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool); 1822 } 1823 1824 /* 1825 * Since partial uptodate is still unimplemented for now, we have to use 1826 * approximate readmore strategies as a start. 1827 */ 1828 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, 1829 struct readahead_control *rac, 1830 erofs_off_t end, 1831 struct page **pagepool, 1832 bool backmost) 1833 { 1834 struct inode *inode = f->inode; 1835 struct erofs_map_blocks *map = &f->map; 1836 erofs_off_t cur; 1837 int err; 1838 1839 if (backmost) { 1840 map->m_la = end; 1841 err = z_erofs_map_blocks_iter(inode, map, 1842 EROFS_GET_BLOCKS_READMORE); 1843 if (err) 1844 return; 1845 1846 /* expend ra for the trailing edge if readahead */ 1847 if (rac) { 1848 loff_t newstart = readahead_pos(rac); 1849 1850 cur = round_up(map->m_la + map->m_llen, PAGE_SIZE); 1851 readahead_expand(rac, newstart, cur - newstart); 1852 return; 1853 } 1854 end = round_up(end, PAGE_SIZE); 1855 } else { 1856 end = round_up(map->m_la, PAGE_SIZE); 1857 1858 if (!map->m_llen) 1859 return; 1860 } 1861 1862 cur = map->m_la + map->m_llen - 1; 1863 while (cur >= end) { 1864 pgoff_t index = cur >> PAGE_SHIFT; 1865 struct page *page; 1866 1867 page = erofs_grab_cache_page_nowait(inode->i_mapping, index); 1868 if (page) { 1869 if (PageUptodate(page)) { 1870 unlock_page(page); 1871 } else { 1872 err = z_erofs_do_read_page(f, page, pagepool); 1873 if (err) 1874 erofs_err(inode->i_sb, 1875 "readmore error at page %lu @ nid %llu", 1876 index, EROFS_I(inode)->nid); 1877 } 1878 put_page(page); 1879 } 1880 1881 if (cur < PAGE_SIZE) 1882 break; 1883 cur = (index << PAGE_SHIFT) - 1; 1884 } 1885 } 1886 1887 static int z_erofs_read_folio(struct file *file, struct folio *folio) 1888 { 1889 struct page *page = &folio->page; 1890 struct inode *const inode = page->mapping->host; 1891 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1892 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1893 struct page *pagepool = NULL; 1894 int err; 1895 1896 trace_erofs_readpage(page, false); 1897 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1898 1899 z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1, 1900 &pagepool, true); 1901 err = z_erofs_do_read_page(&f, page, &pagepool); 1902 z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false); 1903 1904 (void)z_erofs_collector_end(&f); 1905 1906 /* if some compressed cluster ready, need submit them anyway */ 1907 z_erofs_runqueue(&f, &pagepool, 1908 z_erofs_get_sync_decompress_policy(sbi, 0)); 1909 1910 if (err) 1911 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1912 1913 erofs_put_metabuf(&f.map.buf); 1914 erofs_release_pages(&pagepool); 1915 return err; 1916 } 1917 1918 static void z_erofs_readahead(struct readahead_control *rac) 1919 { 1920 struct inode *const inode = rac->mapping->host; 1921 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1922 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1923 struct page *pagepool = NULL, *head = NULL, *page; 1924 unsigned int nr_pages; 1925 1926 f.readahead = true; 1927 f.headoffset = readahead_pos(rac); 1928 1929 z_erofs_pcluster_readmore(&f, rac, f.headoffset + 1930 readahead_length(rac) - 1, &pagepool, true); 1931 nr_pages = readahead_count(rac); 1932 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1933 1934 while ((page = readahead_page(rac))) { 1935 set_page_private(page, (unsigned long)head); 1936 head = page; 1937 } 1938 1939 while (head) { 1940 struct page *page = head; 1941 int err; 1942 1943 /* traversal in reverse order */ 1944 head = (void *)page_private(page); 1945 1946 err = z_erofs_do_read_page(&f, page, &pagepool); 1947 if (err) 1948 erofs_err(inode->i_sb, 1949 "readahead error at page %lu @ nid %llu", 1950 page->index, EROFS_I(inode)->nid); 1951 put_page(page); 1952 } 1953 z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false); 1954 (void)z_erofs_collector_end(&f); 1955 1956 z_erofs_runqueue(&f, &pagepool, 1957 z_erofs_get_sync_decompress_policy(sbi, nr_pages)); 1958 erofs_put_metabuf(&f.map.buf); 1959 erofs_release_pages(&pagepool); 1960 } 1961 1962 const struct address_space_operations z_erofs_aops = { 1963 .read_folio = z_erofs_read_folio, 1964 .readahead = z_erofs_readahead, 1965 }; 1966