1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2018 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 * Copyright (C) 2022 Alibaba Cloud 6 */ 7 #include "compress.h" 8 #include <linux/psi.h> 9 #include <linux/cpuhotplug.h> 10 #include <trace/events/erofs.h> 11 12 #define Z_EROFS_PCLUSTER_MAX_PAGES (Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE) 13 #define Z_EROFS_INLINE_BVECS 2 14 15 /* 16 * let's leave a type here in case of introducing 17 * another tagged pointer later. 18 */ 19 typedef void *z_erofs_next_pcluster_t; 20 21 struct z_erofs_bvec { 22 struct page *page; 23 int offset; 24 unsigned int end; 25 }; 26 27 #define __Z_EROFS_BVSET(name, total) \ 28 struct name { \ 29 /* point to the next page which contains the following bvecs */ \ 30 struct page *nextpage; \ 31 struct z_erofs_bvec bvec[total]; \ 32 } 33 __Z_EROFS_BVSET(z_erofs_bvset,); 34 __Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS); 35 36 /* 37 * Structure fields follow one of the following exclusion rules. 38 * 39 * I: Modifiable by initialization/destruction paths and read-only 40 * for everyone else; 41 * 42 * L: Field should be protected by the pcluster lock; 43 * 44 * A: Field should be accessed / updated in atomic for parallelized code. 45 */ 46 struct z_erofs_pcluster { 47 struct erofs_workgroup obj; 48 struct mutex lock; 49 50 /* A: point to next chained pcluster or TAILs */ 51 z_erofs_next_pcluster_t next; 52 53 /* L: the maximum decompression size of this round */ 54 unsigned int length; 55 56 /* L: total number of bvecs */ 57 unsigned int vcnt; 58 59 /* I: page offset of start position of decompression */ 60 unsigned short pageofs_out; 61 62 /* I: page offset of inline compressed data */ 63 unsigned short pageofs_in; 64 65 union { 66 /* L: inline a certain number of bvec for bootstrap */ 67 struct z_erofs_bvset_inline bvset; 68 69 /* I: can be used to free the pcluster by RCU. */ 70 struct rcu_head rcu; 71 }; 72 73 union { 74 /* I: physical cluster size in pages */ 75 unsigned short pclusterpages; 76 77 /* I: tailpacking inline compressed size */ 78 unsigned short tailpacking_size; 79 }; 80 81 /* I: compression algorithm format */ 82 unsigned char algorithmformat; 83 84 /* L: whether partial decompression or not */ 85 bool partial; 86 87 /* L: indicate several pageofs_outs or not */ 88 bool multibases; 89 90 /* A: compressed bvecs (can be cached or inplaced pages) */ 91 struct z_erofs_bvec compressed_bvecs[]; 92 }; 93 94 /* let's avoid the valid 32-bit kernel addresses */ 95 96 /* the end of a chain of pclusters */ 97 #define Z_EROFS_PCLUSTER_TAIL ((void *)0x5F0ECAFE) 98 #define Z_EROFS_PCLUSTER_NIL (NULL) 99 100 struct z_erofs_decompressqueue { 101 struct super_block *sb; 102 atomic_t pending_bios; 103 z_erofs_next_pcluster_t head; 104 105 union { 106 struct completion done; 107 struct work_struct work; 108 struct kthread_work kthread_work; 109 } u; 110 bool eio, sync; 111 }; 112 113 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl) 114 { 115 return !pcl->obj.index; 116 } 117 118 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl) 119 { 120 if (z_erofs_is_inline_pcluster(pcl)) 121 return 1; 122 return pcl->pclusterpages; 123 } 124 125 /* 126 * bit 30: I/O error occurred on this page 127 * bit 0 - 29: remaining parts to complete this page 128 */ 129 #define Z_EROFS_PAGE_EIO (1 << 30) 130 131 static inline void z_erofs_onlinepage_init(struct page *page) 132 { 133 union { 134 atomic_t o; 135 unsigned long v; 136 } u = { .o = ATOMIC_INIT(1) }; 137 138 set_page_private(page, u.v); 139 smp_wmb(); 140 SetPagePrivate(page); 141 } 142 143 static inline void z_erofs_onlinepage_split(struct page *page) 144 { 145 atomic_inc((atomic_t *)&page->private); 146 } 147 148 static inline void z_erofs_page_mark_eio(struct page *page) 149 { 150 int orig; 151 152 do { 153 orig = atomic_read((atomic_t *)&page->private); 154 } while (atomic_cmpxchg((atomic_t *)&page->private, orig, 155 orig | Z_EROFS_PAGE_EIO) != orig); 156 } 157 158 static inline void z_erofs_onlinepage_endio(struct page *page) 159 { 160 unsigned int v; 161 162 DBG_BUGON(!PagePrivate(page)); 163 v = atomic_dec_return((atomic_t *)&page->private); 164 if (!(v & ~Z_EROFS_PAGE_EIO)) { 165 set_page_private(page, 0); 166 ClearPagePrivate(page); 167 if (!(v & Z_EROFS_PAGE_EIO)) 168 SetPageUptodate(page); 169 unlock_page(page); 170 } 171 } 172 173 #define Z_EROFS_ONSTACK_PAGES 32 174 175 /* 176 * since pclustersize is variable for big pcluster feature, introduce slab 177 * pools implementation for different pcluster sizes. 178 */ 179 struct z_erofs_pcluster_slab { 180 struct kmem_cache *slab; 181 unsigned int maxpages; 182 char name[48]; 183 }; 184 185 #define _PCLP(n) { .maxpages = n } 186 187 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = { 188 _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128), 189 _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES) 190 }; 191 192 struct z_erofs_bvec_iter { 193 struct page *bvpage; 194 struct z_erofs_bvset *bvset; 195 unsigned int nr, cur; 196 }; 197 198 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter) 199 { 200 if (iter->bvpage) 201 kunmap_local(iter->bvset); 202 return iter->bvpage; 203 } 204 205 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter) 206 { 207 unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec; 208 /* have to access nextpage in advance, otherwise it will be unmapped */ 209 struct page *nextpage = iter->bvset->nextpage; 210 struct page *oldpage; 211 212 DBG_BUGON(!nextpage); 213 oldpage = z_erofs_bvec_iter_end(iter); 214 iter->bvpage = nextpage; 215 iter->bvset = kmap_local_page(nextpage); 216 iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec); 217 iter->cur = 0; 218 return oldpage; 219 } 220 221 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter, 222 struct z_erofs_bvset_inline *bvset, 223 unsigned int bootstrap_nr, 224 unsigned int cur) 225 { 226 *iter = (struct z_erofs_bvec_iter) { 227 .nr = bootstrap_nr, 228 .bvset = (struct z_erofs_bvset *)bvset, 229 }; 230 231 while (cur > iter->nr) { 232 cur -= iter->nr; 233 z_erofs_bvset_flip(iter); 234 } 235 iter->cur = cur; 236 } 237 238 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter, 239 struct z_erofs_bvec *bvec, 240 struct page **candidate_bvpage, 241 struct page **pagepool) 242 { 243 if (iter->cur >= iter->nr) { 244 struct page *nextpage = *candidate_bvpage; 245 246 if (!nextpage) { 247 nextpage = erofs_allocpage(pagepool, GFP_NOFS); 248 if (!nextpage) 249 return -ENOMEM; 250 set_page_private(nextpage, Z_EROFS_SHORTLIVED_PAGE); 251 } 252 DBG_BUGON(iter->bvset->nextpage); 253 iter->bvset->nextpage = nextpage; 254 z_erofs_bvset_flip(iter); 255 256 iter->bvset->nextpage = NULL; 257 *candidate_bvpage = NULL; 258 } 259 iter->bvset->bvec[iter->cur++] = *bvec; 260 return 0; 261 } 262 263 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter, 264 struct z_erofs_bvec *bvec, 265 struct page **old_bvpage) 266 { 267 if (iter->cur == iter->nr) 268 *old_bvpage = z_erofs_bvset_flip(iter); 269 else 270 *old_bvpage = NULL; 271 *bvec = iter->bvset->bvec[iter->cur++]; 272 } 273 274 static void z_erofs_destroy_pcluster_pool(void) 275 { 276 int i; 277 278 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 279 if (!pcluster_pool[i].slab) 280 continue; 281 kmem_cache_destroy(pcluster_pool[i].slab); 282 pcluster_pool[i].slab = NULL; 283 } 284 } 285 286 static int z_erofs_create_pcluster_pool(void) 287 { 288 struct z_erofs_pcluster_slab *pcs; 289 struct z_erofs_pcluster *a; 290 unsigned int size; 291 292 for (pcs = pcluster_pool; 293 pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) { 294 size = struct_size(a, compressed_bvecs, pcs->maxpages); 295 296 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages); 297 pcs->slab = kmem_cache_create(pcs->name, size, 0, 298 SLAB_RECLAIM_ACCOUNT, NULL); 299 if (pcs->slab) 300 continue; 301 302 z_erofs_destroy_pcluster_pool(); 303 return -ENOMEM; 304 } 305 return 0; 306 } 307 308 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages) 309 { 310 int i; 311 312 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 313 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 314 struct z_erofs_pcluster *pcl; 315 316 if (nrpages > pcs->maxpages) 317 continue; 318 319 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS); 320 if (!pcl) 321 return ERR_PTR(-ENOMEM); 322 pcl->pclusterpages = nrpages; 323 return pcl; 324 } 325 return ERR_PTR(-EINVAL); 326 } 327 328 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl) 329 { 330 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 331 int i; 332 333 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 334 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 335 336 if (pclusterpages > pcs->maxpages) 337 continue; 338 339 kmem_cache_free(pcs->slab, pcl); 340 return; 341 } 342 DBG_BUGON(1); 343 } 344 345 static struct workqueue_struct *z_erofs_workqueue __read_mostly; 346 347 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 348 static struct kthread_worker __rcu **z_erofs_pcpu_workers; 349 350 static void erofs_destroy_percpu_workers(void) 351 { 352 struct kthread_worker *worker; 353 unsigned int cpu; 354 355 for_each_possible_cpu(cpu) { 356 worker = rcu_dereference_protected( 357 z_erofs_pcpu_workers[cpu], 1); 358 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL); 359 if (worker) 360 kthread_destroy_worker(worker); 361 } 362 kfree(z_erofs_pcpu_workers); 363 } 364 365 static struct kthread_worker *erofs_init_percpu_worker(int cpu) 366 { 367 struct kthread_worker *worker = 368 kthread_create_worker_on_cpu(cpu, 0, "erofs_worker/%u", cpu); 369 370 if (IS_ERR(worker)) 371 return worker; 372 if (IS_ENABLED(CONFIG_EROFS_FS_PCPU_KTHREAD_HIPRI)) 373 sched_set_fifo_low(worker->task); 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 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it 506 * could be dispatched into bypass queue later due to uptodated managed 507 * pages. All related online pages cannot be reused for inplace I/O (or 508 * bvpage) since it can be directly decoded without I/O submission. 509 */ 510 Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE, 511 /* 512 * The current collection has been linked with the owned chain, and 513 * could also be linked with the remaining collections, which means 514 * if the processing page is the tail page of the collection, thus 515 * the current collection can safely use the whole page (since 516 * the previous collection is under control) for in-place I/O, as 517 * illustrated below: 518 * ________________________________________________________________ 519 * | tail (partial) page | head (partial) page | 520 * | (of the current cl) | (of the previous collection) | 521 * | | | 522 * |__PCLUSTER_FOLLOWED___|___________PCLUSTER_FOLLOWED____________| 523 * 524 * [ (*) the above page can be used as inplace I/O. ] 525 */ 526 Z_EROFS_PCLUSTER_FOLLOWED, 527 }; 528 529 struct z_erofs_decompress_frontend { 530 struct inode *const inode; 531 struct erofs_map_blocks map; 532 struct z_erofs_bvec_iter biter; 533 534 struct page *pagepool; 535 struct page *candidate_bvpage; 536 struct z_erofs_pcluster *pcl; 537 z_erofs_next_pcluster_t owned_head; 538 enum z_erofs_pclustermode mode; 539 540 /* used for applying cache strategy on the fly */ 541 bool backmost; 542 erofs_off_t headoffset; 543 544 /* a pointer used to pick up inplace I/O pages */ 545 unsigned int icur; 546 }; 547 548 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 549 .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 550 .mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true } 551 552 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe) 553 { 554 unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy; 555 556 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 557 return false; 558 559 if (fe->backmost) 560 return true; 561 562 if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 563 fe->map.m_la < fe->headoffset) 564 return true; 565 566 return false; 567 } 568 569 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe) 570 { 571 struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode)); 572 struct z_erofs_pcluster *pcl = fe->pcl; 573 bool shouldalloc = z_erofs_should_alloc_cache(fe); 574 bool standalone = true; 575 /* 576 * optimistic allocation without direct reclaim since inplace I/O 577 * can be used if low memory otherwise. 578 */ 579 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 580 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 581 unsigned int i; 582 583 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED) 584 return; 585 586 for (i = 0; i < pcl->pclusterpages; ++i) { 587 struct page *page; 588 void *t; /* mark pages just found for debugging */ 589 struct page *newpage = NULL; 590 591 /* the compressed page was loaded before */ 592 if (READ_ONCE(pcl->compressed_bvecs[i].page)) 593 continue; 594 595 page = find_get_page(mc, pcl->obj.index + i); 596 597 if (page) { 598 t = (void *)((unsigned long)page | 1); 599 } else { 600 /* I/O is needed, no possible to decompress directly */ 601 standalone = false; 602 if (!shouldalloc) 603 continue; 604 605 /* 606 * try to use cached I/O if page allocation 607 * succeeds or fallback to in-place I/O instead 608 * to avoid any direct reclaim. 609 */ 610 newpage = erofs_allocpage(&fe->pagepool, gfp); 611 if (!newpage) 612 continue; 613 set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE); 614 t = (void *)((unsigned long)newpage | 1); 615 } 616 617 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t)) 618 continue; 619 620 if (page) 621 put_page(page); 622 else if (newpage) 623 erofs_pagepool_add(&fe->pagepool, newpage); 624 } 625 626 /* 627 * don't do inplace I/O if all compressed pages are available in 628 * managed cache since it can be moved to the bypass queue instead. 629 */ 630 if (standalone) 631 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 632 } 633 634 /* called by erofs_shrinker to get rid of all compressed_pages */ 635 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 636 struct erofs_workgroup *grp) 637 { 638 struct z_erofs_pcluster *const pcl = 639 container_of(grp, struct z_erofs_pcluster, obj); 640 int i; 641 642 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 643 /* 644 * refcount of workgroup is now freezed as 1, 645 * therefore no need to worry about available decompression users. 646 */ 647 for (i = 0; i < pcl->pclusterpages; ++i) { 648 struct page *page = pcl->compressed_bvecs[i].page; 649 650 if (!page) 651 continue; 652 653 /* block other users from reclaiming or migrating the page */ 654 if (!trylock_page(page)) 655 return -EBUSY; 656 657 if (!erofs_page_is_managed(sbi, page)) 658 continue; 659 660 /* barrier is implied in the following 'unlock_page' */ 661 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 662 detach_page_private(page); 663 unlock_page(page); 664 } 665 return 0; 666 } 667 668 static bool z_erofs_cache_release_folio(struct folio *folio, gfp_t gfp) 669 { 670 struct z_erofs_pcluster *pcl = folio_get_private(folio); 671 bool ret; 672 int i; 673 674 if (!folio_test_private(folio)) 675 return true; 676 677 if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1)) 678 return false; 679 680 ret = false; 681 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 682 for (i = 0; i < pcl->pclusterpages; ++i) { 683 if (pcl->compressed_bvecs[i].page == &folio->page) { 684 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 685 ret = true; 686 break; 687 } 688 } 689 erofs_workgroup_unfreeze(&pcl->obj, 1); 690 691 if (ret) 692 folio_detach_private(folio); 693 return ret; 694 } 695 696 /* 697 * It will be called only on inode eviction. In case that there are still some 698 * decompression requests in progress, wait with rescheduling for a bit here. 699 * An extra lock could be introduced instead but it seems unnecessary. 700 */ 701 static void z_erofs_cache_invalidate_folio(struct folio *folio, 702 size_t offset, size_t length) 703 { 704 const size_t stop = length + offset; 705 706 /* Check for potential overflow in debug mode */ 707 DBG_BUGON(stop > folio_size(folio) || stop < length); 708 709 if (offset == 0 && stop == folio_size(folio)) 710 while (!z_erofs_cache_release_folio(folio, GFP_NOFS)) 711 cond_resched(); 712 } 713 714 static const struct address_space_operations z_erofs_cache_aops = { 715 .release_folio = z_erofs_cache_release_folio, 716 .invalidate_folio = z_erofs_cache_invalidate_folio, 717 }; 718 719 int erofs_init_managed_cache(struct super_block *sb) 720 { 721 struct inode *const inode = new_inode(sb); 722 723 if (!inode) 724 return -ENOMEM; 725 726 set_nlink(inode, 1); 727 inode->i_size = OFFSET_MAX; 728 inode->i_mapping->a_ops = &z_erofs_cache_aops; 729 mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS); 730 EROFS_SB(sb)->managed_cache = inode; 731 return 0; 732 } 733 734 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe, 735 struct z_erofs_bvec *bvec) 736 { 737 struct z_erofs_pcluster *const pcl = fe->pcl; 738 739 while (fe->icur > 0) { 740 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page, 741 NULL, bvec->page)) { 742 pcl->compressed_bvecs[fe->icur] = *bvec; 743 return true; 744 } 745 } 746 return false; 747 } 748 749 /* callers must be with pcluster lock held */ 750 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe, 751 struct z_erofs_bvec *bvec, bool exclusive) 752 { 753 int ret; 754 755 if (exclusive) { 756 /* give priority for inplaceio to use file pages first */ 757 if (z_erofs_try_inplace_io(fe, bvec)) 758 return 0; 759 /* otherwise, check if it can be used as a bvpage */ 760 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED && 761 !fe->candidate_bvpage) 762 fe->candidate_bvpage = bvec->page; 763 } 764 ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage, 765 &fe->pagepool); 766 fe->pcl->vcnt += (ret >= 0); 767 return ret; 768 } 769 770 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f) 771 { 772 struct z_erofs_pcluster *pcl = f->pcl; 773 z_erofs_next_pcluster_t *owned_head = &f->owned_head; 774 775 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 776 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 777 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 778 *owned_head = &pcl->next; 779 /* so we can attach this pcluster to our submission chain. */ 780 f->mode = Z_EROFS_PCLUSTER_FOLLOWED; 781 return; 782 } 783 784 /* type 2, it belongs to an ongoing chain */ 785 f->mode = Z_EROFS_PCLUSTER_INFLIGHT; 786 } 787 788 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe) 789 { 790 struct erofs_map_blocks *map = &fe->map; 791 bool ztailpacking = map->m_flags & EROFS_MAP_META; 792 struct z_erofs_pcluster *pcl; 793 struct erofs_workgroup *grp; 794 int err; 795 796 if (!(map->m_flags & EROFS_MAP_ENCODED) || 797 (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) { 798 DBG_BUGON(1); 799 return -EFSCORRUPTED; 800 } 801 802 /* no available pcluster, let's allocate one */ 803 pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 : 804 map->m_plen >> PAGE_SHIFT); 805 if (IS_ERR(pcl)) 806 return PTR_ERR(pcl); 807 808 atomic_set(&pcl->obj.refcount, 1); 809 pcl->algorithmformat = map->m_algorithmformat; 810 pcl->length = 0; 811 pcl->partial = true; 812 813 /* new pclusters should be claimed as type 1, primary and followed */ 814 pcl->next = fe->owned_head; 815 pcl->pageofs_out = map->m_la & ~PAGE_MASK; 816 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED; 817 818 /* 819 * lock all primary followed works before visible to others 820 * and mutex_trylock *never* fails for a new pcluster. 821 */ 822 mutex_init(&pcl->lock); 823 DBG_BUGON(!mutex_trylock(&pcl->lock)); 824 825 if (ztailpacking) { 826 pcl->obj.index = 0; /* which indicates ztailpacking */ 827 pcl->pageofs_in = erofs_blkoff(fe->inode->i_sb, map->m_pa); 828 pcl->tailpacking_size = map->m_plen; 829 } else { 830 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 831 832 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj); 833 if (IS_ERR(grp)) { 834 err = PTR_ERR(grp); 835 goto err_out; 836 } 837 838 if (grp != &pcl->obj) { 839 fe->pcl = container_of(grp, 840 struct z_erofs_pcluster, obj); 841 err = -EEXIST; 842 goto err_out; 843 } 844 } 845 fe->owned_head = &pcl->next; 846 fe->pcl = pcl; 847 return 0; 848 849 err_out: 850 mutex_unlock(&pcl->lock); 851 z_erofs_free_pcluster(pcl); 852 return err; 853 } 854 855 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe) 856 { 857 struct erofs_map_blocks *map = &fe->map; 858 struct erofs_workgroup *grp = NULL; 859 int ret; 860 861 DBG_BUGON(fe->pcl); 862 863 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */ 864 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL); 865 866 if (!(map->m_flags & EROFS_MAP_META)) { 867 grp = erofs_find_workgroup(fe->inode->i_sb, 868 map->m_pa >> PAGE_SHIFT); 869 } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) { 870 DBG_BUGON(1); 871 return -EFSCORRUPTED; 872 } 873 874 if (grp) { 875 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj); 876 ret = -EEXIST; 877 } else { 878 ret = z_erofs_register_pcluster(fe); 879 } 880 881 if (ret == -EEXIST) { 882 mutex_lock(&fe->pcl->lock); 883 z_erofs_try_to_claim_pcluster(fe); 884 } else if (ret) { 885 return ret; 886 } 887 z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset, 888 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt); 889 /* since file-backed online pages are traversed in reverse order */ 890 fe->icur = z_erofs_pclusterpages(fe->pcl); 891 return 0; 892 } 893 894 /* 895 * keep in mind that no referenced pclusters will be freed 896 * only after a RCU grace period. 897 */ 898 static void z_erofs_rcu_callback(struct rcu_head *head) 899 { 900 z_erofs_free_pcluster(container_of(head, 901 struct z_erofs_pcluster, rcu)); 902 } 903 904 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 905 { 906 struct z_erofs_pcluster *const pcl = 907 container_of(grp, struct z_erofs_pcluster, obj); 908 909 call_rcu(&pcl->rcu, z_erofs_rcu_callback); 910 } 911 912 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe) 913 { 914 struct z_erofs_pcluster *pcl = fe->pcl; 915 916 if (!pcl) 917 return false; 918 919 z_erofs_bvec_iter_end(&fe->biter); 920 mutex_unlock(&pcl->lock); 921 922 if (fe->candidate_bvpage) 923 fe->candidate_bvpage = NULL; 924 925 /* 926 * if all pending pages are added, don't hold its reference 927 * any longer if the pcluster isn't hosted by ourselves. 928 */ 929 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE) 930 erofs_workgroup_put(&pcl->obj); 931 932 fe->pcl = NULL; 933 return true; 934 } 935 936 static int z_erofs_read_fragment(struct inode *inode, erofs_off_t pos, 937 struct page *page, unsigned int pageofs, 938 unsigned int len) 939 { 940 struct super_block *sb = inode->i_sb; 941 struct inode *packed_inode = EROFS_I_SB(inode)->packed_inode; 942 struct erofs_buf buf = __EROFS_BUF_INITIALIZER; 943 u8 *src, *dst; 944 unsigned int i, cnt; 945 946 if (!packed_inode) 947 return -EFSCORRUPTED; 948 949 buf.inode = packed_inode; 950 pos += EROFS_I(inode)->z_fragmentoff; 951 for (i = 0; i < len; i += cnt) { 952 cnt = min_t(unsigned int, len - i, 953 sb->s_blocksize - erofs_blkoff(sb, pos)); 954 src = erofs_bread(&buf, erofs_blknr(sb, pos), EROFS_KMAP); 955 if (IS_ERR(src)) { 956 erofs_put_metabuf(&buf); 957 return PTR_ERR(src); 958 } 959 960 dst = kmap_local_page(page); 961 memcpy(dst + pageofs + i, src + erofs_blkoff(sb, pos), cnt); 962 kunmap_local(dst); 963 pos += cnt; 964 } 965 erofs_put_metabuf(&buf); 966 return 0; 967 } 968 969 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 970 struct page *page) 971 { 972 struct inode *const inode = fe->inode; 973 struct erofs_map_blocks *const map = &fe->map; 974 const loff_t offset = page_offset(page); 975 bool tight = true, exclusive; 976 unsigned int cur, end, spiltted; 977 int err = 0; 978 979 /* register locked file pages as online pages in pack */ 980 z_erofs_onlinepage_init(page); 981 982 spiltted = 0; 983 end = PAGE_SIZE; 984 repeat: 985 cur = end - 1; 986 987 if (offset + cur < map->m_la || 988 offset + cur >= map->m_la + map->m_llen) { 989 if (z_erofs_collector_end(fe)) 990 fe->backmost = false; 991 map->m_la = offset + cur; 992 map->m_llen = 0; 993 err = z_erofs_map_blocks_iter(inode, map, 0); 994 if (err) 995 goto out; 996 } else { 997 if (fe->pcl) 998 goto hitted; 999 /* didn't get a valid pcluster previously (very rare) */ 1000 } 1001 1002 if (!(map->m_flags & EROFS_MAP_MAPPED) || 1003 map->m_flags & EROFS_MAP_FRAGMENT) 1004 goto hitted; 1005 1006 err = z_erofs_collector_begin(fe); 1007 if (err) 1008 goto out; 1009 1010 if (z_erofs_is_inline_pcluster(fe->pcl)) { 1011 void *mp; 1012 1013 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb, 1014 erofs_blknr(inode->i_sb, map->m_pa), 1015 EROFS_NO_KMAP); 1016 if (IS_ERR(mp)) { 1017 err = PTR_ERR(mp); 1018 erofs_err(inode->i_sb, 1019 "failed to get inline page, err %d", err); 1020 goto out; 1021 } 1022 get_page(fe->map.buf.page); 1023 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, 1024 fe->map.buf.page); 1025 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 1026 } else { 1027 /* bind cache first when cached decompression is preferred */ 1028 z_erofs_bind_cache(fe); 1029 } 1030 hitted: 1031 /* 1032 * Ensure the current partial page belongs to this submit chain rather 1033 * than other concurrent submit chains or the noio(bypass) chain since 1034 * those chains are handled asynchronously thus the page cannot be used 1035 * for inplace I/O or bvpage (should be processed in a strict order.) 1036 */ 1037 tight &= (fe->mode > Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE); 1038 1039 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 1040 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 1041 zero_user_segment(page, cur, end); 1042 goto next_part; 1043 } 1044 if (map->m_flags & EROFS_MAP_FRAGMENT) { 1045 unsigned int pageofs, skip, len; 1046 1047 if (offset > map->m_la) { 1048 pageofs = 0; 1049 skip = offset - map->m_la; 1050 } else { 1051 pageofs = map->m_la & ~PAGE_MASK; 1052 skip = 0; 1053 } 1054 len = min_t(unsigned int, map->m_llen - skip, end - cur); 1055 err = z_erofs_read_fragment(inode, skip, page, pageofs, len); 1056 if (err) 1057 goto out; 1058 ++spiltted; 1059 tight = false; 1060 goto next_part; 1061 } 1062 1063 exclusive = (!cur && (!spiltted || tight)); 1064 if (cur) 1065 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED); 1066 1067 err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) { 1068 .page = page, 1069 .offset = offset - map->m_la, 1070 .end = end, 1071 }), exclusive); 1072 if (err) 1073 goto out; 1074 1075 z_erofs_onlinepage_split(page); 1076 /* bump up the number of spiltted parts of a page */ 1077 ++spiltted; 1078 if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK)) 1079 fe->pcl->multibases = true; 1080 if (fe->pcl->length < offset + end - map->m_la) { 1081 fe->pcl->length = offset + end - map->m_la; 1082 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK; 1083 } 1084 if ((map->m_flags & EROFS_MAP_FULL_MAPPED) && 1085 !(map->m_flags & EROFS_MAP_PARTIAL_REF) && 1086 fe->pcl->length == map->m_llen) 1087 fe->pcl->partial = false; 1088 next_part: 1089 /* shorten the remaining extent to update progress */ 1090 map->m_llen = offset + cur - map->m_la; 1091 map->m_flags &= ~EROFS_MAP_FULL_MAPPED; 1092 1093 end = cur; 1094 if (end > 0) 1095 goto repeat; 1096 1097 out: 1098 if (err) 1099 z_erofs_page_mark_eio(page); 1100 z_erofs_onlinepage_endio(page); 1101 return err; 1102 } 1103 1104 static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi, 1105 unsigned int readahead_pages) 1106 { 1107 /* auto: enable for read_folio, disable for readahead */ 1108 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) && 1109 !readahead_pages) 1110 return true; 1111 1112 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) && 1113 (readahead_pages <= sbi->opt.max_sync_decompress_pages)) 1114 return true; 1115 1116 return false; 1117 } 1118 1119 static bool z_erofs_page_is_invalidated(struct page *page) 1120 { 1121 return !page->mapping && !z_erofs_is_shortlived_page(page); 1122 } 1123 1124 struct z_erofs_decompress_backend { 1125 struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES]; 1126 struct super_block *sb; 1127 struct z_erofs_pcluster *pcl; 1128 1129 /* pages with the longest decompressed length for deduplication */ 1130 struct page **decompressed_pages; 1131 /* pages to keep the compressed data */ 1132 struct page **compressed_pages; 1133 1134 struct list_head decompressed_secondary_bvecs; 1135 struct page **pagepool; 1136 unsigned int onstack_used, nr_pages; 1137 }; 1138 1139 struct z_erofs_bvec_item { 1140 struct z_erofs_bvec bvec; 1141 struct list_head list; 1142 }; 1143 1144 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be, 1145 struct z_erofs_bvec *bvec) 1146 { 1147 struct z_erofs_bvec_item *item; 1148 1149 if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) { 1150 unsigned int pgnr; 1151 1152 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT; 1153 DBG_BUGON(pgnr >= be->nr_pages); 1154 if (!be->decompressed_pages[pgnr]) { 1155 be->decompressed_pages[pgnr] = bvec->page; 1156 return; 1157 } 1158 } 1159 1160 /* (cold path) one pcluster is requested multiple times */ 1161 item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL); 1162 item->bvec = *bvec; 1163 list_add(&item->list, &be->decompressed_secondary_bvecs); 1164 } 1165 1166 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be, 1167 int err) 1168 { 1169 unsigned int off0 = be->pcl->pageofs_out; 1170 struct list_head *p, *n; 1171 1172 list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) { 1173 struct z_erofs_bvec_item *bvi; 1174 unsigned int end, cur; 1175 void *dst, *src; 1176 1177 bvi = container_of(p, struct z_erofs_bvec_item, list); 1178 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0; 1179 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset, 1180 bvi->bvec.end); 1181 dst = kmap_local_page(bvi->bvec.page); 1182 while (cur < end) { 1183 unsigned int pgnr, scur, len; 1184 1185 pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT; 1186 DBG_BUGON(pgnr >= be->nr_pages); 1187 1188 scur = bvi->bvec.offset + cur - 1189 ((pgnr << PAGE_SHIFT) - off0); 1190 len = min_t(unsigned int, end - cur, PAGE_SIZE - scur); 1191 if (!be->decompressed_pages[pgnr]) { 1192 err = -EFSCORRUPTED; 1193 cur += len; 1194 continue; 1195 } 1196 src = kmap_local_page(be->decompressed_pages[pgnr]); 1197 memcpy(dst + cur, src + scur, len); 1198 kunmap_local(src); 1199 cur += len; 1200 } 1201 kunmap_local(dst); 1202 if (err) 1203 z_erofs_page_mark_eio(bvi->bvec.page); 1204 z_erofs_onlinepage_endio(bvi->bvec.page); 1205 list_del(p); 1206 kfree(bvi); 1207 } 1208 } 1209 1210 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be) 1211 { 1212 struct z_erofs_pcluster *pcl = be->pcl; 1213 struct z_erofs_bvec_iter biter; 1214 struct page *old_bvpage; 1215 int i; 1216 1217 z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0); 1218 for (i = 0; i < pcl->vcnt; ++i) { 1219 struct z_erofs_bvec bvec; 1220 1221 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage); 1222 1223 if (old_bvpage) 1224 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1225 1226 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page)); 1227 z_erofs_do_decompressed_bvec(be, &bvec); 1228 } 1229 1230 old_bvpage = z_erofs_bvec_iter_end(&biter); 1231 if (old_bvpage) 1232 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1233 } 1234 1235 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be, 1236 bool *overlapped) 1237 { 1238 struct z_erofs_pcluster *pcl = be->pcl; 1239 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1240 int i, err = 0; 1241 1242 *overlapped = false; 1243 for (i = 0; i < pclusterpages; ++i) { 1244 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i]; 1245 struct page *page = bvec->page; 1246 1247 /* compressed pages ought to be present before decompressing */ 1248 if (!page) { 1249 DBG_BUGON(1); 1250 continue; 1251 } 1252 be->compressed_pages[i] = page; 1253 1254 if (z_erofs_is_inline_pcluster(pcl)) { 1255 if (!PageUptodate(page)) 1256 err = -EIO; 1257 continue; 1258 } 1259 1260 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1261 if (!z_erofs_is_shortlived_page(page)) { 1262 if (erofs_page_is_managed(EROFS_SB(be->sb), page)) { 1263 if (!PageUptodate(page)) 1264 err = -EIO; 1265 continue; 1266 } 1267 z_erofs_do_decompressed_bvec(be, bvec); 1268 *overlapped = true; 1269 } 1270 } 1271 1272 if (err) 1273 return err; 1274 return 0; 1275 } 1276 1277 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be, 1278 int err) 1279 { 1280 struct erofs_sb_info *const sbi = EROFS_SB(be->sb); 1281 struct z_erofs_pcluster *pcl = be->pcl; 1282 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1283 const struct z_erofs_decompressor *decompressor = 1284 &erofs_decompressors[pcl->algorithmformat]; 1285 unsigned int i, inputsize; 1286 int err2; 1287 struct page *page; 1288 bool overlapped; 1289 1290 mutex_lock(&pcl->lock); 1291 be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT; 1292 1293 /* allocate (de)compressed page arrays if cannot be kept on stack */ 1294 be->decompressed_pages = NULL; 1295 be->compressed_pages = NULL; 1296 be->onstack_used = 0; 1297 if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) { 1298 be->decompressed_pages = be->onstack_pages; 1299 be->onstack_used = be->nr_pages; 1300 memset(be->decompressed_pages, 0, 1301 sizeof(struct page *) * be->nr_pages); 1302 } 1303 1304 if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES) 1305 be->compressed_pages = be->onstack_pages + be->onstack_used; 1306 1307 if (!be->decompressed_pages) 1308 be->decompressed_pages = 1309 kvcalloc(be->nr_pages, sizeof(struct page *), 1310 GFP_KERNEL | __GFP_NOFAIL); 1311 if (!be->compressed_pages) 1312 be->compressed_pages = 1313 kvcalloc(pclusterpages, sizeof(struct page *), 1314 GFP_KERNEL | __GFP_NOFAIL); 1315 1316 z_erofs_parse_out_bvecs(be); 1317 err2 = z_erofs_parse_in_bvecs(be, &overlapped); 1318 if (err2) 1319 err = err2; 1320 if (err) 1321 goto out; 1322 1323 if (z_erofs_is_inline_pcluster(pcl)) 1324 inputsize = pcl->tailpacking_size; 1325 else 1326 inputsize = pclusterpages * PAGE_SIZE; 1327 1328 err = decompressor->decompress(&(struct z_erofs_decompress_req) { 1329 .sb = be->sb, 1330 .in = be->compressed_pages, 1331 .out = be->decompressed_pages, 1332 .pageofs_in = pcl->pageofs_in, 1333 .pageofs_out = pcl->pageofs_out, 1334 .inputsize = inputsize, 1335 .outputsize = pcl->length, 1336 .alg = pcl->algorithmformat, 1337 .inplace_io = overlapped, 1338 .partial_decoding = pcl->partial, 1339 .fillgaps = pcl->multibases, 1340 }, be->pagepool); 1341 1342 out: 1343 /* must handle all compressed pages before actual file pages */ 1344 if (z_erofs_is_inline_pcluster(pcl)) { 1345 page = pcl->compressed_bvecs[0].page; 1346 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL); 1347 put_page(page); 1348 } else { 1349 for (i = 0; i < pclusterpages; ++i) { 1350 page = pcl->compressed_bvecs[i].page; 1351 1352 if (erofs_page_is_managed(sbi, page)) 1353 continue; 1354 1355 /* recycle all individual short-lived pages */ 1356 (void)z_erofs_put_shortlivedpage(be->pagepool, page); 1357 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 1358 } 1359 } 1360 if (be->compressed_pages < be->onstack_pages || 1361 be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES) 1362 kvfree(be->compressed_pages); 1363 z_erofs_fill_other_copies(be, err); 1364 1365 for (i = 0; i < be->nr_pages; ++i) { 1366 page = be->decompressed_pages[i]; 1367 if (!page) 1368 continue; 1369 1370 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1371 1372 /* recycle all individual short-lived pages */ 1373 if (z_erofs_put_shortlivedpage(be->pagepool, page)) 1374 continue; 1375 if (err) 1376 z_erofs_page_mark_eio(page); 1377 z_erofs_onlinepage_endio(page); 1378 } 1379 1380 if (be->decompressed_pages != be->onstack_pages) 1381 kvfree(be->decompressed_pages); 1382 1383 pcl->length = 0; 1384 pcl->partial = true; 1385 pcl->multibases = false; 1386 pcl->bvset.nextpage = NULL; 1387 pcl->vcnt = 0; 1388 1389 /* pcluster lock MUST be taken before the following line */ 1390 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 1391 mutex_unlock(&pcl->lock); 1392 return err; 1393 } 1394 1395 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 1396 struct page **pagepool) 1397 { 1398 struct z_erofs_decompress_backend be = { 1399 .sb = io->sb, 1400 .pagepool = pagepool, 1401 .decompressed_secondary_bvecs = 1402 LIST_HEAD_INIT(be.decompressed_secondary_bvecs), 1403 }; 1404 z_erofs_next_pcluster_t owned = io->head; 1405 1406 while (owned != Z_EROFS_PCLUSTER_TAIL) { 1407 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1408 1409 be.pcl = container_of(owned, struct z_erofs_pcluster, next); 1410 owned = READ_ONCE(be.pcl->next); 1411 1412 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0); 1413 erofs_workgroup_put(&be.pcl->obj); 1414 } 1415 } 1416 1417 static void z_erofs_decompressqueue_work(struct work_struct *work) 1418 { 1419 struct z_erofs_decompressqueue *bgq = 1420 container_of(work, struct z_erofs_decompressqueue, u.work); 1421 struct page *pagepool = NULL; 1422 1423 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL); 1424 z_erofs_decompress_queue(bgq, &pagepool); 1425 erofs_release_pages(&pagepool); 1426 kvfree(bgq); 1427 } 1428 1429 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1430 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work) 1431 { 1432 z_erofs_decompressqueue_work((struct work_struct *)work); 1433 } 1434 #endif 1435 1436 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 1437 int bios) 1438 { 1439 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 1440 1441 /* wake up the caller thread for sync decompression */ 1442 if (io->sync) { 1443 if (!atomic_add_return(bios, &io->pending_bios)) 1444 complete(&io->u.done); 1445 return; 1446 } 1447 1448 if (atomic_add_return(bios, &io->pending_bios)) 1449 return; 1450 /* Use (kthread_)work and sync decompression for atomic contexts only */ 1451 if (in_atomic() || irqs_disabled()) { 1452 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1453 struct kthread_worker *worker; 1454 1455 rcu_read_lock(); 1456 worker = rcu_dereference( 1457 z_erofs_pcpu_workers[raw_smp_processor_id()]); 1458 if (!worker) { 1459 INIT_WORK(&io->u.work, z_erofs_decompressqueue_work); 1460 queue_work(z_erofs_workqueue, &io->u.work); 1461 } else { 1462 kthread_queue_work(worker, &io->u.kthread_work); 1463 } 1464 rcu_read_unlock(); 1465 #else 1466 queue_work(z_erofs_workqueue, &io->u.work); 1467 #endif 1468 /* enable sync decompression for readahead */ 1469 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) 1470 sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON; 1471 return; 1472 } 1473 z_erofs_decompressqueue_work(&io->u.work); 1474 } 1475 1476 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1477 unsigned int nr, 1478 struct page **pagepool, 1479 struct address_space *mc) 1480 { 1481 const pgoff_t index = pcl->obj.index; 1482 gfp_t gfp = mapping_gfp_mask(mc); 1483 bool tocache = false; 1484 1485 struct address_space *mapping; 1486 struct page *oldpage, *page; 1487 int justfound; 1488 1489 repeat: 1490 page = READ_ONCE(pcl->compressed_bvecs[nr].page); 1491 oldpage = page; 1492 1493 if (!page) 1494 goto out_allocpage; 1495 1496 justfound = (unsigned long)page & 1UL; 1497 page = (struct page *)((unsigned long)page & ~1UL); 1498 1499 /* 1500 * preallocated cached pages, which is used to avoid direct reclaim 1501 * otherwise, it will go inplace I/O path instead. 1502 */ 1503 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1504 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1505 set_page_private(page, 0); 1506 tocache = true; 1507 goto out_tocache; 1508 } 1509 mapping = READ_ONCE(page->mapping); 1510 1511 /* 1512 * file-backed online pages in plcuster are all locked steady, 1513 * therefore it is impossible for `mapping' to be NULL. 1514 */ 1515 if (mapping && mapping != mc) 1516 /* ought to be unmanaged pages */ 1517 goto out; 1518 1519 /* directly return for shortlived page as well */ 1520 if (z_erofs_is_shortlived_page(page)) 1521 goto out; 1522 1523 lock_page(page); 1524 1525 /* only true if page reclaim goes wrong, should never happen */ 1526 DBG_BUGON(justfound && PagePrivate(page)); 1527 1528 /* the page is still in manage cache */ 1529 if (page->mapping == mc) { 1530 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1531 1532 if (!PagePrivate(page)) { 1533 /* 1534 * impossible to be !PagePrivate(page) for 1535 * the current restriction as well if 1536 * the page is already in compressed_bvecs[]. 1537 */ 1538 DBG_BUGON(!justfound); 1539 1540 justfound = 0; 1541 set_page_private(page, (unsigned long)pcl); 1542 SetPagePrivate(page); 1543 } 1544 1545 /* no need to submit io if it is already up-to-date */ 1546 if (PageUptodate(page)) { 1547 unlock_page(page); 1548 page = NULL; 1549 } 1550 goto out; 1551 } 1552 1553 /* 1554 * the managed page has been truncated, it's unsafe to 1555 * reuse this one, let's allocate a new cache-managed page. 1556 */ 1557 DBG_BUGON(page->mapping); 1558 DBG_BUGON(!justfound); 1559 1560 tocache = true; 1561 unlock_page(page); 1562 put_page(page); 1563 out_allocpage: 1564 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1565 if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page, 1566 oldpage, page)) { 1567 erofs_pagepool_add(pagepool, page); 1568 cond_resched(); 1569 goto repeat; 1570 } 1571 out_tocache: 1572 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1573 /* turn into temporary page if fails (1 ref) */ 1574 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1575 goto out; 1576 } 1577 attach_page_private(page, pcl); 1578 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1579 put_page(page); 1580 1581 out: /* the only exit (for tracing and debugging) */ 1582 return page; 1583 } 1584 1585 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb, 1586 struct z_erofs_decompressqueue *fgq, bool *fg) 1587 { 1588 struct z_erofs_decompressqueue *q; 1589 1590 if (fg && !*fg) { 1591 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1592 if (!q) { 1593 *fg = true; 1594 goto fg_out; 1595 } 1596 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1597 kthread_init_work(&q->u.kthread_work, 1598 z_erofs_decompressqueue_kthread_work); 1599 #else 1600 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1601 #endif 1602 } else { 1603 fg_out: 1604 q = fgq; 1605 init_completion(&fgq->u.done); 1606 atomic_set(&fgq->pending_bios, 0); 1607 q->eio = false; 1608 q->sync = true; 1609 } 1610 q->sb = sb; 1611 q->head = Z_EROFS_PCLUSTER_TAIL; 1612 return q; 1613 } 1614 1615 /* define decompression jobqueue types */ 1616 enum { 1617 JQ_BYPASS, 1618 JQ_SUBMIT, 1619 NR_JOBQUEUES, 1620 }; 1621 1622 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1623 z_erofs_next_pcluster_t qtail[], 1624 z_erofs_next_pcluster_t owned_head) 1625 { 1626 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1627 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1628 1629 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL); 1630 1631 WRITE_ONCE(*submit_qtail, owned_head); 1632 WRITE_ONCE(*bypass_qtail, &pcl->next); 1633 1634 qtail[JQ_BYPASS] = &pcl->next; 1635 } 1636 1637 static void z_erofs_decompressqueue_endio(struct bio *bio) 1638 { 1639 struct z_erofs_decompressqueue *q = bio->bi_private; 1640 blk_status_t err = bio->bi_status; 1641 struct bio_vec *bvec; 1642 struct bvec_iter_all iter_all; 1643 1644 bio_for_each_segment_all(bvec, bio, iter_all) { 1645 struct page *page = bvec->bv_page; 1646 1647 DBG_BUGON(PageUptodate(page)); 1648 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1649 1650 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 1651 if (!err) 1652 SetPageUptodate(page); 1653 unlock_page(page); 1654 } 1655 } 1656 if (err) 1657 q->eio = true; 1658 z_erofs_decompress_kickoff(q, -1); 1659 bio_put(bio); 1660 } 1661 1662 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f, 1663 struct z_erofs_decompressqueue *fgq, 1664 bool *force_fg, bool readahead) 1665 { 1666 struct super_block *sb = f->inode->i_sb; 1667 struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb)); 1668 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1669 struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 1670 z_erofs_next_pcluster_t owned_head = f->owned_head; 1671 /* bio is NULL initially, so no need to initialize last_{index,bdev} */ 1672 pgoff_t last_index; 1673 struct block_device *last_bdev; 1674 unsigned int nr_bios = 0; 1675 struct bio *bio = NULL; 1676 unsigned long pflags; 1677 int memstall = 0; 1678 1679 /* 1680 * if managed cache is enabled, bypass jobqueue is needed, 1681 * no need to read from device for all pclusters in this queue. 1682 */ 1683 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1684 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg); 1685 1686 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1687 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1688 1689 /* by default, all need io submission */ 1690 q[JQ_SUBMIT]->head = owned_head; 1691 1692 do { 1693 struct erofs_map_dev mdev; 1694 struct z_erofs_pcluster *pcl; 1695 pgoff_t cur, end; 1696 unsigned int i = 0; 1697 bool bypass = true; 1698 1699 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1700 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1701 owned_head = READ_ONCE(pcl->next); 1702 1703 if (z_erofs_is_inline_pcluster(pcl)) { 1704 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1705 continue; 1706 } 1707 1708 /* no device id here, thus it will always succeed */ 1709 mdev = (struct erofs_map_dev) { 1710 .m_pa = erofs_pos(sb, pcl->obj.index), 1711 }; 1712 (void)erofs_map_dev(sb, &mdev); 1713 1714 cur = erofs_blknr(sb, mdev.m_pa); 1715 end = cur + pcl->pclusterpages; 1716 1717 do { 1718 struct page *page; 1719 1720 page = pickup_page_for_submission(pcl, i++, 1721 &f->pagepool, mc); 1722 if (!page) 1723 continue; 1724 1725 if (bio && (cur != last_index + 1 || 1726 last_bdev != mdev.m_bdev)) { 1727 submit_bio_retry: 1728 submit_bio(bio); 1729 if (memstall) { 1730 psi_memstall_leave(&pflags); 1731 memstall = 0; 1732 } 1733 bio = NULL; 1734 } 1735 1736 if (unlikely(PageWorkingset(page)) && !memstall) { 1737 psi_memstall_enter(&pflags); 1738 memstall = 1; 1739 } 1740 1741 if (!bio) { 1742 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS, 1743 REQ_OP_READ, GFP_NOIO); 1744 bio->bi_end_io = z_erofs_decompressqueue_endio; 1745 1746 last_bdev = mdev.m_bdev; 1747 bio->bi_iter.bi_sector = (sector_t)cur << 1748 (sb->s_blocksize_bits - 9); 1749 bio->bi_private = q[JQ_SUBMIT]; 1750 if (readahead) 1751 bio->bi_opf |= REQ_RAHEAD; 1752 ++nr_bios; 1753 } 1754 1755 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1756 goto submit_bio_retry; 1757 1758 last_index = cur; 1759 bypass = false; 1760 } while (++cur < end); 1761 1762 if (!bypass) 1763 qtail[JQ_SUBMIT] = &pcl->next; 1764 else 1765 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1766 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1767 1768 if (bio) { 1769 submit_bio(bio); 1770 if (memstall) 1771 psi_memstall_leave(&pflags); 1772 } 1773 1774 /* 1775 * although background is preferred, no one is pending for submission. 1776 * don't issue decompression but drop it directly instead. 1777 */ 1778 if (!*force_fg && !nr_bios) { 1779 kvfree(q[JQ_SUBMIT]); 1780 return; 1781 } 1782 z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios); 1783 } 1784 1785 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f, 1786 bool force_fg, bool ra) 1787 { 1788 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1789 1790 if (f->owned_head == Z_EROFS_PCLUSTER_TAIL) 1791 return; 1792 z_erofs_submit_queue(f, io, &force_fg, ra); 1793 1794 /* handle bypass queue (no i/o pclusters) immediately */ 1795 z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool); 1796 1797 if (!force_fg) 1798 return; 1799 1800 /* wait until all bios are completed */ 1801 wait_for_completion_io(&io[JQ_SUBMIT].u.done); 1802 1803 /* handle synchronous decompress queue in the caller context */ 1804 z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool); 1805 } 1806 1807 /* 1808 * Since partial uptodate is still unimplemented for now, we have to use 1809 * approximate readmore strategies as a start. 1810 */ 1811 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, 1812 struct readahead_control *rac, bool backmost) 1813 { 1814 struct inode *inode = f->inode; 1815 struct erofs_map_blocks *map = &f->map; 1816 erofs_off_t cur, end, headoffset = f->headoffset; 1817 int err; 1818 1819 if (backmost) { 1820 if (rac) 1821 end = headoffset + readahead_length(rac) - 1; 1822 else 1823 end = headoffset + PAGE_SIZE - 1; 1824 map->m_la = end; 1825 err = z_erofs_map_blocks_iter(inode, map, 1826 EROFS_GET_BLOCKS_READMORE); 1827 if (err) 1828 return; 1829 1830 /* expand ra for the trailing edge if readahead */ 1831 if (rac) { 1832 cur = round_up(map->m_la + map->m_llen, PAGE_SIZE); 1833 readahead_expand(rac, headoffset, cur - headoffset); 1834 return; 1835 } 1836 end = round_up(end, PAGE_SIZE); 1837 } else { 1838 end = round_up(map->m_la, PAGE_SIZE); 1839 1840 if (!map->m_llen) 1841 return; 1842 } 1843 1844 cur = map->m_la + map->m_llen - 1; 1845 while (cur >= end) { 1846 pgoff_t index = cur >> PAGE_SHIFT; 1847 struct page *page; 1848 1849 page = erofs_grab_cache_page_nowait(inode->i_mapping, index); 1850 if (page) { 1851 if (PageUptodate(page)) { 1852 unlock_page(page); 1853 } else { 1854 err = z_erofs_do_read_page(f, page); 1855 if (err) 1856 erofs_err(inode->i_sb, 1857 "readmore error at page %lu @ nid %llu", 1858 index, EROFS_I(inode)->nid); 1859 } 1860 put_page(page); 1861 } 1862 1863 if (cur < PAGE_SIZE) 1864 break; 1865 cur = (index << PAGE_SHIFT) - 1; 1866 } 1867 } 1868 1869 static int z_erofs_read_folio(struct file *file, struct folio *folio) 1870 { 1871 struct page *page = &folio->page; 1872 struct inode *const inode = page->mapping->host; 1873 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1874 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1875 int err; 1876 1877 trace_erofs_readpage(page, false); 1878 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1879 1880 z_erofs_pcluster_readmore(&f, NULL, true); 1881 err = z_erofs_do_read_page(&f, page); 1882 z_erofs_pcluster_readmore(&f, NULL, false); 1883 (void)z_erofs_collector_end(&f); 1884 1885 /* if some compressed cluster ready, need submit them anyway */ 1886 z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, 0), false); 1887 1888 if (err) 1889 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1890 1891 erofs_put_metabuf(&f.map.buf); 1892 erofs_release_pages(&f.pagepool); 1893 return err; 1894 } 1895 1896 static void z_erofs_readahead(struct readahead_control *rac) 1897 { 1898 struct inode *const inode = rac->mapping->host; 1899 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1900 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1901 struct page *head = NULL, *page; 1902 unsigned int nr_pages; 1903 1904 f.headoffset = readahead_pos(rac); 1905 1906 z_erofs_pcluster_readmore(&f, rac, true); 1907 nr_pages = readahead_count(rac); 1908 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1909 1910 while ((page = readahead_page(rac))) { 1911 set_page_private(page, (unsigned long)head); 1912 head = page; 1913 } 1914 1915 while (head) { 1916 struct page *page = head; 1917 int err; 1918 1919 /* traversal in reverse order */ 1920 head = (void *)page_private(page); 1921 1922 err = z_erofs_do_read_page(&f, page); 1923 if (err) 1924 erofs_err(inode->i_sb, 1925 "readahead error at page %lu @ nid %llu", 1926 page->index, EROFS_I(inode)->nid); 1927 put_page(page); 1928 } 1929 z_erofs_pcluster_readmore(&f, rac, false); 1930 (void)z_erofs_collector_end(&f); 1931 1932 z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, nr_pages), true); 1933 erofs_put_metabuf(&f.map.buf); 1934 erofs_release_pages(&f.pagepool); 1935 } 1936 1937 const struct address_space_operations z_erofs_aops = { 1938 .read_folio = z_erofs_read_folio, 1939 .readahead = z_erofs_readahead, 1940 }; 1941