1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2018 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 * Copyright (C) 2022 Alibaba Cloud 6 */ 7 #include "compress.h" 8 #include <linux/psi.h> 9 #include <linux/cpuhotplug.h> 10 #include <trace/events/erofs.h> 11 12 #define Z_EROFS_PCLUSTER_MAX_PAGES (Z_EROFS_PCLUSTER_MAX_SIZE / PAGE_SIZE) 13 #define Z_EROFS_INLINE_BVECS 2 14 15 /* 16 * let's leave a type here in case of introducing 17 * another tagged pointer later. 18 */ 19 typedef void *z_erofs_next_pcluster_t; 20 21 struct z_erofs_bvec { 22 struct page *page; 23 int offset; 24 unsigned int end; 25 }; 26 27 #define __Z_EROFS_BVSET(name, total) \ 28 struct name { \ 29 /* point to the next page which contains the following bvecs */ \ 30 struct page *nextpage; \ 31 struct z_erofs_bvec bvec[total]; \ 32 } 33 __Z_EROFS_BVSET(z_erofs_bvset,); 34 __Z_EROFS_BVSET(z_erofs_bvset_inline, Z_EROFS_INLINE_BVECS); 35 36 /* 37 * Structure fields follow one of the following exclusion rules. 38 * 39 * I: Modifiable by initialization/destruction paths and read-only 40 * for everyone else; 41 * 42 * L: Field should be protected by the pcluster lock; 43 * 44 * A: Field should be accessed / updated in atomic for parallelized code. 45 */ 46 struct z_erofs_pcluster { 47 struct erofs_workgroup obj; 48 struct mutex lock; 49 50 /* A: point to next chained pcluster or TAILs */ 51 z_erofs_next_pcluster_t next; 52 53 /* L: the maximum decompression size of this round */ 54 unsigned int length; 55 56 /* L: total number of bvecs */ 57 unsigned int vcnt; 58 59 /* I: page offset of start position of decompression */ 60 unsigned short pageofs_out; 61 62 /* I: page offset of inline compressed data */ 63 unsigned short pageofs_in; 64 65 union { 66 /* L: inline a certain number of bvec for bootstrap */ 67 struct z_erofs_bvset_inline bvset; 68 69 /* I: can be used to free the pcluster by RCU. */ 70 struct rcu_head rcu; 71 }; 72 73 union { 74 /* I: physical cluster size in pages */ 75 unsigned short pclusterpages; 76 77 /* I: tailpacking inline compressed size */ 78 unsigned short tailpacking_size; 79 }; 80 81 /* I: compression algorithm format */ 82 unsigned char algorithmformat; 83 84 /* L: whether partial decompression or not */ 85 bool partial; 86 87 /* L: indicate several pageofs_outs or not */ 88 bool multibases; 89 90 /* A: compressed bvecs (can be cached or inplaced pages) */ 91 struct z_erofs_bvec compressed_bvecs[]; 92 }; 93 94 /* the end of a chain of pclusters */ 95 #define Z_EROFS_PCLUSTER_TAIL ((void *) 0x700 + POISON_POINTER_DELTA) 96 #define Z_EROFS_PCLUSTER_NIL (NULL) 97 98 struct z_erofs_decompressqueue { 99 struct super_block *sb; 100 atomic_t pending_bios; 101 z_erofs_next_pcluster_t head; 102 103 union { 104 struct completion done; 105 struct work_struct work; 106 struct kthread_work kthread_work; 107 } u; 108 bool eio, sync; 109 }; 110 111 static inline bool z_erofs_is_inline_pcluster(struct z_erofs_pcluster *pcl) 112 { 113 return !pcl->obj.index; 114 } 115 116 static inline unsigned int z_erofs_pclusterpages(struct z_erofs_pcluster *pcl) 117 { 118 if (z_erofs_is_inline_pcluster(pcl)) 119 return 1; 120 return pcl->pclusterpages; 121 } 122 123 /* 124 * bit 30: I/O error occurred on this page 125 * bit 0 - 29: remaining parts to complete this page 126 */ 127 #define Z_EROFS_PAGE_EIO (1 << 30) 128 129 static inline void z_erofs_onlinepage_init(struct page *page) 130 { 131 union { 132 atomic_t o; 133 unsigned long v; 134 } u = { .o = ATOMIC_INIT(1) }; 135 136 set_page_private(page, u.v); 137 smp_wmb(); 138 SetPagePrivate(page); 139 } 140 141 static inline void z_erofs_onlinepage_split(struct page *page) 142 { 143 atomic_inc((atomic_t *)&page->private); 144 } 145 146 static inline void z_erofs_page_mark_eio(struct page *page) 147 { 148 int orig; 149 150 do { 151 orig = atomic_read((atomic_t *)&page->private); 152 } while (atomic_cmpxchg((atomic_t *)&page->private, orig, 153 orig | Z_EROFS_PAGE_EIO) != orig); 154 } 155 156 static inline void z_erofs_onlinepage_endio(struct page *page) 157 { 158 unsigned int v; 159 160 DBG_BUGON(!PagePrivate(page)); 161 v = atomic_dec_return((atomic_t *)&page->private); 162 if (!(v & ~Z_EROFS_PAGE_EIO)) { 163 set_page_private(page, 0); 164 ClearPagePrivate(page); 165 if (!(v & Z_EROFS_PAGE_EIO)) 166 SetPageUptodate(page); 167 unlock_page(page); 168 } 169 } 170 171 #define Z_EROFS_ONSTACK_PAGES 32 172 173 /* 174 * since pclustersize is variable for big pcluster feature, introduce slab 175 * pools implementation for different pcluster sizes. 176 */ 177 struct z_erofs_pcluster_slab { 178 struct kmem_cache *slab; 179 unsigned int maxpages; 180 char name[48]; 181 }; 182 183 #define _PCLP(n) { .maxpages = n } 184 185 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = { 186 _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128), 187 _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES) 188 }; 189 190 struct z_erofs_bvec_iter { 191 struct page *bvpage; 192 struct z_erofs_bvset *bvset; 193 unsigned int nr, cur; 194 }; 195 196 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter) 197 { 198 if (iter->bvpage) 199 kunmap_local(iter->bvset); 200 return iter->bvpage; 201 } 202 203 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter) 204 { 205 unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec; 206 /* have to access nextpage in advance, otherwise it will be unmapped */ 207 struct page *nextpage = iter->bvset->nextpage; 208 struct page *oldpage; 209 210 DBG_BUGON(!nextpage); 211 oldpage = z_erofs_bvec_iter_end(iter); 212 iter->bvpage = nextpage; 213 iter->bvset = kmap_local_page(nextpage); 214 iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec); 215 iter->cur = 0; 216 return oldpage; 217 } 218 219 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter, 220 struct z_erofs_bvset_inline *bvset, 221 unsigned int bootstrap_nr, 222 unsigned int cur) 223 { 224 *iter = (struct z_erofs_bvec_iter) { 225 .nr = bootstrap_nr, 226 .bvset = (struct z_erofs_bvset *)bvset, 227 }; 228 229 while (cur > iter->nr) { 230 cur -= iter->nr; 231 z_erofs_bvset_flip(iter); 232 } 233 iter->cur = cur; 234 } 235 236 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter, 237 struct z_erofs_bvec *bvec, 238 struct page **candidate_bvpage, 239 struct page **pagepool) 240 { 241 if (iter->cur >= iter->nr) { 242 struct page *nextpage = *candidate_bvpage; 243 244 if (!nextpage) { 245 nextpage = erofs_allocpage(pagepool, GFP_NOFS); 246 if (!nextpage) 247 return -ENOMEM; 248 set_page_private(nextpage, Z_EROFS_SHORTLIVED_PAGE); 249 } 250 DBG_BUGON(iter->bvset->nextpage); 251 iter->bvset->nextpage = nextpage; 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 return worker; 373 } 374 375 static int erofs_init_percpu_workers(void) 376 { 377 struct kthread_worker *worker; 378 unsigned int cpu; 379 380 z_erofs_pcpu_workers = kcalloc(num_possible_cpus(), 381 sizeof(struct kthread_worker *), GFP_ATOMIC); 382 if (!z_erofs_pcpu_workers) 383 return -ENOMEM; 384 385 for_each_online_cpu(cpu) { /* could miss cpu{off,on}line? */ 386 worker = erofs_init_percpu_worker(cpu); 387 if (!IS_ERR(worker)) 388 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker); 389 } 390 return 0; 391 } 392 #else 393 static inline void erofs_destroy_percpu_workers(void) {} 394 static inline int erofs_init_percpu_workers(void) { return 0; } 395 #endif 396 397 #if defined(CONFIG_HOTPLUG_CPU) && defined(CONFIG_EROFS_FS_PCPU_KTHREAD) 398 static DEFINE_SPINLOCK(z_erofs_pcpu_worker_lock); 399 static enum cpuhp_state erofs_cpuhp_state; 400 401 static int erofs_cpu_online(unsigned int cpu) 402 { 403 struct kthread_worker *worker, *old; 404 405 worker = erofs_init_percpu_worker(cpu); 406 if (IS_ERR(worker)) 407 return PTR_ERR(worker); 408 409 spin_lock(&z_erofs_pcpu_worker_lock); 410 old = rcu_dereference_protected(z_erofs_pcpu_workers[cpu], 411 lockdep_is_held(&z_erofs_pcpu_worker_lock)); 412 if (!old) 413 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], worker); 414 spin_unlock(&z_erofs_pcpu_worker_lock); 415 if (old) 416 kthread_destroy_worker(worker); 417 return 0; 418 } 419 420 static int erofs_cpu_offline(unsigned int cpu) 421 { 422 struct kthread_worker *worker; 423 424 spin_lock(&z_erofs_pcpu_worker_lock); 425 worker = rcu_dereference_protected(z_erofs_pcpu_workers[cpu], 426 lockdep_is_held(&z_erofs_pcpu_worker_lock)); 427 rcu_assign_pointer(z_erofs_pcpu_workers[cpu], NULL); 428 spin_unlock(&z_erofs_pcpu_worker_lock); 429 430 synchronize_rcu(); 431 if (worker) 432 kthread_destroy_worker(worker); 433 return 0; 434 } 435 436 static int erofs_cpu_hotplug_init(void) 437 { 438 int state; 439 440 state = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, 441 "fs/erofs:online", erofs_cpu_online, erofs_cpu_offline); 442 if (state < 0) 443 return state; 444 445 erofs_cpuhp_state = state; 446 return 0; 447 } 448 449 static void erofs_cpu_hotplug_destroy(void) 450 { 451 if (erofs_cpuhp_state) 452 cpuhp_remove_state_nocalls(erofs_cpuhp_state); 453 } 454 #else /* !CONFIG_HOTPLUG_CPU || !CONFIG_EROFS_FS_PCPU_KTHREAD */ 455 static inline int erofs_cpu_hotplug_init(void) { return 0; } 456 static inline void erofs_cpu_hotplug_destroy(void) {} 457 #endif 458 459 void z_erofs_exit_zip_subsystem(void) 460 { 461 erofs_cpu_hotplug_destroy(); 462 erofs_destroy_percpu_workers(); 463 destroy_workqueue(z_erofs_workqueue); 464 z_erofs_destroy_pcluster_pool(); 465 } 466 467 int __init z_erofs_init_zip_subsystem(void) 468 { 469 int err = z_erofs_create_pcluster_pool(); 470 471 if (err) 472 goto out_error_pcluster_pool; 473 474 z_erofs_workqueue = alloc_workqueue("erofs_worker", 475 WQ_UNBOUND | WQ_HIGHPRI, num_possible_cpus()); 476 if (!z_erofs_workqueue) { 477 err = -ENOMEM; 478 goto out_error_workqueue_init; 479 } 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 * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it 504 * could be dispatched into bypass queue later due to uptodated managed 505 * pages. All related online pages cannot be reused for inplace I/O (or 506 * bvpage) since it can be directly decoded without I/O submission. 507 */ 508 Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE, 509 /* 510 * The pcluster was just linked to a decompression chain by us. It can 511 * also be linked with the remaining pclusters, which means if the 512 * processing page is the tail page of a pcluster, this pcluster can 513 * safely use the whole page (since the previous pcluster is within the 514 * same chain) for in-place I/O, as illustrated below: 515 * ___________________________________________________ 516 * | tail (partial) page | head (partial) page | 517 * | (of the current pcl) | (of the previous pcl) | 518 * |___PCLUSTER_FOLLOWED___|_____PCLUSTER_FOLLOWED_____| 519 * 520 * [ (*) the page above can be used as inplace I/O. ] 521 */ 522 Z_EROFS_PCLUSTER_FOLLOWED, 523 }; 524 525 struct z_erofs_decompress_frontend { 526 struct inode *const inode; 527 struct erofs_map_blocks map; 528 struct z_erofs_bvec_iter biter; 529 530 struct page *pagepool; 531 struct page *candidate_bvpage; 532 struct z_erofs_pcluster *pcl; 533 z_erofs_next_pcluster_t owned_head; 534 enum z_erofs_pclustermode mode; 535 536 /* used for applying cache strategy on the fly */ 537 bool backmost; 538 erofs_off_t headoffset; 539 540 /* a pointer used to pick up inplace I/O pages */ 541 unsigned int icur; 542 }; 543 544 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 545 .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 546 .mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true } 547 548 static bool z_erofs_should_alloc_cache(struct z_erofs_decompress_frontend *fe) 549 { 550 unsigned int cachestrategy = EROFS_I_SB(fe->inode)->opt.cache_strategy; 551 552 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 553 return false; 554 555 if (fe->backmost) 556 return true; 557 558 if (cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 559 fe->map.m_la < fe->headoffset) 560 return true; 561 562 return false; 563 } 564 565 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe) 566 { 567 struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode)); 568 struct z_erofs_pcluster *pcl = fe->pcl; 569 bool shouldalloc = z_erofs_should_alloc_cache(fe); 570 bool standalone = true; 571 /* 572 * optimistic allocation without direct reclaim since inplace I/O 573 * can be used if low memory otherwise. 574 */ 575 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 576 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 577 unsigned int i; 578 579 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED) 580 return; 581 582 for (i = 0; i < pcl->pclusterpages; ++i) { 583 struct page *page; 584 void *t; /* mark pages just found for debugging */ 585 struct page *newpage = NULL; 586 587 /* the compressed page was loaded before */ 588 if (READ_ONCE(pcl->compressed_bvecs[i].page)) 589 continue; 590 591 page = find_get_page(mc, pcl->obj.index + i); 592 593 if (page) { 594 t = (void *)((unsigned long)page | 1); 595 } else { 596 /* I/O is needed, no possible to decompress directly */ 597 standalone = false; 598 if (!shouldalloc) 599 continue; 600 601 /* 602 * try to use cached I/O if page allocation 603 * succeeds or fallback to in-place I/O instead 604 * to avoid any direct reclaim. 605 */ 606 newpage = erofs_allocpage(&fe->pagepool, gfp); 607 if (!newpage) 608 continue; 609 set_page_private(newpage, Z_EROFS_PREALLOCATED_PAGE); 610 t = (void *)((unsigned long)newpage | 1); 611 } 612 613 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL, t)) 614 continue; 615 616 if (page) 617 put_page(page); 618 else if (newpage) 619 erofs_pagepool_add(&fe->pagepool, newpage); 620 } 621 622 /* 623 * don't do inplace I/O if all compressed pages are available in 624 * managed cache since it can be moved to the bypass queue instead. 625 */ 626 if (standalone) 627 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 628 } 629 630 /* called by erofs_shrinker to get rid of all compressed_pages */ 631 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 632 struct erofs_workgroup *grp) 633 { 634 struct z_erofs_pcluster *const pcl = 635 container_of(grp, struct z_erofs_pcluster, obj); 636 int i; 637 638 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 639 /* 640 * refcount of workgroup is now freezed as 0, 641 * therefore no need to worry about available decompression users. 642 */ 643 for (i = 0; i < pcl->pclusterpages; ++i) { 644 struct page *page = pcl->compressed_bvecs[i].page; 645 646 if (!page) 647 continue; 648 649 /* block other users from reclaiming or migrating the page */ 650 if (!trylock_page(page)) 651 return -EBUSY; 652 653 if (!erofs_page_is_managed(sbi, page)) 654 continue; 655 656 /* barrier is implied in the following 'unlock_page' */ 657 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 658 detach_page_private(page); 659 unlock_page(page); 660 } 661 return 0; 662 } 663 664 static bool z_erofs_cache_release_folio(struct folio *folio, gfp_t gfp) 665 { 666 struct z_erofs_pcluster *pcl = folio_get_private(folio); 667 bool ret; 668 int i; 669 670 if (!folio_test_private(folio)) 671 return true; 672 673 ret = false; 674 spin_lock(&pcl->obj.lockref.lock); 675 if (pcl->obj.lockref.count > 0) 676 goto out; 677 678 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 679 for (i = 0; i < pcl->pclusterpages; ++i) { 680 if (pcl->compressed_bvecs[i].page == &folio->page) { 681 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 682 ret = true; 683 break; 684 } 685 } 686 if (ret) 687 folio_detach_private(folio); 688 out: 689 spin_unlock(&pcl->obj.lockref.lock); 690 return ret; 691 } 692 693 /* 694 * It will be called only on inode eviction. In case that there are still some 695 * decompression requests in progress, wait with rescheduling for a bit here. 696 * An extra lock could be introduced instead but it seems unnecessary. 697 */ 698 static void z_erofs_cache_invalidate_folio(struct folio *folio, 699 size_t offset, size_t length) 700 { 701 const size_t stop = length + offset; 702 703 /* Check for potential overflow in debug mode */ 704 DBG_BUGON(stop > folio_size(folio) || stop < length); 705 706 if (offset == 0 && stop == folio_size(folio)) 707 while (!z_erofs_cache_release_folio(folio, GFP_NOFS)) 708 cond_resched(); 709 } 710 711 static const struct address_space_operations z_erofs_cache_aops = { 712 .release_folio = z_erofs_cache_release_folio, 713 .invalidate_folio = z_erofs_cache_invalidate_folio, 714 }; 715 716 int erofs_init_managed_cache(struct super_block *sb) 717 { 718 struct inode *const inode = new_inode(sb); 719 720 if (!inode) 721 return -ENOMEM; 722 723 set_nlink(inode, 1); 724 inode->i_size = OFFSET_MAX; 725 inode->i_mapping->a_ops = &z_erofs_cache_aops; 726 mapping_set_gfp_mask(inode->i_mapping, GFP_NOFS); 727 EROFS_SB(sb)->managed_cache = inode; 728 return 0; 729 } 730 731 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe, 732 struct z_erofs_bvec *bvec) 733 { 734 struct z_erofs_pcluster *const pcl = fe->pcl; 735 736 while (fe->icur > 0) { 737 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page, 738 NULL, bvec->page)) { 739 pcl->compressed_bvecs[fe->icur] = *bvec; 740 return true; 741 } 742 } 743 return false; 744 } 745 746 /* callers must be with pcluster lock held */ 747 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe, 748 struct z_erofs_bvec *bvec, bool exclusive) 749 { 750 int ret; 751 752 if (exclusive) { 753 /* give priority for inplaceio to use file pages first */ 754 if (z_erofs_try_inplace_io(fe, bvec)) 755 return 0; 756 /* otherwise, check if it can be used as a bvpage */ 757 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED && 758 !fe->candidate_bvpage) 759 fe->candidate_bvpage = bvec->page; 760 } 761 ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage, 762 &fe->pagepool); 763 fe->pcl->vcnt += (ret >= 0); 764 return ret; 765 } 766 767 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f) 768 { 769 struct z_erofs_pcluster *pcl = f->pcl; 770 z_erofs_next_pcluster_t *owned_head = &f->owned_head; 771 772 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 773 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 774 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 775 *owned_head = &pcl->next; 776 /* so we can attach this pcluster to our submission chain. */ 777 f->mode = Z_EROFS_PCLUSTER_FOLLOWED; 778 return; 779 } 780 781 /* type 2, it belongs to an ongoing chain */ 782 f->mode = Z_EROFS_PCLUSTER_INFLIGHT; 783 } 784 785 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe) 786 { 787 struct erofs_map_blocks *map = &fe->map; 788 bool ztailpacking = map->m_flags & EROFS_MAP_META; 789 struct z_erofs_pcluster *pcl; 790 struct erofs_workgroup *grp; 791 int err; 792 793 if (!(map->m_flags & EROFS_MAP_ENCODED) || 794 (!ztailpacking && !(map->m_pa >> PAGE_SHIFT))) { 795 DBG_BUGON(1); 796 return -EFSCORRUPTED; 797 } 798 799 /* no available pcluster, let's allocate one */ 800 pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 : 801 map->m_plen >> PAGE_SHIFT); 802 if (IS_ERR(pcl)) 803 return PTR_ERR(pcl); 804 805 spin_lock_init(&pcl->obj.lockref.lock); 806 pcl->algorithmformat = map->m_algorithmformat; 807 pcl->length = 0; 808 pcl->partial = true; 809 810 /* new pclusters should be claimed as type 1, primary and followed */ 811 pcl->next = fe->owned_head; 812 pcl->pageofs_out = map->m_la & ~PAGE_MASK; 813 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED; 814 815 /* 816 * lock all primary followed works before visible to others 817 * and mutex_trylock *never* fails for a new pcluster. 818 */ 819 mutex_init(&pcl->lock); 820 DBG_BUGON(!mutex_trylock(&pcl->lock)); 821 822 if (ztailpacking) { 823 pcl->obj.index = 0; /* which indicates ztailpacking */ 824 pcl->pageofs_in = erofs_blkoff(fe->inode->i_sb, map->m_pa); 825 pcl->tailpacking_size = map->m_plen; 826 } else { 827 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 828 829 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj); 830 if (IS_ERR(grp)) { 831 err = PTR_ERR(grp); 832 goto err_out; 833 } 834 835 if (grp != &pcl->obj) { 836 fe->pcl = container_of(grp, 837 struct z_erofs_pcluster, obj); 838 err = -EEXIST; 839 goto err_out; 840 } 841 } 842 fe->owned_head = &pcl->next; 843 fe->pcl = pcl; 844 return 0; 845 846 err_out: 847 mutex_unlock(&pcl->lock); 848 z_erofs_free_pcluster(pcl); 849 return err; 850 } 851 852 static int z_erofs_pcluster_begin(struct z_erofs_decompress_frontend *fe) 853 { 854 struct erofs_map_blocks *map = &fe->map; 855 struct super_block *sb = fe->inode->i_sb; 856 erofs_blk_t blknr = erofs_blknr(sb, map->m_pa); 857 struct erofs_workgroup *grp = NULL; 858 int ret; 859 860 DBG_BUGON(fe->pcl); 861 862 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */ 863 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL); 864 865 if (!(map->m_flags & EROFS_MAP_META)) { 866 grp = erofs_find_workgroup(sb, blknr); 867 } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) { 868 DBG_BUGON(1); 869 return -EFSCORRUPTED; 870 } 871 872 if (grp) { 873 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj); 874 ret = -EEXIST; 875 } else { 876 ret = z_erofs_register_pcluster(fe); 877 } 878 879 if (ret == -EEXIST) { 880 mutex_lock(&fe->pcl->lock); 881 z_erofs_try_to_claim_pcluster(fe); 882 } else if (ret) { 883 return ret; 884 } 885 886 z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset, 887 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt); 888 if (!z_erofs_is_inline_pcluster(fe->pcl)) { 889 /* bind cache first when cached decompression is preferred */ 890 z_erofs_bind_cache(fe); 891 } else { 892 void *mptr; 893 894 mptr = erofs_read_metabuf(&map->buf, sb, blknr, EROFS_NO_KMAP); 895 if (IS_ERR(mptr)) { 896 ret = PTR_ERR(mptr); 897 erofs_err(sb, "failed to get inline data %d", ret); 898 return ret; 899 } 900 get_page(map->buf.page); 901 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page, map->buf.page); 902 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE; 903 } 904 /* file-backed inplace I/O pages are traversed in reverse order */ 905 fe->icur = z_erofs_pclusterpages(fe->pcl); 906 return 0; 907 } 908 909 /* 910 * keep in mind that no referenced pclusters will be freed 911 * only after a RCU grace period. 912 */ 913 static void z_erofs_rcu_callback(struct rcu_head *head) 914 { 915 z_erofs_free_pcluster(container_of(head, 916 struct z_erofs_pcluster, rcu)); 917 } 918 919 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 920 { 921 struct z_erofs_pcluster *const pcl = 922 container_of(grp, struct z_erofs_pcluster, obj); 923 924 call_rcu(&pcl->rcu, z_erofs_rcu_callback); 925 } 926 927 static void z_erofs_pcluster_end(struct z_erofs_decompress_frontend *fe) 928 { 929 struct z_erofs_pcluster *pcl = fe->pcl; 930 931 if (!pcl) 932 return; 933 934 z_erofs_bvec_iter_end(&fe->biter); 935 mutex_unlock(&pcl->lock); 936 937 if (fe->candidate_bvpage) 938 fe->candidate_bvpage = NULL; 939 940 /* 941 * if all pending pages are added, don't hold its reference 942 * any longer if the pcluster isn't hosted by ourselves. 943 */ 944 if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE) 945 erofs_workgroup_put(&pcl->obj); 946 947 fe->pcl = NULL; 948 fe->backmost = false; 949 } 950 951 static int z_erofs_read_fragment(struct super_block *sb, struct page *page, 952 unsigned int cur, unsigned int end, erofs_off_t pos) 953 { 954 struct inode *packed_inode = EROFS_SB(sb)->packed_inode; 955 struct erofs_buf buf = __EROFS_BUF_INITIALIZER; 956 unsigned int cnt; 957 u8 *src; 958 959 if (!packed_inode) 960 return -EFSCORRUPTED; 961 962 buf.inode = packed_inode; 963 for (; cur < end; cur += cnt, pos += cnt) { 964 cnt = min_t(unsigned int, end - cur, 965 sb->s_blocksize - erofs_blkoff(sb, pos)); 966 src = erofs_bread(&buf, erofs_blknr(sb, pos), EROFS_KMAP); 967 if (IS_ERR(src)) { 968 erofs_put_metabuf(&buf); 969 return PTR_ERR(src); 970 } 971 memcpy_to_page(page, cur, src + erofs_blkoff(sb, pos), cnt); 972 } 973 erofs_put_metabuf(&buf); 974 return 0; 975 } 976 977 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 978 struct page *page) 979 { 980 struct inode *const inode = fe->inode; 981 struct erofs_map_blocks *const map = &fe->map; 982 const loff_t offset = page_offset(page); 983 bool tight = true, exclusive; 984 unsigned int cur, end, len, spiltted; 985 int err = 0; 986 987 /* register locked file pages as online pages in pack */ 988 z_erofs_onlinepage_init(page); 989 990 spiltted = 0; 991 end = PAGE_SIZE; 992 repeat: 993 cur = end - 1; 994 995 if (offset + cur < map->m_la || 996 offset + cur >= map->m_la + map->m_llen) { 997 z_erofs_pcluster_end(fe); 998 map->m_la = offset + cur; 999 map->m_llen = 0; 1000 err = z_erofs_map_blocks_iter(inode, map, 0); 1001 if (err) 1002 goto out; 1003 } else if (fe->pcl) { 1004 goto hitted; 1005 } 1006 1007 if ((map->m_flags & EROFS_MAP_MAPPED) && 1008 !(map->m_flags & EROFS_MAP_FRAGMENT)) { 1009 err = z_erofs_pcluster_begin(fe); 1010 if (err) 1011 goto out; 1012 } 1013 hitted: 1014 /* 1015 * Ensure the current partial page belongs to this submit chain rather 1016 * than other concurrent submit chains or the noio(bypass) chain since 1017 * those chains are handled asynchronously thus the page cannot be used 1018 * for inplace I/O or bvpage (should be processed in a strict order.) 1019 */ 1020 tight &= (fe->mode > Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE); 1021 1022 cur = end - min_t(erofs_off_t, offset + end - map->m_la, end); 1023 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 1024 zero_user_segment(page, cur, end); 1025 goto next_part; 1026 } 1027 if (map->m_flags & EROFS_MAP_FRAGMENT) { 1028 erofs_off_t fpos = offset + cur - map->m_la; 1029 1030 len = min_t(unsigned int, map->m_llen - fpos, end - cur); 1031 err = z_erofs_read_fragment(inode->i_sb, page, cur, cur + len, 1032 EROFS_I(inode)->z_fragmentoff + fpos); 1033 if (err) 1034 goto out; 1035 ++spiltted; 1036 tight = false; 1037 goto next_part; 1038 } 1039 1040 exclusive = (!cur && (!spiltted || tight)); 1041 if (cur) 1042 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED); 1043 1044 err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) { 1045 .page = page, 1046 .offset = offset - map->m_la, 1047 .end = end, 1048 }), exclusive); 1049 if (err) 1050 goto out; 1051 1052 z_erofs_onlinepage_split(page); 1053 /* bump up the number of spiltted parts of a page */ 1054 ++spiltted; 1055 if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK)) 1056 fe->pcl->multibases = true; 1057 if (fe->pcl->length < offset + end - map->m_la) { 1058 fe->pcl->length = offset + end - map->m_la; 1059 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK; 1060 } 1061 if ((map->m_flags & EROFS_MAP_FULL_MAPPED) && 1062 !(map->m_flags & EROFS_MAP_PARTIAL_REF) && 1063 fe->pcl->length == map->m_llen) 1064 fe->pcl->partial = false; 1065 next_part: 1066 /* shorten the remaining extent to update progress */ 1067 map->m_llen = offset + cur - map->m_la; 1068 map->m_flags &= ~EROFS_MAP_FULL_MAPPED; 1069 1070 end = cur; 1071 if (end > 0) 1072 goto repeat; 1073 1074 out: 1075 if (err) 1076 z_erofs_page_mark_eio(page); 1077 z_erofs_onlinepage_endio(page); 1078 return err; 1079 } 1080 1081 static bool z_erofs_is_sync_decompress(struct erofs_sb_info *sbi, 1082 unsigned int readahead_pages) 1083 { 1084 /* auto: enable for read_folio, disable for readahead */ 1085 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) && 1086 !readahead_pages) 1087 return true; 1088 1089 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) && 1090 (readahead_pages <= sbi->opt.max_sync_decompress_pages)) 1091 return true; 1092 1093 return false; 1094 } 1095 1096 static bool z_erofs_page_is_invalidated(struct page *page) 1097 { 1098 return !page->mapping && !z_erofs_is_shortlived_page(page); 1099 } 1100 1101 struct z_erofs_decompress_backend { 1102 struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES]; 1103 struct super_block *sb; 1104 struct z_erofs_pcluster *pcl; 1105 1106 /* pages with the longest decompressed length for deduplication */ 1107 struct page **decompressed_pages; 1108 /* pages to keep the compressed data */ 1109 struct page **compressed_pages; 1110 1111 struct list_head decompressed_secondary_bvecs; 1112 struct page **pagepool; 1113 unsigned int onstack_used, nr_pages; 1114 }; 1115 1116 struct z_erofs_bvec_item { 1117 struct z_erofs_bvec bvec; 1118 struct list_head list; 1119 }; 1120 1121 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be, 1122 struct z_erofs_bvec *bvec) 1123 { 1124 struct z_erofs_bvec_item *item; 1125 unsigned int pgnr; 1126 1127 if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK) && 1128 (bvec->end == PAGE_SIZE || 1129 bvec->offset + bvec->end == be->pcl->length)) { 1130 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT; 1131 DBG_BUGON(pgnr >= be->nr_pages); 1132 if (!be->decompressed_pages[pgnr]) { 1133 be->decompressed_pages[pgnr] = bvec->page; 1134 return; 1135 } 1136 } 1137 1138 /* (cold path) one pcluster is requested multiple times */ 1139 item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL); 1140 item->bvec = *bvec; 1141 list_add(&item->list, &be->decompressed_secondary_bvecs); 1142 } 1143 1144 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be, 1145 int err) 1146 { 1147 unsigned int off0 = be->pcl->pageofs_out; 1148 struct list_head *p, *n; 1149 1150 list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) { 1151 struct z_erofs_bvec_item *bvi; 1152 unsigned int end, cur; 1153 void *dst, *src; 1154 1155 bvi = container_of(p, struct z_erofs_bvec_item, list); 1156 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0; 1157 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset, 1158 bvi->bvec.end); 1159 dst = kmap_local_page(bvi->bvec.page); 1160 while (cur < end) { 1161 unsigned int pgnr, scur, len; 1162 1163 pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT; 1164 DBG_BUGON(pgnr >= be->nr_pages); 1165 1166 scur = bvi->bvec.offset + cur - 1167 ((pgnr << PAGE_SHIFT) - off0); 1168 len = min_t(unsigned int, end - cur, PAGE_SIZE - scur); 1169 if (!be->decompressed_pages[pgnr]) { 1170 err = -EFSCORRUPTED; 1171 cur += len; 1172 continue; 1173 } 1174 src = kmap_local_page(be->decompressed_pages[pgnr]); 1175 memcpy(dst + cur, src + scur, len); 1176 kunmap_local(src); 1177 cur += len; 1178 } 1179 kunmap_local(dst); 1180 if (err) 1181 z_erofs_page_mark_eio(bvi->bvec.page); 1182 z_erofs_onlinepage_endio(bvi->bvec.page); 1183 list_del(p); 1184 kfree(bvi); 1185 } 1186 } 1187 1188 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be) 1189 { 1190 struct z_erofs_pcluster *pcl = be->pcl; 1191 struct z_erofs_bvec_iter biter; 1192 struct page *old_bvpage; 1193 int i; 1194 1195 z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0); 1196 for (i = 0; i < pcl->vcnt; ++i) { 1197 struct z_erofs_bvec bvec; 1198 1199 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage); 1200 1201 if (old_bvpage) 1202 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1203 1204 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page)); 1205 z_erofs_do_decompressed_bvec(be, &bvec); 1206 } 1207 1208 old_bvpage = z_erofs_bvec_iter_end(&biter); 1209 if (old_bvpage) 1210 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage); 1211 } 1212 1213 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be, 1214 bool *overlapped) 1215 { 1216 struct z_erofs_pcluster *pcl = be->pcl; 1217 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1218 int i, err = 0; 1219 1220 *overlapped = false; 1221 for (i = 0; i < pclusterpages; ++i) { 1222 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i]; 1223 struct page *page = bvec->page; 1224 1225 /* compressed pages ought to be present before decompressing */ 1226 if (!page) { 1227 DBG_BUGON(1); 1228 continue; 1229 } 1230 be->compressed_pages[i] = page; 1231 1232 if (z_erofs_is_inline_pcluster(pcl)) { 1233 if (!PageUptodate(page)) 1234 err = -EIO; 1235 continue; 1236 } 1237 1238 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1239 if (!z_erofs_is_shortlived_page(page)) { 1240 if (erofs_page_is_managed(EROFS_SB(be->sb), page)) { 1241 if (!PageUptodate(page)) 1242 err = -EIO; 1243 continue; 1244 } 1245 z_erofs_do_decompressed_bvec(be, bvec); 1246 *overlapped = true; 1247 } 1248 } 1249 1250 if (err) 1251 return err; 1252 return 0; 1253 } 1254 1255 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be, 1256 int err) 1257 { 1258 struct erofs_sb_info *const sbi = EROFS_SB(be->sb); 1259 struct z_erofs_pcluster *pcl = be->pcl; 1260 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 1261 const struct z_erofs_decompressor *decompressor = 1262 &erofs_decompressors[pcl->algorithmformat]; 1263 unsigned int i, inputsize; 1264 int err2; 1265 struct page *page; 1266 bool overlapped; 1267 1268 mutex_lock(&pcl->lock); 1269 be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT; 1270 1271 /* allocate (de)compressed page arrays if cannot be kept on stack */ 1272 be->decompressed_pages = NULL; 1273 be->compressed_pages = NULL; 1274 be->onstack_used = 0; 1275 if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) { 1276 be->decompressed_pages = be->onstack_pages; 1277 be->onstack_used = be->nr_pages; 1278 memset(be->decompressed_pages, 0, 1279 sizeof(struct page *) * be->nr_pages); 1280 } 1281 1282 if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES) 1283 be->compressed_pages = be->onstack_pages + be->onstack_used; 1284 1285 if (!be->decompressed_pages) 1286 be->decompressed_pages = 1287 kvcalloc(be->nr_pages, sizeof(struct page *), 1288 GFP_KERNEL | __GFP_NOFAIL); 1289 if (!be->compressed_pages) 1290 be->compressed_pages = 1291 kvcalloc(pclusterpages, sizeof(struct page *), 1292 GFP_KERNEL | __GFP_NOFAIL); 1293 1294 z_erofs_parse_out_bvecs(be); 1295 err2 = z_erofs_parse_in_bvecs(be, &overlapped); 1296 if (err2) 1297 err = err2; 1298 if (err) 1299 goto out; 1300 1301 if (z_erofs_is_inline_pcluster(pcl)) 1302 inputsize = pcl->tailpacking_size; 1303 else 1304 inputsize = pclusterpages * PAGE_SIZE; 1305 1306 err = decompressor->decompress(&(struct z_erofs_decompress_req) { 1307 .sb = be->sb, 1308 .in = be->compressed_pages, 1309 .out = be->decompressed_pages, 1310 .pageofs_in = pcl->pageofs_in, 1311 .pageofs_out = pcl->pageofs_out, 1312 .inputsize = inputsize, 1313 .outputsize = pcl->length, 1314 .alg = pcl->algorithmformat, 1315 .inplace_io = overlapped, 1316 .partial_decoding = pcl->partial, 1317 .fillgaps = pcl->multibases, 1318 }, be->pagepool); 1319 1320 out: 1321 /* must handle all compressed pages before actual file pages */ 1322 if (z_erofs_is_inline_pcluster(pcl)) { 1323 page = pcl->compressed_bvecs[0].page; 1324 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL); 1325 put_page(page); 1326 } else { 1327 for (i = 0; i < pclusterpages; ++i) { 1328 page = pcl->compressed_bvecs[i].page; 1329 1330 if (erofs_page_is_managed(sbi, page)) 1331 continue; 1332 1333 /* recycle all individual short-lived pages */ 1334 (void)z_erofs_put_shortlivedpage(be->pagepool, page); 1335 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL); 1336 } 1337 } 1338 if (be->compressed_pages < be->onstack_pages || 1339 be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES) 1340 kvfree(be->compressed_pages); 1341 z_erofs_fill_other_copies(be, err); 1342 1343 for (i = 0; i < be->nr_pages; ++i) { 1344 page = be->decompressed_pages[i]; 1345 if (!page) 1346 continue; 1347 1348 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1349 1350 /* recycle all individual short-lived pages */ 1351 if (z_erofs_put_shortlivedpage(be->pagepool, page)) 1352 continue; 1353 if (err) 1354 z_erofs_page_mark_eio(page); 1355 z_erofs_onlinepage_endio(page); 1356 } 1357 1358 if (be->decompressed_pages != be->onstack_pages) 1359 kvfree(be->decompressed_pages); 1360 1361 pcl->length = 0; 1362 pcl->partial = true; 1363 pcl->multibases = false; 1364 pcl->bvset.nextpage = NULL; 1365 pcl->vcnt = 0; 1366 1367 /* pcluster lock MUST be taken before the following line */ 1368 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 1369 mutex_unlock(&pcl->lock); 1370 return err; 1371 } 1372 1373 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 1374 struct page **pagepool) 1375 { 1376 struct z_erofs_decompress_backend be = { 1377 .sb = io->sb, 1378 .pagepool = pagepool, 1379 .decompressed_secondary_bvecs = 1380 LIST_HEAD_INIT(be.decompressed_secondary_bvecs), 1381 }; 1382 z_erofs_next_pcluster_t owned = io->head; 1383 1384 while (owned != Z_EROFS_PCLUSTER_TAIL) { 1385 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1386 1387 be.pcl = container_of(owned, struct z_erofs_pcluster, next); 1388 owned = READ_ONCE(be.pcl->next); 1389 1390 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0); 1391 erofs_workgroup_put(&be.pcl->obj); 1392 } 1393 } 1394 1395 static void z_erofs_decompressqueue_work(struct work_struct *work) 1396 { 1397 struct z_erofs_decompressqueue *bgq = 1398 container_of(work, struct z_erofs_decompressqueue, u.work); 1399 struct page *pagepool = NULL; 1400 1401 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL); 1402 z_erofs_decompress_queue(bgq, &pagepool); 1403 erofs_release_pages(&pagepool); 1404 kvfree(bgq); 1405 } 1406 1407 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1408 static void z_erofs_decompressqueue_kthread_work(struct kthread_work *work) 1409 { 1410 z_erofs_decompressqueue_work((struct work_struct *)work); 1411 } 1412 #endif 1413 1414 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 1415 int bios) 1416 { 1417 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 1418 1419 /* wake up the caller thread for sync decompression */ 1420 if (io->sync) { 1421 if (!atomic_add_return(bios, &io->pending_bios)) 1422 complete(&io->u.done); 1423 return; 1424 } 1425 1426 if (atomic_add_return(bios, &io->pending_bios)) 1427 return; 1428 /* Use (kthread_)work and sync decompression for atomic contexts only */ 1429 if (!in_task() || irqs_disabled() || rcu_read_lock_any_held()) { 1430 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1431 struct kthread_worker *worker; 1432 1433 rcu_read_lock(); 1434 worker = rcu_dereference( 1435 z_erofs_pcpu_workers[raw_smp_processor_id()]); 1436 if (!worker) { 1437 INIT_WORK(&io->u.work, z_erofs_decompressqueue_work); 1438 queue_work(z_erofs_workqueue, &io->u.work); 1439 } else { 1440 kthread_queue_work(worker, &io->u.kthread_work); 1441 } 1442 rcu_read_unlock(); 1443 #else 1444 queue_work(z_erofs_workqueue, &io->u.work); 1445 #endif 1446 /* enable sync decompression for readahead */ 1447 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) 1448 sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON; 1449 return; 1450 } 1451 z_erofs_decompressqueue_work(&io->u.work); 1452 } 1453 1454 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1455 unsigned int nr, 1456 struct page **pagepool, 1457 struct address_space *mc) 1458 { 1459 const pgoff_t index = pcl->obj.index; 1460 gfp_t gfp = mapping_gfp_mask(mc); 1461 bool tocache = false; 1462 1463 struct address_space *mapping; 1464 struct page *oldpage, *page; 1465 int justfound; 1466 1467 repeat: 1468 page = READ_ONCE(pcl->compressed_bvecs[nr].page); 1469 oldpage = page; 1470 1471 if (!page) 1472 goto out_allocpage; 1473 1474 justfound = (unsigned long)page & 1UL; 1475 page = (struct page *)((unsigned long)page & ~1UL); 1476 1477 /* 1478 * preallocated cached pages, which is used to avoid direct reclaim 1479 * otherwise, it will go inplace I/O path instead. 1480 */ 1481 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1482 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1483 set_page_private(page, 0); 1484 tocache = true; 1485 goto out_tocache; 1486 } 1487 mapping = READ_ONCE(page->mapping); 1488 1489 /* 1490 * file-backed online pages in plcuster are all locked steady, 1491 * therefore it is impossible for `mapping' to be NULL. 1492 */ 1493 if (mapping && mapping != mc) 1494 /* ought to be unmanaged pages */ 1495 goto out; 1496 1497 /* directly return for shortlived page as well */ 1498 if (z_erofs_is_shortlived_page(page)) 1499 goto out; 1500 1501 lock_page(page); 1502 1503 /* only true if page reclaim goes wrong, should never happen */ 1504 DBG_BUGON(justfound && PagePrivate(page)); 1505 1506 /* the page is still in manage cache */ 1507 if (page->mapping == mc) { 1508 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page); 1509 1510 if (!PagePrivate(page)) { 1511 /* 1512 * impossible to be !PagePrivate(page) for 1513 * the current restriction as well if 1514 * the page is already in compressed_bvecs[]. 1515 */ 1516 DBG_BUGON(!justfound); 1517 1518 justfound = 0; 1519 set_page_private(page, (unsigned long)pcl); 1520 SetPagePrivate(page); 1521 } 1522 1523 /* no need to submit io if it is already up-to-date */ 1524 if (PageUptodate(page)) { 1525 unlock_page(page); 1526 page = NULL; 1527 } 1528 goto out; 1529 } 1530 1531 /* 1532 * the managed page has been truncated, it's unsafe to 1533 * reuse this one, let's allocate a new cache-managed page. 1534 */ 1535 DBG_BUGON(page->mapping); 1536 DBG_BUGON(!justfound); 1537 1538 tocache = true; 1539 unlock_page(page); 1540 put_page(page); 1541 out_allocpage: 1542 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1543 if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page, 1544 oldpage, page)) { 1545 erofs_pagepool_add(pagepool, page); 1546 cond_resched(); 1547 goto repeat; 1548 } 1549 out_tocache: 1550 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1551 /* turn into temporary page if fails (1 ref) */ 1552 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1553 goto out; 1554 } 1555 attach_page_private(page, pcl); 1556 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1557 put_page(page); 1558 1559 out: /* the only exit (for tracing and debugging) */ 1560 return page; 1561 } 1562 1563 static struct z_erofs_decompressqueue *jobqueue_init(struct super_block *sb, 1564 struct z_erofs_decompressqueue *fgq, bool *fg) 1565 { 1566 struct z_erofs_decompressqueue *q; 1567 1568 if (fg && !*fg) { 1569 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1570 if (!q) { 1571 *fg = true; 1572 goto fg_out; 1573 } 1574 #ifdef CONFIG_EROFS_FS_PCPU_KTHREAD 1575 kthread_init_work(&q->u.kthread_work, 1576 z_erofs_decompressqueue_kthread_work); 1577 #else 1578 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1579 #endif 1580 } else { 1581 fg_out: 1582 q = fgq; 1583 init_completion(&fgq->u.done); 1584 atomic_set(&fgq->pending_bios, 0); 1585 q->eio = false; 1586 q->sync = true; 1587 } 1588 q->sb = sb; 1589 q->head = Z_EROFS_PCLUSTER_TAIL; 1590 return q; 1591 } 1592 1593 /* define decompression jobqueue types */ 1594 enum { 1595 JQ_BYPASS, 1596 JQ_SUBMIT, 1597 NR_JOBQUEUES, 1598 }; 1599 1600 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1601 z_erofs_next_pcluster_t qtail[], 1602 z_erofs_next_pcluster_t owned_head) 1603 { 1604 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1605 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1606 1607 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL); 1608 1609 WRITE_ONCE(*submit_qtail, owned_head); 1610 WRITE_ONCE(*bypass_qtail, &pcl->next); 1611 1612 qtail[JQ_BYPASS] = &pcl->next; 1613 } 1614 1615 static void z_erofs_decompressqueue_endio(struct bio *bio) 1616 { 1617 struct z_erofs_decompressqueue *q = bio->bi_private; 1618 blk_status_t err = bio->bi_status; 1619 struct bio_vec *bvec; 1620 struct bvec_iter_all iter_all; 1621 1622 bio_for_each_segment_all(bvec, bio, iter_all) { 1623 struct page *page = bvec->bv_page; 1624 1625 DBG_BUGON(PageUptodate(page)); 1626 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1627 1628 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 1629 if (!err) 1630 SetPageUptodate(page); 1631 unlock_page(page); 1632 } 1633 } 1634 if (err) 1635 q->eio = true; 1636 z_erofs_decompress_kickoff(q, -1); 1637 bio_put(bio); 1638 } 1639 1640 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f, 1641 struct z_erofs_decompressqueue *fgq, 1642 bool *force_fg, bool readahead) 1643 { 1644 struct super_block *sb = f->inode->i_sb; 1645 struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb)); 1646 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1647 struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 1648 z_erofs_next_pcluster_t owned_head = f->owned_head; 1649 /* bio is NULL initially, so no need to initialize last_{index,bdev} */ 1650 pgoff_t last_index; 1651 struct block_device *last_bdev; 1652 unsigned int nr_bios = 0; 1653 struct bio *bio = NULL; 1654 unsigned long pflags; 1655 int memstall = 0; 1656 1657 /* 1658 * if managed cache is enabled, bypass jobqueue is needed, 1659 * no need to read from device for all pclusters in this queue. 1660 */ 1661 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1662 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, force_fg); 1663 1664 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1665 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1666 1667 /* by default, all need io submission */ 1668 q[JQ_SUBMIT]->head = owned_head; 1669 1670 do { 1671 struct erofs_map_dev mdev; 1672 struct z_erofs_pcluster *pcl; 1673 pgoff_t cur, end; 1674 unsigned int i = 0; 1675 bool bypass = true; 1676 1677 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1678 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1679 owned_head = READ_ONCE(pcl->next); 1680 1681 if (z_erofs_is_inline_pcluster(pcl)) { 1682 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1683 continue; 1684 } 1685 1686 /* no device id here, thus it will always succeed */ 1687 mdev = (struct erofs_map_dev) { 1688 .m_pa = erofs_pos(sb, pcl->obj.index), 1689 }; 1690 (void)erofs_map_dev(sb, &mdev); 1691 1692 cur = erofs_blknr(sb, mdev.m_pa); 1693 end = cur + pcl->pclusterpages; 1694 1695 do { 1696 struct page *page; 1697 1698 page = pickup_page_for_submission(pcl, i++, 1699 &f->pagepool, mc); 1700 if (!page) 1701 continue; 1702 1703 if (bio && (cur != last_index + 1 || 1704 last_bdev != mdev.m_bdev)) { 1705 submit_bio_retry: 1706 submit_bio(bio); 1707 if (memstall) { 1708 psi_memstall_leave(&pflags); 1709 memstall = 0; 1710 } 1711 bio = NULL; 1712 } 1713 1714 if (unlikely(PageWorkingset(page)) && !memstall) { 1715 psi_memstall_enter(&pflags); 1716 memstall = 1; 1717 } 1718 1719 if (!bio) { 1720 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS, 1721 REQ_OP_READ, GFP_NOIO); 1722 bio->bi_end_io = z_erofs_decompressqueue_endio; 1723 1724 last_bdev = mdev.m_bdev; 1725 bio->bi_iter.bi_sector = (sector_t)cur << 1726 (sb->s_blocksize_bits - 9); 1727 bio->bi_private = q[JQ_SUBMIT]; 1728 if (readahead) 1729 bio->bi_opf |= REQ_RAHEAD; 1730 ++nr_bios; 1731 } 1732 1733 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1734 goto submit_bio_retry; 1735 1736 last_index = cur; 1737 bypass = false; 1738 } while (++cur < end); 1739 1740 if (!bypass) 1741 qtail[JQ_SUBMIT] = &pcl->next; 1742 else 1743 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1744 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1745 1746 if (bio) { 1747 submit_bio(bio); 1748 if (memstall) 1749 psi_memstall_leave(&pflags); 1750 } 1751 1752 /* 1753 * although background is preferred, no one is pending for submission. 1754 * don't issue decompression but drop it directly instead. 1755 */ 1756 if (!*force_fg && !nr_bios) { 1757 kvfree(q[JQ_SUBMIT]); 1758 return; 1759 } 1760 z_erofs_decompress_kickoff(q[JQ_SUBMIT], nr_bios); 1761 } 1762 1763 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f, 1764 bool force_fg, bool ra) 1765 { 1766 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1767 1768 if (f->owned_head == Z_EROFS_PCLUSTER_TAIL) 1769 return; 1770 z_erofs_submit_queue(f, io, &force_fg, ra); 1771 1772 /* handle bypass queue (no i/o pclusters) immediately */ 1773 z_erofs_decompress_queue(&io[JQ_BYPASS], &f->pagepool); 1774 1775 if (!force_fg) 1776 return; 1777 1778 /* wait until all bios are completed */ 1779 wait_for_completion_io(&io[JQ_SUBMIT].u.done); 1780 1781 /* handle synchronous decompress queue in the caller context */ 1782 z_erofs_decompress_queue(&io[JQ_SUBMIT], &f->pagepool); 1783 } 1784 1785 /* 1786 * Since partial uptodate is still unimplemented for now, we have to use 1787 * approximate readmore strategies as a start. 1788 */ 1789 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, 1790 struct readahead_control *rac, bool backmost) 1791 { 1792 struct inode *inode = f->inode; 1793 struct erofs_map_blocks *map = &f->map; 1794 erofs_off_t cur, end, headoffset = f->headoffset; 1795 int err; 1796 1797 if (backmost) { 1798 if (rac) 1799 end = headoffset + readahead_length(rac) - 1; 1800 else 1801 end = headoffset + PAGE_SIZE - 1; 1802 map->m_la = end; 1803 err = z_erofs_map_blocks_iter(inode, map, 1804 EROFS_GET_BLOCKS_READMORE); 1805 if (err) 1806 return; 1807 1808 /* expand ra for the trailing edge if readahead */ 1809 if (rac) { 1810 cur = round_up(map->m_la + map->m_llen, PAGE_SIZE); 1811 readahead_expand(rac, headoffset, cur - headoffset); 1812 return; 1813 } 1814 end = round_up(end, PAGE_SIZE); 1815 } else { 1816 end = round_up(map->m_la, PAGE_SIZE); 1817 1818 if (!map->m_llen) 1819 return; 1820 } 1821 1822 cur = map->m_la + map->m_llen - 1; 1823 while ((cur >= end) && (cur < i_size_read(inode))) { 1824 pgoff_t index = cur >> PAGE_SHIFT; 1825 struct page *page; 1826 1827 page = erofs_grab_cache_page_nowait(inode->i_mapping, index); 1828 if (page) { 1829 if (PageUptodate(page)) 1830 unlock_page(page); 1831 else 1832 (void)z_erofs_do_read_page(f, page); 1833 put_page(page); 1834 } 1835 1836 if (cur < PAGE_SIZE) 1837 break; 1838 cur = (index << PAGE_SHIFT) - 1; 1839 } 1840 } 1841 1842 static int z_erofs_read_folio(struct file *file, struct folio *folio) 1843 { 1844 struct page *page = &folio->page; 1845 struct inode *const inode = page->mapping->host; 1846 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1847 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1848 int err; 1849 1850 trace_erofs_readpage(page, false); 1851 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1852 1853 z_erofs_pcluster_readmore(&f, NULL, true); 1854 err = z_erofs_do_read_page(&f, page); 1855 z_erofs_pcluster_readmore(&f, NULL, false); 1856 z_erofs_pcluster_end(&f); 1857 1858 /* if some compressed cluster ready, need submit them anyway */ 1859 z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, 0), false); 1860 1861 if (err && err != -EINTR) 1862 erofs_err(inode->i_sb, "read error %d @ %lu of nid %llu", 1863 err, folio->index, EROFS_I(inode)->nid); 1864 1865 erofs_put_metabuf(&f.map.buf); 1866 erofs_release_pages(&f.pagepool); 1867 return err; 1868 } 1869 1870 static void z_erofs_readahead(struct readahead_control *rac) 1871 { 1872 struct inode *const inode = rac->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 struct page *head = NULL, *page; 1876 unsigned int nr_pages; 1877 1878 f.headoffset = readahead_pos(rac); 1879 1880 z_erofs_pcluster_readmore(&f, rac, true); 1881 nr_pages = readahead_count(rac); 1882 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1883 1884 while ((page = readahead_page(rac))) { 1885 set_page_private(page, (unsigned long)head); 1886 head = page; 1887 } 1888 1889 while (head) { 1890 struct page *page = head; 1891 int err; 1892 1893 /* traversal in reverse order */ 1894 head = (void *)page_private(page); 1895 1896 err = z_erofs_do_read_page(&f, page); 1897 if (err && err != -EINTR) 1898 erofs_err(inode->i_sb, "readahead error %d @ %lu of nid %llu", 1899 err, page->index, EROFS_I(inode)->nid); 1900 put_page(page); 1901 } 1902 z_erofs_pcluster_readmore(&f, rac, false); 1903 z_erofs_pcluster_end(&f); 1904 1905 z_erofs_runqueue(&f, z_erofs_is_sync_decompress(sbi, nr_pages), true); 1906 erofs_put_metabuf(&f.map.buf); 1907 erofs_release_pages(&f.pagepool); 1908 } 1909 1910 const struct address_space_operations z_erofs_aops = { 1911 .read_folio = z_erofs_read_folio, 1912 .readahead = z_erofs_readahead, 1913 }; 1914