1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2018 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 */ 6 #include "zdata.h" 7 #include "compress.h" 8 #include <linux/prefetch.h> 9 10 #include <trace/events/erofs.h> 11 12 /* 13 * since pclustersize is variable for big pcluster feature, introduce slab 14 * pools implementation for different pcluster sizes. 15 */ 16 struct z_erofs_pcluster_slab { 17 struct kmem_cache *slab; 18 unsigned int maxpages; 19 char name[48]; 20 }; 21 22 #define _PCLP(n) { .maxpages = n } 23 24 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = { 25 _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128), 26 _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES) 27 }; 28 29 static void z_erofs_destroy_pcluster_pool(void) 30 { 31 int i; 32 33 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 34 if (!pcluster_pool[i].slab) 35 continue; 36 kmem_cache_destroy(pcluster_pool[i].slab); 37 pcluster_pool[i].slab = NULL; 38 } 39 } 40 41 static int z_erofs_create_pcluster_pool(void) 42 { 43 struct z_erofs_pcluster_slab *pcs; 44 struct z_erofs_pcluster *a; 45 unsigned int size; 46 47 for (pcs = pcluster_pool; 48 pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) { 49 size = struct_size(a, compressed_pages, pcs->maxpages); 50 51 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages); 52 pcs->slab = kmem_cache_create(pcs->name, size, 0, 53 SLAB_RECLAIM_ACCOUNT, NULL); 54 if (pcs->slab) 55 continue; 56 57 z_erofs_destroy_pcluster_pool(); 58 return -ENOMEM; 59 } 60 return 0; 61 } 62 63 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages) 64 { 65 int i; 66 67 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 68 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 69 struct z_erofs_pcluster *pcl; 70 71 if (nrpages > pcs->maxpages) 72 continue; 73 74 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS); 75 if (!pcl) 76 return ERR_PTR(-ENOMEM); 77 pcl->pclusterpages = nrpages; 78 return pcl; 79 } 80 return ERR_PTR(-EINVAL); 81 } 82 83 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl) 84 { 85 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 86 int i; 87 88 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 89 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 90 91 if (pclusterpages > pcs->maxpages) 92 continue; 93 94 kmem_cache_free(pcs->slab, pcl); 95 return; 96 } 97 DBG_BUGON(1); 98 } 99 100 /* how to allocate cached pages for a pcluster */ 101 enum z_erofs_cache_alloctype { 102 DONTALLOC, /* don't allocate any cached pages */ 103 /* 104 * try to use cached I/O if page allocation succeeds or fallback 105 * to in-place I/O instead to avoid any direct reclaim. 106 */ 107 TRYALLOC, 108 }; 109 110 /* 111 * tagged pointer with 1-bit tag for all compressed pages 112 * tag 0 - the page is just found with an extra page reference 113 */ 114 typedef tagptr1_t compressed_page_t; 115 116 #define tag_compressed_page_justfound(page) \ 117 tagptr_fold(compressed_page_t, page, 1) 118 119 static struct workqueue_struct *z_erofs_workqueue __read_mostly; 120 121 void z_erofs_exit_zip_subsystem(void) 122 { 123 destroy_workqueue(z_erofs_workqueue); 124 z_erofs_destroy_pcluster_pool(); 125 } 126 127 static inline int z_erofs_init_workqueue(void) 128 { 129 const unsigned int onlinecpus = num_possible_cpus(); 130 131 /* 132 * no need to spawn too many threads, limiting threads could minimum 133 * scheduling overhead, perhaps per-CPU threads should be better? 134 */ 135 z_erofs_workqueue = alloc_workqueue("erofs_unzipd", 136 WQ_UNBOUND | WQ_HIGHPRI, 137 onlinecpus + onlinecpus / 4); 138 return z_erofs_workqueue ? 0 : -ENOMEM; 139 } 140 141 int __init z_erofs_init_zip_subsystem(void) 142 { 143 int err = z_erofs_create_pcluster_pool(); 144 145 if (err) 146 return err; 147 err = z_erofs_init_workqueue(); 148 if (err) 149 z_erofs_destroy_pcluster_pool(); 150 return err; 151 } 152 153 enum z_erofs_collectmode { 154 COLLECT_SECONDARY, 155 COLLECT_PRIMARY, 156 /* 157 * The current collection was the tail of an exist chain, in addition 158 * that the previous processed chained collections are all decided to 159 * be hooked up to it. 160 * A new chain will be created for the remaining collections which are 161 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED, 162 * the next collection cannot reuse the whole page safely in 163 * the following scenario: 164 * ________________________________________________________________ 165 * | tail (partial) page | head (partial) page | 166 * | (belongs to the next cl) | (belongs to the current cl) | 167 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________| 168 */ 169 COLLECT_PRIMARY_HOOKED, 170 /* 171 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it 172 * could be dispatched into bypass queue later due to uptodated managed 173 * pages. All related online pages cannot be reused for inplace I/O (or 174 * pagevec) since it can be directly decoded without I/O submission. 175 */ 176 COLLECT_PRIMARY_FOLLOWED_NOINPLACE, 177 /* 178 * The current collection has been linked with the owned chain, and 179 * could also be linked with the remaining collections, which means 180 * if the processing page is the tail page of the collection, thus 181 * the current collection can safely use the whole page (since 182 * the previous collection is under control) for in-place I/O, as 183 * illustrated below: 184 * ________________________________________________________________ 185 * | tail (partial) page | head (partial) page | 186 * | (of the current cl) | (of the previous collection) | 187 * | PRIMARY_FOLLOWED or | | 188 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________| 189 * 190 * [ (*) the above page can be used as inplace I/O. ] 191 */ 192 COLLECT_PRIMARY_FOLLOWED, 193 }; 194 195 struct z_erofs_decompress_frontend { 196 struct inode *const inode; 197 struct erofs_map_blocks map; 198 199 struct z_erofs_pagevec_ctor vector; 200 201 struct z_erofs_pcluster *pcl, *tailpcl; 202 /* a pointer used to pick up inplace I/O pages */ 203 struct page **icpage_ptr; 204 z_erofs_next_pcluster_t owned_head; 205 206 enum z_erofs_collectmode mode; 207 208 bool readahead; 209 /* used for applying cache strategy on the fly */ 210 bool backmost; 211 erofs_off_t headoffset; 212 }; 213 214 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 215 .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 216 .mode = COLLECT_PRIMARY_FOLLOWED, .backmost = true } 217 218 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES]; 219 static DEFINE_MUTEX(z_pagemap_global_lock); 220 221 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe, 222 enum z_erofs_cache_alloctype type, 223 struct page **pagepool) 224 { 225 struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode)); 226 struct z_erofs_pcluster *pcl = fe->pcl; 227 bool standalone = true; 228 /* 229 * optimistic allocation without direct reclaim since inplace I/O 230 * can be used if low memory otherwise. 231 */ 232 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 233 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 234 struct page **pages; 235 pgoff_t index; 236 237 if (fe->mode < COLLECT_PRIMARY_FOLLOWED) 238 return; 239 240 pages = pcl->compressed_pages; 241 index = pcl->obj.index; 242 for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) { 243 struct page *page; 244 compressed_page_t t; 245 struct page *newpage = NULL; 246 247 /* the compressed page was loaded before */ 248 if (READ_ONCE(*pages)) 249 continue; 250 251 page = find_get_page(mc, index); 252 253 if (page) { 254 t = tag_compressed_page_justfound(page); 255 } else { 256 /* I/O is needed, no possible to decompress directly */ 257 standalone = false; 258 switch (type) { 259 case TRYALLOC: 260 newpage = erofs_allocpage(pagepool, gfp); 261 if (!newpage) 262 continue; 263 set_page_private(newpage, 264 Z_EROFS_PREALLOCATED_PAGE); 265 t = tag_compressed_page_justfound(newpage); 266 break; 267 default: /* DONTALLOC */ 268 continue; 269 } 270 } 271 272 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t))) 273 continue; 274 275 if (page) 276 put_page(page); 277 else if (newpage) 278 erofs_pagepool_add(pagepool, newpage); 279 } 280 281 /* 282 * don't do inplace I/O if all compressed pages are available in 283 * managed cache since it can be moved to the bypass queue instead. 284 */ 285 if (standalone) 286 fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 287 } 288 289 /* called by erofs_shrinker to get rid of all compressed_pages */ 290 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 291 struct erofs_workgroup *grp) 292 { 293 struct z_erofs_pcluster *const pcl = 294 container_of(grp, struct z_erofs_pcluster, obj); 295 int i; 296 297 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 298 /* 299 * refcount of workgroup is now freezed as 1, 300 * therefore no need to worry about available decompression users. 301 */ 302 for (i = 0; i < pcl->pclusterpages; ++i) { 303 struct page *page = pcl->compressed_pages[i]; 304 305 if (!page) 306 continue; 307 308 /* block other users from reclaiming or migrating the page */ 309 if (!trylock_page(page)) 310 return -EBUSY; 311 312 if (!erofs_page_is_managed(sbi, page)) 313 continue; 314 315 /* barrier is implied in the following 'unlock_page' */ 316 WRITE_ONCE(pcl->compressed_pages[i], NULL); 317 detach_page_private(page); 318 unlock_page(page); 319 } 320 return 0; 321 } 322 323 int erofs_try_to_free_cached_page(struct page *page) 324 { 325 struct z_erofs_pcluster *const pcl = (void *)page_private(page); 326 int ret = 0; /* 0 - busy */ 327 328 if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) { 329 unsigned int i; 330 331 DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 332 for (i = 0; i < pcl->pclusterpages; ++i) { 333 if (pcl->compressed_pages[i] == page) { 334 WRITE_ONCE(pcl->compressed_pages[i], NULL); 335 ret = 1; 336 break; 337 } 338 } 339 erofs_workgroup_unfreeze(&pcl->obj, 1); 340 341 if (ret) 342 detach_page_private(page); 343 } 344 return ret; 345 } 346 347 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */ 348 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe, 349 struct page *page) 350 { 351 struct z_erofs_pcluster *const pcl = fe->pcl; 352 353 while (fe->icpage_ptr > pcl->compressed_pages) 354 if (!cmpxchg(--fe->icpage_ptr, NULL, page)) 355 return true; 356 return false; 357 } 358 359 /* callers must be with pcluster lock held */ 360 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe, 361 struct page *page, enum z_erofs_page_type type, 362 bool pvec_safereuse) 363 { 364 int ret; 365 366 /* give priority for inplaceio */ 367 if (fe->mode >= COLLECT_PRIMARY && 368 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE && 369 z_erofs_try_inplace_io(fe, page)) 370 return 0; 371 372 ret = z_erofs_pagevec_enqueue(&fe->vector, page, type, 373 pvec_safereuse); 374 fe->pcl->vcnt += (unsigned int)ret; 375 return ret ? 0 : -EAGAIN; 376 } 377 378 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f) 379 { 380 struct z_erofs_pcluster *pcl = f->pcl; 381 z_erofs_next_pcluster_t *owned_head = &f->owned_head; 382 383 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 384 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 385 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 386 *owned_head = &pcl->next; 387 /* so we can attach this pcluster to our submission chain. */ 388 f->mode = COLLECT_PRIMARY_FOLLOWED; 389 return; 390 } 391 392 /* 393 * type 2, link to the end of an existing open chain, be careful 394 * that its submission is controlled by the original attached chain. 395 */ 396 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 397 *owned_head) == Z_EROFS_PCLUSTER_TAIL) { 398 *owned_head = Z_EROFS_PCLUSTER_TAIL; 399 f->mode = COLLECT_PRIMARY_HOOKED; 400 f->tailpcl = NULL; 401 return; 402 } 403 /* type 3, it belongs to a chain, but it isn't the end of the chain */ 404 f->mode = COLLECT_PRIMARY; 405 } 406 407 static int z_erofs_lookup_pcluster(struct z_erofs_decompress_frontend *fe) 408 { 409 struct erofs_map_blocks *map = &fe->map; 410 struct z_erofs_pcluster *pcl = fe->pcl; 411 unsigned int length; 412 413 /* to avoid unexpected loop formed by corrupted images */ 414 if (fe->owned_head == &pcl->next || pcl == fe->tailpcl) { 415 DBG_BUGON(1); 416 return -EFSCORRUPTED; 417 } 418 419 if (pcl->pageofs_out != (map->m_la & ~PAGE_MASK)) { 420 DBG_BUGON(1); 421 return -EFSCORRUPTED; 422 } 423 424 length = READ_ONCE(pcl->length); 425 if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) { 426 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) { 427 DBG_BUGON(1); 428 return -EFSCORRUPTED; 429 } 430 } else { 431 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT; 432 433 if (map->m_flags & EROFS_MAP_FULL_MAPPED) 434 llen |= Z_EROFS_PCLUSTER_FULL_LENGTH; 435 436 while (llen > length && 437 length != cmpxchg_relaxed(&pcl->length, length, llen)) { 438 cpu_relax(); 439 length = READ_ONCE(pcl->length); 440 } 441 } 442 mutex_lock(&pcl->lock); 443 /* used to check tail merging loop due to corrupted images */ 444 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 445 fe->tailpcl = pcl; 446 447 z_erofs_try_to_claim_pcluster(fe); 448 return 0; 449 } 450 451 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe) 452 { 453 struct erofs_map_blocks *map = &fe->map; 454 bool ztailpacking = map->m_flags & EROFS_MAP_META; 455 struct z_erofs_pcluster *pcl; 456 struct erofs_workgroup *grp; 457 int err; 458 459 if (!(map->m_flags & EROFS_MAP_ENCODED)) { 460 DBG_BUGON(1); 461 return -EFSCORRUPTED; 462 } 463 464 /* no available pcluster, let's allocate one */ 465 pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 : 466 map->m_plen >> PAGE_SHIFT); 467 if (IS_ERR(pcl)) 468 return PTR_ERR(pcl); 469 470 atomic_set(&pcl->obj.refcount, 1); 471 pcl->algorithmformat = map->m_algorithmformat; 472 pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) | 473 (map->m_flags & EROFS_MAP_FULL_MAPPED ? 474 Z_EROFS_PCLUSTER_FULL_LENGTH : 0); 475 476 /* new pclusters should be claimed as type 1, primary and followed */ 477 pcl->next = fe->owned_head; 478 pcl->pageofs_out = map->m_la & ~PAGE_MASK; 479 fe->mode = COLLECT_PRIMARY_FOLLOWED; 480 481 /* 482 * lock all primary followed works before visible to others 483 * and mutex_trylock *never* fails for a new pcluster. 484 */ 485 mutex_init(&pcl->lock); 486 DBG_BUGON(!mutex_trylock(&pcl->lock)); 487 488 if (ztailpacking) { 489 pcl->obj.index = 0; /* which indicates ztailpacking */ 490 pcl->pageofs_in = erofs_blkoff(map->m_pa); 491 pcl->tailpacking_size = map->m_plen; 492 } else { 493 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 494 495 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj); 496 if (IS_ERR(grp)) { 497 err = PTR_ERR(grp); 498 goto err_out; 499 } 500 501 if (grp != &pcl->obj) { 502 fe->pcl = container_of(grp, 503 struct z_erofs_pcluster, obj); 504 err = -EEXIST; 505 goto err_out; 506 } 507 } 508 /* used to check tail merging loop due to corrupted images */ 509 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 510 fe->tailpcl = pcl; 511 fe->owned_head = &pcl->next; 512 fe->pcl = pcl; 513 return 0; 514 515 err_out: 516 mutex_unlock(&pcl->lock); 517 z_erofs_free_pcluster(pcl); 518 return err; 519 } 520 521 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe) 522 { 523 struct erofs_map_blocks *map = &fe->map; 524 struct erofs_workgroup *grp; 525 int ret; 526 527 DBG_BUGON(fe->pcl); 528 529 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */ 530 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL); 531 DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 532 533 if (map->m_flags & EROFS_MAP_META) { 534 if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) { 535 DBG_BUGON(1); 536 return -EFSCORRUPTED; 537 } 538 goto tailpacking; 539 } 540 541 grp = erofs_find_workgroup(fe->inode->i_sb, map->m_pa >> PAGE_SHIFT); 542 if (grp) { 543 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj); 544 } else { 545 tailpacking: 546 ret = z_erofs_register_pcluster(fe); 547 if (!ret) 548 goto out; 549 if (ret != -EEXIST) 550 return ret; 551 } 552 553 ret = z_erofs_lookup_pcluster(fe); 554 if (ret) { 555 erofs_workgroup_put(&fe->pcl->obj); 556 return ret; 557 } 558 559 out: 560 z_erofs_pagevec_ctor_init(&fe->vector, Z_EROFS_NR_INLINE_PAGEVECS, 561 fe->pcl->pagevec, fe->pcl->vcnt); 562 /* since file-backed online pages are traversed in reverse order */ 563 fe->icpage_ptr = fe->pcl->compressed_pages + 564 z_erofs_pclusterpages(fe->pcl); 565 return 0; 566 } 567 568 /* 569 * keep in mind that no referenced pclusters will be freed 570 * only after a RCU grace period. 571 */ 572 static void z_erofs_rcu_callback(struct rcu_head *head) 573 { 574 z_erofs_free_pcluster(container_of(head, 575 struct z_erofs_pcluster, rcu)); 576 } 577 578 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 579 { 580 struct z_erofs_pcluster *const pcl = 581 container_of(grp, struct z_erofs_pcluster, obj); 582 583 call_rcu(&pcl->rcu, z_erofs_rcu_callback); 584 } 585 586 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe) 587 { 588 struct z_erofs_pcluster *pcl = fe->pcl; 589 590 if (!pcl) 591 return false; 592 593 z_erofs_pagevec_ctor_exit(&fe->vector, false); 594 mutex_unlock(&pcl->lock); 595 596 /* 597 * if all pending pages are added, don't hold its reference 598 * any longer if the pcluster isn't hosted by ourselves. 599 */ 600 if (fe->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE) 601 erofs_workgroup_put(&pcl->obj); 602 603 fe->pcl = NULL; 604 return true; 605 } 606 607 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe, 608 unsigned int cachestrategy, 609 erofs_off_t la) 610 { 611 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 612 return false; 613 614 if (fe->backmost) 615 return true; 616 617 return cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 618 la < fe->headoffset; 619 } 620 621 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 622 struct page *page, struct page **pagepool) 623 { 624 struct inode *const inode = fe->inode; 625 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 626 struct erofs_map_blocks *const map = &fe->map; 627 const loff_t offset = page_offset(page); 628 bool tight = true; 629 630 enum z_erofs_cache_alloctype cache_strategy; 631 enum z_erofs_page_type page_type; 632 unsigned int cur, end, spiltted, index; 633 int err = 0; 634 635 /* register locked file pages as online pages in pack */ 636 z_erofs_onlinepage_init(page); 637 638 spiltted = 0; 639 end = PAGE_SIZE; 640 repeat: 641 cur = end - 1; 642 643 if (offset + cur < map->m_la || 644 offset + cur >= map->m_la + map->m_llen) { 645 erofs_dbg("out-of-range map @ pos %llu", offset + cur); 646 647 if (z_erofs_collector_end(fe)) 648 fe->backmost = false; 649 map->m_la = offset + cur; 650 map->m_llen = 0; 651 err = z_erofs_map_blocks_iter(inode, map, 0); 652 if (err) 653 goto err_out; 654 } else { 655 if (fe->pcl) 656 goto hitted; 657 /* didn't get a valid pcluster previously (very rare) */ 658 } 659 660 if (!(map->m_flags & EROFS_MAP_MAPPED)) 661 goto hitted; 662 663 err = z_erofs_collector_begin(fe); 664 if (err) 665 goto err_out; 666 667 if (z_erofs_is_inline_pcluster(fe->pcl)) { 668 void *mp; 669 670 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb, 671 erofs_blknr(map->m_pa), EROFS_NO_KMAP); 672 if (IS_ERR(mp)) { 673 err = PTR_ERR(mp); 674 erofs_err(inode->i_sb, 675 "failed to get inline page, err %d", err); 676 goto err_out; 677 } 678 get_page(fe->map.buf.page); 679 WRITE_ONCE(fe->pcl->compressed_pages[0], fe->map.buf.page); 680 fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 681 } else { 682 /* bind cache first when cached decompression is preferred */ 683 if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy, 684 map->m_la)) 685 cache_strategy = TRYALLOC; 686 else 687 cache_strategy = DONTALLOC; 688 689 z_erofs_bind_cache(fe, cache_strategy, pagepool); 690 } 691 hitted: 692 /* 693 * Ensure the current partial page belongs to this submit chain rather 694 * than other concurrent submit chains or the noio(bypass) chain since 695 * those chains are handled asynchronously thus the page cannot be used 696 * for inplace I/O or pagevec (should be processed in strict order.) 697 */ 698 tight &= (fe->mode >= COLLECT_PRIMARY_HOOKED && 699 fe->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE); 700 701 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 702 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 703 zero_user_segment(page, cur, end); 704 goto next_part; 705 } 706 707 /* let's derive page type */ 708 page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD : 709 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 710 (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 711 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED)); 712 713 if (cur) 714 tight &= (fe->mode >= COLLECT_PRIMARY_FOLLOWED); 715 716 retry: 717 err = z_erofs_attach_page(fe, page, page_type, 718 fe->mode >= COLLECT_PRIMARY_FOLLOWED); 719 /* should allocate an additional short-lived page for pagevec */ 720 if (err == -EAGAIN) { 721 struct page *const newpage = 722 alloc_page(GFP_NOFS | __GFP_NOFAIL); 723 724 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE); 725 err = z_erofs_attach_page(fe, newpage, 726 Z_EROFS_PAGE_TYPE_EXCLUSIVE, true); 727 if (!err) 728 goto retry; 729 } 730 731 if (err) 732 goto err_out; 733 734 index = page->index - (map->m_la >> PAGE_SHIFT); 735 736 z_erofs_onlinepage_fixup(page, index, true); 737 738 /* bump up the number of spiltted parts of a page */ 739 ++spiltted; 740 /* also update nr_pages */ 741 fe->pcl->nr_pages = max_t(pgoff_t, fe->pcl->nr_pages, index + 1); 742 next_part: 743 /* can be used for verification */ 744 map->m_llen = offset + cur - map->m_la; 745 746 end = cur; 747 if (end > 0) 748 goto repeat; 749 750 out: 751 z_erofs_onlinepage_endio(page); 752 753 erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 754 __func__, page, spiltted, map->m_llen); 755 return err; 756 757 /* if some error occurred while processing this page */ 758 err_out: 759 SetPageError(page); 760 goto out; 761 } 762 763 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi, 764 unsigned int readahead_pages) 765 { 766 /* auto: enable for read_folio, disable for readahead */ 767 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) && 768 !readahead_pages) 769 return true; 770 771 if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) && 772 (readahead_pages <= sbi->opt.max_sync_decompress_pages)) 773 return true; 774 775 return false; 776 } 777 778 static bool z_erofs_page_is_invalidated(struct page *page) 779 { 780 return !page->mapping && !z_erofs_is_shortlived_page(page); 781 } 782 783 static int z_erofs_decompress_pcluster(struct super_block *sb, 784 struct z_erofs_pcluster *pcl, 785 struct page **pagepool) 786 { 787 struct erofs_sb_info *const sbi = EROFS_SB(sb); 788 unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 789 struct z_erofs_pagevec_ctor ctor; 790 unsigned int i, inputsize, outputsize, llen, nr_pages; 791 struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES]; 792 struct page **pages, **compressed_pages, *page; 793 794 enum z_erofs_page_type page_type; 795 bool overlapped, partial; 796 int err; 797 798 might_sleep(); 799 DBG_BUGON(!READ_ONCE(pcl->nr_pages)); 800 801 mutex_lock(&pcl->lock); 802 nr_pages = pcl->nr_pages; 803 804 if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) { 805 pages = pages_onstack; 806 } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES && 807 mutex_trylock(&z_pagemap_global_lock)) { 808 pages = z_pagemap_global; 809 } else { 810 gfp_t gfp_flags = GFP_KERNEL; 811 812 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES) 813 gfp_flags |= __GFP_NOFAIL; 814 815 pages = kvmalloc_array(nr_pages, sizeof(struct page *), 816 gfp_flags); 817 818 /* fallback to global pagemap for the lowmem scenario */ 819 if (!pages) { 820 mutex_lock(&z_pagemap_global_lock); 821 pages = z_pagemap_global; 822 } 823 } 824 825 for (i = 0; i < nr_pages; ++i) 826 pages[i] = NULL; 827 828 err = 0; 829 z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS, 830 pcl->pagevec, 0); 831 832 for (i = 0; i < pcl->vcnt; ++i) { 833 unsigned int pagenr; 834 835 page = z_erofs_pagevec_dequeue(&ctor, &page_type); 836 837 /* all pages in pagevec ought to be valid */ 838 DBG_BUGON(!page); 839 DBG_BUGON(z_erofs_page_is_invalidated(page)); 840 841 if (z_erofs_put_shortlivedpage(pagepool, page)) 842 continue; 843 844 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD) 845 pagenr = 0; 846 else 847 pagenr = z_erofs_onlinepage_index(page); 848 849 DBG_BUGON(pagenr >= nr_pages); 850 851 /* 852 * currently EROFS doesn't support multiref(dedup), 853 * so here erroring out one multiref page. 854 */ 855 if (pages[pagenr]) { 856 DBG_BUGON(1); 857 SetPageError(pages[pagenr]); 858 z_erofs_onlinepage_endio(pages[pagenr]); 859 err = -EFSCORRUPTED; 860 } 861 pages[pagenr] = page; 862 } 863 z_erofs_pagevec_ctor_exit(&ctor, true); 864 865 overlapped = false; 866 compressed_pages = pcl->compressed_pages; 867 868 for (i = 0; i < pclusterpages; ++i) { 869 unsigned int pagenr; 870 871 page = compressed_pages[i]; 872 /* all compressed pages ought to be valid */ 873 DBG_BUGON(!page); 874 875 if (z_erofs_is_inline_pcluster(pcl)) { 876 if (!PageUptodate(page)) 877 err = -EIO; 878 continue; 879 } 880 881 DBG_BUGON(z_erofs_page_is_invalidated(page)); 882 if (!z_erofs_is_shortlived_page(page)) { 883 if (erofs_page_is_managed(sbi, page)) { 884 if (!PageUptodate(page)) 885 err = -EIO; 886 continue; 887 } 888 889 /* 890 * only if non-head page can be selected 891 * for inplace decompression 892 */ 893 pagenr = z_erofs_onlinepage_index(page); 894 895 DBG_BUGON(pagenr >= nr_pages); 896 if (pages[pagenr]) { 897 DBG_BUGON(1); 898 SetPageError(pages[pagenr]); 899 z_erofs_onlinepage_endio(pages[pagenr]); 900 err = -EFSCORRUPTED; 901 } 902 pages[pagenr] = page; 903 904 overlapped = true; 905 } 906 907 /* PG_error needs checking for all non-managed pages */ 908 if (PageError(page)) { 909 DBG_BUGON(PageUptodate(page)); 910 err = -EIO; 911 } 912 } 913 914 if (err) 915 goto out; 916 917 llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT; 918 if (nr_pages << PAGE_SHIFT >= pcl->pageofs_out + llen) { 919 outputsize = llen; 920 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH); 921 } else { 922 outputsize = (nr_pages << PAGE_SHIFT) - pcl->pageofs_out; 923 partial = true; 924 } 925 926 if (z_erofs_is_inline_pcluster(pcl)) 927 inputsize = pcl->tailpacking_size; 928 else 929 inputsize = pclusterpages * PAGE_SIZE; 930 931 err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 932 .sb = sb, 933 .in = compressed_pages, 934 .out = pages, 935 .pageofs_in = pcl->pageofs_in, 936 .pageofs_out = pcl->pageofs_out, 937 .inputsize = inputsize, 938 .outputsize = outputsize, 939 .alg = pcl->algorithmformat, 940 .inplace_io = overlapped, 941 .partial_decoding = partial 942 }, pagepool); 943 944 out: 945 /* must handle all compressed pages before actual file pages */ 946 if (z_erofs_is_inline_pcluster(pcl)) { 947 page = compressed_pages[0]; 948 WRITE_ONCE(compressed_pages[0], NULL); 949 put_page(page); 950 } else { 951 for (i = 0; i < pclusterpages; ++i) { 952 page = compressed_pages[i]; 953 954 if (erofs_page_is_managed(sbi, page)) 955 continue; 956 957 /* recycle all individual short-lived pages */ 958 (void)z_erofs_put_shortlivedpage(pagepool, page); 959 WRITE_ONCE(compressed_pages[i], NULL); 960 } 961 } 962 963 for (i = 0; i < nr_pages; ++i) { 964 page = pages[i]; 965 if (!page) 966 continue; 967 968 DBG_BUGON(z_erofs_page_is_invalidated(page)); 969 970 /* recycle all individual short-lived pages */ 971 if (z_erofs_put_shortlivedpage(pagepool, page)) 972 continue; 973 974 if (err < 0) 975 SetPageError(page); 976 977 z_erofs_onlinepage_endio(page); 978 } 979 980 if (pages == z_pagemap_global) 981 mutex_unlock(&z_pagemap_global_lock); 982 else if (pages != pages_onstack) 983 kvfree(pages); 984 985 pcl->nr_pages = 0; 986 pcl->vcnt = 0; 987 988 /* pcluster lock MUST be taken before the following line */ 989 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 990 mutex_unlock(&pcl->lock); 991 return err; 992 } 993 994 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 995 struct page **pagepool) 996 { 997 z_erofs_next_pcluster_t owned = io->head; 998 999 while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 1000 struct z_erofs_pcluster *pcl; 1001 1002 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 1003 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 1004 1005 /* no possible that 'owned' equals NULL */ 1006 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1007 1008 pcl = container_of(owned, struct z_erofs_pcluster, next); 1009 owned = READ_ONCE(pcl->next); 1010 1011 z_erofs_decompress_pcluster(io->sb, pcl, pagepool); 1012 erofs_workgroup_put(&pcl->obj); 1013 } 1014 } 1015 1016 static void z_erofs_decompressqueue_work(struct work_struct *work) 1017 { 1018 struct z_erofs_decompressqueue *bgq = 1019 container_of(work, struct z_erofs_decompressqueue, u.work); 1020 struct page *pagepool = NULL; 1021 1022 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1023 z_erofs_decompress_queue(bgq, &pagepool); 1024 1025 erofs_release_pages(&pagepool); 1026 kvfree(bgq); 1027 } 1028 1029 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 1030 bool sync, int bios) 1031 { 1032 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 1033 1034 /* wake up the caller thread for sync decompression */ 1035 if (sync) { 1036 if (!atomic_add_return(bios, &io->pending_bios)) 1037 complete(&io->u.done); 1038 1039 return; 1040 } 1041 1042 if (atomic_add_return(bios, &io->pending_bios)) 1043 return; 1044 /* Use workqueue and sync decompression for atomic contexts only */ 1045 if (in_atomic() || irqs_disabled()) { 1046 queue_work(z_erofs_workqueue, &io->u.work); 1047 /* enable sync decompression for readahead */ 1048 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) 1049 sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON; 1050 return; 1051 } 1052 z_erofs_decompressqueue_work(&io->u.work); 1053 } 1054 1055 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1056 unsigned int nr, 1057 struct page **pagepool, 1058 struct address_space *mc) 1059 { 1060 const pgoff_t index = pcl->obj.index; 1061 gfp_t gfp = mapping_gfp_mask(mc); 1062 bool tocache = false; 1063 1064 struct address_space *mapping; 1065 struct page *oldpage, *page; 1066 1067 compressed_page_t t; 1068 int justfound; 1069 1070 repeat: 1071 page = READ_ONCE(pcl->compressed_pages[nr]); 1072 oldpage = page; 1073 1074 if (!page) 1075 goto out_allocpage; 1076 1077 /* process the target tagged pointer */ 1078 t = tagptr_init(compressed_page_t, page); 1079 justfound = tagptr_unfold_tags(t); 1080 page = tagptr_unfold_ptr(t); 1081 1082 /* 1083 * preallocated cached pages, which is used to avoid direct reclaim 1084 * otherwise, it will go inplace I/O path instead. 1085 */ 1086 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1087 WRITE_ONCE(pcl->compressed_pages[nr], page); 1088 set_page_private(page, 0); 1089 tocache = true; 1090 goto out_tocache; 1091 } 1092 mapping = READ_ONCE(page->mapping); 1093 1094 /* 1095 * file-backed online pages in plcuster are all locked steady, 1096 * therefore it is impossible for `mapping' to be NULL. 1097 */ 1098 if (mapping && mapping != mc) 1099 /* ought to be unmanaged pages */ 1100 goto out; 1101 1102 /* directly return for shortlived page as well */ 1103 if (z_erofs_is_shortlived_page(page)) 1104 goto out; 1105 1106 lock_page(page); 1107 1108 /* only true if page reclaim goes wrong, should never happen */ 1109 DBG_BUGON(justfound && PagePrivate(page)); 1110 1111 /* the page is still in manage cache */ 1112 if (page->mapping == mc) { 1113 WRITE_ONCE(pcl->compressed_pages[nr], page); 1114 1115 ClearPageError(page); 1116 if (!PagePrivate(page)) { 1117 /* 1118 * impossible to be !PagePrivate(page) for 1119 * the current restriction as well if 1120 * the page is already in compressed_pages[]. 1121 */ 1122 DBG_BUGON(!justfound); 1123 1124 justfound = 0; 1125 set_page_private(page, (unsigned long)pcl); 1126 SetPagePrivate(page); 1127 } 1128 1129 /* no need to submit io if it is already up-to-date */ 1130 if (PageUptodate(page)) { 1131 unlock_page(page); 1132 page = NULL; 1133 } 1134 goto out; 1135 } 1136 1137 /* 1138 * the managed page has been truncated, it's unsafe to 1139 * reuse this one, let's allocate a new cache-managed page. 1140 */ 1141 DBG_BUGON(page->mapping); 1142 DBG_BUGON(!justfound); 1143 1144 tocache = true; 1145 unlock_page(page); 1146 put_page(page); 1147 out_allocpage: 1148 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1149 if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) { 1150 erofs_pagepool_add(pagepool, page); 1151 cond_resched(); 1152 goto repeat; 1153 } 1154 out_tocache: 1155 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1156 /* turn into temporary page if fails (1 ref) */ 1157 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1158 goto out; 1159 } 1160 attach_page_private(page, pcl); 1161 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1162 put_page(page); 1163 1164 out: /* the only exit (for tracing and debugging) */ 1165 return page; 1166 } 1167 1168 static struct z_erofs_decompressqueue * 1169 jobqueue_init(struct super_block *sb, 1170 struct z_erofs_decompressqueue *fgq, bool *fg) 1171 { 1172 struct z_erofs_decompressqueue *q; 1173 1174 if (fg && !*fg) { 1175 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1176 if (!q) { 1177 *fg = true; 1178 goto fg_out; 1179 } 1180 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1181 } else { 1182 fg_out: 1183 q = fgq; 1184 init_completion(&fgq->u.done); 1185 atomic_set(&fgq->pending_bios, 0); 1186 } 1187 q->sb = sb; 1188 q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1189 return q; 1190 } 1191 1192 /* define decompression jobqueue types */ 1193 enum { 1194 JQ_BYPASS, 1195 JQ_SUBMIT, 1196 NR_JOBQUEUES, 1197 }; 1198 1199 static void *jobqueueset_init(struct super_block *sb, 1200 struct z_erofs_decompressqueue *q[], 1201 struct z_erofs_decompressqueue *fgq, bool *fg) 1202 { 1203 /* 1204 * if managed cache is enabled, bypass jobqueue is needed, 1205 * no need to read from device for all pclusters in this queue. 1206 */ 1207 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1208 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg); 1209 1210 return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg)); 1211 } 1212 1213 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1214 z_erofs_next_pcluster_t qtail[], 1215 z_erofs_next_pcluster_t owned_head) 1216 { 1217 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1218 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1219 1220 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1221 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1222 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1223 1224 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 1225 1226 WRITE_ONCE(*submit_qtail, owned_head); 1227 WRITE_ONCE(*bypass_qtail, &pcl->next); 1228 1229 qtail[JQ_BYPASS] = &pcl->next; 1230 } 1231 1232 static void z_erofs_decompressqueue_endio(struct bio *bio) 1233 { 1234 tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private); 1235 struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t); 1236 blk_status_t err = bio->bi_status; 1237 struct bio_vec *bvec; 1238 struct bvec_iter_all iter_all; 1239 1240 bio_for_each_segment_all(bvec, bio, iter_all) { 1241 struct page *page = bvec->bv_page; 1242 1243 DBG_BUGON(PageUptodate(page)); 1244 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1245 1246 if (err) 1247 SetPageError(page); 1248 1249 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 1250 if (!err) 1251 SetPageUptodate(page); 1252 unlock_page(page); 1253 } 1254 } 1255 z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1); 1256 bio_put(bio); 1257 } 1258 1259 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f, 1260 struct page **pagepool, 1261 struct z_erofs_decompressqueue *fgq, 1262 bool *force_fg) 1263 { 1264 struct super_block *sb = f->inode->i_sb; 1265 struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb)); 1266 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1267 struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 1268 void *bi_private; 1269 z_erofs_next_pcluster_t owned_head = f->owned_head; 1270 /* bio is NULL initially, so no need to initialize last_{index,bdev} */ 1271 pgoff_t last_index; 1272 struct block_device *last_bdev; 1273 unsigned int nr_bios = 0; 1274 struct bio *bio = NULL; 1275 1276 bi_private = jobqueueset_init(sb, q, fgq, force_fg); 1277 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1278 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1279 1280 /* by default, all need io submission */ 1281 q[JQ_SUBMIT]->head = owned_head; 1282 1283 do { 1284 struct erofs_map_dev mdev; 1285 struct z_erofs_pcluster *pcl; 1286 pgoff_t cur, end; 1287 unsigned int i = 0; 1288 bool bypass = true; 1289 1290 /* no possible 'owned_head' equals the following */ 1291 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1292 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1293 1294 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1295 1296 /* close the main owned chain at first */ 1297 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1298 Z_EROFS_PCLUSTER_TAIL_CLOSED); 1299 if (z_erofs_is_inline_pcluster(pcl)) { 1300 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1301 continue; 1302 } 1303 1304 /* no device id here, thus it will always succeed */ 1305 mdev = (struct erofs_map_dev) { 1306 .m_pa = blknr_to_addr(pcl->obj.index), 1307 }; 1308 (void)erofs_map_dev(sb, &mdev); 1309 1310 cur = erofs_blknr(mdev.m_pa); 1311 end = cur + pcl->pclusterpages; 1312 1313 do { 1314 struct page *page; 1315 1316 page = pickup_page_for_submission(pcl, i++, pagepool, 1317 mc); 1318 if (!page) 1319 continue; 1320 1321 if (bio && (cur != last_index + 1 || 1322 last_bdev != mdev.m_bdev)) { 1323 submit_bio_retry: 1324 submit_bio(bio); 1325 bio = NULL; 1326 } 1327 1328 if (!bio) { 1329 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS, 1330 REQ_OP_READ, GFP_NOIO); 1331 bio->bi_end_io = z_erofs_decompressqueue_endio; 1332 1333 last_bdev = mdev.m_bdev; 1334 bio->bi_iter.bi_sector = (sector_t)cur << 1335 LOG_SECTORS_PER_BLOCK; 1336 bio->bi_private = bi_private; 1337 if (f->readahead) 1338 bio->bi_opf |= REQ_RAHEAD; 1339 ++nr_bios; 1340 } 1341 1342 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1343 goto submit_bio_retry; 1344 1345 last_index = cur; 1346 bypass = false; 1347 } while (++cur < end); 1348 1349 if (!bypass) 1350 qtail[JQ_SUBMIT] = &pcl->next; 1351 else 1352 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1353 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1354 1355 if (bio) 1356 submit_bio(bio); 1357 1358 /* 1359 * although background is preferred, no one is pending for submission. 1360 * don't issue workqueue for decompression but drop it directly instead. 1361 */ 1362 if (!*force_fg && !nr_bios) { 1363 kvfree(q[JQ_SUBMIT]); 1364 return; 1365 } 1366 z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios); 1367 } 1368 1369 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f, 1370 struct page **pagepool, bool force_fg) 1371 { 1372 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1373 1374 if (f->owned_head == Z_EROFS_PCLUSTER_TAIL) 1375 return; 1376 z_erofs_submit_queue(f, pagepool, io, &force_fg); 1377 1378 /* handle bypass queue (no i/o pclusters) immediately */ 1379 z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool); 1380 1381 if (!force_fg) 1382 return; 1383 1384 /* wait until all bios are completed */ 1385 wait_for_completion_io(&io[JQ_SUBMIT].u.done); 1386 1387 /* handle synchronous decompress queue in the caller context */ 1388 z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool); 1389 } 1390 1391 /* 1392 * Since partial uptodate is still unimplemented for now, we have to use 1393 * approximate readmore strategies as a start. 1394 */ 1395 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, 1396 struct readahead_control *rac, 1397 erofs_off_t end, 1398 struct page **pagepool, 1399 bool backmost) 1400 { 1401 struct inode *inode = f->inode; 1402 struct erofs_map_blocks *map = &f->map; 1403 erofs_off_t cur; 1404 int err; 1405 1406 if (backmost) { 1407 map->m_la = end; 1408 err = z_erofs_map_blocks_iter(inode, map, 1409 EROFS_GET_BLOCKS_READMORE); 1410 if (err) 1411 return; 1412 1413 /* expend ra for the trailing edge if readahead */ 1414 if (rac) { 1415 loff_t newstart = readahead_pos(rac); 1416 1417 cur = round_up(map->m_la + map->m_llen, PAGE_SIZE); 1418 readahead_expand(rac, newstart, cur - newstart); 1419 return; 1420 } 1421 end = round_up(end, PAGE_SIZE); 1422 } else { 1423 end = round_up(map->m_la, PAGE_SIZE); 1424 1425 if (!map->m_llen) 1426 return; 1427 } 1428 1429 cur = map->m_la + map->m_llen - 1; 1430 while (cur >= end) { 1431 pgoff_t index = cur >> PAGE_SHIFT; 1432 struct page *page; 1433 1434 page = erofs_grab_cache_page_nowait(inode->i_mapping, index); 1435 if (page) { 1436 if (PageUptodate(page)) { 1437 unlock_page(page); 1438 } else { 1439 err = z_erofs_do_read_page(f, page, pagepool); 1440 if (err) 1441 erofs_err(inode->i_sb, 1442 "readmore error at page %lu @ nid %llu", 1443 index, EROFS_I(inode)->nid); 1444 } 1445 put_page(page); 1446 } 1447 1448 if (cur < PAGE_SIZE) 1449 break; 1450 cur = (index << PAGE_SHIFT) - 1; 1451 } 1452 } 1453 1454 static int z_erofs_read_folio(struct file *file, struct folio *folio) 1455 { 1456 struct page *page = &folio->page; 1457 struct inode *const inode = page->mapping->host; 1458 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1459 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1460 struct page *pagepool = NULL; 1461 int err; 1462 1463 trace_erofs_readpage(page, false); 1464 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1465 1466 z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1, 1467 &pagepool, true); 1468 err = z_erofs_do_read_page(&f, page, &pagepool); 1469 z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false); 1470 1471 (void)z_erofs_collector_end(&f); 1472 1473 /* if some compressed cluster ready, need submit them anyway */ 1474 z_erofs_runqueue(&f, &pagepool, 1475 z_erofs_get_sync_decompress_policy(sbi, 0)); 1476 1477 if (err) 1478 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1479 1480 erofs_put_metabuf(&f.map.buf); 1481 erofs_release_pages(&pagepool); 1482 return err; 1483 } 1484 1485 static void z_erofs_readahead(struct readahead_control *rac) 1486 { 1487 struct inode *const inode = rac->mapping->host; 1488 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1489 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1490 struct page *pagepool = NULL, *head = NULL, *page; 1491 unsigned int nr_pages; 1492 1493 f.readahead = true; 1494 f.headoffset = readahead_pos(rac); 1495 1496 z_erofs_pcluster_readmore(&f, rac, f.headoffset + 1497 readahead_length(rac) - 1, &pagepool, true); 1498 nr_pages = readahead_count(rac); 1499 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1500 1501 while ((page = readahead_page(rac))) { 1502 set_page_private(page, (unsigned long)head); 1503 head = page; 1504 } 1505 1506 while (head) { 1507 struct page *page = head; 1508 int err; 1509 1510 /* traversal in reverse order */ 1511 head = (void *)page_private(page); 1512 1513 err = z_erofs_do_read_page(&f, page, &pagepool); 1514 if (err) 1515 erofs_err(inode->i_sb, 1516 "readahead error at page %lu @ nid %llu", 1517 page->index, EROFS_I(inode)->nid); 1518 put_page(page); 1519 } 1520 z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false); 1521 (void)z_erofs_collector_end(&f); 1522 1523 z_erofs_runqueue(&f, &pagepool, 1524 z_erofs_get_sync_decompress_policy(sbi, nr_pages)); 1525 erofs_put_metabuf(&f.map.buf); 1526 erofs_release_pages(&pagepool); 1527 } 1528 1529 const struct address_space_operations z_erofs_aops = { 1530 .read_folio = z_erofs_read_folio, 1531 .readahead = z_erofs_readahead, 1532 }; 1533