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 int i; 86 87 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 88 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 89 90 if (pcl->pclusterpages > pcs->maxpages) 91 continue; 92 93 kmem_cache_free(pcs->slab, pcl); 94 return; 95 } 96 DBG_BUGON(1); 97 } 98 99 /* 100 * a compressed_pages[] placeholder in order to avoid 101 * being filled with file pages for in-place decompression. 102 */ 103 #define PAGE_UNALLOCATED ((void *)0x5F0E4B1D) 104 105 /* how to allocate cached pages for a pcluster */ 106 enum z_erofs_cache_alloctype { 107 DONTALLOC, /* don't allocate any cached pages */ 108 DELAYEDALLOC, /* delayed allocation (at the time of submitting io) */ 109 /* 110 * try to use cached I/O if page allocation succeeds or fallback 111 * to in-place I/O instead to avoid any direct reclaim. 112 */ 113 TRYALLOC, 114 }; 115 116 /* 117 * tagged pointer with 1-bit tag for all compressed pages 118 * tag 0 - the page is just found with an extra page reference 119 */ 120 typedef tagptr1_t compressed_page_t; 121 122 #define tag_compressed_page_justfound(page) \ 123 tagptr_fold(compressed_page_t, page, 1) 124 125 static struct workqueue_struct *z_erofs_workqueue __read_mostly; 126 127 void z_erofs_exit_zip_subsystem(void) 128 { 129 destroy_workqueue(z_erofs_workqueue); 130 z_erofs_destroy_pcluster_pool(); 131 } 132 133 static inline int z_erofs_init_workqueue(void) 134 { 135 const unsigned int onlinecpus = num_possible_cpus(); 136 137 /* 138 * no need to spawn too many threads, limiting threads could minimum 139 * scheduling overhead, perhaps per-CPU threads should be better? 140 */ 141 z_erofs_workqueue = alloc_workqueue("erofs_unzipd", 142 WQ_UNBOUND | WQ_HIGHPRI, 143 onlinecpus + onlinecpus / 4); 144 return z_erofs_workqueue ? 0 : -ENOMEM; 145 } 146 147 int __init z_erofs_init_zip_subsystem(void) 148 { 149 int err = z_erofs_create_pcluster_pool(); 150 151 if (err) 152 return err; 153 err = z_erofs_init_workqueue(); 154 if (err) 155 z_erofs_destroy_pcluster_pool(); 156 return err; 157 } 158 159 enum z_erofs_collectmode { 160 COLLECT_SECONDARY, 161 COLLECT_PRIMARY, 162 /* 163 * The current collection was the tail of an exist chain, in addition 164 * that the previous processed chained collections are all decided to 165 * be hooked up to it. 166 * A new chain will be created for the remaining collections which are 167 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED, 168 * the next collection cannot reuse the whole page safely in 169 * the following scenario: 170 * ________________________________________________________________ 171 * | tail (partial) page | head (partial) page | 172 * | (belongs to the next cl) | (belongs to the current cl) | 173 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________| 174 */ 175 COLLECT_PRIMARY_HOOKED, 176 /* 177 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it 178 * could be dispatched into bypass queue later due to uptodated managed 179 * pages. All related online pages cannot be reused for inplace I/O (or 180 * pagevec) since it can be directly decoded without I/O submission. 181 */ 182 COLLECT_PRIMARY_FOLLOWED_NOINPLACE, 183 /* 184 * The current collection has been linked with the owned chain, and 185 * could also be linked with the remaining collections, which means 186 * if the processing page is the tail page of the collection, thus 187 * the current collection can safely use the whole page (since 188 * the previous collection is under control) for in-place I/O, as 189 * illustrated below: 190 * ________________________________________________________________ 191 * | tail (partial) page | head (partial) page | 192 * | (of the current cl) | (of the previous collection) | 193 * | PRIMARY_FOLLOWED or | | 194 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________| 195 * 196 * [ (*) the above page can be used as inplace I/O. ] 197 */ 198 COLLECT_PRIMARY_FOLLOWED, 199 }; 200 201 struct z_erofs_collector { 202 struct z_erofs_pagevec_ctor vector; 203 204 struct z_erofs_pcluster *pcl, *tailpcl; 205 struct z_erofs_collection *cl; 206 /* a pointer used to pick up inplace I/O pages */ 207 struct page **icpage_ptr; 208 z_erofs_next_pcluster_t owned_head; 209 210 enum z_erofs_collectmode mode; 211 }; 212 213 struct z_erofs_decompress_frontend { 214 struct inode *const inode; 215 216 struct z_erofs_collector clt; 217 struct erofs_map_blocks map; 218 219 bool readahead; 220 /* used for applying cache strategy on the fly */ 221 bool backmost; 222 erofs_off_t headoffset; 223 }; 224 225 #define COLLECTOR_INIT() { \ 226 .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 227 .mode = COLLECT_PRIMARY_FOLLOWED } 228 229 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 230 .inode = __i, .clt = COLLECTOR_INIT(), \ 231 .backmost = true, } 232 233 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES]; 234 static DEFINE_MUTEX(z_pagemap_global_lock); 235 236 static void preload_compressed_pages(struct z_erofs_collector *clt, 237 struct address_space *mc, 238 enum z_erofs_cache_alloctype type, 239 struct list_head *pagepool) 240 { 241 struct z_erofs_pcluster *pcl = clt->pcl; 242 bool standalone = true; 243 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 244 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 245 struct page **pages; 246 pgoff_t index; 247 248 if (clt->mode < COLLECT_PRIMARY_FOLLOWED) 249 return; 250 251 pages = pcl->compressed_pages; 252 index = pcl->obj.index; 253 for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) { 254 struct page *page; 255 compressed_page_t t; 256 struct page *newpage = NULL; 257 258 /* the compressed page was loaded before */ 259 if (READ_ONCE(*pages)) 260 continue; 261 262 page = find_get_page(mc, index); 263 264 if (page) { 265 t = tag_compressed_page_justfound(page); 266 } else { 267 /* I/O is needed, no possible to decompress directly */ 268 standalone = false; 269 switch (type) { 270 case DELAYEDALLOC: 271 t = tagptr_init(compressed_page_t, 272 PAGE_UNALLOCATED); 273 break; 274 case TRYALLOC: 275 newpage = erofs_allocpage(pagepool, gfp); 276 if (!newpage) 277 continue; 278 set_page_private(newpage, 279 Z_EROFS_PREALLOCATED_PAGE); 280 t = tag_compressed_page_justfound(newpage); 281 break; 282 default: /* DONTALLOC */ 283 continue; 284 } 285 } 286 287 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t))) 288 continue; 289 290 if (page) { 291 put_page(page); 292 } else if (newpage) { 293 set_page_private(newpage, 0); 294 list_add(&newpage->lru, pagepool); 295 } 296 } 297 298 /* 299 * don't do inplace I/O if all compressed pages are available in 300 * managed cache since it can be moved to the bypass queue instead. 301 */ 302 if (standalone) 303 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 304 } 305 306 /* called by erofs_shrinker to get rid of all compressed_pages */ 307 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 308 struct erofs_workgroup *grp) 309 { 310 struct z_erofs_pcluster *const pcl = 311 container_of(grp, struct z_erofs_pcluster, obj); 312 int i; 313 314 /* 315 * refcount of workgroup is now freezed as 1, 316 * therefore no need to worry about available decompression users. 317 */ 318 for (i = 0; i < pcl->pclusterpages; ++i) { 319 struct page *page = pcl->compressed_pages[i]; 320 321 if (!page) 322 continue; 323 324 /* block other users from reclaiming or migrating the page */ 325 if (!trylock_page(page)) 326 return -EBUSY; 327 328 if (!erofs_page_is_managed(sbi, page)) 329 continue; 330 331 /* barrier is implied in the following 'unlock_page' */ 332 WRITE_ONCE(pcl->compressed_pages[i], NULL); 333 detach_page_private(page); 334 unlock_page(page); 335 } 336 return 0; 337 } 338 339 int erofs_try_to_free_cached_page(struct page *page) 340 { 341 struct z_erofs_pcluster *const pcl = (void *)page_private(page); 342 int ret = 0; /* 0 - busy */ 343 344 if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) { 345 unsigned int i; 346 347 for (i = 0; i < pcl->pclusterpages; ++i) { 348 if (pcl->compressed_pages[i] == page) { 349 WRITE_ONCE(pcl->compressed_pages[i], NULL); 350 ret = 1; 351 break; 352 } 353 } 354 erofs_workgroup_unfreeze(&pcl->obj, 1); 355 356 if (ret) 357 detach_page_private(page); 358 } 359 return ret; 360 } 361 362 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */ 363 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt, 364 struct page *page) 365 { 366 struct z_erofs_pcluster *const pcl = clt->pcl; 367 368 while (clt->icpage_ptr > pcl->compressed_pages) 369 if (!cmpxchg(--clt->icpage_ptr, NULL, page)) 370 return true; 371 return false; 372 } 373 374 /* callers must be with collection lock held */ 375 static int z_erofs_attach_page(struct z_erofs_collector *clt, 376 struct page *page, enum z_erofs_page_type type, 377 bool pvec_safereuse) 378 { 379 int ret; 380 381 /* give priority for inplaceio */ 382 if (clt->mode >= COLLECT_PRIMARY && 383 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE && 384 z_erofs_try_inplace_io(clt, page)) 385 return 0; 386 387 ret = z_erofs_pagevec_enqueue(&clt->vector, page, type, 388 pvec_safereuse); 389 clt->cl->vcnt += (unsigned int)ret; 390 return ret ? 0 : -EAGAIN; 391 } 392 393 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt) 394 { 395 struct z_erofs_pcluster *pcl = clt->pcl; 396 z_erofs_next_pcluster_t *owned_head = &clt->owned_head; 397 398 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 399 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 400 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 401 *owned_head = &pcl->next; 402 /* so we can attach this pcluster to our submission chain. */ 403 clt->mode = COLLECT_PRIMARY_FOLLOWED; 404 return; 405 } 406 407 /* 408 * type 2, link to the end of an existing open chain, be careful 409 * that its submission is controlled by the original attached chain. 410 */ 411 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 412 *owned_head) == Z_EROFS_PCLUSTER_TAIL) { 413 *owned_head = Z_EROFS_PCLUSTER_TAIL; 414 clt->mode = COLLECT_PRIMARY_HOOKED; 415 clt->tailpcl = NULL; 416 return; 417 } 418 /* type 3, it belongs to a chain, but it isn't the end of the chain */ 419 clt->mode = COLLECT_PRIMARY; 420 } 421 422 static int z_erofs_lookup_collection(struct z_erofs_collector *clt, 423 struct inode *inode, 424 struct erofs_map_blocks *map) 425 { 426 struct z_erofs_pcluster *pcl = clt->pcl; 427 struct z_erofs_collection *cl; 428 unsigned int length; 429 430 /* to avoid unexpected loop formed by corrupted images */ 431 if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) { 432 DBG_BUGON(1); 433 return -EFSCORRUPTED; 434 } 435 436 cl = z_erofs_primarycollection(pcl); 437 if (cl->pageofs != (map->m_la & ~PAGE_MASK)) { 438 DBG_BUGON(1); 439 return -EFSCORRUPTED; 440 } 441 442 length = READ_ONCE(pcl->length); 443 if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) { 444 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) { 445 DBG_BUGON(1); 446 return -EFSCORRUPTED; 447 } 448 } else { 449 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT; 450 451 if (map->m_flags & EROFS_MAP_FULL_MAPPED) 452 llen |= Z_EROFS_PCLUSTER_FULL_LENGTH; 453 454 while (llen > length && 455 length != cmpxchg_relaxed(&pcl->length, length, llen)) { 456 cpu_relax(); 457 length = READ_ONCE(pcl->length); 458 } 459 } 460 mutex_lock(&cl->lock); 461 /* used to check tail merging loop due to corrupted images */ 462 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 463 clt->tailpcl = pcl; 464 465 z_erofs_try_to_claim_pcluster(clt); 466 clt->cl = cl; 467 return 0; 468 } 469 470 static int z_erofs_register_collection(struct z_erofs_collector *clt, 471 struct inode *inode, 472 struct erofs_map_blocks *map) 473 { 474 struct z_erofs_pcluster *pcl; 475 struct z_erofs_collection *cl; 476 struct erofs_workgroup *grp; 477 int err; 478 479 /* no available pcluster, let's allocate one */ 480 pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT); 481 if (IS_ERR(pcl)) 482 return PTR_ERR(pcl); 483 484 atomic_set(&pcl->obj.refcount, 1); 485 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 486 487 pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) | 488 (map->m_flags & EROFS_MAP_FULL_MAPPED ? 489 Z_EROFS_PCLUSTER_FULL_LENGTH : 0); 490 491 if (map->m_flags & EROFS_MAP_ZIPPED) 492 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4; 493 else 494 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED; 495 496 /* new pclusters should be claimed as type 1, primary and followed */ 497 pcl->next = clt->owned_head; 498 clt->mode = COLLECT_PRIMARY_FOLLOWED; 499 500 cl = z_erofs_primarycollection(pcl); 501 cl->pageofs = map->m_la & ~PAGE_MASK; 502 503 /* 504 * lock all primary followed works before visible to others 505 * and mutex_trylock *never* fails for a new pcluster. 506 */ 507 mutex_init(&cl->lock); 508 DBG_BUGON(!mutex_trylock(&cl->lock)); 509 510 grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj); 511 if (IS_ERR(grp)) { 512 err = PTR_ERR(grp); 513 goto err_out; 514 } 515 516 if (grp != &pcl->obj) { 517 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj); 518 err = -EEXIST; 519 goto err_out; 520 } 521 /* used to check tail merging loop due to corrupted images */ 522 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 523 clt->tailpcl = pcl; 524 clt->owned_head = &pcl->next; 525 clt->pcl = pcl; 526 clt->cl = cl; 527 return 0; 528 529 err_out: 530 mutex_unlock(&cl->lock); 531 z_erofs_free_pcluster(pcl); 532 return err; 533 } 534 535 static int z_erofs_collector_begin(struct z_erofs_collector *clt, 536 struct inode *inode, 537 struct erofs_map_blocks *map) 538 { 539 struct erofs_workgroup *grp; 540 int ret; 541 542 DBG_BUGON(clt->cl); 543 544 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */ 545 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL); 546 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 547 548 if (!PAGE_ALIGNED(map->m_pa)) { 549 DBG_BUGON(1); 550 return -EINVAL; 551 } 552 553 grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT); 554 if (grp) { 555 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj); 556 } else { 557 ret = z_erofs_register_collection(clt, inode, map); 558 559 if (!ret) 560 goto out; 561 if (ret != -EEXIST) 562 return ret; 563 } 564 565 ret = z_erofs_lookup_collection(clt, inode, map); 566 if (ret) { 567 erofs_workgroup_put(&clt->pcl->obj); 568 return ret; 569 } 570 571 out: 572 z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS, 573 clt->cl->pagevec, clt->cl->vcnt); 574 575 /* since file-backed online pages are traversed in reverse order */ 576 clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages; 577 return 0; 578 } 579 580 /* 581 * keep in mind that no referenced pclusters will be freed 582 * only after a RCU grace period. 583 */ 584 static void z_erofs_rcu_callback(struct rcu_head *head) 585 { 586 struct z_erofs_collection *const cl = 587 container_of(head, struct z_erofs_collection, rcu); 588 589 z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster, 590 primary_collection)); 591 } 592 593 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 594 { 595 struct z_erofs_pcluster *const pcl = 596 container_of(grp, struct z_erofs_pcluster, obj); 597 struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl); 598 599 call_rcu(&cl->rcu, z_erofs_rcu_callback); 600 } 601 602 static void z_erofs_collection_put(struct z_erofs_collection *cl) 603 { 604 struct z_erofs_pcluster *const pcl = 605 container_of(cl, struct z_erofs_pcluster, primary_collection); 606 607 erofs_workgroup_put(&pcl->obj); 608 } 609 610 static bool z_erofs_collector_end(struct z_erofs_collector *clt) 611 { 612 struct z_erofs_collection *cl = clt->cl; 613 614 if (!cl) 615 return false; 616 617 z_erofs_pagevec_ctor_exit(&clt->vector, false); 618 mutex_unlock(&cl->lock); 619 620 /* 621 * if all pending pages are added, don't hold its reference 622 * any longer if the pcluster isn't hosted by ourselves. 623 */ 624 if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE) 625 z_erofs_collection_put(cl); 626 627 clt->cl = NULL; 628 return true; 629 } 630 631 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe, 632 unsigned int cachestrategy, 633 erofs_off_t la) 634 { 635 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 636 return false; 637 638 if (fe->backmost) 639 return true; 640 641 return cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 642 la < fe->headoffset; 643 } 644 645 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 646 struct page *page, struct list_head *pagepool) 647 { 648 struct inode *const inode = fe->inode; 649 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 650 struct erofs_map_blocks *const map = &fe->map; 651 struct z_erofs_collector *const clt = &fe->clt; 652 const loff_t offset = page_offset(page); 653 bool tight = true; 654 655 enum z_erofs_cache_alloctype cache_strategy; 656 enum z_erofs_page_type page_type; 657 unsigned int cur, end, spiltted, index; 658 int err = 0; 659 660 /* register locked file pages as online pages in pack */ 661 z_erofs_onlinepage_init(page); 662 663 spiltted = 0; 664 end = PAGE_SIZE; 665 repeat: 666 cur = end - 1; 667 668 /* lucky, within the range of the current map_blocks */ 669 if (offset + cur >= map->m_la && 670 offset + cur < map->m_la + map->m_llen) { 671 /* didn't get a valid collection previously (very rare) */ 672 if (!clt->cl) 673 goto restart_now; 674 goto hitted; 675 } 676 677 /* go ahead the next map_blocks */ 678 erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur); 679 680 if (z_erofs_collector_end(clt)) 681 fe->backmost = false; 682 683 map->m_la = offset + cur; 684 map->m_llen = 0; 685 err = z_erofs_map_blocks_iter(inode, map, 0); 686 if (err) 687 goto err_out; 688 689 restart_now: 690 if (!(map->m_flags & EROFS_MAP_MAPPED)) 691 goto hitted; 692 693 err = z_erofs_collector_begin(clt, inode, map); 694 if (err) 695 goto err_out; 696 697 /* preload all compressed pages (maybe downgrade role if necessary) */ 698 if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la)) 699 cache_strategy = TRYALLOC; 700 else 701 cache_strategy = DONTALLOC; 702 703 preload_compressed_pages(clt, MNGD_MAPPING(sbi), 704 cache_strategy, pagepool); 705 706 hitted: 707 /* 708 * Ensure the current partial page belongs to this submit chain rather 709 * than other concurrent submit chains or the noio(bypass) chain since 710 * those chains are handled asynchronously thus the page cannot be used 711 * for inplace I/O or pagevec (should be processed in strict order.) 712 */ 713 tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED && 714 clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE); 715 716 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 717 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 718 zero_user_segment(page, cur, end); 719 goto next_part; 720 } 721 722 /* let's derive page type */ 723 page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD : 724 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 725 (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 726 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED)); 727 728 if (cur) 729 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED); 730 731 retry: 732 err = z_erofs_attach_page(clt, page, page_type, 733 clt->mode >= COLLECT_PRIMARY_FOLLOWED); 734 /* should allocate an additional short-lived page for pagevec */ 735 if (err == -EAGAIN) { 736 struct page *const newpage = 737 alloc_page(GFP_NOFS | __GFP_NOFAIL); 738 739 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE); 740 err = z_erofs_attach_page(clt, newpage, 741 Z_EROFS_PAGE_TYPE_EXCLUSIVE, true); 742 if (!err) 743 goto retry; 744 } 745 746 if (err) 747 goto err_out; 748 749 index = page->index - (map->m_la >> PAGE_SHIFT); 750 751 z_erofs_onlinepage_fixup(page, index, true); 752 753 /* bump up the number of spiltted parts of a page */ 754 ++spiltted; 755 /* also update nr_pages */ 756 clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1); 757 next_part: 758 /* can be used for verification */ 759 map->m_llen = offset + cur - map->m_la; 760 761 end = cur; 762 if (end > 0) 763 goto repeat; 764 765 out: 766 z_erofs_onlinepage_endio(page); 767 768 erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 769 __func__, page, spiltted, map->m_llen); 770 return err; 771 772 /* if some error occurred while processing this page */ 773 err_out: 774 SetPageError(page); 775 goto out; 776 } 777 778 static void z_erofs_decompressqueue_work(struct work_struct *work); 779 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 780 bool sync, int bios) 781 { 782 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 783 784 /* wake up the caller thread for sync decompression */ 785 if (sync) { 786 unsigned long flags; 787 788 spin_lock_irqsave(&io->u.wait.lock, flags); 789 if (!atomic_add_return(bios, &io->pending_bios)) 790 wake_up_locked(&io->u.wait); 791 spin_unlock_irqrestore(&io->u.wait.lock, flags); 792 return; 793 } 794 795 if (atomic_add_return(bios, &io->pending_bios)) 796 return; 797 /* Use workqueue and sync decompression for atomic contexts only */ 798 if (in_atomic() || irqs_disabled()) { 799 queue_work(z_erofs_workqueue, &io->u.work); 800 sbi->ctx.readahead_sync_decompress = true; 801 return; 802 } 803 z_erofs_decompressqueue_work(&io->u.work); 804 } 805 806 static bool z_erofs_page_is_invalidated(struct page *page) 807 { 808 return !page->mapping && !z_erofs_is_shortlived_page(page); 809 } 810 811 static void z_erofs_decompressqueue_endio(struct bio *bio) 812 { 813 tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private); 814 struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t); 815 blk_status_t err = bio->bi_status; 816 struct bio_vec *bvec; 817 struct bvec_iter_all iter_all; 818 819 bio_for_each_segment_all(bvec, bio, iter_all) { 820 struct page *page = bvec->bv_page; 821 822 DBG_BUGON(PageUptodate(page)); 823 DBG_BUGON(z_erofs_page_is_invalidated(page)); 824 825 if (err) 826 SetPageError(page); 827 828 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 829 if (!err) 830 SetPageUptodate(page); 831 unlock_page(page); 832 } 833 } 834 z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1); 835 bio_put(bio); 836 } 837 838 static int z_erofs_decompress_pcluster(struct super_block *sb, 839 struct z_erofs_pcluster *pcl, 840 struct list_head *pagepool) 841 { 842 struct erofs_sb_info *const sbi = EROFS_SB(sb); 843 struct z_erofs_pagevec_ctor ctor; 844 unsigned int i, inputsize, outputsize, llen, nr_pages; 845 struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES]; 846 struct page **pages, **compressed_pages, *page; 847 848 enum z_erofs_page_type page_type; 849 bool overlapped, partial; 850 struct z_erofs_collection *cl; 851 int err; 852 853 might_sleep(); 854 cl = z_erofs_primarycollection(pcl); 855 DBG_BUGON(!READ_ONCE(cl->nr_pages)); 856 857 mutex_lock(&cl->lock); 858 nr_pages = cl->nr_pages; 859 860 if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) { 861 pages = pages_onstack; 862 } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES && 863 mutex_trylock(&z_pagemap_global_lock)) { 864 pages = z_pagemap_global; 865 } else { 866 gfp_t gfp_flags = GFP_KERNEL; 867 868 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES) 869 gfp_flags |= __GFP_NOFAIL; 870 871 pages = kvmalloc_array(nr_pages, sizeof(struct page *), 872 gfp_flags); 873 874 /* fallback to global pagemap for the lowmem scenario */ 875 if (!pages) { 876 mutex_lock(&z_pagemap_global_lock); 877 pages = z_pagemap_global; 878 } 879 } 880 881 for (i = 0; i < nr_pages; ++i) 882 pages[i] = NULL; 883 884 err = 0; 885 z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS, 886 cl->pagevec, 0); 887 888 for (i = 0; i < cl->vcnt; ++i) { 889 unsigned int pagenr; 890 891 page = z_erofs_pagevec_dequeue(&ctor, &page_type); 892 893 /* all pages in pagevec ought to be valid */ 894 DBG_BUGON(!page); 895 DBG_BUGON(z_erofs_page_is_invalidated(page)); 896 897 if (z_erofs_put_shortlivedpage(pagepool, page)) 898 continue; 899 900 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD) 901 pagenr = 0; 902 else 903 pagenr = z_erofs_onlinepage_index(page); 904 905 DBG_BUGON(pagenr >= nr_pages); 906 907 /* 908 * currently EROFS doesn't support multiref(dedup), 909 * so here erroring out one multiref page. 910 */ 911 if (pages[pagenr]) { 912 DBG_BUGON(1); 913 SetPageError(pages[pagenr]); 914 z_erofs_onlinepage_endio(pages[pagenr]); 915 err = -EFSCORRUPTED; 916 } 917 pages[pagenr] = page; 918 } 919 z_erofs_pagevec_ctor_exit(&ctor, true); 920 921 overlapped = false; 922 compressed_pages = pcl->compressed_pages; 923 924 for (i = 0; i < pcl->pclusterpages; ++i) { 925 unsigned int pagenr; 926 927 page = compressed_pages[i]; 928 929 /* all compressed pages ought to be valid */ 930 DBG_BUGON(!page); 931 DBG_BUGON(z_erofs_page_is_invalidated(page)); 932 933 if (!z_erofs_is_shortlived_page(page)) { 934 if (erofs_page_is_managed(sbi, page)) { 935 if (!PageUptodate(page)) 936 err = -EIO; 937 continue; 938 } 939 940 /* 941 * only if non-head page can be selected 942 * for inplace decompression 943 */ 944 pagenr = z_erofs_onlinepage_index(page); 945 946 DBG_BUGON(pagenr >= nr_pages); 947 if (pages[pagenr]) { 948 DBG_BUGON(1); 949 SetPageError(pages[pagenr]); 950 z_erofs_onlinepage_endio(pages[pagenr]); 951 err = -EFSCORRUPTED; 952 } 953 pages[pagenr] = page; 954 955 overlapped = true; 956 } 957 958 /* PG_error needs checking for all non-managed pages */ 959 if (PageError(page)) { 960 DBG_BUGON(PageUptodate(page)); 961 err = -EIO; 962 } 963 } 964 965 if (err) 966 goto out; 967 968 llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT; 969 if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) { 970 outputsize = llen; 971 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH); 972 } else { 973 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs; 974 partial = true; 975 } 976 977 inputsize = pcl->pclusterpages * PAGE_SIZE; 978 err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 979 .sb = sb, 980 .in = compressed_pages, 981 .out = pages, 982 .pageofs_out = cl->pageofs, 983 .inputsize = inputsize, 984 .outputsize = outputsize, 985 .alg = pcl->algorithmformat, 986 .inplace_io = overlapped, 987 .partial_decoding = partial 988 }, pagepool); 989 990 out: 991 /* must handle all compressed pages before ending pages */ 992 for (i = 0; i < pcl->pclusterpages; ++i) { 993 page = compressed_pages[i]; 994 995 if (erofs_page_is_managed(sbi, page)) 996 continue; 997 998 /* recycle all individual short-lived pages */ 999 (void)z_erofs_put_shortlivedpage(pagepool, page); 1000 1001 WRITE_ONCE(compressed_pages[i], NULL); 1002 } 1003 1004 for (i = 0; i < nr_pages; ++i) { 1005 page = pages[i]; 1006 if (!page) 1007 continue; 1008 1009 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1010 1011 /* recycle all individual short-lived pages */ 1012 if (z_erofs_put_shortlivedpage(pagepool, page)) 1013 continue; 1014 1015 if (err < 0) 1016 SetPageError(page); 1017 1018 z_erofs_onlinepage_endio(page); 1019 } 1020 1021 if (pages == z_pagemap_global) 1022 mutex_unlock(&z_pagemap_global_lock); 1023 else if (pages != pages_onstack) 1024 kvfree(pages); 1025 1026 cl->nr_pages = 0; 1027 cl->vcnt = 0; 1028 1029 /* all cl locks MUST be taken before the following line */ 1030 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 1031 1032 /* all cl locks SHOULD be released right now */ 1033 mutex_unlock(&cl->lock); 1034 1035 z_erofs_collection_put(cl); 1036 return err; 1037 } 1038 1039 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 1040 struct list_head *pagepool) 1041 { 1042 z_erofs_next_pcluster_t owned = io->head; 1043 1044 while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 1045 struct z_erofs_pcluster *pcl; 1046 1047 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 1048 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 1049 1050 /* no possible that 'owned' equals NULL */ 1051 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1052 1053 pcl = container_of(owned, struct z_erofs_pcluster, next); 1054 owned = READ_ONCE(pcl->next); 1055 1056 z_erofs_decompress_pcluster(io->sb, pcl, pagepool); 1057 } 1058 } 1059 1060 static void z_erofs_decompressqueue_work(struct work_struct *work) 1061 { 1062 struct z_erofs_decompressqueue *bgq = 1063 container_of(work, struct z_erofs_decompressqueue, u.work); 1064 LIST_HEAD(pagepool); 1065 1066 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1067 z_erofs_decompress_queue(bgq, &pagepool); 1068 1069 put_pages_list(&pagepool); 1070 kvfree(bgq); 1071 } 1072 1073 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1074 unsigned int nr, 1075 struct list_head *pagepool, 1076 struct address_space *mc, 1077 gfp_t gfp) 1078 { 1079 const pgoff_t index = pcl->obj.index; 1080 bool tocache = false; 1081 1082 struct address_space *mapping; 1083 struct page *oldpage, *page; 1084 1085 compressed_page_t t; 1086 int justfound; 1087 1088 repeat: 1089 page = READ_ONCE(pcl->compressed_pages[nr]); 1090 oldpage = page; 1091 1092 if (!page) 1093 goto out_allocpage; 1094 1095 /* 1096 * the cached page has not been allocated and 1097 * an placeholder is out there, prepare it now. 1098 */ 1099 if (page == PAGE_UNALLOCATED) { 1100 tocache = true; 1101 goto out_allocpage; 1102 } 1103 1104 /* process the target tagged pointer */ 1105 t = tagptr_init(compressed_page_t, page); 1106 justfound = tagptr_unfold_tags(t); 1107 page = tagptr_unfold_ptr(t); 1108 1109 /* 1110 * preallocated cached pages, which is used to avoid direct reclaim 1111 * otherwise, it will go inplace I/O path instead. 1112 */ 1113 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1114 WRITE_ONCE(pcl->compressed_pages[nr], page); 1115 set_page_private(page, 0); 1116 tocache = true; 1117 goto out_tocache; 1118 } 1119 mapping = READ_ONCE(page->mapping); 1120 1121 /* 1122 * file-backed online pages in plcuster are all locked steady, 1123 * therefore it is impossible for `mapping' to be NULL. 1124 */ 1125 if (mapping && mapping != mc) 1126 /* ought to be unmanaged pages */ 1127 goto out; 1128 1129 /* directly return for shortlived page as well */ 1130 if (z_erofs_is_shortlived_page(page)) 1131 goto out; 1132 1133 lock_page(page); 1134 1135 /* only true if page reclaim goes wrong, should never happen */ 1136 DBG_BUGON(justfound && PagePrivate(page)); 1137 1138 /* the page is still in manage cache */ 1139 if (page->mapping == mc) { 1140 WRITE_ONCE(pcl->compressed_pages[nr], page); 1141 1142 ClearPageError(page); 1143 if (!PagePrivate(page)) { 1144 /* 1145 * impossible to be !PagePrivate(page) for 1146 * the current restriction as well if 1147 * the page is already in compressed_pages[]. 1148 */ 1149 DBG_BUGON(!justfound); 1150 1151 justfound = 0; 1152 set_page_private(page, (unsigned long)pcl); 1153 SetPagePrivate(page); 1154 } 1155 1156 /* no need to submit io if it is already up-to-date */ 1157 if (PageUptodate(page)) { 1158 unlock_page(page); 1159 page = NULL; 1160 } 1161 goto out; 1162 } 1163 1164 /* 1165 * the managed page has been truncated, it's unsafe to 1166 * reuse this one, let's allocate a new cache-managed page. 1167 */ 1168 DBG_BUGON(page->mapping); 1169 DBG_BUGON(!justfound); 1170 1171 tocache = true; 1172 unlock_page(page); 1173 put_page(page); 1174 out_allocpage: 1175 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1176 if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) { 1177 list_add(&page->lru, pagepool); 1178 cond_resched(); 1179 goto repeat; 1180 } 1181 out_tocache: 1182 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1183 /* turn into temporary page if fails (1 ref) */ 1184 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1185 goto out; 1186 } 1187 attach_page_private(page, pcl); 1188 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1189 put_page(page); 1190 1191 out: /* the only exit (for tracing and debugging) */ 1192 return page; 1193 } 1194 1195 static struct z_erofs_decompressqueue * 1196 jobqueue_init(struct super_block *sb, 1197 struct z_erofs_decompressqueue *fgq, bool *fg) 1198 { 1199 struct z_erofs_decompressqueue *q; 1200 1201 if (fg && !*fg) { 1202 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1203 if (!q) { 1204 *fg = true; 1205 goto fg_out; 1206 } 1207 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1208 } else { 1209 fg_out: 1210 q = fgq; 1211 init_waitqueue_head(&fgq->u.wait); 1212 atomic_set(&fgq->pending_bios, 0); 1213 } 1214 q->sb = sb; 1215 q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1216 return q; 1217 } 1218 1219 /* define decompression jobqueue types */ 1220 enum { 1221 JQ_BYPASS, 1222 JQ_SUBMIT, 1223 NR_JOBQUEUES, 1224 }; 1225 1226 static void *jobqueueset_init(struct super_block *sb, 1227 struct z_erofs_decompressqueue *q[], 1228 struct z_erofs_decompressqueue *fgq, bool *fg) 1229 { 1230 /* 1231 * if managed cache is enabled, bypass jobqueue is needed, 1232 * no need to read from device for all pclusters in this queue. 1233 */ 1234 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1235 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg); 1236 1237 return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg)); 1238 } 1239 1240 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1241 z_erofs_next_pcluster_t qtail[], 1242 z_erofs_next_pcluster_t owned_head) 1243 { 1244 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1245 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1246 1247 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1248 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1249 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1250 1251 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 1252 1253 WRITE_ONCE(*submit_qtail, owned_head); 1254 WRITE_ONCE(*bypass_qtail, &pcl->next); 1255 1256 qtail[JQ_BYPASS] = &pcl->next; 1257 } 1258 1259 static void z_erofs_submit_queue(struct super_block *sb, 1260 struct z_erofs_decompress_frontend *f, 1261 struct list_head *pagepool, 1262 struct z_erofs_decompressqueue *fgq, 1263 bool *force_fg) 1264 { 1265 struct erofs_sb_info *const sbi = 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->clt.owned_head; 1270 /* since bio will be NULL, no need to initialize last_index */ 1271 pgoff_t last_index; 1272 unsigned int nr_bios = 0; 1273 struct bio *bio = NULL; 1274 1275 bi_private = jobqueueset_init(sb, q, fgq, force_fg); 1276 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1277 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1278 1279 /* by default, all need io submission */ 1280 q[JQ_SUBMIT]->head = owned_head; 1281 1282 do { 1283 struct z_erofs_pcluster *pcl; 1284 pgoff_t cur, end; 1285 unsigned int i = 0; 1286 bool bypass = true; 1287 1288 /* no possible 'owned_head' equals the following */ 1289 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1290 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1291 1292 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1293 1294 cur = pcl->obj.index; 1295 end = cur + pcl->pclusterpages; 1296 1297 /* close the main owned chain at first */ 1298 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1299 Z_EROFS_PCLUSTER_TAIL_CLOSED); 1300 1301 do { 1302 struct page *page; 1303 1304 page = pickup_page_for_submission(pcl, i++, pagepool, 1305 MNGD_MAPPING(sbi), 1306 GFP_NOFS); 1307 if (!page) 1308 continue; 1309 1310 if (bio && cur != last_index + 1) { 1311 submit_bio_retry: 1312 submit_bio(bio); 1313 bio = NULL; 1314 } 1315 1316 if (!bio) { 1317 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS); 1318 1319 bio->bi_end_io = z_erofs_decompressqueue_endio; 1320 bio_set_dev(bio, sb->s_bdev); 1321 bio->bi_iter.bi_sector = (sector_t)cur << 1322 LOG_SECTORS_PER_BLOCK; 1323 bio->bi_private = bi_private; 1324 bio->bi_opf = REQ_OP_READ; 1325 if (f->readahead) 1326 bio->bi_opf |= REQ_RAHEAD; 1327 ++nr_bios; 1328 } 1329 1330 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1331 goto submit_bio_retry; 1332 1333 last_index = cur; 1334 bypass = false; 1335 } while (++cur < end); 1336 1337 if (!bypass) 1338 qtail[JQ_SUBMIT] = &pcl->next; 1339 else 1340 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1341 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1342 1343 if (bio) 1344 submit_bio(bio); 1345 1346 /* 1347 * although background is preferred, no one is pending for submission. 1348 * don't issue workqueue for decompression but drop it directly instead. 1349 */ 1350 if (!*force_fg && !nr_bios) { 1351 kvfree(q[JQ_SUBMIT]); 1352 return; 1353 } 1354 z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios); 1355 } 1356 1357 static void z_erofs_runqueue(struct super_block *sb, 1358 struct z_erofs_decompress_frontend *f, 1359 struct list_head *pagepool, bool force_fg) 1360 { 1361 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1362 1363 if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL) 1364 return; 1365 z_erofs_submit_queue(sb, f, pagepool, io, &force_fg); 1366 1367 /* handle bypass queue (no i/o pclusters) immediately */ 1368 z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool); 1369 1370 if (!force_fg) 1371 return; 1372 1373 /* wait until all bios are completed */ 1374 io_wait_event(io[JQ_SUBMIT].u.wait, 1375 !atomic_read(&io[JQ_SUBMIT].pending_bios)); 1376 1377 /* handle synchronous decompress queue in the caller context */ 1378 z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool); 1379 } 1380 1381 static int z_erofs_readpage(struct file *file, struct page *page) 1382 { 1383 struct inode *const inode = page->mapping->host; 1384 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1385 int err; 1386 LIST_HEAD(pagepool); 1387 1388 trace_erofs_readpage(page, false); 1389 1390 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1391 1392 err = z_erofs_do_read_page(&f, page, &pagepool); 1393 (void)z_erofs_collector_end(&f.clt); 1394 1395 /* if some compressed cluster ready, need submit them anyway */ 1396 z_erofs_runqueue(inode->i_sb, &f, &pagepool, true); 1397 1398 if (err) 1399 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1400 1401 if (f.map.mpage) 1402 put_page(f.map.mpage); 1403 1404 /* clean up the remaining free pages */ 1405 put_pages_list(&pagepool); 1406 return err; 1407 } 1408 1409 static void z_erofs_readahead(struct readahead_control *rac) 1410 { 1411 struct inode *const inode = rac->mapping->host; 1412 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1413 1414 unsigned int nr_pages = readahead_count(rac); 1415 bool sync = (sbi->ctx.readahead_sync_decompress && 1416 nr_pages <= sbi->ctx.max_sync_decompress_pages); 1417 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1418 struct page *page, *head = NULL; 1419 LIST_HEAD(pagepool); 1420 1421 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1422 1423 f.readahead = true; 1424 f.headoffset = readahead_pos(rac); 1425 1426 while ((page = readahead_page(rac))) { 1427 prefetchw(&page->flags); 1428 1429 /* 1430 * A pure asynchronous readahead is indicated if 1431 * a PG_readahead marked page is hitted at first. 1432 * Let's also do asynchronous decompression for this case. 1433 */ 1434 sync &= !(PageReadahead(page) && !head); 1435 1436 set_page_private(page, (unsigned long)head); 1437 head = page; 1438 } 1439 1440 while (head) { 1441 struct page *page = head; 1442 int err; 1443 1444 /* traversal in reverse order */ 1445 head = (void *)page_private(page); 1446 1447 err = z_erofs_do_read_page(&f, page, &pagepool); 1448 if (err) 1449 erofs_err(inode->i_sb, 1450 "readahead error at page %lu @ nid %llu", 1451 page->index, EROFS_I(inode)->nid); 1452 put_page(page); 1453 } 1454 1455 (void)z_erofs_collector_end(&f.clt); 1456 1457 z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync); 1458 1459 if (f.map.mpage) 1460 put_page(f.map.mpage); 1461 1462 /* clean up the remaining free pages */ 1463 put_pages_list(&pagepool); 1464 } 1465 1466 const struct address_space_operations z_erofs_aops = { 1467 .readpage = z_erofs_readpage, 1468 .readahead = z_erofs_readahead, 1469 }; 1470