1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2018 HUAWEI, Inc. 4 * https://www.huawei.com/ 5 * Created by Gao Xiang <gaoxiang25@huawei.com> 6 */ 7 #include "zdata.h" 8 #include "compress.h" 9 #include <linux/prefetch.h> 10 11 #include <trace/events/erofs.h> 12 13 /* 14 * since pclustersize is variable for big pcluster feature, introduce slab 15 * pools implementation for different pcluster sizes. 16 */ 17 struct z_erofs_pcluster_slab { 18 struct kmem_cache *slab; 19 unsigned int maxpages; 20 char name[48]; 21 }; 22 23 #define _PCLP(n) { .maxpages = n } 24 25 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = { 26 _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128), 27 _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES) 28 }; 29 30 static void z_erofs_destroy_pcluster_pool(void) 31 { 32 int i; 33 34 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 35 if (!pcluster_pool[i].slab) 36 continue; 37 kmem_cache_destroy(pcluster_pool[i].slab); 38 pcluster_pool[i].slab = NULL; 39 } 40 } 41 42 static int z_erofs_create_pcluster_pool(void) 43 { 44 struct z_erofs_pcluster_slab *pcs; 45 struct z_erofs_pcluster *a; 46 unsigned int size; 47 48 for (pcs = pcluster_pool; 49 pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) { 50 size = struct_size(a, compressed_pages, pcs->maxpages); 51 52 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages); 53 pcs->slab = kmem_cache_create(pcs->name, size, 0, 54 SLAB_RECLAIM_ACCOUNT, NULL); 55 if (pcs->slab) 56 continue; 57 58 z_erofs_destroy_pcluster_pool(); 59 return -ENOMEM; 60 } 61 return 0; 62 } 63 64 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages) 65 { 66 int i; 67 68 for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 69 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 70 struct z_erofs_pcluster *pcl; 71 72 if (nrpages > pcs->maxpages) 73 continue; 74 75 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS); 76 if (!pcl) 77 return ERR_PTR(-ENOMEM); 78 pcl->pclusterpages = nrpages; 79 return pcl; 80 } 81 return ERR_PTR(-EINVAL); 82 } 83 84 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl) 85 { 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 (pcl->pclusterpages > pcs->maxpages) 92 continue; 93 94 kmem_cache_free(pcs->slab, pcl); 95 return; 96 } 97 DBG_BUGON(1); 98 } 99 100 /* 101 * a compressed_pages[] placeholder in order to avoid 102 * being filled with file pages for in-place decompression. 103 */ 104 #define PAGE_UNALLOCATED ((void *)0x5F0E4B1D) 105 106 /* how to allocate cached pages for a pcluster */ 107 enum z_erofs_cache_alloctype { 108 DONTALLOC, /* don't allocate any cached pages */ 109 DELAYEDALLOC, /* delayed allocation (at the time of submitting io) */ 110 /* 111 * try to use cached I/O if page allocation succeeds or fallback 112 * to in-place I/O instead to avoid any direct reclaim. 113 */ 114 TRYALLOC, 115 }; 116 117 /* 118 * tagged pointer with 1-bit tag for all compressed pages 119 * tag 0 - the page is just found with an extra page reference 120 */ 121 typedef tagptr1_t compressed_page_t; 122 123 #define tag_compressed_page_justfound(page) \ 124 tagptr_fold(compressed_page_t, page, 1) 125 126 static struct workqueue_struct *z_erofs_workqueue __read_mostly; 127 128 void z_erofs_exit_zip_subsystem(void) 129 { 130 destroy_workqueue(z_erofs_workqueue); 131 z_erofs_destroy_pcluster_pool(); 132 } 133 134 static inline int z_erofs_init_workqueue(void) 135 { 136 const unsigned int onlinecpus = num_possible_cpus(); 137 138 /* 139 * no need to spawn too many threads, limiting threads could minimum 140 * scheduling overhead, perhaps per-CPU threads should be better? 141 */ 142 z_erofs_workqueue = alloc_workqueue("erofs_unzipd", 143 WQ_UNBOUND | WQ_HIGHPRI, 144 onlinecpus + onlinecpus / 4); 145 return z_erofs_workqueue ? 0 : -ENOMEM; 146 } 147 148 int __init z_erofs_init_zip_subsystem(void) 149 { 150 int err = z_erofs_create_pcluster_pool(); 151 152 if (err) 153 return err; 154 err = z_erofs_init_workqueue(); 155 if (err) 156 z_erofs_destroy_pcluster_pool(); 157 return err; 158 } 159 160 enum z_erofs_collectmode { 161 COLLECT_SECONDARY, 162 COLLECT_PRIMARY, 163 /* 164 * The current collection was the tail of an exist chain, in addition 165 * that the previous processed chained collections are all decided to 166 * be hooked up to it. 167 * A new chain will be created for the remaining collections which are 168 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED, 169 * the next collection cannot reuse the whole page safely in 170 * the following scenario: 171 * ________________________________________________________________ 172 * | tail (partial) page | head (partial) page | 173 * | (belongs to the next cl) | (belongs to the current cl) | 174 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________| 175 */ 176 COLLECT_PRIMARY_HOOKED, 177 /* 178 * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it 179 * could be dispatched into bypass queue later due to uptodated managed 180 * pages. All related online pages cannot be reused for inplace I/O (or 181 * pagevec) since it can be directly decoded without I/O submission. 182 */ 183 COLLECT_PRIMARY_FOLLOWED_NOINPLACE, 184 /* 185 * The current collection has been linked with the owned chain, and 186 * could also be linked with the remaining collections, which means 187 * if the processing page is the tail page of the collection, thus 188 * the current collection can safely use the whole page (since 189 * the previous collection is under control) for in-place I/O, as 190 * illustrated below: 191 * ________________________________________________________________ 192 * | tail (partial) page | head (partial) page | 193 * | (of the current cl) | (of the previous collection) | 194 * | PRIMARY_FOLLOWED or | | 195 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________| 196 * 197 * [ (*) the above page can be used as inplace I/O. ] 198 */ 199 COLLECT_PRIMARY_FOLLOWED, 200 }; 201 202 struct z_erofs_collector { 203 struct z_erofs_pagevec_ctor vector; 204 205 struct z_erofs_pcluster *pcl, *tailpcl; 206 struct z_erofs_collection *cl; 207 /* a pointer used to pick up inplace I/O pages */ 208 struct page **icpage_ptr; 209 z_erofs_next_pcluster_t owned_head; 210 211 enum z_erofs_collectmode mode; 212 }; 213 214 struct z_erofs_decompress_frontend { 215 struct inode *const inode; 216 217 struct z_erofs_collector clt; 218 struct erofs_map_blocks map; 219 220 bool readahead; 221 /* used for applying cache strategy on the fly */ 222 bool backmost; 223 erofs_off_t headoffset; 224 }; 225 226 #define COLLECTOR_INIT() { \ 227 .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 228 .mode = COLLECT_PRIMARY_FOLLOWED } 229 230 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 231 .inode = __i, .clt = COLLECTOR_INIT(), \ 232 .backmost = true, } 233 234 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES]; 235 static DEFINE_MUTEX(z_pagemap_global_lock); 236 237 static void preload_compressed_pages(struct z_erofs_collector *clt, 238 struct address_space *mc, 239 enum z_erofs_cache_alloctype type, 240 struct list_head *pagepool) 241 { 242 struct z_erofs_pcluster *pcl = clt->pcl; 243 bool standalone = true; 244 gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 245 __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 246 struct page **pages; 247 pgoff_t index; 248 249 if (clt->mode < COLLECT_PRIMARY_FOLLOWED) 250 return; 251 252 pages = pcl->compressed_pages; 253 index = pcl->obj.index; 254 for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) { 255 struct page *page; 256 compressed_page_t t; 257 struct page *newpage = NULL; 258 259 /* the compressed page was loaded before */ 260 if (READ_ONCE(*pages)) 261 continue; 262 263 page = find_get_page(mc, index); 264 265 if (page) { 266 t = tag_compressed_page_justfound(page); 267 } else { 268 /* I/O is needed, no possible to decompress directly */ 269 standalone = false; 270 switch (type) { 271 case DELAYEDALLOC: 272 t = tagptr_init(compressed_page_t, 273 PAGE_UNALLOCATED); 274 break; 275 case TRYALLOC: 276 newpage = erofs_allocpage(pagepool, gfp); 277 if (!newpage) 278 continue; 279 set_page_private(newpage, 280 Z_EROFS_PREALLOCATED_PAGE); 281 t = tag_compressed_page_justfound(newpage); 282 break; 283 default: /* DONTALLOC */ 284 continue; 285 } 286 } 287 288 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t))) 289 continue; 290 291 if (page) { 292 put_page(page); 293 } else if (newpage) { 294 set_page_private(newpage, 0); 295 list_add(&newpage->lru, pagepool); 296 } 297 } 298 299 /* 300 * don't do inplace I/O if all compressed pages are available in 301 * managed cache since it can be moved to the bypass queue instead. 302 */ 303 if (standalone) 304 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 305 } 306 307 /* called by erofs_shrinker to get rid of all compressed_pages */ 308 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 309 struct erofs_workgroup *grp) 310 { 311 struct z_erofs_pcluster *const pcl = 312 container_of(grp, struct z_erofs_pcluster, obj); 313 struct address_space *const mapping = MNGD_MAPPING(sbi); 314 int i; 315 316 /* 317 * refcount of workgroup is now freezed as 1, 318 * therefore no need to worry about available decompression users. 319 */ 320 for (i = 0; i < pcl->pclusterpages; ++i) { 321 struct page *page = pcl->compressed_pages[i]; 322 323 if (!page) 324 continue; 325 326 /* block other users from reclaiming or migrating the page */ 327 if (!trylock_page(page)) 328 return -EBUSY; 329 330 if (page->mapping != mapping) 331 continue; 332 333 /* barrier is implied in the following 'unlock_page' */ 334 WRITE_ONCE(pcl->compressed_pages[i], NULL); 335 detach_page_private(page); 336 unlock_page(page); 337 } 338 return 0; 339 } 340 341 int erofs_try_to_free_cached_page(struct address_space *mapping, 342 struct page *page) 343 { 344 struct z_erofs_pcluster *const pcl = (void *)page_private(page); 345 int ret = 0; /* 0 - busy */ 346 347 if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) { 348 unsigned int i; 349 350 for (i = 0; i < pcl->pclusterpages; ++i) { 351 if (pcl->compressed_pages[i] == page) { 352 WRITE_ONCE(pcl->compressed_pages[i], NULL); 353 ret = 1; 354 break; 355 } 356 } 357 erofs_workgroup_unfreeze(&pcl->obj, 1); 358 359 if (ret) 360 detach_page_private(page); 361 } 362 return ret; 363 } 364 365 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */ 366 static bool z_erofs_try_inplace_io(struct z_erofs_collector *clt, 367 struct page *page) 368 { 369 struct z_erofs_pcluster *const pcl = clt->pcl; 370 371 while (clt->icpage_ptr > pcl->compressed_pages) 372 if (!cmpxchg(--clt->icpage_ptr, NULL, page)) 373 return true; 374 return false; 375 } 376 377 /* callers must be with collection lock held */ 378 static int z_erofs_attach_page(struct z_erofs_collector *clt, 379 struct page *page, 380 enum z_erofs_page_type type) 381 { 382 int ret; 383 384 /* give priority for inplaceio */ 385 if (clt->mode >= COLLECT_PRIMARY && 386 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE && 387 z_erofs_try_inplace_io(clt, page)) 388 return 0; 389 390 ret = z_erofs_pagevec_enqueue(&clt->vector, page, type); 391 clt->cl->vcnt += (unsigned int)ret; 392 393 return ret ? 0 : -EAGAIN; 394 } 395 396 static void z_erofs_try_to_claim_pcluster(struct z_erofs_collector *clt) 397 { 398 struct z_erofs_pcluster *pcl = clt->pcl; 399 z_erofs_next_pcluster_t *owned_head = &clt->owned_head; 400 401 /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 402 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 403 *owned_head) == Z_EROFS_PCLUSTER_NIL) { 404 *owned_head = &pcl->next; 405 /* so we can attach this pcluster to our submission chain. */ 406 clt->mode = COLLECT_PRIMARY_FOLLOWED; 407 return; 408 } 409 410 /* 411 * type 2, link to the end of an existing open chain, be careful 412 * that its submission is controlled by the original attached chain. 413 */ 414 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 415 *owned_head) == Z_EROFS_PCLUSTER_TAIL) { 416 *owned_head = Z_EROFS_PCLUSTER_TAIL; 417 clt->mode = COLLECT_PRIMARY_HOOKED; 418 clt->tailpcl = NULL; 419 return; 420 } 421 /* type 3, it belongs to a chain, but it isn't the end of the chain */ 422 clt->mode = COLLECT_PRIMARY; 423 } 424 425 static int z_erofs_lookup_collection(struct z_erofs_collector *clt, 426 struct inode *inode, 427 struct erofs_map_blocks *map) 428 { 429 struct z_erofs_pcluster *pcl = clt->pcl; 430 struct z_erofs_collection *cl; 431 unsigned int length; 432 433 /* to avoid unexpected loop formed by corrupted images */ 434 if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) { 435 DBG_BUGON(1); 436 return -EFSCORRUPTED; 437 } 438 439 cl = z_erofs_primarycollection(pcl); 440 if (cl->pageofs != (map->m_la & ~PAGE_MASK)) { 441 DBG_BUGON(1); 442 return -EFSCORRUPTED; 443 } 444 445 length = READ_ONCE(pcl->length); 446 if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) { 447 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) { 448 DBG_BUGON(1); 449 return -EFSCORRUPTED; 450 } 451 } else { 452 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT; 453 454 if (map->m_flags & EROFS_MAP_FULL_MAPPED) 455 llen |= Z_EROFS_PCLUSTER_FULL_LENGTH; 456 457 while (llen > length && 458 length != cmpxchg_relaxed(&pcl->length, length, llen)) { 459 cpu_relax(); 460 length = READ_ONCE(pcl->length); 461 } 462 } 463 mutex_lock(&cl->lock); 464 /* used to check tail merging loop due to corrupted images */ 465 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 466 clt->tailpcl = pcl; 467 468 z_erofs_try_to_claim_pcluster(clt); 469 clt->cl = cl; 470 return 0; 471 } 472 473 static int z_erofs_register_collection(struct z_erofs_collector *clt, 474 struct inode *inode, 475 struct erofs_map_blocks *map) 476 { 477 struct z_erofs_pcluster *pcl; 478 struct z_erofs_collection *cl; 479 struct erofs_workgroup *grp; 480 int err; 481 482 /* no available pcluster, let's allocate one */ 483 pcl = z_erofs_alloc_pcluster(map->m_plen >> PAGE_SHIFT); 484 if (IS_ERR(pcl)) 485 return PTR_ERR(pcl); 486 487 atomic_set(&pcl->obj.refcount, 1); 488 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 489 490 pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) | 491 (map->m_flags & EROFS_MAP_FULL_MAPPED ? 492 Z_EROFS_PCLUSTER_FULL_LENGTH : 0); 493 494 if (map->m_flags & EROFS_MAP_ZIPPED) 495 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4; 496 else 497 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED; 498 499 /* new pclusters should be claimed as type 1, primary and followed */ 500 pcl->next = clt->owned_head; 501 clt->mode = COLLECT_PRIMARY_FOLLOWED; 502 503 cl = z_erofs_primarycollection(pcl); 504 cl->pageofs = map->m_la & ~PAGE_MASK; 505 506 /* 507 * lock all primary followed works before visible to others 508 * and mutex_trylock *never* fails for a new pcluster. 509 */ 510 mutex_init(&cl->lock); 511 DBG_BUGON(!mutex_trylock(&cl->lock)); 512 513 grp = erofs_insert_workgroup(inode->i_sb, &pcl->obj); 514 if (IS_ERR(grp)) { 515 err = PTR_ERR(grp); 516 goto err_out; 517 } 518 519 if (grp != &pcl->obj) { 520 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj); 521 err = -EEXIST; 522 goto err_out; 523 } 524 /* used to check tail merging loop due to corrupted images */ 525 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 526 clt->tailpcl = pcl; 527 clt->owned_head = &pcl->next; 528 clt->pcl = pcl; 529 clt->cl = cl; 530 return 0; 531 532 err_out: 533 mutex_unlock(&cl->lock); 534 z_erofs_free_pcluster(pcl); 535 return err; 536 } 537 538 static int z_erofs_collector_begin(struct z_erofs_collector *clt, 539 struct inode *inode, 540 struct erofs_map_blocks *map) 541 { 542 struct erofs_workgroup *grp; 543 int ret; 544 545 DBG_BUGON(clt->cl); 546 547 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */ 548 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL); 549 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 550 551 if (!PAGE_ALIGNED(map->m_pa)) { 552 DBG_BUGON(1); 553 return -EINVAL; 554 } 555 556 grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT); 557 if (grp) { 558 clt->pcl = container_of(grp, struct z_erofs_pcluster, obj); 559 } else { 560 ret = z_erofs_register_collection(clt, inode, map); 561 562 if (!ret) 563 goto out; 564 if (ret != -EEXIST) 565 return ret; 566 } 567 568 ret = z_erofs_lookup_collection(clt, inode, map); 569 if (ret) { 570 erofs_workgroup_put(&clt->pcl->obj); 571 return ret; 572 } 573 574 out: 575 z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS, 576 clt->cl->pagevec, clt->cl->vcnt); 577 578 /* since file-backed online pages are traversed in reverse order */ 579 clt->icpage_ptr = clt->pcl->compressed_pages + clt->pcl->pclusterpages; 580 return 0; 581 } 582 583 /* 584 * keep in mind that no referenced pclusters will be freed 585 * only after a RCU grace period. 586 */ 587 static void z_erofs_rcu_callback(struct rcu_head *head) 588 { 589 struct z_erofs_collection *const cl = 590 container_of(head, struct z_erofs_collection, rcu); 591 592 z_erofs_free_pcluster(container_of(cl, struct z_erofs_pcluster, 593 primary_collection)); 594 } 595 596 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 597 { 598 struct z_erofs_pcluster *const pcl = 599 container_of(grp, struct z_erofs_pcluster, obj); 600 struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl); 601 602 call_rcu(&cl->rcu, z_erofs_rcu_callback); 603 } 604 605 static void z_erofs_collection_put(struct z_erofs_collection *cl) 606 { 607 struct z_erofs_pcluster *const pcl = 608 container_of(cl, struct z_erofs_pcluster, primary_collection); 609 610 erofs_workgroup_put(&pcl->obj); 611 } 612 613 static bool z_erofs_collector_end(struct z_erofs_collector *clt) 614 { 615 struct z_erofs_collection *cl = clt->cl; 616 617 if (!cl) 618 return false; 619 620 z_erofs_pagevec_ctor_exit(&clt->vector, false); 621 mutex_unlock(&cl->lock); 622 623 /* 624 * if all pending pages are added, don't hold its reference 625 * any longer if the pcluster isn't hosted by ourselves. 626 */ 627 if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE) 628 z_erofs_collection_put(cl); 629 630 clt->cl = NULL; 631 return true; 632 } 633 634 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe, 635 unsigned int cachestrategy, 636 erofs_off_t la) 637 { 638 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 639 return false; 640 641 if (fe->backmost) 642 return true; 643 644 return cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 645 la < fe->headoffset; 646 } 647 648 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 649 struct page *page, struct list_head *pagepool) 650 { 651 struct inode *const inode = fe->inode; 652 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 653 struct erofs_map_blocks *const map = &fe->map; 654 struct z_erofs_collector *const clt = &fe->clt; 655 const loff_t offset = page_offset(page); 656 bool tight = true; 657 658 enum z_erofs_cache_alloctype cache_strategy; 659 enum z_erofs_page_type page_type; 660 unsigned int cur, end, spiltted, index; 661 int err = 0; 662 663 /* register locked file pages as online pages in pack */ 664 z_erofs_onlinepage_init(page); 665 666 spiltted = 0; 667 end = PAGE_SIZE; 668 repeat: 669 cur = end - 1; 670 671 /* lucky, within the range of the current map_blocks */ 672 if (offset + cur >= map->m_la && 673 offset + cur < map->m_la + map->m_llen) { 674 /* didn't get a valid collection previously (very rare) */ 675 if (!clt->cl) 676 goto restart_now; 677 goto hitted; 678 } 679 680 /* go ahead the next map_blocks */ 681 erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur); 682 683 if (z_erofs_collector_end(clt)) 684 fe->backmost = false; 685 686 map->m_la = offset + cur; 687 map->m_llen = 0; 688 err = z_erofs_map_blocks_iter(inode, map, 0); 689 if (err) 690 goto err_out; 691 692 restart_now: 693 if (!(map->m_flags & EROFS_MAP_MAPPED)) 694 goto hitted; 695 696 err = z_erofs_collector_begin(clt, inode, map); 697 if (err) 698 goto err_out; 699 700 /* preload all compressed pages (maybe downgrade role if necessary) */ 701 if (should_alloc_managed_pages(fe, sbi->ctx.cache_strategy, map->m_la)) 702 cache_strategy = TRYALLOC; 703 else 704 cache_strategy = DONTALLOC; 705 706 preload_compressed_pages(clt, MNGD_MAPPING(sbi), 707 cache_strategy, pagepool); 708 709 hitted: 710 /* 711 * Ensure the current partial page belongs to this submit chain rather 712 * than other concurrent submit chains or the noio(bypass) chain since 713 * those chains are handled asynchronously thus the page cannot be used 714 * for inplace I/O or pagevec (should be processed in strict order.) 715 */ 716 tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED && 717 clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE); 718 719 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 720 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 721 zero_user_segment(page, cur, end); 722 goto next_part; 723 } 724 725 /* let's derive page type */ 726 page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD : 727 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 728 (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 729 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED)); 730 731 if (cur) 732 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED); 733 734 retry: 735 err = z_erofs_attach_page(clt, page, page_type); 736 /* should allocate an additional short-lived page for pagevec */ 737 if (err == -EAGAIN) { 738 struct page *const newpage = 739 alloc_page(GFP_NOFS | __GFP_NOFAIL); 740 741 set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE); 742 err = z_erofs_attach_page(clt, newpage, 743 Z_EROFS_PAGE_TYPE_EXCLUSIVE); 744 if (!err) 745 goto retry; 746 } 747 748 if (err) 749 goto err_out; 750 751 index = page->index - (map->m_la >> PAGE_SHIFT); 752 753 z_erofs_onlinepage_fixup(page, index, true); 754 755 /* bump up the number of spiltted parts of a page */ 756 ++spiltted; 757 /* also update nr_pages */ 758 clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1); 759 next_part: 760 /* can be used for verification */ 761 map->m_llen = offset + cur - map->m_la; 762 763 end = cur; 764 if (end > 0) 765 goto repeat; 766 767 out: 768 z_erofs_onlinepage_endio(page); 769 770 erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 771 __func__, page, spiltted, map->m_llen); 772 return err; 773 774 /* if some error occurred while processing this page */ 775 err_out: 776 SetPageError(page); 777 goto out; 778 } 779 780 static void z_erofs_decompressqueue_work(struct work_struct *work); 781 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 782 bool sync, int bios) 783 { 784 struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 785 786 /* wake up the caller thread for sync decompression */ 787 if (sync) { 788 unsigned long flags; 789 790 spin_lock_irqsave(&io->u.wait.lock, flags); 791 if (!atomic_add_return(bios, &io->pending_bios)) 792 wake_up_locked(&io->u.wait); 793 spin_unlock_irqrestore(&io->u.wait.lock, flags); 794 return; 795 } 796 797 if (atomic_add_return(bios, &io->pending_bios)) 798 return; 799 /* Use workqueue and sync decompression for atomic contexts only */ 800 if (in_atomic() || irqs_disabled()) { 801 queue_work(z_erofs_workqueue, &io->u.work); 802 sbi->ctx.readahead_sync_decompress = true; 803 return; 804 } 805 z_erofs_decompressqueue_work(&io->u.work); 806 } 807 808 static bool z_erofs_page_is_invalidated(struct page *page) 809 { 810 return !page->mapping && !z_erofs_is_shortlived_page(page); 811 } 812 813 static void z_erofs_decompressqueue_endio(struct bio *bio) 814 { 815 tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private); 816 struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t); 817 blk_status_t err = bio->bi_status; 818 struct bio_vec *bvec; 819 struct bvec_iter_all iter_all; 820 821 bio_for_each_segment_all(bvec, bio, iter_all) { 822 struct page *page = bvec->bv_page; 823 824 DBG_BUGON(PageUptodate(page)); 825 DBG_BUGON(z_erofs_page_is_invalidated(page)); 826 827 if (err) 828 SetPageError(page); 829 830 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 831 if (!err) 832 SetPageUptodate(page); 833 unlock_page(page); 834 } 835 } 836 z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1); 837 bio_put(bio); 838 } 839 840 static int z_erofs_decompress_pcluster(struct super_block *sb, 841 struct z_erofs_pcluster *pcl, 842 struct list_head *pagepool) 843 { 844 struct erofs_sb_info *const sbi = EROFS_SB(sb); 845 struct z_erofs_pagevec_ctor ctor; 846 unsigned int i, inputsize, outputsize, llen, nr_pages; 847 struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES]; 848 struct page **pages, **compressed_pages, *page; 849 850 enum z_erofs_page_type page_type; 851 bool overlapped, partial; 852 struct z_erofs_collection *cl; 853 int err; 854 855 might_sleep(); 856 cl = z_erofs_primarycollection(pcl); 857 DBG_BUGON(!READ_ONCE(cl->nr_pages)); 858 859 mutex_lock(&cl->lock); 860 nr_pages = cl->nr_pages; 861 862 if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) { 863 pages = pages_onstack; 864 } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES && 865 mutex_trylock(&z_pagemap_global_lock)) { 866 pages = z_pagemap_global; 867 } else { 868 gfp_t gfp_flags = GFP_KERNEL; 869 870 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES) 871 gfp_flags |= __GFP_NOFAIL; 872 873 pages = kvmalloc_array(nr_pages, sizeof(struct page *), 874 gfp_flags); 875 876 /* fallback to global pagemap for the lowmem scenario */ 877 if (!pages) { 878 mutex_lock(&z_pagemap_global_lock); 879 pages = z_pagemap_global; 880 } 881 } 882 883 for (i = 0; i < nr_pages; ++i) 884 pages[i] = NULL; 885 886 err = 0; 887 z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS, 888 cl->pagevec, 0); 889 890 for (i = 0; i < cl->vcnt; ++i) { 891 unsigned int pagenr; 892 893 page = z_erofs_pagevec_dequeue(&ctor, &page_type); 894 895 /* all pages in pagevec ought to be valid */ 896 DBG_BUGON(!page); 897 DBG_BUGON(z_erofs_page_is_invalidated(page)); 898 899 if (z_erofs_put_shortlivedpage(pagepool, page)) 900 continue; 901 902 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD) 903 pagenr = 0; 904 else 905 pagenr = z_erofs_onlinepage_index(page); 906 907 DBG_BUGON(pagenr >= nr_pages); 908 909 /* 910 * currently EROFS doesn't support multiref(dedup), 911 * so here erroring out one multiref page. 912 */ 913 if (pages[pagenr]) { 914 DBG_BUGON(1); 915 SetPageError(pages[pagenr]); 916 z_erofs_onlinepage_endio(pages[pagenr]); 917 err = -EFSCORRUPTED; 918 } 919 pages[pagenr] = page; 920 } 921 z_erofs_pagevec_ctor_exit(&ctor, true); 922 923 overlapped = false; 924 compressed_pages = pcl->compressed_pages; 925 926 for (i = 0; i < pcl->pclusterpages; ++i) { 927 unsigned int pagenr; 928 929 page = compressed_pages[i]; 930 931 /* all compressed pages ought to be valid */ 932 DBG_BUGON(!page); 933 DBG_BUGON(z_erofs_page_is_invalidated(page)); 934 935 if (!z_erofs_is_shortlived_page(page)) { 936 if (erofs_page_is_managed(sbi, page)) { 937 if (!PageUptodate(page)) 938 err = -EIO; 939 continue; 940 } 941 942 /* 943 * only if non-head page can be selected 944 * for inplace decompression 945 */ 946 pagenr = z_erofs_onlinepage_index(page); 947 948 DBG_BUGON(pagenr >= nr_pages); 949 if (pages[pagenr]) { 950 DBG_BUGON(1); 951 SetPageError(pages[pagenr]); 952 z_erofs_onlinepage_endio(pages[pagenr]); 953 err = -EFSCORRUPTED; 954 } 955 pages[pagenr] = page; 956 957 overlapped = true; 958 } 959 960 /* PG_error needs checking for all non-managed pages */ 961 if (PageError(page)) { 962 DBG_BUGON(PageUptodate(page)); 963 err = -EIO; 964 } 965 } 966 967 if (err) 968 goto out; 969 970 llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT; 971 if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) { 972 outputsize = llen; 973 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH); 974 } else { 975 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs; 976 partial = true; 977 } 978 979 inputsize = pcl->pclusterpages * PAGE_SIZE; 980 err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 981 .sb = sb, 982 .in = compressed_pages, 983 .out = pages, 984 .pageofs_out = cl->pageofs, 985 .inputsize = inputsize, 986 .outputsize = outputsize, 987 .alg = pcl->algorithmformat, 988 .inplace_io = overlapped, 989 .partial_decoding = partial 990 }, pagepool); 991 992 out: 993 /* must handle all compressed pages before ending pages */ 994 for (i = 0; i < pcl->pclusterpages; ++i) { 995 page = compressed_pages[i]; 996 997 if (erofs_page_is_managed(sbi, page)) 998 continue; 999 1000 /* recycle all individual short-lived pages */ 1001 (void)z_erofs_put_shortlivedpage(pagepool, page); 1002 1003 WRITE_ONCE(compressed_pages[i], NULL); 1004 } 1005 1006 for (i = 0; i < nr_pages; ++i) { 1007 page = pages[i]; 1008 if (!page) 1009 continue; 1010 1011 DBG_BUGON(z_erofs_page_is_invalidated(page)); 1012 1013 /* recycle all individual short-lived pages */ 1014 if (z_erofs_put_shortlivedpage(pagepool, page)) 1015 continue; 1016 1017 if (err < 0) 1018 SetPageError(page); 1019 1020 z_erofs_onlinepage_endio(page); 1021 } 1022 1023 if (pages == z_pagemap_global) 1024 mutex_unlock(&z_pagemap_global_lock); 1025 else if (pages != pages_onstack) 1026 kvfree(pages); 1027 1028 cl->nr_pages = 0; 1029 cl->vcnt = 0; 1030 1031 /* all cl locks MUST be taken before the following line */ 1032 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 1033 1034 /* all cl locks SHOULD be released right now */ 1035 mutex_unlock(&cl->lock); 1036 1037 z_erofs_collection_put(cl); 1038 return err; 1039 } 1040 1041 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 1042 struct list_head *pagepool) 1043 { 1044 z_erofs_next_pcluster_t owned = io->head; 1045 1046 while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 1047 struct z_erofs_pcluster *pcl; 1048 1049 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 1050 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 1051 1052 /* no possible that 'owned' equals NULL */ 1053 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 1054 1055 pcl = container_of(owned, struct z_erofs_pcluster, next); 1056 owned = READ_ONCE(pcl->next); 1057 1058 z_erofs_decompress_pcluster(io->sb, pcl, pagepool); 1059 } 1060 } 1061 1062 static void z_erofs_decompressqueue_work(struct work_struct *work) 1063 { 1064 struct z_erofs_decompressqueue *bgq = 1065 container_of(work, struct z_erofs_decompressqueue, u.work); 1066 LIST_HEAD(pagepool); 1067 1068 DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1069 z_erofs_decompress_queue(bgq, &pagepool); 1070 1071 put_pages_list(&pagepool); 1072 kvfree(bgq); 1073 } 1074 1075 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 1076 unsigned int nr, 1077 struct list_head *pagepool, 1078 struct address_space *mc, 1079 gfp_t gfp) 1080 { 1081 const pgoff_t index = pcl->obj.index; 1082 bool tocache = false; 1083 1084 struct address_space *mapping; 1085 struct page *oldpage, *page; 1086 1087 compressed_page_t t; 1088 int justfound; 1089 1090 repeat: 1091 page = READ_ONCE(pcl->compressed_pages[nr]); 1092 oldpage = page; 1093 1094 if (!page) 1095 goto out_allocpage; 1096 1097 /* 1098 * the cached page has not been allocated and 1099 * an placeholder is out there, prepare it now. 1100 */ 1101 if (page == PAGE_UNALLOCATED) { 1102 tocache = true; 1103 goto out_allocpage; 1104 } 1105 1106 /* process the target tagged pointer */ 1107 t = tagptr_init(compressed_page_t, page); 1108 justfound = tagptr_unfold_tags(t); 1109 page = tagptr_unfold_ptr(t); 1110 1111 /* 1112 * preallocated cached pages, which is used to avoid direct reclaim 1113 * otherwise, it will go inplace I/O path instead. 1114 */ 1115 if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 1116 WRITE_ONCE(pcl->compressed_pages[nr], page); 1117 set_page_private(page, 0); 1118 tocache = true; 1119 goto out_tocache; 1120 } 1121 mapping = READ_ONCE(page->mapping); 1122 1123 /* 1124 * file-backed online pages in plcuster are all locked steady, 1125 * therefore it is impossible for `mapping' to be NULL. 1126 */ 1127 if (mapping && mapping != mc) 1128 /* ought to be unmanaged pages */ 1129 goto out; 1130 1131 /* directly return for shortlived page as well */ 1132 if (z_erofs_is_shortlived_page(page)) 1133 goto out; 1134 1135 lock_page(page); 1136 1137 /* only true if page reclaim goes wrong, should never happen */ 1138 DBG_BUGON(justfound && PagePrivate(page)); 1139 1140 /* the page is still in manage cache */ 1141 if (page->mapping == mc) { 1142 WRITE_ONCE(pcl->compressed_pages[nr], page); 1143 1144 ClearPageError(page); 1145 if (!PagePrivate(page)) { 1146 /* 1147 * impossible to be !PagePrivate(page) for 1148 * the current restriction as well if 1149 * the page is already in compressed_pages[]. 1150 */ 1151 DBG_BUGON(!justfound); 1152 1153 justfound = 0; 1154 set_page_private(page, (unsigned long)pcl); 1155 SetPagePrivate(page); 1156 } 1157 1158 /* no need to submit io if it is already up-to-date */ 1159 if (PageUptodate(page)) { 1160 unlock_page(page); 1161 page = NULL; 1162 } 1163 goto out; 1164 } 1165 1166 /* 1167 * the managed page has been truncated, it's unsafe to 1168 * reuse this one, let's allocate a new cache-managed page. 1169 */ 1170 DBG_BUGON(page->mapping); 1171 DBG_BUGON(!justfound); 1172 1173 tocache = true; 1174 unlock_page(page); 1175 put_page(page); 1176 out_allocpage: 1177 page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 1178 if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) { 1179 list_add(&page->lru, pagepool); 1180 cond_resched(); 1181 goto repeat; 1182 } 1183 out_tocache: 1184 if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1185 /* turn into temporary page if fails (1 ref) */ 1186 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1187 goto out; 1188 } 1189 attach_page_private(page, pcl); 1190 /* drop a refcount added by allocpage (then we have 2 refs here) */ 1191 put_page(page); 1192 1193 out: /* the only exit (for tracing and debugging) */ 1194 return page; 1195 } 1196 1197 static struct z_erofs_decompressqueue * 1198 jobqueue_init(struct super_block *sb, 1199 struct z_erofs_decompressqueue *fgq, bool *fg) 1200 { 1201 struct z_erofs_decompressqueue *q; 1202 1203 if (fg && !*fg) { 1204 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1205 if (!q) { 1206 *fg = true; 1207 goto fg_out; 1208 } 1209 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1210 } else { 1211 fg_out: 1212 q = fgq; 1213 init_waitqueue_head(&fgq->u.wait); 1214 atomic_set(&fgq->pending_bios, 0); 1215 } 1216 q->sb = sb; 1217 q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1218 return q; 1219 } 1220 1221 /* define decompression jobqueue types */ 1222 enum { 1223 JQ_BYPASS, 1224 JQ_SUBMIT, 1225 NR_JOBQUEUES, 1226 }; 1227 1228 static void *jobqueueset_init(struct super_block *sb, 1229 struct z_erofs_decompressqueue *q[], 1230 struct z_erofs_decompressqueue *fgq, bool *fg) 1231 { 1232 /* 1233 * if managed cache is enabled, bypass jobqueue is needed, 1234 * no need to read from device for all pclusters in this queue. 1235 */ 1236 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1237 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg); 1238 1239 return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg)); 1240 } 1241 1242 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1243 z_erofs_next_pcluster_t qtail[], 1244 z_erofs_next_pcluster_t owned_head) 1245 { 1246 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1247 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1248 1249 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1250 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1251 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1252 1253 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 1254 1255 WRITE_ONCE(*submit_qtail, owned_head); 1256 WRITE_ONCE(*bypass_qtail, &pcl->next); 1257 1258 qtail[JQ_BYPASS] = &pcl->next; 1259 } 1260 1261 static void z_erofs_submit_queue(struct super_block *sb, 1262 struct z_erofs_decompress_frontend *f, 1263 struct list_head *pagepool, 1264 struct z_erofs_decompressqueue *fgq, 1265 bool *force_fg) 1266 { 1267 struct erofs_sb_info *const sbi = EROFS_SB(sb); 1268 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1269 struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 1270 void *bi_private; 1271 z_erofs_next_pcluster_t owned_head = f->clt.owned_head; 1272 /* since bio will be NULL, no need to initialize last_index */ 1273 pgoff_t last_index; 1274 unsigned int nr_bios = 0; 1275 struct bio *bio = NULL; 1276 1277 bi_private = jobqueueset_init(sb, q, fgq, force_fg); 1278 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1279 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1280 1281 /* by default, all need io submission */ 1282 q[JQ_SUBMIT]->head = owned_head; 1283 1284 do { 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 cur = pcl->obj.index; 1297 end = cur + pcl->pclusterpages; 1298 1299 /* close the main owned chain at first */ 1300 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1301 Z_EROFS_PCLUSTER_TAIL_CLOSED); 1302 1303 do { 1304 struct page *page; 1305 1306 page = pickup_page_for_submission(pcl, i++, pagepool, 1307 MNGD_MAPPING(sbi), 1308 GFP_NOFS); 1309 if (!page) 1310 continue; 1311 1312 if (bio && cur != last_index + 1) { 1313 submit_bio_retry: 1314 submit_bio(bio); 1315 bio = NULL; 1316 } 1317 1318 if (!bio) { 1319 bio = bio_alloc(GFP_NOIO, BIO_MAX_VECS); 1320 1321 bio->bi_end_io = z_erofs_decompressqueue_endio; 1322 bio_set_dev(bio, sb->s_bdev); 1323 bio->bi_iter.bi_sector = (sector_t)cur << 1324 LOG_SECTORS_PER_BLOCK; 1325 bio->bi_private = bi_private; 1326 bio->bi_opf = REQ_OP_READ; 1327 if (f->readahead) 1328 bio->bi_opf |= REQ_RAHEAD; 1329 ++nr_bios; 1330 } 1331 1332 if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 1333 goto submit_bio_retry; 1334 1335 last_index = cur; 1336 bypass = false; 1337 } while (++cur < end); 1338 1339 if (!bypass) 1340 qtail[JQ_SUBMIT] = &pcl->next; 1341 else 1342 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1343 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1344 1345 if (bio) 1346 submit_bio(bio); 1347 1348 /* 1349 * although background is preferred, no one is pending for submission. 1350 * don't issue workqueue for decompression but drop it directly instead. 1351 */ 1352 if (!*force_fg && !nr_bios) { 1353 kvfree(q[JQ_SUBMIT]); 1354 return; 1355 } 1356 z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios); 1357 } 1358 1359 static void z_erofs_runqueue(struct super_block *sb, 1360 struct z_erofs_decompress_frontend *f, 1361 struct list_head *pagepool, bool force_fg) 1362 { 1363 struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 1364 1365 if (f->clt.owned_head == Z_EROFS_PCLUSTER_TAIL) 1366 return; 1367 z_erofs_submit_queue(sb, f, pagepool, io, &force_fg); 1368 1369 /* handle bypass queue (no i/o pclusters) immediately */ 1370 z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool); 1371 1372 if (!force_fg) 1373 return; 1374 1375 /* wait until all bios are completed */ 1376 io_wait_event(io[JQ_SUBMIT].u.wait, 1377 !atomic_read(&io[JQ_SUBMIT].pending_bios)); 1378 1379 /* handle synchronous decompress queue in the caller context */ 1380 z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool); 1381 } 1382 1383 static int z_erofs_readpage(struct file *file, struct page *page) 1384 { 1385 struct inode *const inode = page->mapping->host; 1386 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1387 int err; 1388 LIST_HEAD(pagepool); 1389 1390 trace_erofs_readpage(page, false); 1391 1392 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1393 1394 err = z_erofs_do_read_page(&f, page, &pagepool); 1395 (void)z_erofs_collector_end(&f.clt); 1396 1397 /* if some compressed cluster ready, need submit them anyway */ 1398 z_erofs_runqueue(inode->i_sb, &f, &pagepool, true); 1399 1400 if (err) 1401 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1402 1403 if (f.map.mpage) 1404 put_page(f.map.mpage); 1405 1406 /* clean up the remaining free pages */ 1407 put_pages_list(&pagepool); 1408 return err; 1409 } 1410 1411 static void z_erofs_readahead(struct readahead_control *rac) 1412 { 1413 struct inode *const inode = rac->mapping->host; 1414 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1415 1416 unsigned int nr_pages = readahead_count(rac); 1417 bool sync = (sbi->ctx.readahead_sync_decompress && 1418 nr_pages <= sbi->ctx.max_sync_decompress_pages); 1419 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1420 struct page *page, *head = NULL; 1421 LIST_HEAD(pagepool); 1422 1423 trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 1424 1425 f.readahead = true; 1426 f.headoffset = readahead_pos(rac); 1427 1428 while ((page = readahead_page(rac))) { 1429 prefetchw(&page->flags); 1430 1431 /* 1432 * A pure asynchronous readahead is indicated if 1433 * a PG_readahead marked page is hitted at first. 1434 * Let's also do asynchronous decompression for this case. 1435 */ 1436 sync &= !(PageReadahead(page) && !head); 1437 1438 set_page_private(page, (unsigned long)head); 1439 head = page; 1440 } 1441 1442 while (head) { 1443 struct page *page = head; 1444 int err; 1445 1446 /* traversal in reverse order */ 1447 head = (void *)page_private(page); 1448 1449 err = z_erofs_do_read_page(&f, page, &pagepool); 1450 if (err) 1451 erofs_err(inode->i_sb, 1452 "readahead error at page %lu @ nid %llu", 1453 page->index, EROFS_I(inode)->nid); 1454 put_page(page); 1455 } 1456 1457 (void)z_erofs_collector_end(&f.clt); 1458 1459 z_erofs_runqueue(inode->i_sb, &f, &pagepool, sync); 1460 1461 if (f.map.mpage) 1462 put_page(f.map.mpage); 1463 1464 /* clean up the remaining free pages */ 1465 put_pages_list(&pagepool); 1466 } 1467 1468 const struct address_space_operations z_erofs_aops = { 1469 .readpage = z_erofs_readpage, 1470 .readahead = z_erofs_readahead, 1471 }; 1472 1473