1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2018 HUAWEI, Inc. 4 * http://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 * a compressed_pages[] placeholder in order to avoid 15 * being filled with file pages for in-place decompression. 16 */ 17 #define PAGE_UNALLOCATED ((void *)0x5F0E4B1D) 18 19 /* how to allocate cached pages for a pcluster */ 20 enum z_erofs_cache_alloctype { 21 DONTALLOC, /* don't allocate any cached pages */ 22 DELAYEDALLOC, /* delayed allocation (at the time of submitting io) */ 23 }; 24 25 /* 26 * tagged pointer with 1-bit tag for all compressed pages 27 * tag 0 - the page is just found with an extra page reference 28 */ 29 typedef tagptr1_t compressed_page_t; 30 31 #define tag_compressed_page_justfound(page) \ 32 tagptr_fold(compressed_page_t, page, 1) 33 34 static struct workqueue_struct *z_erofs_workqueue __read_mostly; 35 static struct kmem_cache *pcluster_cachep __read_mostly; 36 37 void z_erofs_exit_zip_subsystem(void) 38 { 39 destroy_workqueue(z_erofs_workqueue); 40 kmem_cache_destroy(pcluster_cachep); 41 } 42 43 static inline int z_erofs_init_workqueue(void) 44 { 45 const unsigned int onlinecpus = num_possible_cpus(); 46 const unsigned int flags = WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE; 47 48 /* 49 * no need to spawn too many threads, limiting threads could minimum 50 * scheduling overhead, perhaps per-CPU threads should be better? 51 */ 52 z_erofs_workqueue = alloc_workqueue("erofs_unzipd", flags, 53 onlinecpus + onlinecpus / 4); 54 return z_erofs_workqueue ? 0 : -ENOMEM; 55 } 56 57 static void z_erofs_pcluster_init_once(void *ptr) 58 { 59 struct z_erofs_pcluster *pcl = ptr; 60 struct z_erofs_collection *cl = z_erofs_primarycollection(pcl); 61 unsigned int i; 62 63 mutex_init(&cl->lock); 64 cl->nr_pages = 0; 65 cl->vcnt = 0; 66 for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i) 67 pcl->compressed_pages[i] = NULL; 68 } 69 70 static void z_erofs_pcluster_init_always(struct z_erofs_pcluster *pcl) 71 { 72 struct z_erofs_collection *cl = z_erofs_primarycollection(pcl); 73 74 atomic_set(&pcl->obj.refcount, 1); 75 76 DBG_BUGON(cl->nr_pages); 77 DBG_BUGON(cl->vcnt); 78 } 79 80 int __init z_erofs_init_zip_subsystem(void) 81 { 82 pcluster_cachep = kmem_cache_create("erofs_compress", 83 Z_EROFS_WORKGROUP_SIZE, 0, 84 SLAB_RECLAIM_ACCOUNT, 85 z_erofs_pcluster_init_once); 86 if (pcluster_cachep) { 87 if (!z_erofs_init_workqueue()) 88 return 0; 89 90 kmem_cache_destroy(pcluster_cachep); 91 } 92 return -ENOMEM; 93 } 94 95 enum z_erofs_collectmode { 96 COLLECT_SECONDARY, 97 COLLECT_PRIMARY, 98 /* 99 * The current collection was the tail of an exist chain, in addition 100 * that the previous processed chained collections are all decided to 101 * be hooked up to it. 102 * A new chain will be created for the remaining collections which are 103 * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED, 104 * the next collection cannot reuse the whole page safely in 105 * the following scenario: 106 * ________________________________________________________________ 107 * | tail (partial) page | head (partial) page | 108 * | (belongs to the next cl) | (belongs to the current cl) | 109 * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________| 110 */ 111 COLLECT_PRIMARY_HOOKED, 112 COLLECT_PRIMARY_FOLLOWED_NOINPLACE, 113 /* 114 * The current collection has been linked with the owned chain, and 115 * could also be linked with the remaining collections, which means 116 * if the processing page is the tail page of the collection, thus 117 * the current collection can safely use the whole page (since 118 * the previous collection is under control) for in-place I/O, as 119 * illustrated below: 120 * ________________________________________________________________ 121 * | tail (partial) page | head (partial) page | 122 * | (of the current cl) | (of the previous collection) | 123 * | PRIMARY_FOLLOWED or | | 124 * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________| 125 * 126 * [ (*) the above page can be used as inplace I/O. ] 127 */ 128 COLLECT_PRIMARY_FOLLOWED, 129 }; 130 131 struct z_erofs_collector { 132 struct z_erofs_pagevec_ctor vector; 133 134 struct z_erofs_pcluster *pcl, *tailpcl; 135 struct z_erofs_collection *cl; 136 struct page **compressedpages; 137 z_erofs_next_pcluster_t owned_head; 138 139 enum z_erofs_collectmode mode; 140 }; 141 142 struct z_erofs_decompress_frontend { 143 struct inode *const inode; 144 145 struct z_erofs_collector clt; 146 struct erofs_map_blocks map; 147 148 /* used for applying cache strategy on the fly */ 149 bool backmost; 150 erofs_off_t headoffset; 151 }; 152 153 #define COLLECTOR_INIT() { \ 154 .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 155 .mode = COLLECT_PRIMARY_FOLLOWED } 156 157 #define DECOMPRESS_FRONTEND_INIT(__i) { \ 158 .inode = __i, .clt = COLLECTOR_INIT(), \ 159 .backmost = true, } 160 161 static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES]; 162 static DEFINE_MUTEX(z_pagemap_global_lock); 163 164 static void preload_compressed_pages(struct z_erofs_collector *clt, 165 struct address_space *mc, 166 enum z_erofs_cache_alloctype type, 167 struct list_head *pagepool) 168 { 169 const struct z_erofs_pcluster *pcl = clt->pcl; 170 const unsigned int clusterpages = BIT(pcl->clusterbits); 171 struct page **pages = clt->compressedpages; 172 pgoff_t index = pcl->obj.index + (pages - pcl->compressed_pages); 173 bool standalone = true; 174 175 if (clt->mode < COLLECT_PRIMARY_FOLLOWED) 176 return; 177 178 for (; pages < pcl->compressed_pages + clusterpages; ++pages) { 179 struct page *page; 180 compressed_page_t t; 181 182 /* the compressed page was loaded before */ 183 if (READ_ONCE(*pages)) 184 continue; 185 186 page = find_get_page(mc, index); 187 188 if (page) { 189 t = tag_compressed_page_justfound(page); 190 } else if (type == DELAYEDALLOC) { 191 t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED); 192 } else { /* DONTALLOC */ 193 if (standalone) 194 clt->compressedpages = pages; 195 standalone = false; 196 continue; 197 } 198 199 if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t))) 200 continue; 201 202 if (page) 203 put_page(page); 204 } 205 206 if (standalone) /* downgrade to PRIMARY_FOLLOWED_NOINPLACE */ 207 clt->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 208 } 209 210 /* called by erofs_shrinker to get rid of all compressed_pages */ 211 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 212 struct erofs_workgroup *grp) 213 { 214 struct z_erofs_pcluster *const pcl = 215 container_of(grp, struct z_erofs_pcluster, obj); 216 struct address_space *const mapping = MNGD_MAPPING(sbi); 217 const unsigned int clusterpages = BIT(pcl->clusterbits); 218 int i; 219 220 /* 221 * refcount of workgroup is now freezed as 1, 222 * therefore no need to worry about available decompression users. 223 */ 224 for (i = 0; i < clusterpages; ++i) { 225 struct page *page = pcl->compressed_pages[i]; 226 227 if (!page) 228 continue; 229 230 /* block other users from reclaiming or migrating the page */ 231 if (!trylock_page(page)) 232 return -EBUSY; 233 234 if (page->mapping != mapping) 235 continue; 236 237 /* barrier is implied in the following 'unlock_page' */ 238 WRITE_ONCE(pcl->compressed_pages[i], NULL); 239 set_page_private(page, 0); 240 ClearPagePrivate(page); 241 242 unlock_page(page); 243 put_page(page); 244 } 245 return 0; 246 } 247 248 int erofs_try_to_free_cached_page(struct address_space *mapping, 249 struct page *page) 250 { 251 struct z_erofs_pcluster *const pcl = (void *)page_private(page); 252 const unsigned int clusterpages = BIT(pcl->clusterbits); 253 int ret = 0; /* 0 - busy */ 254 255 if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) { 256 unsigned int i; 257 258 for (i = 0; i < clusterpages; ++i) { 259 if (pcl->compressed_pages[i] == page) { 260 WRITE_ONCE(pcl->compressed_pages[i], NULL); 261 ret = 1; 262 break; 263 } 264 } 265 erofs_workgroup_unfreeze(&pcl->obj, 1); 266 267 if (ret) { 268 ClearPagePrivate(page); 269 put_page(page); 270 } 271 } 272 return ret; 273 } 274 275 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */ 276 static inline bool z_erofs_try_inplace_io(struct z_erofs_collector *clt, 277 struct page *page) 278 { 279 struct z_erofs_pcluster *const pcl = clt->pcl; 280 const unsigned int clusterpages = BIT(pcl->clusterbits); 281 282 while (clt->compressedpages < pcl->compressed_pages + clusterpages) { 283 if (!cmpxchg(clt->compressedpages++, NULL, page)) 284 return true; 285 } 286 return false; 287 } 288 289 /* callers must be with collection lock held */ 290 static int z_erofs_attach_page(struct z_erofs_collector *clt, 291 struct page *page, 292 enum z_erofs_page_type type) 293 { 294 int ret; 295 bool occupied; 296 297 /* give priority for inplaceio */ 298 if (clt->mode >= COLLECT_PRIMARY && 299 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE && 300 z_erofs_try_inplace_io(clt, page)) 301 return 0; 302 303 ret = z_erofs_pagevec_enqueue(&clt->vector, 304 page, type, &occupied); 305 clt->cl->vcnt += (unsigned int)ret; 306 307 return ret ? 0 : -EAGAIN; 308 } 309 310 static enum z_erofs_collectmode 311 try_to_claim_pcluster(struct z_erofs_pcluster *pcl, 312 z_erofs_next_pcluster_t *owned_head) 313 { 314 /* let's claim these following types of pclusters */ 315 retry: 316 if (pcl->next == Z_EROFS_PCLUSTER_NIL) { 317 /* type 1, nil pcluster */ 318 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 319 *owned_head) != Z_EROFS_PCLUSTER_NIL) 320 goto retry; 321 322 *owned_head = &pcl->next; 323 /* lucky, I am the followee :) */ 324 return COLLECT_PRIMARY_FOLLOWED; 325 } else if (pcl->next == Z_EROFS_PCLUSTER_TAIL) { 326 /* 327 * type 2, link to the end of a existing open chain, 328 * be careful that its submission itself is governed 329 * by the original owned chain. 330 */ 331 if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 332 *owned_head) != Z_EROFS_PCLUSTER_TAIL) 333 goto retry; 334 *owned_head = Z_EROFS_PCLUSTER_TAIL; 335 return COLLECT_PRIMARY_HOOKED; 336 } 337 return COLLECT_PRIMARY; /* :( better luck next time */ 338 } 339 340 static int z_erofs_lookup_collection(struct z_erofs_collector *clt, 341 struct inode *inode, 342 struct erofs_map_blocks *map) 343 { 344 struct erofs_workgroup *grp; 345 struct z_erofs_pcluster *pcl; 346 struct z_erofs_collection *cl; 347 unsigned int length; 348 bool tag; 349 350 grp = erofs_find_workgroup(inode->i_sb, map->m_pa >> PAGE_SHIFT, &tag); 351 if (!grp) 352 return -ENOENT; 353 354 pcl = container_of(grp, struct z_erofs_pcluster, obj); 355 if (clt->owned_head == &pcl->next || pcl == clt->tailpcl) { 356 DBG_BUGON(1); 357 erofs_workgroup_put(grp); 358 return -EFSCORRUPTED; 359 } 360 361 cl = z_erofs_primarycollection(pcl); 362 if (cl->pageofs != (map->m_la & ~PAGE_MASK)) { 363 DBG_BUGON(1); 364 erofs_workgroup_put(grp); 365 return -EFSCORRUPTED; 366 } 367 368 length = READ_ONCE(pcl->length); 369 if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) { 370 if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) { 371 DBG_BUGON(1); 372 erofs_workgroup_put(grp); 373 return -EFSCORRUPTED; 374 } 375 } else { 376 unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT; 377 378 if (map->m_flags & EROFS_MAP_FULL_MAPPED) 379 llen |= Z_EROFS_PCLUSTER_FULL_LENGTH; 380 381 while (llen > length && 382 length != cmpxchg_relaxed(&pcl->length, length, llen)) { 383 cpu_relax(); 384 length = READ_ONCE(pcl->length); 385 } 386 } 387 mutex_lock(&cl->lock); 388 /* used to check tail merging loop due to corrupted images */ 389 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 390 clt->tailpcl = pcl; 391 clt->mode = try_to_claim_pcluster(pcl, &clt->owned_head); 392 /* clean tailpcl if the current owned_head is Z_EROFS_PCLUSTER_TAIL */ 393 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 394 clt->tailpcl = NULL; 395 clt->pcl = pcl; 396 clt->cl = cl; 397 return 0; 398 } 399 400 static int z_erofs_register_collection(struct z_erofs_collector *clt, 401 struct inode *inode, 402 struct erofs_map_blocks *map) 403 { 404 struct z_erofs_pcluster *pcl; 405 struct z_erofs_collection *cl; 406 int err; 407 408 /* no available workgroup, let's allocate one */ 409 pcl = kmem_cache_alloc(pcluster_cachep, GFP_NOFS); 410 if (!pcl) 411 return -ENOMEM; 412 413 z_erofs_pcluster_init_always(pcl); 414 pcl->obj.index = map->m_pa >> PAGE_SHIFT; 415 416 pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) | 417 (map->m_flags & EROFS_MAP_FULL_MAPPED ? 418 Z_EROFS_PCLUSTER_FULL_LENGTH : 0); 419 420 if (map->m_flags & EROFS_MAP_ZIPPED) 421 pcl->algorithmformat = Z_EROFS_COMPRESSION_LZ4; 422 else 423 pcl->algorithmformat = Z_EROFS_COMPRESSION_SHIFTED; 424 425 pcl->clusterbits = EROFS_I(inode)->z_physical_clusterbits[0]; 426 pcl->clusterbits -= PAGE_SHIFT; 427 428 /* new pclusters should be claimed as type 1, primary and followed */ 429 pcl->next = clt->owned_head; 430 clt->mode = COLLECT_PRIMARY_FOLLOWED; 431 432 cl = z_erofs_primarycollection(pcl); 433 cl->pageofs = map->m_la & ~PAGE_MASK; 434 435 /* 436 * lock all primary followed works before visible to others 437 * and mutex_trylock *never* fails for a new pcluster. 438 */ 439 mutex_trylock(&cl->lock); 440 441 err = erofs_register_workgroup(inode->i_sb, &pcl->obj, 0); 442 if (err) { 443 mutex_unlock(&cl->lock); 444 kmem_cache_free(pcluster_cachep, pcl); 445 return -EAGAIN; 446 } 447 /* used to check tail merging loop due to corrupted images */ 448 if (clt->owned_head == Z_EROFS_PCLUSTER_TAIL) 449 clt->tailpcl = pcl; 450 clt->owned_head = &pcl->next; 451 clt->pcl = pcl; 452 clt->cl = cl; 453 return 0; 454 } 455 456 static int z_erofs_collector_begin(struct z_erofs_collector *clt, 457 struct inode *inode, 458 struct erofs_map_blocks *map) 459 { 460 int ret; 461 462 DBG_BUGON(clt->cl); 463 464 /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous collection */ 465 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_NIL); 466 DBG_BUGON(clt->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 467 468 if (!PAGE_ALIGNED(map->m_pa)) { 469 DBG_BUGON(1); 470 return -EINVAL; 471 } 472 473 repeat: 474 ret = z_erofs_lookup_collection(clt, inode, map); 475 if (ret == -ENOENT) { 476 ret = z_erofs_register_collection(clt, inode, map); 477 478 /* someone registered at the same time, give another try */ 479 if (ret == -EAGAIN) { 480 cond_resched(); 481 goto repeat; 482 } 483 } 484 485 if (ret) 486 return ret; 487 488 z_erofs_pagevec_ctor_init(&clt->vector, Z_EROFS_NR_INLINE_PAGEVECS, 489 clt->cl->pagevec, clt->cl->vcnt); 490 491 clt->compressedpages = clt->pcl->compressed_pages; 492 if (clt->mode <= COLLECT_PRIMARY) /* cannot do in-place I/O */ 493 clt->compressedpages += Z_EROFS_CLUSTER_MAX_PAGES; 494 return 0; 495 } 496 497 /* 498 * keep in mind that no referenced pclusters will be freed 499 * only after a RCU grace period. 500 */ 501 static void z_erofs_rcu_callback(struct rcu_head *head) 502 { 503 struct z_erofs_collection *const cl = 504 container_of(head, struct z_erofs_collection, rcu); 505 506 kmem_cache_free(pcluster_cachep, 507 container_of(cl, struct z_erofs_pcluster, 508 primary_collection)); 509 } 510 511 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 512 { 513 struct z_erofs_pcluster *const pcl = 514 container_of(grp, struct z_erofs_pcluster, obj); 515 struct z_erofs_collection *const cl = z_erofs_primarycollection(pcl); 516 517 call_rcu(&cl->rcu, z_erofs_rcu_callback); 518 } 519 520 static void z_erofs_collection_put(struct z_erofs_collection *cl) 521 { 522 struct z_erofs_pcluster *const pcl = 523 container_of(cl, struct z_erofs_pcluster, primary_collection); 524 525 erofs_workgroup_put(&pcl->obj); 526 } 527 528 static bool z_erofs_collector_end(struct z_erofs_collector *clt) 529 { 530 struct z_erofs_collection *cl = clt->cl; 531 532 if (!cl) 533 return false; 534 535 z_erofs_pagevec_ctor_exit(&clt->vector, false); 536 mutex_unlock(&cl->lock); 537 538 /* 539 * if all pending pages are added, don't hold its reference 540 * any longer if the pcluster isn't hosted by ourselves. 541 */ 542 if (clt->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE) 543 z_erofs_collection_put(cl); 544 545 clt->cl = NULL; 546 return true; 547 } 548 549 static inline struct page *__stagingpage_alloc(struct list_head *pagepool, 550 gfp_t gfp) 551 { 552 struct page *page = erofs_allocpage(pagepool, gfp, true); 553 554 page->mapping = Z_EROFS_MAPPING_STAGING; 555 return page; 556 } 557 558 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe, 559 unsigned int cachestrategy, 560 erofs_off_t la) 561 { 562 if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 563 return false; 564 565 if (fe->backmost) 566 return true; 567 568 return cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 569 la < fe->headoffset; 570 } 571 572 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 573 struct page *page, 574 struct list_head *pagepool) 575 { 576 struct inode *const inode = fe->inode; 577 struct erofs_sb_info *const sbi __maybe_unused = EROFS_I_SB(inode); 578 struct erofs_map_blocks *const map = &fe->map; 579 struct z_erofs_collector *const clt = &fe->clt; 580 const loff_t offset = page_offset(page); 581 bool tight = true; 582 583 enum z_erofs_cache_alloctype cache_strategy; 584 enum z_erofs_page_type page_type; 585 unsigned int cur, end, spiltted, index; 586 int err = 0; 587 588 /* register locked file pages as online pages in pack */ 589 z_erofs_onlinepage_init(page); 590 591 spiltted = 0; 592 end = PAGE_SIZE; 593 repeat: 594 cur = end - 1; 595 596 /* lucky, within the range of the current map_blocks */ 597 if (offset + cur >= map->m_la && 598 offset + cur < map->m_la + map->m_llen) { 599 /* didn't get a valid collection previously (very rare) */ 600 if (!clt->cl) 601 goto restart_now; 602 goto hitted; 603 } 604 605 /* go ahead the next map_blocks */ 606 erofs_dbg("%s: [out-of-range] pos %llu", __func__, offset + cur); 607 608 if (z_erofs_collector_end(clt)) 609 fe->backmost = false; 610 611 map->m_la = offset + cur; 612 map->m_llen = 0; 613 err = z_erofs_map_blocks_iter(inode, map, 0); 614 if (err) 615 goto err_out; 616 617 restart_now: 618 if (!(map->m_flags & EROFS_MAP_MAPPED)) 619 goto hitted; 620 621 err = z_erofs_collector_begin(clt, inode, map); 622 if (err) 623 goto err_out; 624 625 /* preload all compressed pages (maybe downgrade role if necessary) */ 626 if (should_alloc_managed_pages(fe, sbi->cache_strategy, map->m_la)) 627 cache_strategy = DELAYEDALLOC; 628 else 629 cache_strategy = DONTALLOC; 630 631 preload_compressed_pages(clt, MNGD_MAPPING(sbi), 632 cache_strategy, pagepool); 633 634 hitted: 635 /* 636 * Ensure the current partial page belongs to this submit chain rather 637 * than other concurrent submit chains or the noio(bypass) chain since 638 * those chains are handled asynchronously thus the page cannot be used 639 * for inplace I/O or pagevec (should be processed in strict order.) 640 */ 641 tight &= (clt->mode >= COLLECT_PRIMARY_HOOKED && 642 clt->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE); 643 644 cur = end - min_t(unsigned int, offset + end - map->m_la, end); 645 if (!(map->m_flags & EROFS_MAP_MAPPED)) { 646 zero_user_segment(page, cur, end); 647 goto next_part; 648 } 649 650 /* let's derive page type */ 651 page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD : 652 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 653 (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 654 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED)); 655 656 if (cur) 657 tight &= (clt->mode >= COLLECT_PRIMARY_FOLLOWED); 658 659 retry: 660 err = z_erofs_attach_page(clt, page, page_type); 661 /* should allocate an additional staging page for pagevec */ 662 if (err == -EAGAIN) { 663 struct page *const newpage = 664 __stagingpage_alloc(pagepool, GFP_NOFS); 665 666 err = z_erofs_attach_page(clt, newpage, 667 Z_EROFS_PAGE_TYPE_EXCLUSIVE); 668 if (!err) 669 goto retry; 670 } 671 672 if (err) 673 goto err_out; 674 675 index = page->index - (map->m_la >> PAGE_SHIFT); 676 677 z_erofs_onlinepage_fixup(page, index, true); 678 679 /* bump up the number of spiltted parts of a page */ 680 ++spiltted; 681 /* also update nr_pages */ 682 clt->cl->nr_pages = max_t(pgoff_t, clt->cl->nr_pages, index + 1); 683 next_part: 684 /* can be used for verification */ 685 map->m_llen = offset + cur - map->m_la; 686 687 end = cur; 688 if (end > 0) 689 goto repeat; 690 691 out: 692 z_erofs_onlinepage_endio(page); 693 694 erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 695 __func__, page, spiltted, map->m_llen); 696 return err; 697 698 /* if some error occurred while processing this page */ 699 err_out: 700 SetPageError(page); 701 goto out; 702 } 703 704 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios) 705 { 706 tagptr1_t t = tagptr_init(tagptr1_t, ptr); 707 struct z_erofs_unzip_io *io = tagptr_unfold_ptr(t); 708 bool background = tagptr_unfold_tags(t); 709 710 if (!background) { 711 unsigned long flags; 712 713 spin_lock_irqsave(&io->u.wait.lock, flags); 714 if (!atomic_add_return(bios, &io->pending_bios)) 715 wake_up_locked(&io->u.wait); 716 spin_unlock_irqrestore(&io->u.wait.lock, flags); 717 return; 718 } 719 720 if (!atomic_add_return(bios, &io->pending_bios)) 721 queue_work(z_erofs_workqueue, &io->u.work); 722 } 723 724 static inline void z_erofs_vle_read_endio(struct bio *bio) 725 { 726 struct erofs_sb_info *sbi = NULL; 727 blk_status_t err = bio->bi_status; 728 struct bio_vec *bvec; 729 struct bvec_iter_all iter_all; 730 731 bio_for_each_segment_all(bvec, bio, iter_all) { 732 struct page *page = bvec->bv_page; 733 bool cachemngd = false; 734 735 DBG_BUGON(PageUptodate(page)); 736 DBG_BUGON(!page->mapping); 737 738 if (!sbi && !z_erofs_page_is_staging(page)) 739 sbi = EROFS_SB(page->mapping->host->i_sb); 740 741 /* sbi should already be gotten if the page is managed */ 742 if (sbi) 743 cachemngd = erofs_page_is_managed(sbi, page); 744 745 if (err) 746 SetPageError(page); 747 else if (cachemngd) 748 SetPageUptodate(page); 749 750 if (cachemngd) 751 unlock_page(page); 752 } 753 754 z_erofs_vle_unzip_kickoff(bio->bi_private, -1); 755 bio_put(bio); 756 } 757 758 static int z_erofs_decompress_pcluster(struct super_block *sb, 759 struct z_erofs_pcluster *pcl, 760 struct list_head *pagepool) 761 { 762 struct erofs_sb_info *const sbi = EROFS_SB(sb); 763 const unsigned int clusterpages = BIT(pcl->clusterbits); 764 struct z_erofs_pagevec_ctor ctor; 765 unsigned int i, outputsize, llen, nr_pages; 766 struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES]; 767 struct page **pages, **compressed_pages, *page; 768 769 enum z_erofs_page_type page_type; 770 bool overlapped, partial; 771 struct z_erofs_collection *cl; 772 int err; 773 774 might_sleep(); 775 cl = z_erofs_primarycollection(pcl); 776 DBG_BUGON(!READ_ONCE(cl->nr_pages)); 777 778 mutex_lock(&cl->lock); 779 nr_pages = cl->nr_pages; 780 781 if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) { 782 pages = pages_onstack; 783 } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES && 784 mutex_trylock(&z_pagemap_global_lock)) { 785 pages = z_pagemap_global; 786 } else { 787 gfp_t gfp_flags = GFP_KERNEL; 788 789 if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES) 790 gfp_flags |= __GFP_NOFAIL; 791 792 pages = kvmalloc_array(nr_pages, sizeof(struct page *), 793 gfp_flags); 794 795 /* fallback to global pagemap for the lowmem scenario */ 796 if (!pages) { 797 mutex_lock(&z_pagemap_global_lock); 798 pages = z_pagemap_global; 799 } 800 } 801 802 for (i = 0; i < nr_pages; ++i) 803 pages[i] = NULL; 804 805 err = 0; 806 z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS, 807 cl->pagevec, 0); 808 809 for (i = 0; i < cl->vcnt; ++i) { 810 unsigned int pagenr; 811 812 page = z_erofs_pagevec_dequeue(&ctor, &page_type); 813 814 /* all pages in pagevec ought to be valid */ 815 DBG_BUGON(!page); 816 DBG_BUGON(!page->mapping); 817 818 if (z_erofs_put_stagingpage(pagepool, page)) 819 continue; 820 821 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD) 822 pagenr = 0; 823 else 824 pagenr = z_erofs_onlinepage_index(page); 825 826 DBG_BUGON(pagenr >= nr_pages); 827 828 /* 829 * currently EROFS doesn't support multiref(dedup), 830 * so here erroring out one multiref page. 831 */ 832 if (pages[pagenr]) { 833 DBG_BUGON(1); 834 SetPageError(pages[pagenr]); 835 z_erofs_onlinepage_endio(pages[pagenr]); 836 err = -EFSCORRUPTED; 837 } 838 pages[pagenr] = page; 839 } 840 z_erofs_pagevec_ctor_exit(&ctor, true); 841 842 overlapped = false; 843 compressed_pages = pcl->compressed_pages; 844 845 for (i = 0; i < clusterpages; ++i) { 846 unsigned int pagenr; 847 848 page = compressed_pages[i]; 849 850 /* all compressed pages ought to be valid */ 851 DBG_BUGON(!page); 852 DBG_BUGON(!page->mapping); 853 854 if (!z_erofs_page_is_staging(page)) { 855 if (erofs_page_is_managed(sbi, page)) { 856 if (!PageUptodate(page)) 857 err = -EIO; 858 continue; 859 } 860 861 /* 862 * only if non-head page can be selected 863 * for inplace decompression 864 */ 865 pagenr = z_erofs_onlinepage_index(page); 866 867 DBG_BUGON(pagenr >= nr_pages); 868 if (pages[pagenr]) { 869 DBG_BUGON(1); 870 SetPageError(pages[pagenr]); 871 z_erofs_onlinepage_endio(pages[pagenr]); 872 err = -EFSCORRUPTED; 873 } 874 pages[pagenr] = page; 875 876 overlapped = true; 877 } 878 879 /* PG_error needs checking for inplaced and staging pages */ 880 if (PageError(page)) { 881 DBG_BUGON(PageUptodate(page)); 882 err = -EIO; 883 } 884 } 885 886 if (err) 887 goto out; 888 889 llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT; 890 if (nr_pages << PAGE_SHIFT >= cl->pageofs + llen) { 891 outputsize = llen; 892 partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH); 893 } else { 894 outputsize = (nr_pages << PAGE_SHIFT) - cl->pageofs; 895 partial = true; 896 } 897 898 err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 899 .sb = sb, 900 .in = compressed_pages, 901 .out = pages, 902 .pageofs_out = cl->pageofs, 903 .inputsize = PAGE_SIZE, 904 .outputsize = outputsize, 905 .alg = pcl->algorithmformat, 906 .inplace_io = overlapped, 907 .partial_decoding = partial 908 }, pagepool); 909 910 out: 911 /* must handle all compressed pages before endding pages */ 912 for (i = 0; i < clusterpages; ++i) { 913 page = compressed_pages[i]; 914 915 if (erofs_page_is_managed(sbi, page)) 916 continue; 917 918 /* recycle all individual staging pages */ 919 (void)z_erofs_put_stagingpage(pagepool, page); 920 921 WRITE_ONCE(compressed_pages[i], NULL); 922 } 923 924 for (i = 0; i < nr_pages; ++i) { 925 page = pages[i]; 926 if (!page) 927 continue; 928 929 DBG_BUGON(!page->mapping); 930 931 /* recycle all individual staging pages */ 932 if (z_erofs_put_stagingpage(pagepool, page)) 933 continue; 934 935 if (err < 0) 936 SetPageError(page); 937 938 z_erofs_onlinepage_endio(page); 939 } 940 941 if (pages == z_pagemap_global) 942 mutex_unlock(&z_pagemap_global_lock); 943 else if (pages != pages_onstack) 944 kvfree(pages); 945 946 cl->nr_pages = 0; 947 cl->vcnt = 0; 948 949 /* all cl locks MUST be taken before the following line */ 950 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 951 952 /* all cl locks SHOULD be released right now */ 953 mutex_unlock(&cl->lock); 954 955 z_erofs_collection_put(cl); 956 return err; 957 } 958 959 static void z_erofs_vle_unzip_all(struct super_block *sb, 960 struct z_erofs_unzip_io *io, 961 struct list_head *pagepool) 962 { 963 z_erofs_next_pcluster_t owned = io->head; 964 965 while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 966 struct z_erofs_pcluster *pcl; 967 968 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 969 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 970 971 /* no possible that 'owned' equals NULL */ 972 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 973 974 pcl = container_of(owned, struct z_erofs_pcluster, next); 975 owned = READ_ONCE(pcl->next); 976 977 z_erofs_decompress_pcluster(sb, pcl, pagepool); 978 } 979 } 980 981 static void z_erofs_vle_unzip_wq(struct work_struct *work) 982 { 983 struct z_erofs_unzip_io_sb *iosb = 984 container_of(work, struct z_erofs_unzip_io_sb, io.u.work); 985 LIST_HEAD(pagepool); 986 987 DBG_BUGON(iosb->io.head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 988 z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &pagepool); 989 990 put_pages_list(&pagepool); 991 kvfree(iosb); 992 } 993 994 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 995 unsigned int nr, 996 struct list_head *pagepool, 997 struct address_space *mc, 998 gfp_t gfp) 999 { 1000 /* determined at compile time to avoid too many #ifdefs */ 1001 const bool nocache = __builtin_constant_p(mc) ? !mc : false; 1002 const pgoff_t index = pcl->obj.index; 1003 bool tocache = false; 1004 1005 struct address_space *mapping; 1006 struct page *oldpage, *page; 1007 1008 compressed_page_t t; 1009 int justfound; 1010 1011 repeat: 1012 page = READ_ONCE(pcl->compressed_pages[nr]); 1013 oldpage = page; 1014 1015 if (!page) 1016 goto out_allocpage; 1017 1018 /* 1019 * the cached page has not been allocated and 1020 * an placeholder is out there, prepare it now. 1021 */ 1022 if (!nocache && page == PAGE_UNALLOCATED) { 1023 tocache = true; 1024 goto out_allocpage; 1025 } 1026 1027 /* process the target tagged pointer */ 1028 t = tagptr_init(compressed_page_t, page); 1029 justfound = tagptr_unfold_tags(t); 1030 page = tagptr_unfold_ptr(t); 1031 1032 mapping = READ_ONCE(page->mapping); 1033 1034 /* 1035 * if managed cache is disabled, it's no way to 1036 * get such a cached-like page. 1037 */ 1038 if (nocache) { 1039 /* if managed cache is disabled, it is impossible `justfound' */ 1040 DBG_BUGON(justfound); 1041 1042 /* and it should be locked, not uptodate, and not truncated */ 1043 DBG_BUGON(!PageLocked(page)); 1044 DBG_BUGON(PageUptodate(page)); 1045 DBG_BUGON(!mapping); 1046 goto out; 1047 } 1048 1049 /* 1050 * unmanaged (file) pages are all locked solidly, 1051 * therefore it is impossible for `mapping' to be NULL. 1052 */ 1053 if (mapping && mapping != mc) 1054 /* ought to be unmanaged pages */ 1055 goto out; 1056 1057 lock_page(page); 1058 1059 /* only true if page reclaim goes wrong, should never happen */ 1060 DBG_BUGON(justfound && PagePrivate(page)); 1061 1062 /* the page is still in manage cache */ 1063 if (page->mapping == mc) { 1064 WRITE_ONCE(pcl->compressed_pages[nr], page); 1065 1066 ClearPageError(page); 1067 if (!PagePrivate(page)) { 1068 /* 1069 * impossible to be !PagePrivate(page) for 1070 * the current restriction as well if 1071 * the page is already in compressed_pages[]. 1072 */ 1073 DBG_BUGON(!justfound); 1074 1075 justfound = 0; 1076 set_page_private(page, (unsigned long)pcl); 1077 SetPagePrivate(page); 1078 } 1079 1080 /* no need to submit io if it is already up-to-date */ 1081 if (PageUptodate(page)) { 1082 unlock_page(page); 1083 page = NULL; 1084 } 1085 goto out; 1086 } 1087 1088 /* 1089 * the managed page has been truncated, it's unsafe to 1090 * reuse this one, let's allocate a new cache-managed page. 1091 */ 1092 DBG_BUGON(page->mapping); 1093 DBG_BUGON(!justfound); 1094 1095 tocache = true; 1096 unlock_page(page); 1097 put_page(page); 1098 out_allocpage: 1099 page = __stagingpage_alloc(pagepool, gfp); 1100 if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) { 1101 list_add(&page->lru, pagepool); 1102 cpu_relax(); 1103 goto repeat; 1104 } 1105 if (nocache || !tocache) 1106 goto out; 1107 if (add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1108 page->mapping = Z_EROFS_MAPPING_STAGING; 1109 goto out; 1110 } 1111 1112 set_page_private(page, (unsigned long)pcl); 1113 SetPagePrivate(page); 1114 out: /* the only exit (for tracing and debugging) */ 1115 return page; 1116 } 1117 1118 static struct z_erofs_unzip_io *jobqueue_init(struct super_block *sb, 1119 struct z_erofs_unzip_io *io, 1120 bool foreground) 1121 { 1122 struct z_erofs_unzip_io_sb *iosb; 1123 1124 if (foreground) { 1125 /* waitqueue available for foreground io */ 1126 DBG_BUGON(!io); 1127 1128 init_waitqueue_head(&io->u.wait); 1129 atomic_set(&io->pending_bios, 0); 1130 goto out; 1131 } 1132 1133 iosb = kvzalloc(sizeof(*iosb), GFP_KERNEL | __GFP_NOFAIL); 1134 DBG_BUGON(!iosb); 1135 1136 /* initialize fields in the allocated descriptor */ 1137 io = &iosb->io; 1138 iosb->sb = sb; 1139 INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq); 1140 out: 1141 io->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1142 return io; 1143 } 1144 1145 /* define decompression jobqueue types */ 1146 enum { 1147 JQ_BYPASS, 1148 JQ_SUBMIT, 1149 NR_JOBQUEUES, 1150 }; 1151 1152 static void *jobqueueset_init(struct super_block *sb, 1153 z_erofs_next_pcluster_t qtail[], 1154 struct z_erofs_unzip_io *q[], 1155 struct z_erofs_unzip_io *fgq, 1156 bool forcefg) 1157 { 1158 /* 1159 * if managed cache is enabled, bypass jobqueue is needed, 1160 * no need to read from device for all pclusters in this queue. 1161 */ 1162 q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true); 1163 qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1164 1165 q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg); 1166 qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 1167 1168 return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg)); 1169 } 1170 1171 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 1172 z_erofs_next_pcluster_t qtail[], 1173 z_erofs_next_pcluster_t owned_head) 1174 { 1175 z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 1176 z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 1177 1178 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1179 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1180 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1181 1182 WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 1183 1184 WRITE_ONCE(*submit_qtail, owned_head); 1185 WRITE_ONCE(*bypass_qtail, &pcl->next); 1186 1187 qtail[JQ_BYPASS] = &pcl->next; 1188 } 1189 1190 static bool postsubmit_is_all_bypassed(struct z_erofs_unzip_io *q[], 1191 unsigned int nr_bios, 1192 bool force_fg) 1193 { 1194 /* 1195 * although background is preferred, no one is pending for submission. 1196 * don't issue workqueue for decompression but drop it directly instead. 1197 */ 1198 if (force_fg || nr_bios) 1199 return false; 1200 1201 kvfree(container_of(q[JQ_SUBMIT], struct z_erofs_unzip_io_sb, io)); 1202 return true; 1203 } 1204 1205 static bool z_erofs_vle_submit_all(struct super_block *sb, 1206 z_erofs_next_pcluster_t owned_head, 1207 struct list_head *pagepool, 1208 struct z_erofs_unzip_io *fgq, 1209 bool force_fg) 1210 { 1211 struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb); 1212 z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1213 struct z_erofs_unzip_io *q[NR_JOBQUEUES]; 1214 struct bio *bio; 1215 void *bi_private; 1216 /* since bio will be NULL, no need to initialize last_index */ 1217 pgoff_t uninitialized_var(last_index); 1218 bool force_submit = false; 1219 unsigned int nr_bios; 1220 1221 if (owned_head == Z_EROFS_PCLUSTER_TAIL) 1222 return false; 1223 1224 force_submit = false; 1225 bio = NULL; 1226 nr_bios = 0; 1227 bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg); 1228 1229 /* by default, all need io submission */ 1230 q[JQ_SUBMIT]->head = owned_head; 1231 1232 do { 1233 struct z_erofs_pcluster *pcl; 1234 unsigned int clusterpages; 1235 pgoff_t first_index; 1236 struct page *page; 1237 unsigned int i = 0, bypass = 0; 1238 int err; 1239 1240 /* no possible 'owned_head' equals the following */ 1241 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 1242 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 1243 1244 pcl = container_of(owned_head, struct z_erofs_pcluster, next); 1245 1246 clusterpages = BIT(pcl->clusterbits); 1247 1248 /* close the main owned chain at first */ 1249 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1250 Z_EROFS_PCLUSTER_TAIL_CLOSED); 1251 1252 first_index = pcl->obj.index; 1253 force_submit |= (first_index != last_index + 1); 1254 1255 repeat: 1256 page = pickup_page_for_submission(pcl, i, pagepool, 1257 MNGD_MAPPING(sbi), 1258 GFP_NOFS); 1259 if (!page) { 1260 force_submit = true; 1261 ++bypass; 1262 goto skippage; 1263 } 1264 1265 if (bio && force_submit) { 1266 submit_bio_retry: 1267 submit_bio(bio); 1268 bio = NULL; 1269 } 1270 1271 if (!bio) { 1272 bio = bio_alloc(GFP_NOIO, BIO_MAX_PAGES); 1273 1274 bio->bi_end_io = z_erofs_vle_read_endio; 1275 bio_set_dev(bio, sb->s_bdev); 1276 bio->bi_iter.bi_sector = (sector_t)(first_index + i) << 1277 LOG_SECTORS_PER_BLOCK; 1278 bio->bi_private = bi_private; 1279 bio->bi_opf = REQ_OP_READ; 1280 1281 ++nr_bios; 1282 } 1283 1284 err = bio_add_page(bio, page, PAGE_SIZE, 0); 1285 if (err < PAGE_SIZE) 1286 goto submit_bio_retry; 1287 1288 force_submit = false; 1289 last_index = first_index + i; 1290 skippage: 1291 if (++i < clusterpages) 1292 goto repeat; 1293 1294 if (bypass < clusterpages) 1295 qtail[JQ_SUBMIT] = &pcl->next; 1296 else 1297 move_to_bypass_jobqueue(pcl, qtail, owned_head); 1298 } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 1299 1300 if (bio) 1301 submit_bio(bio); 1302 1303 if (postsubmit_is_all_bypassed(q, nr_bios, force_fg)) 1304 return true; 1305 1306 z_erofs_vle_unzip_kickoff(bi_private, nr_bios); 1307 return true; 1308 } 1309 1310 static void z_erofs_submit_and_unzip(struct super_block *sb, 1311 struct z_erofs_collector *clt, 1312 struct list_head *pagepool, 1313 bool force_fg) 1314 { 1315 struct z_erofs_unzip_io io[NR_JOBQUEUES]; 1316 1317 if (!z_erofs_vle_submit_all(sb, clt->owned_head, 1318 pagepool, io, force_fg)) 1319 return; 1320 1321 /* decompress no I/O pclusters immediately */ 1322 z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool); 1323 1324 if (!force_fg) 1325 return; 1326 1327 /* wait until all bios are completed */ 1328 wait_event(io[JQ_SUBMIT].u.wait, 1329 !atomic_read(&io[JQ_SUBMIT].pending_bios)); 1330 1331 /* let's synchronous decompression */ 1332 z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool); 1333 } 1334 1335 static int z_erofs_vle_normalaccess_readpage(struct file *file, 1336 struct page *page) 1337 { 1338 struct inode *const inode = page->mapping->host; 1339 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1340 int err; 1341 LIST_HEAD(pagepool); 1342 1343 trace_erofs_readpage(page, false); 1344 1345 f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 1346 1347 err = z_erofs_do_read_page(&f, page, &pagepool); 1348 (void)z_erofs_collector_end(&f.clt); 1349 1350 /* if some compressed cluster ready, need submit them anyway */ 1351 z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, true); 1352 1353 if (err) 1354 erofs_err(inode->i_sb, "failed to read, err [%d]", err); 1355 1356 if (f.map.mpage) 1357 put_page(f.map.mpage); 1358 1359 /* clean up the remaining free pages */ 1360 put_pages_list(&pagepool); 1361 return err; 1362 } 1363 1364 static bool should_decompress_synchronously(struct erofs_sb_info *sbi, 1365 unsigned int nr) 1366 { 1367 return nr <= sbi->max_sync_decompress_pages; 1368 } 1369 1370 static int z_erofs_vle_normalaccess_readpages(struct file *filp, 1371 struct address_space *mapping, 1372 struct list_head *pages, 1373 unsigned int nr_pages) 1374 { 1375 struct inode *const inode = mapping->host; 1376 struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 1377 1378 bool sync = should_decompress_synchronously(sbi, nr_pages); 1379 struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1380 gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL); 1381 struct page *head = NULL; 1382 LIST_HEAD(pagepool); 1383 1384 trace_erofs_readpages(mapping->host, lru_to_page(pages), 1385 nr_pages, false); 1386 1387 f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT; 1388 1389 for (; nr_pages; --nr_pages) { 1390 struct page *page = lru_to_page(pages); 1391 1392 prefetchw(&page->flags); 1393 list_del(&page->lru); 1394 1395 /* 1396 * A pure asynchronous readahead is indicated if 1397 * a PG_readahead marked page is hitted at first. 1398 * Let's also do asynchronous decompression for this case. 1399 */ 1400 sync &= !(PageReadahead(page) && !head); 1401 1402 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) { 1403 list_add(&page->lru, &pagepool); 1404 continue; 1405 } 1406 1407 set_page_private(page, (unsigned long)head); 1408 head = page; 1409 } 1410 1411 while (head) { 1412 struct page *page = head; 1413 int err; 1414 1415 /* traversal in reverse order */ 1416 head = (void *)page_private(page); 1417 1418 err = z_erofs_do_read_page(&f, page, &pagepool); 1419 if (err) 1420 erofs_err(inode->i_sb, 1421 "readahead error at page %lu @ nid %llu", 1422 page->index, EROFS_I(inode)->nid); 1423 put_page(page); 1424 } 1425 1426 (void)z_erofs_collector_end(&f.clt); 1427 1428 z_erofs_submit_and_unzip(inode->i_sb, &f.clt, &pagepool, sync); 1429 1430 if (f.map.mpage) 1431 put_page(f.map.mpage); 1432 1433 /* clean up the remaining free pages */ 1434 put_pages_list(&pagepool); 1435 return 0; 1436 } 1437 1438 const struct address_space_operations z_erofs_vle_normalaccess_aops = { 1439 .readpage = z_erofs_vle_normalaccess_readpage, 1440 .readpages = z_erofs_vle_normalaccess_readpages, 1441 }; 1442 1443