drivers/staging/erofs/unzip_vle.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * linux/drivers/staging/erofs/unzip_vle.c
   4  *
   5  * Copyright (C) 2018 HUAWEI, Inc.
   6  *             http://www.huawei.com/
   7  * Created by Gao Xiang <gaoxiang25@huawei.com>
   8  *
   9  * This file is subject to the terms and conditions of the GNU General Public
  10  * License.  See the file COPYING in the main directory of the Linux
  11  * distribution for more details.
  12  */
  13 #include "unzip_vle.h"
  14 #include <linux/prefetch.h>
  15
  16 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
  17 static struct kmem_cache *z_erofs_workgroup_cachep __read_mostly;
  18
  19 void z_erofs_exit_zip_subsystem(void)
  20 {
  21         BUG_ON(z_erofs_workqueue == NULL);
  22         BUG_ON(z_erofs_workgroup_cachep == NULL);
  23
  24         destroy_workqueue(z_erofs_workqueue);
  25         kmem_cache_destroy(z_erofs_workgroup_cachep);
  26 }
  27
  28 static inline int init_unzip_workqueue(void)
  29 {
  30         const unsigned onlinecpus = num_possible_cpus();
  31
  32         /*
  33          * we don't need too many threads, limiting threads
  34          * could improve scheduling performance.
  35          */
  36         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
  37                 WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE,
  38                 onlinecpus + onlinecpus / 4);
  39
  40         return z_erofs_workqueue != NULL ? 0 : -ENOMEM;
  41 }
  42
  43 int z_erofs_init_zip_subsystem(void)
  44 {
  45         z_erofs_workgroup_cachep =
  46                 kmem_cache_create("erofs_compress",
  47                 Z_EROFS_WORKGROUP_SIZE, 0,
  48                 SLAB_RECLAIM_ACCOUNT, NULL);
  49
  50         if (z_erofs_workgroup_cachep != NULL) {
  51                 if (!init_unzip_workqueue())
  52                         return 0;
  53
  54                 kmem_cache_destroy(z_erofs_workgroup_cachep);
  55         }
  56         return -ENOMEM;
  57 }
  58
  59 enum z_erofs_vle_work_role {
  60         Z_EROFS_VLE_WORK_SECONDARY,
  61         Z_EROFS_VLE_WORK_PRIMARY,
  62         /*
  63          * The current work has at least been linked with the following
  64          * processed chained works, which means if the processing page
  65          * is the tail partial page of the work, the current work can
  66          * safely use the whole page, as illustrated below:
  67          * +--------------+-------------------------------------------+
  68          * |  tail page   |      head page (of the previous work)     |
  69          * +--------------+-------------------------------------------+
  70          *   /\  which belongs to the current work
  71          * [  (*) this page can be used for the current work itself.  ]
  72          */
  73         Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
  74         Z_EROFS_VLE_WORK_MAX
  75 };
  76
  77 struct z_erofs_vle_work_builder {
  78         enum z_erofs_vle_work_role role;
  79         /*
  80          * 'hosted = false' means that the current workgroup doesn't belong to
  81          * the owned chained workgroups. In the other words, it is none of our
  82          * business to submit this workgroup.
  83          */
  84         bool hosted;
  85
  86         struct z_erofs_vle_workgroup *grp;
  87         struct z_erofs_vle_work *work;
  88         struct z_erofs_pagevec_ctor vector;
  89
  90         /* pages used for reading the compressed data */
  91         struct page **compressed_pages;
  92         unsigned compressed_deficit;
  93 };
  94
  95 #define VLE_WORK_BUILDER_INIT() \
  96         { .work = NULL, .role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED }
  97
  98 #ifdef EROFS_FS_HAS_MANAGED_CACHE
  99
 100 static bool grab_managed_cache_pages(struct address_space *mapping,
 101                                      erofs_blk_t start,
 102                                      struct page **compressed_pages,
 103                                      int clusterblks,
 104                                      bool reserve_allocation)
 105 {
 106         bool noio = true;
 107         unsigned int i;
 108
 109         /* TODO: optimize by introducing find_get_pages_range */
 110         for (i = 0; i < clusterblks; ++i) {
 111                 struct page *page, *found;
 112
 113                 if (READ_ONCE(compressed_pages[i]) != NULL)
 114                         continue;
 115
 116                 page = found = find_get_page(mapping, start + i);
 117                 if (found == NULL) {
 118                         noio = false;
 119                         if (!reserve_allocation)
 120                                 continue;
 121                         page = EROFS_UNALLOCATED_CACHED_PAGE;
 122                 }
 123
 124                 if (NULL == cmpxchg(compressed_pages + i, NULL, page))
 125                         continue;
 126
 127                 if (found != NULL)
 128                         put_page(found);
 129         }
 130         return noio;
 131 }
 132
 133 /* called by erofs_shrinker to get rid of all compressed_pages */
 134 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 135                                        struct erofs_workgroup *egrp)
 136 {
 137         struct z_erofs_vle_workgroup *const grp =
 138                 container_of(egrp, struct z_erofs_vle_workgroup, obj);
 139         struct address_space *const mapping = sbi->managed_cache->i_mapping;
 140         const int clusterpages = erofs_clusterpages(sbi);
 141         int i;
 142
 143         /*
 144          * refcount of workgroup is now freezed as 1,
 145          * therefore no need to worry about available decompression users.
 146          */
 147         for (i = 0; i < clusterpages; ++i) {
 148                 struct page *page = grp->compressed_pages[i];
 149
 150                 if (page == NULL || page->mapping != mapping)
 151                         continue;
 152
 153                 /* block other users from reclaiming or migrating the page */
 154                 if (!trylock_page(page))
 155                         return -EBUSY;
 156
 157                 /* barrier is implied in the following 'unlock_page' */
 158                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 159
 160                 set_page_private(page, 0);
 161                 ClearPagePrivate(page);
 162
 163                 unlock_page(page);
 164                 put_page(page);
 165         }
 166         return 0;
 167 }
 168
 169 int erofs_try_to_free_cached_page(struct address_space *mapping,
 170                                   struct page *page)
 171 {
 172         struct erofs_sb_info *const sbi = EROFS_SB(mapping->host->i_sb);
 173         const unsigned int clusterpages = erofs_clusterpages(sbi);
 174
 175         struct z_erofs_vle_workgroup *grp;
 176         int ret = 0;    /* 0 - busy */
 177
 178         /* prevent the workgroup from being freed */
 179         rcu_read_lock();
 180         grp = (void *)page_private(page);
 181
 182         if (erofs_workgroup_try_to_freeze(&grp->obj, 1)) {
 183                 unsigned int i;
 184
 185                 for (i = 0; i < clusterpages; ++i) {
 186                         if (grp->compressed_pages[i] == page) {
 187                                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 188                                 ret = 1;
 189                                 break;
 190                         }
 191                 }
 192                 erofs_workgroup_unfreeze(&grp->obj, 1);
 193         }
 194         rcu_read_unlock();
 195
 196         if (ret) {
 197                 ClearPagePrivate(page);
 198                 put_page(page);
 199         }
 200         return ret;
 201 }
 202 #endif
 203
 204 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 205 static inline bool try_to_reuse_as_compressed_page(
 206         struct z_erofs_vle_work_builder *b,
 207         struct page *page)
 208 {
 209         while (b->compressed_deficit) {
 210                 --b->compressed_deficit;
 211                 if (NULL == cmpxchg(b->compressed_pages++, NULL, page))
 212                         return true;
 213         }
 214
 215         return false;
 216 }
 217
 218 /* callers must be with work->lock held */
 219 static int z_erofs_vle_work_add_page(
 220         struct z_erofs_vle_work_builder *builder,
 221         struct page *page,
 222         enum z_erofs_page_type type)
 223 {
 224         int ret;
 225         bool occupied;
 226
 227         /* give priority for the compressed data storage */
 228         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY &&
 229                 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 230                 try_to_reuse_as_compressed_page(builder, page))
 231                 return 0;
 232
 233         ret = z_erofs_pagevec_ctor_enqueue(&builder->vector,
 234                 page, type, &occupied);
 235         builder->work->vcnt += (unsigned)ret;
 236
 237         return ret ? 0 : -EAGAIN;
 238 }
 239
 240 static inline bool try_to_claim_workgroup(
 241         struct z_erofs_vle_workgroup *grp,
 242         z_erofs_vle_owned_workgrp_t *owned_head,
 243         bool *hosted)
 244 {
 245         DBG_BUGON(*hosted == true);
 246
 247         /* let's claim these following types of workgroup */
 248 retry:
 249         if (grp->next == Z_EROFS_VLE_WORKGRP_NIL) {
 250                 /* type 1, nil workgroup */
 251                 if (Z_EROFS_VLE_WORKGRP_NIL != cmpxchg(&grp->next,
 252                         Z_EROFS_VLE_WORKGRP_NIL, *owned_head))
 253                         goto retry;
 254
 255                 *owned_head = grp;
 256                 *hosted = true;
 257         } else if (grp->next == Z_EROFS_VLE_WORKGRP_TAIL) {
 258                 /*
 259                  * type 2, link to the end of a existing open chain,
 260                  * be careful that its submission itself is governed
 261                  * by the original owned chain.
 262                  */
 263                 if (Z_EROFS_VLE_WORKGRP_TAIL != cmpxchg(&grp->next,
 264                         Z_EROFS_VLE_WORKGRP_TAIL, *owned_head))
 265                         goto retry;
 266
 267                 *owned_head = Z_EROFS_VLE_WORKGRP_TAIL;
 268         } else
 269                 return false;   /* :( better luck next time */
 270
 271         return true;    /* lucky, I am the followee :) */
 272 }
 273
 274 static struct z_erofs_vle_work *
 275 z_erofs_vle_work_lookup(struct super_block *sb,
 276                         pgoff_t idx, unsigned pageofs,
 277                         struct z_erofs_vle_workgroup **grp_ret,
 278                         enum z_erofs_vle_work_role *role,
 279                         z_erofs_vle_owned_workgrp_t *owned_head,
 280                         bool *hosted)
 281 {
 282         bool tag, primary;
 283         struct erofs_workgroup *egrp;
 284         struct z_erofs_vle_workgroup *grp;
 285         struct z_erofs_vle_work *work;
 286
 287         egrp = erofs_find_workgroup(sb, idx, &tag);
 288         if (egrp == NULL) {
 289                 *grp_ret = NULL;
 290                 return NULL;
 291         }
 292
 293         *grp_ret = grp = container_of(egrp,
 294                 struct z_erofs_vle_workgroup, obj);
 295
 296 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 297         work = z_erofs_vle_grab_work(grp, pageofs);
 298         primary = true;
 299 #else
 300         BUG();
 301 #endif
 302
 303         DBG_BUGON(work->pageofs != pageofs);
 304
 305         /*
 306          * lock must be taken first to avoid grp->next == NIL between
 307          * claiming workgroup and adding pages:
 308          *                        grp->next != NIL
 309          *   grp->next = NIL
 310          *   mutex_unlock_all
 311          *                        mutex_lock(&work->lock)
 312          *                        add all pages to pagevec
 313          *
 314          * [correct locking case 1]:
 315          *   mutex_lock(grp->work[a])
 316          *   ...
 317          *   mutex_lock(grp->work[b])     mutex_lock(grp->work[c])
 318          *   ...                          *role = SECONDARY
 319          *                                add all pages to pagevec
 320          *                                ...
 321          *                                mutex_unlock(grp->work[c])
 322          *   mutex_lock(grp->work[c])
 323          *   ...
 324          *   grp->next = NIL
 325          *   mutex_unlock_all
 326          *
 327          * [correct locking case 2]:
 328          *   mutex_lock(grp->work[b])
 329          *   ...
 330          *   mutex_lock(grp->work[a])
 331          *   ...
 332          *   mutex_lock(grp->work[c])
 333          *   ...
 334          *   grp->next = NIL
 335          *   mutex_unlock_all
 336          *                                mutex_lock(grp->work[a])
 337          *                                *role = PRIMARY_OWNER
 338          *                                add all pages to pagevec
 339          *                                ...
 340          */
 341         mutex_lock(&work->lock);
 342
 343         *hosted = false;
 344         if (!primary)
 345                 *role = Z_EROFS_VLE_WORK_SECONDARY;
 346         /* claim the workgroup if possible */
 347         else if (try_to_claim_workgroup(grp, owned_head, hosted))
 348                 *role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 349         else
 350                 *role = Z_EROFS_VLE_WORK_PRIMARY;
 351
 352         return work;
 353 }
 354
 355 static struct z_erofs_vle_work *
 356 z_erofs_vle_work_register(struct super_block *sb,
 357                           struct z_erofs_vle_workgroup **grp_ret,
 358                           struct erofs_map_blocks *map,
 359                           pgoff_t index, unsigned pageofs,
 360                           enum z_erofs_vle_work_role *role,
 361                           z_erofs_vle_owned_workgrp_t *owned_head,
 362                           bool *hosted)
 363 {
 364         bool newgrp = false;
 365         struct z_erofs_vle_workgroup *grp = *grp_ret;
 366         struct z_erofs_vle_work *work;
 367
 368 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 369         BUG_ON(grp != NULL);
 370 #else
 371         if (grp != NULL)
 372                 goto skip;
 373 #endif
 374         /* no available workgroup, let's allocate one */
 375         grp = kmem_cache_zalloc(z_erofs_workgroup_cachep, GFP_NOFS);
 376         if (unlikely(grp == NULL))
 377                 return ERR_PTR(-ENOMEM);
 378
 379         grp->obj.index = index;
 380         grp->llen = map->m_llen;
 381
 382         z_erofs_vle_set_workgrp_fmt(grp,
 383                 (map->m_flags & EROFS_MAP_ZIPPED) ?
 384                         Z_EROFS_VLE_WORKGRP_FMT_LZ4 :
 385                         Z_EROFS_VLE_WORKGRP_FMT_PLAIN);
 386         atomic_set(&grp->obj.refcount, 1);
 387
 388         /* new workgrps have been claimed as type 1 */
 389         WRITE_ONCE(grp->next, *owned_head);
 390         /* primary and followed work for all new workgrps */
 391         *role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 392         /* it should be submitted by ourselves */
 393         *hosted = true;
 394
 395         newgrp = true;
 396 #ifdef CONFIG_EROFS_FS_ZIP_MULTIREF
 397 skip:
 398         /* currently unimplemented */
 399         BUG();
 400 #else
 401         work = z_erofs_vle_grab_primary_work(grp);
 402 #endif
 403         work->pageofs = pageofs;
 404
 405         mutex_init(&work->lock);
 406
 407         if (newgrp) {
 408                 int err = erofs_register_workgroup(sb, &grp->obj, 0);
 409
 410                 if (err) {
 411                         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 412                         return ERR_PTR(-EAGAIN);
 413                 }
 414         }
 415
 416         *owned_head = *grp_ret = grp;
 417
 418         mutex_lock(&work->lock);
 419         return work;
 420 }
 421
 422 static inline void __update_workgrp_llen(struct z_erofs_vle_workgroup *grp,
 423                                          unsigned int llen)
 424 {
 425         while (1) {
 426                 unsigned int orig_llen = grp->llen;
 427
 428                 if (orig_llen >= llen || orig_llen ==
 429                         cmpxchg(&grp->llen, orig_llen, llen))
 430                         break;
 431         }
 432 }
 433
 434 #define builder_is_followed(builder) \
 435         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 436
 437 static int z_erofs_vle_work_iter_begin(struct z_erofs_vle_work_builder *builder,
 438                                        struct super_block *sb,
 439                                        struct erofs_map_blocks *map,
 440                                        z_erofs_vle_owned_workgrp_t *owned_head)
 441 {
 442         const unsigned clusterpages = erofs_clusterpages(EROFS_SB(sb));
 443         const erofs_blk_t index = erofs_blknr(map->m_pa);
 444         const unsigned pageofs = map->m_la & ~PAGE_MASK;
 445         struct z_erofs_vle_workgroup *grp;
 446         struct z_erofs_vle_work *work;
 447
 448         DBG_BUGON(builder->work != NULL);
 449
 450         /* must be Z_EROFS_WORK_TAIL or the next chained work */
 451         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_NIL);
 452         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
 453
 454         DBG_BUGON(erofs_blkoff(map->m_pa));
 455
 456 repeat:
 457         work = z_erofs_vle_work_lookup(sb, index,
 458                 pageofs, &grp, &builder->role, owned_head, &builder->hosted);
 459         if (work != NULL) {
 460                 __update_workgrp_llen(grp, map->m_llen);
 461                 goto got_it;
 462         }
 463
 464         work = z_erofs_vle_work_register(sb, &grp, map, index, pageofs,
 465                 &builder->role, owned_head, &builder->hosted);
 466
 467         if (unlikely(work == ERR_PTR(-EAGAIN)))
 468                 goto repeat;
 469
 470         if (unlikely(IS_ERR(work)))
 471                 return PTR_ERR(work);
 472 got_it:
 473         z_erofs_pagevec_ctor_init(&builder->vector,
 474                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, work->vcnt);
 475
 476         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY) {
 477                 /* enable possibly in-place decompression */
 478                 builder->compressed_pages = grp->compressed_pages;
 479                 builder->compressed_deficit = clusterpages;
 480         } else {
 481                 builder->compressed_pages = NULL;
 482                 builder->compressed_deficit = 0;
 483         }
 484
 485         builder->grp = grp;
 486         builder->work = work;
 487         return 0;
 488 }
 489
 490 /*
 491  * keep in mind that no referenced workgroups will be freed
 492  * only after a RCU grace period, so rcu_read_lock() could
 493  * prevent a workgroup from being freed.
 494  */
 495 static void z_erofs_rcu_callback(struct rcu_head *head)
 496 {
 497         struct z_erofs_vle_work *work = container_of(head,
 498                 struct z_erofs_vle_work, rcu);
 499         struct z_erofs_vle_workgroup *grp =
 500                 z_erofs_vle_work_workgroup(work, true);
 501
 502         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 503 }
 504
 505 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 506 {
 507         struct z_erofs_vle_workgroup *const vgrp = container_of(grp,
 508                 struct z_erofs_vle_workgroup, obj);
 509         struct z_erofs_vle_work *const work = &vgrp->work;
 510
 511         call_rcu(&work->rcu, z_erofs_rcu_callback);
 512 }
 513
 514 static void __z_erofs_vle_work_release(struct z_erofs_vle_workgroup *grp,
 515         struct z_erofs_vle_work *work __maybe_unused)
 516 {
 517         erofs_workgroup_put(&grp->obj);
 518 }
 519
 520 void z_erofs_vle_work_release(struct z_erofs_vle_work *work)
 521 {
 522         struct z_erofs_vle_workgroup *grp =
 523                 z_erofs_vle_work_workgroup(work, true);
 524
 525         __z_erofs_vle_work_release(grp, work);
 526 }
 527
 528 static inline bool
 529 z_erofs_vle_work_iter_end(struct z_erofs_vle_work_builder *builder)
 530 {
 531         struct z_erofs_vle_work *work = builder->work;
 532
 533         if (work == NULL)
 534                 return false;
 535
 536         z_erofs_pagevec_ctor_exit(&builder->vector, false);
 537         mutex_unlock(&work->lock);
 538
 539         /*
 540          * if all pending pages are added, don't hold work reference
 541          * any longer if the current work isn't hosted by ourselves.
 542          */
 543         if (!builder->hosted)
 544                 __z_erofs_vle_work_release(builder->grp, work);
 545
 546         builder->work = NULL;
 547         builder->grp = NULL;
 548         return true;
 549 }
 550
 551 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 552                                                gfp_t gfp)
 553 {
 554         struct page *page = erofs_allocpage(pagepool, gfp);
 555
 556         if (unlikely(page == NULL))
 557                 return NULL;
 558
 559         page->mapping = Z_EROFS_MAPPING_STAGING;
 560         return page;
 561 }
 562
 563 struct z_erofs_vle_frontend {
 564         struct inode *const inode;
 565
 566         struct z_erofs_vle_work_builder builder;
 567         struct erofs_map_blocks_iter m_iter;
 568
 569         z_erofs_vle_owned_workgrp_t owned_head;
 570
 571         bool initial;
 572 #if (EROFS_FS_ZIP_CACHE_LVL >= 2)
 573         erofs_off_t cachedzone_la;
 574 #endif
 575 };
 576
 577 #define VLE_FRONTEND_INIT(__i) { \
 578         .inode = __i, \
 579         .m_iter = { \
 580                 { .m_llen = 0, .m_plen = 0 }, \
 581                 .mpage = NULL \
 582         }, \
 583         .builder = VLE_WORK_BUILDER_INIT(), \
 584         .owned_head = Z_EROFS_VLE_WORKGRP_TAIL, \
 585         .initial = true, }
 586
 587 static int z_erofs_do_read_page(struct z_erofs_vle_frontend *fe,
 588                                 struct page *page,
 589                                 struct list_head *page_pool)
 590 {
 591         struct super_block *const sb = fe->inode->i_sb;
 592         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
 593         struct erofs_map_blocks_iter *const m = &fe->m_iter;
 594         struct erofs_map_blocks *const map = &m->map;
 595         struct z_erofs_vle_work_builder *const builder = &fe->builder;
 596         const loff_t offset = page_offset(page);
 597
 598         bool tight = builder_is_followed(builder);
 599         struct z_erofs_vle_work *work = builder->work;
 600
 601 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 602         struct address_space *const mngda = sbi->managed_cache->i_mapping;
 603         struct z_erofs_vle_workgroup *grp;
 604         bool noio_outoforder;
 605 #endif
 606
 607         enum z_erofs_page_type page_type;
 608         unsigned cur, end, spiltted, index;
 609         int err;
 610
 611         /* register locked file pages as online pages in pack */
 612         z_erofs_onlinepage_init(page);
 613
 614         spiltted = 0;
 615         end = PAGE_SIZE;
 616 repeat:
 617         cur = end - 1;
 618
 619         /* lucky, within the range of the current map_blocks */
 620         if (offset + cur >= map->m_la &&
 621                 offset + cur < map->m_la + map->m_llen)
 622                 goto hitted;
 623
 624         /* go ahead the next map_blocks */
 625         debugln("%s: [out-of-range] pos %llu", __func__, offset + cur);
 626
 627         if (!z_erofs_vle_work_iter_end(builder))
 628                 fe->initial = false;
 629
 630         map->m_la = offset + cur;
 631         map->m_llen = 0;
 632         err = erofs_map_blocks_iter(fe->inode, map, &m->mpage, 0);
 633         if (unlikely(err))
 634                 goto err_out;
 635
 636         /* deal with hole (FIXME! broken now) */
 637         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED)))
 638                 goto hitted;
 639
 640         DBG_BUGON(map->m_plen != 1 << sbi->clusterbits);
 641         BUG_ON(erofs_blkoff(map->m_pa));
 642
 643         err = z_erofs_vle_work_iter_begin(builder, sb, map, &fe->owned_head);
 644         if (unlikely(err))
 645                 goto err_out;
 646
 647 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 648         grp = fe->builder.grp;
 649
 650         /* let's do out-of-order decompression for noio */
 651         noio_outoforder = grab_managed_cache_pages(mngda,
 652                 erofs_blknr(map->m_pa),
 653                 grp->compressed_pages, erofs_blknr(map->m_plen),
 654                 /* compressed page caching selection strategy */
 655                 fe->initial | (EROFS_FS_ZIP_CACHE_LVL >= 2 ?
 656                         map->m_la < fe->cachedzone_la : 0));
 657
 658         if (noio_outoforder && builder_is_followed(builder))
 659                 builder->role = Z_EROFS_VLE_WORK_PRIMARY;
 660 #endif
 661
 662         tight &= builder_is_followed(builder);
 663         work = builder->work;
 664 hitted:
 665         cur = end - min_t(unsigned, offset + end - map->m_la, end);
 666         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED))) {
 667                 zero_user_segment(page, cur, end);
 668                 goto next_part;
 669         }
 670
 671         /* let's derive page type */
 672         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 673                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 674                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 675                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 676
 677 retry:
 678         err = z_erofs_vle_work_add_page(builder, page, page_type);
 679         /* should allocate an additional staging page for pagevec */
 680         if (err == -EAGAIN) {
 681                 struct page *const newpage =
 682                         __stagingpage_alloc(page_pool, GFP_NOFS);
 683
 684                 err = z_erofs_vle_work_add_page(builder,
 685                         newpage, Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 686                 if (!err)
 687                         goto retry;
 688         }
 689
 690         if (unlikely(err))
 691                 goto err_out;
 692
 693         index = page->index - map->m_la / PAGE_SIZE;
 694
 695         /* FIXME! avoid the last relundant fixup & endio */
 696         z_erofs_onlinepage_fixup(page, index, true);
 697         ++spiltted;
 698
 699         /* also update nr_pages and increase queued_pages */
 700         work->nr_pages = max_t(pgoff_t, work->nr_pages, index + 1);
 701 next_part:
 702         /* can be used for verification */
 703         map->m_llen = offset + cur - map->m_la;
 704
 705         end = cur;
 706         if (end > 0)
 707                 goto repeat;
 708
 709         /* FIXME! avoid the last relundant fixup & endio */
 710         z_erofs_onlinepage_endio(page);
 711
 712         debugln("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 713                 __func__, page, spiltted, map->m_llen);
 714         return 0;
 715
 716 err_out:
 717         /* TODO: the missing error handing cases */
 718         return err;
 719 }
 720
 721 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 722 {
 723         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 724         struct z_erofs_vle_unzip_io *io = tagptr_unfold_ptr(t);
 725         bool background = tagptr_unfold_tags(t);
 726
 727         if (atomic_add_return(bios, &io->pending_bios))
 728                 return;
 729
 730         if (background)
 731                 queue_work(z_erofs_workqueue, &io->u.work);
 732         else
 733                 wake_up(&io->u.wait);
 734 }
 735
 736 static inline void z_erofs_vle_read_endio(struct bio *bio)
 737 {
 738         const blk_status_t err = bio->bi_status;
 739         unsigned i;
 740         struct bio_vec *bvec;
 741 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 742         struct address_space *mngda = NULL;
 743 #endif
 744
 745         bio_for_each_segment_all(bvec, bio, i) {
 746                 struct page *page = bvec->bv_page;
 747                 bool cachemngd = false;
 748
 749                 DBG_BUGON(PageUptodate(page));
 750                 BUG_ON(page->mapping == NULL);
 751
 752 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 753                 if (unlikely(mngda == NULL && !z_erofs_is_stagingpage(page))) {
 754                         struct inode *const inode = page->mapping->host;
 755                         struct super_block *const sb = inode->i_sb;
 756
 757                         mngda = EROFS_SB(sb)->managed_cache->i_mapping;
 758                 }
 759
 760                 /*
 761                  * If mngda has not gotten, it equals NULL,
 762                  * however, page->mapping never be NULL if working properly.
 763                  */
 764                 cachemngd = (page->mapping == mngda);
 765 #endif
 766
 767                 if (unlikely(err))
 768                         SetPageError(page);
 769                 else if (cachemngd)
 770                         SetPageUptodate(page);
 771
 772                 if (cachemngd)
 773                         unlock_page(page);
 774         }
 775
 776         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 777         bio_put(bio);
 778 }
 779
 780 static struct page *z_pagemap_global[Z_EROFS_VLE_VMAP_GLOBAL_PAGES];
 781 static DEFINE_MUTEX(z_pagemap_global_lock);
 782
 783 static int z_erofs_vle_unzip(struct super_block *sb,
 784         struct z_erofs_vle_workgroup *grp,
 785         struct list_head *page_pool)
 786 {
 787         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 788 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 789         struct address_space *const mngda = sbi->managed_cache->i_mapping;
 790 #endif
 791         const unsigned clusterpages = erofs_clusterpages(sbi);
 792
 793         struct z_erofs_pagevec_ctor ctor;
 794         unsigned nr_pages;
 795 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 796         unsigned sparsemem_pages = 0;
 797 #endif
 798         struct page *pages_onstack[Z_EROFS_VLE_VMAP_ONSTACK_PAGES];
 799         struct page **pages, **compressed_pages, *page;
 800         unsigned i, llen;
 801
 802         enum z_erofs_page_type page_type;
 803         bool overlapped;
 804         struct z_erofs_vle_work *work;
 805         void *vout;
 806         int err;
 807
 808         might_sleep();
 809 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 810         work = z_erofs_vle_grab_primary_work(grp);
 811 #else
 812         BUG();
 813 #endif
 814         BUG_ON(!READ_ONCE(work->nr_pages));
 815
 816         mutex_lock(&work->lock);
 817         nr_pages = work->nr_pages;
 818
 819         if (likely(nr_pages <= Z_EROFS_VLE_VMAP_ONSTACK_PAGES))
 820                 pages = pages_onstack;
 821         else if (nr_pages <= Z_EROFS_VLE_VMAP_GLOBAL_PAGES &&
 822                 mutex_trylock(&z_pagemap_global_lock))
 823                 pages = z_pagemap_global;
 824         else {
 825 repeat:
 826                 pages = kvmalloc_array(nr_pages,
 827                         sizeof(struct page *), GFP_KERNEL);
 828
 829                 /* fallback to global pagemap for the lowmem scenario */
 830                 if (unlikely(pages == NULL)) {
 831                         if (nr_pages > Z_EROFS_VLE_VMAP_GLOBAL_PAGES)
 832                                 goto repeat;
 833                         else {
 834                                 mutex_lock(&z_pagemap_global_lock);
 835                                 pages = z_pagemap_global;
 836                         }
 837                 }
 838         }
 839
 840         for (i = 0; i < nr_pages; ++i)
 841                 pages[i] = NULL;
 842
 843         z_erofs_pagevec_ctor_init(&ctor,
 844                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, 0);
 845
 846         for (i = 0; i < work->vcnt; ++i) {
 847                 unsigned pagenr;
 848
 849                 page = z_erofs_pagevec_ctor_dequeue(&ctor, &page_type);
 850
 851                 /* all pages in pagevec ought to be valid */
 852                 DBG_BUGON(page == NULL);
 853                 DBG_BUGON(page->mapping == NULL);
 854
 855                 if (z_erofs_gather_if_stagingpage(page_pool, page))
 856                         continue;
 857
 858                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 859                         pagenr = 0;
 860                 else
 861                         pagenr = z_erofs_onlinepage_index(page);
 862
 863                 BUG_ON(pagenr >= nr_pages);
 864
 865 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 866                 BUG_ON(pages[pagenr] != NULL);
 867                 ++sparsemem_pages;
 868 #endif
 869                 pages[pagenr] = page;
 870         }
 871
 872         z_erofs_pagevec_ctor_exit(&ctor, true);
 873
 874         overlapped = false;
 875         compressed_pages = grp->compressed_pages;
 876
 877         for (i = 0; i < clusterpages; ++i) {
 878                 unsigned pagenr;
 879
 880                 page = compressed_pages[i];
 881
 882                 /* all compressed pages ought to be valid */
 883                 DBG_BUGON(page == NULL);
 884                 DBG_BUGON(page->mapping == NULL);
 885
 886                 if (z_erofs_is_stagingpage(page))
 887                         continue;
 888 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 889                 else if (page->mapping == mngda) {
 890                         BUG_ON(PageLocked(page));
 891                         BUG_ON(!PageUptodate(page));
 892                         continue;
 893                 }
 894 #endif
 895
 896                 /* only non-head page could be reused as a compressed page */
 897                 pagenr = z_erofs_onlinepage_index(page);
 898
 899                 BUG_ON(pagenr >= nr_pages);
 900 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 901                 BUG_ON(pages[pagenr] != NULL);
 902                 ++sparsemem_pages;
 903 #endif
 904                 pages[pagenr] = page;
 905
 906                 overlapped = true;
 907         }
 908
 909         llen = (nr_pages << PAGE_SHIFT) - work->pageofs;
 910
 911         if (z_erofs_vle_workgrp_fmt(grp) == Z_EROFS_VLE_WORKGRP_FMT_PLAIN) {
 912                 /* FIXME! this should be fixed in the future */
 913                 BUG_ON(grp->llen != llen);
 914
 915                 err = z_erofs_vle_plain_copy(compressed_pages, clusterpages,
 916                         pages, nr_pages, work->pageofs);
 917                 goto out;
 918         }
 919
 920         if (llen > grp->llen)
 921                 llen = grp->llen;
 922
 923         err = z_erofs_vle_unzip_fast_percpu(compressed_pages,
 924                 clusterpages, pages, llen, work->pageofs,
 925                 z_erofs_onlinepage_endio);
 926         if (err != -ENOTSUPP)
 927                 goto out_percpu;
 928
 929 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 930         if (sparsemem_pages >= nr_pages) {
 931                 BUG_ON(sparsemem_pages > nr_pages);
 932                 goto skip_allocpage;
 933         }
 934 #endif
 935
 936         for (i = 0; i < nr_pages; ++i) {
 937                 if (pages[i] != NULL)
 938                         continue;
 939
 940                 pages[i] = __stagingpage_alloc(page_pool, GFP_NOFS);
 941         }
 942
 943 #ifndef CONFIG_EROFS_FS_ZIP_MULTIREF
 944 skip_allocpage:
 945 #endif
 946         vout = erofs_vmap(pages, nr_pages);
 947
 948         err = z_erofs_vle_unzip_vmap(compressed_pages,
 949                 clusterpages, vout, llen, work->pageofs, overlapped);
 950
 951         erofs_vunmap(vout, nr_pages);
 952
 953 out:
 954         for (i = 0; i < nr_pages; ++i) {
 955                 page = pages[i];
 956                 DBG_BUGON(page->mapping == NULL);
 957
 958                 /* recycle all individual staging pages */
 959                 if (z_erofs_gather_if_stagingpage(page_pool, page))
 960                         continue;
 961
 962                 if (unlikely(err < 0))
 963                         SetPageError(page);
 964
 965                 z_erofs_onlinepage_endio(page);
 966         }
 967
 968 out_percpu:
 969         for (i = 0; i < clusterpages; ++i) {
 970                 page = compressed_pages[i];
 971
 972 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 973                 if (page->mapping == mngda)
 974                         continue;
 975 #endif
 976                 /* recycle all individual staging pages */
 977                 (void)z_erofs_gather_if_stagingpage(page_pool, page);
 978
 979                 WRITE_ONCE(compressed_pages[i], NULL);
 980         }
 981
 982         if (pages == z_pagemap_global)
 983                 mutex_unlock(&z_pagemap_global_lock);
 984         else if (unlikely(pages != pages_onstack))
 985                 kvfree(pages);
 986
 987         work->nr_pages = 0;
 988         work->vcnt = 0;
 989
 990         /* all work locks MUST be taken before the following line */
 991
 992         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_NIL);
 993
 994         /* all work locks SHOULD be released right now */
 995         mutex_unlock(&work->lock);
 996
 997         z_erofs_vle_work_release(work);
 998         return err;
 999 }
1000
1001 static void z_erofs_vle_unzip_all(struct super_block *sb,
1002                                   struct z_erofs_vle_unzip_io *io,
1003                                   struct list_head *page_pool)
1004 {
1005         z_erofs_vle_owned_workgrp_t owned = io->head;
1006
1007         while (owned != Z_EROFS_VLE_WORKGRP_TAIL_CLOSED) {
1008                 struct z_erofs_vle_workgroup *grp;
1009
1010                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1011                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_TAIL);
1012
1013                 /* no possible that 'owned' equals NULL */
1014                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_NIL);
1015
1016                 grp = owned;
1017                 owned = READ_ONCE(grp->next);
1018
1019                 z_erofs_vle_unzip(sb, grp, page_pool);
1020         }
1021 }
1022
1023 static void z_erofs_vle_unzip_wq(struct work_struct *work)
1024 {
1025         struct z_erofs_vle_unzip_io_sb *iosb = container_of(work,
1026                 struct z_erofs_vle_unzip_io_sb, io.u.work);
1027         LIST_HEAD(page_pool);
1028
1029         BUG_ON(iosb->io.head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1030         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &page_pool);
1031
1032         put_pages_list(&page_pool);
1033         kvfree(iosb);
1034 }
1035
1036 static inline struct z_erofs_vle_unzip_io *
1037 prepare_io_handler(struct super_block *sb,
1038                    struct z_erofs_vle_unzip_io *io,
1039                    bool background)
1040 {
1041         struct z_erofs_vle_unzip_io_sb *iosb;
1042
1043         if (!background) {
1044                 /* waitqueue available for foreground io */
1045                 BUG_ON(io == NULL);
1046
1047                 init_waitqueue_head(&io->u.wait);
1048                 atomic_set(&io->pending_bios, 0);
1049                 goto out;
1050         }
1051
1052         if (io != NULL)
1053                 BUG();
1054         else {
1055                 /* allocate extra io descriptor for background io */
1056                 iosb = kvzalloc(sizeof(struct z_erofs_vle_unzip_io_sb),
1057                         GFP_KERNEL | __GFP_NOFAIL);
1058                 BUG_ON(iosb == NULL);
1059
1060                 io = &iosb->io;
1061         }
1062
1063         iosb->sb = sb;
1064         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1065 out:
1066         io->head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1067         return io;
1068 }
1069
1070 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1071 /* true - unlocked (noio), false - locked (need submit io) */
1072 static inline bool recover_managed_page(struct z_erofs_vle_workgroup *grp,
1073                                         struct page *page)
1074 {
1075         wait_on_page_locked(page);
1076         if (PagePrivate(page) && PageUptodate(page))
1077                 return true;
1078
1079         lock_page(page);
1080         if (unlikely(!PagePrivate(page))) {
1081                 set_page_private(page, (unsigned long)grp);
1082                 SetPagePrivate(page);
1083         }
1084         if (unlikely(PageUptodate(page))) {
1085                 unlock_page(page);
1086                 return true;
1087         }
1088         return false;
1089 }
1090
1091 #define __FSIO_1 1
1092 #else
1093 #define __FSIO_1 0
1094 #endif
1095
1096 static bool z_erofs_vle_submit_all(struct super_block *sb,
1097                                    z_erofs_vle_owned_workgrp_t owned_head,
1098                                    struct list_head *pagepool,
1099                                    struct z_erofs_vle_unzip_io *fg_io,
1100                                    bool force_fg)
1101 {
1102         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1103         const unsigned clusterpages = erofs_clusterpages(sbi);
1104         const gfp_t gfp = GFP_NOFS;
1105 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1106         struct address_space *const mngda = sbi->managed_cache->i_mapping;
1107         struct z_erofs_vle_workgroup *lstgrp_noio = NULL, *lstgrp_io = NULL;
1108 #endif
1109         struct z_erofs_vle_unzip_io *ios[1 + __FSIO_1];
1110         struct bio *bio;
1111         tagptr1_t bi_private;
1112         /* since bio will be NULL, no need to initialize last_index */
1113         pgoff_t uninitialized_var(last_index);
1114         bool force_submit = false;
1115         unsigned nr_bios;
1116
1117         if (unlikely(owned_head == Z_EROFS_VLE_WORKGRP_TAIL))
1118                 return false;
1119
1120         /*
1121          * force_fg == 1, (io, fg_io[0]) no io, (io, fg_io[1]) need submit io
1122          * force_fg == 0, (io, fg_io[0]) no io; (io[1], bg_io) need submit io
1123          */
1124 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1125         ios[0] = prepare_io_handler(sb, fg_io + 0, false);
1126 #endif
1127
1128         if (force_fg) {
1129                 ios[__FSIO_1] = prepare_io_handler(sb, fg_io + __FSIO_1, false);
1130                 bi_private = tagptr_fold(tagptr1_t, ios[__FSIO_1], 0);
1131         } else {
1132                 ios[__FSIO_1] = prepare_io_handler(sb, NULL, true);
1133                 bi_private = tagptr_fold(tagptr1_t, ios[__FSIO_1], 1);
1134         }
1135
1136         nr_bios = 0;
1137         force_submit = false;
1138         bio = NULL;
1139
1140         /* by default, all need io submission */
1141         ios[__FSIO_1]->head = owned_head;
1142
1143         do {
1144                 struct z_erofs_vle_workgroup *grp;
1145                 struct page **compressed_pages, *oldpage, *page;
1146                 pgoff_t first_index;
1147                 unsigned i = 0;
1148 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1149                 unsigned int noio = 0;
1150                 bool cachemngd;
1151 #endif
1152                 int err;
1153
1154                 /* no possible 'owned_head' equals the following */
1155                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1156                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_NIL);
1157
1158                 grp = owned_head;
1159
1160                 /* close the main owned chain at first */
1161                 owned_head = cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
1162                         Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1163
1164                 first_index = grp->obj.index;
1165                 compressed_pages = grp->compressed_pages;
1166
1167                 force_submit |= (first_index != last_index + 1);
1168 repeat:
1169                 /* fulfill all compressed pages */
1170                 oldpage = page = READ_ONCE(compressed_pages[i]);
1171
1172 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1173                 cachemngd = false;
1174
1175                 if (page == EROFS_UNALLOCATED_CACHED_PAGE) {
1176                         cachemngd = true;
1177                         goto do_allocpage;
1178                 } else if (page != NULL) {
1179                         if (page->mapping != mngda)
1180                                 BUG_ON(PageUptodate(page));
1181                         else if (recover_managed_page(grp, page)) {
1182                                 /* page is uptodate, skip io submission */
1183                                 force_submit = true;
1184                                 ++noio;
1185                                 goto skippage;
1186                         }
1187                 } else {
1188 do_allocpage:
1189 #else
1190                 if (page != NULL)
1191                         BUG_ON(PageUptodate(page));
1192                 else {
1193 #endif
1194                         page = __stagingpage_alloc(pagepool, gfp);
1195
1196                         if (oldpage != cmpxchg(compressed_pages + i,
1197                                 oldpage, page)) {
1198                                 list_add(&page->lru, pagepool);
1199                                 goto repeat;
1200 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1201                         } else if (cachemngd && !add_to_page_cache_lru(page,
1202                                 mngda, first_index + i, gfp)) {
1203                                 set_page_private(page, (unsigned long)grp);
1204                                 SetPagePrivate(page);
1205 #endif
1206                         }
1207                 }
1208
1209                 if (bio != NULL && force_submit) {
1210 submit_bio_retry:
1211                         __submit_bio(bio, REQ_OP_READ, 0);
1212                         bio = NULL;
1213                 }
1214
1215                 if (bio == NULL) {
1216                         bio = prepare_bio(sb, first_index + i,
1217                                 BIO_MAX_PAGES, z_erofs_vle_read_endio);
1218                         bio->bi_private = tagptr_cast_ptr(bi_private);
1219
1220                         ++nr_bios;
1221                 }
1222
1223                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1224                 if (err < PAGE_SIZE)
1225                         goto submit_bio_retry;
1226
1227                 force_submit = false;
1228                 last_index = first_index + i;
1229 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1230 skippage:
1231 #endif
1232                 if (++i < clusterpages)
1233                         goto repeat;
1234
1235 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1236                 if (noio < clusterpages) {
1237                         lstgrp_io = grp;
1238                 } else {
1239                         z_erofs_vle_owned_workgrp_t iogrp_next =
1240                                 owned_head == Z_EROFS_VLE_WORKGRP_TAIL ?
1241                                 Z_EROFS_VLE_WORKGRP_TAIL_CLOSED :
1242                                 owned_head;
1243
1244                         if (lstgrp_io == NULL)
1245                                 ios[1]->head = iogrp_next;
1246                         else
1247                                 WRITE_ONCE(lstgrp_io->next, iogrp_next);
1248
1249                         if (lstgrp_noio == NULL)
1250                                 ios[0]->head = grp;
1251                         else
1252                                 WRITE_ONCE(lstgrp_noio->next, grp);
1253
1254                         lstgrp_noio = grp;
1255                 }
1256 #endif
1257         } while (owned_head != Z_EROFS_VLE_WORKGRP_TAIL);
1258
1259         if (bio != NULL)
1260                 __submit_bio(bio, REQ_OP_READ, 0);
1261
1262 #ifndef EROFS_FS_HAS_MANAGED_CACHE
1263         BUG_ON(!nr_bios);
1264 #else
1265         if (lstgrp_noio != NULL)
1266                 WRITE_ONCE(lstgrp_noio->next, Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1267
1268         if (!force_fg && !nr_bios) {
1269                 kvfree(container_of(ios[1],
1270                         struct z_erofs_vle_unzip_io_sb, io));
1271                 return true;
1272         }
1273 #endif
1274
1275         z_erofs_vle_unzip_kickoff(tagptr_cast_ptr(bi_private), nr_bios);
1276         return true;
1277 }
1278
1279 static void z_erofs_submit_and_unzip(struct z_erofs_vle_frontend *f,
1280                                      struct list_head *pagepool,
1281                                      bool force_fg)
1282 {
1283         struct super_block *sb = f->inode->i_sb;
1284         struct z_erofs_vle_unzip_io io[1 + __FSIO_1];
1285
1286         if (!z_erofs_vle_submit_all(sb, f->owned_head, pagepool, io, force_fg))
1287                 return;
1288
1289 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1290         z_erofs_vle_unzip_all(sb, &io[0], pagepool);
1291 #endif
1292         if (!force_fg)
1293                 return;
1294
1295         /* wait until all bios are completed */
1296         wait_event(io[__FSIO_1].u.wait,
1297                 !atomic_read(&io[__FSIO_1].pending_bios));
1298
1299         /* let's synchronous decompression */
1300         z_erofs_vle_unzip_all(sb, &io[__FSIO_1], pagepool);
1301 }
1302
1303 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1304                                              struct page *page)
1305 {
1306         struct inode *const inode = page->mapping->host;
1307         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1308         int err;
1309         LIST_HEAD(pagepool);
1310
1311 #if (EROFS_FS_ZIP_CACHE_LVL >= 2)
1312         f.cachedzone_la = page->index << PAGE_SHIFT;
1313 #endif
1314         err = z_erofs_do_read_page(&f, page, &pagepool);
1315         (void)z_erofs_vle_work_iter_end(&f.builder);
1316
1317         if (err) {
1318                 errln("%s, failed to read, err [%d]", __func__, err);
1319                 goto out;
1320         }
1321
1322         z_erofs_submit_and_unzip(&f, &pagepool, true);
1323 out:
1324         if (f.m_iter.mpage != NULL)
1325                 put_page(f.m_iter.mpage);
1326
1327         /* clean up the remaining free pages */
1328         put_pages_list(&pagepool);
1329         return 0;
1330 }
1331
1332 static inline int __z_erofs_vle_normalaccess_readpages(
1333         struct file *filp,
1334         struct address_space *mapping,
1335         struct list_head *pages, unsigned nr_pages, bool sync)
1336 {
1337         struct inode *const inode = mapping->host;
1338
1339         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1340         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1341         struct page *head = NULL;
1342         LIST_HEAD(pagepool);
1343
1344 #if (EROFS_FS_ZIP_CACHE_LVL >= 2)
1345         f.cachedzone_la = lru_to_page(pages)->index << PAGE_SHIFT;
1346 #endif
1347         for (; nr_pages; --nr_pages) {
1348                 struct page *page = lru_to_page(pages);
1349
1350                 prefetchw(&page->flags);
1351                 list_del(&page->lru);
1352
1353                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1354                         list_add(&page->lru, &pagepool);
1355                         continue;
1356                 }
1357
1358                 BUG_ON(PagePrivate(page));
1359                 set_page_private(page, (unsigned long)head);
1360                 head = page;
1361         }
1362
1363         while (head != NULL) {
1364                 struct page *page = head;
1365                 int err;
1366
1367                 /* traversal in reverse order */
1368                 head = (void *)page_private(page);
1369
1370                 err = z_erofs_do_read_page(&f, page, &pagepool);
1371                 if (err) {
1372                         struct erofs_vnode *vi = EROFS_V(inode);
1373
1374                         errln("%s, readahead error at page %lu of nid %llu",
1375                                 __func__, page->index, vi->nid);
1376                 }
1377
1378                 put_page(page);
1379         }
1380
1381         (void)z_erofs_vle_work_iter_end(&f.builder);
1382
1383         z_erofs_submit_and_unzip(&f, &pagepool, sync);
1384
1385         if (f.m_iter.mpage != NULL)
1386                 put_page(f.m_iter.mpage);
1387
1388         /* clean up the remaining free pages */
1389         put_pages_list(&pagepool);
1390         return 0;
1391 }
1392
1393 static int z_erofs_vle_normalaccess_readpages(
1394         struct file *filp,
1395         struct address_space *mapping,
1396         struct list_head *pages, unsigned nr_pages)
1397 {
1398         return __z_erofs_vle_normalaccess_readpages(filp,
1399                 mapping, pages, nr_pages,
1400                 nr_pages < 4 /* sync */);
1401 }
1402
1403 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1404         .readpage = z_erofs_vle_normalaccess_readpage,
1405         .readpages = z_erofs_vle_normalaccess_readpages,
1406 };
1407
1408 #define __vle_cluster_advise(x, bit, bits) \
1409         ((le16_to_cpu(x) >> (bit)) & ((1 << (bits)) - 1))
1410
1411 #define __vle_cluster_type(advise) __vle_cluster_advise(advise, \
1412         Z_EROFS_VLE_DI_CLUSTER_TYPE_BIT, Z_EROFS_VLE_DI_CLUSTER_TYPE_BITS)
1413
1414 enum {
1415         Z_EROFS_VLE_CLUSTER_TYPE_PLAIN,
1416         Z_EROFS_VLE_CLUSTER_TYPE_HEAD,
1417         Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD,
1418         Z_EROFS_VLE_CLUSTER_TYPE_RESERVED,
1419         Z_EROFS_VLE_CLUSTER_TYPE_MAX
1420 };
1421
1422 #define vle_cluster_type(di)    \
1423         __vle_cluster_type((di)->di_advise)
1424
1425 static inline unsigned
1426 vle_compressed_index_clusterofs(unsigned clustersize,
1427         struct z_erofs_vle_decompressed_index *di)
1428 {
1429         debugln("%s, vle=%pK, advise=%x (type %u), clusterofs=%x blkaddr=%x",
1430                 __func__, di, di->di_advise, vle_cluster_type(di),
1431                 di->di_clusterofs, di->di_u.blkaddr);
1432
1433         switch (vle_cluster_type(di)) {
1434         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1435                 break;
1436         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1437         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1438                 return di->di_clusterofs;
1439         default:
1440                 BUG_ON(1);
1441         }
1442         return clustersize;
1443 }
1444
1445 static inline erofs_blk_t
1446 vle_extent_blkaddr(struct inode *inode, pgoff_t index)
1447 {
1448         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1449         struct erofs_vnode *vi = EROFS_V(inode);
1450
1451         unsigned ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1452                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1453                 index * sizeof(struct z_erofs_vle_decompressed_index);
1454
1455         return erofs_blknr(iloc(sbi, vi->nid) + ofs);
1456 }
1457
1458 static inline unsigned int
1459 vle_extent_blkoff(struct inode *inode, pgoff_t index)
1460 {
1461         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1462         struct erofs_vnode *vi = EROFS_V(inode);
1463
1464         unsigned ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1465                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1466                 index * sizeof(struct z_erofs_vle_decompressed_index);
1467
1468         return erofs_blkoff(iloc(sbi, vi->nid) + ofs);
1469 }
1470
1471 /*
1472  * Variable-sized Logical Extent (Fixed Physical Cluster) Compression Mode
1473  * ---
1474  * VLE compression mode attempts to compress a number of logical data into
1475  * a physical cluster with a fixed size.
1476  * VLE compression mode uses "struct z_erofs_vle_decompressed_index".
1477  */
1478 static erofs_off_t vle_get_logical_extent_head(
1479         struct inode *inode,
1480         struct page **page_iter,
1481         void **kaddr_iter,
1482         unsigned lcn,   /* logical cluster number */
1483         erofs_blk_t *pcn,
1484         unsigned *flags)
1485 {
1486         /* for extent meta */
1487         struct page *page = *page_iter;
1488         erofs_blk_t blkaddr = vle_extent_blkaddr(inode, lcn);
1489         struct z_erofs_vle_decompressed_index *di;
1490         unsigned long long ofs;
1491         const unsigned int clusterbits = EROFS_SB(inode->i_sb)->clusterbits;
1492         const unsigned int clustersize = 1 << clusterbits;
1493
1494         if (page->index != blkaddr) {
1495                 kunmap_atomic(*kaddr_iter);
1496                 unlock_page(page);
1497                 put_page(page);
1498
1499                 *page_iter = page = erofs_get_meta_page(inode->i_sb,
1500                         blkaddr, false);
1501                 *kaddr_iter = kmap_atomic(page);
1502         }
1503
1504         di = *kaddr_iter + vle_extent_blkoff(inode, lcn);
1505         switch (vle_cluster_type(di)) {
1506         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1507                 BUG_ON(!di->di_u.delta[0]);
1508                 BUG_ON(lcn < di->di_u.delta[0]);
1509
1510                 ofs = vle_get_logical_extent_head(inode,
1511                         page_iter, kaddr_iter,
1512                         lcn - di->di_u.delta[0], pcn, flags);
1513                 break;
1514         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1515                 *flags ^= EROFS_MAP_ZIPPED;
1516         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1517                 /* clustersize should be a power of two */
1518                 ofs = ((unsigned long long)lcn << clusterbits) +
1519                         (le16_to_cpu(di->di_clusterofs) & (clustersize - 1));
1520                 *pcn = le32_to_cpu(di->di_u.blkaddr);
1521                 break;
1522         default:
1523                 BUG_ON(1);
1524         }
1525         return ofs;
1526 }
1527
1528 int z_erofs_map_blocks_iter(struct inode *inode,
1529         struct erofs_map_blocks *map,
1530         struct page **mpage_ret, int flags)
1531 {
1532         /* logicial extent (start, end) offset */
1533         unsigned long long ofs, end;
1534         struct z_erofs_vle_decompressed_index *di;
1535         erofs_blk_t e_blkaddr, pcn;
1536         unsigned lcn, logical_cluster_ofs, cluster_type;
1537         u32 ofs_rem;
1538         struct page *mpage = *mpage_ret;
1539         void *kaddr;
1540         bool initial;
1541         const unsigned int clusterbits = EROFS_SB(inode->i_sb)->clusterbits;
1542         const unsigned int clustersize = 1 << clusterbits;
1543         int err = 0;
1544
1545         /* if both m_(l,p)len are 0, regularize l_lblk, l_lofs, etc... */
1546         initial = !map->m_llen;
1547
1548         /* when trying to read beyond EOF, leave it unmapped */
1549         if (unlikely(map->m_la >= inode->i_size)) {
1550                 BUG_ON(!initial);
1551                 map->m_llen = map->m_la + 1 - inode->i_size;
1552                 map->m_la = inode->i_size - 1;
1553                 map->m_flags = 0;
1554                 goto out;
1555         }
1556
1557         debugln("%s, m_la %llu m_llen %llu --- start", __func__,
1558                 map->m_la, map->m_llen);
1559
1560         ofs = map->m_la + map->m_llen;
1561
1562         /* clustersize should be power of two */
1563         lcn = ofs >> clusterbits;
1564         ofs_rem = ofs & (clustersize - 1);
1565
1566         e_blkaddr = vle_extent_blkaddr(inode, lcn);
1567
1568         if (mpage == NULL || mpage->index != e_blkaddr) {
1569                 if (mpage != NULL)
1570                         put_page(mpage);
1571
1572                 mpage = erofs_get_meta_page(inode->i_sb, e_blkaddr, false);
1573                 *mpage_ret = mpage;
1574         } else {
1575                 lock_page(mpage);
1576                 DBG_BUGON(!PageUptodate(mpage));
1577         }
1578
1579         kaddr = kmap_atomic(mpage);
1580         di = kaddr + vle_extent_blkoff(inode, lcn);
1581
1582         debugln("%s, lcn %u e_blkaddr %u e_blkoff %u", __func__, lcn,
1583                 e_blkaddr, vle_extent_blkoff(inode, lcn));
1584
1585         logical_cluster_ofs = vle_compressed_index_clusterofs(clustersize, di);
1586         if (!initial) {
1587                 /* [walking mode] 'map' has been already initialized */
1588                 map->m_llen += logical_cluster_ofs;
1589                 goto unmap_out;
1590         }
1591
1592         /* by default, compressed */
1593         map->m_flags |= EROFS_MAP_ZIPPED;
1594
1595         end = (u64)(lcn + 1) * clustersize;
1596
1597         cluster_type = vle_cluster_type(di);
1598
1599         switch (cluster_type) {
1600         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1601                 if (ofs_rem >= logical_cluster_ofs)
1602                         map->m_flags ^= EROFS_MAP_ZIPPED;
1603                 /* fallthrough */
1604         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1605                 if (ofs_rem == logical_cluster_ofs) {
1606                         pcn = le32_to_cpu(di->di_u.blkaddr);
1607                         goto exact_hitted;
1608                 }
1609
1610                 if (ofs_rem > logical_cluster_ofs) {
1611                         ofs = lcn * clustersize | logical_cluster_ofs;
1612                         pcn = le32_to_cpu(di->di_u.blkaddr);
1613                         break;
1614                 }
1615
1616                 /* logical cluster number should be >= 1 */
1617                 if (unlikely(!lcn)) {
1618                         errln("invalid logical cluster 0 at nid %llu",
1619                                 EROFS_V(inode)->nid);
1620                         err = -EIO;
1621                         goto unmap_out;
1622                 }
1623                 end = (lcn-- * clustersize) | logical_cluster_ofs;
1624                 /* fallthrough */
1625         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1626                 /* get the correspoinding first chunk */
1627                 ofs = vle_get_logical_extent_head(inode, mpage_ret,
1628                         &kaddr, lcn, &pcn, &map->m_flags);
1629                 mpage = *mpage_ret;
1630                 break;
1631         default:
1632                 errln("unknown cluster type %u at offset %llu of nid %llu",
1633                         cluster_type, ofs, EROFS_V(inode)->nid);
1634                 err = -EIO;
1635                 goto unmap_out;
1636         }
1637
1638         map->m_la = ofs;
1639 exact_hitted:
1640         map->m_llen = end - ofs;
1641         map->m_plen = clustersize;
1642         map->m_pa = blknr_to_addr(pcn);
1643         map->m_flags |= EROFS_MAP_MAPPED;
1644 unmap_out:
1645         kunmap_atomic(kaddr);
1646         unlock_page(mpage);
1647 out:
1648         debugln("%s, m_la %llu m_pa %llu m_llen %llu m_plen %llu m_flags 0%o",
1649                 __func__, map->m_la, map->m_pa,
1650                 map->m_llen, map->m_plen, map->m_flags);
1651
1652         /* aggressively BUG_ON iff CONFIG_EROFS_FS_DEBUG is on */
1653         DBG_BUGON(err < 0);
1654         return err;
1655 }
1656