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 #include <trace/events/erofs.h>
  17
  18 /*
  19  * a compressed_pages[] placeholder in order to avoid
  20  * being filled with file pages for in-place decompression.
  21  */
  22 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
  23
  24 /* how to allocate cached pages for a workgroup */
  25 enum z_erofs_cache_alloctype {
  26         DONTALLOC,      /* don't allocate any cached pages */
  27         DELAYEDALLOC,   /* delayed allocation (at the time of submitting io) */
  28 };
  29
  30 /*
  31  * tagged pointer with 1-bit tag for all compressed pages
  32  * tag 0 - the page is just found with an extra page reference
  33  */
  34 typedef tagptr1_t compressed_page_t;
  35
  36 #define tag_compressed_page_justfound(page) \
  37         tagptr_fold(compressed_page_t, page, 1)
  38
  39 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
  40 static struct kmem_cache *z_erofs_workgroup_cachep __read_mostly;
  41
  42 void z_erofs_exit_zip_subsystem(void)
  43 {
  44         destroy_workqueue(z_erofs_workqueue);
  45         kmem_cache_destroy(z_erofs_workgroup_cachep);
  46 }
  47
  48 static inline int init_unzip_workqueue(void)
  49 {
  50         const unsigned int onlinecpus = num_possible_cpus();
  51
  52         /*
  53          * we don't need too many threads, limiting threads
  54          * could improve scheduling performance.
  55          */
  56         z_erofs_workqueue =
  57                 alloc_workqueue("erofs_unzipd",
  58                                 WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE,
  59                                 onlinecpus + onlinecpus / 4);
  60
  61         return z_erofs_workqueue ? 0 : -ENOMEM;
  62 }
  63
  64 static void init_once(void *ptr)
  65 {
  66         struct z_erofs_vle_workgroup *grp = ptr;
  67         struct z_erofs_vle_work *const work =
  68                 z_erofs_vle_grab_primary_work(grp);
  69         unsigned int i;
  70
  71         mutex_init(&work->lock);
  72         work->nr_pages = 0;
  73         work->vcnt = 0;
  74         for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i)
  75                 grp->compressed_pages[i] = NULL;
  76 }
  77
  78 static void init_always(struct z_erofs_vle_workgroup *grp)
  79 {
  80         struct z_erofs_vle_work *const work =
  81                 z_erofs_vle_grab_primary_work(grp);
  82
  83         atomic_set(&grp->obj.refcount, 1);
  84         grp->flags = 0;
  85
  86         DBG_BUGON(work->nr_pages);
  87         DBG_BUGON(work->vcnt);
  88 }
  89
  90 int __init z_erofs_init_zip_subsystem(void)
  91 {
  92         z_erofs_workgroup_cachep =
  93                 kmem_cache_create("erofs_compress",
  94                                   Z_EROFS_WORKGROUP_SIZE, 0,
  95                                   SLAB_RECLAIM_ACCOUNT, init_once);
  96
  97         if (z_erofs_workgroup_cachep) {
  98                 if (!init_unzip_workqueue())
  99                         return 0;
 100
 101                 kmem_cache_destroy(z_erofs_workgroup_cachep);
 102         }
 103         return -ENOMEM;
 104 }
 105
 106 enum z_erofs_vle_work_role {
 107         Z_EROFS_VLE_WORK_SECONDARY,
 108         Z_EROFS_VLE_WORK_PRIMARY,
 109         /*
 110          * The current work has at least been linked with the following
 111          * processed chained works, which means if the processing page
 112          * is the tail partial page of the work, the current work can
 113          * safely use the whole page, as illustrated below:
 114          * +--------------+-------------------------------------------+
 115          * |  tail page   |      head page (of the previous work)     |
 116          * +--------------+-------------------------------------------+
 117          *   /\  which belongs to the current work
 118          * [  (*) this page can be used for the current work itself.  ]
 119          */
 120         Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
 121         Z_EROFS_VLE_WORK_MAX
 122 };
 123
 124 struct z_erofs_vle_work_builder {
 125         enum z_erofs_vle_work_role role;
 126         /*
 127          * 'hosted = false' means that the current workgroup doesn't belong to
 128          * the owned chained workgroups. In the other words, it is none of our
 129          * business to submit this workgroup.
 130          */
 131         bool hosted;
 132
 133         struct z_erofs_vle_workgroup *grp;
 134         struct z_erofs_vle_work *work;
 135         struct z_erofs_pagevec_ctor vector;
 136
 137         /* pages used for reading the compressed data */
 138         struct page **compressed_pages;
 139         unsigned int compressed_deficit;
 140 };
 141
 142 #define VLE_WORK_BUILDER_INIT() \
 143         { .work = NULL, .role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED }
 144
 145 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 146 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 147                                      struct address_space *mc,
 148                                      pgoff_t index,
 149                                      unsigned int clusterpages,
 150                                      enum z_erofs_cache_alloctype type,
 151                                      struct list_head *pagepool,
 152                                      gfp_t gfp)
 153 {
 154         struct page **const pages = bl->compressed_pages;
 155         const unsigned int remaining = bl->compressed_deficit;
 156         bool standalone = true;
 157         unsigned int i, j = 0;
 158
 159         if (bl->role < Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 160                 return;
 161
 162         gfp = mapping_gfp_constraint(mc, gfp) & ~__GFP_RECLAIM;
 163
 164         index += clusterpages - remaining;
 165
 166         for (i = 0; i < remaining; ++i) {
 167                 struct page *page;
 168                 compressed_page_t t;
 169
 170                 /* the compressed page was loaded before */
 171                 if (READ_ONCE(pages[i]))
 172                         continue;
 173
 174                 page = find_get_page(mc, index + i);
 175
 176                 if (page) {
 177                         t = tag_compressed_page_justfound(page);
 178                 } else if (type == DELAYEDALLOC) {
 179                         t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
 180                 } else {        /* DONTALLOC */
 181                         if (standalone)
 182                                 j = i;
 183                         standalone = false;
 184                         continue;
 185                 }
 186
 187                 if (!cmpxchg_relaxed(&pages[i], NULL, tagptr_cast_ptr(t)))
 188                         continue;
 189
 190                 if (page)
 191                         put_page(page);
 192         }
 193         bl->compressed_pages += j;
 194         bl->compressed_deficit = remaining - j;
 195
 196         if (standalone)
 197                 bl->role = Z_EROFS_VLE_WORK_PRIMARY;
 198 }
 199
 200 /* called by erofs_shrinker to get rid of all compressed_pages */
 201 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 202                                        struct erofs_workgroup *egrp)
 203 {
 204         struct z_erofs_vle_workgroup *const grp =
 205                 container_of(egrp, struct z_erofs_vle_workgroup, obj);
 206         struct address_space *const mapping = MNGD_MAPPING(sbi);
 207         const int clusterpages = erofs_clusterpages(sbi);
 208         int i;
 209
 210         /*
 211          * refcount of workgroup is now freezed as 1,
 212          * therefore no need to worry about available decompression users.
 213          */
 214         for (i = 0; i < clusterpages; ++i) {
 215                 struct page *page = grp->compressed_pages[i];
 216
 217                 if (!page || page->mapping != mapping)
 218                         continue;
 219
 220                 /* block other users from reclaiming or migrating the page */
 221                 if (!trylock_page(page))
 222                         return -EBUSY;
 223
 224                 /* barrier is implied in the following 'unlock_page' */
 225                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 226
 227                 set_page_private(page, 0);
 228                 ClearPagePrivate(page);
 229
 230                 unlock_page(page);
 231                 put_page(page);
 232         }
 233         return 0;
 234 }
 235
 236 int erofs_try_to_free_cached_page(struct address_space *mapping,
 237                                   struct page *page)
 238 {
 239         struct erofs_sb_info *const sbi = EROFS_SB(mapping->host->i_sb);
 240         const unsigned int clusterpages = erofs_clusterpages(sbi);
 241
 242         struct z_erofs_vle_workgroup *grp;
 243         int ret = 0;    /* 0 - busy */
 244
 245         /* prevent the workgroup from being freed */
 246         rcu_read_lock();
 247         grp = (void *)page_private(page);
 248
 249         if (erofs_workgroup_try_to_freeze(&grp->obj, 1)) {
 250                 unsigned int i;
 251
 252                 for (i = 0; i < clusterpages; ++i) {
 253                         if (grp->compressed_pages[i] == page) {
 254                                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 255                                 ret = 1;
 256                                 break;
 257                         }
 258                 }
 259                 erofs_workgroup_unfreeze(&grp->obj, 1);
 260         }
 261         rcu_read_unlock();
 262
 263         if (ret) {
 264                 ClearPagePrivate(page);
 265                 put_page(page);
 266         }
 267         return ret;
 268 }
 269 #else
 270 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 271                                      struct address_space *mc,
 272                                      pgoff_t index,
 273                                      unsigned int clusterpages,
 274                                      enum z_erofs_cache_alloctype type,
 275                                      struct list_head *pagepool,
 276                                      gfp_t gfp)
 277 {
 278         /* nowhere to load compressed pages from */
 279 }
 280 #endif
 281
 282 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 283 static inline bool try_to_reuse_as_compressed_page(
 284         struct z_erofs_vle_work_builder *b,
 285         struct page *page)
 286 {
 287         while (b->compressed_deficit) {
 288                 --b->compressed_deficit;
 289                 if (!cmpxchg(b->compressed_pages++, NULL, page))
 290                         return true;
 291         }
 292
 293         return false;
 294 }
 295
 296 /* callers must be with work->lock held */
 297 static int z_erofs_vle_work_add_page(
 298         struct z_erofs_vle_work_builder *builder,
 299         struct page *page,
 300         enum z_erofs_page_type type)
 301 {
 302         int ret;
 303         bool occupied;
 304
 305         /* give priority for the compressed data storage */
 306         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY &&
 307                 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 308                 try_to_reuse_as_compressed_page(builder, page))
 309                 return 0;
 310
 311         ret = z_erofs_pagevec_ctor_enqueue(&builder->vector,
 312                 page, type, &occupied);
 313         builder->work->vcnt += (unsigned int)ret;
 314
 315         return ret ? 0 : -EAGAIN;
 316 }
 317
 318 static inline bool try_to_claim_workgroup(
 319         struct z_erofs_vle_workgroup *grp,
 320         z_erofs_vle_owned_workgrp_t *owned_head,
 321         bool *hosted)
 322 {
 323         DBG_BUGON(*hosted == true);
 324
 325         /* let's claim these following types of workgroup */
 326 retry:
 327         if (grp->next == Z_EROFS_VLE_WORKGRP_NIL) {
 328                 /* type 1, nil workgroup */
 329                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_NIL,
 330                             *owned_head) != Z_EROFS_VLE_WORKGRP_NIL)
 331                         goto retry;
 332
 333                 *owned_head = &grp->next;
 334                 *hosted = true;
 335         } else if (grp->next == Z_EROFS_VLE_WORKGRP_TAIL) {
 336                 /*
 337                  * type 2, link to the end of a existing open chain,
 338                  * be careful that its submission itself is governed
 339                  * by the original owned chain.
 340                  */
 341                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
 342                             *owned_head) != Z_EROFS_VLE_WORKGRP_TAIL)
 343                         goto retry;
 344
 345                 *owned_head = Z_EROFS_VLE_WORKGRP_TAIL;
 346         } else
 347                 return false;   /* :( better luck next time */
 348
 349         return true;    /* lucky, I am the followee :) */
 350 }
 351
 352 struct z_erofs_vle_work_finder {
 353         struct super_block *sb;
 354         pgoff_t idx;
 355         unsigned int pageofs;
 356
 357         struct z_erofs_vle_workgroup **grp_ret;
 358         enum z_erofs_vle_work_role *role;
 359         z_erofs_vle_owned_workgrp_t *owned_head;
 360         bool *hosted;
 361 };
 362
 363 static struct z_erofs_vle_work *
 364 z_erofs_vle_work_lookup(const struct z_erofs_vle_work_finder *f)
 365 {
 366         bool tag, primary;
 367         struct erofs_workgroup *egrp;
 368         struct z_erofs_vle_workgroup *grp;
 369         struct z_erofs_vle_work *work;
 370
 371         egrp = erofs_find_workgroup(f->sb, f->idx, &tag);
 372         if (!egrp) {
 373                 *f->grp_ret = NULL;
 374                 return NULL;
 375         }
 376
 377         grp = container_of(egrp, struct z_erofs_vle_workgroup, obj);
 378         *f->grp_ret = grp;
 379
 380         work = z_erofs_vle_grab_work(grp, f->pageofs);
 381         /* if multiref is disabled, `primary' is always true */
 382         primary = true;
 383
 384         DBG_BUGON(work->pageofs != f->pageofs);
 385
 386         /*
 387          * lock must be taken first to avoid grp->next == NIL between
 388          * claiming workgroup and adding pages:
 389          *                        grp->next != NIL
 390          *   grp->next = NIL
 391          *   mutex_unlock_all
 392          *                        mutex_lock(&work->lock)
 393          *                        add all pages to pagevec
 394          *
 395          * [correct locking case 1]:
 396          *   mutex_lock(grp->work[a])
 397          *   ...
 398          *   mutex_lock(grp->work[b])     mutex_lock(grp->work[c])
 399          *   ...                          *role = SECONDARY
 400          *                                add all pages to pagevec
 401          *                                ...
 402          *                                mutex_unlock(grp->work[c])
 403          *   mutex_lock(grp->work[c])
 404          *   ...
 405          *   grp->next = NIL
 406          *   mutex_unlock_all
 407          *
 408          * [correct locking case 2]:
 409          *   mutex_lock(grp->work[b])
 410          *   ...
 411          *   mutex_lock(grp->work[a])
 412          *   ...
 413          *   mutex_lock(grp->work[c])
 414          *   ...
 415          *   grp->next = NIL
 416          *   mutex_unlock_all
 417          *                                mutex_lock(grp->work[a])
 418          *                                *role = PRIMARY_OWNER
 419          *                                add all pages to pagevec
 420          *                                ...
 421          */
 422         mutex_lock(&work->lock);
 423
 424         *f->hosted = false;
 425         if (!primary)
 426                 *f->role = Z_EROFS_VLE_WORK_SECONDARY;
 427         /* claim the workgroup if possible */
 428         else if (try_to_claim_workgroup(grp, f->owned_head, f->hosted))
 429                 *f->role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 430         else
 431                 *f->role = Z_EROFS_VLE_WORK_PRIMARY;
 432
 433         return work;
 434 }
 435
 436 static struct z_erofs_vle_work *
 437 z_erofs_vle_work_register(const struct z_erofs_vle_work_finder *f,
 438                           struct erofs_map_blocks *map)
 439 {
 440         bool gnew = false;
 441         struct z_erofs_vle_workgroup *grp = *f->grp_ret;
 442         struct z_erofs_vle_work *work;
 443
 444         /* if multiref is disabled, grp should never be nullptr */
 445         if (unlikely(grp)) {
 446                 DBG_BUGON(1);
 447                 return ERR_PTR(-EINVAL);
 448         }
 449
 450         /* no available workgroup, let's allocate one */
 451         grp = kmem_cache_alloc(z_erofs_workgroup_cachep, GFP_NOFS);
 452         if (unlikely(!grp))
 453                 return ERR_PTR(-ENOMEM);
 454
 455         init_always(grp);
 456         grp->obj.index = f->idx;
 457         grp->llen = map->m_llen;
 458
 459         z_erofs_vle_set_workgrp_fmt(grp,
 460                 (map->m_flags & EROFS_MAP_ZIPPED) ?
 461                         Z_EROFS_VLE_WORKGRP_FMT_LZ4 :
 462                         Z_EROFS_VLE_WORKGRP_FMT_PLAIN);
 463
 464         /* new workgrps have been claimed as type 1 */
 465         WRITE_ONCE(grp->next, *f->owned_head);
 466         /* primary and followed work for all new workgrps */
 467         *f->role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 468         /* it should be submitted by ourselves */
 469         *f->hosted = true;
 470
 471         gnew = true;
 472         work = z_erofs_vle_grab_primary_work(grp);
 473         work->pageofs = f->pageofs;
 474
 475         /*
 476          * lock all primary followed works before visible to others
 477          * and mutex_trylock *never* fails for a new workgroup.
 478          */
 479         mutex_trylock(&work->lock);
 480
 481         if (gnew) {
 482                 int err = erofs_register_workgroup(f->sb, &grp->obj, 0);
 483
 484                 if (err) {
 485                         mutex_unlock(&work->lock);
 486                         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 487                         return ERR_PTR(-EAGAIN);
 488                 }
 489         }
 490
 491         *f->owned_head = &grp->next;
 492         *f->grp_ret = grp;
 493         return work;
 494 }
 495
 496 #define builder_is_followed(builder) \
 497         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 498
 499 static int z_erofs_vle_work_iter_begin(struct z_erofs_vle_work_builder *builder,
 500                                        struct super_block *sb,
 501                                        struct erofs_map_blocks *map,
 502                                        z_erofs_vle_owned_workgrp_t *owned_head)
 503 {
 504         const unsigned int clusterpages = erofs_clusterpages(EROFS_SB(sb));
 505         struct z_erofs_vle_workgroup *grp;
 506         const struct z_erofs_vle_work_finder finder = {
 507                 .sb = sb,
 508                 .idx = erofs_blknr(map->m_pa),
 509                 .pageofs = map->m_la & ~PAGE_MASK,
 510                 .grp_ret = &grp,
 511                 .role = &builder->role,
 512                 .owned_head = owned_head,
 513                 .hosted = &builder->hosted
 514         };
 515         struct z_erofs_vle_work *work;
 516
 517         DBG_BUGON(builder->work);
 518
 519         /* must be Z_EROFS_WORK_TAIL or the next chained work */
 520         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_NIL);
 521         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
 522
 523         DBG_BUGON(erofs_blkoff(map->m_pa));
 524
 525 repeat:
 526         work = z_erofs_vle_work_lookup(&finder);
 527         if (work) {
 528                 unsigned int orig_llen;
 529
 530                 /* increase workgroup `llen' if needed */
 531                 while ((orig_llen = READ_ONCE(grp->llen)) < map->m_llen &&
 532                        orig_llen != cmpxchg_relaxed(&grp->llen,
 533                                                     orig_llen, map->m_llen))
 534                         cpu_relax();
 535                 goto got_it;
 536         }
 537
 538         work = z_erofs_vle_work_register(&finder, map);
 539         if (unlikely(work == ERR_PTR(-EAGAIN)))
 540                 goto repeat;
 541
 542         if (unlikely(IS_ERR(work)))
 543                 return PTR_ERR(work);
 544 got_it:
 545         z_erofs_pagevec_ctor_init(&builder->vector,
 546                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, work->vcnt);
 547
 548         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY) {
 549                 /* enable possibly in-place decompression */
 550                 builder->compressed_pages = grp->compressed_pages;
 551                 builder->compressed_deficit = clusterpages;
 552         } else {
 553                 builder->compressed_pages = NULL;
 554                 builder->compressed_deficit = 0;
 555         }
 556
 557         builder->grp = grp;
 558         builder->work = work;
 559         return 0;
 560 }
 561
 562 /*
 563  * keep in mind that no referenced workgroups will be freed
 564  * only after a RCU grace period, so rcu_read_lock() could
 565  * prevent a workgroup from being freed.
 566  */
 567 static void z_erofs_rcu_callback(struct rcu_head *head)
 568 {
 569         struct z_erofs_vle_work *work = container_of(head,
 570                 struct z_erofs_vle_work, rcu);
 571         struct z_erofs_vle_workgroup *grp =
 572                 z_erofs_vle_work_workgroup(work, true);
 573
 574         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 575 }
 576
 577 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 578 {
 579         struct z_erofs_vle_workgroup *const vgrp = container_of(grp,
 580                 struct z_erofs_vle_workgroup, obj);
 581         struct z_erofs_vle_work *const work = &vgrp->work;
 582
 583         call_rcu(&work->rcu, z_erofs_rcu_callback);
 584 }
 585
 586 static void __z_erofs_vle_work_release(struct z_erofs_vle_workgroup *grp,
 587         struct z_erofs_vle_work *work __maybe_unused)
 588 {
 589         erofs_workgroup_put(&grp->obj);
 590 }
 591
 592 void z_erofs_vle_work_release(struct z_erofs_vle_work *work)
 593 {
 594         struct z_erofs_vle_workgroup *grp =
 595                 z_erofs_vle_work_workgroup(work, true);
 596
 597         __z_erofs_vle_work_release(grp, work);
 598 }
 599
 600 static inline bool
 601 z_erofs_vle_work_iter_end(struct z_erofs_vle_work_builder *builder)
 602 {
 603         struct z_erofs_vle_work *work = builder->work;
 604
 605         if (!work)
 606                 return false;
 607
 608         z_erofs_pagevec_ctor_exit(&builder->vector, false);
 609         mutex_unlock(&work->lock);
 610
 611         /*
 612          * if all pending pages are added, don't hold work reference
 613          * any longer if the current work isn't hosted by ourselves.
 614          */
 615         if (!builder->hosted)
 616                 __z_erofs_vle_work_release(builder->grp, work);
 617
 618         builder->work = NULL;
 619         builder->grp = NULL;
 620         return true;
 621 }
 622
 623 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 624                                                gfp_t gfp)
 625 {
 626         struct page *page = erofs_allocpage(pagepool, gfp);
 627
 628         if (unlikely(!page))
 629                 return NULL;
 630
 631         page->mapping = Z_EROFS_MAPPING_STAGING;
 632         return page;
 633 }
 634
 635 struct z_erofs_vle_frontend {
 636         struct inode *const inode;
 637
 638         struct z_erofs_vle_work_builder builder;
 639         struct erofs_map_blocks_iter m_iter;
 640
 641         z_erofs_vle_owned_workgrp_t owned_head;
 642
 643         /* used for applying cache strategy on the fly */
 644         bool backmost;
 645         erofs_off_t headoffset;
 646 };
 647
 648 #define VLE_FRONTEND_INIT(__i) { \
 649         .inode = __i, \
 650         .m_iter = { \
 651                 { .m_llen = 0, .m_plen = 0 }, \
 652                 .mpage = NULL \
 653         }, \
 654         .builder = VLE_WORK_BUILDER_INIT(), \
 655         .owned_head = Z_EROFS_VLE_WORKGRP_TAIL, \
 656         .backmost = true, }
 657
 658 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 659 static inline bool
 660 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 661 {
 662         if (fe->backmost)
 663                 return true;
 664
 665         if (EROFS_FS_ZIP_CACHE_LVL >= 2)
 666                 return la < fe->headoffset;
 667
 668         return false;
 669 }
 670 #else
 671 static inline bool
 672 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 673 {
 674         return false;
 675 }
 676 #endif
 677
 678 static int z_erofs_do_read_page(struct z_erofs_vle_frontend *fe,
 679                                 struct page *page,
 680                                 struct list_head *page_pool)
 681 {
 682         struct super_block *const sb = fe->inode->i_sb;
 683         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
 684         struct erofs_map_blocks_iter *const m = &fe->m_iter;
 685         struct erofs_map_blocks *const map = &m->map;
 686         struct z_erofs_vle_work_builder *const builder = &fe->builder;
 687         const loff_t offset = page_offset(page);
 688
 689         bool tight = builder_is_followed(builder);
 690         struct z_erofs_vle_work *work = builder->work;
 691
 692         enum z_erofs_cache_alloctype cache_strategy;
 693         enum z_erofs_page_type page_type;
 694         unsigned int cur, end, spiltted, index;
 695         int err = 0;
 696
 697         /* register locked file pages as online pages in pack */
 698         z_erofs_onlinepage_init(page);
 699
 700         spiltted = 0;
 701         end = PAGE_SIZE;
 702 repeat:
 703         cur = end - 1;
 704
 705         /* lucky, within the range of the current map_blocks */
 706         if (offset + cur >= map->m_la &&
 707                 offset + cur < map->m_la + map->m_llen)
 708                 goto hitted;
 709
 710         /* go ahead the next map_blocks */
 711         debugln("%s: [out-of-range] pos %llu", __func__, offset + cur);
 712
 713         if (z_erofs_vle_work_iter_end(builder))
 714                 fe->backmost = false;
 715
 716         map->m_la = offset + cur;
 717         map->m_llen = 0;
 718         err = erofs_map_blocks_iter(fe->inode, map, &m->mpage, 0);
 719         if (unlikely(err))
 720                 goto err_out;
 721
 722         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED)))
 723                 goto hitted;
 724
 725         DBG_BUGON(map->m_plen != 1 << sbi->clusterbits);
 726         DBG_BUGON(erofs_blkoff(map->m_pa));
 727
 728         err = z_erofs_vle_work_iter_begin(builder, sb, map, &fe->owned_head);
 729         if (unlikely(err))
 730                 goto err_out;
 731
 732         /* preload all compressed pages (maybe downgrade role if necessary) */
 733         if (should_alloc_managed_pages(fe, map->m_la))
 734                 cache_strategy = DELAYEDALLOC;
 735         else
 736                 cache_strategy = DONTALLOC;
 737
 738         preload_compressed_pages(builder, MNGD_MAPPING(sbi),
 739                                  map->m_pa / PAGE_SIZE,
 740                                  map->m_plen / PAGE_SIZE,
 741                                  cache_strategy, page_pool, GFP_KERNEL);
 742
 743         tight &= builder_is_followed(builder);
 744         work = builder->work;
 745 hitted:
 746         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 747         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED))) {
 748                 zero_user_segment(page, cur, end);
 749                 goto next_part;
 750         }
 751
 752         /* let's derive page type */
 753         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 754                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 755                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 756                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 757
 758 retry:
 759         err = z_erofs_vle_work_add_page(builder, page, page_type);
 760         /* should allocate an additional staging page for pagevec */
 761         if (err == -EAGAIN) {
 762                 struct page *const newpage =
 763                         __stagingpage_alloc(page_pool, GFP_NOFS);
 764
 765                 err = z_erofs_vle_work_add_page(builder,
 766                         newpage, Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 767                 if (likely(!err))
 768                         goto retry;
 769         }
 770
 771         if (unlikely(err))
 772                 goto err_out;
 773
 774         index = page->index - map->m_la / PAGE_SIZE;
 775
 776         /* FIXME! avoid the last relundant fixup & endio */
 777         z_erofs_onlinepage_fixup(page, index, true);
 778
 779         /* bump up the number of spiltted parts of a page */
 780         ++spiltted;
 781         /* also update nr_pages */
 782         work->nr_pages = max_t(pgoff_t, work->nr_pages, index + 1);
 783 next_part:
 784         /* can be used for verification */
 785         map->m_llen = offset + cur - map->m_la;
 786
 787         end = cur;
 788         if (end > 0)
 789                 goto repeat;
 790
 791 out:
 792         /* FIXME! avoid the last relundant fixup & endio */
 793         z_erofs_onlinepage_endio(page);
 794
 795         debugln("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 796                 __func__, page, spiltted, map->m_llen);
 797         return err;
 798
 799         /* if some error occurred while processing this page */
 800 err_out:
 801         SetPageError(page);
 802         goto out;
 803 }
 804
 805 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 806 {
 807         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 808         struct z_erofs_vle_unzip_io *io = tagptr_unfold_ptr(t);
 809         bool background = tagptr_unfold_tags(t);
 810
 811         if (!background) {
 812                 unsigned long flags;
 813
 814                 spin_lock_irqsave(&io->u.wait.lock, flags);
 815                 if (!atomic_add_return(bios, &io->pending_bios))
 816                         wake_up_locked(&io->u.wait);
 817                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 818                 return;
 819         }
 820
 821         if (!atomic_add_return(bios, &io->pending_bios))
 822                 queue_work(z_erofs_workqueue, &io->u.work);
 823 }
 824
 825 static inline void z_erofs_vle_read_endio(struct bio *bio)
 826 {
 827         const blk_status_t err = bio->bi_status;
 828         unsigned int i;
 829         struct bio_vec *bvec;
 830 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 831         struct address_space *mc = NULL;
 832 #endif
 833
 834         bio_for_each_segment_all(bvec, bio, i) {
 835                 struct page *page = bvec->bv_page;
 836                 bool cachemngd = false;
 837
 838                 DBG_BUGON(PageUptodate(page));
 839                 DBG_BUGON(!page->mapping);
 840
 841 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 842                 if (unlikely(!mc && !z_erofs_is_stagingpage(page))) {
 843                         struct inode *const inode = page->mapping->host;
 844                         struct super_block *const sb = inode->i_sb;
 845
 846                         mc = MNGD_MAPPING(EROFS_SB(sb));
 847                 }
 848
 849                 /*
 850                  * If mc has not gotten, it equals NULL,
 851                  * however, page->mapping never be NULL if working properly.
 852                  */
 853                 cachemngd = (page->mapping == mc);
 854 #endif
 855
 856                 if (unlikely(err))
 857                         SetPageError(page);
 858                 else if (cachemngd)
 859                         SetPageUptodate(page);
 860
 861                 if (cachemngd)
 862                         unlock_page(page);
 863         }
 864
 865         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 866         bio_put(bio);
 867 }
 868
 869 static struct page *z_pagemap_global[Z_EROFS_VLE_VMAP_GLOBAL_PAGES];
 870 static DEFINE_MUTEX(z_pagemap_global_lock);
 871
 872 static int z_erofs_vle_unzip(struct super_block *sb,
 873         struct z_erofs_vle_workgroup *grp,
 874         struct list_head *page_pool)
 875 {
 876         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 877         const unsigned int clusterpages = erofs_clusterpages(sbi);
 878
 879         struct z_erofs_pagevec_ctor ctor;
 880         unsigned int nr_pages;
 881         unsigned int sparsemem_pages = 0;
 882         struct page *pages_onstack[Z_EROFS_VLE_VMAP_ONSTACK_PAGES];
 883         struct page **pages, **compressed_pages, *page;
 884         unsigned int i, llen;
 885
 886         enum z_erofs_page_type page_type;
 887         bool overlapped;
 888         struct z_erofs_vle_work *work;
 889         void *vout;
 890         int err;
 891
 892         might_sleep();
 893         work = z_erofs_vle_grab_primary_work(grp);
 894         DBG_BUGON(!READ_ONCE(work->nr_pages));
 895
 896         mutex_lock(&work->lock);
 897         nr_pages = work->nr_pages;
 898
 899         if (likely(nr_pages <= Z_EROFS_VLE_VMAP_ONSTACK_PAGES))
 900                 pages = pages_onstack;
 901         else if (nr_pages <= Z_EROFS_VLE_VMAP_GLOBAL_PAGES &&
 902                 mutex_trylock(&z_pagemap_global_lock))
 903                 pages = z_pagemap_global;
 904         else {
 905 repeat:
 906                 pages = kvmalloc_array(nr_pages,
 907                         sizeof(struct page *), GFP_KERNEL);
 908
 909                 /* fallback to global pagemap for the lowmem scenario */
 910                 if (unlikely(!pages)) {
 911                         if (nr_pages > Z_EROFS_VLE_VMAP_GLOBAL_PAGES)
 912                                 goto repeat;
 913                         else {
 914                                 mutex_lock(&z_pagemap_global_lock);
 915                                 pages = z_pagemap_global;
 916                         }
 917                 }
 918         }
 919
 920         for (i = 0; i < nr_pages; ++i)
 921                 pages[i] = NULL;
 922
 923         z_erofs_pagevec_ctor_init(&ctor,
 924                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, 0);
 925
 926         for (i = 0; i < work->vcnt; ++i) {
 927                 unsigned int pagenr;
 928
 929                 page = z_erofs_pagevec_ctor_dequeue(&ctor, &page_type);
 930
 931                 /* all pages in pagevec ought to be valid */
 932                 DBG_BUGON(!page);
 933                 DBG_BUGON(!page->mapping);
 934
 935                 if (z_erofs_gather_if_stagingpage(page_pool, page))
 936                         continue;
 937
 938                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 939                         pagenr = 0;
 940                 else
 941                         pagenr = z_erofs_onlinepage_index(page);
 942
 943                 DBG_BUGON(pagenr >= nr_pages);
 944                 DBG_BUGON(pages[pagenr]);
 945
 946                 pages[pagenr] = page;
 947         }
 948         sparsemem_pages = i;
 949
 950         z_erofs_pagevec_ctor_exit(&ctor, true);
 951
 952         overlapped = false;
 953         compressed_pages = grp->compressed_pages;
 954
 955         for (i = 0; i < clusterpages; ++i) {
 956                 unsigned int pagenr;
 957
 958                 page = compressed_pages[i];
 959
 960                 /* all compressed pages ought to be valid */
 961                 DBG_BUGON(!page);
 962                 DBG_BUGON(!page->mapping);
 963
 964                 if (z_erofs_is_stagingpage(page))
 965                         continue;
 966 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 967                 if (page->mapping == MNGD_MAPPING(sbi)) {
 968                         DBG_BUGON(!PageUptodate(page));
 969                         continue;
 970                 }
 971 #endif
 972
 973                 /* only non-head page could be reused as a compressed page */
 974                 pagenr = z_erofs_onlinepage_index(page);
 975
 976                 DBG_BUGON(pagenr >= nr_pages);
 977                 DBG_BUGON(pages[pagenr]);
 978                 ++sparsemem_pages;
 979                 pages[pagenr] = page;
 980
 981                 overlapped = true;
 982         }
 983
 984         llen = (nr_pages << PAGE_SHIFT) - work->pageofs;
 985
 986         if (z_erofs_vle_workgrp_fmt(grp) == Z_EROFS_VLE_WORKGRP_FMT_PLAIN) {
 987                 err = z_erofs_vle_plain_copy(compressed_pages, clusterpages,
 988                         pages, nr_pages, work->pageofs);
 989                 goto out;
 990         }
 991
 992         if (llen > grp->llen)
 993                 llen = grp->llen;
 994
 995         err = z_erofs_vle_unzip_fast_percpu(compressed_pages,
 996                 clusterpages, pages, llen, work->pageofs,
 997                 z_erofs_onlinepage_endio);
 998         if (err != -ENOTSUPP)
 999                 goto out_percpu;
1000
1001         if (sparsemem_pages >= nr_pages)
1002                 goto skip_allocpage;
1003
1004         for (i = 0; i < nr_pages; ++i) {
1005                 if (pages[i])
1006                         continue;
1007
1008                 pages[i] = __stagingpage_alloc(page_pool, GFP_NOFS);
1009         }
1010
1011 skip_allocpage:
1012         vout = erofs_vmap(pages, nr_pages);
1013
1014         err = z_erofs_vle_unzip_vmap(compressed_pages,
1015                 clusterpages, vout, llen, work->pageofs, overlapped);
1016
1017         erofs_vunmap(vout, nr_pages);
1018
1019 out:
1020         for (i = 0; i < nr_pages; ++i) {
1021                 page = pages[i];
1022                 DBG_BUGON(!page->mapping);
1023
1024                 /* recycle all individual staging pages */
1025                 if (z_erofs_gather_if_stagingpage(page_pool, page))
1026                         continue;
1027
1028                 if (unlikely(err < 0))
1029                         SetPageError(page);
1030
1031                 z_erofs_onlinepage_endio(page);
1032         }
1033
1034 out_percpu:
1035         for (i = 0; i < clusterpages; ++i) {
1036                 page = compressed_pages[i];
1037
1038 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1039                 if (page->mapping == MNGD_MAPPING(sbi))
1040                         continue;
1041 #endif
1042                 /* recycle all individual staging pages */
1043                 (void)z_erofs_gather_if_stagingpage(page_pool, page);
1044
1045                 WRITE_ONCE(compressed_pages[i], NULL);
1046         }
1047
1048         if (pages == z_pagemap_global)
1049                 mutex_unlock(&z_pagemap_global_lock);
1050         else if (unlikely(pages != pages_onstack))
1051                 kvfree(pages);
1052
1053         work->nr_pages = 0;
1054         work->vcnt = 0;
1055
1056         /* all work locks MUST be taken before the following line */
1057
1058         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_NIL);
1059
1060         /* all work locks SHOULD be released right now */
1061         mutex_unlock(&work->lock);
1062
1063         z_erofs_vle_work_release(work);
1064         return err;
1065 }
1066
1067 static void z_erofs_vle_unzip_all(struct super_block *sb,
1068                                   struct z_erofs_vle_unzip_io *io,
1069                                   struct list_head *page_pool)
1070 {
1071         z_erofs_vle_owned_workgrp_t owned = io->head;
1072
1073         while (owned != Z_EROFS_VLE_WORKGRP_TAIL_CLOSED) {
1074                 struct z_erofs_vle_workgroup *grp;
1075
1076                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1077                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_TAIL);
1078
1079                 /* no possible that 'owned' equals NULL */
1080                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_NIL);
1081
1082                 grp = container_of(owned, struct z_erofs_vle_workgroup, next);
1083                 owned = READ_ONCE(grp->next);
1084
1085                 z_erofs_vle_unzip(sb, grp, page_pool);
1086         }
1087 }
1088
1089 static void z_erofs_vle_unzip_wq(struct work_struct *work)
1090 {
1091         struct z_erofs_vle_unzip_io_sb *iosb = container_of(work,
1092                 struct z_erofs_vle_unzip_io_sb, io.u.work);
1093         LIST_HEAD(page_pool);
1094
1095         DBG_BUGON(iosb->io.head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1096         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &page_pool);
1097
1098         put_pages_list(&page_pool);
1099         kvfree(iosb);
1100 }
1101
1102 static struct page *
1103 pickup_page_for_submission(struct z_erofs_vle_workgroup *grp,
1104                            unsigned int nr,
1105                            struct list_head *pagepool,
1106                            struct address_space *mc,
1107                            gfp_t gfp)
1108 {
1109         /* determined at compile time to avoid too many #ifdefs */
1110         const bool nocache = __builtin_constant_p(mc) ? !mc : false;
1111         const pgoff_t index = grp->obj.index;
1112         bool tocache = false;
1113
1114         struct address_space *mapping;
1115         struct page *oldpage, *page;
1116
1117         compressed_page_t t;
1118         int justfound;
1119
1120 repeat:
1121         page = READ_ONCE(grp->compressed_pages[nr]);
1122         oldpage = page;
1123
1124         if (!page)
1125                 goto out_allocpage;
1126
1127         /*
1128          * the cached page has not been allocated and
1129          * an placeholder is out there, prepare it now.
1130          */
1131         if (!nocache && page == PAGE_UNALLOCATED) {
1132                 tocache = true;
1133                 goto out_allocpage;
1134         }
1135
1136         /* process the target tagged pointer */
1137         t = tagptr_init(compressed_page_t, page);
1138         justfound = tagptr_unfold_tags(t);
1139         page = tagptr_unfold_ptr(t);
1140
1141         mapping = READ_ONCE(page->mapping);
1142
1143         /*
1144          * if managed cache is disabled, it's no way to
1145          * get such a cached-like page.
1146          */
1147         if (nocache) {
1148                 /* if managed cache is disabled, it is impossible `justfound' */
1149                 DBG_BUGON(justfound);
1150
1151                 /* and it should be locked, not uptodate, and not truncated */
1152                 DBG_BUGON(!PageLocked(page));
1153                 DBG_BUGON(PageUptodate(page));
1154                 DBG_BUGON(!mapping);
1155                 goto out;
1156         }
1157
1158         /*
1159          * unmanaged (file) pages are all locked solidly,
1160          * therefore it is impossible for `mapping' to be NULL.
1161          */
1162         if (mapping && mapping != mc)
1163                 /* ought to be unmanaged pages */
1164                 goto out;
1165
1166         lock_page(page);
1167
1168         /* only true if page reclaim goes wrong, should never happen */
1169         DBG_BUGON(justfound && PagePrivate(page));
1170
1171         /* the page is still in manage cache */
1172         if (page->mapping == mc) {
1173                 WRITE_ONCE(grp->compressed_pages[nr], page);
1174
1175                 if (!PagePrivate(page)) {
1176                         /*
1177                          * impossible to be !PagePrivate(page) for
1178                          * the current restriction as well if
1179                          * the page is already in compressed_pages[].
1180                          */
1181                         DBG_BUGON(!justfound);
1182
1183                         justfound = 0;
1184                         set_page_private(page, (unsigned long)grp);
1185                         SetPagePrivate(page);
1186                 }
1187
1188                 /* no need to submit io if it is already up-to-date */
1189                 if (PageUptodate(page)) {
1190                         unlock_page(page);
1191                         page = NULL;
1192                 }
1193                 goto out;
1194         }
1195
1196         /*
1197          * the managed page has been truncated, it's unsafe to
1198          * reuse this one, let's allocate a new cache-managed page.
1199          */
1200         DBG_BUGON(page->mapping);
1201         DBG_BUGON(!justfound);
1202
1203         tocache = true;
1204         unlock_page(page);
1205         put_page(page);
1206 out_allocpage:
1207         page = __stagingpage_alloc(pagepool, gfp);
1208         if (oldpage != cmpxchg(&grp->compressed_pages[nr], oldpage, page)) {
1209                 list_add(&page->lru, pagepool);
1210                 cpu_relax();
1211                 goto repeat;
1212         }
1213         if (nocache || !tocache)
1214                 goto out;
1215         if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1216                 page->mapping = Z_EROFS_MAPPING_STAGING;
1217                 goto out;
1218         }
1219
1220         set_page_private(page, (unsigned long)grp);
1221         SetPagePrivate(page);
1222 out:    /* the only exit (for tracing and debugging) */
1223         return page;
1224 }
1225
1226 static struct z_erofs_vle_unzip_io *
1227 jobqueue_init(struct super_block *sb,
1228               struct z_erofs_vle_unzip_io *io,
1229               bool foreground)
1230 {
1231         struct z_erofs_vle_unzip_io_sb *iosb;
1232
1233         if (foreground) {
1234                 /* waitqueue available for foreground io */
1235                 DBG_BUGON(!io);
1236
1237                 init_waitqueue_head(&io->u.wait);
1238                 atomic_set(&io->pending_bios, 0);
1239                 goto out;
1240         }
1241
1242         iosb = kvzalloc(sizeof(struct z_erofs_vle_unzip_io_sb),
1243                         GFP_KERNEL | __GFP_NOFAIL);
1244         DBG_BUGON(!iosb);
1245
1246         /* initialize fields in the allocated descriptor */
1247         io = &iosb->io;
1248         iosb->sb = sb;
1249         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1250 out:
1251         io->head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1252         return io;
1253 }
1254
1255 /* define workgroup jobqueue types */
1256 enum {
1257 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1258         JQ_BYPASS,
1259 #endif
1260         JQ_SUBMIT,
1261         NR_JOBQUEUES,
1262 };
1263
1264 static void *jobqueueset_init(struct super_block *sb,
1265                               z_erofs_vle_owned_workgrp_t qtail[],
1266                               struct z_erofs_vle_unzip_io *q[],
1267                               struct z_erofs_vle_unzip_io *fgq,
1268                               bool forcefg)
1269 {
1270 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1271         /*
1272          * if managed cache is enabled, bypass jobqueue is needed,
1273          * no need to read from device for all workgroups in this queue.
1274          */
1275         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1276         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1277 #endif
1278
1279         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1280         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1281
1282         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1283 }
1284
1285 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1286 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1287                                     z_erofs_vle_owned_workgrp_t qtail[],
1288                                     z_erofs_vle_owned_workgrp_t owned_head)
1289 {
1290         z_erofs_vle_owned_workgrp_t *const submit_qtail = qtail[JQ_SUBMIT];
1291         z_erofs_vle_owned_workgrp_t *const bypass_qtail = qtail[JQ_BYPASS];
1292
1293         DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1294         if (owned_head == Z_EROFS_VLE_WORKGRP_TAIL)
1295                 owned_head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1296
1297         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1298
1299         WRITE_ONCE(*submit_qtail, owned_head);
1300         WRITE_ONCE(*bypass_qtail, &grp->next);
1301
1302         qtail[JQ_BYPASS] = &grp->next;
1303 }
1304
1305 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1306                                        unsigned int nr_bios,
1307                                        bool force_fg)
1308 {
1309         /*
1310          * although background is preferred, no one is pending for submission.
1311          * don't issue workqueue for decompression but drop it directly instead.
1312          */
1313         if (force_fg || nr_bios)
1314                 return false;
1315
1316         kvfree(container_of(q[JQ_SUBMIT],
1317                             struct z_erofs_vle_unzip_io_sb,
1318                             io));
1319         return true;
1320 }
1321 #else
1322 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1323                                     z_erofs_vle_owned_workgrp_t qtail[],
1324                                     z_erofs_vle_owned_workgrp_t owned_head)
1325 {
1326         /* impossible to bypass submission for managed cache disabled */
1327         DBG_BUGON(1);
1328 }
1329
1330 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1331                                        unsigned int nr_bios,
1332                                        bool force_fg)
1333 {
1334         /* bios should be >0 if managed cache is disabled */
1335         DBG_BUGON(!nr_bios);
1336         return false;
1337 }
1338 #endif
1339
1340 static bool z_erofs_vle_submit_all(struct super_block *sb,
1341                                    z_erofs_vle_owned_workgrp_t owned_head,
1342                                    struct list_head *pagepool,
1343                                    struct z_erofs_vle_unzip_io *fgq,
1344                                    bool force_fg)
1345 {
1346         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1347         const unsigned int clusterpages = erofs_clusterpages(sbi);
1348         const gfp_t gfp = GFP_NOFS;
1349
1350         z_erofs_vle_owned_workgrp_t qtail[NR_JOBQUEUES];
1351         struct z_erofs_vle_unzip_io *q[NR_JOBQUEUES];
1352         struct bio *bio;
1353         void *bi_private;
1354         /* since bio will be NULL, no need to initialize last_index */
1355         pgoff_t uninitialized_var(last_index);
1356         bool force_submit = false;
1357         unsigned int nr_bios;
1358
1359         if (unlikely(owned_head == Z_EROFS_VLE_WORKGRP_TAIL))
1360                 return false;
1361
1362         force_submit = false;
1363         bio = NULL;
1364         nr_bios = 0;
1365         bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1366
1367         /* by default, all need io submission */
1368         q[JQ_SUBMIT]->head = owned_head;
1369
1370         do {
1371                 struct z_erofs_vle_workgroup *grp;
1372                 pgoff_t first_index;
1373                 struct page *page;
1374                 unsigned int i = 0, bypass = 0;
1375                 int err;
1376
1377                 /* no possible 'owned_head' equals the following */
1378                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1379                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_NIL);
1380
1381                 grp = container_of(owned_head,
1382                                    struct z_erofs_vle_workgroup, next);
1383
1384                 /* close the main owned chain at first */
1385                 owned_head = cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
1386                                      Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1387
1388                 first_index = grp->obj.index;
1389                 force_submit |= (first_index != last_index + 1);
1390
1391 repeat:
1392                 page = pickup_page_for_submission(grp, i, pagepool,
1393                                                   MNGD_MAPPING(sbi), gfp);
1394                 if (!page) {
1395                         force_submit = true;
1396                         ++bypass;
1397                         goto skippage;
1398                 }
1399
1400                 if (bio && force_submit) {
1401 submit_bio_retry:
1402                         __submit_bio(bio, REQ_OP_READ, 0);
1403                         bio = NULL;
1404                 }
1405
1406                 if (!bio) {
1407                         bio = erofs_grab_bio(sb, first_index + i,
1408                                              BIO_MAX_PAGES,
1409                                              z_erofs_vle_read_endio, true);
1410                         bio->bi_private = bi_private;
1411
1412                         ++nr_bios;
1413                 }
1414
1415                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1416                 if (err < PAGE_SIZE)
1417                         goto submit_bio_retry;
1418
1419                 force_submit = false;
1420                 last_index = first_index + i;
1421 skippage:
1422                 if (++i < clusterpages)
1423                         goto repeat;
1424
1425                 if (bypass < clusterpages)
1426                         qtail[JQ_SUBMIT] = &grp->next;
1427                 else
1428                         move_to_bypass_jobqueue(grp, qtail, owned_head);
1429         } while (owned_head != Z_EROFS_VLE_WORKGRP_TAIL);
1430
1431         if (bio)
1432                 __submit_bio(bio, REQ_OP_READ, 0);
1433
1434         if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1435                 return true;
1436
1437         z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1438         return true;
1439 }
1440
1441 static void z_erofs_submit_and_unzip(struct z_erofs_vle_frontend *f,
1442                                      struct list_head *pagepool,
1443                                      bool force_fg)
1444 {
1445         struct super_block *sb = f->inode->i_sb;
1446         struct z_erofs_vle_unzip_io io[NR_JOBQUEUES];
1447
1448         if (!z_erofs_vle_submit_all(sb, f->owned_head, pagepool, io, force_fg))
1449                 return;
1450
1451 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1452         z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1453 #endif
1454         if (!force_fg)
1455                 return;
1456
1457         /* wait until all bios are completed */
1458         wait_event(io[JQ_SUBMIT].u.wait,
1459                    !atomic_read(&io[JQ_SUBMIT].pending_bios));
1460
1461         /* let's synchronous decompression */
1462         z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1463 }
1464
1465 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1466                                              struct page *page)
1467 {
1468         struct inode *const inode = page->mapping->host;
1469         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1470         int err;
1471         LIST_HEAD(pagepool);
1472
1473         trace_erofs_readpage(page, false);
1474
1475         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1476
1477         err = z_erofs_do_read_page(&f, page, &pagepool);
1478         (void)z_erofs_vle_work_iter_end(&f.builder);
1479
1480         if (err) {
1481                 errln("%s, failed to read, err [%d]", __func__, err);
1482                 goto out;
1483         }
1484
1485         z_erofs_submit_and_unzip(&f, &pagepool, true);
1486 out:
1487         if (f.m_iter.mpage)
1488                 put_page(f.m_iter.mpage);
1489
1490         /* clean up the remaining free pages */
1491         put_pages_list(&pagepool);
1492         return 0;
1493 }
1494
1495 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1496                                               struct address_space *mapping,
1497                                               struct list_head *pages,
1498                                               unsigned int nr_pages)
1499 {
1500         struct inode *const inode = mapping->host;
1501         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1502
1503         bool sync = __should_decompress_synchronously(sbi, nr_pages);
1504         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1505         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1506         struct page *head = NULL;
1507         LIST_HEAD(pagepool);
1508
1509         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1510                               nr_pages, false);
1511
1512         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1513
1514         for (; nr_pages; --nr_pages) {
1515                 struct page *page = lru_to_page(pages);
1516
1517                 prefetchw(&page->flags);
1518                 list_del(&page->lru);
1519
1520                 /*
1521                  * A pure asynchronous readahead is indicated if
1522                  * a PG_readahead marked page is hitted at first.
1523                  * Let's also do asynchronous decompression for this case.
1524                  */
1525                 sync &= !(PageReadahead(page) && !head);
1526
1527                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1528                         list_add(&page->lru, &pagepool);
1529                         continue;
1530                 }
1531
1532                 set_page_private(page, (unsigned long)head);
1533                 head = page;
1534         }
1535
1536         while (head) {
1537                 struct page *page = head;
1538                 int err;
1539
1540                 /* traversal in reverse order */
1541                 head = (void *)page_private(page);
1542
1543                 err = z_erofs_do_read_page(&f, page, &pagepool);
1544                 if (err) {
1545                         struct erofs_vnode *vi = EROFS_V(inode);
1546
1547                         errln("%s, readahead error at page %lu of nid %llu",
1548                                 __func__, page->index, vi->nid);
1549                 }
1550
1551                 put_page(page);
1552         }
1553
1554         (void)z_erofs_vle_work_iter_end(&f.builder);
1555
1556         z_erofs_submit_and_unzip(&f, &pagepool, sync);
1557
1558         if (f.m_iter.mpage)
1559                 put_page(f.m_iter.mpage);
1560
1561         /* clean up the remaining free pages */
1562         put_pages_list(&pagepool);
1563         return 0;
1564 }
1565
1566 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1567         .readpage = z_erofs_vle_normalaccess_readpage,
1568         .readpages = z_erofs_vle_normalaccess_readpages,
1569 };
1570
1571 /*
1572  * Variable-sized Logical Extent (Fixed Physical Cluster) Compression Mode
1573  * ---
1574  * VLE compression mode attempts to compress a number of logical data into
1575  * a physical cluster with a fixed size.
1576  * VLE compression mode uses "struct z_erofs_vle_decompressed_index".
1577  */
1578 #define __vle_cluster_advise(x, bit, bits) \
1579         ((le16_to_cpu(x) >> (bit)) & ((1 << (bits)) - 1))
1580
1581 #define __vle_cluster_type(advise) __vle_cluster_advise(advise, \
1582         Z_EROFS_VLE_DI_CLUSTER_TYPE_BIT, Z_EROFS_VLE_DI_CLUSTER_TYPE_BITS)
1583
1584 #define vle_cluster_type(di)    \
1585         __vle_cluster_type((di)->di_advise)
1586
1587 static int
1588 vle_decompressed_index_clusterofs(unsigned int *clusterofs,
1589                                   unsigned int clustersize,
1590                                   struct z_erofs_vle_decompressed_index *di)
1591 {
1592         switch (vle_cluster_type(di)) {
1593         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1594                 *clusterofs = clustersize;
1595                 break;
1596         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1597         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1598                 *clusterofs = le16_to_cpu(di->di_clusterofs);
1599                 break;
1600         default:
1601                 DBG_BUGON(1);
1602                 return -EIO;
1603         }
1604         return 0;
1605 }
1606
1607 static inline erofs_blk_t
1608 vle_extent_blkaddr(struct inode *inode, pgoff_t index)
1609 {
1610         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1611         struct erofs_vnode *vi = EROFS_V(inode);
1612
1613         unsigned int ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1614                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1615                 index * sizeof(struct z_erofs_vle_decompressed_index);
1616
1617         return erofs_blknr(iloc(sbi, vi->nid) + ofs);
1618 }
1619
1620 static inline unsigned int
1621 vle_extent_blkoff(struct inode *inode, pgoff_t index)
1622 {
1623         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1624         struct erofs_vnode *vi = EROFS_V(inode);
1625
1626         unsigned int ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1627                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1628                 index * sizeof(struct z_erofs_vle_decompressed_index);
1629
1630         return erofs_blkoff(iloc(sbi, vi->nid) + ofs);
1631 }
1632
1633 struct vle_map_blocks_iter_ctx {
1634         struct inode *inode;
1635         struct super_block *sb;
1636         unsigned int clusterbits;
1637
1638         struct page **mpage_ret;
1639         void **kaddr_ret;
1640 };
1641
1642 static int
1643 vle_get_logical_extent_head(const struct vle_map_blocks_iter_ctx *ctx,
1644                             unsigned int lcn,   /* logical cluster number */
1645                             unsigned long long *ofs,
1646                             erofs_blk_t *pblk,
1647                             unsigned int *flags)
1648 {
1649         const unsigned int clustersize = 1 << ctx->clusterbits;
1650         const erofs_blk_t mblk = vle_extent_blkaddr(ctx->inode, lcn);
1651         struct page *mpage = *ctx->mpage_ret;   /* extent metapage */
1652
1653         struct z_erofs_vle_decompressed_index *di;
1654         unsigned int cluster_type, delta0;
1655
1656         if (mpage->index != mblk) {
1657                 kunmap_atomic(*ctx->kaddr_ret);
1658                 unlock_page(mpage);
1659                 put_page(mpage);
1660
1661                 mpage = erofs_get_meta_page(ctx->sb, mblk, false);
1662                 if (IS_ERR(mpage)) {
1663                         *ctx->mpage_ret = NULL;
1664                         return PTR_ERR(mpage);
1665                 }
1666                 *ctx->mpage_ret = mpage;
1667                 *ctx->kaddr_ret = kmap_atomic(mpage);
1668         }
1669
1670         di = *ctx->kaddr_ret + vle_extent_blkoff(ctx->inode, lcn);
1671
1672         cluster_type = vle_cluster_type(di);
1673         switch (cluster_type) {
1674         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1675                 delta0 = le16_to_cpu(di->di_u.delta[0]);
1676                 if (unlikely(!delta0 || delta0 > lcn)) {
1677                         errln("invalid NONHEAD dl0 %u at lcn %u of nid %llu",
1678                               delta0, lcn, EROFS_V(ctx->inode)->nid);
1679                         DBG_BUGON(1);
1680                         return -EIO;
1681                 }
1682                 return vle_get_logical_extent_head(ctx,
1683                         lcn - delta0, ofs, pblk, flags);
1684         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1685                 *flags ^= EROFS_MAP_ZIPPED;
1686                 /* fallthrough */
1687         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1688                 /* clustersize should be a power of two */
1689                 *ofs = ((u64)lcn << ctx->clusterbits) +
1690                         (le16_to_cpu(di->di_clusterofs) & (clustersize - 1));
1691                 *pblk = le32_to_cpu(di->di_u.blkaddr);
1692                 break;
1693         default:
1694                 errln("unknown cluster type %u at lcn %u of nid %llu",
1695                       cluster_type, lcn, EROFS_V(ctx->inode)->nid);
1696                 DBG_BUGON(1);
1697                 return -EIO;
1698         }
1699         return 0;
1700 }
1701
1702 int z_erofs_map_blocks_iter(struct inode *inode,
1703         struct erofs_map_blocks *map,
1704         struct page **mpage_ret, int flags)
1705 {
1706         void *kaddr;
1707         const struct vle_map_blocks_iter_ctx ctx = {
1708                 .inode = inode,
1709                 .sb = inode->i_sb,
1710                 .clusterbits = EROFS_I_SB(inode)->clusterbits,
1711                 .mpage_ret = mpage_ret,
1712                 .kaddr_ret = &kaddr
1713         };
1714         const unsigned int clustersize = 1 << ctx.clusterbits;
1715         /* if both m_(l,p)len are 0, regularize l_lblk, l_lofs, etc... */
1716         const bool initial = !map->m_llen;
1717
1718         /* logicial extent (start, end) offset */
1719         unsigned long long ofs, end;
1720         unsigned int lcn;
1721         u32 ofs_rem;
1722
1723         /* initialize `pblk' to keep gcc from printing foolish warnings */
1724         erofs_blk_t mblk, pblk = 0;
1725         struct page *mpage = *mpage_ret;
1726         struct z_erofs_vle_decompressed_index *di;
1727         unsigned int cluster_type, logical_cluster_ofs;
1728         int err = 0;
1729
1730         trace_z_erofs_map_blocks_iter_enter(inode, map, flags);
1731
1732         /* when trying to read beyond EOF, leave it unmapped */
1733         if (unlikely(map->m_la >= inode->i_size)) {
1734                 DBG_BUGON(!initial);
1735                 map->m_llen = map->m_la + 1 - inode->i_size;
1736                 map->m_la = inode->i_size;
1737                 map->m_flags = 0;
1738                 goto out;
1739         }
1740
1741         debugln("%s, m_la %llu m_llen %llu --- start", __func__,
1742                 map->m_la, map->m_llen);
1743
1744         ofs = map->m_la + map->m_llen;
1745
1746         /* clustersize should be power of two */
1747         lcn = ofs >> ctx.clusterbits;
1748         ofs_rem = ofs & (clustersize - 1);
1749
1750         mblk = vle_extent_blkaddr(inode, lcn);
1751
1752         if (!mpage || mpage->index != mblk) {
1753                 if (mpage)
1754                         put_page(mpage);
1755
1756                 mpage = erofs_get_meta_page(ctx.sb, mblk, false);
1757                 if (IS_ERR(mpage)) {
1758                         err = PTR_ERR(mpage);
1759                         goto out;
1760                 }
1761                 *mpage_ret = mpage;
1762         } else {
1763                 lock_page(mpage);
1764                 DBG_BUGON(!PageUptodate(mpage));
1765         }
1766
1767         kaddr = kmap_atomic(mpage);
1768         di = kaddr + vle_extent_blkoff(inode, lcn);
1769
1770         debugln("%s, lcn %u mblk %u e_blkoff %u", __func__, lcn,
1771                 mblk, vle_extent_blkoff(inode, lcn));
1772
1773         err = vle_decompressed_index_clusterofs(&logical_cluster_ofs,
1774                                                 clustersize, di);
1775         if (unlikely(err))
1776                 goto unmap_out;
1777
1778         if (!initial) {
1779                 /* [walking mode] 'map' has been already initialized */
1780                 map->m_llen += logical_cluster_ofs;
1781                 goto unmap_out;
1782         }
1783
1784         /* by default, compressed */
1785         map->m_flags |= EROFS_MAP_ZIPPED;
1786
1787         end = ((u64)lcn + 1) * clustersize;
1788
1789         cluster_type = vle_cluster_type(di);
1790
1791         switch (cluster_type) {
1792         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1793                 if (ofs_rem >= logical_cluster_ofs)
1794                         map->m_flags ^= EROFS_MAP_ZIPPED;
1795                 /* fallthrough */
1796         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1797                 if (ofs_rem == logical_cluster_ofs) {
1798                         pblk = le32_to_cpu(di->di_u.blkaddr);
1799                         goto exact_hitted;
1800                 }
1801
1802                 if (ofs_rem > logical_cluster_ofs) {
1803                         ofs = (u64)lcn * clustersize | logical_cluster_ofs;
1804                         pblk = le32_to_cpu(di->di_u.blkaddr);
1805                         break;
1806                 }
1807
1808                 /* logical cluster number should be >= 1 */
1809                 if (unlikely(!lcn)) {
1810                         errln("invalid logical cluster 0 at nid %llu",
1811                                 EROFS_V(inode)->nid);
1812                         err = -EIO;
1813                         goto unmap_out;
1814                 }
1815                 end = ((u64)lcn-- * clustersize) | logical_cluster_ofs;
1816                 /* fallthrough */
1817         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1818                 /* get the correspoinding first chunk */
1819                 err = vle_get_logical_extent_head(&ctx, lcn, &ofs,
1820                                                   &pblk, &map->m_flags);
1821                 mpage = *mpage_ret;
1822
1823                 if (unlikely(err)) {
1824                         if (mpage)
1825                                 goto unmap_out;
1826                         goto out;
1827                 }
1828                 break;
1829         default:
1830                 errln("unknown cluster type %u at offset %llu of nid %llu",
1831                         cluster_type, ofs, EROFS_V(inode)->nid);
1832                 err = -EIO;
1833                 goto unmap_out;
1834         }
1835
1836         map->m_la = ofs;
1837 exact_hitted:
1838         map->m_llen = end - ofs;
1839         map->m_plen = clustersize;
1840         map->m_pa = blknr_to_addr(pblk);
1841         map->m_flags |= EROFS_MAP_MAPPED;
1842 unmap_out:
1843         kunmap_atomic(kaddr);
1844         unlock_page(mpage);
1845 out:
1846         debugln("%s, m_la %llu m_pa %llu m_llen %llu m_plen %llu m_flags 0%o",
1847                 __func__, map->m_la, map->m_pa,
1848                 map->m_llen, map->m_plen, map->m_flags);
1849
1850         trace_z_erofs_map_blocks_iter_exit(inode, map, flags, err);
1851
1852         /* aggressively BUG_ON iff CONFIG_EROFS_FS_DEBUG is on */
1853         DBG_BUGON(err < 0 && err != -ENOMEM);
1854         return err;
1855 }
1856