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 "compress.h"
  15 #include <linux/prefetch.h>
  16
  17 #include <trace/events/erofs.h>
  18
  19 /*
  20  * a compressed_pages[] placeholder in order to avoid
  21  * being filled with file pages for in-place decompression.
  22  */
  23 #define PAGE_UNALLOCATED     ((void *)0x5F0E4B1D)
  24
  25 /* how to allocate cached pages for a workgroup */
  26 enum z_erofs_cache_alloctype {
  27         DONTALLOC,      /* don't allocate any cached pages */
  28         DELAYEDALLOC,   /* delayed allocation (at the time of submitting io) */
  29 };
  30
  31 /*
  32  * tagged pointer with 1-bit tag for all compressed pages
  33  * tag 0 - the page is just found with an extra page reference
  34  */
  35 typedef tagptr1_t compressed_page_t;
  36
  37 #define tag_compressed_page_justfound(page) \
  38         tagptr_fold(compressed_page_t, page, 1)
  39
  40 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
  41 static struct kmem_cache *z_erofs_workgroup_cachep __read_mostly;
  42
  43 void z_erofs_exit_zip_subsystem(void)
  44 {
  45         destroy_workqueue(z_erofs_workqueue);
  46         kmem_cache_destroy(z_erofs_workgroup_cachep);
  47 }
  48
  49 static inline int init_unzip_workqueue(void)
  50 {
  51         const unsigned int onlinecpus = num_possible_cpus();
  52
  53         /*
  54          * we don't need too many threads, limiting threads
  55          * could improve scheduling performance.
  56          */
  57         z_erofs_workqueue =
  58                 alloc_workqueue("erofs_unzipd",
  59                                 WQ_UNBOUND | WQ_HIGHPRI | WQ_CPU_INTENSIVE,
  60                                 onlinecpus + onlinecpus / 4);
  61
  62         return z_erofs_workqueue ? 0 : -ENOMEM;
  63 }
  64
  65 static void init_once(void *ptr)
  66 {
  67         struct z_erofs_vle_workgroup *grp = ptr;
  68         struct z_erofs_vle_work *const work =
  69                 z_erofs_vle_grab_primary_work(grp);
  70         unsigned int i;
  71
  72         mutex_init(&work->lock);
  73         work->nr_pages = 0;
  74         work->vcnt = 0;
  75         for (i = 0; i < Z_EROFS_CLUSTER_MAX_PAGES; ++i)
  76                 grp->compressed_pages[i] = NULL;
  77 }
  78
  79 static void init_always(struct z_erofs_vle_workgroup *grp)
  80 {
  81         struct z_erofs_vle_work *const work =
  82                 z_erofs_vle_grab_primary_work(grp);
  83
  84         atomic_set(&grp->obj.refcount, 1);
  85         grp->flags = 0;
  86
  87         DBG_BUGON(work->nr_pages);
  88         DBG_BUGON(work->vcnt);
  89 }
  90
  91 int __init z_erofs_init_zip_subsystem(void)
  92 {
  93         z_erofs_workgroup_cachep =
  94                 kmem_cache_create("erofs_compress",
  95                                   Z_EROFS_WORKGROUP_SIZE, 0,
  96                                   SLAB_RECLAIM_ACCOUNT, init_once);
  97
  98         if (z_erofs_workgroup_cachep) {
  99                 if (!init_unzip_workqueue())
 100                         return 0;
 101
 102                 kmem_cache_destroy(z_erofs_workgroup_cachep);
 103         }
 104         return -ENOMEM;
 105 }
 106
 107 enum z_erofs_vle_work_role {
 108         Z_EROFS_VLE_WORK_SECONDARY,
 109         Z_EROFS_VLE_WORK_PRIMARY,
 110         /*
 111          * The current work was the tail of an exist chain, and the previous
 112          * processed chained works are all decided to be hooked up to it.
 113          * A new chain should be created for the remaining unprocessed works,
 114          * therefore different from Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
 115          * the next work cannot reuse the whole page in the following scenario:
 116          *  ________________________________________________________________
 117          * |      tail (partial) page     |       head (partial) page       |
 118          * |  (belongs to the next work)  |  (belongs to the current work)  |
 119          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 120          */
 121         Z_EROFS_VLE_WORK_PRIMARY_HOOKED,
 122         /*
 123          * The current work has been linked with the processed chained works,
 124          * and could be also linked with the potential remaining works, which
 125          * means if the processing page is the tail partial page of the work,
 126          * the current work can safely use the whole page (since the next work
 127          * is under control) for in-place decompression, as illustrated below:
 128          *  ________________________________________________________________
 129          * |  tail (partial) page  |          head (partial) page           |
 130          * | (of the current work) |         (of the previous work)         |
 131          * |  PRIMARY_FOLLOWED or  |                                        |
 132          * |_____PRIMARY_HOOKED____|____________PRIMARY_FOLLOWED____________|
 133          *
 134          * [  (*) the above page can be used for the current work itself.  ]
 135          */
 136         Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
 137         Z_EROFS_VLE_WORK_MAX
 138 };
 139
 140 struct z_erofs_vle_work_builder {
 141         enum z_erofs_vle_work_role role;
 142         /*
 143          * 'hosted = false' means that the current workgroup doesn't belong to
 144          * the owned chained workgroups. In the other words, it is none of our
 145          * business to submit this workgroup.
 146          */
 147         bool hosted;
 148
 149         struct z_erofs_vle_workgroup *grp;
 150         struct z_erofs_vle_work *work;
 151         struct z_erofs_pagevec_ctor vector;
 152
 153         /* pages used for reading the compressed data */
 154         struct page **compressed_pages;
 155         unsigned int compressed_deficit;
 156 };
 157
 158 #define VLE_WORK_BUILDER_INIT() \
 159         { .work = NULL, .role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED }
 160
 161 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 162 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 163                                      struct address_space *mc,
 164                                      pgoff_t index,
 165                                      unsigned int clusterpages,
 166                                      enum z_erofs_cache_alloctype type,
 167                                      struct list_head *pagepool,
 168                                      gfp_t gfp)
 169 {
 170         struct page **const pages = bl->compressed_pages;
 171         const unsigned int remaining = bl->compressed_deficit;
 172         bool standalone = true;
 173         unsigned int i, j = 0;
 174
 175         if (bl->role < Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 176                 return;
 177
 178         gfp = mapping_gfp_constraint(mc, gfp) & ~__GFP_RECLAIM;
 179
 180         index += clusterpages - remaining;
 181
 182         for (i = 0; i < remaining; ++i) {
 183                 struct page *page;
 184                 compressed_page_t t;
 185
 186                 /* the compressed page was loaded before */
 187                 if (READ_ONCE(pages[i]))
 188                         continue;
 189
 190                 page = find_get_page(mc, index + i);
 191
 192                 if (page) {
 193                         t = tag_compressed_page_justfound(page);
 194                 } else if (type == DELAYEDALLOC) {
 195                         t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
 196                 } else {        /* DONTALLOC */
 197                         if (standalone)
 198                                 j = i;
 199                         standalone = false;
 200                         continue;
 201                 }
 202
 203                 if (!cmpxchg_relaxed(&pages[i], NULL, tagptr_cast_ptr(t)))
 204                         continue;
 205
 206                 if (page)
 207                         put_page(page);
 208         }
 209         bl->compressed_pages += j;
 210         bl->compressed_deficit = remaining - j;
 211
 212         if (standalone)
 213                 bl->role = Z_EROFS_VLE_WORK_PRIMARY;
 214 }
 215
 216 /* called by erofs_shrinker to get rid of all compressed_pages */
 217 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 218                                        struct erofs_workgroup *egrp)
 219 {
 220         struct z_erofs_vle_workgroup *const grp =
 221                 container_of(egrp, struct z_erofs_vle_workgroup, obj);
 222         struct address_space *const mapping = MNGD_MAPPING(sbi);
 223         const int clusterpages = erofs_clusterpages(sbi);
 224         int i;
 225
 226         /*
 227          * refcount of workgroup is now freezed as 1,
 228          * therefore no need to worry about available decompression users.
 229          */
 230         for (i = 0; i < clusterpages; ++i) {
 231                 struct page *page = grp->compressed_pages[i];
 232
 233                 if (!page || page->mapping != mapping)
 234                         continue;
 235
 236                 /* block other users from reclaiming or migrating the page */
 237                 if (!trylock_page(page))
 238                         return -EBUSY;
 239
 240                 /* barrier is implied in the following 'unlock_page' */
 241                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 242
 243                 set_page_private(page, 0);
 244                 ClearPagePrivate(page);
 245
 246                 unlock_page(page);
 247                 put_page(page);
 248         }
 249         return 0;
 250 }
 251
 252 int erofs_try_to_free_cached_page(struct address_space *mapping,
 253                                   struct page *page)
 254 {
 255         struct erofs_sb_info *const sbi = EROFS_SB(mapping->host->i_sb);
 256         const unsigned int clusterpages = erofs_clusterpages(sbi);
 257         struct z_erofs_vle_workgroup *const grp = (void *)page_private(page);
 258         int ret = 0;    /* 0 - busy */
 259
 260         if (erofs_workgroup_try_to_freeze(&grp->obj, 1)) {
 261                 unsigned int i;
 262
 263                 for (i = 0; i < clusterpages; ++i) {
 264                         if (grp->compressed_pages[i] == page) {
 265                                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 266                                 ret = 1;
 267                                 break;
 268                         }
 269                 }
 270                 erofs_workgroup_unfreeze(&grp->obj, 1);
 271
 272                 if (ret) {
 273                         ClearPagePrivate(page);
 274                         put_page(page);
 275                 }
 276         }
 277         return ret;
 278 }
 279 #else
 280 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 281                                      struct address_space *mc,
 282                                      pgoff_t index,
 283                                      unsigned int clusterpages,
 284                                      enum z_erofs_cache_alloctype type,
 285                                      struct list_head *pagepool,
 286                                      gfp_t gfp)
 287 {
 288         /* nowhere to load compressed pages from */
 289 }
 290 #endif
 291
 292 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 293 static inline bool try_to_reuse_as_compressed_page(
 294         struct z_erofs_vle_work_builder *b,
 295         struct page *page)
 296 {
 297         while (b->compressed_deficit) {
 298                 --b->compressed_deficit;
 299                 if (!cmpxchg(b->compressed_pages++, NULL, page))
 300                         return true;
 301         }
 302
 303         return false;
 304 }
 305
 306 /* callers must be with work->lock held */
 307 static int z_erofs_vle_work_add_page(
 308         struct z_erofs_vle_work_builder *builder,
 309         struct page *page,
 310         enum z_erofs_page_type type)
 311 {
 312         int ret;
 313         bool occupied;
 314
 315         /* give priority for the compressed data storage */
 316         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY &&
 317             type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 318             try_to_reuse_as_compressed_page(builder, page))
 319                 return 0;
 320
 321         ret = z_erofs_pagevec_ctor_enqueue(&builder->vector,
 322                                            page, type, &occupied);
 323         builder->work->vcnt += (unsigned int)ret;
 324
 325         return ret ? 0 : -EAGAIN;
 326 }
 327
 328 static enum z_erofs_vle_work_role
 329 try_to_claim_workgroup(struct z_erofs_vle_workgroup *grp,
 330                        z_erofs_vle_owned_workgrp_t *owned_head,
 331                        bool *hosted)
 332 {
 333         DBG_BUGON(*hosted);
 334
 335         /* let's claim these following types of workgroup */
 336 retry:
 337         if (grp->next == Z_EROFS_VLE_WORKGRP_NIL) {
 338                 /* type 1, nil workgroup */
 339                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_NIL,
 340                             *owned_head) != Z_EROFS_VLE_WORKGRP_NIL)
 341                         goto retry;
 342
 343                 *owned_head = &grp->next;
 344                 *hosted = true;
 345                 /* lucky, I am the followee :) */
 346                 return Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 347
 348         } else if (grp->next == Z_EROFS_VLE_WORKGRP_TAIL) {
 349                 /*
 350                  * type 2, link to the end of a existing open chain,
 351                  * be careful that its submission itself is governed
 352                  * by the original owned chain.
 353                  */
 354                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
 355                             *owned_head) != Z_EROFS_VLE_WORKGRP_TAIL)
 356                         goto retry;
 357                 *owned_head = Z_EROFS_VLE_WORKGRP_TAIL;
 358                 return Z_EROFS_VLE_WORK_PRIMARY_HOOKED;
 359         }
 360
 361         return Z_EROFS_VLE_WORK_PRIMARY; /* :( better luck next time */
 362 }
 363
 364 struct z_erofs_vle_work_finder {
 365         struct super_block *sb;
 366         pgoff_t idx;
 367         unsigned int pageofs;
 368
 369         struct z_erofs_vle_workgroup **grp_ret;
 370         enum z_erofs_vle_work_role *role;
 371         z_erofs_vle_owned_workgrp_t *owned_head;
 372         bool *hosted;
 373 };
 374
 375 static struct z_erofs_vle_work *
 376 z_erofs_vle_work_lookup(const struct z_erofs_vle_work_finder *f)
 377 {
 378         bool tag, primary;
 379         struct erofs_workgroup *egrp;
 380         struct z_erofs_vle_workgroup *grp;
 381         struct z_erofs_vle_work *work;
 382
 383         egrp = erofs_find_workgroup(f->sb, f->idx, &tag);
 384         if (!egrp) {
 385                 *f->grp_ret = NULL;
 386                 return NULL;
 387         }
 388
 389         grp = container_of(egrp, struct z_erofs_vle_workgroup, obj);
 390         *f->grp_ret = grp;
 391
 392         work = z_erofs_vle_grab_work(grp, f->pageofs);
 393         /* if multiref is disabled, `primary' is always true */
 394         primary = true;
 395
 396         DBG_BUGON(work->pageofs != f->pageofs);
 397
 398         /*
 399          * lock must be taken first to avoid grp->next == NIL between
 400          * claiming workgroup and adding pages:
 401          *                        grp->next != NIL
 402          *   grp->next = NIL
 403          *   mutex_unlock_all
 404          *                        mutex_lock(&work->lock)
 405          *                        add all pages to pagevec
 406          *
 407          * [correct locking case 1]:
 408          *   mutex_lock(grp->work[a])
 409          *   ...
 410          *   mutex_lock(grp->work[b])     mutex_lock(grp->work[c])
 411          *   ...                          *role = SECONDARY
 412          *                                add all pages to pagevec
 413          *                                ...
 414          *                                mutex_unlock(grp->work[c])
 415          *   mutex_lock(grp->work[c])
 416          *   ...
 417          *   grp->next = NIL
 418          *   mutex_unlock_all
 419          *
 420          * [correct locking case 2]:
 421          *   mutex_lock(grp->work[b])
 422          *   ...
 423          *   mutex_lock(grp->work[a])
 424          *   ...
 425          *   mutex_lock(grp->work[c])
 426          *   ...
 427          *   grp->next = NIL
 428          *   mutex_unlock_all
 429          *                                mutex_lock(grp->work[a])
 430          *                                *role = PRIMARY_OWNER
 431          *                                add all pages to pagevec
 432          *                                ...
 433          */
 434         mutex_lock(&work->lock);
 435
 436         *f->hosted = false;
 437         if (!primary)
 438                 *f->role = Z_EROFS_VLE_WORK_SECONDARY;
 439         else    /* claim the workgroup if possible */
 440                 *f->role = try_to_claim_workgroup(grp, f->owned_head,
 441                                                   f->hosted);
 442         return work;
 443 }
 444
 445 static struct z_erofs_vle_work *
 446 z_erofs_vle_work_register(const struct z_erofs_vle_work_finder *f,
 447                           struct erofs_map_blocks *map)
 448 {
 449         bool gnew = false;
 450         struct z_erofs_vle_workgroup *grp = *f->grp_ret;
 451         struct z_erofs_vle_work *work;
 452
 453         /* if multiref is disabled, grp should never be nullptr */
 454         if (unlikely(grp)) {
 455                 DBG_BUGON(1);
 456                 return ERR_PTR(-EINVAL);
 457         }
 458
 459         /* no available workgroup, let's allocate one */
 460         grp = kmem_cache_alloc(z_erofs_workgroup_cachep, GFP_NOFS);
 461         if (unlikely(!grp))
 462                 return ERR_PTR(-ENOMEM);
 463
 464         init_always(grp);
 465         grp->obj.index = f->idx;
 466         grp->llen = map->m_llen;
 467
 468         z_erofs_vle_set_workgrp_fmt(grp, (map->m_flags & EROFS_MAP_ZIPPED) ?
 469                                     Z_EROFS_VLE_WORKGRP_FMT_LZ4 :
 470                                     Z_EROFS_VLE_WORKGRP_FMT_PLAIN);
 471
 472         if (map->m_flags & EROFS_MAP_FULL_MAPPED)
 473                 grp->flags |= Z_EROFS_VLE_WORKGRP_FULL_LENGTH;
 474
 475         /* new workgrps have been claimed as type 1 */
 476         WRITE_ONCE(grp->next, *f->owned_head);
 477         /* primary and followed work for all new workgrps */
 478         *f->role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 479         /* it should be submitted by ourselves */
 480         *f->hosted = true;
 481
 482         gnew = true;
 483         work = z_erofs_vle_grab_primary_work(grp);
 484         work->pageofs = f->pageofs;
 485
 486         /*
 487          * lock all primary followed works before visible to others
 488          * and mutex_trylock *never* fails for a new workgroup.
 489          */
 490         mutex_trylock(&work->lock);
 491
 492         if (gnew) {
 493                 int err = erofs_register_workgroup(f->sb, &grp->obj, 0);
 494
 495                 if (err) {
 496                         mutex_unlock(&work->lock);
 497                         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 498                         return ERR_PTR(-EAGAIN);
 499                 }
 500         }
 501
 502         *f->owned_head = &grp->next;
 503         *f->grp_ret = grp;
 504         return work;
 505 }
 506
 507 #define builder_is_hooked(builder) \
 508         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_HOOKED)
 509
 510 #define builder_is_followed(builder) \
 511         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 512
 513 static int z_erofs_vle_work_iter_begin(struct z_erofs_vle_work_builder *builder,
 514                                        struct super_block *sb,
 515                                        struct erofs_map_blocks *map,
 516                                        z_erofs_vle_owned_workgrp_t *owned_head)
 517 {
 518         const unsigned int clusterpages = erofs_clusterpages(EROFS_SB(sb));
 519         struct z_erofs_vle_workgroup *grp;
 520         const struct z_erofs_vle_work_finder finder = {
 521                 .sb = sb,
 522                 .idx = erofs_blknr(map->m_pa),
 523                 .pageofs = map->m_la & ~PAGE_MASK,
 524                 .grp_ret = &grp,
 525                 .role = &builder->role,
 526                 .owned_head = owned_head,
 527                 .hosted = &builder->hosted
 528         };
 529         struct z_erofs_vle_work *work;
 530
 531         DBG_BUGON(builder->work);
 532
 533         /* must be Z_EROFS_WORK_TAIL or the next chained work */
 534         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_NIL);
 535         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
 536
 537         DBG_BUGON(erofs_blkoff(map->m_pa));
 538
 539 repeat:
 540         work = z_erofs_vle_work_lookup(&finder);
 541         if (work) {
 542                 unsigned int orig_llen;
 543
 544                 /* increase workgroup `llen' if needed */
 545                 while ((orig_llen = READ_ONCE(grp->llen)) < map->m_llen &&
 546                        orig_llen != cmpxchg_relaxed(&grp->llen,
 547                                                     orig_llen, map->m_llen))
 548                         cpu_relax();
 549                 goto got_it;
 550         }
 551
 552         work = z_erofs_vle_work_register(&finder, map);
 553         if (unlikely(work == ERR_PTR(-EAGAIN)))
 554                 goto repeat;
 555
 556         if (IS_ERR(work))
 557                 return PTR_ERR(work);
 558 got_it:
 559         z_erofs_pagevec_ctor_init(&builder->vector, Z_EROFS_NR_INLINE_PAGEVECS,
 560                                   work->pagevec, work->vcnt);
 561
 562         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY) {
 563                 /* enable possibly in-place decompression */
 564                 builder->compressed_pages = grp->compressed_pages;
 565                 builder->compressed_deficit = clusterpages;
 566         } else {
 567                 builder->compressed_pages = NULL;
 568                 builder->compressed_deficit = 0;
 569         }
 570
 571         builder->grp = grp;
 572         builder->work = work;
 573         return 0;
 574 }
 575
 576 /*
 577  * keep in mind that no referenced workgroups will be freed
 578  * only after a RCU grace period, so rcu_read_lock() could
 579  * prevent a workgroup from being freed.
 580  */
 581 static void z_erofs_rcu_callback(struct rcu_head *head)
 582 {
 583         struct z_erofs_vle_work *work = container_of(head,
 584                 struct z_erofs_vle_work, rcu);
 585         struct z_erofs_vle_workgroup *grp =
 586                 z_erofs_vle_work_workgroup(work, true);
 587
 588         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 589 }
 590
 591 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 592 {
 593         struct z_erofs_vle_workgroup *const vgrp = container_of(grp,
 594                 struct z_erofs_vle_workgroup, obj);
 595         struct z_erofs_vle_work *const work = &vgrp->work;
 596
 597         call_rcu(&work->rcu, z_erofs_rcu_callback);
 598 }
 599
 600 static void
 601 __z_erofs_vle_work_release(struct z_erofs_vle_workgroup *grp,
 602                            struct z_erofs_vle_work *work __maybe_unused)
 603 {
 604         erofs_workgroup_put(&grp->obj);
 605 }
 606
 607 static void z_erofs_vle_work_release(struct z_erofs_vle_work *work)
 608 {
 609         struct z_erofs_vle_workgroup *grp =
 610                 z_erofs_vle_work_workgroup(work, true);
 611
 612         __z_erofs_vle_work_release(grp, work);
 613 }
 614
 615 static inline bool
 616 z_erofs_vle_work_iter_end(struct z_erofs_vle_work_builder *builder)
 617 {
 618         struct z_erofs_vle_work *work = builder->work;
 619
 620         if (!work)
 621                 return false;
 622
 623         z_erofs_pagevec_ctor_exit(&builder->vector, false);
 624         mutex_unlock(&work->lock);
 625
 626         /*
 627          * if all pending pages are added, don't hold work reference
 628          * any longer if the current work isn't hosted by ourselves.
 629          */
 630         if (!builder->hosted)
 631                 __z_erofs_vle_work_release(builder->grp, work);
 632
 633         builder->work = NULL;
 634         builder->grp = NULL;
 635         return true;
 636 }
 637
 638 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 639                                                gfp_t gfp)
 640 {
 641         struct page *page = erofs_allocpage(pagepool, gfp);
 642
 643         if (unlikely(!page))
 644                 return NULL;
 645
 646         page->mapping = Z_EROFS_MAPPING_STAGING;
 647         return page;
 648 }
 649
 650 struct z_erofs_vle_frontend {
 651         struct inode *const inode;
 652
 653         struct z_erofs_vle_work_builder builder;
 654         struct erofs_map_blocks map;
 655
 656         z_erofs_vle_owned_workgrp_t owned_head;
 657
 658         /* used for applying cache strategy on the fly */
 659         bool backmost;
 660         erofs_off_t headoffset;
 661 };
 662
 663 #define VLE_FRONTEND_INIT(__i) { \
 664         .inode = __i, \
 665         .map = { \
 666                 .m_llen = 0, \
 667                 .m_plen = 0, \
 668                 .mpage = NULL \
 669         }, \
 670         .builder = VLE_WORK_BUILDER_INIT(), \
 671         .owned_head = Z_EROFS_VLE_WORKGRP_TAIL, \
 672         .backmost = true, }
 673
 674 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 675 static inline bool
 676 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 677 {
 678         if (fe->backmost)
 679                 return true;
 680
 681         if (EROFS_FS_ZIP_CACHE_LVL >= 2)
 682                 return la < fe->headoffset;
 683
 684         return false;
 685 }
 686 #else
 687 static inline bool
 688 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 689 {
 690         return false;
 691 }
 692 #endif
 693
 694 static int z_erofs_do_read_page(struct z_erofs_vle_frontend *fe,
 695                                 struct page *page,
 696                                 struct list_head *page_pool)
 697 {
 698         struct super_block *const sb = fe->inode->i_sb;
 699         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
 700         struct erofs_map_blocks *const map = &fe->map;
 701         struct z_erofs_vle_work_builder *const builder = &fe->builder;
 702         const loff_t offset = page_offset(page);
 703
 704         bool tight = builder_is_hooked(builder);
 705         struct z_erofs_vle_work *work = builder->work;
 706
 707         enum z_erofs_cache_alloctype cache_strategy;
 708         enum z_erofs_page_type page_type;
 709         unsigned int cur, end, spiltted, index;
 710         int err = 0;
 711
 712         /* register locked file pages as online pages in pack */
 713         z_erofs_onlinepage_init(page);
 714
 715         spiltted = 0;
 716         end = PAGE_SIZE;
 717 repeat:
 718         cur = end - 1;
 719
 720         /* lucky, within the range of the current map_blocks */
 721         if (offset + cur >= map->m_la &&
 722             offset + cur < map->m_la + map->m_llen) {
 723                 /* didn't get a valid unzip work previously (very rare) */
 724                 if (!builder->work)
 725                         goto restart_now;
 726                 goto hitted;
 727         }
 728
 729         /* go ahead the next map_blocks */
 730         debugln("%s: [out-of-range] pos %llu", __func__, offset + cur);
 731
 732         if (z_erofs_vle_work_iter_end(builder))
 733                 fe->backmost = false;
 734
 735         map->m_la = offset + cur;
 736         map->m_llen = 0;
 737         err = z_erofs_map_blocks_iter(fe->inode, map, 0);
 738         if (unlikely(err))
 739                 goto err_out;
 740
 741 restart_now:
 742         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED)))
 743                 goto hitted;
 744
 745         DBG_BUGON(map->m_plen != 1 << sbi->clusterbits);
 746         DBG_BUGON(erofs_blkoff(map->m_pa));
 747
 748         err = z_erofs_vle_work_iter_begin(builder, sb, map, &fe->owned_head);
 749         if (unlikely(err))
 750                 goto err_out;
 751
 752         /* preload all compressed pages (maybe downgrade role if necessary) */
 753         if (should_alloc_managed_pages(fe, map->m_la))
 754                 cache_strategy = DELAYEDALLOC;
 755         else
 756                 cache_strategy = DONTALLOC;
 757
 758         preload_compressed_pages(builder, MNGD_MAPPING(sbi),
 759                                  map->m_pa / PAGE_SIZE,
 760                                  map->m_plen / PAGE_SIZE,
 761                                  cache_strategy, page_pool, GFP_KERNEL);
 762
 763         tight &= builder_is_hooked(builder);
 764         work = builder->work;
 765 hitted:
 766         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 767         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED))) {
 768                 zero_user_segment(page, cur, end);
 769                 goto next_part;
 770         }
 771
 772         /* let's derive page type */
 773         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 774                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 775                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 776                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 777
 778         if (cur)
 779                 tight &= builder_is_followed(builder);
 780
 781 retry:
 782         err = z_erofs_vle_work_add_page(builder, page, page_type);
 783         /* should allocate an additional staging page for pagevec */
 784         if (err == -EAGAIN) {
 785                 struct page *const newpage =
 786                         __stagingpage_alloc(page_pool, GFP_NOFS);
 787
 788                 err = z_erofs_vle_work_add_page(builder, newpage,
 789                                                 Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 790                 if (likely(!err))
 791                         goto retry;
 792         }
 793
 794         if (unlikely(err))
 795                 goto err_out;
 796
 797         index = page->index - map->m_la / PAGE_SIZE;
 798
 799         /* FIXME! avoid the last relundant fixup & endio */
 800         z_erofs_onlinepage_fixup(page, index, true);
 801
 802         /* bump up the number of spiltted parts of a page */
 803         ++spiltted;
 804         /* also update nr_pages */
 805         work->nr_pages = max_t(pgoff_t, work->nr_pages, index + 1);
 806 next_part:
 807         /* can be used for verification */
 808         map->m_llen = offset + cur - map->m_la;
 809
 810         end = cur;
 811         if (end > 0)
 812                 goto repeat;
 813
 814 out:
 815         /* FIXME! avoid the last relundant fixup & endio */
 816         z_erofs_onlinepage_endio(page);
 817
 818         debugln("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 819                 __func__, page, spiltted, map->m_llen);
 820         return err;
 821
 822         /* if some error occurred while processing this page */
 823 err_out:
 824         SetPageError(page);
 825         goto out;
 826 }
 827
 828 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 829 {
 830         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 831         struct z_erofs_vle_unzip_io *io = tagptr_unfold_ptr(t);
 832         bool background = tagptr_unfold_tags(t);
 833
 834         if (!background) {
 835                 unsigned long flags;
 836
 837                 spin_lock_irqsave(&io->u.wait.lock, flags);
 838                 if (!atomic_add_return(bios, &io->pending_bios))
 839                         wake_up_locked(&io->u.wait);
 840                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 841                 return;
 842         }
 843
 844         if (!atomic_add_return(bios, &io->pending_bios))
 845                 queue_work(z_erofs_workqueue, &io->u.work);
 846 }
 847
 848 static inline void z_erofs_vle_read_endio(struct bio *bio)
 849 {
 850         struct erofs_sb_info *sbi = NULL;
 851         blk_status_t err = bio->bi_status;
 852         struct bio_vec *bvec;
 853         struct bvec_iter_all iter_all;
 854
 855         bio_for_each_segment_all(bvec, bio, iter_all) {
 856                 struct page *page = bvec->bv_page;
 857                 bool cachemngd = false;
 858
 859                 DBG_BUGON(PageUptodate(page));
 860                 DBG_BUGON(!page->mapping);
 861
 862                 if (unlikely(!sbi && !z_erofs_page_is_staging(page))) {
 863                         sbi = EROFS_SB(page->mapping->host->i_sb);
 864
 865                         if (time_to_inject(sbi, FAULT_READ_IO)) {
 866                                 erofs_show_injection_info(FAULT_READ_IO);
 867                                 err = BLK_STS_IOERR;
 868                         }
 869                 }
 870
 871                 /* sbi should already be gotten if the page is managed */
 872                 if (sbi)
 873                         cachemngd = erofs_page_is_managed(sbi, page);
 874
 875                 if (unlikely(err))
 876                         SetPageError(page);
 877                 else if (cachemngd)
 878                         SetPageUptodate(page);
 879
 880                 if (cachemngd)
 881                         unlock_page(page);
 882         }
 883
 884         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 885         bio_put(bio);
 886 }
 887
 888 static struct page *z_pagemap_global[Z_EROFS_VLE_VMAP_GLOBAL_PAGES];
 889 static DEFINE_MUTEX(z_pagemap_global_lock);
 890
 891 static int z_erofs_vle_unzip(struct super_block *sb,
 892                              struct z_erofs_vle_workgroup *grp,
 893                              struct list_head *page_pool)
 894 {
 895         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 896         const unsigned int clusterpages = erofs_clusterpages(sbi);
 897
 898         struct z_erofs_pagevec_ctor ctor;
 899         unsigned int nr_pages;
 900         unsigned int sparsemem_pages = 0;
 901         struct page *pages_onstack[Z_EROFS_VLE_VMAP_ONSTACK_PAGES];
 902         struct page **pages, **compressed_pages, *page;
 903         unsigned int algorithm;
 904         unsigned int i, outputsize;
 905
 906         enum z_erofs_page_type page_type;
 907         bool overlapped, partial;
 908         struct z_erofs_vle_work *work;
 909         int err;
 910
 911         might_sleep();
 912         work = z_erofs_vle_grab_primary_work(grp);
 913         DBG_BUGON(!READ_ONCE(work->nr_pages));
 914
 915         mutex_lock(&work->lock);
 916         nr_pages = work->nr_pages;
 917
 918         if (likely(nr_pages <= Z_EROFS_VLE_VMAP_ONSTACK_PAGES))
 919                 pages = pages_onstack;
 920         else if (nr_pages <= Z_EROFS_VLE_VMAP_GLOBAL_PAGES &&
 921                  mutex_trylock(&z_pagemap_global_lock))
 922                 pages = z_pagemap_global;
 923         else {
 924 repeat:
 925                 pages = kvmalloc_array(nr_pages, sizeof(struct page *),
 926                                        GFP_KERNEL);
 927
 928                 /* fallback to global pagemap for the lowmem scenario */
 929                 if (unlikely(!pages)) {
 930                         if (nr_pages > Z_EROFS_VLE_VMAP_GLOBAL_PAGES)
 931                                 goto repeat;
 932                         else {
 933                                 mutex_lock(&z_pagemap_global_lock);
 934                                 pages = z_pagemap_global;
 935                         }
 936                 }
 937         }
 938
 939         for (i = 0; i < nr_pages; ++i)
 940                 pages[i] = NULL;
 941
 942         z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS,
 943                                   work->pagevec, 0);
 944
 945         for (i = 0; i < work->vcnt; ++i) {
 946                 unsigned int pagenr;
 947
 948                 page = z_erofs_pagevec_ctor_dequeue(&ctor, &page_type);
 949
 950                 /* all pages in pagevec ought to be valid */
 951                 DBG_BUGON(!page);
 952                 DBG_BUGON(!page->mapping);
 953
 954                 if (z_erofs_put_stagingpage(page_pool, page))
 955                         continue;
 956
 957                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 958                         pagenr = 0;
 959                 else
 960                         pagenr = z_erofs_onlinepage_index(page);
 961
 962                 DBG_BUGON(pagenr >= nr_pages);
 963                 DBG_BUGON(pages[pagenr]);
 964
 965                 pages[pagenr] = page;
 966         }
 967         sparsemem_pages = i;
 968
 969         z_erofs_pagevec_ctor_exit(&ctor, true);
 970
 971         overlapped = false;
 972         compressed_pages = grp->compressed_pages;
 973
 974         err = 0;
 975         for (i = 0; i < clusterpages; ++i) {
 976                 unsigned int pagenr;
 977
 978                 page = compressed_pages[i];
 979
 980                 /* all compressed pages ought to be valid */
 981                 DBG_BUGON(!page);
 982                 DBG_BUGON(!page->mapping);
 983
 984                 if (!z_erofs_page_is_staging(page)) {
 985                         if (erofs_page_is_managed(sbi, page)) {
 986                                 if (unlikely(!PageUptodate(page)))
 987                                         err = -EIO;
 988                                 continue;
 989                         }
 990
 991                         /*
 992                          * only if non-head page can be selected
 993                          * for inplace decompression
 994                          */
 995                         pagenr = z_erofs_onlinepage_index(page);
 996
 997                         DBG_BUGON(pagenr >= nr_pages);
 998                         DBG_BUGON(pages[pagenr]);
 999                         ++sparsemem_pages;
1000                         pages[pagenr] = page;
1001
1002                         overlapped = true;
1003                 }
1004
1005                 /* PG_error needs checking for inplaced and staging pages */
1006                 if (unlikely(PageError(page))) {
1007                         DBG_BUGON(PageUptodate(page));
1008                         err = -EIO;
1009                 }
1010         }
1011
1012         if (unlikely(err))
1013                 goto out;
1014
1015         if (nr_pages << PAGE_SHIFT >= work->pageofs + grp->llen) {
1016                 outputsize = grp->llen;
1017                 partial = !(grp->flags & Z_EROFS_VLE_WORKGRP_FULL_LENGTH);
1018         } else {
1019                 outputsize = (nr_pages << PAGE_SHIFT) - work->pageofs;
1020                 partial = true;
1021         }
1022
1023         if (z_erofs_vle_workgrp_fmt(grp) == Z_EROFS_VLE_WORKGRP_FMT_PLAIN)
1024                 algorithm = Z_EROFS_COMPRESSION_SHIFTED;
1025         else
1026                 algorithm = Z_EROFS_COMPRESSION_LZ4;
1027
1028         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1029                                         .sb = sb,
1030                                         .in = compressed_pages,
1031                                         .out = pages,
1032                                         .pageofs_out = work->pageofs,
1033                                         .inputsize = PAGE_SIZE,
1034                                         .outputsize = outputsize,
1035                                         .alg = algorithm,
1036                                         .inplace_io = overlapped,
1037                                         .partial_decoding = partial
1038                                  }, page_pool);
1039
1040 out:
1041         /* must handle all compressed pages before endding pages */
1042         for (i = 0; i < clusterpages; ++i) {
1043                 page = compressed_pages[i];
1044
1045                 if (erofs_page_is_managed(sbi, page))
1046                         continue;
1047
1048                 /* recycle all individual staging pages */
1049                 (void)z_erofs_put_stagingpage(page_pool, page);
1050
1051                 WRITE_ONCE(compressed_pages[i], NULL);
1052         }
1053
1054         for (i = 0; i < nr_pages; ++i) {
1055                 page = pages[i];
1056                 if (!page)
1057                         continue;
1058
1059                 DBG_BUGON(!page->mapping);
1060
1061                 /* recycle all individual staging pages */
1062                 if (z_erofs_put_stagingpage(page_pool, page))
1063                         continue;
1064
1065                 if (unlikely(err < 0))
1066                         SetPageError(page);
1067
1068                 z_erofs_onlinepage_endio(page);
1069         }
1070
1071         if (pages == z_pagemap_global)
1072                 mutex_unlock(&z_pagemap_global_lock);
1073         else if (unlikely(pages != pages_onstack))
1074                 kvfree(pages);
1075
1076         work->nr_pages = 0;
1077         work->vcnt = 0;
1078
1079         /* all work locks MUST be taken before the following line */
1080
1081         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_NIL);
1082
1083         /* all work locks SHOULD be released right now */
1084         mutex_unlock(&work->lock);
1085
1086         z_erofs_vle_work_release(work);
1087         return err;
1088 }
1089
1090 static void z_erofs_vle_unzip_all(struct super_block *sb,
1091                                   struct z_erofs_vle_unzip_io *io,
1092                                   struct list_head *page_pool)
1093 {
1094         z_erofs_vle_owned_workgrp_t owned = io->head;
1095
1096         while (owned != Z_EROFS_VLE_WORKGRP_TAIL_CLOSED) {
1097                 struct z_erofs_vle_workgroup *grp;
1098
1099                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1100                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_TAIL);
1101
1102                 /* no possible that 'owned' equals NULL */
1103                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_NIL);
1104
1105                 grp = container_of(owned, struct z_erofs_vle_workgroup, next);
1106                 owned = READ_ONCE(grp->next);
1107
1108                 z_erofs_vle_unzip(sb, grp, page_pool);
1109         }
1110 }
1111
1112 static void z_erofs_vle_unzip_wq(struct work_struct *work)
1113 {
1114         struct z_erofs_vle_unzip_io_sb *iosb = container_of(work,
1115                 struct z_erofs_vle_unzip_io_sb, io.u.work);
1116         LIST_HEAD(page_pool);
1117
1118         DBG_BUGON(iosb->io.head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1119         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &page_pool);
1120
1121         put_pages_list(&page_pool);
1122         kvfree(iosb);
1123 }
1124
1125 static struct page *
1126 pickup_page_for_submission(struct z_erofs_vle_workgroup *grp,
1127                            unsigned int nr,
1128                            struct list_head *pagepool,
1129                            struct address_space *mc,
1130                            gfp_t gfp)
1131 {
1132         /* determined at compile time to avoid too many #ifdefs */
1133         const bool nocache = __builtin_constant_p(mc) ? !mc : false;
1134         const pgoff_t index = grp->obj.index;
1135         bool tocache = false;
1136
1137         struct address_space *mapping;
1138         struct page *oldpage, *page;
1139
1140         compressed_page_t t;
1141         int justfound;
1142
1143 repeat:
1144         page = READ_ONCE(grp->compressed_pages[nr]);
1145         oldpage = page;
1146
1147         if (!page)
1148                 goto out_allocpage;
1149
1150         /*
1151          * the cached page has not been allocated and
1152          * an placeholder is out there, prepare it now.
1153          */
1154         if (!nocache && page == PAGE_UNALLOCATED) {
1155                 tocache = true;
1156                 goto out_allocpage;
1157         }
1158
1159         /* process the target tagged pointer */
1160         t = tagptr_init(compressed_page_t, page);
1161         justfound = tagptr_unfold_tags(t);
1162         page = tagptr_unfold_ptr(t);
1163
1164         mapping = READ_ONCE(page->mapping);
1165
1166         /*
1167          * if managed cache is disabled, it's no way to
1168          * get such a cached-like page.
1169          */
1170         if (nocache) {
1171                 /* if managed cache is disabled, it is impossible `justfound' */
1172                 DBG_BUGON(justfound);
1173
1174                 /* and it should be locked, not uptodate, and not truncated */
1175                 DBG_BUGON(!PageLocked(page));
1176                 DBG_BUGON(PageUptodate(page));
1177                 DBG_BUGON(!mapping);
1178                 goto out;
1179         }
1180
1181         /*
1182          * unmanaged (file) pages are all locked solidly,
1183          * therefore it is impossible for `mapping' to be NULL.
1184          */
1185         if (mapping && mapping != mc)
1186                 /* ought to be unmanaged pages */
1187                 goto out;
1188
1189         lock_page(page);
1190
1191         /* only true if page reclaim goes wrong, should never happen */
1192         DBG_BUGON(justfound && PagePrivate(page));
1193
1194         /* the page is still in manage cache */
1195         if (page->mapping == mc) {
1196                 WRITE_ONCE(grp->compressed_pages[nr], page);
1197
1198                 ClearPageError(page);
1199                 if (!PagePrivate(page)) {
1200                         /*
1201                          * impossible to be !PagePrivate(page) for
1202                          * the current restriction as well if
1203                          * the page is already in compressed_pages[].
1204                          */
1205                         DBG_BUGON(!justfound);
1206
1207                         justfound = 0;
1208                         set_page_private(page, (unsigned long)grp);
1209                         SetPagePrivate(page);
1210                 }
1211
1212                 /* no need to submit io if it is already up-to-date */
1213                 if (PageUptodate(page)) {
1214                         unlock_page(page);
1215                         page = NULL;
1216                 }
1217                 goto out;
1218         }
1219
1220         /*
1221          * the managed page has been truncated, it's unsafe to
1222          * reuse this one, let's allocate a new cache-managed page.
1223          */
1224         DBG_BUGON(page->mapping);
1225         DBG_BUGON(!justfound);
1226
1227         tocache = true;
1228         unlock_page(page);
1229         put_page(page);
1230 out_allocpage:
1231         page = __stagingpage_alloc(pagepool, gfp);
1232         if (oldpage != cmpxchg(&grp->compressed_pages[nr], oldpage, page)) {
1233                 list_add(&page->lru, pagepool);
1234                 cpu_relax();
1235                 goto repeat;
1236         }
1237         if (nocache || !tocache)
1238                 goto out;
1239         if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1240                 page->mapping = Z_EROFS_MAPPING_STAGING;
1241                 goto out;
1242         }
1243
1244         set_page_private(page, (unsigned long)grp);
1245         SetPagePrivate(page);
1246 out:    /* the only exit (for tracing and debugging) */
1247         return page;
1248 }
1249
1250 static struct z_erofs_vle_unzip_io *
1251 jobqueue_init(struct super_block *sb,
1252               struct z_erofs_vle_unzip_io *io,
1253               bool foreground)
1254 {
1255         struct z_erofs_vle_unzip_io_sb *iosb;
1256
1257         if (foreground) {
1258                 /* waitqueue available for foreground io */
1259                 DBG_BUGON(!io);
1260
1261                 init_waitqueue_head(&io->u.wait);
1262                 atomic_set(&io->pending_bios, 0);
1263                 goto out;
1264         }
1265
1266         iosb = kvzalloc(sizeof(*iosb), GFP_KERNEL | __GFP_NOFAIL);
1267         DBG_BUGON(!iosb);
1268
1269         /* initialize fields in the allocated descriptor */
1270         io = &iosb->io;
1271         iosb->sb = sb;
1272         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1273 out:
1274         io->head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1275         return io;
1276 }
1277
1278 /* define workgroup jobqueue types */
1279 enum {
1280 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1281         JQ_BYPASS,
1282 #endif
1283         JQ_SUBMIT,
1284         NR_JOBQUEUES,
1285 };
1286
1287 static void *jobqueueset_init(struct super_block *sb,
1288                               z_erofs_vle_owned_workgrp_t qtail[],
1289                               struct z_erofs_vle_unzip_io *q[],
1290                               struct z_erofs_vle_unzip_io *fgq,
1291                               bool forcefg)
1292 {
1293 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1294         /*
1295          * if managed cache is enabled, bypass jobqueue is needed,
1296          * no need to read from device for all workgroups in this queue.
1297          */
1298         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1299         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1300 #endif
1301
1302         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1303         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1304
1305         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1306 }
1307
1308 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1309 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1310                                     z_erofs_vle_owned_workgrp_t qtail[],
1311                                     z_erofs_vle_owned_workgrp_t owned_head)
1312 {
1313         z_erofs_vle_owned_workgrp_t *const submit_qtail = qtail[JQ_SUBMIT];
1314         z_erofs_vle_owned_workgrp_t *const bypass_qtail = qtail[JQ_BYPASS];
1315
1316         DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1317         if (owned_head == Z_EROFS_VLE_WORKGRP_TAIL)
1318                 owned_head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1319
1320         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1321
1322         WRITE_ONCE(*submit_qtail, owned_head);
1323         WRITE_ONCE(*bypass_qtail, &grp->next);
1324
1325         qtail[JQ_BYPASS] = &grp->next;
1326 }
1327
1328 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1329                                        unsigned int nr_bios,
1330                                        bool force_fg)
1331 {
1332         /*
1333          * although background is preferred, no one is pending for submission.
1334          * don't issue workqueue for decompression but drop it directly instead.
1335          */
1336         if (force_fg || nr_bios)
1337                 return false;
1338
1339         kvfree(container_of(q[JQ_SUBMIT],
1340                             struct z_erofs_vle_unzip_io_sb,
1341                             io));
1342         return true;
1343 }
1344 #else
1345 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1346                                     z_erofs_vle_owned_workgrp_t qtail[],
1347                                     z_erofs_vle_owned_workgrp_t owned_head)
1348 {
1349         /* impossible to bypass submission for managed cache disabled */
1350         DBG_BUGON(1);
1351 }
1352
1353 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1354                                        unsigned int nr_bios,
1355                                        bool force_fg)
1356 {
1357         /* bios should be >0 if managed cache is disabled */
1358         DBG_BUGON(!nr_bios);
1359         return false;
1360 }
1361 #endif
1362
1363 static bool z_erofs_vle_submit_all(struct super_block *sb,
1364                                    z_erofs_vle_owned_workgrp_t owned_head,
1365                                    struct list_head *pagepool,
1366                                    struct z_erofs_vle_unzip_io *fgq,
1367                                    bool force_fg)
1368 {
1369         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1370         const unsigned int clusterpages = erofs_clusterpages(sbi);
1371         const gfp_t gfp = GFP_NOFS;
1372
1373         z_erofs_vle_owned_workgrp_t qtail[NR_JOBQUEUES];
1374         struct z_erofs_vle_unzip_io *q[NR_JOBQUEUES];
1375         struct bio *bio;
1376         void *bi_private;
1377         /* since bio will be NULL, no need to initialize last_index */
1378         pgoff_t uninitialized_var(last_index);
1379         bool force_submit = false;
1380         unsigned int nr_bios;
1381
1382         if (unlikely(owned_head == Z_EROFS_VLE_WORKGRP_TAIL))
1383                 return false;
1384
1385         force_submit = false;
1386         bio = NULL;
1387         nr_bios = 0;
1388         bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1389
1390         /* by default, all need io submission */
1391         q[JQ_SUBMIT]->head = owned_head;
1392
1393         do {
1394                 struct z_erofs_vle_workgroup *grp;
1395                 pgoff_t first_index;
1396                 struct page *page;
1397                 unsigned int i = 0, bypass = 0;
1398                 int err;
1399
1400                 /* no possible 'owned_head' equals the following */
1401                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1402                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_NIL);
1403
1404                 grp = container_of(owned_head,
1405                                    struct z_erofs_vle_workgroup, next);
1406
1407                 /* close the main owned chain at first */
1408                 owned_head = cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
1409                                      Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1410
1411                 first_index = grp->obj.index;
1412                 force_submit |= (first_index != last_index + 1);
1413
1414 repeat:
1415                 page = pickup_page_for_submission(grp, i, pagepool,
1416                                                   MNGD_MAPPING(sbi), gfp);
1417                 if (!page) {
1418                         force_submit = true;
1419                         ++bypass;
1420                         goto skippage;
1421                 }
1422
1423                 if (bio && force_submit) {
1424 submit_bio_retry:
1425                         __submit_bio(bio, REQ_OP_READ, 0);
1426                         bio = NULL;
1427                 }
1428
1429                 if (!bio) {
1430                         bio = erofs_grab_bio(sb, first_index + i,
1431                                              BIO_MAX_PAGES, bi_private,
1432                                              z_erofs_vle_read_endio, true);
1433                         ++nr_bios;
1434                 }
1435
1436                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1437                 if (err < PAGE_SIZE)
1438                         goto submit_bio_retry;
1439
1440                 force_submit = false;
1441                 last_index = first_index + i;
1442 skippage:
1443                 if (++i < clusterpages)
1444                         goto repeat;
1445
1446                 if (bypass < clusterpages)
1447                         qtail[JQ_SUBMIT] = &grp->next;
1448                 else
1449                         move_to_bypass_jobqueue(grp, qtail, owned_head);
1450         } while (owned_head != Z_EROFS_VLE_WORKGRP_TAIL);
1451
1452         if (bio)
1453                 __submit_bio(bio, REQ_OP_READ, 0);
1454
1455         if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1456                 return true;
1457
1458         z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1459         return true;
1460 }
1461
1462 static void z_erofs_submit_and_unzip(struct z_erofs_vle_frontend *f,
1463                                      struct list_head *pagepool,
1464                                      bool force_fg)
1465 {
1466         struct super_block *sb = f->inode->i_sb;
1467         struct z_erofs_vle_unzip_io io[NR_JOBQUEUES];
1468
1469         if (!z_erofs_vle_submit_all(sb, f->owned_head, pagepool, io, force_fg))
1470                 return;
1471
1472 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1473         z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1474 #endif
1475         if (!force_fg)
1476                 return;
1477
1478         /* wait until all bios are completed */
1479         wait_event(io[JQ_SUBMIT].u.wait,
1480                    !atomic_read(&io[JQ_SUBMIT].pending_bios));
1481
1482         /* let's synchronous decompression */
1483         z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1484 }
1485
1486 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1487                                              struct page *page)
1488 {
1489         struct inode *const inode = page->mapping->host;
1490         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1491         int err;
1492         LIST_HEAD(pagepool);
1493
1494         trace_erofs_readpage(page, false);
1495
1496         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1497
1498         err = z_erofs_do_read_page(&f, page, &pagepool);
1499         (void)z_erofs_vle_work_iter_end(&f.builder);
1500
1501         if (err) {
1502                 errln("%s, failed to read, err [%d]", __func__, err);
1503                 goto out;
1504         }
1505
1506         z_erofs_submit_and_unzip(&f, &pagepool, true);
1507 out:
1508         if (f.map.mpage)
1509                 put_page(f.map.mpage);
1510
1511         /* clean up the remaining free pages */
1512         put_pages_list(&pagepool);
1513         return 0;
1514 }
1515
1516 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1517                                               struct address_space *mapping,
1518                                               struct list_head *pages,
1519                                               unsigned int nr_pages)
1520 {
1521         struct inode *const inode = mapping->host;
1522         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1523
1524         bool sync = __should_decompress_synchronously(sbi, nr_pages);
1525         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1526         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1527         struct page *head = NULL;
1528         LIST_HEAD(pagepool);
1529
1530         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1531                               nr_pages, false);
1532
1533         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1534
1535         for (; nr_pages; --nr_pages) {
1536                 struct page *page = lru_to_page(pages);
1537
1538                 prefetchw(&page->flags);
1539                 list_del(&page->lru);
1540
1541                 /*
1542                  * A pure asynchronous readahead is indicated if
1543                  * a PG_readahead marked page is hitted at first.
1544                  * Let's also do asynchronous decompression for this case.
1545                  */
1546                 sync &= !(PageReadahead(page) && !head);
1547
1548                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1549                         list_add(&page->lru, &pagepool);
1550                         continue;
1551                 }
1552
1553                 set_page_private(page, (unsigned long)head);
1554                 head = page;
1555         }
1556
1557         while (head) {
1558                 struct page *page = head;
1559                 int err;
1560
1561                 /* traversal in reverse order */
1562                 head = (void *)page_private(page);
1563
1564                 err = z_erofs_do_read_page(&f, page, &pagepool);
1565                 if (err) {
1566                         struct erofs_vnode *vi = EROFS_V(inode);
1567
1568                         errln("%s, readahead error at page %lu of nid %llu",
1569                               __func__, page->index, vi->nid);
1570                 }
1571
1572                 put_page(page);
1573         }
1574
1575         (void)z_erofs_vle_work_iter_end(&f.builder);
1576
1577         z_erofs_submit_and_unzip(&f, &pagepool, sync);
1578
1579         if (f.map.mpage)
1580                 put_page(f.map.mpage);
1581
1582         /* clean up the remaining free pages */
1583         put_pages_list(&pagepool);
1584         return 0;
1585 }
1586
1587 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1588         .readpage = z_erofs_vle_normalaccess_readpage,
1589         .readpages = z_erofs_vle_normalaccess_readpages,
1590 };
1591