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 was the tail of an exist chain, and the previous
 111          * processed chained works are all decided to be hooked up to it.
 112          * A new chain should be created for the remaining unprocessed works,
 113          * therefore different from Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
 114          * the next work cannot reuse the whole page in the following scenario:
 115          *  ________________________________________________________________
 116          * |      tail (partial) page     |       head (partial) page       |
 117          * |  (belongs to the next work)  |  (belongs to the current work)  |
 118          * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________|
 119          */
 120         Z_EROFS_VLE_WORK_PRIMARY_HOOKED,
 121         /*
 122          * The current work has been linked with the processed chained works,
 123          * and could be also linked with the potential remaining works, which
 124          * means if the processing page is the tail partial page of the work,
 125          * the current work can safely use the whole page (since the next work
 126          * is under control) for in-place decompression, as illustrated below:
 127          *  ________________________________________________________________
 128          * |  tail (partial) page  |          head (partial) page           |
 129          * | (of the current work) |         (of the previous work)         |
 130          * |  PRIMARY_FOLLOWED or  |                                        |
 131          * |_____PRIMARY_HOOKED____|____________PRIMARY_FOLLOWED____________|
 132          *
 133          * [  (*) the above page can be used for the current work itself.  ]
 134          */
 135         Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED,
 136         Z_EROFS_VLE_WORK_MAX
 137 };
 138
 139 struct z_erofs_vle_work_builder {
 140         enum z_erofs_vle_work_role role;
 141         /*
 142          * 'hosted = false' means that the current workgroup doesn't belong to
 143          * the owned chained workgroups. In the other words, it is none of our
 144          * business to submit this workgroup.
 145          */
 146         bool hosted;
 147
 148         struct z_erofs_vle_workgroup *grp;
 149         struct z_erofs_vle_work *work;
 150         struct z_erofs_pagevec_ctor vector;
 151
 152         /* pages used for reading the compressed data */
 153         struct page **compressed_pages;
 154         unsigned int compressed_deficit;
 155 };
 156
 157 #define VLE_WORK_BUILDER_INIT() \
 158         { .work = NULL, .role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED }
 159
 160 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 161 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 162                                      struct address_space *mc,
 163                                      pgoff_t index,
 164                                      unsigned int clusterpages,
 165                                      enum z_erofs_cache_alloctype type,
 166                                      struct list_head *pagepool,
 167                                      gfp_t gfp)
 168 {
 169         struct page **const pages = bl->compressed_pages;
 170         const unsigned int remaining = bl->compressed_deficit;
 171         bool standalone = true;
 172         unsigned int i, j = 0;
 173
 174         if (bl->role < Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 175                 return;
 176
 177         gfp = mapping_gfp_constraint(mc, gfp) & ~__GFP_RECLAIM;
 178
 179         index += clusterpages - remaining;
 180
 181         for (i = 0; i < remaining; ++i) {
 182                 struct page *page;
 183                 compressed_page_t t;
 184
 185                 /* the compressed page was loaded before */
 186                 if (READ_ONCE(pages[i]))
 187                         continue;
 188
 189                 page = find_get_page(mc, index + i);
 190
 191                 if (page) {
 192                         t = tag_compressed_page_justfound(page);
 193                 } else if (type == DELAYEDALLOC) {
 194                         t = tagptr_init(compressed_page_t, PAGE_UNALLOCATED);
 195                 } else {        /* DONTALLOC */
 196                         if (standalone)
 197                                 j = i;
 198                         standalone = false;
 199                         continue;
 200                 }
 201
 202                 if (!cmpxchg_relaxed(&pages[i], NULL, tagptr_cast_ptr(t)))
 203                         continue;
 204
 205                 if (page)
 206                         put_page(page);
 207         }
 208         bl->compressed_pages += j;
 209         bl->compressed_deficit = remaining - j;
 210
 211         if (standalone)
 212                 bl->role = Z_EROFS_VLE_WORK_PRIMARY;
 213 }
 214
 215 /* called by erofs_shrinker to get rid of all compressed_pages */
 216 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 217                                        struct erofs_workgroup *egrp)
 218 {
 219         struct z_erofs_vle_workgroup *const grp =
 220                 container_of(egrp, struct z_erofs_vle_workgroup, obj);
 221         struct address_space *const mapping = MNGD_MAPPING(sbi);
 222         const int clusterpages = erofs_clusterpages(sbi);
 223         int i;
 224
 225         /*
 226          * refcount of workgroup is now freezed as 1,
 227          * therefore no need to worry about available decompression users.
 228          */
 229         for (i = 0; i < clusterpages; ++i) {
 230                 struct page *page = grp->compressed_pages[i];
 231
 232                 if (!page || page->mapping != mapping)
 233                         continue;
 234
 235                 /* block other users from reclaiming or migrating the page */
 236                 if (!trylock_page(page))
 237                         return -EBUSY;
 238
 239                 /* barrier is implied in the following 'unlock_page' */
 240                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 241
 242                 set_page_private(page, 0);
 243                 ClearPagePrivate(page);
 244
 245                 unlock_page(page);
 246                 put_page(page);
 247         }
 248         return 0;
 249 }
 250
 251 int erofs_try_to_free_cached_page(struct address_space *mapping,
 252                                   struct page *page)
 253 {
 254         struct erofs_sb_info *const sbi = EROFS_SB(mapping->host->i_sb);
 255         const unsigned int clusterpages = erofs_clusterpages(sbi);
 256         struct z_erofs_vle_workgroup *const grp = (void *)page_private(page);
 257         int ret = 0;    /* 0 - busy */
 258
 259         if (erofs_workgroup_try_to_freeze(&grp->obj, 1)) {
 260                 unsigned int i;
 261
 262                 for (i = 0; i < clusterpages; ++i) {
 263                         if (grp->compressed_pages[i] == page) {
 264                                 WRITE_ONCE(grp->compressed_pages[i], NULL);
 265                                 ret = 1;
 266                                 break;
 267                         }
 268                 }
 269                 erofs_workgroup_unfreeze(&grp->obj, 1);
 270
 271                 if (ret) {
 272                         ClearPagePrivate(page);
 273                         put_page(page);
 274                 }
 275         }
 276         return ret;
 277 }
 278 #else
 279 static void preload_compressed_pages(struct z_erofs_vle_work_builder *bl,
 280                                      struct address_space *mc,
 281                                      pgoff_t index,
 282                                      unsigned int clusterpages,
 283                                      enum z_erofs_cache_alloctype type,
 284                                      struct list_head *pagepool,
 285                                      gfp_t gfp)
 286 {
 287         /* nowhere to load compressed pages from */
 288 }
 289 #endif
 290
 291 /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */
 292 static inline bool try_to_reuse_as_compressed_page(
 293         struct z_erofs_vle_work_builder *b,
 294         struct page *page)
 295 {
 296         while (b->compressed_deficit) {
 297                 --b->compressed_deficit;
 298                 if (!cmpxchg(b->compressed_pages++, NULL, page))
 299                         return true;
 300         }
 301
 302         return false;
 303 }
 304
 305 /* callers must be with work->lock held */
 306 static int z_erofs_vle_work_add_page(
 307         struct z_erofs_vle_work_builder *builder,
 308         struct page *page,
 309         enum z_erofs_page_type type)
 310 {
 311         int ret;
 312         bool occupied;
 313
 314         /* give priority for the compressed data storage */
 315         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY &&
 316                 type == Z_EROFS_PAGE_TYPE_EXCLUSIVE &&
 317                 try_to_reuse_as_compressed_page(builder, page))
 318                 return 0;
 319
 320         ret = z_erofs_pagevec_ctor_enqueue(&builder->vector,
 321                 page, type, &occupied);
 322         builder->work->vcnt += (unsigned int)ret;
 323
 324         return ret ? 0 : -EAGAIN;
 325 }
 326
 327 static enum z_erofs_vle_work_role
 328 try_to_claim_workgroup(struct z_erofs_vle_workgroup *grp,
 329                        z_erofs_vle_owned_workgrp_t *owned_head,
 330                        bool *hosted)
 331 {
 332         DBG_BUGON(*hosted == true);
 333
 334         /* let's claim these following types of workgroup */
 335 retry:
 336         if (grp->next == Z_EROFS_VLE_WORKGRP_NIL) {
 337                 /* type 1, nil workgroup */
 338                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_NIL,
 339                             *owned_head) != Z_EROFS_VLE_WORKGRP_NIL)
 340                         goto retry;
 341
 342                 *owned_head = &grp->next;
 343                 *hosted = true;
 344                 /* lucky, I am the followee :) */
 345                 return Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 346
 347         } else if (grp->next == Z_EROFS_VLE_WORKGRP_TAIL) {
 348                 /*
 349                  * type 2, link to the end of a existing open chain,
 350                  * be careful that its submission itself is governed
 351                  * by the original owned chain.
 352                  */
 353                 if (cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
 354                             *owned_head) != Z_EROFS_VLE_WORKGRP_TAIL)
 355                         goto retry;
 356                 *owned_head = Z_EROFS_VLE_WORKGRP_TAIL;
 357                 return Z_EROFS_VLE_WORK_PRIMARY_HOOKED;
 358         }
 359
 360         return Z_EROFS_VLE_WORK_PRIMARY; /* :( better luck next time */
 361 }
 362
 363 struct z_erofs_vle_work_finder {
 364         struct super_block *sb;
 365         pgoff_t idx;
 366         unsigned int pageofs;
 367
 368         struct z_erofs_vle_workgroup **grp_ret;
 369         enum z_erofs_vle_work_role *role;
 370         z_erofs_vle_owned_workgrp_t *owned_head;
 371         bool *hosted;
 372 };
 373
 374 static struct z_erofs_vle_work *
 375 z_erofs_vle_work_lookup(const struct z_erofs_vle_work_finder *f)
 376 {
 377         bool tag, primary;
 378         struct erofs_workgroup *egrp;
 379         struct z_erofs_vle_workgroup *grp;
 380         struct z_erofs_vle_work *work;
 381
 382         egrp = erofs_find_workgroup(f->sb, f->idx, &tag);
 383         if (!egrp) {
 384                 *f->grp_ret = NULL;
 385                 return NULL;
 386         }
 387
 388         grp = container_of(egrp, struct z_erofs_vle_workgroup, obj);
 389         *f->grp_ret = grp;
 390
 391         work = z_erofs_vle_grab_work(grp, f->pageofs);
 392         /* if multiref is disabled, `primary' is always true */
 393         primary = true;
 394
 395         DBG_BUGON(work->pageofs != f->pageofs);
 396
 397         /*
 398          * lock must be taken first to avoid grp->next == NIL between
 399          * claiming workgroup and adding pages:
 400          *                        grp->next != NIL
 401          *   grp->next = NIL
 402          *   mutex_unlock_all
 403          *                        mutex_lock(&work->lock)
 404          *                        add all pages to pagevec
 405          *
 406          * [correct locking case 1]:
 407          *   mutex_lock(grp->work[a])
 408          *   ...
 409          *   mutex_lock(grp->work[b])     mutex_lock(grp->work[c])
 410          *   ...                          *role = SECONDARY
 411          *                                add all pages to pagevec
 412          *                                ...
 413          *                                mutex_unlock(grp->work[c])
 414          *   mutex_lock(grp->work[c])
 415          *   ...
 416          *   grp->next = NIL
 417          *   mutex_unlock_all
 418          *
 419          * [correct locking case 2]:
 420          *   mutex_lock(grp->work[b])
 421          *   ...
 422          *   mutex_lock(grp->work[a])
 423          *   ...
 424          *   mutex_lock(grp->work[c])
 425          *   ...
 426          *   grp->next = NIL
 427          *   mutex_unlock_all
 428          *                                mutex_lock(grp->work[a])
 429          *                                *role = PRIMARY_OWNER
 430          *                                add all pages to pagevec
 431          *                                ...
 432          */
 433         mutex_lock(&work->lock);
 434
 435         *f->hosted = false;
 436         if (!primary)
 437                 *f->role = Z_EROFS_VLE_WORK_SECONDARY;
 438         else    /* claim the workgroup if possible */
 439                 *f->role = try_to_claim_workgroup(grp, f->owned_head,
 440                                                   f->hosted);
 441         return work;
 442 }
 443
 444 static struct z_erofs_vle_work *
 445 z_erofs_vle_work_register(const struct z_erofs_vle_work_finder *f,
 446                           struct erofs_map_blocks *map)
 447 {
 448         bool gnew = false;
 449         struct z_erofs_vle_workgroup *grp = *f->grp_ret;
 450         struct z_erofs_vle_work *work;
 451
 452         /* if multiref is disabled, grp should never be nullptr */
 453         if (unlikely(grp)) {
 454                 DBG_BUGON(1);
 455                 return ERR_PTR(-EINVAL);
 456         }
 457
 458         /* no available workgroup, let's allocate one */
 459         grp = kmem_cache_alloc(z_erofs_workgroup_cachep, GFP_NOFS);
 460         if (unlikely(!grp))
 461                 return ERR_PTR(-ENOMEM);
 462
 463         init_always(grp);
 464         grp->obj.index = f->idx;
 465         grp->llen = map->m_llen;
 466
 467         z_erofs_vle_set_workgrp_fmt(grp,
 468                 (map->m_flags & EROFS_MAP_ZIPPED) ?
 469                         Z_EROFS_VLE_WORKGRP_FMT_LZ4 :
 470                         Z_EROFS_VLE_WORKGRP_FMT_PLAIN);
 471
 472         /* new workgrps have been claimed as type 1 */
 473         WRITE_ONCE(grp->next, *f->owned_head);
 474         /* primary and followed work for all new workgrps */
 475         *f->role = Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED;
 476         /* it should be submitted by ourselves */
 477         *f->hosted = true;
 478
 479         gnew = true;
 480         work = z_erofs_vle_grab_primary_work(grp);
 481         work->pageofs = f->pageofs;
 482
 483         /*
 484          * lock all primary followed works before visible to others
 485          * and mutex_trylock *never* fails for a new workgroup.
 486          */
 487         mutex_trylock(&work->lock);
 488
 489         if (gnew) {
 490                 int err = erofs_register_workgroup(f->sb, &grp->obj, 0);
 491
 492                 if (err) {
 493                         mutex_unlock(&work->lock);
 494                         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 495                         return ERR_PTR(-EAGAIN);
 496                 }
 497         }
 498
 499         *f->owned_head = &grp->next;
 500         *f->grp_ret = grp;
 501         return work;
 502 }
 503
 504 #define builder_is_hooked(builder) \
 505         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_HOOKED)
 506
 507 #define builder_is_followed(builder) \
 508         ((builder)->role >= Z_EROFS_VLE_WORK_PRIMARY_FOLLOWED)
 509
 510 static int z_erofs_vle_work_iter_begin(struct z_erofs_vle_work_builder *builder,
 511                                        struct super_block *sb,
 512                                        struct erofs_map_blocks *map,
 513                                        z_erofs_vle_owned_workgrp_t *owned_head)
 514 {
 515         const unsigned int clusterpages = erofs_clusterpages(EROFS_SB(sb));
 516         struct z_erofs_vle_workgroup *grp;
 517         const struct z_erofs_vle_work_finder finder = {
 518                 .sb = sb,
 519                 .idx = erofs_blknr(map->m_pa),
 520                 .pageofs = map->m_la & ~PAGE_MASK,
 521                 .grp_ret = &grp,
 522                 .role = &builder->role,
 523                 .owned_head = owned_head,
 524                 .hosted = &builder->hosted
 525         };
 526         struct z_erofs_vle_work *work;
 527
 528         DBG_BUGON(builder->work);
 529
 530         /* must be Z_EROFS_WORK_TAIL or the next chained work */
 531         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_NIL);
 532         DBG_BUGON(*owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
 533
 534         DBG_BUGON(erofs_blkoff(map->m_pa));
 535
 536 repeat:
 537         work = z_erofs_vle_work_lookup(&finder);
 538         if (work) {
 539                 unsigned int orig_llen;
 540
 541                 /* increase workgroup `llen' if needed */
 542                 while ((orig_llen = READ_ONCE(grp->llen)) < map->m_llen &&
 543                        orig_llen != cmpxchg_relaxed(&grp->llen,
 544                                                     orig_llen, map->m_llen))
 545                         cpu_relax();
 546                 goto got_it;
 547         }
 548
 549         work = z_erofs_vle_work_register(&finder, map);
 550         if (unlikely(work == ERR_PTR(-EAGAIN)))
 551                 goto repeat;
 552
 553         if (IS_ERR(work))
 554                 return PTR_ERR(work);
 555 got_it:
 556         z_erofs_pagevec_ctor_init(&builder->vector,
 557                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, work->vcnt);
 558
 559         if (builder->role >= Z_EROFS_VLE_WORK_PRIMARY) {
 560                 /* enable possibly in-place decompression */
 561                 builder->compressed_pages = grp->compressed_pages;
 562                 builder->compressed_deficit = clusterpages;
 563         } else {
 564                 builder->compressed_pages = NULL;
 565                 builder->compressed_deficit = 0;
 566         }
 567
 568         builder->grp = grp;
 569         builder->work = work;
 570         return 0;
 571 }
 572
 573 /*
 574  * keep in mind that no referenced workgroups will be freed
 575  * only after a RCU grace period, so rcu_read_lock() could
 576  * prevent a workgroup from being freed.
 577  */
 578 static void z_erofs_rcu_callback(struct rcu_head *head)
 579 {
 580         struct z_erofs_vle_work *work = container_of(head,
 581                 struct z_erofs_vle_work, rcu);
 582         struct z_erofs_vle_workgroup *grp =
 583                 z_erofs_vle_work_workgroup(work, true);
 584
 585         kmem_cache_free(z_erofs_workgroup_cachep, grp);
 586 }
 587
 588 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 589 {
 590         struct z_erofs_vle_workgroup *const vgrp = container_of(grp,
 591                 struct z_erofs_vle_workgroup, obj);
 592         struct z_erofs_vle_work *const work = &vgrp->work;
 593
 594         call_rcu(&work->rcu, z_erofs_rcu_callback);
 595 }
 596
 597 static void __z_erofs_vle_work_release(struct z_erofs_vle_workgroup *grp,
 598         struct z_erofs_vle_work *work __maybe_unused)
 599 {
 600         erofs_workgroup_put(&grp->obj);
 601 }
 602
 603 static void z_erofs_vle_work_release(struct z_erofs_vle_work *work)
 604 {
 605         struct z_erofs_vle_workgroup *grp =
 606                 z_erofs_vle_work_workgroup(work, true);
 607
 608         __z_erofs_vle_work_release(grp, work);
 609 }
 610
 611 static inline bool
 612 z_erofs_vle_work_iter_end(struct z_erofs_vle_work_builder *builder)
 613 {
 614         struct z_erofs_vle_work *work = builder->work;
 615
 616         if (!work)
 617                 return false;
 618
 619         z_erofs_pagevec_ctor_exit(&builder->vector, false);
 620         mutex_unlock(&work->lock);
 621
 622         /*
 623          * if all pending pages are added, don't hold work reference
 624          * any longer if the current work isn't hosted by ourselves.
 625          */
 626         if (!builder->hosted)
 627                 __z_erofs_vle_work_release(builder->grp, work);
 628
 629         builder->work = NULL;
 630         builder->grp = NULL;
 631         return true;
 632 }
 633
 634 static inline struct page *__stagingpage_alloc(struct list_head *pagepool,
 635                                                gfp_t gfp)
 636 {
 637         struct page *page = erofs_allocpage(pagepool, gfp);
 638
 639         if (unlikely(!page))
 640                 return NULL;
 641
 642         page->mapping = Z_EROFS_MAPPING_STAGING;
 643         return page;
 644 }
 645
 646 struct z_erofs_vle_frontend {
 647         struct inode *const inode;
 648
 649         struct z_erofs_vle_work_builder builder;
 650         struct erofs_map_blocks map;
 651
 652         z_erofs_vle_owned_workgrp_t owned_head;
 653
 654         /* used for applying cache strategy on the fly */
 655         bool backmost;
 656         erofs_off_t headoffset;
 657 };
 658
 659 #define VLE_FRONTEND_INIT(__i) { \
 660         .inode = __i, \
 661         .map = { \
 662                 .m_llen = 0, \
 663                 .m_plen = 0, \
 664                 .mpage = NULL \
 665         }, \
 666         .builder = VLE_WORK_BUILDER_INIT(), \
 667         .owned_head = Z_EROFS_VLE_WORKGRP_TAIL, \
 668         .backmost = true, }
 669
 670 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 671 static inline bool
 672 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 673 {
 674         if (fe->backmost)
 675                 return true;
 676
 677         if (EROFS_FS_ZIP_CACHE_LVL >= 2)
 678                 return la < fe->headoffset;
 679
 680         return false;
 681 }
 682 #else
 683 static inline bool
 684 should_alloc_managed_pages(struct z_erofs_vle_frontend *fe, erofs_off_t la)
 685 {
 686         return false;
 687 }
 688 #endif
 689
 690 static int z_erofs_do_read_page(struct z_erofs_vle_frontend *fe,
 691                                 struct page *page,
 692                                 struct list_head *page_pool)
 693 {
 694         struct super_block *const sb = fe->inode->i_sb;
 695         struct erofs_sb_info *const sbi __maybe_unused = EROFS_SB(sb);
 696         struct erofs_map_blocks *const map = &fe->map;
 697         struct z_erofs_vle_work_builder *const builder = &fe->builder;
 698         const loff_t offset = page_offset(page);
 699
 700         bool tight = builder_is_hooked(builder);
 701         struct z_erofs_vle_work *work = builder->work;
 702
 703         enum z_erofs_cache_alloctype cache_strategy;
 704         enum z_erofs_page_type page_type;
 705         unsigned int cur, end, spiltted, index;
 706         int err = 0;
 707
 708         /* register locked file pages as online pages in pack */
 709         z_erofs_onlinepage_init(page);
 710
 711         spiltted = 0;
 712         end = PAGE_SIZE;
 713 repeat:
 714         cur = end - 1;
 715
 716         /* lucky, within the range of the current map_blocks */
 717         if (offset + cur >= map->m_la &&
 718                 offset + cur < map->m_la + map->m_llen) {
 719                 /* didn't get a valid unzip work previously (very rare) */
 720                 if (!builder->work)
 721                         goto restart_now;
 722                 goto hitted;
 723         }
 724
 725         /* go ahead the next map_blocks */
 726         debugln("%s: [out-of-range] pos %llu", __func__, offset + cur);
 727
 728         if (z_erofs_vle_work_iter_end(builder))
 729                 fe->backmost = false;
 730
 731         map->m_la = offset + cur;
 732         map->m_llen = 0;
 733         err = z_erofs_map_blocks_iter(fe->inode, map, 0);
 734         if (unlikely(err))
 735                 goto err_out;
 736
 737 restart_now:
 738         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED)))
 739                 goto hitted;
 740
 741         DBG_BUGON(map->m_plen != 1 << sbi->clusterbits);
 742         DBG_BUGON(erofs_blkoff(map->m_pa));
 743
 744         err = z_erofs_vle_work_iter_begin(builder, sb, map, &fe->owned_head);
 745         if (unlikely(err))
 746                 goto err_out;
 747
 748         /* preload all compressed pages (maybe downgrade role if necessary) */
 749         if (should_alloc_managed_pages(fe, map->m_la))
 750                 cache_strategy = DELAYEDALLOC;
 751         else
 752                 cache_strategy = DONTALLOC;
 753
 754         preload_compressed_pages(builder, MNGD_MAPPING(sbi),
 755                                  map->m_pa / PAGE_SIZE,
 756                                  map->m_plen / PAGE_SIZE,
 757                                  cache_strategy, page_pool, GFP_KERNEL);
 758
 759         tight &= builder_is_hooked(builder);
 760         work = builder->work;
 761 hitted:
 762         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 763         if (unlikely(!(map->m_flags & EROFS_MAP_MAPPED))) {
 764                 zero_user_segment(page, cur, end);
 765                 goto next_part;
 766         }
 767
 768         /* let's derive page type */
 769         page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD :
 770                 (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 771                         (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE :
 772                                 Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED));
 773
 774         if (cur)
 775                 tight &= builder_is_followed(builder);
 776
 777 retry:
 778         err = z_erofs_vle_work_add_page(builder, page, page_type);
 779         /* should allocate an additional staging page for pagevec */
 780         if (err == -EAGAIN) {
 781                 struct page *const newpage =
 782                         __stagingpage_alloc(page_pool, GFP_NOFS);
 783
 784                 err = z_erofs_vle_work_add_page(builder,
 785                         newpage, Z_EROFS_PAGE_TYPE_EXCLUSIVE);
 786                 if (likely(!err))
 787                         goto retry;
 788         }
 789
 790         if (unlikely(err))
 791                 goto err_out;
 792
 793         index = page->index - map->m_la / PAGE_SIZE;
 794
 795         /* FIXME! avoid the last relundant fixup & endio */
 796         z_erofs_onlinepage_fixup(page, index, true);
 797
 798         /* bump up the number of spiltted parts of a page */
 799         ++spiltted;
 800         /* also update nr_pages */
 801         work->nr_pages = max_t(pgoff_t, work->nr_pages, index + 1);
 802 next_part:
 803         /* can be used for verification */
 804         map->m_llen = offset + cur - map->m_la;
 805
 806         end = cur;
 807         if (end > 0)
 808                 goto repeat;
 809
 810 out:
 811         /* FIXME! avoid the last relundant fixup & endio */
 812         z_erofs_onlinepage_endio(page);
 813
 814         debugln("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 815                 __func__, page, spiltted, map->m_llen);
 816         return err;
 817
 818         /* if some error occurred while processing this page */
 819 err_out:
 820         SetPageError(page);
 821         goto out;
 822 }
 823
 824 static void z_erofs_vle_unzip_kickoff(void *ptr, int bios)
 825 {
 826         tagptr1_t t = tagptr_init(tagptr1_t, ptr);
 827         struct z_erofs_vle_unzip_io *io = tagptr_unfold_ptr(t);
 828         bool background = tagptr_unfold_tags(t);
 829
 830         if (!background) {
 831                 unsigned long flags;
 832
 833                 spin_lock_irqsave(&io->u.wait.lock, flags);
 834                 if (!atomic_add_return(bios, &io->pending_bios))
 835                         wake_up_locked(&io->u.wait);
 836                 spin_unlock_irqrestore(&io->u.wait.lock, flags);
 837                 return;
 838         }
 839
 840         if (!atomic_add_return(bios, &io->pending_bios))
 841                 queue_work(z_erofs_workqueue, &io->u.work);
 842 }
 843
 844 static inline void z_erofs_vle_read_endio(struct bio *bio)
 845 {
 846         const blk_status_t err = bio->bi_status;
 847         unsigned int i;
 848         struct bio_vec *bvec;
 849 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 850         struct address_space *mc = NULL;
 851 #endif
 852         struct bvec_iter_all iter_all;
 853
 854         bio_for_each_segment_all(bvec, bio, i, iter_all) {
 855                 struct page *page = bvec->bv_page;
 856                 bool cachemngd = false;
 857
 858                 DBG_BUGON(PageUptodate(page));
 859                 DBG_BUGON(!page->mapping);
 860
 861 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 862                 if (unlikely(!mc && !z_erofs_is_stagingpage(page))) {
 863                         struct inode *const inode = page->mapping->host;
 864                         struct super_block *const sb = inode->i_sb;
 865
 866                         mc = MNGD_MAPPING(EROFS_SB(sb));
 867                 }
 868
 869                 /*
 870                  * If mc has not gotten, it equals NULL,
 871                  * however, page->mapping never be NULL if working properly.
 872                  */
 873                 cachemngd = (page->mapping == mc);
 874 #endif
 875
 876                 if (unlikely(err))
 877                         SetPageError(page);
 878                 else if (cachemngd)
 879                         SetPageUptodate(page);
 880
 881                 if (cachemngd)
 882                         unlock_page(page);
 883         }
 884
 885         z_erofs_vle_unzip_kickoff(bio->bi_private, -1);
 886         bio_put(bio);
 887 }
 888
 889 static struct page *z_pagemap_global[Z_EROFS_VLE_VMAP_GLOBAL_PAGES];
 890 static DEFINE_MUTEX(z_pagemap_global_lock);
 891
 892 static int z_erofs_vle_unzip(struct super_block *sb,
 893         struct z_erofs_vle_workgroup *grp,
 894         struct list_head *page_pool)
 895 {
 896         struct erofs_sb_info *const sbi = EROFS_SB(sb);
 897         const unsigned int clusterpages = erofs_clusterpages(sbi);
 898
 899         struct z_erofs_pagevec_ctor ctor;
 900         unsigned int nr_pages;
 901         unsigned int sparsemem_pages = 0;
 902         struct page *pages_onstack[Z_EROFS_VLE_VMAP_ONSTACK_PAGES];
 903         struct page **pages, **compressed_pages, *page;
 904         unsigned int i, llen;
 905
 906         enum z_erofs_page_type page_type;
 907         bool overlapped;
 908         struct z_erofs_vle_work *work;
 909         void *vout;
 910         int err;
 911
 912         might_sleep();
 913         work = z_erofs_vle_grab_primary_work(grp);
 914         DBG_BUGON(!READ_ONCE(work->nr_pages));
 915
 916         mutex_lock(&work->lock);
 917         nr_pages = work->nr_pages;
 918
 919         if (likely(nr_pages <= Z_EROFS_VLE_VMAP_ONSTACK_PAGES))
 920                 pages = pages_onstack;
 921         else if (nr_pages <= Z_EROFS_VLE_VMAP_GLOBAL_PAGES &&
 922                 mutex_trylock(&z_pagemap_global_lock))
 923                 pages = z_pagemap_global;
 924         else {
 925 repeat:
 926                 pages = kvmalloc_array(nr_pages,
 927                         sizeof(struct page *), GFP_KERNEL);
 928
 929                 /* fallback to global pagemap for the lowmem scenario */
 930                 if (unlikely(!pages)) {
 931                         if (nr_pages > Z_EROFS_VLE_VMAP_GLOBAL_PAGES)
 932                                 goto repeat;
 933                         else {
 934                                 mutex_lock(&z_pagemap_global_lock);
 935                                 pages = z_pagemap_global;
 936                         }
 937                 }
 938         }
 939
 940         for (i = 0; i < nr_pages; ++i)
 941                 pages[i] = NULL;
 942
 943         z_erofs_pagevec_ctor_init(&ctor,
 944                 Z_EROFS_VLE_INLINE_PAGEVECS, work->pagevec, 0);
 945
 946         for (i = 0; i < work->vcnt; ++i) {
 947                 unsigned int pagenr;
 948
 949                 page = z_erofs_pagevec_ctor_dequeue(&ctor, &page_type);
 950
 951                 /* all pages in pagevec ought to be valid */
 952                 DBG_BUGON(!page);
 953                 DBG_BUGON(!page->mapping);
 954
 955                 if (z_erofs_gather_if_stagingpage(page_pool, page))
 956                         continue;
 957
 958                 if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD)
 959                         pagenr = 0;
 960                 else
 961                         pagenr = z_erofs_onlinepage_index(page);
 962
 963                 DBG_BUGON(pagenr >= nr_pages);
 964                 DBG_BUGON(pages[pagenr]);
 965
 966                 pages[pagenr] = page;
 967         }
 968         sparsemem_pages = i;
 969
 970         z_erofs_pagevec_ctor_exit(&ctor, true);
 971
 972         overlapped = false;
 973         compressed_pages = grp->compressed_pages;
 974
 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_is_stagingpage(page))
 985                         continue;
 986 #ifdef EROFS_FS_HAS_MANAGED_CACHE
 987                 if (page->mapping == MNGD_MAPPING(sbi)) {
 988                         DBG_BUGON(!PageUptodate(page));
 989                         continue;
 990                 }
 991 #endif
 992
 993                 /* only non-head page could be reused as a compressed page */
 994                 pagenr = z_erofs_onlinepage_index(page);
 995
 996                 DBG_BUGON(pagenr >= nr_pages);
 997                 DBG_BUGON(pages[pagenr]);
 998                 ++sparsemem_pages;
 999                 pages[pagenr] = page;
1000
1001                 overlapped = true;
1002         }
1003
1004         llen = (nr_pages << PAGE_SHIFT) - work->pageofs;
1005
1006         if (z_erofs_vle_workgrp_fmt(grp) == Z_EROFS_VLE_WORKGRP_FMT_PLAIN) {
1007                 err = z_erofs_vle_plain_copy(compressed_pages, clusterpages,
1008                         pages, nr_pages, work->pageofs);
1009                 goto out;
1010         }
1011
1012         if (llen > grp->llen)
1013                 llen = grp->llen;
1014
1015         err = z_erofs_vle_unzip_fast_percpu(compressed_pages, clusterpages,
1016                                             pages, llen, work->pageofs);
1017         if (err != -ENOTSUPP)
1018                 goto out;
1019
1020         if (sparsemem_pages >= nr_pages)
1021                 goto skip_allocpage;
1022
1023         for (i = 0; i < nr_pages; ++i) {
1024                 if (pages[i])
1025                         continue;
1026
1027                 pages[i] = __stagingpage_alloc(page_pool, GFP_NOFS);
1028         }
1029
1030 skip_allocpage:
1031         vout = erofs_vmap(pages, nr_pages);
1032
1033         err = z_erofs_vle_unzip_vmap(compressed_pages,
1034                 clusterpages, vout, llen, work->pageofs, overlapped);
1035
1036         erofs_vunmap(vout, nr_pages);
1037
1038 out:
1039         /* must handle all compressed pages before endding pages */
1040         for (i = 0; i < clusterpages; ++i) {
1041                 page = compressed_pages[i];
1042
1043 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1044                 if (page->mapping == MNGD_MAPPING(sbi))
1045                         continue;
1046 #endif
1047                 /* recycle all individual staging pages */
1048                 (void)z_erofs_gather_if_stagingpage(page_pool, page);
1049
1050                 WRITE_ONCE(compressed_pages[i], NULL);
1051         }
1052
1053         for (i = 0; i < nr_pages; ++i) {
1054                 page = pages[i];
1055                 if (!page)
1056                         continue;
1057
1058                 DBG_BUGON(!page->mapping);
1059
1060                 /* recycle all individual staging pages */
1061                 if (z_erofs_gather_if_stagingpage(page_pool, page))
1062                         continue;
1063
1064                 if (unlikely(err < 0))
1065                         SetPageError(page);
1066
1067                 z_erofs_onlinepage_endio(page);
1068         }
1069
1070         if (pages == z_pagemap_global)
1071                 mutex_unlock(&z_pagemap_global_lock);
1072         else if (unlikely(pages != pages_onstack))
1073                 kvfree(pages);
1074
1075         work->nr_pages = 0;
1076         work->vcnt = 0;
1077
1078         /* all work locks MUST be taken before the following line */
1079
1080         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_NIL);
1081
1082         /* all work locks SHOULD be released right now */
1083         mutex_unlock(&work->lock);
1084
1085         z_erofs_vle_work_release(work);
1086         return err;
1087 }
1088
1089 static void z_erofs_vle_unzip_all(struct super_block *sb,
1090                                   struct z_erofs_vle_unzip_io *io,
1091                                   struct list_head *page_pool)
1092 {
1093         z_erofs_vle_owned_workgrp_t owned = io->head;
1094
1095         while (owned != Z_EROFS_VLE_WORKGRP_TAIL_CLOSED) {
1096                 struct z_erofs_vle_workgroup *grp;
1097
1098                 /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1099                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_TAIL);
1100
1101                 /* no possible that 'owned' equals NULL */
1102                 DBG_BUGON(owned == Z_EROFS_VLE_WORKGRP_NIL);
1103
1104                 grp = container_of(owned, struct z_erofs_vle_workgroup, next);
1105                 owned = READ_ONCE(grp->next);
1106
1107                 z_erofs_vle_unzip(sb, grp, page_pool);
1108         }
1109 }
1110
1111 static void z_erofs_vle_unzip_wq(struct work_struct *work)
1112 {
1113         struct z_erofs_vle_unzip_io_sb *iosb = container_of(work,
1114                 struct z_erofs_vle_unzip_io_sb, io.u.work);
1115         LIST_HEAD(page_pool);
1116
1117         DBG_BUGON(iosb->io.head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1118         z_erofs_vle_unzip_all(iosb->sb, &iosb->io, &page_pool);
1119
1120         put_pages_list(&page_pool);
1121         kvfree(iosb);
1122 }
1123
1124 static struct page *
1125 pickup_page_for_submission(struct z_erofs_vle_workgroup *grp,
1126                            unsigned int nr,
1127                            struct list_head *pagepool,
1128                            struct address_space *mc,
1129                            gfp_t gfp)
1130 {
1131         /* determined at compile time to avoid too many #ifdefs */
1132         const bool nocache = __builtin_constant_p(mc) ? !mc : false;
1133         const pgoff_t index = grp->obj.index;
1134         bool tocache = false;
1135
1136         struct address_space *mapping;
1137         struct page *oldpage, *page;
1138
1139         compressed_page_t t;
1140         int justfound;
1141
1142 repeat:
1143         page = READ_ONCE(grp->compressed_pages[nr]);
1144         oldpage = page;
1145
1146         if (!page)
1147                 goto out_allocpage;
1148
1149         /*
1150          * the cached page has not been allocated and
1151          * an placeholder is out there, prepare it now.
1152          */
1153         if (!nocache && page == PAGE_UNALLOCATED) {
1154                 tocache = true;
1155                 goto out_allocpage;
1156         }
1157
1158         /* process the target tagged pointer */
1159         t = tagptr_init(compressed_page_t, page);
1160         justfound = tagptr_unfold_tags(t);
1161         page = tagptr_unfold_ptr(t);
1162
1163         mapping = READ_ONCE(page->mapping);
1164
1165         /*
1166          * if managed cache is disabled, it's no way to
1167          * get such a cached-like page.
1168          */
1169         if (nocache) {
1170                 /* if managed cache is disabled, it is impossible `justfound' */
1171                 DBG_BUGON(justfound);
1172
1173                 /* and it should be locked, not uptodate, and not truncated */
1174                 DBG_BUGON(!PageLocked(page));
1175                 DBG_BUGON(PageUptodate(page));
1176                 DBG_BUGON(!mapping);
1177                 goto out;
1178         }
1179
1180         /*
1181          * unmanaged (file) pages are all locked solidly,
1182          * therefore it is impossible for `mapping' to be NULL.
1183          */
1184         if (mapping && mapping != mc)
1185                 /* ought to be unmanaged pages */
1186                 goto out;
1187
1188         lock_page(page);
1189
1190         /* only true if page reclaim goes wrong, should never happen */
1191         DBG_BUGON(justfound && PagePrivate(page));
1192
1193         /* the page is still in manage cache */
1194         if (page->mapping == mc) {
1195                 WRITE_ONCE(grp->compressed_pages[nr], page);
1196
1197                 if (!PagePrivate(page)) {
1198                         /*
1199                          * impossible to be !PagePrivate(page) for
1200                          * the current restriction as well if
1201                          * the page is already in compressed_pages[].
1202                          */
1203                         DBG_BUGON(!justfound);
1204
1205                         justfound = 0;
1206                         set_page_private(page, (unsigned long)grp);
1207                         SetPagePrivate(page);
1208                 }
1209
1210                 /* no need to submit io if it is already up-to-date */
1211                 if (PageUptodate(page)) {
1212                         unlock_page(page);
1213                         page = NULL;
1214                 }
1215                 goto out;
1216         }
1217
1218         /*
1219          * the managed page has been truncated, it's unsafe to
1220          * reuse this one, let's allocate a new cache-managed page.
1221          */
1222         DBG_BUGON(page->mapping);
1223         DBG_BUGON(!justfound);
1224
1225         tocache = true;
1226         unlock_page(page);
1227         put_page(page);
1228 out_allocpage:
1229         page = __stagingpage_alloc(pagepool, gfp);
1230         if (oldpage != cmpxchg(&grp->compressed_pages[nr], oldpage, page)) {
1231                 list_add(&page->lru, pagepool);
1232                 cpu_relax();
1233                 goto repeat;
1234         }
1235         if (nocache || !tocache)
1236                 goto out;
1237         if (add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1238                 page->mapping = Z_EROFS_MAPPING_STAGING;
1239                 goto out;
1240         }
1241
1242         set_page_private(page, (unsigned long)grp);
1243         SetPagePrivate(page);
1244 out:    /* the only exit (for tracing and debugging) */
1245         return page;
1246 }
1247
1248 static struct z_erofs_vle_unzip_io *
1249 jobqueue_init(struct super_block *sb,
1250               struct z_erofs_vle_unzip_io *io,
1251               bool foreground)
1252 {
1253         struct z_erofs_vle_unzip_io_sb *iosb;
1254
1255         if (foreground) {
1256                 /* waitqueue available for foreground io */
1257                 DBG_BUGON(!io);
1258
1259                 init_waitqueue_head(&io->u.wait);
1260                 atomic_set(&io->pending_bios, 0);
1261                 goto out;
1262         }
1263
1264         iosb = kvzalloc(sizeof(struct z_erofs_vle_unzip_io_sb),
1265                         GFP_KERNEL | __GFP_NOFAIL);
1266         DBG_BUGON(!iosb);
1267
1268         /* initialize fields in the allocated descriptor */
1269         io = &iosb->io;
1270         iosb->sb = sb;
1271         INIT_WORK(&io->u.work, z_erofs_vle_unzip_wq);
1272 out:
1273         io->head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1274         return io;
1275 }
1276
1277 /* define workgroup jobqueue types */
1278 enum {
1279 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1280         JQ_BYPASS,
1281 #endif
1282         JQ_SUBMIT,
1283         NR_JOBQUEUES,
1284 };
1285
1286 static void *jobqueueset_init(struct super_block *sb,
1287                               z_erofs_vle_owned_workgrp_t qtail[],
1288                               struct z_erofs_vle_unzip_io *q[],
1289                               struct z_erofs_vle_unzip_io *fgq,
1290                               bool forcefg)
1291 {
1292 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1293         /*
1294          * if managed cache is enabled, bypass jobqueue is needed,
1295          * no need to read from device for all workgroups in this queue.
1296          */
1297         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, true);
1298         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1299 #endif
1300
1301         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, forcefg);
1302         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1303
1304         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], !forcefg));
1305 }
1306
1307 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1308 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1309                                     z_erofs_vle_owned_workgrp_t qtail[],
1310                                     z_erofs_vle_owned_workgrp_t owned_head)
1311 {
1312         z_erofs_vle_owned_workgrp_t *const submit_qtail = qtail[JQ_SUBMIT];
1313         z_erofs_vle_owned_workgrp_t *const bypass_qtail = qtail[JQ_BYPASS];
1314
1315         DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1316         if (owned_head == Z_EROFS_VLE_WORKGRP_TAIL)
1317                 owned_head = Z_EROFS_VLE_WORKGRP_TAIL_CLOSED;
1318
1319         WRITE_ONCE(grp->next, Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1320
1321         WRITE_ONCE(*submit_qtail, owned_head);
1322         WRITE_ONCE(*bypass_qtail, &grp->next);
1323
1324         qtail[JQ_BYPASS] = &grp->next;
1325 }
1326
1327 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1328                                        unsigned int nr_bios,
1329                                        bool force_fg)
1330 {
1331         /*
1332          * although background is preferred, no one is pending for submission.
1333          * don't issue workqueue for decompression but drop it directly instead.
1334          */
1335         if (force_fg || nr_bios)
1336                 return false;
1337
1338         kvfree(container_of(q[JQ_SUBMIT],
1339                             struct z_erofs_vle_unzip_io_sb,
1340                             io));
1341         return true;
1342 }
1343 #else
1344 static void move_to_bypass_jobqueue(struct z_erofs_vle_workgroup *grp,
1345                                     z_erofs_vle_owned_workgrp_t qtail[],
1346                                     z_erofs_vle_owned_workgrp_t owned_head)
1347 {
1348         /* impossible to bypass submission for managed cache disabled */
1349         DBG_BUGON(1);
1350 }
1351
1352 static bool postsubmit_is_all_bypassed(struct z_erofs_vle_unzip_io *q[],
1353                                        unsigned int nr_bios,
1354                                        bool force_fg)
1355 {
1356         /* bios should be >0 if managed cache is disabled */
1357         DBG_BUGON(!nr_bios);
1358         return false;
1359 }
1360 #endif
1361
1362 static bool z_erofs_vle_submit_all(struct super_block *sb,
1363                                    z_erofs_vle_owned_workgrp_t owned_head,
1364                                    struct list_head *pagepool,
1365                                    struct z_erofs_vle_unzip_io *fgq,
1366                                    bool force_fg)
1367 {
1368         struct erofs_sb_info *const sbi = EROFS_SB(sb);
1369         const unsigned int clusterpages = erofs_clusterpages(sbi);
1370         const gfp_t gfp = GFP_NOFS;
1371
1372         z_erofs_vle_owned_workgrp_t qtail[NR_JOBQUEUES];
1373         struct z_erofs_vle_unzip_io *q[NR_JOBQUEUES];
1374         struct bio *bio;
1375         void *bi_private;
1376         /* since bio will be NULL, no need to initialize last_index */
1377         pgoff_t uninitialized_var(last_index);
1378         bool force_submit = false;
1379         unsigned int nr_bios;
1380
1381         if (unlikely(owned_head == Z_EROFS_VLE_WORKGRP_TAIL))
1382                 return false;
1383
1384         force_submit = false;
1385         bio = NULL;
1386         nr_bios = 0;
1387         bi_private = jobqueueset_init(sb, qtail, q, fgq, force_fg);
1388
1389         /* by default, all need io submission */
1390         q[JQ_SUBMIT]->head = owned_head;
1391
1392         do {
1393                 struct z_erofs_vle_workgroup *grp;
1394                 pgoff_t first_index;
1395                 struct page *page;
1396                 unsigned int i = 0, bypass = 0;
1397                 int err;
1398
1399                 /* no possible 'owned_head' equals the following */
1400                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1401                 DBG_BUGON(owned_head == Z_EROFS_VLE_WORKGRP_NIL);
1402
1403                 grp = container_of(owned_head,
1404                                    struct z_erofs_vle_workgroup, next);
1405
1406                 /* close the main owned chain at first */
1407                 owned_head = cmpxchg(&grp->next, Z_EROFS_VLE_WORKGRP_TAIL,
1408                                      Z_EROFS_VLE_WORKGRP_TAIL_CLOSED);
1409
1410                 first_index = grp->obj.index;
1411                 force_submit |= (first_index != last_index + 1);
1412
1413 repeat:
1414                 page = pickup_page_for_submission(grp, i, pagepool,
1415                                                   MNGD_MAPPING(sbi), gfp);
1416                 if (!page) {
1417                         force_submit = true;
1418                         ++bypass;
1419                         goto skippage;
1420                 }
1421
1422                 if (bio && force_submit) {
1423 submit_bio_retry:
1424                         __submit_bio(bio, REQ_OP_READ, 0);
1425                         bio = NULL;
1426                 }
1427
1428                 if (!bio) {
1429                         bio = erofs_grab_bio(sb, first_index + i,
1430                                              BIO_MAX_PAGES,
1431                                              z_erofs_vle_read_endio, true);
1432                         bio->bi_private = bi_private;
1433
1434                         ++nr_bios;
1435                 }
1436
1437                 err = bio_add_page(bio, page, PAGE_SIZE, 0);
1438                 if (err < PAGE_SIZE)
1439                         goto submit_bio_retry;
1440
1441                 force_submit = false;
1442                 last_index = first_index + i;
1443 skippage:
1444                 if (++i < clusterpages)
1445                         goto repeat;
1446
1447                 if (bypass < clusterpages)
1448                         qtail[JQ_SUBMIT] = &grp->next;
1449                 else
1450                         move_to_bypass_jobqueue(grp, qtail, owned_head);
1451         } while (owned_head != Z_EROFS_VLE_WORKGRP_TAIL);
1452
1453         if (bio)
1454                 __submit_bio(bio, REQ_OP_READ, 0);
1455
1456         if (postsubmit_is_all_bypassed(q, nr_bios, force_fg))
1457                 return true;
1458
1459         z_erofs_vle_unzip_kickoff(bi_private, nr_bios);
1460         return true;
1461 }
1462
1463 static void z_erofs_submit_and_unzip(struct z_erofs_vle_frontend *f,
1464                                      struct list_head *pagepool,
1465                                      bool force_fg)
1466 {
1467         struct super_block *sb = f->inode->i_sb;
1468         struct z_erofs_vle_unzip_io io[NR_JOBQUEUES];
1469
1470         if (!z_erofs_vle_submit_all(sb, f->owned_head, pagepool, io, force_fg))
1471                 return;
1472
1473 #ifdef EROFS_FS_HAS_MANAGED_CACHE
1474         z_erofs_vle_unzip_all(sb, &io[JQ_BYPASS], pagepool);
1475 #endif
1476         if (!force_fg)
1477                 return;
1478
1479         /* wait until all bios are completed */
1480         wait_event(io[JQ_SUBMIT].u.wait,
1481                    !atomic_read(&io[JQ_SUBMIT].pending_bios));
1482
1483         /* let's synchronous decompression */
1484         z_erofs_vle_unzip_all(sb, &io[JQ_SUBMIT], pagepool);
1485 }
1486
1487 static int z_erofs_vle_normalaccess_readpage(struct file *file,
1488                                              struct page *page)
1489 {
1490         struct inode *const inode = page->mapping->host;
1491         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1492         int err;
1493         LIST_HEAD(pagepool);
1494
1495         trace_erofs_readpage(page, false);
1496
1497         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1498
1499         err = z_erofs_do_read_page(&f, page, &pagepool);
1500         (void)z_erofs_vle_work_iter_end(&f.builder);
1501
1502         if (err) {
1503                 errln("%s, failed to read, err [%d]", __func__, err);
1504                 goto out;
1505         }
1506
1507         z_erofs_submit_and_unzip(&f, &pagepool, true);
1508 out:
1509         if (f.map.mpage)
1510                 put_page(f.map.mpage);
1511
1512         /* clean up the remaining free pages */
1513         put_pages_list(&pagepool);
1514         return 0;
1515 }
1516
1517 static int z_erofs_vle_normalaccess_readpages(struct file *filp,
1518                                               struct address_space *mapping,
1519                                               struct list_head *pages,
1520                                               unsigned int nr_pages)
1521 {
1522         struct inode *const inode = mapping->host;
1523         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1524
1525         bool sync = __should_decompress_synchronously(sbi, nr_pages);
1526         struct z_erofs_vle_frontend f = VLE_FRONTEND_INIT(inode);
1527         gfp_t gfp = mapping_gfp_constraint(mapping, GFP_KERNEL);
1528         struct page *head = NULL;
1529         LIST_HEAD(pagepool);
1530
1531         trace_erofs_readpages(mapping->host, lru_to_page(pages),
1532                               nr_pages, false);
1533
1534         f.headoffset = (erofs_off_t)lru_to_page(pages)->index << PAGE_SHIFT;
1535
1536         for (; nr_pages; --nr_pages) {
1537                 struct page *page = lru_to_page(pages);
1538
1539                 prefetchw(&page->flags);
1540                 list_del(&page->lru);
1541
1542                 /*
1543                  * A pure asynchronous readahead is indicated if
1544                  * a PG_readahead marked page is hitted at first.
1545                  * Let's also do asynchronous decompression for this case.
1546                  */
1547                 sync &= !(PageReadahead(page) && !head);
1548
1549                 if (add_to_page_cache_lru(page, mapping, page->index, gfp)) {
1550                         list_add(&page->lru, &pagepool);
1551                         continue;
1552                 }
1553
1554                 set_page_private(page, (unsigned long)head);
1555                 head = page;
1556         }
1557
1558         while (head) {
1559                 struct page *page = head;
1560                 int err;
1561
1562                 /* traversal in reverse order */
1563                 head = (void *)page_private(page);
1564
1565                 err = z_erofs_do_read_page(&f, page, &pagepool);
1566                 if (err) {
1567                         struct erofs_vnode *vi = EROFS_V(inode);
1568
1569                         errln("%s, readahead error at page %lu of nid %llu",
1570                                 __func__, page->index, vi->nid);
1571                 }
1572
1573                 put_page(page);
1574         }
1575
1576         (void)z_erofs_vle_work_iter_end(&f.builder);
1577
1578         z_erofs_submit_and_unzip(&f, &pagepool, sync);
1579
1580         if (f.map.mpage)
1581                 put_page(f.map.mpage);
1582
1583         /* clean up the remaining free pages */
1584         put_pages_list(&pagepool);
1585         return 0;
1586 }
1587
1588 const struct address_space_operations z_erofs_vle_normalaccess_aops = {
1589         .readpage = z_erofs_vle_normalaccess_readpage,
1590         .readpages = z_erofs_vle_normalaccess_readpages,
1591 };
1592
1593 /*
1594  * Variable-sized Logical Extent (Fixed Physical Cluster) Compression Mode
1595  * ---
1596  * VLE compression mode attempts to compress a number of logical data into
1597  * a physical cluster with a fixed size.
1598  * VLE compression mode uses "struct z_erofs_vle_decompressed_index".
1599  */
1600 #define __vle_cluster_advise(x, bit, bits) \
1601         ((le16_to_cpu(x) >> (bit)) & ((1 << (bits)) - 1))
1602
1603 #define __vle_cluster_type(advise) __vle_cluster_advise(advise, \
1604         Z_EROFS_VLE_DI_CLUSTER_TYPE_BIT, Z_EROFS_VLE_DI_CLUSTER_TYPE_BITS)
1605
1606 #define vle_cluster_type(di)    \
1607         __vle_cluster_type((di)->di_advise)
1608
1609 static int
1610 vle_decompressed_index_clusterofs(unsigned int *clusterofs,
1611                                   unsigned int clustersize,
1612                                   struct z_erofs_vle_decompressed_index *di)
1613 {
1614         switch (vle_cluster_type(di)) {
1615         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1616                 *clusterofs = clustersize;
1617                 break;
1618         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1619         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1620                 *clusterofs = le16_to_cpu(di->di_clusterofs);
1621                 break;
1622         default:
1623                 DBG_BUGON(1);
1624                 return -EIO;
1625         }
1626         return 0;
1627 }
1628
1629 static inline erofs_blk_t
1630 vle_extent_blkaddr(struct inode *inode, pgoff_t index)
1631 {
1632         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1633         struct erofs_vnode *vi = EROFS_V(inode);
1634
1635         unsigned int ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1636                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1637                 index * sizeof(struct z_erofs_vle_decompressed_index);
1638
1639         return erofs_blknr(iloc(sbi, vi->nid) + ofs);
1640 }
1641
1642 static inline unsigned int
1643 vle_extent_blkoff(struct inode *inode, pgoff_t index)
1644 {
1645         struct erofs_sb_info *sbi = EROFS_I_SB(inode);
1646         struct erofs_vnode *vi = EROFS_V(inode);
1647
1648         unsigned int ofs = Z_EROFS_VLE_EXTENT_ALIGN(vi->inode_isize +
1649                 vi->xattr_isize) + sizeof(struct erofs_extent_header) +
1650                 index * sizeof(struct z_erofs_vle_decompressed_index);
1651
1652         return erofs_blkoff(iloc(sbi, vi->nid) + ofs);
1653 }
1654
1655 struct vle_map_blocks_iter_ctx {
1656         struct inode *inode;
1657         struct super_block *sb;
1658         unsigned int clusterbits;
1659
1660         struct page **mpage_ret;
1661         void **kaddr_ret;
1662 };
1663
1664 static int
1665 vle_get_logical_extent_head(const struct vle_map_blocks_iter_ctx *ctx,
1666                             unsigned int lcn,   /* logical cluster number */
1667                             unsigned long long *ofs,
1668                             erofs_blk_t *pblk,
1669                             unsigned int *flags)
1670 {
1671         const unsigned int clustersize = 1 << ctx->clusterbits;
1672         const erofs_blk_t mblk = vle_extent_blkaddr(ctx->inode, lcn);
1673         struct page *mpage = *ctx->mpage_ret;   /* extent metapage */
1674
1675         struct z_erofs_vle_decompressed_index *di;
1676         unsigned int cluster_type, delta0;
1677
1678         if (mpage->index != mblk) {
1679                 kunmap_atomic(*ctx->kaddr_ret);
1680                 unlock_page(mpage);
1681                 put_page(mpage);
1682
1683                 mpage = erofs_get_meta_page(ctx->sb, mblk, false);
1684                 if (IS_ERR(mpage)) {
1685                         *ctx->mpage_ret = NULL;
1686                         return PTR_ERR(mpage);
1687                 }
1688                 *ctx->mpage_ret = mpage;
1689                 *ctx->kaddr_ret = kmap_atomic(mpage);
1690         }
1691
1692         di = *ctx->kaddr_ret + vle_extent_blkoff(ctx->inode, lcn);
1693
1694         cluster_type = vle_cluster_type(di);
1695         switch (cluster_type) {
1696         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1697                 delta0 = le16_to_cpu(di->di_u.delta[0]);
1698                 if (unlikely(!delta0 || delta0 > lcn)) {
1699                         errln("invalid NONHEAD dl0 %u at lcn %u of nid %llu",
1700                               delta0, lcn, EROFS_V(ctx->inode)->nid);
1701                         DBG_BUGON(1);
1702                         return -EIO;
1703                 }
1704                 return vle_get_logical_extent_head(ctx,
1705                         lcn - delta0, ofs, pblk, flags);
1706         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1707                 *flags ^= EROFS_MAP_ZIPPED;
1708                 /* fallthrough */
1709         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1710                 /* clustersize should be a power of two */
1711                 *ofs = ((u64)lcn << ctx->clusterbits) +
1712                         (le16_to_cpu(di->di_clusterofs) & (clustersize - 1));
1713                 *pblk = le32_to_cpu(di->di_u.blkaddr);
1714                 break;
1715         default:
1716                 errln("unknown cluster type %u at lcn %u of nid %llu",
1717                       cluster_type, lcn, EROFS_V(ctx->inode)->nid);
1718                 DBG_BUGON(1);
1719                 return -EIO;
1720         }
1721         return 0;
1722 }
1723
1724 int z_erofs_map_blocks_iter(struct inode *inode,
1725         struct erofs_map_blocks *map,
1726         int flags)
1727 {
1728         void *kaddr;
1729         const struct vle_map_blocks_iter_ctx ctx = {
1730                 .inode = inode,
1731                 .sb = inode->i_sb,
1732                 .clusterbits = EROFS_I_SB(inode)->clusterbits,
1733                 .mpage_ret = &map->mpage,
1734                 .kaddr_ret = &kaddr
1735         };
1736         const unsigned int clustersize = 1 << ctx.clusterbits;
1737         /* if both m_(l,p)len are 0, regularize l_lblk, l_lofs, etc... */
1738         const bool initial = !map->m_llen;
1739
1740         /* logicial extent (start, end) offset */
1741         unsigned long long ofs, end;
1742         unsigned int lcn;
1743         u32 ofs_rem;
1744
1745         /* initialize `pblk' to keep gcc from printing foolish warnings */
1746         erofs_blk_t mblk, pblk = 0;
1747         struct page *mpage = map->mpage;
1748         struct z_erofs_vle_decompressed_index *di;
1749         unsigned int cluster_type, logical_cluster_ofs;
1750         int err = 0;
1751
1752         trace_z_erofs_map_blocks_iter_enter(inode, map, flags);
1753
1754         /* when trying to read beyond EOF, leave it unmapped */
1755         if (unlikely(map->m_la >= inode->i_size)) {
1756                 DBG_BUGON(!initial);
1757                 map->m_llen = map->m_la + 1 - inode->i_size;
1758                 map->m_la = inode->i_size;
1759                 map->m_flags = 0;
1760                 goto out;
1761         }
1762
1763         debugln("%s, m_la %llu m_llen %llu --- start", __func__,
1764                 map->m_la, map->m_llen);
1765
1766         ofs = map->m_la + map->m_llen;
1767
1768         /* clustersize should be power of two */
1769         lcn = ofs >> ctx.clusterbits;
1770         ofs_rem = ofs & (clustersize - 1);
1771
1772         mblk = vle_extent_blkaddr(inode, lcn);
1773
1774         if (!mpage || mpage->index != mblk) {
1775                 if (mpage)
1776                         put_page(mpage);
1777
1778                 mpage = erofs_get_meta_page(ctx.sb, mblk, false);
1779                 if (IS_ERR(mpage)) {
1780                         err = PTR_ERR(mpage);
1781                         goto out;
1782                 }
1783                 map->mpage = mpage;
1784         } else {
1785                 lock_page(mpage);
1786                 DBG_BUGON(!PageUptodate(mpage));
1787         }
1788
1789         kaddr = kmap_atomic(mpage);
1790         di = kaddr + vle_extent_blkoff(inode, lcn);
1791
1792         debugln("%s, lcn %u mblk %u e_blkoff %u", __func__, lcn,
1793                 mblk, vle_extent_blkoff(inode, lcn));
1794
1795         err = vle_decompressed_index_clusterofs(&logical_cluster_ofs,
1796                                                 clustersize, di);
1797         if (unlikely(err))
1798                 goto unmap_out;
1799
1800         if (!initial) {
1801                 /* [walking mode] 'map' has been already initialized */
1802                 map->m_llen += logical_cluster_ofs;
1803                 goto unmap_out;
1804         }
1805
1806         /* by default, compressed */
1807         map->m_flags |= EROFS_MAP_ZIPPED;
1808
1809         end = ((u64)lcn + 1) * clustersize;
1810
1811         cluster_type = vle_cluster_type(di);
1812
1813         switch (cluster_type) {
1814         case Z_EROFS_VLE_CLUSTER_TYPE_PLAIN:
1815                 if (ofs_rem >= logical_cluster_ofs)
1816                         map->m_flags ^= EROFS_MAP_ZIPPED;
1817                 /* fallthrough */
1818         case Z_EROFS_VLE_CLUSTER_TYPE_HEAD:
1819                 if (ofs_rem == logical_cluster_ofs) {
1820                         pblk = le32_to_cpu(di->di_u.blkaddr);
1821                         goto exact_hitted;
1822                 }
1823
1824                 if (ofs_rem > logical_cluster_ofs) {
1825                         ofs = (u64)lcn * clustersize | logical_cluster_ofs;
1826                         pblk = le32_to_cpu(di->di_u.blkaddr);
1827                         break;
1828                 }
1829
1830                 /* logical cluster number should be >= 1 */
1831                 if (unlikely(!lcn)) {
1832                         errln("invalid logical cluster 0 at nid %llu",
1833                                 EROFS_V(inode)->nid);
1834                         err = -EIO;
1835                         goto unmap_out;
1836                 }
1837                 end = ((u64)lcn-- * clustersize) | logical_cluster_ofs;
1838                 /* fallthrough */
1839         case Z_EROFS_VLE_CLUSTER_TYPE_NONHEAD:
1840                 /* get the correspoinding first chunk */
1841                 err = vle_get_logical_extent_head(&ctx, lcn, &ofs,
1842                                                   &pblk, &map->m_flags);
1843                 mpage = map->mpage;
1844
1845                 if (unlikely(err)) {
1846                         if (mpage)
1847                                 goto unmap_out;
1848                         goto out;
1849                 }
1850                 break;
1851         default:
1852                 errln("unknown cluster type %u at offset %llu of nid %llu",
1853                         cluster_type, ofs, EROFS_V(inode)->nid);
1854                 err = -EIO;
1855                 goto unmap_out;
1856         }
1857
1858         map->m_la = ofs;
1859 exact_hitted:
1860         map->m_llen = end - ofs;
1861         map->m_plen = clustersize;
1862         map->m_pa = blknr_to_addr(pblk);
1863         map->m_flags |= EROFS_MAP_MAPPED;
1864 unmap_out:
1865         kunmap_atomic(kaddr);
1866         unlock_page(mpage);
1867 out:
1868         debugln("%s, m_la %llu m_pa %llu m_llen %llu m_plen %llu m_flags 0%o",
1869                 __func__, map->m_la, map->m_pa,
1870                 map->m_llen, map->m_plen, map->m_flags);
1871
1872         trace_z_erofs_map_blocks_iter_exit(inode, map, flags, err);
1873
1874         /* aggressively BUG_ON iff CONFIG_EROFS_FS_DEBUG is on */
1875         DBG_BUGON(err < 0 && err != -ENOMEM);
1876         return err;
1877 }
1878