fs/erofs/zdata.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2018 HUAWEI, Inc.
   4  *             https://www.huawei.com/
   5  * Copyright (C) 2022 Alibaba Cloud
   6  */
   7 #include "zdata.h"
   8 #include "compress.h"
   9 #include <linux/prefetch.h>
  10
  11 #include <trace/events/erofs.h>
  12
  13 /*
  14  * since pclustersize is variable for big pcluster feature, introduce slab
  15  * pools implementation for different pcluster sizes.
  16  */
  17 struct z_erofs_pcluster_slab {
  18         struct kmem_cache *slab;
  19         unsigned int maxpages;
  20         char name[48];
  21 };
  22
  23 #define _PCLP(n) { .maxpages = n }
  24
  25 static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = {
  26         _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128),
  27         _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES)
  28 };
  29
  30 struct z_erofs_bvec_iter {
  31         struct page *bvpage;
  32         struct z_erofs_bvset *bvset;
  33         unsigned int nr, cur;
  34 };
  35
  36 static struct page *z_erofs_bvec_iter_end(struct z_erofs_bvec_iter *iter)
  37 {
  38         if (iter->bvpage)
  39                 kunmap_local(iter->bvset);
  40         return iter->bvpage;
  41 }
  42
  43 static struct page *z_erofs_bvset_flip(struct z_erofs_bvec_iter *iter)
  44 {
  45         unsigned long base = (unsigned long)((struct z_erofs_bvset *)0)->bvec;
  46         /* have to access nextpage in advance, otherwise it will be unmapped */
  47         struct page *nextpage = iter->bvset->nextpage;
  48         struct page *oldpage;
  49
  50         DBG_BUGON(!nextpage);
  51         oldpage = z_erofs_bvec_iter_end(iter);
  52         iter->bvpage = nextpage;
  53         iter->bvset = kmap_local_page(nextpage);
  54         iter->nr = (PAGE_SIZE - base) / sizeof(struct z_erofs_bvec);
  55         iter->cur = 0;
  56         return oldpage;
  57 }
  58
  59 static void z_erofs_bvec_iter_begin(struct z_erofs_bvec_iter *iter,
  60                                     struct z_erofs_bvset_inline *bvset,
  61                                     unsigned int bootstrap_nr,
  62                                     unsigned int cur)
  63 {
  64         *iter = (struct z_erofs_bvec_iter) {
  65                 .nr = bootstrap_nr,
  66                 .bvset = (struct z_erofs_bvset *)bvset,
  67         };
  68
  69         while (cur > iter->nr) {
  70                 cur -= iter->nr;
  71                 z_erofs_bvset_flip(iter);
  72         }
  73         iter->cur = cur;
  74 }
  75
  76 static int z_erofs_bvec_enqueue(struct z_erofs_bvec_iter *iter,
  77                                 struct z_erofs_bvec *bvec,
  78                                 struct page **candidate_bvpage)
  79 {
  80         if (iter->cur == iter->nr) {
  81                 if (!*candidate_bvpage)
  82                         return -EAGAIN;
  83
  84                 DBG_BUGON(iter->bvset->nextpage);
  85                 iter->bvset->nextpage = *candidate_bvpage;
  86                 z_erofs_bvset_flip(iter);
  87
  88                 iter->bvset->nextpage = NULL;
  89                 *candidate_bvpage = NULL;
  90         }
  91         iter->bvset->bvec[iter->cur++] = *bvec;
  92         return 0;
  93 }
  94
  95 static void z_erofs_bvec_dequeue(struct z_erofs_bvec_iter *iter,
  96                                  struct z_erofs_bvec *bvec,
  97                                  struct page **old_bvpage)
  98 {
  99         if (iter->cur == iter->nr)
 100                 *old_bvpage = z_erofs_bvset_flip(iter);
 101         else
 102                 *old_bvpage = NULL;
 103         *bvec = iter->bvset->bvec[iter->cur++];
 104 }
 105
 106 static void z_erofs_destroy_pcluster_pool(void)
 107 {
 108         int i;
 109
 110         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 111                 if (!pcluster_pool[i].slab)
 112                         continue;
 113                 kmem_cache_destroy(pcluster_pool[i].slab);
 114                 pcluster_pool[i].slab = NULL;
 115         }
 116 }
 117
 118 static int z_erofs_create_pcluster_pool(void)
 119 {
 120         struct z_erofs_pcluster_slab *pcs;
 121         struct z_erofs_pcluster *a;
 122         unsigned int size;
 123
 124         for (pcs = pcluster_pool;
 125              pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) {
 126                 size = struct_size(a, compressed_bvecs, pcs->maxpages);
 127
 128                 sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages);
 129                 pcs->slab = kmem_cache_create(pcs->name, size, 0,
 130                                               SLAB_RECLAIM_ACCOUNT, NULL);
 131                 if (pcs->slab)
 132                         continue;
 133
 134                 z_erofs_destroy_pcluster_pool();
 135                 return -ENOMEM;
 136         }
 137         return 0;
 138 }
 139
 140 static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages)
 141 {
 142         int i;
 143
 144         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 145                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 146                 struct z_erofs_pcluster *pcl;
 147
 148                 if (nrpages > pcs->maxpages)
 149                         continue;
 150
 151                 pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS);
 152                 if (!pcl)
 153                         return ERR_PTR(-ENOMEM);
 154                 pcl->pclusterpages = nrpages;
 155                 return pcl;
 156         }
 157         return ERR_PTR(-EINVAL);
 158 }
 159
 160 static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl)
 161 {
 162         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 163         int i;
 164
 165         for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) {
 166                 struct z_erofs_pcluster_slab *pcs = pcluster_pool + i;
 167
 168                 if (pclusterpages > pcs->maxpages)
 169                         continue;
 170
 171                 kmem_cache_free(pcs->slab, pcl);
 172                 return;
 173         }
 174         DBG_BUGON(1);
 175 }
 176
 177 /* how to allocate cached pages for a pcluster */
 178 enum z_erofs_cache_alloctype {
 179         DONTALLOC,      /* don't allocate any cached pages */
 180         /*
 181          * try to use cached I/O if page allocation succeeds or fallback
 182          * to in-place I/O instead to avoid any direct reclaim.
 183          */
 184         TRYALLOC,
 185 };
 186
 187 /*
 188  * tagged pointer with 1-bit tag for all compressed pages
 189  * tag 0 - the page is just found with an extra page reference
 190  */
 191 typedef tagptr1_t compressed_page_t;
 192
 193 #define tag_compressed_page_justfound(page) \
 194         tagptr_fold(compressed_page_t, page, 1)
 195
 196 static struct workqueue_struct *z_erofs_workqueue __read_mostly;
 197
 198 void z_erofs_exit_zip_subsystem(void)
 199 {
 200         destroy_workqueue(z_erofs_workqueue);
 201         z_erofs_destroy_pcluster_pool();
 202 }
 203
 204 static inline int z_erofs_init_workqueue(void)
 205 {
 206         const unsigned int onlinecpus = num_possible_cpus();
 207
 208         /*
 209          * no need to spawn too many threads, limiting threads could minimum
 210          * scheduling overhead, perhaps per-CPU threads should be better?
 211          */
 212         z_erofs_workqueue = alloc_workqueue("erofs_unzipd",
 213                                             WQ_UNBOUND | WQ_HIGHPRI,
 214                                             onlinecpus + onlinecpus / 4);
 215         return z_erofs_workqueue ? 0 : -ENOMEM;
 216 }
 217
 218 int __init z_erofs_init_zip_subsystem(void)
 219 {
 220         int err = z_erofs_create_pcluster_pool();
 221
 222         if (err)
 223                 return err;
 224         err = z_erofs_init_workqueue();
 225         if (err)
 226                 z_erofs_destroy_pcluster_pool();
 227         return err;
 228 }
 229
 230 enum z_erofs_pclustermode {
 231         Z_EROFS_PCLUSTER_INFLIGHT,
 232         /*
 233          * The current pclusters was the tail of an exist chain, in addition
 234          * that the previous processed chained pclusters are all decided to
 235          * be hooked up to it.
 236          * A new chain will be created for the remaining pclusters which are
 237          * not processed yet, so different from Z_EROFS_PCLUSTER_FOLLOWED,
 238          * the next pcluster cannot reuse the whole page safely for inplace I/O
 239          * in the following scenario:
 240          *  ________________________________________________________________
 241          * |      tail (partial) page     |       head (partial) page       |
 242          * |   (belongs to the next pcl)  |   (belongs to the current pcl)  |
 243          * |_______PCLUSTER_FOLLOWED______|________PCLUSTER_HOOKED__________|
 244          */
 245         Z_EROFS_PCLUSTER_HOOKED,
 246         /*
 247          * a weak form of Z_EROFS_PCLUSTER_FOLLOWED, the difference is that it
 248          * could be dispatched into bypass queue later due to uptodated managed
 249          * pages. All related online pages cannot be reused for inplace I/O (or
 250          * bvpage) since it can be directly decoded without I/O submission.
 251          */
 252         Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE,
 253         /*
 254          * The current collection has been linked with the owned chain, and
 255          * could also be linked with the remaining collections, which means
 256          * if the processing page is the tail page of the collection, thus
 257          * the current collection can safely use the whole page (since
 258          * the previous collection is under control) for in-place I/O, as
 259          * illustrated below:
 260          *  ________________________________________________________________
 261          * |  tail (partial) page |          head (partial) page           |
 262          * |  (of the current cl) |      (of the previous collection)      |
 263          * | PCLUSTER_FOLLOWED or |                                        |
 264          * |_____PCLUSTER_HOOKED__|___________PCLUSTER_FOLLOWED____________|
 265          *
 266          * [  (*) the above page can be used as inplace I/O.               ]
 267          */
 268         Z_EROFS_PCLUSTER_FOLLOWED,
 269 };
 270
 271 struct z_erofs_decompress_frontend {
 272         struct inode *const inode;
 273         struct erofs_map_blocks map;
 274         struct z_erofs_bvec_iter biter;
 275
 276         struct page *candidate_bvpage;
 277         struct z_erofs_pcluster *pcl, *tailpcl;
 278         z_erofs_next_pcluster_t owned_head;
 279         enum z_erofs_pclustermode mode;
 280
 281         bool readahead;
 282         /* used for applying cache strategy on the fly */
 283         bool backmost;
 284         erofs_off_t headoffset;
 285
 286         /* a pointer used to pick up inplace I/O pages */
 287         unsigned int icur;
 288 };
 289
 290 #define DECOMPRESS_FRONTEND_INIT(__i) { \
 291         .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \
 292         .mode = Z_EROFS_PCLUSTER_FOLLOWED, .backmost = true }
 293
 294 static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe,
 295                                enum z_erofs_cache_alloctype type,
 296                                struct page **pagepool)
 297 {
 298         struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode));
 299         struct z_erofs_pcluster *pcl = fe->pcl;
 300         bool standalone = true;
 301         /*
 302          * optimistic allocation without direct reclaim since inplace I/O
 303          * can be used if low memory otherwise.
 304          */
 305         gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) |
 306                         __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN;
 307         unsigned int i;
 308
 309         if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED)
 310                 return;
 311
 312         for (i = 0; i < pcl->pclusterpages; ++i) {
 313                 struct page *page;
 314                 compressed_page_t t;
 315                 struct page *newpage = NULL;
 316
 317                 /* the compressed page was loaded before */
 318                 if (READ_ONCE(pcl->compressed_bvecs[i].page))
 319                         continue;
 320
 321                 page = find_get_page(mc, pcl->obj.index + i);
 322
 323                 if (page) {
 324                         t = tag_compressed_page_justfound(page);
 325                 } else {
 326                         /* I/O is needed, no possible to decompress directly */
 327                         standalone = false;
 328                         switch (type) {
 329                         case TRYALLOC:
 330                                 newpage = erofs_allocpage(pagepool, gfp);
 331                                 if (!newpage)
 332                                         continue;
 333                                 set_page_private(newpage,
 334                                                  Z_EROFS_PREALLOCATED_PAGE);
 335                                 t = tag_compressed_page_justfound(newpage);
 336                                 break;
 337                         default:        /* DONTALLOC */
 338                                 continue;
 339                         }
 340                 }
 341
 342                 if (!cmpxchg_relaxed(&pcl->compressed_bvecs[i].page, NULL,
 343                                      tagptr_cast_ptr(t)))
 344                         continue;
 345
 346                 if (page)
 347                         put_page(page);
 348                 else if (newpage)
 349                         erofs_pagepool_add(pagepool, newpage);
 350         }
 351
 352         /*
 353          * don't do inplace I/O if all compressed pages are available in
 354          * managed cache since it can be moved to the bypass queue instead.
 355          */
 356         if (standalone)
 357                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
 358 }
 359
 360 /* called by erofs_shrinker to get rid of all compressed_pages */
 361 int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi,
 362                                        struct erofs_workgroup *grp)
 363 {
 364         struct z_erofs_pcluster *const pcl =
 365                 container_of(grp, struct z_erofs_pcluster, obj);
 366         int i;
 367
 368         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 369         /*
 370          * refcount of workgroup is now freezed as 1,
 371          * therefore no need to worry about available decompression users.
 372          */
 373         for (i = 0; i < pcl->pclusterpages; ++i) {
 374                 struct page *page = pcl->compressed_bvecs[i].page;
 375
 376                 if (!page)
 377                         continue;
 378
 379                 /* block other users from reclaiming or migrating the page */
 380                 if (!trylock_page(page))
 381                         return -EBUSY;
 382
 383                 if (!erofs_page_is_managed(sbi, page))
 384                         continue;
 385
 386                 /* barrier is implied in the following 'unlock_page' */
 387                 WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 388                 detach_page_private(page);
 389                 unlock_page(page);
 390         }
 391         return 0;
 392 }
 393
 394 int erofs_try_to_free_cached_page(struct page *page)
 395 {
 396         struct z_erofs_pcluster *const pcl = (void *)page_private(page);
 397         int ret, i;
 398
 399         if (!erofs_workgroup_try_to_freeze(&pcl->obj, 1))
 400                 return 0;
 401
 402         ret = 0;
 403         DBG_BUGON(z_erofs_is_inline_pcluster(pcl));
 404         for (i = 0; i < pcl->pclusterpages; ++i) {
 405                 if (pcl->compressed_bvecs[i].page == page) {
 406                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
 407                         ret = 1;
 408                         break;
 409                 }
 410         }
 411         erofs_workgroup_unfreeze(&pcl->obj, 1);
 412         if (ret)
 413                 detach_page_private(page);
 414         return ret;
 415 }
 416
 417 static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe,
 418                                    struct z_erofs_bvec *bvec)
 419 {
 420         struct z_erofs_pcluster *const pcl = fe->pcl;
 421
 422         while (fe->icur > 0) {
 423                 if (!cmpxchg(&pcl->compressed_bvecs[--fe->icur].page,
 424                              NULL, bvec->page)) {
 425                         pcl->compressed_bvecs[fe->icur] = *bvec;
 426                         return true;
 427                 }
 428         }
 429         return false;
 430 }
 431
 432 /* callers must be with pcluster lock held */
 433 static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe,
 434                                struct z_erofs_bvec *bvec, bool exclusive)
 435 {
 436         int ret;
 437
 438         if (exclusive) {
 439                 /* give priority for inplaceio to use file pages first */
 440                 if (z_erofs_try_inplace_io(fe, bvec))
 441                         return 0;
 442                 /* otherwise, check if it can be used as a bvpage */
 443                 if (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED &&
 444                     !fe->candidate_bvpage)
 445                         fe->candidate_bvpage = bvec->page;
 446         }
 447         ret = z_erofs_bvec_enqueue(&fe->biter, bvec, &fe->candidate_bvpage);
 448         fe->pcl->vcnt += (ret >= 0);
 449         return ret;
 450 }
 451
 452 static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f)
 453 {
 454         struct z_erofs_pcluster *pcl = f->pcl;
 455         z_erofs_next_pcluster_t *owned_head = &f->owned_head;
 456
 457         /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */
 458         if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL,
 459                     *owned_head) == Z_EROFS_PCLUSTER_NIL) {
 460                 *owned_head = &pcl->next;
 461                 /* so we can attach this pcluster to our submission chain. */
 462                 f->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 463                 return;
 464         }
 465
 466         /*
 467          * type 2, link to the end of an existing open chain, be careful
 468          * that its submission is controlled by the original attached chain.
 469          */
 470         if (*owned_head != &pcl->next && pcl != f->tailpcl &&
 471             cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
 472                     *owned_head) == Z_EROFS_PCLUSTER_TAIL) {
 473                 *owned_head = Z_EROFS_PCLUSTER_TAIL;
 474                 f->mode = Z_EROFS_PCLUSTER_HOOKED;
 475                 f->tailpcl = NULL;
 476                 return;
 477         }
 478         /* type 3, it belongs to a chain, but it isn't the end of the chain */
 479         f->mode = Z_EROFS_PCLUSTER_INFLIGHT;
 480 }
 481
 482 static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe)
 483 {
 484         struct erofs_map_blocks *map = &fe->map;
 485         bool ztailpacking = map->m_flags & EROFS_MAP_META;
 486         struct z_erofs_pcluster *pcl;
 487         struct erofs_workgroup *grp;
 488         int err;
 489
 490         if (!(map->m_flags & EROFS_MAP_ENCODED)) {
 491                 DBG_BUGON(1);
 492                 return -EFSCORRUPTED;
 493         }
 494
 495         /* no available pcluster, let's allocate one */
 496         pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 :
 497                                      map->m_plen >> PAGE_SHIFT);
 498         if (IS_ERR(pcl))
 499                 return PTR_ERR(pcl);
 500
 501         atomic_set(&pcl->obj.refcount, 1);
 502         pcl->algorithmformat = map->m_algorithmformat;
 503         pcl->length = 0;
 504         pcl->partial = true;
 505
 506         /* new pclusters should be claimed as type 1, primary and followed */
 507         pcl->next = fe->owned_head;
 508         pcl->pageofs_out = map->m_la & ~PAGE_MASK;
 509         fe->mode = Z_EROFS_PCLUSTER_FOLLOWED;
 510
 511         /*
 512          * lock all primary followed works before visible to others
 513          * and mutex_trylock *never* fails for a new pcluster.
 514          */
 515         mutex_init(&pcl->lock);
 516         DBG_BUGON(!mutex_trylock(&pcl->lock));
 517
 518         if (ztailpacking) {
 519                 pcl->obj.index = 0;     /* which indicates ztailpacking */
 520                 pcl->pageofs_in = erofs_blkoff(map->m_pa);
 521                 pcl->tailpacking_size = map->m_plen;
 522         } else {
 523                 pcl->obj.index = map->m_pa >> PAGE_SHIFT;
 524
 525                 grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj);
 526                 if (IS_ERR(grp)) {
 527                         err = PTR_ERR(grp);
 528                         goto err_out;
 529                 }
 530
 531                 if (grp != &pcl->obj) {
 532                         fe->pcl = container_of(grp,
 533                                         struct z_erofs_pcluster, obj);
 534                         err = -EEXIST;
 535                         goto err_out;
 536                 }
 537         }
 538         /* used to check tail merging loop due to corrupted images */
 539         if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
 540                 fe->tailpcl = pcl;
 541         fe->owned_head = &pcl->next;
 542         fe->pcl = pcl;
 543         return 0;
 544
 545 err_out:
 546         mutex_unlock(&pcl->lock);
 547         z_erofs_free_pcluster(pcl);
 548         return err;
 549 }
 550
 551 static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe)
 552 {
 553         struct erofs_map_blocks *map = &fe->map;
 554         struct erofs_workgroup *grp = NULL;
 555         int ret;
 556
 557         DBG_BUGON(fe->pcl);
 558
 559         /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */
 560         DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL);
 561         DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
 562
 563         if (!(map->m_flags & EROFS_MAP_META)) {
 564                 grp = erofs_find_workgroup(fe->inode->i_sb,
 565                                            map->m_pa >> PAGE_SHIFT);
 566         } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) {
 567                 DBG_BUGON(1);
 568                 return -EFSCORRUPTED;
 569         }
 570
 571         if (grp) {
 572                 fe->pcl = container_of(grp, struct z_erofs_pcluster, obj);
 573                 ret = -EEXIST;
 574         } else {
 575                 ret = z_erofs_register_pcluster(fe);
 576         }
 577
 578         if (ret == -EEXIST) {
 579                 mutex_lock(&fe->pcl->lock);
 580                 /* used to check tail merging loop due to corrupted images */
 581                 if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL)
 582                         fe->tailpcl = fe->pcl;
 583
 584                 z_erofs_try_to_claim_pcluster(fe);
 585         } else if (ret) {
 586                 return ret;
 587         }
 588         z_erofs_bvec_iter_begin(&fe->biter, &fe->pcl->bvset,
 589                                 Z_EROFS_INLINE_BVECS, fe->pcl->vcnt);
 590         /* since file-backed online pages are traversed in reverse order */
 591         fe->icur = z_erofs_pclusterpages(fe->pcl);
 592         return 0;
 593 }
 594
 595 /*
 596  * keep in mind that no referenced pclusters will be freed
 597  * only after a RCU grace period.
 598  */
 599 static void z_erofs_rcu_callback(struct rcu_head *head)
 600 {
 601         z_erofs_free_pcluster(container_of(head,
 602                         struct z_erofs_pcluster, rcu));
 603 }
 604
 605 void erofs_workgroup_free_rcu(struct erofs_workgroup *grp)
 606 {
 607         struct z_erofs_pcluster *const pcl =
 608                 container_of(grp, struct z_erofs_pcluster, obj);
 609
 610         call_rcu(&pcl->rcu, z_erofs_rcu_callback);
 611 }
 612
 613 static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe)
 614 {
 615         struct z_erofs_pcluster *pcl = fe->pcl;
 616
 617         if (!pcl)
 618                 return false;
 619
 620         z_erofs_bvec_iter_end(&fe->biter);
 621         mutex_unlock(&pcl->lock);
 622
 623         if (fe->candidate_bvpage) {
 624                 DBG_BUGON(z_erofs_is_shortlived_page(fe->candidate_bvpage));
 625                 fe->candidate_bvpage = NULL;
 626         }
 627
 628         /*
 629          * if all pending pages are added, don't hold its reference
 630          * any longer if the pcluster isn't hosted by ourselves.
 631          */
 632         if (fe->mode < Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE)
 633                 erofs_workgroup_put(&pcl->obj);
 634
 635         fe->pcl = NULL;
 636         return true;
 637 }
 638
 639 static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe,
 640                                        unsigned int cachestrategy,
 641                                        erofs_off_t la)
 642 {
 643         if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED)
 644                 return false;
 645
 646         if (fe->backmost)
 647                 return true;
 648
 649         return cachestrategy >= EROFS_ZIP_CACHE_READAROUND &&
 650                 la < fe->headoffset;
 651 }
 652
 653 static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe,
 654                                 struct page *page, struct page **pagepool)
 655 {
 656         struct inode *const inode = fe->inode;
 657         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
 658         struct erofs_map_blocks *const map = &fe->map;
 659         const loff_t offset = page_offset(page);
 660         bool tight = true, exclusive;
 661
 662         enum z_erofs_cache_alloctype cache_strategy;
 663         unsigned int cur, end, spiltted;
 664         int err = 0;
 665
 666         /* register locked file pages as online pages in pack */
 667         z_erofs_onlinepage_init(page);
 668
 669         spiltted = 0;
 670         end = PAGE_SIZE;
 671 repeat:
 672         cur = end - 1;
 673
 674         if (offset + cur < map->m_la ||
 675             offset + cur >= map->m_la + map->m_llen) {
 676                 erofs_dbg("out-of-range map @ pos %llu", offset + cur);
 677
 678                 if (z_erofs_collector_end(fe))
 679                         fe->backmost = false;
 680                 map->m_la = offset + cur;
 681                 map->m_llen = 0;
 682                 err = z_erofs_map_blocks_iter(inode, map, 0);
 683                 if (err)
 684                         goto out;
 685         } else {
 686                 if (fe->pcl)
 687                         goto hitted;
 688                 /* didn't get a valid pcluster previously (very rare) */
 689         }
 690
 691         if (!(map->m_flags & EROFS_MAP_MAPPED))
 692                 goto hitted;
 693
 694         err = z_erofs_collector_begin(fe);
 695         if (err)
 696                 goto out;
 697
 698         if (z_erofs_is_inline_pcluster(fe->pcl)) {
 699                 void *mp;
 700
 701                 mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb,
 702                                         erofs_blknr(map->m_pa), EROFS_NO_KMAP);
 703                 if (IS_ERR(mp)) {
 704                         err = PTR_ERR(mp);
 705                         erofs_err(inode->i_sb,
 706                                   "failed to get inline page, err %d", err);
 707                         goto out;
 708                 }
 709                 get_page(fe->map.buf.page);
 710                 WRITE_ONCE(fe->pcl->compressed_bvecs[0].page,
 711                            fe->map.buf.page);
 712                 fe->mode = Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE;
 713         } else {
 714                 /* bind cache first when cached decompression is preferred */
 715                 if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy,
 716                                                map->m_la))
 717                         cache_strategy = TRYALLOC;
 718                 else
 719                         cache_strategy = DONTALLOC;
 720
 721                 z_erofs_bind_cache(fe, cache_strategy, pagepool);
 722         }
 723 hitted:
 724         /*
 725          * Ensure the current partial page belongs to this submit chain rather
 726          * than other concurrent submit chains or the noio(bypass) chain since
 727          * those chains are handled asynchronously thus the page cannot be used
 728          * for inplace I/O or bvpage (should be processed in a strict order.)
 729          */
 730         tight &= (fe->mode >= Z_EROFS_PCLUSTER_HOOKED &&
 731                   fe->mode != Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE);
 732
 733         cur = end - min_t(unsigned int, offset + end - map->m_la, end);
 734         if (!(map->m_flags & EROFS_MAP_MAPPED)) {
 735                 zero_user_segment(page, cur, end);
 736                 goto next_part;
 737         }
 738
 739         exclusive = (!cur && (!spiltted || tight));
 740         if (cur)
 741                 tight &= (fe->mode >= Z_EROFS_PCLUSTER_FOLLOWED);
 742
 743 retry:
 744         err = z_erofs_attach_page(fe, &((struct z_erofs_bvec) {
 745                                         .page = page,
 746                                         .offset = offset - map->m_la,
 747                                         .end = end,
 748                                   }), exclusive);
 749         /* should allocate an additional short-lived page for bvset */
 750         if (err == -EAGAIN && !fe->candidate_bvpage) {
 751                 fe->candidate_bvpage = alloc_page(GFP_NOFS | __GFP_NOFAIL);
 752                 set_page_private(fe->candidate_bvpage,
 753                                  Z_EROFS_SHORTLIVED_PAGE);
 754                 goto retry;
 755         }
 756
 757         if (err) {
 758                 DBG_BUGON(err == -EAGAIN && fe->candidate_bvpage);
 759                 goto out;
 760         }
 761
 762         z_erofs_onlinepage_split(page);
 763         /* bump up the number of spiltted parts of a page */
 764         ++spiltted;
 765         if (fe->pcl->pageofs_out != (map->m_la & ~PAGE_MASK))
 766                 fe->pcl->multibases = true;
 767
 768         if ((map->m_flags & EROFS_MAP_FULL_MAPPED) &&
 769             fe->pcl->length == map->m_llen)
 770                 fe->pcl->partial = false;
 771         if (fe->pcl->length < offset + end - map->m_la) {
 772                 fe->pcl->length = offset + end - map->m_la;
 773                 fe->pcl->pageofs_out = map->m_la & ~PAGE_MASK;
 774         }
 775 next_part:
 776         /* shorten the remaining extent to update progress */
 777         map->m_llen = offset + cur - map->m_la;
 778         map->m_flags &= ~EROFS_MAP_FULL_MAPPED;
 779
 780         end = cur;
 781         if (end > 0)
 782                 goto repeat;
 783
 784 out:
 785         if (err)
 786                 z_erofs_page_mark_eio(page);
 787         z_erofs_onlinepage_endio(page);
 788
 789         erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu",
 790                   __func__, page, spiltted, map->m_llen);
 791         return err;
 792 }
 793
 794 static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi,
 795                                        unsigned int readahead_pages)
 796 {
 797         /* auto: enable for read_folio, disable for readahead */
 798         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) &&
 799             !readahead_pages)
 800                 return true;
 801
 802         if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) &&
 803             (readahead_pages <= sbi->opt.max_sync_decompress_pages))
 804                 return true;
 805
 806         return false;
 807 }
 808
 809 static bool z_erofs_page_is_invalidated(struct page *page)
 810 {
 811         return !page->mapping && !z_erofs_is_shortlived_page(page);
 812 }
 813
 814 struct z_erofs_decompress_backend {
 815         struct page *onstack_pages[Z_EROFS_ONSTACK_PAGES];
 816         struct super_block *sb;
 817         struct z_erofs_pcluster *pcl;
 818
 819         /* pages with the longest decompressed length for deduplication */
 820         struct page **decompressed_pages;
 821         /* pages to keep the compressed data */
 822         struct page **compressed_pages;
 823
 824         struct list_head decompressed_secondary_bvecs;
 825         struct page **pagepool;
 826         unsigned int onstack_used, nr_pages;
 827 };
 828
 829 struct z_erofs_bvec_item {
 830         struct z_erofs_bvec bvec;
 831         struct list_head list;
 832 };
 833
 834 static void z_erofs_do_decompressed_bvec(struct z_erofs_decompress_backend *be,
 835                                          struct z_erofs_bvec *bvec)
 836 {
 837         struct z_erofs_bvec_item *item;
 838
 839         if (!((bvec->offset + be->pcl->pageofs_out) & ~PAGE_MASK)) {
 840                 unsigned int pgnr;
 841                 struct page *oldpage;
 842
 843                 pgnr = (bvec->offset + be->pcl->pageofs_out) >> PAGE_SHIFT;
 844                 DBG_BUGON(pgnr >= be->nr_pages);
 845                 oldpage = be->decompressed_pages[pgnr];
 846                 be->decompressed_pages[pgnr] = bvec->page;
 847
 848                 if (!oldpage)
 849                         return;
 850         }
 851
 852         /* (cold path) one pcluster is requested multiple times */
 853         item = kmalloc(sizeof(*item), GFP_KERNEL | __GFP_NOFAIL);
 854         item->bvec = *bvec;
 855         list_add(&item->list, &be->decompressed_secondary_bvecs);
 856 }
 857
 858 static void z_erofs_fill_other_copies(struct z_erofs_decompress_backend *be,
 859                                       int err)
 860 {
 861         unsigned int off0 = be->pcl->pageofs_out;
 862         struct list_head *p, *n;
 863
 864         list_for_each_safe(p, n, &be->decompressed_secondary_bvecs) {
 865                 struct z_erofs_bvec_item *bvi;
 866                 unsigned int end, cur;
 867                 void *dst, *src;
 868
 869                 bvi = container_of(p, struct z_erofs_bvec_item, list);
 870                 cur = bvi->bvec.offset < 0 ? -bvi->bvec.offset : 0;
 871                 end = min_t(unsigned int, be->pcl->length - bvi->bvec.offset,
 872                             bvi->bvec.end);
 873                 dst = kmap_local_page(bvi->bvec.page);
 874                 while (cur < end) {
 875                         unsigned int pgnr, scur, len;
 876
 877                         pgnr = (bvi->bvec.offset + cur + off0) >> PAGE_SHIFT;
 878                         DBG_BUGON(pgnr >= be->nr_pages);
 879
 880                         scur = bvi->bvec.offset + cur -
 881                                         ((pgnr << PAGE_SHIFT) - off0);
 882                         len = min_t(unsigned int, end - cur, PAGE_SIZE - scur);
 883                         if (!be->decompressed_pages[pgnr]) {
 884                                 err = -EFSCORRUPTED;
 885                                 cur += len;
 886                                 continue;
 887                         }
 888                         src = kmap_local_page(be->decompressed_pages[pgnr]);
 889                         memcpy(dst + cur, src + scur, len);
 890                         kunmap_local(src);
 891                         cur += len;
 892                 }
 893                 kunmap_local(dst);
 894                 if (err)
 895                         z_erofs_page_mark_eio(bvi->bvec.page);
 896                 z_erofs_onlinepage_endio(bvi->bvec.page);
 897                 list_del(p);
 898                 kfree(bvi);
 899         }
 900 }
 901
 902 static void z_erofs_parse_out_bvecs(struct z_erofs_decompress_backend *be)
 903 {
 904         struct z_erofs_pcluster *pcl = be->pcl;
 905         struct z_erofs_bvec_iter biter;
 906         struct page *old_bvpage;
 907         int i;
 908
 909         z_erofs_bvec_iter_begin(&biter, &pcl->bvset, Z_EROFS_INLINE_BVECS, 0);
 910         for (i = 0; i < pcl->vcnt; ++i) {
 911                 struct z_erofs_bvec bvec;
 912
 913                 z_erofs_bvec_dequeue(&biter, &bvec, &old_bvpage);
 914
 915                 if (old_bvpage)
 916                         z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
 917
 918                 DBG_BUGON(z_erofs_page_is_invalidated(bvec.page));
 919                 z_erofs_do_decompressed_bvec(be, &bvec);
 920         }
 921
 922         old_bvpage = z_erofs_bvec_iter_end(&biter);
 923         if (old_bvpage)
 924                 z_erofs_put_shortlivedpage(be->pagepool, old_bvpage);
 925 }
 926
 927 static int z_erofs_parse_in_bvecs(struct z_erofs_decompress_backend *be,
 928                                   bool *overlapped)
 929 {
 930         struct z_erofs_pcluster *pcl = be->pcl;
 931         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 932         int i, err = 0;
 933
 934         *overlapped = false;
 935         for (i = 0; i < pclusterpages; ++i) {
 936                 struct z_erofs_bvec *bvec = &pcl->compressed_bvecs[i];
 937                 struct page *page = bvec->page;
 938
 939                 /* compressed pages ought to be present before decompressing */
 940                 if (!page) {
 941                         DBG_BUGON(1);
 942                         continue;
 943                 }
 944                 be->compressed_pages[i] = page;
 945
 946                 if (z_erofs_is_inline_pcluster(pcl)) {
 947                         if (!PageUptodate(page))
 948                                 err = -EIO;
 949                         continue;
 950                 }
 951
 952                 DBG_BUGON(z_erofs_page_is_invalidated(page));
 953                 if (!z_erofs_is_shortlived_page(page)) {
 954                         if (erofs_page_is_managed(EROFS_SB(be->sb), page)) {
 955                                 if (!PageUptodate(page))
 956                                         err = -EIO;
 957                                 continue;
 958                         }
 959                         z_erofs_do_decompressed_bvec(be, bvec);
 960                         *overlapped = true;
 961                 }
 962         }
 963
 964         if (err)
 965                 return err;
 966         return 0;
 967 }
 968
 969 static int z_erofs_decompress_pcluster(struct z_erofs_decompress_backend *be,
 970                                        int err)
 971 {
 972         struct erofs_sb_info *const sbi = EROFS_SB(be->sb);
 973         struct z_erofs_pcluster *pcl = be->pcl;
 974         unsigned int pclusterpages = z_erofs_pclusterpages(pcl);
 975         unsigned int i, inputsize;
 976         int err2;
 977         struct page *page;
 978         bool overlapped;
 979
 980         mutex_lock(&pcl->lock);
 981         be->nr_pages = PAGE_ALIGN(pcl->length + pcl->pageofs_out) >> PAGE_SHIFT;
 982
 983         /* allocate (de)compressed page arrays if cannot be kept on stack */
 984         be->decompressed_pages = NULL;
 985         be->compressed_pages = NULL;
 986         be->onstack_used = 0;
 987         if (be->nr_pages <= Z_EROFS_ONSTACK_PAGES) {
 988                 be->decompressed_pages = be->onstack_pages;
 989                 be->onstack_used = be->nr_pages;
 990                 memset(be->decompressed_pages, 0,
 991                        sizeof(struct page *) * be->nr_pages);
 992         }
 993
 994         if (pclusterpages + be->onstack_used <= Z_EROFS_ONSTACK_PAGES)
 995                 be->compressed_pages = be->onstack_pages + be->onstack_used;
 996
 997         if (!be->decompressed_pages)
 998                 be->decompressed_pages =
 999                         kvcalloc(be->nr_pages, sizeof(struct page *),
1000                                  GFP_KERNEL | __GFP_NOFAIL);
1001         if (!be->compressed_pages)
1002                 be->compressed_pages =
1003                         kvcalloc(pclusterpages, sizeof(struct page *),
1004                                  GFP_KERNEL | __GFP_NOFAIL);
1005
1006         z_erofs_parse_out_bvecs(be);
1007         err2 = z_erofs_parse_in_bvecs(be, &overlapped);
1008         if (err2)
1009                 err = err2;
1010         if (err)
1011                 goto out;
1012
1013         if (z_erofs_is_inline_pcluster(pcl))
1014                 inputsize = pcl->tailpacking_size;
1015         else
1016                 inputsize = pclusterpages * PAGE_SIZE;
1017
1018         err = z_erofs_decompress(&(struct z_erofs_decompress_req) {
1019                                         .sb = be->sb,
1020                                         .in = be->compressed_pages,
1021                                         .out = be->decompressed_pages,
1022                                         .pageofs_in = pcl->pageofs_in,
1023                                         .pageofs_out = pcl->pageofs_out,
1024                                         .inputsize = inputsize,
1025                                         .outputsize = pcl->length,
1026                                         .alg = pcl->algorithmformat,
1027                                         .inplace_io = overlapped,
1028                                         .partial_decoding = pcl->partial,
1029                                         .fillgaps = pcl->multibases,
1030                                  }, be->pagepool);
1031
1032 out:
1033         /* must handle all compressed pages before actual file pages */
1034         if (z_erofs_is_inline_pcluster(pcl)) {
1035                 page = pcl->compressed_bvecs[0].page;
1036                 WRITE_ONCE(pcl->compressed_bvecs[0].page, NULL);
1037                 put_page(page);
1038         } else {
1039                 for (i = 0; i < pclusterpages; ++i) {
1040                         page = pcl->compressed_bvecs[i].page;
1041
1042                         if (erofs_page_is_managed(sbi, page))
1043                                 continue;
1044
1045                         /* recycle all individual short-lived pages */
1046                         (void)z_erofs_put_shortlivedpage(be->pagepool, page);
1047                         WRITE_ONCE(pcl->compressed_bvecs[i].page, NULL);
1048                 }
1049         }
1050         if (be->compressed_pages < be->onstack_pages ||
1051             be->compressed_pages >= be->onstack_pages + Z_EROFS_ONSTACK_PAGES)
1052                 kvfree(be->compressed_pages);
1053         z_erofs_fill_other_copies(be, err);
1054
1055         for (i = 0; i < be->nr_pages; ++i) {
1056                 page = be->decompressed_pages[i];
1057                 if (!page)
1058                         continue;
1059
1060                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1061
1062                 /* recycle all individual short-lived pages */
1063                 if (z_erofs_put_shortlivedpage(be->pagepool, page))
1064                         continue;
1065                 if (err)
1066                         z_erofs_page_mark_eio(page);
1067                 z_erofs_onlinepage_endio(page);
1068         }
1069
1070         if (be->decompressed_pages != be->onstack_pages)
1071                 kvfree(be->decompressed_pages);
1072
1073         pcl->length = 0;
1074         pcl->partial = true;
1075         pcl->multibases = false;
1076         pcl->bvset.nextpage = NULL;
1077         pcl->vcnt = 0;
1078
1079         /* pcluster lock MUST be taken before the following line */
1080         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL);
1081         mutex_unlock(&pcl->lock);
1082         return err;
1083 }
1084
1085 static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io,
1086                                      struct page **pagepool)
1087 {
1088         struct z_erofs_decompress_backend be = {
1089                 .sb = io->sb,
1090                 .pagepool = pagepool,
1091                 .decompressed_secondary_bvecs =
1092                         LIST_HEAD_INIT(be.decompressed_secondary_bvecs),
1093         };
1094         z_erofs_next_pcluster_t owned = io->head;
1095
1096         while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) {
1097                 /* impossible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */
1098                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL);
1099                 /* impossible that 'owned' equals Z_EROFS_PCLUSTER_NIL */
1100                 DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL);
1101
1102                 be.pcl = container_of(owned, struct z_erofs_pcluster, next);
1103                 owned = READ_ONCE(be.pcl->next);
1104
1105                 z_erofs_decompress_pcluster(&be, io->eio ? -EIO : 0);
1106                 erofs_workgroup_put(&be.pcl->obj);
1107         }
1108 }
1109
1110 static void z_erofs_decompressqueue_work(struct work_struct *work)
1111 {
1112         struct z_erofs_decompressqueue *bgq =
1113                 container_of(work, struct z_erofs_decompressqueue, u.work);
1114         struct page *pagepool = NULL;
1115
1116         DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1117         z_erofs_decompress_queue(bgq, &pagepool);
1118
1119         erofs_release_pages(&pagepool);
1120         kvfree(bgq);
1121 }
1122
1123 static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io,
1124                                        bool sync, int bios)
1125 {
1126         struct erofs_sb_info *const sbi = EROFS_SB(io->sb);
1127
1128         /* wake up the caller thread for sync decompression */
1129         if (sync) {
1130                 if (!atomic_add_return(bios, &io->pending_bios))
1131                         complete(&io->u.done);
1132                 return;
1133         }
1134
1135         if (atomic_add_return(bios, &io->pending_bios))
1136                 return;
1137         /* Use workqueue and sync decompression for atomic contexts only */
1138         if (in_atomic() || irqs_disabled()) {
1139                 queue_work(z_erofs_workqueue, &io->u.work);
1140                 /* enable sync decompression for readahead */
1141                 if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO)
1142                         sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON;
1143                 return;
1144         }
1145         z_erofs_decompressqueue_work(&io->u.work);
1146 }
1147
1148 static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl,
1149                                                unsigned int nr,
1150                                                struct page **pagepool,
1151                                                struct address_space *mc)
1152 {
1153         const pgoff_t index = pcl->obj.index;
1154         gfp_t gfp = mapping_gfp_mask(mc);
1155         bool tocache = false;
1156
1157         struct address_space *mapping;
1158         struct page *oldpage, *page;
1159
1160         compressed_page_t t;
1161         int justfound;
1162
1163 repeat:
1164         page = READ_ONCE(pcl->compressed_bvecs[nr].page);
1165         oldpage = page;
1166
1167         if (!page)
1168                 goto out_allocpage;
1169
1170         /* process the target tagged pointer */
1171         t = tagptr_init(compressed_page_t, page);
1172         justfound = tagptr_unfold_tags(t);
1173         page = tagptr_unfold_ptr(t);
1174
1175         /*
1176          * preallocated cached pages, which is used to avoid direct reclaim
1177          * otherwise, it will go inplace I/O path instead.
1178          */
1179         if (page->private == Z_EROFS_PREALLOCATED_PAGE) {
1180                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1181                 set_page_private(page, 0);
1182                 tocache = true;
1183                 goto out_tocache;
1184         }
1185         mapping = READ_ONCE(page->mapping);
1186
1187         /*
1188          * file-backed online pages in plcuster are all locked steady,
1189          * therefore it is impossible for `mapping' to be NULL.
1190          */
1191         if (mapping && mapping != mc)
1192                 /* ought to be unmanaged pages */
1193                 goto out;
1194
1195         /* directly return for shortlived page as well */
1196         if (z_erofs_is_shortlived_page(page))
1197                 goto out;
1198
1199         lock_page(page);
1200
1201         /* only true if page reclaim goes wrong, should never happen */
1202         DBG_BUGON(justfound && PagePrivate(page));
1203
1204         /* the page is still in manage cache */
1205         if (page->mapping == mc) {
1206                 WRITE_ONCE(pcl->compressed_bvecs[nr].page, page);
1207
1208                 if (!PagePrivate(page)) {
1209                         /*
1210                          * impossible to be !PagePrivate(page) for
1211                          * the current restriction as well if
1212                          * the page is already in compressed_bvecs[].
1213                          */
1214                         DBG_BUGON(!justfound);
1215
1216                         justfound = 0;
1217                         set_page_private(page, (unsigned long)pcl);
1218                         SetPagePrivate(page);
1219                 }
1220
1221                 /* no need to submit io if it is already up-to-date */
1222                 if (PageUptodate(page)) {
1223                         unlock_page(page);
1224                         page = NULL;
1225                 }
1226                 goto out;
1227         }
1228
1229         /*
1230          * the managed page has been truncated, it's unsafe to
1231          * reuse this one, let's allocate a new cache-managed page.
1232          */
1233         DBG_BUGON(page->mapping);
1234         DBG_BUGON(!justfound);
1235
1236         tocache = true;
1237         unlock_page(page);
1238         put_page(page);
1239 out_allocpage:
1240         page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL);
1241         if (oldpage != cmpxchg(&pcl->compressed_bvecs[nr].page,
1242                                oldpage, page)) {
1243                 erofs_pagepool_add(pagepool, page);
1244                 cond_resched();
1245                 goto repeat;
1246         }
1247 out_tocache:
1248         if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) {
1249                 /* turn into temporary page if fails (1 ref) */
1250                 set_page_private(page, Z_EROFS_SHORTLIVED_PAGE);
1251                 goto out;
1252         }
1253         attach_page_private(page, pcl);
1254         /* drop a refcount added by allocpage (then we have 2 refs here) */
1255         put_page(page);
1256
1257 out:    /* the only exit (for tracing and debugging) */
1258         return page;
1259 }
1260
1261 static struct z_erofs_decompressqueue *
1262 jobqueue_init(struct super_block *sb,
1263               struct z_erofs_decompressqueue *fgq, bool *fg)
1264 {
1265         struct z_erofs_decompressqueue *q;
1266
1267         if (fg && !*fg) {
1268                 q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN);
1269                 if (!q) {
1270                         *fg = true;
1271                         goto fg_out;
1272                 }
1273                 INIT_WORK(&q->u.work, z_erofs_decompressqueue_work);
1274         } else {
1275 fg_out:
1276                 q = fgq;
1277                 init_completion(&fgq->u.done);
1278                 atomic_set(&fgq->pending_bios, 0);
1279                 q->eio = false;
1280         }
1281         q->sb = sb;
1282         q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1283         return q;
1284 }
1285
1286 /* define decompression jobqueue types */
1287 enum {
1288         JQ_BYPASS,
1289         JQ_SUBMIT,
1290         NR_JOBQUEUES,
1291 };
1292
1293 static void *jobqueueset_init(struct super_block *sb,
1294                               struct z_erofs_decompressqueue *q[],
1295                               struct z_erofs_decompressqueue *fgq, bool *fg)
1296 {
1297         /*
1298          * if managed cache is enabled, bypass jobqueue is needed,
1299          * no need to read from device for all pclusters in this queue.
1300          */
1301         q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL);
1302         q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg);
1303
1304         return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg));
1305 }
1306
1307 static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl,
1308                                     z_erofs_next_pcluster_t qtail[],
1309                                     z_erofs_next_pcluster_t owned_head)
1310 {
1311         z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT];
1312         z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS];
1313
1314         DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1315         if (owned_head == Z_EROFS_PCLUSTER_TAIL)
1316                 owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED;
1317
1318         WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED);
1319
1320         WRITE_ONCE(*submit_qtail, owned_head);
1321         WRITE_ONCE(*bypass_qtail, &pcl->next);
1322
1323         qtail[JQ_BYPASS] = &pcl->next;
1324 }
1325
1326 static void z_erofs_decompressqueue_endio(struct bio *bio)
1327 {
1328         tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private);
1329         struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t);
1330         blk_status_t err = bio->bi_status;
1331         struct bio_vec *bvec;
1332         struct bvec_iter_all iter_all;
1333
1334         bio_for_each_segment_all(bvec, bio, iter_all) {
1335                 struct page *page = bvec->bv_page;
1336
1337                 DBG_BUGON(PageUptodate(page));
1338                 DBG_BUGON(z_erofs_page_is_invalidated(page));
1339
1340                 if (erofs_page_is_managed(EROFS_SB(q->sb), page)) {
1341                         if (!err)
1342                                 SetPageUptodate(page);
1343                         unlock_page(page);
1344                 }
1345         }
1346         if (err)
1347                 q->eio = true;
1348         z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1);
1349         bio_put(bio);
1350 }
1351
1352 static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f,
1353                                  struct page **pagepool,
1354                                  struct z_erofs_decompressqueue *fgq,
1355                                  bool *force_fg)
1356 {
1357         struct super_block *sb = f->inode->i_sb;
1358         struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb));
1359         z_erofs_next_pcluster_t qtail[NR_JOBQUEUES];
1360         struct z_erofs_decompressqueue *q[NR_JOBQUEUES];
1361         void *bi_private;
1362         z_erofs_next_pcluster_t owned_head = f->owned_head;
1363         /* bio is NULL initially, so no need to initialize last_{index,bdev} */
1364         pgoff_t last_index;
1365         struct block_device *last_bdev;
1366         unsigned int nr_bios = 0;
1367         struct bio *bio = NULL;
1368
1369         bi_private = jobqueueset_init(sb, q, fgq, force_fg);
1370         qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head;
1371         qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head;
1372
1373         /* by default, all need io submission */
1374         q[JQ_SUBMIT]->head = owned_head;
1375
1376         do {
1377                 struct erofs_map_dev mdev;
1378                 struct z_erofs_pcluster *pcl;
1379                 pgoff_t cur, end;
1380                 unsigned int i = 0;
1381                 bool bypass = true;
1382
1383                 /* no possible 'owned_head' equals the following */
1384                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED);
1385                 DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL);
1386
1387                 pcl = container_of(owned_head, struct z_erofs_pcluster, next);
1388
1389                 /* close the main owned chain at first */
1390                 owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL,
1391                                      Z_EROFS_PCLUSTER_TAIL_CLOSED);
1392                 if (z_erofs_is_inline_pcluster(pcl)) {
1393                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1394                         continue;
1395                 }
1396
1397                 /* no device id here, thus it will always succeed */
1398                 mdev = (struct erofs_map_dev) {
1399                         .m_pa = blknr_to_addr(pcl->obj.index),
1400                 };
1401                 (void)erofs_map_dev(sb, &mdev);
1402
1403                 cur = erofs_blknr(mdev.m_pa);
1404                 end = cur + pcl->pclusterpages;
1405
1406                 do {
1407                         struct page *page;
1408
1409                         page = pickup_page_for_submission(pcl, i++, pagepool,
1410                                                           mc);
1411                         if (!page)
1412                                 continue;
1413
1414                         if (bio && (cur != last_index + 1 ||
1415                                     last_bdev != mdev.m_bdev)) {
1416 submit_bio_retry:
1417                                 submit_bio(bio);
1418                                 bio = NULL;
1419                         }
1420
1421                         if (!bio) {
1422                                 bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS,
1423                                                 REQ_OP_READ, GFP_NOIO);
1424                                 bio->bi_end_io = z_erofs_decompressqueue_endio;
1425
1426                                 last_bdev = mdev.m_bdev;
1427                                 bio->bi_iter.bi_sector = (sector_t)cur <<
1428                                         LOG_SECTORS_PER_BLOCK;
1429                                 bio->bi_private = bi_private;
1430                                 if (f->readahead)
1431                                         bio->bi_opf |= REQ_RAHEAD;
1432                                 ++nr_bios;
1433                         }
1434
1435                         if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE)
1436                                 goto submit_bio_retry;
1437
1438                         last_index = cur;
1439                         bypass = false;
1440                 } while (++cur < end);
1441
1442                 if (!bypass)
1443                         qtail[JQ_SUBMIT] = &pcl->next;
1444                 else
1445                         move_to_bypass_jobqueue(pcl, qtail, owned_head);
1446         } while (owned_head != Z_EROFS_PCLUSTER_TAIL);
1447
1448         if (bio)
1449                 submit_bio(bio);
1450
1451         /*
1452          * although background is preferred, no one is pending for submission.
1453          * don't issue workqueue for decompression but drop it directly instead.
1454          */
1455         if (!*force_fg && !nr_bios) {
1456                 kvfree(q[JQ_SUBMIT]);
1457                 return;
1458         }
1459         z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios);
1460 }
1461
1462 static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f,
1463                              struct page **pagepool, bool force_fg)
1464 {
1465         struct z_erofs_decompressqueue io[NR_JOBQUEUES];
1466
1467         if (f->owned_head == Z_EROFS_PCLUSTER_TAIL)
1468                 return;
1469         z_erofs_submit_queue(f, pagepool, io, &force_fg);
1470
1471         /* handle bypass queue (no i/o pclusters) immediately */
1472         z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool);
1473
1474         if (!force_fg)
1475                 return;
1476
1477         /* wait until all bios are completed */
1478         wait_for_completion_io(&io[JQ_SUBMIT].u.done);
1479
1480         /* handle synchronous decompress queue in the caller context */
1481         z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool);
1482 }
1483
1484 /*
1485  * Since partial uptodate is still unimplemented for now, we have to use
1486  * approximate readmore strategies as a start.
1487  */
1488 static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f,
1489                                       struct readahead_control *rac,
1490                                       erofs_off_t end,
1491                                       struct page **pagepool,
1492                                       bool backmost)
1493 {
1494         struct inode *inode = f->inode;
1495         struct erofs_map_blocks *map = &f->map;
1496         erofs_off_t cur;
1497         int err;
1498
1499         if (backmost) {
1500                 map->m_la = end;
1501                 err = z_erofs_map_blocks_iter(inode, map,
1502                                               EROFS_GET_BLOCKS_READMORE);
1503                 if (err)
1504                         return;
1505
1506                 /* expend ra for the trailing edge if readahead */
1507                 if (rac) {
1508                         loff_t newstart = readahead_pos(rac);
1509
1510                         cur = round_up(map->m_la + map->m_llen, PAGE_SIZE);
1511                         readahead_expand(rac, newstart, cur - newstart);
1512                         return;
1513                 }
1514                 end = round_up(end, PAGE_SIZE);
1515         } else {
1516                 end = round_up(map->m_la, PAGE_SIZE);
1517
1518                 if (!map->m_llen)
1519                         return;
1520         }
1521
1522         cur = map->m_la + map->m_llen - 1;
1523         while (cur >= end) {
1524                 pgoff_t index = cur >> PAGE_SHIFT;
1525                 struct page *page;
1526
1527                 page = erofs_grab_cache_page_nowait(inode->i_mapping, index);
1528                 if (page) {
1529                         if (PageUptodate(page)) {
1530                                 unlock_page(page);
1531                         } else {
1532                                 err = z_erofs_do_read_page(f, page, pagepool);
1533                                 if (err)
1534                                         erofs_err(inode->i_sb,
1535                                                   "readmore error at page %lu @ nid %llu",
1536                                                   index, EROFS_I(inode)->nid);
1537                         }
1538                         put_page(page);
1539                 }
1540
1541                 if (cur < PAGE_SIZE)
1542                         break;
1543                 cur = (index << PAGE_SHIFT) - 1;
1544         }
1545 }
1546
1547 static int z_erofs_read_folio(struct file *file, struct folio *folio)
1548 {
1549         struct page *page = &folio->page;
1550         struct inode *const inode = page->mapping->host;
1551         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1552         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1553         struct page *pagepool = NULL;
1554         int err;
1555
1556         trace_erofs_readpage(page, false);
1557         f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT;
1558
1559         z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1,
1560                                   &pagepool, true);
1561         err = z_erofs_do_read_page(&f, page, &pagepool);
1562         z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false);
1563
1564         (void)z_erofs_collector_end(&f);
1565
1566         /* if some compressed cluster ready, need submit them anyway */
1567         z_erofs_runqueue(&f, &pagepool,
1568                          z_erofs_get_sync_decompress_policy(sbi, 0));
1569
1570         if (err)
1571                 erofs_err(inode->i_sb, "failed to read, err [%d]", err);
1572
1573         erofs_put_metabuf(&f.map.buf);
1574         erofs_release_pages(&pagepool);
1575         return err;
1576 }
1577
1578 static void z_erofs_readahead(struct readahead_control *rac)
1579 {
1580         struct inode *const inode = rac->mapping->host;
1581         struct erofs_sb_info *const sbi = EROFS_I_SB(inode);
1582         struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode);
1583         struct page *pagepool = NULL, *head = NULL, *page;
1584         unsigned int nr_pages;
1585
1586         f.readahead = true;
1587         f.headoffset = readahead_pos(rac);
1588
1589         z_erofs_pcluster_readmore(&f, rac, f.headoffset +
1590                                   readahead_length(rac) - 1, &pagepool, true);
1591         nr_pages = readahead_count(rac);
1592         trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false);
1593
1594         while ((page = readahead_page(rac))) {
1595                 set_page_private(page, (unsigned long)head);
1596                 head = page;
1597         }
1598
1599         while (head) {
1600                 struct page *page = head;
1601                 int err;
1602
1603                 /* traversal in reverse order */
1604                 head = (void *)page_private(page);
1605
1606                 err = z_erofs_do_read_page(&f, page, &pagepool);
1607                 if (err)
1608                         erofs_err(inode->i_sb,
1609                                   "readahead error at page %lu @ nid %llu",
1610                                   page->index, EROFS_I(inode)->nid);
1611                 put_page(page);
1612         }
1613         z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false);
1614         (void)z_erofs_collector_end(&f);
1615
1616         z_erofs_runqueue(&f, &pagepool,
1617                          z_erofs_get_sync_decompress_policy(sbi, nr_pages));
1618         erofs_put_metabuf(&f.map.buf);
1619         erofs_release_pages(&pagepool);
1620 }
1621
1622 const struct address_space_operations z_erofs_aops = {
1623         .read_folio = z_erofs_read_folio,
1624         .readahead = z_erofs_readahead,
1625 };