fs/bcachefs/io_read.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Some low level IO code, and hacks for various block layer limitations
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <[email protected]>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "alloc_background.h"
  11 #include "alloc_foreground.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "checksum.h"
  15 #include "clock.h"
  16 #include "compress.h"
  17 #include "data_update.h"
  18 #include "disk_groups.h"
  19 #include "ec.h"
  20 #include "error.h"
  21 #include "io_read.h"
  22 #include "io_misc.h"
  23 #include "io_write.h"
  24 #include "subvolume.h"
  25 #include "trace.h"
  26
  27 #include <linux/sched/mm.h>
  28
  29 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
  30
  31 static bool bch2_target_congested(struct bch_fs *c, u16 target)
  32 {
  33         const struct bch_devs_mask *devs;
  34         unsigned d, nr = 0, total = 0;
  35         u64 now = local_clock(), last;
  36         s64 congested;
  37         struct bch_dev *ca;
  38
  39         if (!target)
  40                 return false;
  41
  42         rcu_read_lock();
  43         devs = bch2_target_to_mask(c, target) ?:
  44                 &c->rw_devs[BCH_DATA_user];
  45
  46         for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
  47                 ca = rcu_dereference(c->devs[d]);
  48                 if (!ca)
  49                         continue;
  50
  51                 congested = atomic_read(&ca->congested);
  52                 last = READ_ONCE(ca->congested_last);
  53                 if (time_after64(now, last))
  54                         congested -= (now - last) >> 12;
  55
  56                 total += max(congested, 0LL);
  57                 nr++;
  58         }
  59         rcu_read_unlock();
  60
  61         return bch2_rand_range(nr * CONGESTED_MAX) < total;
  62 }
  63
  64 #else
  65
  66 static bool bch2_target_congested(struct bch_fs *c, u16 target)
  67 {
  68         return false;
  69 }
  70
  71 #endif
  72
  73 /* Cache promotion on read */
  74
  75 struct promote_op {
  76         struct rcu_head         rcu;
  77         u64                     start_time;
  78
  79         struct rhash_head       hash;
  80         struct bpos             pos;
  81
  82         struct data_update      write;
  83         struct bio_vec          bi_inline_vecs[]; /* must be last */
  84 };
  85
  86 static const struct rhashtable_params bch_promote_params = {
  87         .head_offset    = offsetof(struct promote_op, hash),
  88         .key_offset     = offsetof(struct promote_op, pos),
  89         .key_len        = sizeof(struct bpos),
  90 };
  91
  92 static inline int should_promote(struct bch_fs *c, struct bkey_s_c k,
  93                                   struct bpos pos,
  94                                   struct bch_io_opts opts,
  95                                   unsigned flags)
  96 {
  97         BUG_ON(!opts.promote_target);
  98
  99         if (!(flags & BCH_READ_MAY_PROMOTE))
 100                 return -BCH_ERR_nopromote_may_not;
 101
 102         if (bch2_bkey_has_target(c, k, opts.promote_target))
 103                 return -BCH_ERR_nopromote_already_promoted;
 104
 105         if (bkey_extent_is_unwritten(k))
 106                 return -BCH_ERR_nopromote_unwritten;
 107
 108         if (bch2_target_congested(c, opts.promote_target))
 109                 return -BCH_ERR_nopromote_congested;
 110
 111         if (rhashtable_lookup_fast(&c->promote_table, &pos,
 112                                    bch_promote_params))
 113                 return -BCH_ERR_nopromote_in_flight;
 114
 115         return 0;
 116 }
 117
 118 static void promote_free(struct bch_fs *c, struct promote_op *op)
 119 {
 120         int ret;
 121
 122         bch2_data_update_exit(&op->write);
 123
 124         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
 125                                      bch_promote_params);
 126         BUG_ON(ret);
 127         bch2_write_ref_put(c, BCH_WRITE_REF_promote);
 128         kfree_rcu(op, rcu);
 129 }
 130
 131 static void promote_done(struct bch_write_op *wop)
 132 {
 133         struct promote_op *op =
 134                 container_of(wop, struct promote_op, write.op);
 135         struct bch_fs *c = op->write.op.c;
 136
 137         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
 138                                op->start_time);
 139         promote_free(c, op);
 140 }
 141
 142 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
 143 {
 144         struct bio *bio = &op->write.op.wbio.bio;
 145
 146         trace_and_count(op->write.op.c, read_promote, &rbio->bio);
 147
 148         /* we now own pages: */
 149         BUG_ON(!rbio->bounce);
 150         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
 151
 152         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
 153                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
 154         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
 155
 156         bch2_data_update_read_done(&op->write, rbio->pick.crc);
 157 }
 158
 159 static struct promote_op *__promote_alloc(struct btree_trans *trans,
 160                                           enum btree_id btree_id,
 161                                           struct bkey_s_c k,
 162                                           struct bpos pos,
 163                                           struct extent_ptr_decoded *pick,
 164                                           struct bch_io_opts opts,
 165                                           unsigned sectors,
 166                                           struct bch_read_bio **rbio)
 167 {
 168         struct bch_fs *c = trans->c;
 169         struct promote_op *op = NULL;
 170         struct bio *bio;
 171         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
 172         int ret;
 173
 174         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_promote))
 175                 return ERR_PTR(-BCH_ERR_nopromote_no_writes);
 176
 177         op = kzalloc(struct_size(op, bi_inline_vecs, pages), GFP_KERNEL);
 178         if (!op) {
 179                 ret = -BCH_ERR_nopromote_enomem;
 180                 goto err;
 181         }
 182
 183         op->start_time = local_clock();
 184         op->pos = pos;
 185
 186         /*
 187          * We don't use the mempool here because extents that aren't
 188          * checksummed or compressed can be too big for the mempool:
 189          */
 190         *rbio = kzalloc(sizeof(struct bch_read_bio) +
 191                         sizeof(struct bio_vec) * pages,
 192                         GFP_KERNEL);
 193         if (!*rbio) {
 194                 ret = -BCH_ERR_nopromote_enomem;
 195                 goto err;
 196         }
 197
 198         rbio_init(&(*rbio)->bio, opts);
 199         bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
 200
 201         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9, GFP_KERNEL)) {
 202                 ret = -BCH_ERR_nopromote_enomem;
 203                 goto err;
 204         }
 205
 206         (*rbio)->bounce         = true;
 207         (*rbio)->split          = true;
 208         (*rbio)->kmalloc        = true;
 209
 210         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
 211                                           bch_promote_params)) {
 212                 ret = -BCH_ERR_nopromote_in_flight;
 213                 goto err;
 214         }
 215
 216         bio = &op->write.op.wbio.bio;
 217         bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
 218
 219         ret = bch2_data_update_init(trans, NULL, NULL, &op->write,
 220                         writepoint_hashed((unsigned long) current),
 221                         opts,
 222                         (struct data_update_opts) {
 223                                 .target         = opts.promote_target,
 224                                 .extra_replicas = 1,
 225                                 .write_flags    = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
 226                         },
 227                         btree_id, k);
 228         /*
 229          * possible errors: -BCH_ERR_nocow_lock_blocked,
 230          * -BCH_ERR_ENOSPC_disk_reservation:
 231          */
 232         if (ret) {
 233                 BUG_ON(rhashtable_remove_fast(&c->promote_table, &op->hash,
 234                                               bch_promote_params));
 235                 goto err;
 236         }
 237
 238         op->write.op.end_io = promote_done;
 239
 240         return op;
 241 err:
 242         if (*rbio)
 243                 bio_free_pages(&(*rbio)->bio);
 244         kfree(*rbio);
 245         *rbio = NULL;
 246         kfree(op);
 247         bch2_write_ref_put(c, BCH_WRITE_REF_promote);
 248         return ERR_PTR(ret);
 249 }
 250
 251 noinline
 252 static struct promote_op *promote_alloc(struct btree_trans *trans,
 253                                         struct bvec_iter iter,
 254                                         struct bkey_s_c k,
 255                                         struct extent_ptr_decoded *pick,
 256                                         struct bch_io_opts opts,
 257                                         unsigned flags,
 258                                         struct bch_read_bio **rbio,
 259                                         bool *bounce,
 260                                         bool *read_full)
 261 {
 262         struct bch_fs *c = trans->c;
 263         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
 264         /* data might have to be decompressed in the write path: */
 265         unsigned sectors = promote_full
 266                 ? max(pick->crc.compressed_size, pick->crc.live_size)
 267                 : bvec_iter_sectors(iter);
 268         struct bpos pos = promote_full
 269                 ? bkey_start_pos(k.k)
 270                 : POS(k.k->p.inode, iter.bi_sector);
 271         struct promote_op *promote;
 272         int ret;
 273
 274         ret = should_promote(c, k, pos, opts, flags);
 275         if (ret)
 276                 goto nopromote;
 277
 278         promote = __promote_alloc(trans,
 279                                   k.k->type == KEY_TYPE_reflink_v
 280                                   ? BTREE_ID_reflink
 281                                   : BTREE_ID_extents,
 282                                   k, pos, pick, opts, sectors, rbio);
 283         ret = PTR_ERR_OR_ZERO(promote);
 284         if (ret)
 285                 goto nopromote;
 286
 287         *bounce         = true;
 288         *read_full      = promote_full;
 289         return promote;
 290 nopromote:
 291         trace_read_nopromote(c, ret);
 292         return NULL;
 293 }
 294
 295 /* Read */
 296
 297 #define READ_RETRY_AVOID        1
 298 #define READ_RETRY              2
 299 #define READ_ERR                3
 300
 301 enum rbio_context {
 302         RBIO_CONTEXT_NULL,
 303         RBIO_CONTEXT_HIGHPRI,
 304         RBIO_CONTEXT_UNBOUND,
 305 };
 306
 307 static inline struct bch_read_bio *
 308 bch2_rbio_parent(struct bch_read_bio *rbio)
 309 {
 310         return rbio->split ? rbio->parent : rbio;
 311 }
 312
 313 __always_inline
 314 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
 315                            enum rbio_context context,
 316                            struct workqueue_struct *wq)
 317 {
 318         if (context <= rbio->context) {
 319                 fn(&rbio->work);
 320         } else {
 321                 rbio->work.func         = fn;
 322                 rbio->context           = context;
 323                 queue_work(wq, &rbio->work);
 324         }
 325 }
 326
 327 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
 328 {
 329         BUG_ON(rbio->bounce && !rbio->split);
 330
 331         if (rbio->promote)
 332                 promote_free(rbio->c, rbio->promote);
 333         rbio->promote = NULL;
 334
 335         if (rbio->bounce)
 336                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
 337
 338         if (rbio->split) {
 339                 struct bch_read_bio *parent = rbio->parent;
 340
 341                 if (rbio->kmalloc)
 342                         kfree(rbio);
 343                 else
 344                         bio_put(&rbio->bio);
 345
 346                 rbio = parent;
 347         }
 348
 349         return rbio;
 350 }
 351
 352 /*
 353  * Only called on a top level bch_read_bio to complete an entire read request,
 354  * not a split:
 355  */
 356 static void bch2_rbio_done(struct bch_read_bio *rbio)
 357 {
 358         if (rbio->start_time)
 359                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
 360                                        rbio->start_time);
 361         bio_endio(&rbio->bio);
 362 }
 363
 364 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
 365                                      struct bvec_iter bvec_iter,
 366                                      struct bch_io_failures *failed,
 367                                      unsigned flags)
 368 {
 369         struct btree_trans *trans = bch2_trans_get(c);
 370         struct btree_iter iter;
 371         struct bkey_buf sk;
 372         struct bkey_s_c k;
 373         int ret;
 374
 375         flags &= ~BCH_READ_LAST_FRAGMENT;
 376         flags |= BCH_READ_MUST_CLONE;
 377
 378         bch2_bkey_buf_init(&sk);
 379
 380         bch2_trans_iter_init(trans, &iter, rbio->data_btree,
 381                              rbio->read_pos, BTREE_ITER_slots);
 382 retry:
 383         rbio->bio.bi_status = 0;
 384
 385         k = bch2_btree_iter_peek_slot(&iter);
 386         if (bkey_err(k))
 387                 goto err;
 388
 389         bch2_bkey_buf_reassemble(&sk, c, k);
 390         k = bkey_i_to_s_c(sk.k);
 391         bch2_trans_unlock(trans);
 392
 393         if (!bch2_bkey_matches_ptr(c, k,
 394                                    rbio->pick.ptr,
 395                                    rbio->data_pos.offset -
 396                                    rbio->pick.crc.offset)) {
 397                 /* extent we wanted to read no longer exists: */
 398                 rbio->hole = true;
 399                 goto out;
 400         }
 401
 402         ret = __bch2_read_extent(trans, rbio, bvec_iter,
 403                                  rbio->read_pos,
 404                                  rbio->data_btree,
 405                                  k, 0, failed, flags);
 406         if (ret == READ_RETRY)
 407                 goto retry;
 408         if (ret)
 409                 goto err;
 410 out:
 411         bch2_rbio_done(rbio);
 412         bch2_trans_iter_exit(trans, &iter);
 413         bch2_trans_put(trans);
 414         bch2_bkey_buf_exit(&sk, c);
 415         return;
 416 err:
 417         rbio->bio.bi_status = BLK_STS_IOERR;
 418         goto out;
 419 }
 420
 421 static void bch2_rbio_retry(struct work_struct *work)
 422 {
 423         struct bch_read_bio *rbio =
 424                 container_of(work, struct bch_read_bio, work);
 425         struct bch_fs *c        = rbio->c;
 426         struct bvec_iter iter   = rbio->bvec_iter;
 427         unsigned flags          = rbio->flags;
 428         subvol_inum inum = {
 429                 .subvol = rbio->subvol,
 430                 .inum   = rbio->read_pos.inode,
 431         };
 432         struct bch_io_failures failed = { .nr = 0 };
 433
 434         trace_and_count(c, read_retry, &rbio->bio);
 435
 436         if (rbio->retry == READ_RETRY_AVOID)
 437                 bch2_mark_io_failure(&failed, &rbio->pick);
 438
 439         rbio->bio.bi_status = 0;
 440
 441         rbio = bch2_rbio_free(rbio);
 442
 443         flags |= BCH_READ_IN_RETRY;
 444         flags &= ~BCH_READ_MAY_PROMOTE;
 445
 446         if (flags & BCH_READ_NODECODE) {
 447                 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
 448         } else {
 449                 flags &= ~BCH_READ_LAST_FRAGMENT;
 450                 flags |= BCH_READ_MUST_CLONE;
 451
 452                 __bch2_read(c, rbio, iter, inum, &failed, flags);
 453         }
 454 }
 455
 456 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
 457                             blk_status_t error)
 458 {
 459         rbio->retry = retry;
 460
 461         if (rbio->flags & BCH_READ_IN_RETRY)
 462                 return;
 463
 464         if (retry == READ_ERR) {
 465                 rbio = bch2_rbio_free(rbio);
 466
 467                 rbio->bio.bi_status = error;
 468                 bch2_rbio_done(rbio);
 469         } else {
 470                 bch2_rbio_punt(rbio, bch2_rbio_retry,
 471                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
 472         }
 473 }
 474
 475 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
 476                                    struct bch_read_bio *rbio)
 477 {
 478         struct bch_fs *c = rbio->c;
 479         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
 480         struct bch_extent_crc_unpacked new_crc;
 481         struct btree_iter iter;
 482         struct bkey_i *new;
 483         struct bkey_s_c k;
 484         int ret = 0;
 485
 486         if (crc_is_compressed(rbio->pick.crc))
 487                 return 0;
 488
 489         k = bch2_bkey_get_iter(trans, &iter, rbio->data_btree, rbio->data_pos,
 490                                BTREE_ITER_slots|BTREE_ITER_intent);
 491         if ((ret = bkey_err(k)))
 492                 goto out;
 493
 494         if (bversion_cmp(k.k->version, rbio->version) ||
 495             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
 496                 goto out;
 497
 498         /* Extent was merged? */
 499         if (bkey_start_offset(k.k) < data_offset ||
 500             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
 501                 goto out;
 502
 503         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
 504                         rbio->pick.crc, NULL, &new_crc,
 505                         bkey_start_offset(k.k) - data_offset, k.k->size,
 506                         rbio->pick.crc.csum_type)) {
 507                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
 508                 ret = 0;
 509                 goto out;
 510         }
 511
 512         /*
 513          * going to be temporarily appending another checksum entry:
 514          */
 515         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
 516                                  sizeof(struct bch_extent_crc128));
 517         if ((ret = PTR_ERR_OR_ZERO(new)))
 518                 goto out;
 519
 520         bkey_reassemble(new, k);
 521
 522         if (!bch2_bkey_narrow_crcs(new, new_crc))
 523                 goto out;
 524
 525         ret = bch2_trans_update(trans, &iter, new,
 526                                 BTREE_UPDATE_internal_snapshot_node);
 527 out:
 528         bch2_trans_iter_exit(trans, &iter);
 529         return ret;
 530 }
 531
 532 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
 533 {
 534         bch2_trans_do(rbio->c, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 535                       __bch2_rbio_narrow_crcs(trans, rbio));
 536 }
 537
 538 /* Inner part that may run in process context */
 539 static void __bch2_read_endio(struct work_struct *work)
 540 {
 541         struct bch_read_bio *rbio =
 542                 container_of(work, struct bch_read_bio, work);
 543         struct bch_fs *c        = rbio->c;
 544         struct bio *src         = &rbio->bio;
 545         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
 546         struct bvec_iter dst_iter = rbio->bvec_iter;
 547         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
 548         struct nonce nonce = extent_nonce(rbio->version, crc);
 549         unsigned nofs_flags;
 550         struct bch_csum csum;
 551         int ret;
 552
 553         nofs_flags = memalloc_nofs_save();
 554
 555         /* Reset iterator for checksumming and copying bounced data: */
 556         if (rbio->bounce) {
 557                 src->bi_iter.bi_size            = crc.compressed_size << 9;
 558                 src->bi_iter.bi_idx             = 0;
 559                 src->bi_iter.bi_bvec_done       = 0;
 560         } else {
 561                 src->bi_iter                    = rbio->bvec_iter;
 562         }
 563
 564         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
 565         if (bch2_crc_cmp(csum, rbio->pick.crc.csum) && !c->opts.no_data_io)
 566                 goto csum_err;
 567
 568         /*
 569          * XXX
 570          * We need to rework the narrow_crcs path to deliver the read completion
 571          * first, and then punt to a different workqueue, otherwise we're
 572          * holding up reads while doing btree updates which is bad for memory
 573          * reclaim.
 574          */
 575         if (unlikely(rbio->narrow_crcs))
 576                 bch2_rbio_narrow_crcs(rbio);
 577
 578         if (rbio->flags & BCH_READ_NODECODE)
 579                 goto nodecode;
 580
 581         /* Adjust crc to point to subset of data we want: */
 582         crc.offset     += rbio->offset_into_extent;
 583         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
 584
 585         if (crc_is_compressed(crc)) {
 586                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
 587                 if (ret)
 588                         goto decrypt_err;
 589
 590                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc) &&
 591                     !c->opts.no_data_io)
 592                         goto decompression_err;
 593         } else {
 594                 /* don't need to decrypt the entire bio: */
 595                 nonce = nonce_add(nonce, crc.offset << 9);
 596                 bio_advance(src, crc.offset << 9);
 597
 598                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
 599                 src->bi_iter.bi_size = dst_iter.bi_size;
 600
 601                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
 602                 if (ret)
 603                         goto decrypt_err;
 604
 605                 if (rbio->bounce) {
 606                         struct bvec_iter src_iter = src->bi_iter;
 607
 608                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
 609                 }
 610         }
 611
 612         if (rbio->promote) {
 613                 /*
 614                  * Re encrypt data we decrypted, so it's consistent with
 615                  * rbio->crc:
 616                  */
 617                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
 618                 if (ret)
 619                         goto decrypt_err;
 620
 621                 promote_start(rbio->promote, rbio);
 622                 rbio->promote = NULL;
 623         }
 624 nodecode:
 625         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
 626                 rbio = bch2_rbio_free(rbio);
 627                 bch2_rbio_done(rbio);
 628         }
 629 out:
 630         memalloc_nofs_restore(nofs_flags);
 631         return;
 632 csum_err:
 633         /*
 634          * Checksum error: if the bio wasn't bounced, we may have been
 635          * reading into buffers owned by userspace (that userspace can
 636          * scribble over) - retry the read, bouncing it this time:
 637          */
 638         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
 639                 rbio->flags |= BCH_READ_MUST_BOUNCE;
 640                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
 641                 goto out;
 642         }
 643
 644         struct printbuf buf = PRINTBUF;
 645         buf.atomic++;
 646         prt_str(&buf, "data ");
 647         bch2_csum_err_msg(&buf, crc.csum_type, rbio->pick.crc.csum, csum);
 648
 649         struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
 650         if (ca) {
 651                 bch_err_inum_offset_ratelimited(ca,
 652                         rbio->read_pos.inode,
 653                         rbio->read_pos.offset << 9,
 654                         "data %s", buf.buf);
 655                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
 656         }
 657         printbuf_exit(&buf);
 658         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
 659         goto out;
 660 decompression_err:
 661         bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
 662                                         rbio->read_pos.offset << 9,
 663                                         "decompression error");
 664         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
 665         goto out;
 666 decrypt_err:
 667         bch_err_inum_offset_ratelimited(c, rbio->read_pos.inode,
 668                                         rbio->read_pos.offset << 9,
 669                                         "decrypt error");
 670         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
 671         goto out;
 672 }
 673
 674 static void bch2_read_endio(struct bio *bio)
 675 {
 676         struct bch_read_bio *rbio =
 677                 container_of(bio, struct bch_read_bio, bio);
 678         struct bch_fs *c        = rbio->c;
 679         struct bch_dev *ca = rbio->have_ioref ? bch2_dev_have_ref(c, rbio->pick.ptr.dev) : NULL;
 680         struct workqueue_struct *wq = NULL;
 681         enum rbio_context context = RBIO_CONTEXT_NULL;
 682
 683         if (rbio->have_ioref) {
 684                 bch2_latency_acct(ca, rbio->submit_time, READ);
 685                 percpu_ref_put(&ca->io_ref);
 686         }
 687
 688         if (!rbio->split)
 689                 rbio->bio.bi_end_io = rbio->end_io;
 690
 691         if (bio->bi_status) {
 692                 if (ca) {
 693                         bch_err_inum_offset_ratelimited(ca,
 694                                 rbio->read_pos.inode,
 695                                 rbio->read_pos.offset,
 696                                 "data read error: %s",
 697                                 bch2_blk_status_to_str(bio->bi_status));
 698                         bch2_io_error(ca, BCH_MEMBER_ERROR_read);
 699                 }
 700                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
 701                 return;
 702         }
 703
 704         if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
 705             (ca && dev_ptr_stale(ca, &rbio->pick.ptr))) {
 706                 trace_and_count(c, read_reuse_race, &rbio->bio);
 707
 708                 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
 709                         bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
 710                 else
 711                         bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
 712                 return;
 713         }
 714
 715         if (rbio->narrow_crcs ||
 716             rbio->promote ||
 717             crc_is_compressed(rbio->pick.crc) ||
 718             bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
 719                 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
 720         else if (rbio->pick.crc.csum_type)
 721                 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
 722
 723         bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
 724 }
 725
 726 int __bch2_read_indirect_extent(struct btree_trans *trans,
 727                                 unsigned *offset_into_extent,
 728                                 struct bkey_buf *orig_k)
 729 {
 730         struct btree_iter iter;
 731         struct bkey_s_c k;
 732         u64 reflink_offset;
 733         int ret;
 734
 735         reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
 736                 *offset_into_extent;
 737
 738         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_reflink,
 739                                POS(0, reflink_offset), 0);
 740         ret = bkey_err(k);
 741         if (ret)
 742                 goto err;
 743
 744         if (k.k->type != KEY_TYPE_reflink_v &&
 745             k.k->type != KEY_TYPE_indirect_inline_data) {
 746                 bch_err_inum_offset_ratelimited(trans->c,
 747                         orig_k->k->k.p.inode,
 748                         orig_k->k->k.p.offset << 9,
 749                         "%llu len %u points to nonexistent indirect extent %llu",
 750                         orig_k->k->k.p.offset,
 751                         orig_k->k->k.size,
 752                         reflink_offset);
 753                 bch2_inconsistent_error(trans->c);
 754                 ret = -EIO;
 755                 goto err;
 756         }
 757
 758         *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
 759         bch2_bkey_buf_reassemble(orig_k, trans->c, k);
 760 err:
 761         bch2_trans_iter_exit(trans, &iter);
 762         return ret;
 763 }
 764
 765 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
 766                                                    struct bch_dev *ca,
 767                                                    struct bkey_s_c k,
 768                                                    struct bch_extent_ptr ptr)
 769 {
 770         struct bch_fs *c = trans->c;
 771         struct btree_iter iter;
 772         struct printbuf buf = PRINTBUF;
 773         int ret;
 774
 775         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 776                              PTR_BUCKET_POS(ca, &ptr),
 777                              BTREE_ITER_cached);
 778
 779         prt_printf(&buf, "Attempting to read from stale dirty pointer:\n");
 780         printbuf_indent_add(&buf, 2);
 781
 782         bch2_bkey_val_to_text(&buf, c, k);
 783         prt_newline(&buf);
 784
 785         prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
 786
 787         ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
 788         if (!ret) {
 789                 prt_newline(&buf);
 790                 bch2_bkey_val_to_text(&buf, c, k);
 791         }
 792
 793         bch2_fs_inconsistent(c, "%s", buf.buf);
 794
 795         bch2_trans_iter_exit(trans, &iter);
 796         printbuf_exit(&buf);
 797 }
 798
 799 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
 800                        struct bvec_iter iter, struct bpos read_pos,
 801                        enum btree_id data_btree, struct bkey_s_c k,
 802                        unsigned offset_into_extent,
 803                        struct bch_io_failures *failed, unsigned flags)
 804 {
 805         struct bch_fs *c = trans->c;
 806         struct extent_ptr_decoded pick;
 807         struct bch_read_bio *rbio = NULL;
 808         struct promote_op *promote = NULL;
 809         bool bounce = false, read_full = false, narrow_crcs = false;
 810         struct bpos data_pos = bkey_start_pos(k.k);
 811         int pick_ret;
 812
 813         if (bkey_extent_is_inline_data(k.k)) {
 814                 unsigned bytes = min_t(unsigned, iter.bi_size,
 815                                        bkey_inline_data_bytes(k.k));
 816
 817                 swap(iter.bi_size, bytes);
 818                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
 819                 swap(iter.bi_size, bytes);
 820                 bio_advance_iter(&orig->bio, &iter, bytes);
 821                 zero_fill_bio_iter(&orig->bio, iter);
 822                 goto out_read_done;
 823         }
 824 retry_pick:
 825         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
 826
 827         /* hole or reservation - just zero fill: */
 828         if (!pick_ret)
 829                 goto hole;
 830
 831         if (pick_ret < 0) {
 832                 bch_err_inum_offset_ratelimited(c,
 833                                 read_pos.inode, read_pos.offset << 9,
 834                                 "no device to read from");
 835                 goto err;
 836         }
 837
 838         struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
 839
 840         /*
 841          * Stale dirty pointers are treated as IO errors, but @failed isn't
 842          * allocated unless we're in the retry path - so if we're not in the
 843          * retry path, don't check here, it'll be caught in bch2_read_endio()
 844          * and we'll end up in the retry path:
 845          */
 846         if ((flags & BCH_READ_IN_RETRY) &&
 847             !pick.ptr.cached &&
 848             ca &&
 849             unlikely(dev_ptr_stale(ca, &pick.ptr))) {
 850                 read_from_stale_dirty_pointer(trans, ca, k, pick.ptr);
 851                 bch2_mark_io_failure(failed, &pick);
 852                 percpu_ref_put(&ca->io_ref);
 853                 goto retry_pick;
 854         }
 855
 856         /*
 857          * Unlock the iterator while the btree node's lock is still in
 858          * cache, before doing the IO:
 859          */
 860         bch2_trans_unlock(trans);
 861
 862         if (flags & BCH_READ_NODECODE) {
 863                 /*
 864                  * can happen if we retry, and the extent we were going to read
 865                  * has been merged in the meantime:
 866                  */
 867                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS) {
 868                         if (ca)
 869                                 percpu_ref_put(&ca->io_ref);
 870                         goto hole;
 871                 }
 872
 873                 iter.bi_size    = pick.crc.compressed_size << 9;
 874                 goto get_bio;
 875         }
 876
 877         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
 878             bio_flagged(&orig->bio, BIO_CHAIN))
 879                 flags |= BCH_READ_MUST_CLONE;
 880
 881         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
 882                 bch2_can_narrow_extent_crcs(k, pick.crc);
 883
 884         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
 885                 flags |= BCH_READ_MUST_BOUNCE;
 886
 887         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
 888
 889         if (crc_is_compressed(pick.crc) ||
 890             (pick.crc.csum_type != BCH_CSUM_none &&
 891              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
 892               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
 893                (flags & BCH_READ_USER_MAPPED)) ||
 894               (flags & BCH_READ_MUST_BOUNCE)))) {
 895                 read_full = true;
 896                 bounce = true;
 897         }
 898
 899         if (orig->opts.promote_target)
 900                 promote = promote_alloc(trans, iter, k, &pick, orig->opts, flags,
 901                                         &rbio, &bounce, &read_full);
 902
 903         if (!read_full) {
 904                 EBUG_ON(crc_is_compressed(pick.crc));
 905                 EBUG_ON(pick.crc.csum_type &&
 906                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
 907                          bvec_iter_sectors(iter) != pick.crc.live_size ||
 908                          pick.crc.offset ||
 909                          offset_into_extent));
 910
 911                 data_pos.offset += offset_into_extent;
 912                 pick.ptr.offset += pick.crc.offset +
 913                         offset_into_extent;
 914                 offset_into_extent              = 0;
 915                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
 916                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
 917                 pick.crc.offset                 = 0;
 918                 pick.crc.live_size              = bvec_iter_sectors(iter);
 919         }
 920 get_bio:
 921         if (rbio) {
 922                 /*
 923                  * promote already allocated bounce rbio:
 924                  * promote needs to allocate a bio big enough for uncompressing
 925                  * data in the write path, but we're not going to use it all
 926                  * here:
 927                  */
 928                 EBUG_ON(rbio->bio.bi_iter.bi_size <
 929                        pick.crc.compressed_size << 9);
 930                 rbio->bio.bi_iter.bi_size =
 931                         pick.crc.compressed_size << 9;
 932         } else if (bounce) {
 933                 unsigned sectors = pick.crc.compressed_size;
 934
 935                 rbio = rbio_init(bio_alloc_bioset(NULL,
 936                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
 937                                                   0,
 938                                                   GFP_NOFS,
 939                                                   &c->bio_read_split),
 940                                  orig->opts);
 941
 942                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
 943                 rbio->bounce    = true;
 944                 rbio->split     = true;
 945         } else if (flags & BCH_READ_MUST_CLONE) {
 946                 /*
 947                  * Have to clone if there were any splits, due to error
 948                  * reporting issues (if a split errored, and retrying didn't
 949                  * work, when it reports the error to its parent (us) we don't
 950                  * know if the error was from our bio, and we should retry, or
 951                  * from the whole bio, in which case we don't want to retry and
 952                  * lose the error)
 953                  */
 954                 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOFS,
 955                                                  &c->bio_read_split),
 956                                  orig->opts);
 957                 rbio->bio.bi_iter = iter;
 958                 rbio->split     = true;
 959         } else {
 960                 rbio = orig;
 961                 rbio->bio.bi_iter = iter;
 962                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
 963         }
 964
 965         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
 966
 967         rbio->c                 = c;
 968         rbio->submit_time       = local_clock();
 969         if (rbio->split)
 970                 rbio->parent    = orig;
 971         else
 972                 rbio->end_io    = orig->bio.bi_end_io;
 973         rbio->bvec_iter         = iter;
 974         rbio->offset_into_extent= offset_into_extent;
 975         rbio->flags             = flags;
 976         rbio->have_ioref        = ca != NULL;
 977         rbio->narrow_crcs       = narrow_crcs;
 978         rbio->hole              = 0;
 979         rbio->retry             = 0;
 980         rbio->context           = 0;
 981         /* XXX: only initialize this if needed */
 982         rbio->devs_have         = bch2_bkey_devs(k);
 983         rbio->pick              = pick;
 984         rbio->subvol            = orig->subvol;
 985         rbio->read_pos          = read_pos;
 986         rbio->data_btree        = data_btree;
 987         rbio->data_pos          = data_pos;
 988         rbio->version           = k.k->version;
 989         rbio->promote           = promote;
 990         INIT_WORK(&rbio->work, NULL);
 991
 992         rbio->bio.bi_opf        = orig->bio.bi_opf;
 993         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
 994         rbio->bio.bi_end_io     = bch2_read_endio;
 995
 996         if (rbio->bounce)
 997                 trace_and_count(c, read_bounce, &rbio->bio);
 998
 999         this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
1000         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
1001
1002         /*
1003          * If it's being moved internally, we don't want to flag it as a cache
1004          * hit:
1005          */
1006         if (ca && pick.ptr.cached && !(flags & BCH_READ_NODECODE))
1007                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
1008                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
1009
1010         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
1011                 bio_inc_remaining(&orig->bio);
1012                 trace_and_count(c, read_split, &orig->bio);
1013         }
1014
1015         if (!rbio->pick.idx) {
1016                 if (!rbio->have_ioref) {
1017                         bch_err_inum_offset_ratelimited(c,
1018                                         read_pos.inode,
1019                                         read_pos.offset << 9,
1020                                         "no device to read from");
1021                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1022                         goto out;
1023                 }
1024
1025                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
1026                              bio_sectors(&rbio->bio));
1027                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
1028
1029                 if (unlikely(c->opts.no_data_io)) {
1030                         if (likely(!(flags & BCH_READ_IN_RETRY)))
1031                                 bio_endio(&rbio->bio);
1032                 } else {
1033                         if (likely(!(flags & BCH_READ_IN_RETRY)))
1034                                 submit_bio(&rbio->bio);
1035                         else
1036                                 submit_bio_wait(&rbio->bio);
1037                 }
1038
1039                 /*
1040                  * We just submitted IO which may block, we expect relock fail
1041                  * events and shouldn't count them:
1042                  */
1043                 trans->notrace_relock_fail = true;
1044         } else {
1045                 /* Attempting reconstruct read: */
1046                 if (bch2_ec_read_extent(trans, rbio)) {
1047                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1048                         goto out;
1049                 }
1050
1051                 if (likely(!(flags & BCH_READ_IN_RETRY)))
1052                         bio_endio(&rbio->bio);
1053         }
1054 out:
1055         if (likely(!(flags & BCH_READ_IN_RETRY))) {
1056                 return 0;
1057         } else {
1058                 int ret;
1059
1060                 rbio->context = RBIO_CONTEXT_UNBOUND;
1061                 bch2_read_endio(&rbio->bio);
1062
1063                 ret = rbio->retry;
1064                 rbio = bch2_rbio_free(rbio);
1065
1066                 if (ret == READ_RETRY_AVOID) {
1067                         bch2_mark_io_failure(failed, &pick);
1068                         ret = READ_RETRY;
1069                 }
1070
1071                 if (!ret)
1072                         goto out_read_done;
1073
1074                 return ret;
1075         }
1076
1077 err:
1078         if (flags & BCH_READ_IN_RETRY)
1079                 return READ_ERR;
1080
1081         orig->bio.bi_status = BLK_STS_IOERR;
1082         goto out_read_done;
1083
1084 hole:
1085         /*
1086          * won't normally happen in the BCH_READ_NODECODE
1087          * (bch2_move_extent()) path, but if we retry and the extent we wanted
1088          * to read no longer exists we have to signal that:
1089          */
1090         if (flags & BCH_READ_NODECODE)
1091                 orig->hole = true;
1092
1093         zero_fill_bio_iter(&orig->bio, iter);
1094 out_read_done:
1095         if (flags & BCH_READ_LAST_FRAGMENT)
1096                 bch2_rbio_done(orig);
1097         return 0;
1098 }
1099
1100 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
1101                  struct bvec_iter bvec_iter, subvol_inum inum,
1102                  struct bch_io_failures *failed, unsigned flags)
1103 {
1104         struct btree_trans *trans = bch2_trans_get(c);
1105         struct btree_iter iter;
1106         struct bkey_buf sk;
1107         struct bkey_s_c k;
1108         u32 snapshot;
1109         int ret;
1110
1111         BUG_ON(flags & BCH_READ_NODECODE);
1112
1113         bch2_bkey_buf_init(&sk);
1114 retry:
1115         bch2_trans_begin(trans);
1116         iter = (struct btree_iter) { NULL };
1117
1118         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
1119         if (ret)
1120                 goto err;
1121
1122         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1123                              SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
1124                              BTREE_ITER_slots);
1125         while (1) {
1126                 unsigned bytes, sectors, offset_into_extent;
1127                 enum btree_id data_btree = BTREE_ID_extents;
1128
1129                 /*
1130                  * read_extent -> io_time_reset may cause a transaction restart
1131                  * without returning an error, we need to check for that here:
1132                  */
1133                 ret = bch2_trans_relock(trans);
1134                 if (ret)
1135                         break;
1136
1137                 bch2_btree_iter_set_pos(&iter,
1138                                 POS(inum.inum, bvec_iter.bi_sector));
1139
1140                 k = bch2_btree_iter_peek_slot(&iter);
1141                 ret = bkey_err(k);
1142                 if (ret)
1143                         break;
1144
1145                 offset_into_extent = iter.pos.offset -
1146                         bkey_start_offset(k.k);
1147                 sectors = k.k->size - offset_into_extent;
1148
1149                 bch2_bkey_buf_reassemble(&sk, c, k);
1150
1151                 ret = bch2_read_indirect_extent(trans, &data_btree,
1152                                         &offset_into_extent, &sk);
1153                 if (ret)
1154                         break;
1155
1156                 k = bkey_i_to_s_c(sk.k);
1157
1158                 /*
1159                  * With indirect extents, the amount of data to read is the min
1160                  * of the original extent and the indirect extent:
1161                  */
1162                 sectors = min(sectors, k.k->size - offset_into_extent);
1163
1164                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
1165                 swap(bvec_iter.bi_size, bytes);
1166
1167                 if (bvec_iter.bi_size == bytes)
1168                         flags |= BCH_READ_LAST_FRAGMENT;
1169
1170                 ret = __bch2_read_extent(trans, rbio, bvec_iter, iter.pos,
1171                                          data_btree, k,
1172                                          offset_into_extent, failed, flags);
1173                 if (ret)
1174                         break;
1175
1176                 if (flags & BCH_READ_LAST_FRAGMENT)
1177                         break;
1178
1179                 swap(bvec_iter.bi_size, bytes);
1180                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
1181
1182                 ret = btree_trans_too_many_iters(trans);
1183                 if (ret)
1184                         break;
1185         }
1186 err:
1187         bch2_trans_iter_exit(trans, &iter);
1188
1189         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1190             ret == READ_RETRY ||
1191             ret == READ_RETRY_AVOID)
1192                 goto retry;
1193
1194         bch2_trans_put(trans);
1195         bch2_bkey_buf_exit(&sk, c);
1196
1197         if (ret) {
1198                 bch_err_inum_offset_ratelimited(c, inum.inum,
1199                                                 bvec_iter.bi_sector << 9,
1200                                                 "read error %i from btree lookup", ret);
1201                 rbio->bio.bi_status = BLK_STS_IOERR;
1202                 bch2_rbio_done(rbio);
1203         }
1204 }
1205
1206 void bch2_fs_io_read_exit(struct bch_fs *c)
1207 {
1208         if (c->promote_table.tbl)
1209                 rhashtable_destroy(&c->promote_table);
1210         bioset_exit(&c->bio_read_split);
1211         bioset_exit(&c->bio_read);
1212 }
1213
1214 int bch2_fs_io_read_init(struct bch_fs *c)
1215 {
1216         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
1217                         BIOSET_NEED_BVECS))
1218                 return -BCH_ERR_ENOMEM_bio_read_init;
1219
1220         if (bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
1221                         BIOSET_NEED_BVECS))
1222                 return -BCH_ERR_ENOMEM_bio_read_split_init;
1223
1224         if (rhashtable_init(&c->promote_table, &bch_promote_params))
1225                 return -BCH_ERR_ENOMEM_promote_table_init;
1226
1227         return 0;
1228 }