git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/io.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 <kent.overstreet@gmail.com>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "alloc_background.h"
  11 #include "alloc_foreground.h"
  12 #include "bkey_buf.h"
  13 #include "bset.h"
  14 #include "btree_update.h"
  15 #include "buckets.h"
  16 #include "checksum.h"
  17 #include "compress.h"
  18 #include "clock.h"
  19 #include "debug.h"
  20 #include "disk_groups.h"
  21 #include "ec.h"
  22 #include "error.h"
  23 #include "extent_update.h"
  24 #include "inode.h"
  25 #include "io.h"
  26 #include "journal.h"
  27 #include "keylist.h"
  28 #include "move.h"
  29 #include "rebalance.h"
  30 #include "subvolume.h"
  31 #include "super.h"
  32 #include "super-io.h"
  33
  34 #include <linux/blkdev.h>
  35 #include <linux/random.h>
  36 #include <linux/sched/mm.h>
  37
  38 #include <trace/events/bcachefs.h>
  39
  40 const char *bch2_blk_status_to_str(blk_status_t status)
  41 {
  42         if (status == BLK_STS_REMOVED)
  43                 return "device removed";
  44         return blk_status_to_str(status);
  45 }
  46
  47 static bool bch2_target_congested(struct bch_fs *c, u16 target)
  48 {
  49         const struct bch_devs_mask *devs;
  50         unsigned d, nr = 0, total = 0;
  51         u64 now = local_clock(), last;
  52         s64 congested;
  53         struct bch_dev *ca;
  54
  55         if (!target)
  56                 return false;
  57
  58         rcu_read_lock();
  59         devs = bch2_target_to_mask(c, target) ?:
  60                 &c->rw_devs[BCH_DATA_user];
  61
  62         for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
  63                 ca = rcu_dereference(c->devs[d]);
  64                 if (!ca)
  65                         continue;
  66
  67                 congested = atomic_read(&ca->congested);
  68                 last = READ_ONCE(ca->congested_last);
  69                 if (time_after64(now, last))
  70                         congested -= (now - last) >> 12;
  71
  72                 total += max(congested, 0LL);
  73                 nr++;
  74         }
  75         rcu_read_unlock();
  76
  77         return bch2_rand_range(nr * CONGESTED_MAX) < total;
  78 }
  79
  80 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
  81                                        u64 now, int rw)
  82 {
  83         u64 latency_capable =
  84                 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
  85         /* ideally we'd be taking into account the device's variance here: */
  86         u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
  87         s64 latency_over = io_latency - latency_threshold;
  88
  89         if (latency_threshold && latency_over > 0) {
  90                 /*
  91                  * bump up congested by approximately latency_over * 4 /
  92                  * latency_threshold - we don't need much accuracy here so don't
  93                  * bother with the divide:
  94                  */
  95                 if (atomic_read(&ca->congested) < CONGESTED_MAX)
  96                         atomic_add(latency_over >>
  97                                    max_t(int, ilog2(latency_threshold) - 2, 0),
  98                                    &ca->congested);
  99
 100                 ca->congested_last = now;
 101         } else if (atomic_read(&ca->congested) > 0) {
 102                 atomic_dec(&ca->congested);
 103         }
 104 }
 105
 106 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
 107 {
 108         atomic64_t *latency = &ca->cur_latency[rw];
 109         u64 now = local_clock();
 110         u64 io_latency = time_after64(now, submit_time)
 111                 ? now - submit_time
 112                 : 0;
 113         u64 old, new, v = atomic64_read(latency);
 114
 115         do {
 116                 old = v;
 117
 118                 /*
 119                  * If the io latency was reasonably close to the current
 120                  * latency, skip doing the update and atomic operation - most of
 121                  * the time:
 122                  */
 123                 if (abs((int) (old - io_latency)) < (old >> 1) &&
 124                     now & ~(~0U << 5))
 125                         break;
 126
 127                 new = ewma_add(old, io_latency, 5);
 128         } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
 129
 130         bch2_congested_acct(ca, io_latency, now, rw);
 131
 132         __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
 133 }
 134
 135 /* Allocate, free from mempool: */
 136
 137 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
 138 {
 139         struct bvec_iter_all iter;
 140         struct bio_vec *bv;
 141
 142         bio_for_each_segment_all(bv, bio, iter)
 143                 if (bv->bv_page != ZERO_PAGE(0))
 144                         mempool_free(bv->bv_page, &c->bio_bounce_pages);
 145         bio->bi_vcnt = 0;
 146 }
 147
 148 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
 149 {
 150         struct page *page;
 151
 152         if (likely(!*using_mempool)) {
 153                 page = alloc_page(GFP_NOIO);
 154                 if (unlikely(!page)) {
 155                         mutex_lock(&c->bio_bounce_pages_lock);
 156                         *using_mempool = true;
 157                         goto pool_alloc;
 158
 159                 }
 160         } else {
 161 pool_alloc:
 162                 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
 163         }
 164
 165         return page;
 166 }
 167
 168 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
 169                                size_t size)
 170 {
 171         bool using_mempool = false;
 172
 173         while (size) {
 174                 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
 175                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 176
 177                 BUG_ON(!bio_add_page(bio, page, len, 0));
 178                 size -= len;
 179         }
 180
 181         if (using_mempool)
 182                 mutex_unlock(&c->bio_bounce_pages_lock);
 183 }
 184
 185 /* Extent update path: */
 186
 187 int bch2_sum_sector_overwrites(struct btree_trans *trans,
 188                                struct btree_iter *extent_iter,
 189                                struct bkey_i *new,
 190                                bool *usage_increasing,
 191                                s64 *i_sectors_delta,
 192                                s64 *disk_sectors_delta)
 193 {
 194         struct bch_fs *c = trans->c;
 195         struct btree_iter iter;
 196         struct bkey_s_c old;
 197         unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
 198         bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
 199         int ret = 0;
 200
 201         *usage_increasing       = false;
 202         *i_sectors_delta        = 0;
 203         *disk_sectors_delta     = 0;
 204
 205         bch2_trans_copy_iter(&iter, extent_iter);
 206
 207         for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) {
 208                 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 209                         max(bkey_start_offset(&new->k),
 210                             bkey_start_offset(old.k));
 211
 212                 *i_sectors_delta += sectors *
 213                         (bkey_extent_is_allocation(&new->k) -
 214                          bkey_extent_is_allocation(old.k));
 215
 216                 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
 217                 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
 218                         ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 219                         : 0;
 220
 221                 if (!*usage_increasing &&
 222                     (new->k.p.snapshot != old.k->p.snapshot ||
 223                      new_replicas > bch2_bkey_replicas(c, old) ||
 224                      (!new_compressed && bch2_bkey_sectors_compressed(old))))
 225                         *usage_increasing = true;
 226
 227                 if (bkey_cmp(old.k->p, new->k.p) >= 0)
 228                         break;
 229         }
 230
 231         bch2_trans_iter_exit(trans, &iter);
 232         return ret;
 233 }
 234
 235 int bch2_extent_update(struct btree_trans *trans,
 236                        subvol_inum inum,
 237                        struct btree_iter *iter,
 238                        struct bkey_i *k,
 239                        struct disk_reservation *disk_res,
 240                        u64 *journal_seq,
 241                        u64 new_i_size,
 242                        s64 *i_sectors_delta_total,
 243                        bool check_enospc)
 244 {
 245         struct btree_iter inode_iter;
 246         struct bch_inode_unpacked inode_u;
 247         struct bpos next_pos;
 248         bool usage_increasing;
 249         s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 250         int ret;
 251
 252         /*
 253          * This traverses us the iterator without changing iter->path->pos to
 254          * search_key() (which is pos + 1 for extents): we want there to be a
 255          * path already traversed at iter->pos because
 256          * bch2_trans_extent_update() will use it to attempt extent merging
 257          */
 258         ret = __bch2_btree_iter_traverse(iter);
 259         if (ret)
 260                 return ret;
 261
 262         ret = bch2_extent_trim_atomic(trans, iter, k);
 263         if (ret)
 264                 return ret;
 265
 266         new_i_size = min(k->k.p.offset << 9, new_i_size);
 267         next_pos = k->k.p;
 268
 269         ret = bch2_sum_sector_overwrites(trans, iter, k,
 270                         &usage_increasing,
 271                         &i_sectors_delta,
 272                         &disk_sectors_delta);
 273         if (ret)
 274                 return ret;
 275
 276         if (disk_res &&
 277             disk_sectors_delta > (s64) disk_res->sectors) {
 278                 ret = bch2_disk_reservation_add(trans->c, disk_res,
 279                                         disk_sectors_delta - disk_res->sectors,
 280                                         !check_enospc || !usage_increasing
 281                                         ? BCH_DISK_RESERVATION_NOFAIL : 0);
 282                 if (ret)
 283                         return ret;
 284         }
 285
 286         ret = bch2_inode_peek(trans, &inode_iter, &inode_u, inum,
 287                               BTREE_ITER_INTENT);
 288         if (ret)
 289                 return ret;
 290
 291         if (!(inode_u.bi_flags & BCH_INODE_I_SIZE_DIRTY) &&
 292             new_i_size > inode_u.bi_size)
 293                 inode_u.bi_size = new_i_size;
 294
 295         inode_u.bi_sectors += i_sectors_delta;
 296
 297         ret =   bch2_trans_update(trans, iter, k, 0) ?:
 298                 bch2_inode_write(trans, &inode_iter, &inode_u) ?:
 299                 bch2_trans_commit(trans, disk_res, journal_seq,
 300                                 BTREE_INSERT_NOCHECK_RW|
 301                                 BTREE_INSERT_NOFAIL);
 302         bch2_trans_iter_exit(trans, &inode_iter);
 303
 304         if (ret)
 305                 return ret;
 306
 307         if (i_sectors_delta_total)
 308                 *i_sectors_delta_total += i_sectors_delta;
 309         bch2_btree_iter_set_pos(iter, next_pos);
 310
 311         return 0;
 312 }
 313
 314 /*
 315  * Returns -EINTR if we had to drop locks:
 316  */
 317 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 318                    subvol_inum inum, u64 end,
 319                    s64 *i_sectors_delta)
 320 {
 321         struct bch_fs *c        = trans->c;
 322         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
 323         struct bpos end_pos = POS(inum.inum, end);
 324         struct bkey_s_c k;
 325         int ret = 0, ret2 = 0;
 326         u32 snapshot;
 327
 328         while (!ret || ret == -EINTR) {
 329                 struct disk_reservation disk_res =
 330                         bch2_disk_reservation_init(c, 0);
 331                 struct bkey_i delete;
 332
 333                 if (ret)
 334                         ret2 = ret;
 335
 336                 bch2_trans_begin(trans);
 337
 338                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 339                 if (ret)
 340                         continue;
 341
 342                 bch2_btree_iter_set_snapshot(iter, snapshot);
 343
 344                 k = bch2_btree_iter_peek(iter);
 345                 if (bkey_cmp(iter->pos, end_pos) >= 0) {
 346                         bch2_btree_iter_set_pos(iter, end_pos);
 347                         break;
 348                 }
 349
 350                 ret = bkey_err(k);
 351                 if (ret)
 352                         continue;
 353
 354                 bkey_init(&delete.k);
 355                 delete.k.p = iter->pos;
 356
 357                 /* create the biggest key we can */
 358                 bch2_key_resize(&delete.k, max_sectors);
 359                 bch2_cut_back(end_pos, &delete);
 360
 361                 ret = bch2_extent_update(trans, inum, iter, &delete,
 362                                 &disk_res, NULL,
 363                                 0, i_sectors_delta, false);
 364                 bch2_disk_reservation_put(c, &disk_res);
 365         }
 366
 367         return ret ?: ret2;
 368 }
 369
 370 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
 371                 s64 *i_sectors_delta)
 372 {
 373         struct btree_trans trans;
 374         struct btree_iter iter;
 375         int ret;
 376
 377         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 378         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 379                              POS(inum.inum, start),
 380                              BTREE_ITER_INTENT);
 381
 382         ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
 383
 384         bch2_trans_iter_exit(&trans, &iter);
 385         bch2_trans_exit(&trans);
 386
 387         return ret == -EINTR ? 0 : ret;
 388 }
 389
 390 int bch2_write_index_default(struct bch_write_op *op)
 391 {
 392         struct bch_fs *c = op->c;
 393         struct bkey_buf sk;
 394         struct open_bucket *ec_ob = ec_open_bucket(c, &op->open_buckets);
 395         struct keylist *keys = &op->insert_keys;
 396         struct bkey_i *k = bch2_keylist_front(keys);
 397         struct btree_trans trans;
 398         struct btree_iter iter;
 399         subvol_inum inum = {
 400                 .subvol = op->subvol,
 401                 .inum   = k->k.p.inode,
 402         };
 403         int ret;
 404
 405         BUG_ON(!inum.subvol);
 406
 407         bch2_bkey_buf_init(&sk);
 408         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 409
 410         do {
 411                 bch2_trans_begin(&trans);
 412
 413                 k = bch2_keylist_front(keys);
 414                 bch2_bkey_buf_copy(&sk, c, k);
 415
 416                 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
 417                                                   &sk.k->k.p.snapshot);
 418                 if (ret == -EINTR)
 419                         continue;
 420                 if (ret)
 421                         break;
 422
 423                 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 424                                      bkey_start_pos(&sk.k->k),
 425                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 426
 427                 ret = bch2_extent_update(&trans, inum, &iter, sk.k,
 428                                          &op->res, op_journal_seq(op),
 429                                          op->new_i_size, &op->i_sectors_delta,
 430                                          op->flags & BCH_WRITE_CHECK_ENOSPC);
 431                 bch2_trans_iter_exit(&trans, &iter);
 432
 433                 if (ret == -EINTR)
 434                         continue;
 435                 if (ret)
 436                         break;
 437
 438                 if (ec_ob)
 439                         bch2_ob_add_backpointer(c, ec_ob, &sk.k->k);
 440
 441                 if (bkey_cmp(iter.pos, k->k.p) >= 0)
 442                         bch2_keylist_pop_front(&op->insert_keys);
 443                 else
 444                         bch2_cut_front(iter.pos, k);
 445         } while (!bch2_keylist_empty(keys));
 446
 447         bch2_trans_exit(&trans);
 448         bch2_bkey_buf_exit(&sk, c);
 449
 450         return ret;
 451 }
 452
 453 /* Writes */
 454
 455 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 456                                enum bch_data_type type,
 457                                const struct bkey_i *k)
 458 {
 459         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 460         const struct bch_extent_ptr *ptr;
 461         struct bch_write_bio *n;
 462         struct bch_dev *ca;
 463
 464         BUG_ON(c->opts.nochanges);
 465
 466         bkey_for_each_ptr(ptrs, ptr) {
 467                 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
 468                        !c->devs[ptr->dev]);
 469
 470                 ca = bch_dev_bkey_exists(c, ptr->dev);
 471
 472                 if (to_entry(ptr + 1) < ptrs.end) {
 473                         n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
 474                                                 GFP_NOIO, &ca->replica_set));
 475
 476                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 477                         n->bio.bi_private       = wbio->bio.bi_private;
 478                         n->parent               = wbio;
 479                         n->split                = true;
 480                         n->bounce               = false;
 481                         n->put_bio              = true;
 482                         n->bio.bi_opf           = wbio->bio.bi_opf;
 483                         bio_inc_remaining(&wbio->bio);
 484                 } else {
 485                         n = wbio;
 486                         n->split                = false;
 487                 }
 488
 489                 n->c                    = c;
 490                 n->dev                  = ptr->dev;
 491                 n->have_ioref           = bch2_dev_get_ioref(ca,
 492                                         type == BCH_DATA_btree ? READ : WRITE);
 493                 n->submit_time          = local_clock();
 494                 n->bio.bi_iter.bi_sector = ptr->offset;
 495
 496                 if (likely(n->have_ioref)) {
 497                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 498                                      bio_sectors(&n->bio));
 499
 500                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 501                         submit_bio(&n->bio);
 502                 } else {
 503                         n->bio.bi_status        = BLK_STS_REMOVED;
 504                         bio_endio(&n->bio);
 505                 }
 506         }
 507 }
 508
 509 static void __bch2_write(struct closure *);
 510
 511 static void bch2_write_done(struct closure *cl)
 512 {
 513         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 514         struct bch_fs *c = op->c;
 515
 516         if (!op->error && (op->flags & BCH_WRITE_FLUSH))
 517                 op->error = bch2_journal_error(&c->journal);
 518
 519         bch2_disk_reservation_put(c, &op->res);
 520         percpu_ref_put(&c->writes);
 521         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 522
 523         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 524
 525         if (op->end_io) {
 526                 EBUG_ON(cl->parent);
 527                 closure_debug_destroy(cl);
 528                 op->end_io(op);
 529         } else {
 530                 closure_return(cl);
 531         }
 532 }
 533
 534 /**
 535  * bch_write_index - after a write, update index to point to new data
 536  */
 537 static void __bch2_write_index(struct bch_write_op *op)
 538 {
 539         struct bch_fs *c = op->c;
 540         struct keylist *keys = &op->insert_keys;
 541         struct bch_extent_ptr *ptr;
 542         struct bkey_i *src, *dst = keys->keys, *n, *k;
 543         unsigned dev;
 544         int ret;
 545
 546         for (src = keys->keys; src != keys->top; src = n) {
 547                 n = bkey_next(src);
 548
 549                 if (bkey_extent_is_direct_data(&src->k)) {
 550                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 551                                             test_bit(ptr->dev, op->failed.d));
 552
 553                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src))) {
 554                                 ret = -EIO;
 555                                 goto err;
 556                         }
 557                 }
 558
 559                 if (dst != src)
 560                         memmove_u64s_down(dst, src, src->u64s);
 561                 dst = bkey_next(dst);
 562         }
 563
 564         keys->top = dst;
 565
 566         /*
 567          * probably not the ideal place to hook this in, but I don't
 568          * particularly want to plumb io_opts all the way through the btree
 569          * update stack right now
 570          */
 571         for_each_keylist_key(keys, k) {
 572                 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
 573
 574                 if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
 575                         bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
 576
 577         }
 578
 579         if (!bch2_keylist_empty(keys)) {
 580                 u64 sectors_start = keylist_sectors(keys);
 581                 int ret = op->index_update_fn(op);
 582
 583                 BUG_ON(ret == -EINTR);
 584                 BUG_ON(keylist_sectors(keys) && !ret);
 585
 586                 op->written += sectors_start - keylist_sectors(keys);
 587
 588                 if (ret) {
 589                         bch_err_inum_ratelimited(c, op->pos.inode,
 590                                 "write error %i from btree update", ret);
 591                         op->error = ret;
 592                 }
 593         }
 594 out:
 595         /* If some a bucket wasn't written, we can't erasure code it: */
 596         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 597                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 598
 599         bch2_open_buckets_put(c, &op->open_buckets);
 600         return;
 601 err:
 602         keys->top = keys->keys;
 603         op->error = ret;
 604         goto out;
 605 }
 606
 607 static void bch2_write_index(struct closure *cl)
 608 {
 609         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 610         struct bch_fs *c = op->c;
 611
 612         __bch2_write_index(op);
 613
 614         if (!(op->flags & BCH_WRITE_DONE)) {
 615                 continue_at(cl, __bch2_write, index_update_wq(op));
 616         } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
 617                 bch2_journal_flush_seq_async(&c->journal,
 618                                              *op_journal_seq(op),
 619                                              cl);
 620                 continue_at(cl, bch2_write_done, index_update_wq(op));
 621         } else {
 622                 continue_at_nobarrier(cl, bch2_write_done, NULL);
 623         }
 624 }
 625
 626 static void bch2_write_endio(struct bio *bio)
 627 {
 628         struct closure *cl              = bio->bi_private;
 629         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 630         struct bch_write_bio *wbio      = to_wbio(bio);
 631         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 632         struct bch_fs *c                = wbio->c;
 633         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
 634
 635         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
 636                                     op->pos.inode,
 637                                     op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
 638                                     "data write error: %s",
 639                                bch2_blk_status_to_str(bio->bi_status)))
 640                 set_bit(wbio->dev, op->failed.d);
 641
 642         if (wbio->have_ioref) {
 643                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 644                 percpu_ref_put(&ca->io_ref);
 645         }
 646
 647         if (wbio->bounce)
 648                 bch2_bio_free_pages_pool(c, bio);
 649
 650         if (wbio->put_bio)
 651                 bio_put(bio);
 652
 653         if (parent)
 654                 bio_endio(&parent->bio);
 655         else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
 656                 closure_put(cl);
 657         else
 658                 continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
 659 }
 660
 661 static void init_append_extent(struct bch_write_op *op,
 662                                struct write_point *wp,
 663                                struct bversion version,
 664                                struct bch_extent_crc_unpacked crc)
 665 {
 666         struct bch_fs *c = op->c;
 667         struct bkey_i_extent *e;
 668
 669         op->pos.offset += crc.uncompressed_size;
 670
 671         e = bkey_extent_init(op->insert_keys.top);
 672         e->k.p          = op->pos;
 673         e->k.size       = crc.uncompressed_size;
 674         e->k.version    = version;
 675
 676         if (crc.csum_type ||
 677             crc.compression_type ||
 678             crc.nonce)
 679                 bch2_extent_crc_append(&e->k_i, crc);
 680
 681         bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, crc.compressed_size,
 682                                        op->flags & BCH_WRITE_CACHED);
 683
 684         bch2_keylist_push(&op->insert_keys);
 685 }
 686
 687 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 688                                         struct write_point *wp,
 689                                         struct bio *src,
 690                                         bool *page_alloc_failed,
 691                                         void *buf)
 692 {
 693         struct bch_write_bio *wbio;
 694         struct bio *bio;
 695         unsigned output_available =
 696                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 697         unsigned pages = DIV_ROUND_UP(output_available +
 698                                       (buf
 699                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 700                                        : 0), PAGE_SIZE);
 701
 702         pages = min(pages, BIO_MAX_VECS);
 703
 704         bio = bio_alloc_bioset(NULL, pages, 0,
 705                                GFP_NOIO, &c->bio_write);
 706         wbio                    = wbio_init(bio);
 707         wbio->put_bio           = true;
 708         /* copy WRITE_SYNC flag */
 709         wbio->bio.bi_opf        = src->bi_opf;
 710
 711         if (buf) {
 712                 bch2_bio_map(bio, buf, output_available);
 713                 return bio;
 714         }
 715
 716         wbio->bounce            = true;
 717
 718         /*
 719          * We can't use mempool for more than c->sb.encoded_extent_max
 720          * worth of pages, but we'd like to allocate more if we can:
 721          */
 722         bch2_bio_alloc_pages_pool(c, bio,
 723                                   min_t(unsigned, output_available,
 724                                         c->opts.encoded_extent_max));
 725
 726         if (bio->bi_iter.bi_size < output_available)
 727                 *page_alloc_failed =
 728                         bch2_bio_alloc_pages(bio,
 729                                              output_available -
 730                                              bio->bi_iter.bi_size,
 731                                              GFP_NOFS) != 0;
 732
 733         return bio;
 734 }
 735
 736 static int bch2_write_rechecksum(struct bch_fs *c,
 737                                  struct bch_write_op *op,
 738                                  unsigned new_csum_type)
 739 {
 740         struct bio *bio = &op->wbio.bio;
 741         struct bch_extent_crc_unpacked new_crc;
 742         int ret;
 743
 744         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 745
 746         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 747             bch2_csum_type_is_encryption(new_csum_type))
 748                 new_csum_type = op->crc.csum_type;
 749
 750         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 751                                   NULL, &new_crc,
 752                                   op->crc.offset, op->crc.live_size,
 753                                   new_csum_type);
 754         if (ret)
 755                 return ret;
 756
 757         bio_advance(bio, op->crc.offset << 9);
 758         bio->bi_iter.bi_size = op->crc.live_size << 9;
 759         op->crc = new_crc;
 760         return 0;
 761 }
 762
 763 static int bch2_write_decrypt(struct bch_write_op *op)
 764 {
 765         struct bch_fs *c = op->c;
 766         struct nonce nonce = extent_nonce(op->version, op->crc);
 767         struct bch_csum csum;
 768         int ret;
 769
 770         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 771                 return 0;
 772
 773         /*
 774          * If we need to decrypt data in the write path, we'll no longer be able
 775          * to verify the existing checksum (poly1305 mac, in this case) after
 776          * it's decrypted - this is the last point we'll be able to reverify the
 777          * checksum:
 778          */
 779         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 780         if (bch2_crc_cmp(op->crc.csum, csum))
 781                 return -EIO;
 782
 783         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 784         op->crc.csum_type = 0;
 785         op->crc.csum = (struct bch_csum) { 0, 0 };
 786         return ret;
 787 }
 788
 789 static enum prep_encoded_ret {
 790         PREP_ENCODED_OK,
 791         PREP_ENCODED_ERR,
 792         PREP_ENCODED_CHECKSUM_ERR,
 793         PREP_ENCODED_DO_WRITE,
 794 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 795 {
 796         struct bch_fs *c = op->c;
 797         struct bio *bio = &op->wbio.bio;
 798
 799         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 800                 return PREP_ENCODED_OK;
 801
 802         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 803
 804         /* Can we just write the entire extent as is? */
 805         if (op->crc.uncompressed_size == op->crc.live_size &&
 806             op->crc.compressed_size <= wp->sectors_free &&
 807             (op->crc.compression_type == op->compression_type ||
 808              op->incompressible)) {
 809                 if (!crc_is_compressed(op->crc) &&
 810                     op->csum_type != op->crc.csum_type &&
 811                     bch2_write_rechecksum(c, op, op->csum_type))
 812                         return PREP_ENCODED_CHECKSUM_ERR;
 813
 814                 return PREP_ENCODED_DO_WRITE;
 815         }
 816
 817         /*
 818          * If the data is compressed and we couldn't write the entire extent as
 819          * is, we have to decompress it:
 820          */
 821         if (crc_is_compressed(op->crc)) {
 822                 struct bch_csum csum;
 823
 824                 if (bch2_write_decrypt(op))
 825                         return PREP_ENCODED_CHECKSUM_ERR;
 826
 827                 /* Last point we can still verify checksum: */
 828                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 829                                          extent_nonce(op->version, op->crc),
 830                                          bio);
 831                 if (bch2_crc_cmp(op->crc.csum, csum))
 832                         return PREP_ENCODED_CHECKSUM_ERR;
 833
 834                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 835                         return PREP_ENCODED_ERR;
 836         }
 837
 838         /*
 839          * No longer have compressed data after this point - data might be
 840          * encrypted:
 841          */
 842
 843         /*
 844          * If the data is checksummed and we're only writing a subset,
 845          * rechecksum and adjust bio to point to currently live data:
 846          */
 847         if ((op->crc.live_size != op->crc.uncompressed_size ||
 848              op->crc.csum_type != op->csum_type) &&
 849             bch2_write_rechecksum(c, op, op->csum_type))
 850                 return PREP_ENCODED_CHECKSUM_ERR;
 851
 852         /*
 853          * If we want to compress the data, it has to be decrypted:
 854          */
 855         if ((op->compression_type ||
 856              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 857              bch2_csum_type_is_encryption(op->csum_type)) &&
 858             bch2_write_decrypt(op))
 859                 return PREP_ENCODED_CHECKSUM_ERR;
 860
 861         return PREP_ENCODED_OK;
 862 }
 863
 864 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 865                              struct bio **_dst)
 866 {
 867         struct bch_fs *c = op->c;
 868         struct bio *src = &op->wbio.bio, *dst = src;
 869         struct bvec_iter saved_iter;
 870         void *ec_buf;
 871         unsigned total_output = 0, total_input = 0;
 872         bool bounce = false;
 873         bool page_alloc_failed = false;
 874         int ret, more = 0;
 875
 876         BUG_ON(!bio_sectors(src));
 877
 878         ec_buf = bch2_writepoint_ec_buf(c, wp);
 879
 880         switch (bch2_write_prep_encoded_data(op, wp)) {
 881         case PREP_ENCODED_OK:
 882                 break;
 883         case PREP_ENCODED_ERR:
 884                 ret = -EIO;
 885                 goto err;
 886         case PREP_ENCODED_CHECKSUM_ERR:
 887                 goto csum_err;
 888         case PREP_ENCODED_DO_WRITE:
 889                 /* XXX look for bug here */
 890                 if (ec_buf) {
 891                         dst = bch2_write_bio_alloc(c, wp, src,
 892                                                    &page_alloc_failed,
 893                                                    ec_buf);
 894                         bio_copy_data(dst, src);
 895                         bounce = true;
 896                 }
 897                 init_append_extent(op, wp, op->version, op->crc);
 898                 goto do_write;
 899         }
 900
 901         if (ec_buf ||
 902             op->compression_type ||
 903             (op->csum_type &&
 904              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 905             (bch2_csum_type_is_encryption(op->csum_type) &&
 906              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 907                 dst = bch2_write_bio_alloc(c, wp, src,
 908                                            &page_alloc_failed,
 909                                            ec_buf);
 910                 bounce = true;
 911         }
 912
 913         saved_iter = dst->bi_iter;
 914
 915         do {
 916                 struct bch_extent_crc_unpacked crc =
 917                         (struct bch_extent_crc_unpacked) { 0 };
 918                 struct bversion version = op->version;
 919                 size_t dst_len, src_len;
 920
 921                 if (page_alloc_failed &&
 922                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 923                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 924                         break;
 925
 926                 BUG_ON(op->compression_type &&
 927                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 928                        bch2_csum_type_is_encryption(op->crc.csum_type));
 929                 BUG_ON(op->compression_type && !bounce);
 930
 931                 crc.compression_type = op->incompressible
 932                         ? BCH_COMPRESSION_TYPE_incompressible
 933                         : op->compression_type
 934                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 935                                             op->compression_type)
 936                         : 0;
 937                 if (!crc_is_compressed(crc)) {
 938                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 939                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 940
 941                         if (op->csum_type)
 942                                 dst_len = min_t(unsigned, dst_len,
 943                                                 c->opts.encoded_extent_max);
 944
 945                         if (bounce) {
 946                                 swap(dst->bi_iter.bi_size, dst_len);
 947                                 bio_copy_data(dst, src);
 948                                 swap(dst->bi_iter.bi_size, dst_len);
 949                         }
 950
 951                         src_len = dst_len;
 952                 }
 953
 954                 BUG_ON(!src_len || !dst_len);
 955
 956                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 957                         if (bversion_zero(version)) {
 958                                 version.lo = atomic64_inc_return(&c->key_version);
 959                         } else {
 960                                 crc.nonce = op->nonce;
 961                                 op->nonce += src_len >> 9;
 962                         }
 963                 }
 964
 965                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 966                     !crc_is_compressed(crc) &&
 967                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
 968                     bch2_csum_type_is_encryption(op->csum_type)) {
 969                         /*
 970                          * Note: when we're using rechecksum(), we need to be
 971                          * checksumming @src because it has all the data our
 972                          * existing checksum covers - if we bounced (because we
 973                          * were trying to compress), @dst will only have the
 974                          * part of the data the new checksum will cover.
 975                          *
 976                          * But normally we want to be checksumming post bounce,
 977                          * because part of the reason for bouncing is so the
 978                          * data can't be modified (by userspace) while it's in
 979                          * flight.
 980                          */
 981                         if (bch2_rechecksum_bio(c, src, version, op->crc,
 982                                         &crc, &op->crc,
 983                                         src_len >> 9,
 984                                         bio_sectors(src) - (src_len >> 9),
 985                                         op->csum_type))
 986                                 goto csum_err;
 987                 } else {
 988                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 989                             bch2_rechecksum_bio(c, src, version, op->crc,
 990                                         NULL, &op->crc,
 991                                         src_len >> 9,
 992                                         bio_sectors(src) - (src_len >> 9),
 993                                         op->crc.csum_type))
 994                                 goto csum_err;
 995
 996                         crc.compressed_size     = dst_len >> 9;
 997                         crc.uncompressed_size   = src_len >> 9;
 998                         crc.live_size           = src_len >> 9;
 999
1000                         swap(dst->bi_iter.bi_size, dst_len);
1001                         ret = bch2_encrypt_bio(c, op->csum_type,
1002                                                extent_nonce(version, crc), dst);
1003                         if (ret)
1004                                 goto err;
1005
1006                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1007                                          extent_nonce(version, crc), dst);
1008                         crc.csum_type = op->csum_type;
1009                         swap(dst->bi_iter.bi_size, dst_len);
1010                 }
1011
1012                 init_append_extent(op, wp, version, crc);
1013
1014                 if (dst != src)
1015                         bio_advance(dst, dst_len);
1016                 bio_advance(src, src_len);
1017                 total_output    += dst_len;
1018                 total_input     += src_len;
1019         } while (dst->bi_iter.bi_size &&
1020                  src->bi_iter.bi_size &&
1021                  wp->sectors_free &&
1022                  !bch2_keylist_realloc(&op->insert_keys,
1023                                       op->inline_keys,
1024                                       ARRAY_SIZE(op->inline_keys),
1025                                       BKEY_EXTENT_U64s_MAX));
1026
1027         more = src->bi_iter.bi_size != 0;
1028
1029         dst->bi_iter = saved_iter;
1030
1031         if (dst == src && more) {
1032                 BUG_ON(total_output != total_input);
1033
1034                 dst = bio_split(src, total_input >> 9,
1035                                 GFP_NOIO, &c->bio_write);
1036                 wbio_init(dst)->put_bio = true;
1037                 /* copy WRITE_SYNC flag */
1038                 dst->bi_opf             = src->bi_opf;
1039         }
1040
1041         dst->bi_iter.bi_size = total_output;
1042 do_write:
1043         *_dst = dst;
1044         return more;
1045 csum_err:
1046         bch_err(c, "error verifying existing checksum while "
1047                 "rewriting existing data (memory corruption?)");
1048         ret = -EIO;
1049 err:
1050         if (to_wbio(dst)->bounce)
1051                 bch2_bio_free_pages_pool(c, dst);
1052         if (to_wbio(dst)->put_bio)
1053                 bio_put(dst);
1054
1055         return ret;
1056 }
1057
1058 static void __bch2_write(struct closure *cl)
1059 {
1060         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1061         struct bch_fs *c = op->c;
1062         struct write_point *wp;
1063         struct bio *bio = NULL;
1064         bool skip_put = true;
1065         unsigned nofs_flags;
1066         int ret;
1067
1068         nofs_flags = memalloc_nofs_save();
1069 again:
1070         memset(&op->failed, 0, sizeof(op->failed));
1071
1072         do {
1073                 struct bkey_i *key_to_write;
1074                 unsigned key_to_write_offset = op->insert_keys.top_p -
1075                         op->insert_keys.keys_p;
1076
1077                 /* +1 for possible cache device: */
1078                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1079                     ARRAY_SIZE(op->open_buckets.v))
1080                         goto flush_io;
1081
1082                 if (bch2_keylist_realloc(&op->insert_keys,
1083                                         op->inline_keys,
1084                                         ARRAY_SIZE(op->inline_keys),
1085                                         BKEY_EXTENT_U64s_MAX))
1086                         goto flush_io;
1087
1088                 if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
1089                     percpu_ref_is_dying(&c->writes)) {
1090                         ret = -EROFS;
1091                         goto err;
1092                 }
1093
1094                 /*
1095                  * The copygc thread is now global, which means it's no longer
1096                  * freeing up space on specific disks, which means that
1097                  * allocations for specific disks may hang arbitrarily long:
1098                  */
1099                 wp = bch2_alloc_sectors_start(c,
1100                         op->target,
1101                         op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1102                         op->write_point,
1103                         &op->devs_have,
1104                         op->nr_replicas,
1105                         op->nr_replicas_required,
1106                         op->alloc_reserve,
1107                         op->flags,
1108                         (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1109                                       BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1110                 EBUG_ON(!wp);
1111
1112                 if (unlikely(IS_ERR(wp))) {
1113                         if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1114                                 ret = PTR_ERR(wp);
1115                                 goto err;
1116                         }
1117
1118                         goto flush_io;
1119                 }
1120
1121                 /*
1122                  * It's possible for the allocator to fail, put us on the
1123                  * freelist waitlist, and then succeed in one of various retry
1124                  * paths: if that happens, we need to disable the skip_put
1125                  * optimization because otherwise there won't necessarily be a
1126                  * barrier before we free the bch_write_op:
1127                  */
1128                 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1129                         skip_put = false;
1130
1131                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1132                 ret = bch2_write_extent(op, wp, &bio);
1133                 bch2_alloc_sectors_done(c, wp);
1134
1135                 if (ret < 0)
1136                         goto err;
1137
1138                 if (ret) {
1139                         skip_put = false;
1140                 } else {
1141                         /*
1142                          * for the skip_put optimization this has to be set
1143                          * before we submit the bio:
1144                          */
1145                         op->flags |= BCH_WRITE_DONE;
1146                 }
1147
1148                 bio->bi_end_io  = bch2_write_endio;
1149                 bio->bi_private = &op->cl;
1150                 bio->bi_opf |= REQ_OP_WRITE;
1151
1152                 if (!skip_put)
1153                         closure_get(bio->bi_private);
1154                 else
1155                         op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1156
1157                 key_to_write = (void *) (op->insert_keys.keys_p +
1158                                          key_to_write_offset);
1159
1160                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1161                                           key_to_write);
1162         } while (ret);
1163
1164         if (!skip_put)
1165                 continue_at(cl, bch2_write_index, index_update_wq(op));
1166 out:
1167         memalloc_nofs_restore(nofs_flags);
1168         return;
1169 err:
1170         op->error = ret;
1171         op->flags |= BCH_WRITE_DONE;
1172
1173         continue_at(cl, bch2_write_index, index_update_wq(op));
1174         goto out;
1175 flush_io:
1176         /*
1177          * If the write can't all be submitted at once, we generally want to
1178          * block synchronously as that signals backpressure to the caller.
1179          *
1180          * However, if we're running out of a workqueue, we can't block here
1181          * because we'll be blocking other work items from completing:
1182          */
1183         if (current->flags & PF_WQ_WORKER) {
1184                 continue_at(cl, bch2_write_index, index_update_wq(op));
1185                 goto out;
1186         }
1187
1188         closure_sync(cl);
1189
1190         if (!bch2_keylist_empty(&op->insert_keys)) {
1191                 __bch2_write_index(op);
1192
1193                 if (op->error) {
1194                         op->flags |= BCH_WRITE_DONE;
1195                         continue_at_nobarrier(cl, bch2_write_done, NULL);
1196                         goto out;
1197                 }
1198         }
1199
1200         goto again;
1201 }
1202
1203 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1204 {
1205         struct closure *cl = &op->cl;
1206         struct bio *bio = &op->wbio.bio;
1207         struct bvec_iter iter;
1208         struct bkey_i_inline_data *id;
1209         unsigned sectors;
1210         int ret;
1211
1212         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1213
1214         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1215                                    ARRAY_SIZE(op->inline_keys),
1216                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1217         if (ret) {
1218                 op->error = ret;
1219                 goto err;
1220         }
1221
1222         sectors = bio_sectors(bio);
1223         op->pos.offset += sectors;
1224
1225         id = bkey_inline_data_init(op->insert_keys.top);
1226         id->k.p         = op->pos;
1227         id->k.version   = op->version;
1228         id->k.size      = sectors;
1229
1230         iter = bio->bi_iter;
1231         iter.bi_size = data_len;
1232         memcpy_from_bio(id->v.data, bio, iter);
1233
1234         while (data_len & 7)
1235                 id->v.data[data_len++] = '\0';
1236         set_bkey_val_bytes(&id->k, data_len);
1237         bch2_keylist_push(&op->insert_keys);
1238
1239         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1240         op->flags |= BCH_WRITE_DONE;
1241
1242         continue_at_nobarrier(cl, bch2_write_index, NULL);
1243         return;
1244 err:
1245         bch2_write_done(&op->cl);
1246 }
1247
1248 /**
1249  * bch_write - handle a write to a cache device or flash only volume
1250  *
1251  * This is the starting point for any data to end up in a cache device; it could
1252  * be from a normal write, or a writeback write, or a write to a flash only
1253  * volume - it's also used by the moving garbage collector to compact data in
1254  * mostly empty buckets.
1255  *
1256  * It first writes the data to the cache, creating a list of keys to be inserted
1257  * (if the data won't fit in a single open bucket, there will be multiple keys);
1258  * after the data is written it calls bch_journal, and after the keys have been
1259  * added to the next journal write they're inserted into the btree.
1260  *
1261  * If op->discard is true, instead of inserting the data it invalidates the
1262  * region of the cache represented by op->bio and op->inode.
1263  */
1264 void bch2_write(struct closure *cl)
1265 {
1266         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1267         struct bio *bio = &op->wbio.bio;
1268         struct bch_fs *c = op->c;
1269         unsigned data_len;
1270
1271         BUG_ON(!op->nr_replicas);
1272         BUG_ON(!op->write_point.v);
1273         BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1274
1275         op->start_time = local_clock();
1276         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1277         wbio_init(bio)->put_bio = false;
1278
1279         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1280                 bch_err_inum_ratelimited(c, op->pos.inode,
1281                                          "misaligned write");
1282                 op->error = -EIO;
1283                 goto err;
1284         }
1285
1286         if (c->opts.nochanges ||
1287             !percpu_ref_tryget(&c->writes)) {
1288                 op->error = -EROFS;
1289                 goto err;
1290         }
1291
1292         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1293         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1294
1295         data_len = min_t(u64, bio->bi_iter.bi_size,
1296                          op->new_i_size - (op->pos.offset << 9));
1297
1298         if (c->opts.inline_data &&
1299             data_len <= min(block_bytes(c) / 2, 1024U)) {
1300                 bch2_write_data_inline(op, data_len);
1301                 return;
1302         }
1303
1304         continue_at_nobarrier(cl, __bch2_write, NULL);
1305         return;
1306 err:
1307         bch2_disk_reservation_put(c, &op->res);
1308
1309         if (op->end_io) {
1310                 EBUG_ON(cl->parent);
1311                 closure_debug_destroy(cl);
1312                 op->end_io(op);
1313         } else {
1314                 closure_return(cl);
1315         }
1316 }
1317
1318 /* Cache promotion on read */
1319
1320 struct promote_op {
1321         struct closure          cl;
1322         struct rcu_head         rcu;
1323         u64                     start_time;
1324
1325         struct rhash_head       hash;
1326         struct bpos             pos;
1327
1328         struct migrate_write    write;
1329         struct bio_vec          bi_inline_vecs[0]; /* must be last */
1330 };
1331
1332 static const struct rhashtable_params bch_promote_params = {
1333         .head_offset    = offsetof(struct promote_op, hash),
1334         .key_offset     = offsetof(struct promote_op, pos),
1335         .key_len        = sizeof(struct bpos),
1336 };
1337
1338 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1339                                   struct bpos pos,
1340                                   struct bch_io_opts opts,
1341                                   unsigned flags)
1342 {
1343         if (!(flags & BCH_READ_MAY_PROMOTE))
1344                 return false;
1345
1346         if (!opts.promote_target)
1347                 return false;
1348
1349         if (bch2_bkey_has_target(c, k, opts.promote_target))
1350                 return false;
1351
1352         if (bch2_target_congested(c, opts.promote_target)) {
1353                 /* XXX trace this */
1354                 return false;
1355         }
1356
1357         if (rhashtable_lookup_fast(&c->promote_table, &pos,
1358                                    bch_promote_params))
1359                 return false;
1360
1361         return true;
1362 }
1363
1364 static void promote_free(struct bch_fs *c, struct promote_op *op)
1365 {
1366         int ret;
1367
1368         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1369                                      bch_promote_params);
1370         BUG_ON(ret);
1371         percpu_ref_put(&c->writes);
1372         kfree_rcu(op, rcu);
1373 }
1374
1375 static void promote_done(struct closure *cl)
1376 {
1377         struct promote_op *op =
1378                 container_of(cl, struct promote_op, cl);
1379         struct bch_fs *c = op->write.op.c;
1380
1381         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1382                                op->start_time);
1383
1384         bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1385         promote_free(c, op);
1386 }
1387
1388 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1389 {
1390         struct bch_fs *c = rbio->c;
1391         struct closure *cl = &op->cl;
1392         struct bio *bio = &op->write.op.wbio.bio;
1393
1394         trace_promote(&rbio->bio);
1395
1396         /* we now own pages: */
1397         BUG_ON(!rbio->bounce);
1398         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1399
1400         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1401                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1402         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1403
1404         bch2_migrate_read_done(&op->write, rbio);
1405
1406         closure_init(cl, NULL);
1407         closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
1408         closure_return_with_destructor(cl, promote_done);
1409 }
1410
1411 static struct promote_op *__promote_alloc(struct bch_fs *c,
1412                                           enum btree_id btree_id,
1413                                           struct bkey_s_c k,
1414                                           struct bpos pos,
1415                                           struct extent_ptr_decoded *pick,
1416                                           struct bch_io_opts opts,
1417                                           unsigned sectors,
1418                                           struct bch_read_bio **rbio)
1419 {
1420         struct promote_op *op = NULL;
1421         struct bio *bio;
1422         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1423         int ret;
1424
1425         if (!percpu_ref_tryget(&c->writes))
1426                 return NULL;
1427
1428         op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1429         if (!op)
1430                 goto err;
1431
1432         op->start_time = local_clock();
1433         op->pos = pos;
1434
1435         /*
1436          * We don't use the mempool here because extents that aren't
1437          * checksummed or compressed can be too big for the mempool:
1438          */
1439         *rbio = kzalloc(sizeof(struct bch_read_bio) +
1440                         sizeof(struct bio_vec) * pages,
1441                         GFP_NOIO);
1442         if (!*rbio)
1443                 goto err;
1444
1445         rbio_init(&(*rbio)->bio, opts);
1446         bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
1447
1448         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1449                                  GFP_NOIO))
1450                 goto err;
1451
1452         (*rbio)->bounce         = true;
1453         (*rbio)->split          = true;
1454         (*rbio)->kmalloc        = true;
1455
1456         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1457                                           bch_promote_params))
1458                 goto err;
1459
1460         bio = &op->write.op.wbio.bio;
1461         bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
1462
1463         ret = bch2_migrate_write_init(c, &op->write,
1464                         writepoint_hashed((unsigned long) current),
1465                         opts,
1466                         DATA_PROMOTE,
1467                         (struct data_opts) {
1468                                 .target         = opts.promote_target,
1469                                 .nr_replicas    = 1,
1470                         },
1471                         btree_id, k);
1472         BUG_ON(ret);
1473
1474         return op;
1475 err:
1476         if (*rbio)
1477                 bio_free_pages(&(*rbio)->bio);
1478         kfree(*rbio);
1479         *rbio = NULL;
1480         kfree(op);
1481         percpu_ref_put(&c->writes);
1482         return NULL;
1483 }
1484
1485 noinline
1486 static struct promote_op *promote_alloc(struct bch_fs *c,
1487                                                struct bvec_iter iter,
1488                                                struct bkey_s_c k,
1489                                                struct extent_ptr_decoded *pick,
1490                                                struct bch_io_opts opts,
1491                                                unsigned flags,
1492                                                struct bch_read_bio **rbio,
1493                                                bool *bounce,
1494                                                bool *read_full)
1495 {
1496         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1497         /* data might have to be decompressed in the write path: */
1498         unsigned sectors = promote_full
1499                 ? max(pick->crc.compressed_size, pick->crc.live_size)
1500                 : bvec_iter_sectors(iter);
1501         struct bpos pos = promote_full
1502                 ? bkey_start_pos(k.k)
1503                 : POS(k.k->p.inode, iter.bi_sector);
1504         struct promote_op *promote;
1505
1506         if (!should_promote(c, k, pos, opts, flags))
1507                 return NULL;
1508
1509         promote = __promote_alloc(c,
1510                                   k.k->type == KEY_TYPE_reflink_v
1511                                   ? BTREE_ID_reflink
1512                                   : BTREE_ID_extents,
1513                                   k, pos, pick, opts, sectors, rbio);
1514         if (!promote)
1515                 return NULL;
1516
1517         *bounce         = true;
1518         *read_full      = promote_full;
1519         return promote;
1520 }
1521
1522 /* Read */
1523
1524 #define READ_RETRY_AVOID        1
1525 #define READ_RETRY              2
1526 #define READ_ERR                3
1527
1528 enum rbio_context {
1529         RBIO_CONTEXT_NULL,
1530         RBIO_CONTEXT_HIGHPRI,
1531         RBIO_CONTEXT_UNBOUND,
1532 };
1533
1534 static inline struct bch_read_bio *
1535 bch2_rbio_parent(struct bch_read_bio *rbio)
1536 {
1537         return rbio->split ? rbio->parent : rbio;
1538 }
1539
1540 __always_inline
1541 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1542                            enum rbio_context context,
1543                            struct workqueue_struct *wq)
1544 {
1545         if (context <= rbio->context) {
1546                 fn(&rbio->work);
1547         } else {
1548                 rbio->work.func         = fn;
1549                 rbio->context           = context;
1550                 queue_work(wq, &rbio->work);
1551         }
1552 }
1553
1554 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1555 {
1556         BUG_ON(rbio->bounce && !rbio->split);
1557
1558         if (rbio->promote)
1559                 promote_free(rbio->c, rbio->promote);
1560         rbio->promote = NULL;
1561
1562         if (rbio->bounce)
1563                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1564
1565         if (rbio->split) {
1566                 struct bch_read_bio *parent = rbio->parent;
1567
1568                 if (rbio->kmalloc)
1569                         kfree(rbio);
1570                 else
1571                         bio_put(&rbio->bio);
1572
1573                 rbio = parent;
1574         }
1575
1576         return rbio;
1577 }
1578
1579 /*
1580  * Only called on a top level bch_read_bio to complete an entire read request,
1581  * not a split:
1582  */
1583 static void bch2_rbio_done(struct bch_read_bio *rbio)
1584 {
1585         if (rbio->start_time)
1586                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1587                                        rbio->start_time);
1588         bio_endio(&rbio->bio);
1589 }
1590
1591 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1592                                      struct bvec_iter bvec_iter,
1593                                      struct bch_io_failures *failed,
1594                                      unsigned flags)
1595 {
1596         struct btree_trans trans;
1597         struct btree_iter iter;
1598         struct bkey_buf sk;
1599         struct bkey_s_c k;
1600         int ret;
1601
1602         flags &= ~BCH_READ_LAST_FRAGMENT;
1603         flags |= BCH_READ_MUST_CLONE;
1604
1605         bch2_bkey_buf_init(&sk);
1606         bch2_trans_init(&trans, c, 0, 0);
1607
1608         bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1609                              rbio->read_pos, BTREE_ITER_SLOTS);
1610 retry:
1611         rbio->bio.bi_status = 0;
1612
1613         k = bch2_btree_iter_peek_slot(&iter);
1614         if (bkey_err(k))
1615                 goto err;
1616
1617         bch2_bkey_buf_reassemble(&sk, c, k);
1618         k = bkey_i_to_s_c(sk.k);
1619         bch2_trans_unlock(&trans);
1620
1621         if (!bch2_bkey_matches_ptr(c, k,
1622                                    rbio->pick.ptr,
1623                                    rbio->data_pos.offset -
1624                                    rbio->pick.crc.offset)) {
1625                 /* extent we wanted to read no longer exists: */
1626                 rbio->hole = true;
1627                 goto out;
1628         }
1629
1630         ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1631                                  rbio->read_pos,
1632                                  rbio->data_btree,
1633                                  k, 0, failed, flags);
1634         if (ret == READ_RETRY)
1635                 goto retry;
1636         if (ret)
1637                 goto err;
1638 out:
1639         bch2_rbio_done(rbio);
1640         bch2_trans_iter_exit(&trans, &iter);
1641         bch2_trans_exit(&trans);
1642         bch2_bkey_buf_exit(&sk, c);
1643         return;
1644 err:
1645         rbio->bio.bi_status = BLK_STS_IOERR;
1646         goto out;
1647 }
1648
1649 static void bch2_rbio_retry(struct work_struct *work)
1650 {
1651         struct bch_read_bio *rbio =
1652                 container_of(work, struct bch_read_bio, work);
1653         struct bch_fs *c        = rbio->c;
1654         struct bvec_iter iter   = rbio->bvec_iter;
1655         unsigned flags          = rbio->flags;
1656         subvol_inum inum = {
1657                 .subvol = rbio->subvol,
1658                 .inum   = rbio->read_pos.inode,
1659         };
1660         struct bch_io_failures failed = { .nr = 0 };
1661
1662         trace_read_retry(&rbio->bio);
1663
1664         if (rbio->retry == READ_RETRY_AVOID)
1665                 bch2_mark_io_failure(&failed, &rbio->pick);
1666
1667         rbio->bio.bi_status = 0;
1668
1669         rbio = bch2_rbio_free(rbio);
1670
1671         flags |= BCH_READ_IN_RETRY;
1672         flags &= ~BCH_READ_MAY_PROMOTE;
1673
1674         if (flags & BCH_READ_NODECODE) {
1675                 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1676         } else {
1677                 flags &= ~BCH_READ_LAST_FRAGMENT;
1678                 flags |= BCH_READ_MUST_CLONE;
1679
1680                 __bch2_read(c, rbio, iter, inum, &failed, flags);
1681         }
1682 }
1683
1684 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1685                             blk_status_t error)
1686 {
1687         rbio->retry = retry;
1688
1689         if (rbio->flags & BCH_READ_IN_RETRY)
1690                 return;
1691
1692         if (retry == READ_ERR) {
1693                 rbio = bch2_rbio_free(rbio);
1694
1695                 rbio->bio.bi_status = error;
1696                 bch2_rbio_done(rbio);
1697         } else {
1698                 bch2_rbio_punt(rbio, bch2_rbio_retry,
1699                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1700         }
1701 }
1702
1703 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1704                                    struct bch_read_bio *rbio)
1705 {
1706         struct bch_fs *c = rbio->c;
1707         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1708         struct bch_extent_crc_unpacked new_crc;
1709         struct btree_iter iter;
1710         struct bkey_i *new;
1711         struct bkey_s_c k;
1712         int ret = 0;
1713
1714         if (crc_is_compressed(rbio->pick.crc))
1715                 return 0;
1716
1717         bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1718                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1719         k = bch2_btree_iter_peek_slot(&iter);
1720         if ((ret = bkey_err(k)))
1721                 goto out;
1722
1723         if (bversion_cmp(k.k->version, rbio->version) ||
1724             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1725                 goto out;
1726
1727         /* Extent was merged? */
1728         if (bkey_start_offset(k.k) < data_offset ||
1729             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1730                 goto out;
1731
1732         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1733                         rbio->pick.crc, NULL, &new_crc,
1734                         bkey_start_offset(k.k) - data_offset, k.k->size,
1735                         rbio->pick.crc.csum_type)) {
1736                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1737                 ret = 0;
1738                 goto out;
1739         }
1740
1741         /*
1742          * going to be temporarily appending another checksum entry:
1743          */
1744         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1745                                  sizeof(struct bch_extent_crc128));
1746         if ((ret = PTR_ERR_OR_ZERO(new)))
1747                 goto out;
1748
1749         bkey_reassemble(new, k);
1750
1751         if (!bch2_bkey_narrow_crcs(new, new_crc))
1752                 goto out;
1753
1754         ret = bch2_trans_update(trans, &iter, new,
1755                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1756 out:
1757         bch2_trans_iter_exit(trans, &iter);
1758         return ret;
1759 }
1760
1761 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1762 {
1763         bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1764                       __bch2_rbio_narrow_crcs(&trans, rbio));
1765 }
1766
1767 /* Inner part that may run in process context */
1768 static void __bch2_read_endio(struct work_struct *work)
1769 {
1770         struct bch_read_bio *rbio =
1771                 container_of(work, struct bch_read_bio, work);
1772         struct bch_fs *c        = rbio->c;
1773         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1774         struct bio *src         = &rbio->bio;
1775         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
1776         struct bvec_iter dst_iter = rbio->bvec_iter;
1777         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1778         struct nonce nonce = extent_nonce(rbio->version, crc);
1779         unsigned nofs_flags;
1780         struct bch_csum csum;
1781         int ret;
1782
1783         nofs_flags = memalloc_nofs_save();
1784
1785         /* Reset iterator for checksumming and copying bounced data: */
1786         if (rbio->bounce) {
1787                 src->bi_iter.bi_size            = crc.compressed_size << 9;
1788                 src->bi_iter.bi_idx             = 0;
1789                 src->bi_iter.bi_bvec_done       = 0;
1790         } else {
1791                 src->bi_iter                    = rbio->bvec_iter;
1792         }
1793
1794         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1795         if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1796                 goto csum_err;
1797
1798         /*
1799          * XXX
1800          * We need to rework the narrow_crcs path to deliver the read completion
1801          * first, and then punt to a different workqueue, otherwise we're
1802          * holding up reads while doing btree updates which is bad for memory
1803          * reclaim.
1804          */
1805         if (unlikely(rbio->narrow_crcs))
1806                 bch2_rbio_narrow_crcs(rbio);
1807
1808         if (rbio->flags & BCH_READ_NODECODE)
1809                 goto nodecode;
1810
1811         /* Adjust crc to point to subset of data we want: */
1812         crc.offset     += rbio->offset_into_extent;
1813         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
1814
1815         if (crc_is_compressed(crc)) {
1816                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1817                 if (ret)
1818                         goto decrypt_err;
1819
1820                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1821                         goto decompression_err;
1822         } else {
1823                 /* don't need to decrypt the entire bio: */
1824                 nonce = nonce_add(nonce, crc.offset << 9);
1825                 bio_advance(src, crc.offset << 9);
1826
1827                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1828                 src->bi_iter.bi_size = dst_iter.bi_size;
1829
1830                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1831                 if (ret)
1832                         goto decrypt_err;
1833
1834                 if (rbio->bounce) {
1835                         struct bvec_iter src_iter = src->bi_iter;
1836                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1837                 }
1838         }
1839
1840         if (rbio->promote) {
1841                 /*
1842                  * Re encrypt data we decrypted, so it's consistent with
1843                  * rbio->crc:
1844                  */
1845                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1846                 if (ret)
1847                         goto decrypt_err;
1848
1849                 promote_start(rbio->promote, rbio);
1850                 rbio->promote = NULL;
1851         }
1852 nodecode:
1853         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1854                 rbio = bch2_rbio_free(rbio);
1855                 bch2_rbio_done(rbio);
1856         }
1857 out:
1858         memalloc_nofs_restore(nofs_flags);
1859         return;
1860 csum_err:
1861         /*
1862          * Checksum error: if the bio wasn't bounced, we may have been
1863          * reading into buffers owned by userspace (that userspace can
1864          * scribble over) - retry the read, bouncing it this time:
1865          */
1866         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1867                 rbio->flags |= BCH_READ_MUST_BOUNCE;
1868                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1869                 goto out;
1870         }
1871
1872         bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1873                 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1874                 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1875                 csum.hi, csum.lo, crc.csum_type);
1876         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1877         goto out;
1878 decompression_err:
1879         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1880                                  "decompression error");
1881         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1882         goto out;
1883 decrypt_err:
1884         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1885                                  "decrypt error");
1886         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1887         goto out;
1888 }
1889
1890 static void bch2_read_endio(struct bio *bio)
1891 {
1892         struct bch_read_bio *rbio =
1893                 container_of(bio, struct bch_read_bio, bio);
1894         struct bch_fs *c        = rbio->c;
1895         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1896         struct workqueue_struct *wq = NULL;
1897         enum rbio_context context = RBIO_CONTEXT_NULL;
1898
1899         if (rbio->have_ioref) {
1900                 bch2_latency_acct(ca, rbio->submit_time, READ);
1901                 percpu_ref_put(&ca->io_ref);
1902         }
1903
1904         if (!rbio->split)
1905                 rbio->bio.bi_end_io = rbio->end_io;
1906
1907         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1908                                     rbio->read_pos.inode,
1909                                     rbio->read_pos.offset,
1910                                     "data read error: %s",
1911                                bch2_blk_status_to_str(bio->bi_status))) {
1912                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1913                 return;
1914         }
1915
1916         if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1917             ptr_stale(ca, &rbio->pick.ptr)) {
1918                 atomic_long_inc(&c->read_realloc_races);
1919
1920                 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1921                         bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1922                 else
1923                         bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1924                 return;
1925         }
1926
1927         if (rbio->narrow_crcs ||
1928             crc_is_compressed(rbio->pick.crc) ||
1929             bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1930                 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1931         else if (rbio->pick.crc.csum_type)
1932                 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1933
1934         bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1935 }
1936
1937 int __bch2_read_indirect_extent(struct btree_trans *trans,
1938                                 unsigned *offset_into_extent,
1939                                 struct bkey_buf *orig_k)
1940 {
1941         struct btree_iter iter;
1942         struct bkey_s_c k;
1943         u64 reflink_offset;
1944         int ret;
1945
1946         reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1947                 *offset_into_extent;
1948
1949         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
1950                              POS(0, reflink_offset),
1951                              BTREE_ITER_SLOTS);
1952         k = bch2_btree_iter_peek_slot(&iter);
1953         ret = bkey_err(k);
1954         if (ret)
1955                 goto err;
1956
1957         if (k.k->type != KEY_TYPE_reflink_v &&
1958             k.k->type != KEY_TYPE_indirect_inline_data) {
1959                 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
1960                         "%llu len %u points to nonexistent indirect extent %llu",
1961                         orig_k->k->k.p.offset,
1962                         orig_k->k->k.size,
1963                         reflink_offset);
1964                 bch2_inconsistent_error(trans->c);
1965                 ret = -EIO;
1966                 goto err;
1967         }
1968
1969         *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
1970         bch2_bkey_buf_reassemble(orig_k, trans->c, k);
1971 err:
1972         bch2_trans_iter_exit(trans, &iter);
1973         return ret;
1974 }
1975
1976 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
1977                                                    struct bkey_s_c k,
1978                                                    struct bch_extent_ptr ptr)
1979 {
1980         struct bch_fs *c = trans->c;
1981         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
1982         struct btree_iter iter;
1983         struct printbuf buf = PRINTBUF;
1984         int ret;
1985
1986         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1987                              PTR_BUCKET_POS(c, &ptr),
1988                              BTREE_ITER_CACHED);
1989
1990         prt_printf(&buf, "Attempting to read from stale dirty pointer:");
1991         printbuf_indent_add(&buf, 2);
1992         prt_newline(&buf);
1993
1994         bch2_bkey_val_to_text(&buf, c, k);
1995         prt_newline(&buf);
1996
1997         prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
1998
1999         ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
2000         if (!ret) {
2001                 prt_newline(&buf);
2002                 bch2_bkey_val_to_text(&buf, c, k);
2003         }
2004
2005         bch2_fs_inconsistent(c, "%s", buf.buf);
2006
2007         bch2_trans_iter_exit(trans, &iter);
2008         printbuf_exit(&buf);
2009 }
2010
2011 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2012                        struct bvec_iter iter, struct bpos read_pos,
2013                        enum btree_id data_btree, struct bkey_s_c k,
2014                        unsigned offset_into_extent,
2015                        struct bch_io_failures *failed, unsigned flags)
2016 {
2017         struct bch_fs *c = trans->c;
2018         struct extent_ptr_decoded pick;
2019         struct bch_read_bio *rbio = NULL;
2020         struct bch_dev *ca = NULL;
2021         struct promote_op *promote = NULL;
2022         bool bounce = false, read_full = false, narrow_crcs = false;
2023         struct bpos data_pos = bkey_start_pos(k.k);
2024         int pick_ret;
2025
2026         if (bkey_extent_is_inline_data(k.k)) {
2027                 unsigned bytes = min_t(unsigned, iter.bi_size,
2028                                        bkey_inline_data_bytes(k.k));
2029
2030                 swap(iter.bi_size, bytes);
2031                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2032                 swap(iter.bi_size, bytes);
2033                 bio_advance_iter(&orig->bio, &iter, bytes);
2034                 zero_fill_bio_iter(&orig->bio, iter);
2035                 goto out_read_done;
2036         }
2037 retry_pick:
2038         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2039
2040         /* hole or reservation - just zero fill: */
2041         if (!pick_ret)
2042                 goto hole;
2043
2044         if (pick_ret < 0) {
2045                 bch_err_inum_ratelimited(c, k.k->p.inode,
2046                                          "no device to read from");
2047                 goto err;
2048         }
2049
2050         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2051
2052         /*
2053          * Stale dirty pointers are treated as IO errors, but @failed isn't
2054          * allocated unless we're in the retry path - so if we're not in the
2055          * retry path, don't check here, it'll be caught in bch2_read_endio()
2056          * and we'll end up in the retry path:
2057          */
2058         if ((flags & BCH_READ_IN_RETRY) &&
2059             !pick.ptr.cached &&
2060             unlikely(ptr_stale(ca, &pick.ptr))) {
2061                 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2062                 bch2_mark_io_failure(failed, &pick);
2063                 goto retry_pick;
2064         }
2065
2066         /*
2067          * Unlock the iterator while the btree node's lock is still in
2068          * cache, before doing the IO:
2069          */
2070         bch2_trans_unlock(trans);
2071
2072         if (flags & BCH_READ_NODECODE) {
2073                 /*
2074                  * can happen if we retry, and the extent we were going to read
2075                  * has been merged in the meantime:
2076                  */
2077                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2078                         goto hole;
2079
2080                 iter.bi_size    = pick.crc.compressed_size << 9;
2081                 goto get_bio;
2082         }
2083
2084         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2085             bio_flagged(&orig->bio, BIO_CHAIN))
2086                 flags |= BCH_READ_MUST_CLONE;
2087
2088         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2089                 bch2_can_narrow_extent_crcs(k, pick.crc);
2090
2091         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2092                 flags |= BCH_READ_MUST_BOUNCE;
2093
2094         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2095
2096         if (crc_is_compressed(pick.crc) ||
2097             (pick.crc.csum_type != BCH_CSUM_none &&
2098              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2099               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2100                (flags & BCH_READ_USER_MAPPED)) ||
2101               (flags & BCH_READ_MUST_BOUNCE)))) {
2102                 read_full = true;
2103                 bounce = true;
2104         }
2105
2106         if (orig->opts.promote_target)
2107                 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2108                                         &rbio, &bounce, &read_full);
2109
2110         if (!read_full) {
2111                 EBUG_ON(crc_is_compressed(pick.crc));
2112                 EBUG_ON(pick.crc.csum_type &&
2113                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2114                          bvec_iter_sectors(iter) != pick.crc.live_size ||
2115                          pick.crc.offset ||
2116                          offset_into_extent));
2117
2118                 data_pos.offset += offset_into_extent;
2119                 pick.ptr.offset += pick.crc.offset +
2120                         offset_into_extent;
2121                 offset_into_extent              = 0;
2122                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
2123                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
2124                 pick.crc.offset                 = 0;
2125                 pick.crc.live_size              = bvec_iter_sectors(iter);
2126                 offset_into_extent              = 0;
2127         }
2128 get_bio:
2129         if (rbio) {
2130                 /*
2131                  * promote already allocated bounce rbio:
2132                  * promote needs to allocate a bio big enough for uncompressing
2133                  * data in the write path, but we're not going to use it all
2134                  * here:
2135                  */
2136                 EBUG_ON(rbio->bio.bi_iter.bi_size <
2137                        pick.crc.compressed_size << 9);
2138                 rbio->bio.bi_iter.bi_size =
2139                         pick.crc.compressed_size << 9;
2140         } else if (bounce) {
2141                 unsigned sectors = pick.crc.compressed_size;
2142
2143                 rbio = rbio_init(bio_alloc_bioset(NULL,
2144                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
2145                                                   0,
2146                                                   GFP_NOIO,
2147                                                   &c->bio_read_split),
2148                                  orig->opts);
2149
2150                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2151                 rbio->bounce    = true;
2152                 rbio->split     = true;
2153         } else if (flags & BCH_READ_MUST_CLONE) {
2154                 /*
2155                  * Have to clone if there were any splits, due to error
2156                  * reporting issues (if a split errored, and retrying didn't
2157                  * work, when it reports the error to its parent (us) we don't
2158                  * know if the error was from our bio, and we should retry, or
2159                  * from the whole bio, in which case we don't want to retry and
2160                  * lose the error)
2161                  */
2162                 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOIO,
2163                                                  &c->bio_read_split),
2164                                  orig->opts);
2165                 rbio->bio.bi_iter = iter;
2166                 rbio->split     = true;
2167         } else {
2168                 rbio = orig;
2169                 rbio->bio.bi_iter = iter;
2170                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2171         }
2172
2173         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2174
2175         rbio->c                 = c;
2176         rbio->submit_time       = local_clock();
2177         if (rbio->split)
2178                 rbio->parent    = orig;
2179         else
2180                 rbio->end_io    = orig->bio.bi_end_io;
2181         rbio->bvec_iter         = iter;
2182         rbio->offset_into_extent= offset_into_extent;
2183         rbio->flags             = flags;
2184         rbio->have_ioref        = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2185         rbio->narrow_crcs       = narrow_crcs;
2186         rbio->hole              = 0;
2187         rbio->retry             = 0;
2188         rbio->context           = 0;
2189         /* XXX: only initialize this if needed */
2190         rbio->devs_have         = bch2_bkey_devs(k);
2191         rbio->pick              = pick;
2192         rbio->subvol            = orig->subvol;
2193         rbio->read_pos          = read_pos;
2194         rbio->data_btree        = data_btree;
2195         rbio->data_pos          = data_pos;
2196         rbio->version           = k.k->version;
2197         rbio->promote           = promote;
2198         INIT_WORK(&rbio->work, NULL);
2199
2200         rbio->bio.bi_opf        = orig->bio.bi_opf;
2201         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2202         rbio->bio.bi_end_io     = bch2_read_endio;
2203
2204         if (rbio->bounce)
2205                 trace_read_bounce(&rbio->bio);
2206
2207         this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
2208         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2209
2210         /*
2211          * If it's being moved internally, we don't want to flag it as a cache
2212          * hit:
2213          */
2214         if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2215                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2216                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
2217
2218         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2219                 bio_inc_remaining(&orig->bio);
2220                 trace_read_split(&orig->bio);
2221         }
2222
2223         if (!rbio->pick.idx) {
2224                 if (!rbio->have_ioref) {
2225                         bch_err_inum_ratelimited(c, k.k->p.inode,
2226                                                  "no device to read from");
2227                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2228                         goto out;
2229                 }
2230
2231                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2232                              bio_sectors(&rbio->bio));
2233                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2234
2235                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2236                         submit_bio(&rbio->bio);
2237                 else
2238                         submit_bio_wait(&rbio->bio);
2239         } else {
2240                 /* Attempting reconstruct read: */
2241                 if (bch2_ec_read_extent(c, rbio)) {
2242                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2243                         goto out;
2244                 }
2245
2246                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2247                         bio_endio(&rbio->bio);
2248         }
2249 out:
2250         if (likely(!(flags & BCH_READ_IN_RETRY))) {
2251                 return 0;
2252         } else {
2253                 int ret;
2254
2255                 rbio->context = RBIO_CONTEXT_UNBOUND;
2256                 bch2_read_endio(&rbio->bio);
2257
2258                 ret = rbio->retry;
2259                 rbio = bch2_rbio_free(rbio);
2260
2261                 if (ret == READ_RETRY_AVOID) {
2262                         bch2_mark_io_failure(failed, &pick);
2263                         ret = READ_RETRY;
2264                 }
2265
2266                 if (!ret)
2267                         goto out_read_done;
2268
2269                 return ret;
2270         }
2271
2272 err:
2273         if (flags & BCH_READ_IN_RETRY)
2274                 return READ_ERR;
2275
2276         orig->bio.bi_status = BLK_STS_IOERR;
2277         goto out_read_done;
2278
2279 hole:
2280         /*
2281          * won't normally happen in the BCH_READ_NODECODE
2282          * (bch2_move_extent()) path, but if we retry and the extent we wanted
2283          * to read no longer exists we have to signal that:
2284          */
2285         if (flags & BCH_READ_NODECODE)
2286                 orig->hole = true;
2287
2288         zero_fill_bio_iter(&orig->bio, iter);
2289 out_read_done:
2290         if (flags & BCH_READ_LAST_FRAGMENT)
2291                 bch2_rbio_done(orig);
2292         return 0;
2293 }
2294
2295 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2296                  struct bvec_iter bvec_iter, subvol_inum inum,
2297                  struct bch_io_failures *failed, unsigned flags)
2298 {
2299         struct btree_trans trans;
2300         struct btree_iter iter;
2301         struct bkey_buf sk;
2302         struct bkey_s_c k;
2303         u32 snapshot;
2304         int ret;
2305
2306         BUG_ON(flags & BCH_READ_NODECODE);
2307
2308         bch2_bkey_buf_init(&sk);
2309         bch2_trans_init(&trans, c, 0, 0);
2310 retry:
2311         bch2_trans_begin(&trans);
2312         iter = (struct btree_iter) { NULL };
2313
2314         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2315         if (ret)
2316                 goto err;
2317
2318         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2319                              SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2320                              BTREE_ITER_SLOTS);
2321         while (1) {
2322                 unsigned bytes, sectors, offset_into_extent;
2323                 enum btree_id data_btree = BTREE_ID_extents;
2324
2325                 /*
2326                  * read_extent -> io_time_reset may cause a transaction restart
2327                  * without returning an error, we need to check for that here:
2328                  */
2329                 if (!bch2_trans_relock(&trans)) {
2330                         ret = -EINTR;
2331                         break;
2332                 }
2333
2334                 bch2_btree_iter_set_pos(&iter,
2335                                 POS(inum.inum, bvec_iter.bi_sector));
2336
2337                 k = bch2_btree_iter_peek_slot(&iter);
2338                 ret = bkey_err(k);
2339                 if (ret)
2340                         break;
2341
2342                 offset_into_extent = iter.pos.offset -
2343                         bkey_start_offset(k.k);
2344                 sectors = k.k->size - offset_into_extent;
2345
2346                 bch2_bkey_buf_reassemble(&sk, c, k);
2347
2348                 ret = bch2_read_indirect_extent(&trans, &data_btree,
2349                                         &offset_into_extent, &sk);
2350                 if (ret)
2351                         break;
2352
2353                 k = bkey_i_to_s_c(sk.k);
2354
2355                 /*
2356                  * With indirect extents, the amount of data to read is the min
2357                  * of the original extent and the indirect extent:
2358                  */
2359                 sectors = min(sectors, k.k->size - offset_into_extent);
2360
2361                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2362                 swap(bvec_iter.bi_size, bytes);
2363
2364                 if (bvec_iter.bi_size == bytes)
2365                         flags |= BCH_READ_LAST_FRAGMENT;
2366
2367                 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2368                                          data_btree, k,
2369                                          offset_into_extent, failed, flags);
2370                 if (ret)
2371                         break;
2372
2373                 if (flags & BCH_READ_LAST_FRAGMENT)
2374                         break;
2375
2376                 swap(bvec_iter.bi_size, bytes);
2377                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2378
2379                 ret = btree_trans_too_many_iters(&trans);
2380                 if (ret)
2381                         break;
2382         }
2383 err:
2384         bch2_trans_iter_exit(&trans, &iter);
2385
2386         if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
2387                 goto retry;
2388
2389         bch2_trans_exit(&trans);
2390         bch2_bkey_buf_exit(&sk, c);
2391
2392         if (ret) {
2393                 bch_err_inum_ratelimited(c, inum.inum,
2394                                          "read error %i from btree lookup", ret);
2395                 rbio->bio.bi_status = BLK_STS_IOERR;
2396                 bch2_rbio_done(rbio);
2397         }
2398 }
2399
2400 void bch2_fs_io_exit(struct bch_fs *c)
2401 {
2402         if (c->promote_table.tbl)
2403                 rhashtable_destroy(&c->promote_table);
2404         mempool_exit(&c->bio_bounce_pages);
2405         bioset_exit(&c->bio_write);
2406         bioset_exit(&c->bio_read_split);
2407         bioset_exit(&c->bio_read);
2408 }
2409
2410 int bch2_fs_io_init(struct bch_fs *c)
2411 {
2412         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2413                         BIOSET_NEED_BVECS) ||
2414             bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2415                         BIOSET_NEED_BVECS) ||
2416             bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2417                         BIOSET_NEED_BVECS) ||
2418             mempool_init_page_pool(&c->bio_bounce_pages,
2419                                    max_t(unsigned,
2420                                          c->opts.btree_node_size,
2421                                          c->opts.encoded_extent_max) /
2422                                    PAGE_SIZE, 0) ||
2423             rhashtable_init(&c->promote_table, &bch_promote_params))
2424                 return -ENOMEM;
2425
2426         return 0;
2427 }