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_clone_fast(&wbio->bio, GFP_NOIO,
 474                                                    &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(GFP_NOIO, pages, &c->bio_write);
 705         wbio                    = wbio_init(bio);
 706         wbio->put_bio           = true;
 707         /* copy WRITE_SYNC flag */
 708         wbio->bio.bi_opf        = src->bi_opf;
 709
 710         if (buf) {
 711                 bch2_bio_map(bio, buf, output_available);
 712                 return bio;
 713         }
 714
 715         wbio->bounce            = true;
 716
 717         /*
 718          * We can't use mempool for more than c->sb.encoded_extent_max
 719          * worth of pages, but we'd like to allocate more if we can:
 720          */
 721         bch2_bio_alloc_pages_pool(c, bio,
 722                                   min_t(unsigned, output_available,
 723                                         c->opts.encoded_extent_max));
 724
 725         if (bio->bi_iter.bi_size < output_available)
 726                 *page_alloc_failed =
 727                         bch2_bio_alloc_pages(bio,
 728                                              output_available -
 729                                              bio->bi_iter.bi_size,
 730                                              GFP_NOFS) != 0;
 731
 732         return bio;
 733 }
 734
 735 static int bch2_write_rechecksum(struct bch_fs *c,
 736                                  struct bch_write_op *op,
 737                                  unsigned new_csum_type)
 738 {
 739         struct bio *bio = &op->wbio.bio;
 740         struct bch_extent_crc_unpacked new_crc;
 741         int ret;
 742
 743         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 744
 745         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 746             bch2_csum_type_is_encryption(new_csum_type))
 747                 new_csum_type = op->crc.csum_type;
 748
 749         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 750                                   NULL, &new_crc,
 751                                   op->crc.offset, op->crc.live_size,
 752                                   new_csum_type);
 753         if (ret)
 754                 return ret;
 755
 756         bio_advance(bio, op->crc.offset << 9);
 757         bio->bi_iter.bi_size = op->crc.live_size << 9;
 758         op->crc = new_crc;
 759         return 0;
 760 }
 761
 762 static int bch2_write_decrypt(struct bch_write_op *op)
 763 {
 764         struct bch_fs *c = op->c;
 765         struct nonce nonce = extent_nonce(op->version, op->crc);
 766         struct bch_csum csum;
 767         int ret;
 768
 769         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 770                 return 0;
 771
 772         /*
 773          * If we need to decrypt data in the write path, we'll no longer be able
 774          * to verify the existing checksum (poly1305 mac, in this case) after
 775          * it's decrypted - this is the last point we'll be able to reverify the
 776          * checksum:
 777          */
 778         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 779         if (bch2_crc_cmp(op->crc.csum, csum))
 780                 return -EIO;
 781
 782         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 783         op->crc.csum_type = 0;
 784         op->crc.csum = (struct bch_csum) { 0, 0 };
 785         return ret;
 786 }
 787
 788 static enum prep_encoded_ret {
 789         PREP_ENCODED_OK,
 790         PREP_ENCODED_ERR,
 791         PREP_ENCODED_CHECKSUM_ERR,
 792         PREP_ENCODED_DO_WRITE,
 793 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 794 {
 795         struct bch_fs *c = op->c;
 796         struct bio *bio = &op->wbio.bio;
 797
 798         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 799                 return PREP_ENCODED_OK;
 800
 801         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 802
 803         /* Can we just write the entire extent as is? */
 804         if (op->crc.uncompressed_size == op->crc.live_size &&
 805             op->crc.compressed_size <= wp->sectors_free &&
 806             (op->crc.compression_type == op->compression_type ||
 807              op->incompressible)) {
 808                 if (!crc_is_compressed(op->crc) &&
 809                     op->csum_type != op->crc.csum_type &&
 810                     bch2_write_rechecksum(c, op, op->csum_type))
 811                         return PREP_ENCODED_CHECKSUM_ERR;
 812
 813                 return PREP_ENCODED_DO_WRITE;
 814         }
 815
 816         /*
 817          * If the data is compressed and we couldn't write the entire extent as
 818          * is, we have to decompress it:
 819          */
 820         if (crc_is_compressed(op->crc)) {
 821                 struct bch_csum csum;
 822
 823                 if (bch2_write_decrypt(op))
 824                         return PREP_ENCODED_CHECKSUM_ERR;
 825
 826                 /* Last point we can still verify checksum: */
 827                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 828                                          extent_nonce(op->version, op->crc),
 829                                          bio);
 830                 if (bch2_crc_cmp(op->crc.csum, csum))
 831                         return PREP_ENCODED_CHECKSUM_ERR;
 832
 833                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 834                         return PREP_ENCODED_ERR;
 835         }
 836
 837         /*
 838          * No longer have compressed data after this point - data might be
 839          * encrypted:
 840          */
 841
 842         /*
 843          * If the data is checksummed and we're only writing a subset,
 844          * rechecksum and adjust bio to point to currently live data:
 845          */
 846         if ((op->crc.live_size != op->crc.uncompressed_size ||
 847              op->crc.csum_type != op->csum_type) &&
 848             bch2_write_rechecksum(c, op, op->csum_type))
 849                 return PREP_ENCODED_CHECKSUM_ERR;
 850
 851         /*
 852          * If we want to compress the data, it has to be decrypted:
 853          */
 854         if ((op->compression_type ||
 855              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 856              bch2_csum_type_is_encryption(op->csum_type)) &&
 857             bch2_write_decrypt(op))
 858                 return PREP_ENCODED_CHECKSUM_ERR;
 859
 860         return PREP_ENCODED_OK;
 861 }
 862
 863 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 864                              struct bio **_dst)
 865 {
 866         struct bch_fs *c = op->c;
 867         struct bio *src = &op->wbio.bio, *dst = src;
 868         struct bvec_iter saved_iter;
 869         void *ec_buf;
 870         unsigned total_output = 0, total_input = 0;
 871         bool bounce = false;
 872         bool page_alloc_failed = false;
 873         int ret, more = 0;
 874
 875         BUG_ON(!bio_sectors(src));
 876
 877         ec_buf = bch2_writepoint_ec_buf(c, wp);
 878
 879         switch (bch2_write_prep_encoded_data(op, wp)) {
 880         case PREP_ENCODED_OK:
 881                 break;
 882         case PREP_ENCODED_ERR:
 883                 ret = -EIO;
 884                 goto err;
 885         case PREP_ENCODED_CHECKSUM_ERR:
 886                 goto csum_err;
 887         case PREP_ENCODED_DO_WRITE:
 888                 /* XXX look for bug here */
 889                 if (ec_buf) {
 890                         dst = bch2_write_bio_alloc(c, wp, src,
 891                                                    &page_alloc_failed,
 892                                                    ec_buf);
 893                         bio_copy_data(dst, src);
 894                         bounce = true;
 895                 }
 896                 init_append_extent(op, wp, op->version, op->crc);
 897                 goto do_write;
 898         }
 899
 900         if (ec_buf ||
 901             op->compression_type ||
 902             (op->csum_type &&
 903              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 904             (bch2_csum_type_is_encryption(op->csum_type) &&
 905              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 906                 dst = bch2_write_bio_alloc(c, wp, src,
 907                                            &page_alloc_failed,
 908                                            ec_buf);
 909                 bounce = true;
 910         }
 911
 912         saved_iter = dst->bi_iter;
 913
 914         do {
 915                 struct bch_extent_crc_unpacked crc =
 916                         (struct bch_extent_crc_unpacked) { 0 };
 917                 struct bversion version = op->version;
 918                 size_t dst_len, src_len;
 919
 920                 if (page_alloc_failed &&
 921                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 922                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 923                         break;
 924
 925                 BUG_ON(op->compression_type &&
 926                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 927                        bch2_csum_type_is_encryption(op->crc.csum_type));
 928                 BUG_ON(op->compression_type && !bounce);
 929
 930                 crc.compression_type = op->incompressible
 931                         ? BCH_COMPRESSION_TYPE_incompressible
 932                         : op->compression_type
 933                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 934                                             op->compression_type)
 935                         : 0;
 936                 if (!crc_is_compressed(crc)) {
 937                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 938                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 939
 940                         if (op->csum_type)
 941                                 dst_len = min_t(unsigned, dst_len,
 942                                                 c->opts.encoded_extent_max);
 943
 944                         if (bounce) {
 945                                 swap(dst->bi_iter.bi_size, dst_len);
 946                                 bio_copy_data(dst, src);
 947                                 swap(dst->bi_iter.bi_size, dst_len);
 948                         }
 949
 950                         src_len = dst_len;
 951                 }
 952
 953                 BUG_ON(!src_len || !dst_len);
 954
 955                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 956                         if (bversion_zero(version)) {
 957                                 version.lo = atomic64_inc_return(&c->key_version);
 958                         } else {
 959                                 crc.nonce = op->nonce;
 960                                 op->nonce += src_len >> 9;
 961                         }
 962                 }
 963
 964                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 965                     !crc_is_compressed(crc) &&
 966                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
 967                     bch2_csum_type_is_encryption(op->csum_type)) {
 968                         /*
 969                          * Note: when we're using rechecksum(), we need to be
 970                          * checksumming @src because it has all the data our
 971                          * existing checksum covers - if we bounced (because we
 972                          * were trying to compress), @dst will only have the
 973                          * part of the data the new checksum will cover.
 974                          *
 975                          * But normally we want to be checksumming post bounce,
 976                          * because part of the reason for bouncing is so the
 977                          * data can't be modified (by userspace) while it's in
 978                          * flight.
 979                          */
 980                         if (bch2_rechecksum_bio(c, src, version, op->crc,
 981                                         &crc, &op->crc,
 982                                         src_len >> 9,
 983                                         bio_sectors(src) - (src_len >> 9),
 984                                         op->csum_type))
 985                                 goto csum_err;
 986                 } else {
 987                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 988                             bch2_rechecksum_bio(c, src, version, op->crc,
 989                                         NULL, &op->crc,
 990                                         src_len >> 9,
 991                                         bio_sectors(src) - (src_len >> 9),
 992                                         op->crc.csum_type))
 993                                 goto csum_err;
 994
 995                         crc.compressed_size     = dst_len >> 9;
 996                         crc.uncompressed_size   = src_len >> 9;
 997                         crc.live_size           = src_len >> 9;
 998
 999                         swap(dst->bi_iter.bi_size, dst_len);
1000                         ret = bch2_encrypt_bio(c, op->csum_type,
1001                                                extent_nonce(version, crc), dst);
1002                         if (ret)
1003                                 goto err;
1004
1005                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1006                                          extent_nonce(version, crc), dst);
1007                         crc.csum_type = op->csum_type;
1008                         swap(dst->bi_iter.bi_size, dst_len);
1009                 }
1010
1011                 init_append_extent(op, wp, version, crc);
1012
1013                 if (dst != src)
1014                         bio_advance(dst, dst_len);
1015                 bio_advance(src, src_len);
1016                 total_output    += dst_len;
1017                 total_input     += src_len;
1018         } while (dst->bi_iter.bi_size &&
1019                  src->bi_iter.bi_size &&
1020                  wp->sectors_free &&
1021                  !bch2_keylist_realloc(&op->insert_keys,
1022                                       op->inline_keys,
1023                                       ARRAY_SIZE(op->inline_keys),
1024                                       BKEY_EXTENT_U64s_MAX));
1025
1026         more = src->bi_iter.bi_size != 0;
1027
1028         dst->bi_iter = saved_iter;
1029
1030         if (dst == src && more) {
1031                 BUG_ON(total_output != total_input);
1032
1033                 dst = bio_split(src, total_input >> 9,
1034                                 GFP_NOIO, &c->bio_write);
1035                 wbio_init(dst)->put_bio = true;
1036                 /* copy WRITE_SYNC flag */
1037                 dst->bi_opf             = src->bi_opf;
1038         }
1039
1040         dst->bi_iter.bi_size = total_output;
1041 do_write:
1042         *_dst = dst;
1043         return more;
1044 csum_err:
1045         bch_err(c, "error verifying existing checksum while "
1046                 "rewriting existing data (memory corruption?)");
1047         ret = -EIO;
1048 err:
1049         if (to_wbio(dst)->bounce)
1050                 bch2_bio_free_pages_pool(c, dst);
1051         if (to_wbio(dst)->put_bio)
1052                 bio_put(dst);
1053
1054         return ret;
1055 }
1056
1057 static void __bch2_write(struct closure *cl)
1058 {
1059         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1060         struct bch_fs *c = op->c;
1061         struct write_point *wp;
1062         struct bio *bio = NULL;
1063         bool skip_put = true;
1064         unsigned nofs_flags;
1065         int ret;
1066
1067         nofs_flags = memalloc_nofs_save();
1068 again:
1069         memset(&op->failed, 0, sizeof(op->failed));
1070
1071         do {
1072                 struct bkey_i *key_to_write;
1073                 unsigned key_to_write_offset = op->insert_keys.top_p -
1074                         op->insert_keys.keys_p;
1075
1076                 /* +1 for possible cache device: */
1077                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1078                     ARRAY_SIZE(op->open_buckets.v))
1079                         goto flush_io;
1080
1081                 if (bch2_keylist_realloc(&op->insert_keys,
1082                                         op->inline_keys,
1083                                         ARRAY_SIZE(op->inline_keys),
1084                                         BKEY_EXTENT_U64s_MAX))
1085                         goto flush_io;
1086
1087                 if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
1088                     percpu_ref_is_dying(&c->writes)) {
1089                         ret = -EROFS;
1090                         goto err;
1091                 }
1092
1093                 /*
1094                  * The copygc thread is now global, which means it's no longer
1095                  * freeing up space on specific disks, which means that
1096                  * allocations for specific disks may hang arbitrarily long:
1097                  */
1098                 wp = bch2_alloc_sectors_start(c,
1099                         op->target,
1100                         op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1101                         op->write_point,
1102                         &op->devs_have,
1103                         op->nr_replicas,
1104                         op->nr_replicas_required,
1105                         op->alloc_reserve,
1106                         op->flags,
1107                         (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1108                                       BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1109                 EBUG_ON(!wp);
1110
1111                 if (unlikely(IS_ERR(wp))) {
1112                         if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1113                                 ret = PTR_ERR(wp);
1114                                 goto err;
1115                         }
1116
1117                         goto flush_io;
1118                 }
1119
1120                 /*
1121                  * It's possible for the allocator to fail, put us on the
1122                  * freelist waitlist, and then succeed in one of various retry
1123                  * paths: if that happens, we need to disable the skip_put
1124                  * optimization because otherwise there won't necessarily be a
1125                  * barrier before we free the bch_write_op:
1126                  */
1127                 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1128                         skip_put = false;
1129
1130                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1131                 ret = bch2_write_extent(op, wp, &bio);
1132                 bch2_alloc_sectors_done(c, wp);
1133
1134                 if (ret < 0)
1135                         goto err;
1136
1137                 if (ret) {
1138                         skip_put = false;
1139                 } else {
1140                         /*
1141                          * for the skip_put optimization this has to be set
1142                          * before we submit the bio:
1143                          */
1144                         op->flags |= BCH_WRITE_DONE;
1145                 }
1146
1147                 bio->bi_end_io  = bch2_write_endio;
1148                 bio->bi_private = &op->cl;
1149                 bio->bi_opf |= REQ_OP_WRITE;
1150
1151                 if (!skip_put)
1152                         closure_get(bio->bi_private);
1153                 else
1154                         op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1155
1156                 key_to_write = (void *) (op->insert_keys.keys_p +
1157                                          key_to_write_offset);
1158
1159                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1160                                           key_to_write);
1161         } while (ret);
1162
1163         if (!skip_put)
1164                 continue_at(cl, bch2_write_index, index_update_wq(op));
1165 out:
1166         memalloc_nofs_restore(nofs_flags);
1167         return;
1168 err:
1169         op->error = ret;
1170         op->flags |= BCH_WRITE_DONE;
1171
1172         continue_at(cl, bch2_write_index, index_update_wq(op));
1173         goto out;
1174 flush_io:
1175         /*
1176          * If the write can't all be submitted at once, we generally want to
1177          * block synchronously as that signals backpressure to the caller.
1178          *
1179          * However, if we're running out of a workqueue, we can't block here
1180          * because we'll be blocking other work items from completing:
1181          */
1182         if (current->flags & PF_WQ_WORKER) {
1183                 continue_at(cl, bch2_write_index, index_update_wq(op));
1184                 goto out;
1185         }
1186
1187         closure_sync(cl);
1188
1189         if (!bch2_keylist_empty(&op->insert_keys)) {
1190                 __bch2_write_index(op);
1191
1192                 if (op->error) {
1193                         op->flags |= BCH_WRITE_DONE;
1194                         continue_at_nobarrier(cl, bch2_write_done, NULL);
1195                         goto out;
1196                 }
1197         }
1198
1199         goto again;
1200 }
1201
1202 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1203 {
1204         struct closure *cl = &op->cl;
1205         struct bio *bio = &op->wbio.bio;
1206         struct bvec_iter iter;
1207         struct bkey_i_inline_data *id;
1208         unsigned sectors;
1209         int ret;
1210
1211         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1212
1213         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1214                                    ARRAY_SIZE(op->inline_keys),
1215                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1216         if (ret) {
1217                 op->error = ret;
1218                 goto err;
1219         }
1220
1221         sectors = bio_sectors(bio);
1222         op->pos.offset += sectors;
1223
1224         id = bkey_inline_data_init(op->insert_keys.top);
1225         id->k.p         = op->pos;
1226         id->k.version   = op->version;
1227         id->k.size      = sectors;
1228
1229         iter = bio->bi_iter;
1230         iter.bi_size = data_len;
1231         memcpy_from_bio(id->v.data, bio, iter);
1232
1233         while (data_len & 7)
1234                 id->v.data[data_len++] = '\0';
1235         set_bkey_val_bytes(&id->k, data_len);
1236         bch2_keylist_push(&op->insert_keys);
1237
1238         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1239         op->flags |= BCH_WRITE_DONE;
1240
1241         continue_at_nobarrier(cl, bch2_write_index, NULL);
1242         return;
1243 err:
1244         bch2_write_done(&op->cl);
1245 }
1246
1247 /**
1248  * bch_write - handle a write to a cache device or flash only volume
1249  *
1250  * This is the starting point for any data to end up in a cache device; it could
1251  * be from a normal write, or a writeback write, or a write to a flash only
1252  * volume - it's also used by the moving garbage collector to compact data in
1253  * mostly empty buckets.
1254  *
1255  * It first writes the data to the cache, creating a list of keys to be inserted
1256  * (if the data won't fit in a single open bucket, there will be multiple keys);
1257  * after the data is written it calls bch_journal, and after the keys have been
1258  * added to the next journal write they're inserted into the btree.
1259  *
1260  * If op->discard is true, instead of inserting the data it invalidates the
1261  * region of the cache represented by op->bio and op->inode.
1262  */
1263 void bch2_write(struct closure *cl)
1264 {
1265         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1266         struct bio *bio = &op->wbio.bio;
1267         struct bch_fs *c = op->c;
1268         unsigned data_len;
1269
1270         BUG_ON(!op->nr_replicas);
1271         BUG_ON(!op->write_point.v);
1272         BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1273
1274         op->start_time = local_clock();
1275         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1276         wbio_init(bio)->put_bio = false;
1277
1278         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1279                 bch_err_inum_ratelimited(c, op->pos.inode,
1280                                          "misaligned write");
1281                 op->error = -EIO;
1282                 goto err;
1283         }
1284
1285         if (c->opts.nochanges ||
1286             !percpu_ref_tryget(&c->writes)) {
1287                 op->error = -EROFS;
1288                 goto err;
1289         }
1290
1291         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1292         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1293
1294         data_len = min_t(u64, bio->bi_iter.bi_size,
1295                          op->new_i_size - (op->pos.offset << 9));
1296
1297         if (c->opts.inline_data &&
1298             data_len <= min(block_bytes(c) / 2, 1024U)) {
1299                 bch2_write_data_inline(op, data_len);
1300                 return;
1301         }
1302
1303         continue_at_nobarrier(cl, __bch2_write, NULL);
1304         return;
1305 err:
1306         bch2_disk_reservation_put(c, &op->res);
1307
1308         if (op->end_io) {
1309                 EBUG_ON(cl->parent);
1310                 closure_debug_destroy(cl);
1311                 op->end_io(op);
1312         } else {
1313                 closure_return(cl);
1314         }
1315 }
1316
1317 /* Cache promotion on read */
1318
1319 struct promote_op {
1320         struct closure          cl;
1321         struct rcu_head         rcu;
1322         u64                     start_time;
1323
1324         struct rhash_head       hash;
1325         struct bpos             pos;
1326
1327         struct migrate_write    write;
1328         struct bio_vec          bi_inline_vecs[0]; /* must be last */
1329 };
1330
1331 static const struct rhashtable_params bch_promote_params = {
1332         .head_offset    = offsetof(struct promote_op, hash),
1333         .key_offset     = offsetof(struct promote_op, pos),
1334         .key_len        = sizeof(struct bpos),
1335 };
1336
1337 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1338                                   struct bpos pos,
1339                                   struct bch_io_opts opts,
1340                                   unsigned flags)
1341 {
1342         if (!(flags & BCH_READ_MAY_PROMOTE))
1343                 return false;
1344
1345         if (!opts.promote_target)
1346                 return false;
1347
1348         if (bch2_bkey_has_target(c, k, opts.promote_target))
1349                 return false;
1350
1351         if (bch2_target_congested(c, opts.promote_target)) {
1352                 /* XXX trace this */
1353                 return false;
1354         }
1355
1356         if (rhashtable_lookup_fast(&c->promote_table, &pos,
1357                                    bch_promote_params))
1358                 return false;
1359
1360         return true;
1361 }
1362
1363 static void promote_free(struct bch_fs *c, struct promote_op *op)
1364 {
1365         int ret;
1366
1367         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1368                                      bch_promote_params);
1369         BUG_ON(ret);
1370         percpu_ref_put(&c->writes);
1371         kfree_rcu(op, rcu);
1372 }
1373
1374 static void promote_done(struct closure *cl)
1375 {
1376         struct promote_op *op =
1377                 container_of(cl, struct promote_op, cl);
1378         struct bch_fs *c = op->write.op.c;
1379
1380         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1381                                op->start_time);
1382
1383         bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1384         promote_free(c, op);
1385 }
1386
1387 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1388 {
1389         struct bch_fs *c = rbio->c;
1390         struct closure *cl = &op->cl;
1391         struct bio *bio = &op->write.op.wbio.bio;
1392
1393         trace_promote(&rbio->bio);
1394
1395         /* we now own pages: */
1396         BUG_ON(!rbio->bounce);
1397         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1398
1399         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1400                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1401         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1402
1403         bch2_migrate_read_done(&op->write, rbio);
1404
1405         closure_init(cl, NULL);
1406         closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
1407         closure_return_with_destructor(cl, promote_done);
1408 }
1409
1410 static struct promote_op *__promote_alloc(struct bch_fs *c,
1411                                           enum btree_id btree_id,
1412                                           struct bkey_s_c k,
1413                                           struct bpos pos,
1414                                           struct extent_ptr_decoded *pick,
1415                                           struct bch_io_opts opts,
1416                                           unsigned sectors,
1417                                           struct bch_read_bio **rbio)
1418 {
1419         struct promote_op *op = NULL;
1420         struct bio *bio;
1421         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1422         int ret;
1423
1424         if (!percpu_ref_tryget(&c->writes))
1425                 return NULL;
1426
1427         op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1428         if (!op)
1429                 goto err;
1430
1431         op->start_time = local_clock();
1432         op->pos = pos;
1433
1434         /*
1435          * We don't use the mempool here because extents that aren't
1436          * checksummed or compressed can be too big for the mempool:
1437          */
1438         *rbio = kzalloc(sizeof(struct bch_read_bio) +
1439                         sizeof(struct bio_vec) * pages,
1440                         GFP_NOIO);
1441         if (!*rbio)
1442                 goto err;
1443
1444         rbio_init(&(*rbio)->bio, opts);
1445         bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages);
1446
1447         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1448                                  GFP_NOIO))
1449                 goto err;
1450
1451         (*rbio)->bounce         = true;
1452         (*rbio)->split          = true;
1453         (*rbio)->kmalloc        = true;
1454
1455         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1456                                           bch_promote_params))
1457                 goto err;
1458
1459         bio = &op->write.op.wbio.bio;
1460         bio_init(bio, bio->bi_inline_vecs, pages);
1461
1462         ret = bch2_migrate_write_init(c, &op->write,
1463                         writepoint_hashed((unsigned long) current),
1464                         opts,
1465                         DATA_PROMOTE,
1466                         (struct data_opts) {
1467                                 .target         = opts.promote_target,
1468                                 .nr_replicas    = 1,
1469                         },
1470                         btree_id, k);
1471         BUG_ON(ret);
1472
1473         return op;
1474 err:
1475         if (*rbio)
1476                 bio_free_pages(&(*rbio)->bio);
1477         kfree(*rbio);
1478         *rbio = NULL;
1479         kfree(op);
1480         percpu_ref_put(&c->writes);
1481         return NULL;
1482 }
1483
1484 noinline
1485 static struct promote_op *promote_alloc(struct bch_fs *c,
1486                                                struct bvec_iter iter,
1487                                                struct bkey_s_c k,
1488                                                struct extent_ptr_decoded *pick,
1489                                                struct bch_io_opts opts,
1490                                                unsigned flags,
1491                                                struct bch_read_bio **rbio,
1492                                                bool *bounce,
1493                                                bool *read_full)
1494 {
1495         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1496         /* data might have to be decompressed in the write path: */
1497         unsigned sectors = promote_full
1498                 ? max(pick->crc.compressed_size, pick->crc.live_size)
1499                 : bvec_iter_sectors(iter);
1500         struct bpos pos = promote_full
1501                 ? bkey_start_pos(k.k)
1502                 : POS(k.k->p.inode, iter.bi_sector);
1503         struct promote_op *promote;
1504
1505         if (!should_promote(c, k, pos, opts, flags))
1506                 return NULL;
1507
1508         promote = __promote_alloc(c,
1509                                   k.k->type == KEY_TYPE_reflink_v
1510                                   ? BTREE_ID_reflink
1511                                   : BTREE_ID_extents,
1512                                   k, pos, pick, opts, sectors, rbio);
1513         if (!promote)
1514                 return NULL;
1515
1516         *bounce         = true;
1517         *read_full      = promote_full;
1518         return promote;
1519 }
1520
1521 /* Read */
1522
1523 #define READ_RETRY_AVOID        1
1524 #define READ_RETRY              2
1525 #define READ_ERR                3
1526
1527 enum rbio_context {
1528         RBIO_CONTEXT_NULL,
1529         RBIO_CONTEXT_HIGHPRI,
1530         RBIO_CONTEXT_UNBOUND,
1531 };
1532
1533 static inline struct bch_read_bio *
1534 bch2_rbio_parent(struct bch_read_bio *rbio)
1535 {
1536         return rbio->split ? rbio->parent : rbio;
1537 }
1538
1539 __always_inline
1540 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1541                            enum rbio_context context,
1542                            struct workqueue_struct *wq)
1543 {
1544         if (context <= rbio->context) {
1545                 fn(&rbio->work);
1546         } else {
1547                 rbio->work.func         = fn;
1548                 rbio->context           = context;
1549                 queue_work(wq, &rbio->work);
1550         }
1551 }
1552
1553 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1554 {
1555         BUG_ON(rbio->bounce && !rbio->split);
1556
1557         if (rbio->promote)
1558                 promote_free(rbio->c, rbio->promote);
1559         rbio->promote = NULL;
1560
1561         if (rbio->bounce)
1562                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1563
1564         if (rbio->split) {
1565                 struct bch_read_bio *parent = rbio->parent;
1566
1567                 if (rbio->kmalloc)
1568                         kfree(rbio);
1569                 else
1570                         bio_put(&rbio->bio);
1571
1572                 rbio = parent;
1573         }
1574
1575         return rbio;
1576 }
1577
1578 /*
1579  * Only called on a top level bch_read_bio to complete an entire read request,
1580  * not a split:
1581  */
1582 static void bch2_rbio_done(struct bch_read_bio *rbio)
1583 {
1584         if (rbio->start_time)
1585                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1586                                        rbio->start_time);
1587         bio_endio(&rbio->bio);
1588 }
1589
1590 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1591                                      struct bvec_iter bvec_iter,
1592                                      struct bch_io_failures *failed,
1593                                      unsigned flags)
1594 {
1595         struct btree_trans trans;
1596         struct btree_iter iter;
1597         struct bkey_buf sk;
1598         struct bkey_s_c k;
1599         int ret;
1600
1601         flags &= ~BCH_READ_LAST_FRAGMENT;
1602         flags |= BCH_READ_MUST_CLONE;
1603
1604         bch2_bkey_buf_init(&sk);
1605         bch2_trans_init(&trans, c, 0, 0);
1606
1607         bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1608                              rbio->read_pos, BTREE_ITER_SLOTS);
1609 retry:
1610         rbio->bio.bi_status = 0;
1611
1612         k = bch2_btree_iter_peek_slot(&iter);
1613         if (bkey_err(k))
1614                 goto err;
1615
1616         bch2_bkey_buf_reassemble(&sk, c, k);
1617         k = bkey_i_to_s_c(sk.k);
1618         bch2_trans_unlock(&trans);
1619
1620         if (!bch2_bkey_matches_ptr(c, k,
1621                                    rbio->pick.ptr,
1622                                    rbio->data_pos.offset -
1623                                    rbio->pick.crc.offset)) {
1624                 /* extent we wanted to read no longer exists: */
1625                 rbio->hole = true;
1626                 goto out;
1627         }
1628
1629         ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1630                                  rbio->read_pos,
1631                                  rbio->data_btree,
1632                                  k, 0, failed, flags);
1633         if (ret == READ_RETRY)
1634                 goto retry;
1635         if (ret)
1636                 goto err;
1637 out:
1638         bch2_rbio_done(rbio);
1639         bch2_trans_iter_exit(&trans, &iter);
1640         bch2_trans_exit(&trans);
1641         bch2_bkey_buf_exit(&sk, c);
1642         return;
1643 err:
1644         rbio->bio.bi_status = BLK_STS_IOERR;
1645         goto out;
1646 }
1647
1648 static void bch2_rbio_retry(struct work_struct *work)
1649 {
1650         struct bch_read_bio *rbio =
1651                 container_of(work, struct bch_read_bio, work);
1652         struct bch_fs *c        = rbio->c;
1653         struct bvec_iter iter   = rbio->bvec_iter;
1654         unsigned flags          = rbio->flags;
1655         subvol_inum inum = {
1656                 .subvol = rbio->subvol,
1657                 .inum   = rbio->read_pos.inode,
1658         };
1659         struct bch_io_failures failed = { .nr = 0 };
1660
1661         trace_read_retry(&rbio->bio);
1662
1663         if (rbio->retry == READ_RETRY_AVOID)
1664                 bch2_mark_io_failure(&failed, &rbio->pick);
1665
1666         rbio->bio.bi_status = 0;
1667
1668         rbio = bch2_rbio_free(rbio);
1669
1670         flags |= BCH_READ_IN_RETRY;
1671         flags &= ~BCH_READ_MAY_PROMOTE;
1672
1673         if (flags & BCH_READ_NODECODE) {
1674                 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1675         } else {
1676                 flags &= ~BCH_READ_LAST_FRAGMENT;
1677                 flags |= BCH_READ_MUST_CLONE;
1678
1679                 __bch2_read(c, rbio, iter, inum, &failed, flags);
1680         }
1681 }
1682
1683 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1684                             blk_status_t error)
1685 {
1686         rbio->retry = retry;
1687
1688         if (rbio->flags & BCH_READ_IN_RETRY)
1689                 return;
1690
1691         if (retry == READ_ERR) {
1692                 rbio = bch2_rbio_free(rbio);
1693
1694                 rbio->bio.bi_status = error;
1695                 bch2_rbio_done(rbio);
1696         } else {
1697                 bch2_rbio_punt(rbio, bch2_rbio_retry,
1698                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1699         }
1700 }
1701
1702 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1703                                    struct bch_read_bio *rbio)
1704 {
1705         struct bch_fs *c = rbio->c;
1706         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1707         struct bch_extent_crc_unpacked new_crc;
1708         struct btree_iter iter;
1709         struct bkey_i *new;
1710         struct bkey_s_c k;
1711         int ret = 0;
1712
1713         if (crc_is_compressed(rbio->pick.crc))
1714                 return 0;
1715
1716         bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1717                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1718         k = bch2_btree_iter_peek_slot(&iter);
1719         if ((ret = bkey_err(k)))
1720                 goto out;
1721
1722         if (bversion_cmp(k.k->version, rbio->version) ||
1723             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1724                 goto out;
1725
1726         /* Extent was merged? */
1727         if (bkey_start_offset(k.k) < data_offset ||
1728             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1729                 goto out;
1730
1731         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1732                         rbio->pick.crc, NULL, &new_crc,
1733                         bkey_start_offset(k.k) - data_offset, k.k->size,
1734                         rbio->pick.crc.csum_type)) {
1735                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1736                 ret = 0;
1737                 goto out;
1738         }
1739
1740         /*
1741          * going to be temporarily appending another checksum entry:
1742          */
1743         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1744                                  sizeof(struct bch_extent_crc128));
1745         if ((ret = PTR_ERR_OR_ZERO(new)))
1746                 goto out;
1747
1748         bkey_reassemble(new, k);
1749
1750         if (!bch2_bkey_narrow_crcs(new, new_crc))
1751                 goto out;
1752
1753         ret = bch2_trans_update(trans, &iter, new,
1754                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1755 out:
1756         bch2_trans_iter_exit(trans, &iter);
1757         return ret;
1758 }
1759
1760 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1761 {
1762         bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1763                       __bch2_rbio_narrow_crcs(&trans, rbio));
1764 }
1765
1766 /* Inner part that may run in process context */
1767 static void __bch2_read_endio(struct work_struct *work)
1768 {
1769         struct bch_read_bio *rbio =
1770                 container_of(work, struct bch_read_bio, work);
1771         struct bch_fs *c        = rbio->c;
1772         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1773         struct bio *src         = &rbio->bio;
1774         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
1775         struct bvec_iter dst_iter = rbio->bvec_iter;
1776         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1777         struct nonce nonce = extent_nonce(rbio->version, crc);
1778         unsigned nofs_flags;
1779         struct bch_csum csum;
1780         int ret;
1781
1782         nofs_flags = memalloc_nofs_save();
1783
1784         /* Reset iterator for checksumming and copying bounced data: */
1785         if (rbio->bounce) {
1786                 src->bi_iter.bi_size            = crc.compressed_size << 9;
1787                 src->bi_iter.bi_idx             = 0;
1788                 src->bi_iter.bi_bvec_done       = 0;
1789         } else {
1790                 src->bi_iter                    = rbio->bvec_iter;
1791         }
1792
1793         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1794         if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1795                 goto csum_err;
1796
1797         /*
1798          * XXX
1799          * We need to rework the narrow_crcs path to deliver the read completion
1800          * first, and then punt to a different workqueue, otherwise we're
1801          * holding up reads while doing btree updates which is bad for memory
1802          * reclaim.
1803          */
1804         if (unlikely(rbio->narrow_crcs))
1805                 bch2_rbio_narrow_crcs(rbio);
1806
1807         if (rbio->flags & BCH_READ_NODECODE)
1808                 goto nodecode;
1809
1810         /* Adjust crc to point to subset of data we want: */
1811         crc.offset     += rbio->offset_into_extent;
1812         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
1813
1814         if (crc_is_compressed(crc)) {
1815                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1816                 if (ret)
1817                         goto decrypt_err;
1818
1819                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1820                         goto decompression_err;
1821         } else {
1822                 /* don't need to decrypt the entire bio: */
1823                 nonce = nonce_add(nonce, crc.offset << 9);
1824                 bio_advance(src, crc.offset << 9);
1825
1826                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1827                 src->bi_iter.bi_size = dst_iter.bi_size;
1828
1829                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1830                 if (ret)
1831                         goto decrypt_err;
1832
1833                 if (rbio->bounce) {
1834                         struct bvec_iter src_iter = src->bi_iter;
1835                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1836                 }
1837         }
1838
1839         if (rbio->promote) {
1840                 /*
1841                  * Re encrypt data we decrypted, so it's consistent with
1842                  * rbio->crc:
1843                  */
1844                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1845                 if (ret)
1846                         goto decrypt_err;
1847
1848                 promote_start(rbio->promote, rbio);
1849                 rbio->promote = NULL;
1850         }
1851 nodecode:
1852         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1853                 rbio = bch2_rbio_free(rbio);
1854                 bch2_rbio_done(rbio);
1855         }
1856 out:
1857         memalloc_nofs_restore(nofs_flags);
1858         return;
1859 csum_err:
1860         /*
1861          * Checksum error: if the bio wasn't bounced, we may have been
1862          * reading into buffers owned by userspace (that userspace can
1863          * scribble over) - retry the read, bouncing it this time:
1864          */
1865         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1866                 rbio->flags |= BCH_READ_MUST_BOUNCE;
1867                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1868                 goto out;
1869         }
1870
1871         bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1872                 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1873                 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1874                 csum.hi, csum.lo, crc.csum_type);
1875         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1876         goto out;
1877 decompression_err:
1878         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1879                                  "decompression error");
1880         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1881         goto out;
1882 decrypt_err:
1883         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1884                                  "decrypt error");
1885         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1886         goto out;
1887 }
1888
1889 static void bch2_read_endio(struct bio *bio)
1890 {
1891         struct bch_read_bio *rbio =
1892                 container_of(bio, struct bch_read_bio, bio);
1893         struct bch_fs *c        = rbio->c;
1894         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1895         struct workqueue_struct *wq = NULL;
1896         enum rbio_context context = RBIO_CONTEXT_NULL;
1897
1898         if (rbio->have_ioref) {
1899                 bch2_latency_acct(ca, rbio->submit_time, READ);
1900                 percpu_ref_put(&ca->io_ref);
1901         }
1902
1903         if (!rbio->split)
1904                 rbio->bio.bi_end_io = rbio->end_io;
1905
1906         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1907                                     rbio->read_pos.inode,
1908                                     rbio->read_pos.offset,
1909                                     "data read error: %s",
1910                                bch2_blk_status_to_str(bio->bi_status))) {
1911                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1912                 return;
1913         }
1914
1915         if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1916             ptr_stale(ca, &rbio->pick.ptr)) {
1917                 atomic_long_inc(&c->read_realloc_races);
1918
1919                 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1920                         bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1921                 else
1922                         bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1923                 return;
1924         }
1925
1926         if (rbio->narrow_crcs ||
1927             crc_is_compressed(rbio->pick.crc) ||
1928             bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1929                 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1930         else if (rbio->pick.crc.csum_type)
1931                 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1932
1933         bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1934 }
1935
1936 int __bch2_read_indirect_extent(struct btree_trans *trans,
1937                                 unsigned *offset_into_extent,
1938                                 struct bkey_buf *orig_k)
1939 {
1940         struct btree_iter iter;
1941         struct bkey_s_c k;
1942         u64 reflink_offset;
1943         int ret;
1944
1945         reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1946                 *offset_into_extent;
1947
1948         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
1949                              POS(0, reflink_offset),
1950                              BTREE_ITER_SLOTS);
1951         k = bch2_btree_iter_peek_slot(&iter);
1952         ret = bkey_err(k);
1953         if (ret)
1954                 goto err;
1955
1956         if (k.k->type != KEY_TYPE_reflink_v &&
1957             k.k->type != KEY_TYPE_indirect_inline_data) {
1958                 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
1959                         "%llu len %u points to nonexistent indirect extent %llu",
1960                         orig_k->k->k.p.offset,
1961                         orig_k->k->k.size,
1962                         reflink_offset);
1963                 bch2_inconsistent_error(trans->c);
1964                 ret = -EIO;
1965                 goto err;
1966         }
1967
1968         *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
1969         bch2_bkey_buf_reassemble(orig_k, trans->c, k);
1970 err:
1971         bch2_trans_iter_exit(trans, &iter);
1972         return ret;
1973 }
1974
1975 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
1976                                                    struct bkey_s_c k,
1977                                                    struct bch_extent_ptr ptr)
1978 {
1979         struct bch_fs *c = trans->c;
1980         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
1981         struct btree_iter iter;
1982         struct printbuf buf = PRINTBUF;
1983         int ret;
1984
1985         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1986                              PTR_BUCKET_POS(c, &ptr),
1987                              BTREE_ITER_CACHED);
1988
1989         pr_buf(&buf, "Attempting to read from stale dirty pointer:");
1990         pr_indent_push(&buf, 2);
1991         pr_newline(&buf);
1992
1993         bch2_bkey_val_to_text(&buf, c, k);
1994         pr_newline(&buf);
1995
1996         pr_buf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
1997
1998         ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
1999         if (!ret) {
2000                 pr_newline(&buf);
2001                 bch2_bkey_val_to_text(&buf, c, k);
2002         }
2003
2004         bch2_fs_inconsistent(c, "%s", buf.buf);
2005
2006         bch2_trans_iter_exit(trans, &iter);
2007         printbuf_exit(&buf);
2008 }
2009
2010 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2011                        struct bvec_iter iter, struct bpos read_pos,
2012                        enum btree_id data_btree, struct bkey_s_c k,
2013                        unsigned offset_into_extent,
2014                        struct bch_io_failures *failed, unsigned flags)
2015 {
2016         struct bch_fs *c = trans->c;
2017         struct extent_ptr_decoded pick;
2018         struct bch_read_bio *rbio = NULL;
2019         struct bch_dev *ca = NULL;
2020         struct promote_op *promote = NULL;
2021         bool bounce = false, read_full = false, narrow_crcs = false;
2022         struct bpos data_pos = bkey_start_pos(k.k);
2023         int pick_ret;
2024
2025         if (bkey_extent_is_inline_data(k.k)) {
2026                 unsigned bytes = min_t(unsigned, iter.bi_size,
2027                                        bkey_inline_data_bytes(k.k));
2028
2029                 swap(iter.bi_size, bytes);
2030                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2031                 swap(iter.bi_size, bytes);
2032                 bio_advance_iter(&orig->bio, &iter, bytes);
2033                 zero_fill_bio_iter(&orig->bio, iter);
2034                 goto out_read_done;
2035         }
2036 retry_pick:
2037         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2038
2039         /* hole or reservation - just zero fill: */
2040         if (!pick_ret)
2041                 goto hole;
2042
2043         if (pick_ret < 0) {
2044                 bch_err_inum_ratelimited(c, k.k->p.inode,
2045                                          "no device to read from");
2046                 goto err;
2047         }
2048
2049         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2050
2051         /*
2052          * Stale dirty pointers are treated as IO errors, but @failed isn't
2053          * allocated unless we're in the retry path - so if we're not in the
2054          * retry path, don't check here, it'll be caught in bch2_read_endio()
2055          * and we'll end up in the retry path:
2056          */
2057         if ((flags & BCH_READ_IN_RETRY) &&
2058             !pick.ptr.cached &&
2059             unlikely(ptr_stale(ca, &pick.ptr))) {
2060                 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2061                 bch2_mark_io_failure(failed, &pick);
2062                 goto retry_pick;
2063         }
2064
2065         /*
2066          * Unlock the iterator while the btree node's lock is still in
2067          * cache, before doing the IO:
2068          */
2069         bch2_trans_unlock(trans);
2070
2071         if (flags & BCH_READ_NODECODE) {
2072                 /*
2073                  * can happen if we retry, and the extent we were going to read
2074                  * has been merged in the meantime:
2075                  */
2076                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2077                         goto hole;
2078
2079                 iter.bi_size    = pick.crc.compressed_size << 9;
2080                 goto get_bio;
2081         }
2082
2083         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2084             bio_flagged(&orig->bio, BIO_CHAIN))
2085                 flags |= BCH_READ_MUST_CLONE;
2086
2087         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2088                 bch2_can_narrow_extent_crcs(k, pick.crc);
2089
2090         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2091                 flags |= BCH_READ_MUST_BOUNCE;
2092
2093         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2094
2095         if (crc_is_compressed(pick.crc) ||
2096             (pick.crc.csum_type != BCH_CSUM_none &&
2097              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2098               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2099                (flags & BCH_READ_USER_MAPPED)) ||
2100               (flags & BCH_READ_MUST_BOUNCE)))) {
2101                 read_full = true;
2102                 bounce = true;
2103         }
2104
2105         if (orig->opts.promote_target)
2106                 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2107                                         &rbio, &bounce, &read_full);
2108
2109         if (!read_full) {
2110                 EBUG_ON(crc_is_compressed(pick.crc));
2111                 EBUG_ON(pick.crc.csum_type &&
2112                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2113                          bvec_iter_sectors(iter) != pick.crc.live_size ||
2114                          pick.crc.offset ||
2115                          offset_into_extent));
2116
2117                 data_pos.offset += offset_into_extent;
2118                 pick.ptr.offset += pick.crc.offset +
2119                         offset_into_extent;
2120                 offset_into_extent              = 0;
2121                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
2122                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
2123                 pick.crc.offset                 = 0;
2124                 pick.crc.live_size              = bvec_iter_sectors(iter);
2125                 offset_into_extent              = 0;
2126         }
2127 get_bio:
2128         if (rbio) {
2129                 /*
2130                  * promote already allocated bounce rbio:
2131                  * promote needs to allocate a bio big enough for uncompressing
2132                  * data in the write path, but we're not going to use it all
2133                  * here:
2134                  */
2135                 EBUG_ON(rbio->bio.bi_iter.bi_size <
2136                        pick.crc.compressed_size << 9);
2137                 rbio->bio.bi_iter.bi_size =
2138                         pick.crc.compressed_size << 9;
2139         } else if (bounce) {
2140                 unsigned sectors = pick.crc.compressed_size;
2141
2142                 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
2143                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
2144                                                   &c->bio_read_split),
2145                                  orig->opts);
2146
2147                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2148                 rbio->bounce    = true;
2149                 rbio->split     = true;
2150         } else if (flags & BCH_READ_MUST_CLONE) {
2151                 /*
2152                  * Have to clone if there were any splits, due to error
2153                  * reporting issues (if a split errored, and retrying didn't
2154                  * work, when it reports the error to its parent (us) we don't
2155                  * know if the error was from our bio, and we should retry, or
2156                  * from the whole bio, in which case we don't want to retry and
2157                  * lose the error)
2158                  */
2159                 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
2160                                                 &c->bio_read_split),
2161                                  orig->opts);
2162                 rbio->bio.bi_iter = iter;
2163                 rbio->split     = true;
2164         } else {
2165                 rbio = orig;
2166                 rbio->bio.bi_iter = iter;
2167                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2168         }
2169
2170         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2171
2172         rbio->c                 = c;
2173         rbio->submit_time       = local_clock();
2174         if (rbio->split)
2175                 rbio->parent    = orig;
2176         else
2177                 rbio->end_io    = orig->bio.bi_end_io;
2178         rbio->bvec_iter         = iter;
2179         rbio->offset_into_extent= offset_into_extent;
2180         rbio->flags             = flags;
2181         rbio->have_ioref        = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2182         rbio->narrow_crcs       = narrow_crcs;
2183         rbio->hole              = 0;
2184         rbio->retry             = 0;
2185         rbio->context           = 0;
2186         /* XXX: only initialize this if needed */
2187         rbio->devs_have         = bch2_bkey_devs(k);
2188         rbio->pick              = pick;
2189         rbio->subvol            = orig->subvol;
2190         rbio->read_pos          = read_pos;
2191         rbio->data_btree        = data_btree;
2192         rbio->data_pos          = data_pos;
2193         rbio->version           = k.k->version;
2194         rbio->promote           = promote;
2195         INIT_WORK(&rbio->work, NULL);
2196
2197         rbio->bio.bi_opf        = orig->bio.bi_opf;
2198         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2199         rbio->bio.bi_end_io     = bch2_read_endio;
2200
2201         if (rbio->bounce)
2202                 trace_read_bounce(&rbio->bio);
2203
2204         this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
2205         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2206
2207         /*
2208          * If it's being moved internally, we don't want to flag it as a cache
2209          * hit:
2210          */
2211         if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2212                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2213                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
2214
2215         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2216                 bio_inc_remaining(&orig->bio);
2217                 trace_read_split(&orig->bio);
2218         }
2219
2220         if (!rbio->pick.idx) {
2221                 if (!rbio->have_ioref) {
2222                         bch_err_inum_ratelimited(c, k.k->p.inode,
2223                                                  "no device to read from");
2224                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2225                         goto out;
2226                 }
2227
2228                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2229                              bio_sectors(&rbio->bio));
2230                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2231
2232                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2233                         submit_bio(&rbio->bio);
2234                 else
2235                         submit_bio_wait(&rbio->bio);
2236         } else {
2237                 /* Attempting reconstruct read: */
2238                 if (bch2_ec_read_extent(c, rbio)) {
2239                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2240                         goto out;
2241                 }
2242
2243                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2244                         bio_endio(&rbio->bio);
2245         }
2246 out:
2247         if (likely(!(flags & BCH_READ_IN_RETRY))) {
2248                 return 0;
2249         } else {
2250                 int ret;
2251
2252                 rbio->context = RBIO_CONTEXT_UNBOUND;
2253                 bch2_read_endio(&rbio->bio);
2254
2255                 ret = rbio->retry;
2256                 rbio = bch2_rbio_free(rbio);
2257
2258                 if (ret == READ_RETRY_AVOID) {
2259                         bch2_mark_io_failure(failed, &pick);
2260                         ret = READ_RETRY;
2261                 }
2262
2263                 if (!ret)
2264                         goto out_read_done;
2265
2266                 return ret;
2267         }
2268
2269 err:
2270         if (flags & BCH_READ_IN_RETRY)
2271                 return READ_ERR;
2272
2273         orig->bio.bi_status = BLK_STS_IOERR;
2274         goto out_read_done;
2275
2276 hole:
2277         /*
2278          * won't normally happen in the BCH_READ_NODECODE
2279          * (bch2_move_extent()) path, but if we retry and the extent we wanted
2280          * to read no longer exists we have to signal that:
2281          */
2282         if (flags & BCH_READ_NODECODE)
2283                 orig->hole = true;
2284
2285         zero_fill_bio_iter(&orig->bio, iter);
2286 out_read_done:
2287         if (flags & BCH_READ_LAST_FRAGMENT)
2288                 bch2_rbio_done(orig);
2289         return 0;
2290 }
2291
2292 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2293                  struct bvec_iter bvec_iter, subvol_inum inum,
2294                  struct bch_io_failures *failed, unsigned flags)
2295 {
2296         struct btree_trans trans;
2297         struct btree_iter iter;
2298         struct bkey_buf sk;
2299         struct bkey_s_c k;
2300         u32 snapshot;
2301         int ret;
2302
2303         BUG_ON(flags & BCH_READ_NODECODE);
2304
2305         bch2_bkey_buf_init(&sk);
2306         bch2_trans_init(&trans, c, 0, 0);
2307 retry:
2308         bch2_trans_begin(&trans);
2309         iter = (struct btree_iter) { NULL };
2310
2311         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2312         if (ret)
2313                 goto err;
2314
2315         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2316                              SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2317                              BTREE_ITER_SLOTS);
2318         while (1) {
2319                 unsigned bytes, sectors, offset_into_extent;
2320                 enum btree_id data_btree = BTREE_ID_extents;
2321
2322                 /*
2323                  * read_extent -> io_time_reset may cause a transaction restart
2324                  * without returning an error, we need to check for that here:
2325                  */
2326                 if (!bch2_trans_relock(&trans)) {
2327                         ret = -EINTR;
2328                         break;
2329                 }
2330
2331                 bch2_btree_iter_set_pos(&iter,
2332                                 POS(inum.inum, bvec_iter.bi_sector));
2333
2334                 k = bch2_btree_iter_peek_slot(&iter);
2335                 ret = bkey_err(k);
2336                 if (ret)
2337                         break;
2338
2339                 offset_into_extent = iter.pos.offset -
2340                         bkey_start_offset(k.k);
2341                 sectors = k.k->size - offset_into_extent;
2342
2343                 bch2_bkey_buf_reassemble(&sk, c, k);
2344
2345                 ret = bch2_read_indirect_extent(&trans, &data_btree,
2346                                         &offset_into_extent, &sk);
2347                 if (ret)
2348                         break;
2349
2350                 k = bkey_i_to_s_c(sk.k);
2351
2352                 /*
2353                  * With indirect extents, the amount of data to read is the min
2354                  * of the original extent and the indirect extent:
2355                  */
2356                 sectors = min(sectors, k.k->size - offset_into_extent);
2357
2358                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2359                 swap(bvec_iter.bi_size, bytes);
2360
2361                 if (bvec_iter.bi_size == bytes)
2362                         flags |= BCH_READ_LAST_FRAGMENT;
2363
2364                 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2365                                          data_btree, k,
2366                                          offset_into_extent, failed, flags);
2367                 if (ret)
2368                         break;
2369
2370                 if (flags & BCH_READ_LAST_FRAGMENT)
2371                         break;
2372
2373                 swap(bvec_iter.bi_size, bytes);
2374                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2375
2376                 ret = btree_trans_too_many_iters(&trans);
2377                 if (ret)
2378                         break;
2379         }
2380 err:
2381         bch2_trans_iter_exit(&trans, &iter);
2382
2383         if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
2384                 goto retry;
2385
2386         bch2_trans_exit(&trans);
2387         bch2_bkey_buf_exit(&sk, c);
2388
2389         if (ret) {
2390                 bch_err_inum_ratelimited(c, inum.inum,
2391                                          "read error %i from btree lookup", ret);
2392                 rbio->bio.bi_status = BLK_STS_IOERR;
2393                 bch2_rbio_done(rbio);
2394         }
2395 }
2396
2397 void bch2_fs_io_exit(struct bch_fs *c)
2398 {
2399         if (c->promote_table.tbl)
2400                 rhashtable_destroy(&c->promote_table);
2401         mempool_exit(&c->bio_bounce_pages);
2402         bioset_exit(&c->bio_write);
2403         bioset_exit(&c->bio_read_split);
2404         bioset_exit(&c->bio_read);
2405 }
2406
2407 int bch2_fs_io_init(struct bch_fs *c)
2408 {
2409         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2410                         BIOSET_NEED_BVECS) ||
2411             bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2412                         BIOSET_NEED_BVECS) ||
2413             bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2414                         BIOSET_NEED_BVECS) ||
2415             mempool_init_page_pool(&c->bio_bounce_pages,
2416                                    max_t(unsigned,
2417                                          c->opts.btree_node_size,
2418                                          c->opts.encoded_extent_max) /
2419                                    PAGE_SIZE, 0) ||
2420             rhashtable_init(&c->promote_table, &bch_promote_params))
2421                 return -ENOMEM;
2422
2423         return 0;
2424 }