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         struct open_bucket *ob;
 669         unsigned i;
 670
 671         BUG_ON(crc.compressed_size > wp->sectors_free);
 672         wp->sectors_free -= crc.compressed_size;
 673         op->pos.offset += crc.uncompressed_size;
 674
 675         e = bkey_extent_init(op->insert_keys.top);
 676         e->k.p          = op->pos;
 677         e->k.size       = crc.uncompressed_size;
 678         e->k.version    = version;
 679
 680         if (crc.csum_type ||
 681             crc.compression_type ||
 682             crc.nonce)
 683                 bch2_extent_crc_append(&e->k_i, crc);
 684
 685         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 686                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 687                 union bch_extent_entry *end =
 688                         bkey_val_end(bkey_i_to_s(&e->k_i));
 689
 690                 end->ptr = ob->ptr;
 691                 end->ptr.type = 1 << BCH_EXTENT_ENTRY_ptr;
 692                 end->ptr.cached = !ca->mi.durability ||
 693                         (op->flags & BCH_WRITE_CACHED) != 0;
 694                 end->ptr.offset += ca->mi.bucket_size - ob->sectors_free;
 695
 696                 e->k.u64s++;
 697
 698                 BUG_ON(crc.compressed_size > ob->sectors_free);
 699                 ob->sectors_free -= crc.compressed_size;
 700         }
 701
 702         bch2_keylist_push(&op->insert_keys);
 703 }
 704
 705 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 706                                         struct write_point *wp,
 707                                         struct bio *src,
 708                                         bool *page_alloc_failed,
 709                                         void *buf)
 710 {
 711         struct bch_write_bio *wbio;
 712         struct bio *bio;
 713         unsigned output_available =
 714                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 715         unsigned pages = DIV_ROUND_UP(output_available +
 716                                       (buf
 717                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 718                                        : 0), PAGE_SIZE);
 719
 720         pages = min(pages, BIO_MAX_VECS);
 721
 722         bio = bio_alloc_bioset(GFP_NOIO, pages, &c->bio_write);
 723         wbio                    = wbio_init(bio);
 724         wbio->put_bio           = true;
 725         /* copy WRITE_SYNC flag */
 726         wbio->bio.bi_opf        = src->bi_opf;
 727
 728         if (buf) {
 729                 bch2_bio_map(bio, buf, output_available);
 730                 return bio;
 731         }
 732
 733         wbio->bounce            = true;
 734
 735         /*
 736          * We can't use mempool for more than c->sb.encoded_extent_max
 737          * worth of pages, but we'd like to allocate more if we can:
 738          */
 739         bch2_bio_alloc_pages_pool(c, bio,
 740                                   min_t(unsigned, output_available,
 741                                         c->opts.encoded_extent_max));
 742
 743         if (bio->bi_iter.bi_size < output_available)
 744                 *page_alloc_failed =
 745                         bch2_bio_alloc_pages(bio,
 746                                              output_available -
 747                                              bio->bi_iter.bi_size,
 748                                              GFP_NOFS) != 0;
 749
 750         return bio;
 751 }
 752
 753 static int bch2_write_rechecksum(struct bch_fs *c,
 754                                  struct bch_write_op *op,
 755                                  unsigned new_csum_type)
 756 {
 757         struct bio *bio = &op->wbio.bio;
 758         struct bch_extent_crc_unpacked new_crc;
 759         int ret;
 760
 761         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 762
 763         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 764             bch2_csum_type_is_encryption(new_csum_type))
 765                 new_csum_type = op->crc.csum_type;
 766
 767         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 768                                   NULL, &new_crc,
 769                                   op->crc.offset, op->crc.live_size,
 770                                   new_csum_type);
 771         if (ret)
 772                 return ret;
 773
 774         bio_advance(bio, op->crc.offset << 9);
 775         bio->bi_iter.bi_size = op->crc.live_size << 9;
 776         op->crc = new_crc;
 777         return 0;
 778 }
 779
 780 static int bch2_write_decrypt(struct bch_write_op *op)
 781 {
 782         struct bch_fs *c = op->c;
 783         struct nonce nonce = extent_nonce(op->version, op->crc);
 784         struct bch_csum csum;
 785
 786         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 787                 return 0;
 788
 789         /*
 790          * If we need to decrypt data in the write path, we'll no longer be able
 791          * to verify the existing checksum (poly1305 mac, in this case) after
 792          * it's decrypted - this is the last point we'll be able to reverify the
 793          * checksum:
 794          */
 795         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 796         if (bch2_crc_cmp(op->crc.csum, csum))
 797                 return -EIO;
 798
 799         bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 800         op->crc.csum_type = 0;
 801         op->crc.csum = (struct bch_csum) { 0, 0 };
 802         return 0;
 803 }
 804
 805 static enum prep_encoded_ret {
 806         PREP_ENCODED_OK,
 807         PREP_ENCODED_ERR,
 808         PREP_ENCODED_CHECKSUM_ERR,
 809         PREP_ENCODED_DO_WRITE,
 810 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 811 {
 812         struct bch_fs *c = op->c;
 813         struct bio *bio = &op->wbio.bio;
 814
 815         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 816                 return PREP_ENCODED_OK;
 817
 818         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 819
 820         /* Can we just write the entire extent as is? */
 821         if (op->crc.uncompressed_size == op->crc.live_size &&
 822             op->crc.compressed_size <= wp->sectors_free &&
 823             (op->crc.compression_type == op->compression_type ||
 824              op->incompressible)) {
 825                 if (!crc_is_compressed(op->crc) &&
 826                     op->csum_type != op->crc.csum_type &&
 827                     bch2_write_rechecksum(c, op, op->csum_type))
 828                         return PREP_ENCODED_CHECKSUM_ERR;
 829
 830                 return PREP_ENCODED_DO_WRITE;
 831         }
 832
 833         /*
 834          * If the data is compressed and we couldn't write the entire extent as
 835          * is, we have to decompress it:
 836          */
 837         if (crc_is_compressed(op->crc)) {
 838                 struct bch_csum csum;
 839
 840                 if (bch2_write_decrypt(op))
 841                         return PREP_ENCODED_CHECKSUM_ERR;
 842
 843                 /* Last point we can still verify checksum: */
 844                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 845                                          extent_nonce(op->version, op->crc),
 846                                          bio);
 847                 if (bch2_crc_cmp(op->crc.csum, csum))
 848                         return PREP_ENCODED_CHECKSUM_ERR;
 849
 850                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 851                         return PREP_ENCODED_ERR;
 852         }
 853
 854         /*
 855          * No longer have compressed data after this point - data might be
 856          * encrypted:
 857          */
 858
 859         /*
 860          * If the data is checksummed and we're only writing a subset,
 861          * rechecksum and adjust bio to point to currently live data:
 862          */
 863         if ((op->crc.live_size != op->crc.uncompressed_size ||
 864              op->crc.csum_type != op->csum_type) &&
 865             bch2_write_rechecksum(c, op, op->csum_type))
 866                 return PREP_ENCODED_CHECKSUM_ERR;
 867
 868         /*
 869          * If we want to compress the data, it has to be decrypted:
 870          */
 871         if ((op->compression_type ||
 872              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 873              bch2_csum_type_is_encryption(op->csum_type)) &&
 874             bch2_write_decrypt(op))
 875                 return PREP_ENCODED_CHECKSUM_ERR;
 876
 877         return PREP_ENCODED_OK;
 878 }
 879
 880 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 881                              struct bio **_dst)
 882 {
 883         struct bch_fs *c = op->c;
 884         struct bio *src = &op->wbio.bio, *dst = src;
 885         struct bvec_iter saved_iter;
 886         void *ec_buf;
 887         unsigned total_output = 0, total_input = 0;
 888         bool bounce = false;
 889         bool page_alloc_failed = false;
 890         int ret, more = 0;
 891
 892         BUG_ON(!bio_sectors(src));
 893
 894         ec_buf = bch2_writepoint_ec_buf(c, wp);
 895
 896         switch (bch2_write_prep_encoded_data(op, wp)) {
 897         case PREP_ENCODED_OK:
 898                 break;
 899         case PREP_ENCODED_ERR:
 900                 ret = -EIO;
 901                 goto err;
 902         case PREP_ENCODED_CHECKSUM_ERR:
 903                 goto csum_err;
 904         case PREP_ENCODED_DO_WRITE:
 905                 /* XXX look for bug here */
 906                 if (ec_buf) {
 907                         dst = bch2_write_bio_alloc(c, wp, src,
 908                                                    &page_alloc_failed,
 909                                                    ec_buf);
 910                         bio_copy_data(dst, src);
 911                         bounce = true;
 912                 }
 913                 init_append_extent(op, wp, op->version, op->crc);
 914                 goto do_write;
 915         }
 916
 917         if (ec_buf ||
 918             op->compression_type ||
 919             (op->csum_type &&
 920              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 921             (bch2_csum_type_is_encryption(op->csum_type) &&
 922              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 923                 dst = bch2_write_bio_alloc(c, wp, src,
 924                                            &page_alloc_failed,
 925                                            ec_buf);
 926                 bounce = true;
 927         }
 928
 929         saved_iter = dst->bi_iter;
 930
 931         do {
 932                 struct bch_extent_crc_unpacked crc =
 933                         (struct bch_extent_crc_unpacked) { 0 };
 934                 struct bversion version = op->version;
 935                 size_t dst_len, src_len;
 936
 937                 if (page_alloc_failed &&
 938                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 939                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 940                         break;
 941
 942                 BUG_ON(op->compression_type &&
 943                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 944                        bch2_csum_type_is_encryption(op->crc.csum_type));
 945                 BUG_ON(op->compression_type && !bounce);
 946
 947                 crc.compression_type = op->incompressible
 948                         ? BCH_COMPRESSION_TYPE_incompressible
 949                         : op->compression_type
 950                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 951                                             op->compression_type)
 952                         : 0;
 953                 if (!crc_is_compressed(crc)) {
 954                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 955                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 956
 957                         if (op->csum_type)
 958                                 dst_len = min_t(unsigned, dst_len,
 959                                                 c->opts.encoded_extent_max);
 960
 961                         if (bounce) {
 962                                 swap(dst->bi_iter.bi_size, dst_len);
 963                                 bio_copy_data(dst, src);
 964                                 swap(dst->bi_iter.bi_size, dst_len);
 965                         }
 966
 967                         src_len = dst_len;
 968                 }
 969
 970                 BUG_ON(!src_len || !dst_len);
 971
 972                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 973                         if (bversion_zero(version)) {
 974                                 version.lo = atomic64_inc_return(&c->key_version);
 975                         } else {
 976                                 crc.nonce = op->nonce;
 977                                 op->nonce += src_len >> 9;
 978                         }
 979                 }
 980
 981                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 982                     !crc_is_compressed(crc) &&
 983                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
 984                     bch2_csum_type_is_encryption(op->csum_type)) {
 985                         /*
 986                          * Note: when we're using rechecksum(), we need to be
 987                          * checksumming @src because it has all the data our
 988                          * existing checksum covers - if we bounced (because we
 989                          * were trying to compress), @dst will only have the
 990                          * part of the data the new checksum will cover.
 991                          *
 992                          * But normally we want to be checksumming post bounce,
 993                          * because part of the reason for bouncing is so the
 994                          * data can't be modified (by userspace) while it's in
 995                          * flight.
 996                          */
 997                         if (bch2_rechecksum_bio(c, src, version, op->crc,
 998                                         &crc, &op->crc,
 999                                         src_len >> 9,
1000                                         bio_sectors(src) - (src_len >> 9),
1001                                         op->csum_type))
1002                                 goto csum_err;
1003                 } else {
1004                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1005                             bch2_rechecksum_bio(c, src, version, op->crc,
1006                                         NULL, &op->crc,
1007                                         src_len >> 9,
1008                                         bio_sectors(src) - (src_len >> 9),
1009                                         op->crc.csum_type))
1010                                 goto csum_err;
1011
1012                         crc.compressed_size     = dst_len >> 9;
1013                         crc.uncompressed_size   = src_len >> 9;
1014                         crc.live_size           = src_len >> 9;
1015
1016                         swap(dst->bi_iter.bi_size, dst_len);
1017                         bch2_encrypt_bio(c, op->csum_type,
1018                                          extent_nonce(version, crc), dst);
1019                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1020                                          extent_nonce(version, crc), dst);
1021                         crc.csum_type = op->csum_type;
1022                         swap(dst->bi_iter.bi_size, dst_len);
1023                 }
1024
1025                 init_append_extent(op, wp, version, crc);
1026
1027                 if (dst != src)
1028                         bio_advance(dst, dst_len);
1029                 bio_advance(src, src_len);
1030                 total_output    += dst_len;
1031                 total_input     += src_len;
1032         } while (dst->bi_iter.bi_size &&
1033                  src->bi_iter.bi_size &&
1034                  wp->sectors_free &&
1035                  !bch2_keylist_realloc(&op->insert_keys,
1036                                       op->inline_keys,
1037                                       ARRAY_SIZE(op->inline_keys),
1038                                       BKEY_EXTENT_U64s_MAX));
1039
1040         more = src->bi_iter.bi_size != 0;
1041
1042         dst->bi_iter = saved_iter;
1043
1044         if (dst == src && more) {
1045                 BUG_ON(total_output != total_input);
1046
1047                 dst = bio_split(src, total_input >> 9,
1048                                 GFP_NOIO, &c->bio_write);
1049                 wbio_init(dst)->put_bio = true;
1050                 /* copy WRITE_SYNC flag */
1051                 dst->bi_opf             = src->bi_opf;
1052         }
1053
1054         dst->bi_iter.bi_size = total_output;
1055 do_write:
1056         *_dst = dst;
1057         return more;
1058 csum_err:
1059         bch_err(c, "error verifying existing checksum while "
1060                 "rewriting existing data (memory corruption?)");
1061         ret = -EIO;
1062 err:
1063         if (to_wbio(dst)->bounce)
1064                 bch2_bio_free_pages_pool(c, dst);
1065         if (to_wbio(dst)->put_bio)
1066                 bio_put(dst);
1067
1068         return ret;
1069 }
1070
1071 static void __bch2_write(struct closure *cl)
1072 {
1073         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1074         struct bch_fs *c = op->c;
1075         struct write_point *wp;
1076         struct bio *bio;
1077         bool skip_put = true;
1078         unsigned nofs_flags;
1079         int ret;
1080
1081         nofs_flags = memalloc_nofs_save();
1082 again:
1083         memset(&op->failed, 0, sizeof(op->failed));
1084
1085         do {
1086                 struct bkey_i *key_to_write;
1087                 unsigned key_to_write_offset = op->insert_keys.top_p -
1088                         op->insert_keys.keys_p;
1089
1090                 /* +1 for possible cache device: */
1091                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1092                     ARRAY_SIZE(op->open_buckets.v))
1093                         goto flush_io;
1094
1095                 if (bch2_keylist_realloc(&op->insert_keys,
1096                                         op->inline_keys,
1097                                         ARRAY_SIZE(op->inline_keys),
1098                                         BKEY_EXTENT_U64s_MAX))
1099                         goto flush_io;
1100
1101                 if ((op->flags & BCH_WRITE_FROM_INTERNAL) &&
1102                     percpu_ref_is_dying(&c->writes)) {
1103                         ret = -EROFS;
1104                         goto err;
1105                 }
1106
1107                 /*
1108                  * The copygc thread is now global, which means it's no longer
1109                  * freeing up space on specific disks, which means that
1110                  * allocations for specific disks may hang arbitrarily long:
1111                  */
1112                 wp = bch2_alloc_sectors_start(c,
1113                         op->target,
1114                         op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1115                         op->write_point,
1116                         &op->devs_have,
1117                         op->nr_replicas,
1118                         op->nr_replicas_required,
1119                         op->alloc_reserve,
1120                         op->flags,
1121                         (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1122                                       BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1123                 EBUG_ON(!wp);
1124
1125                 if (unlikely(IS_ERR(wp))) {
1126                         if (unlikely(PTR_ERR(wp) != -EAGAIN)) {
1127                                 ret = PTR_ERR(wp);
1128                                 goto err;
1129                         }
1130
1131                         goto flush_io;
1132                 }
1133
1134                 /*
1135                  * It's possible for the allocator to fail, put us on the
1136                  * freelist waitlist, and then succeed in one of various retry
1137                  * paths: if that happens, we need to disable the skip_put
1138                  * optimization because otherwise there won't necessarily be a
1139                  * barrier before we free the bch_write_op:
1140                  */
1141                 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1142                         skip_put = false;
1143
1144                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1145                 ret = bch2_write_extent(op, wp, &bio);
1146                 bch2_alloc_sectors_done(c, wp);
1147
1148                 if (ret < 0)
1149                         goto err;
1150
1151                 if (ret) {
1152                         skip_put = false;
1153                 } else {
1154                         /*
1155                          * for the skip_put optimization this has to be set
1156                          * before we submit the bio:
1157                          */
1158                         op->flags |= BCH_WRITE_DONE;
1159                 }
1160
1161                 bio->bi_end_io  = bch2_write_endio;
1162                 bio->bi_private = &op->cl;
1163                 bio->bi_opf |= REQ_OP_WRITE;
1164
1165                 if (!skip_put)
1166                         closure_get(bio->bi_private);
1167                 else
1168                         op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1169
1170                 key_to_write = (void *) (op->insert_keys.keys_p +
1171                                          key_to_write_offset);
1172
1173                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1174                                           key_to_write);
1175         } while (ret);
1176
1177         if (!skip_put)
1178                 continue_at(cl, bch2_write_index, index_update_wq(op));
1179 out:
1180         memalloc_nofs_restore(nofs_flags);
1181         return;
1182 err:
1183         op->error = ret;
1184         op->flags |= BCH_WRITE_DONE;
1185
1186         continue_at(cl, bch2_write_index, index_update_wq(op));
1187         goto out;
1188 flush_io:
1189         /*
1190          * If the write can't all be submitted at once, we generally want to
1191          * block synchronously as that signals backpressure to the caller.
1192          *
1193          * However, if we're running out of a workqueue, we can't block here
1194          * because we'll be blocking other work items from completing:
1195          */
1196         if (current->flags & PF_WQ_WORKER) {
1197                 continue_at(cl, bch2_write_index, index_update_wq(op));
1198                 goto out;
1199         }
1200
1201         closure_sync(cl);
1202
1203         if (!bch2_keylist_empty(&op->insert_keys)) {
1204                 __bch2_write_index(op);
1205
1206                 if (op->error) {
1207                         op->flags |= BCH_WRITE_DONE;
1208                         continue_at_nobarrier(cl, bch2_write_done, NULL);
1209                         goto out;
1210                 }
1211         }
1212
1213         goto again;
1214 }
1215
1216 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1217 {
1218         struct closure *cl = &op->cl;
1219         struct bio *bio = &op->wbio.bio;
1220         struct bvec_iter iter;
1221         struct bkey_i_inline_data *id;
1222         unsigned sectors;
1223         int ret;
1224
1225         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1226
1227         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1228                                    ARRAY_SIZE(op->inline_keys),
1229                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1230         if (ret) {
1231                 op->error = ret;
1232                 goto err;
1233         }
1234
1235         sectors = bio_sectors(bio);
1236         op->pos.offset += sectors;
1237
1238         id = bkey_inline_data_init(op->insert_keys.top);
1239         id->k.p         = op->pos;
1240         id->k.version   = op->version;
1241         id->k.size      = sectors;
1242
1243         iter = bio->bi_iter;
1244         iter.bi_size = data_len;
1245         memcpy_from_bio(id->v.data, bio, iter);
1246
1247         while (data_len & 7)
1248                 id->v.data[data_len++] = '\0';
1249         set_bkey_val_bytes(&id->k, data_len);
1250         bch2_keylist_push(&op->insert_keys);
1251
1252         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1253         op->flags |= BCH_WRITE_DONE;
1254
1255         continue_at_nobarrier(cl, bch2_write_index, NULL);
1256         return;
1257 err:
1258         bch2_write_done(&op->cl);
1259 }
1260
1261 /**
1262  * bch_write - handle a write to a cache device or flash only volume
1263  *
1264  * This is the starting point for any data to end up in a cache device; it could
1265  * be from a normal write, or a writeback write, or a write to a flash only
1266  * volume - it's also used by the moving garbage collector to compact data in
1267  * mostly empty buckets.
1268  *
1269  * It first writes the data to the cache, creating a list of keys to be inserted
1270  * (if the data won't fit in a single open bucket, there will be multiple keys);
1271  * after the data is written it calls bch_journal, and after the keys have been
1272  * added to the next journal write they're inserted into the btree.
1273  *
1274  * If op->discard is true, instead of inserting the data it invalidates the
1275  * region of the cache represented by op->bio and op->inode.
1276  */
1277 void bch2_write(struct closure *cl)
1278 {
1279         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1280         struct bio *bio = &op->wbio.bio;
1281         struct bch_fs *c = op->c;
1282         unsigned data_len;
1283
1284         BUG_ON(!op->nr_replicas);
1285         BUG_ON(!op->write_point.v);
1286         BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1287
1288         op->start_time = local_clock();
1289         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1290         wbio_init(bio)->put_bio = false;
1291
1292         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1293                 bch_err_inum_ratelimited(c, op->pos.inode,
1294                                          "misaligned write");
1295                 op->error = -EIO;
1296                 goto err;
1297         }
1298
1299         if (c->opts.nochanges ||
1300             !percpu_ref_tryget(&c->writes)) {
1301                 op->error = -EROFS;
1302                 goto err;
1303         }
1304
1305         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1306
1307         data_len = min_t(u64, bio->bi_iter.bi_size,
1308                          op->new_i_size - (op->pos.offset << 9));
1309
1310         if (c->opts.inline_data &&
1311             data_len <= min(block_bytes(c) / 2, 1024U)) {
1312                 bch2_write_data_inline(op, data_len);
1313                 return;
1314         }
1315
1316         continue_at_nobarrier(cl, __bch2_write, NULL);
1317         return;
1318 err:
1319         bch2_disk_reservation_put(c, &op->res);
1320
1321         if (op->end_io) {
1322                 EBUG_ON(cl->parent);
1323                 closure_debug_destroy(cl);
1324                 op->end_io(op);
1325         } else {
1326                 closure_return(cl);
1327         }
1328 }
1329
1330 /* Cache promotion on read */
1331
1332 struct promote_op {
1333         struct closure          cl;
1334         struct rcu_head         rcu;
1335         u64                     start_time;
1336
1337         struct rhash_head       hash;
1338         struct bpos             pos;
1339
1340         struct migrate_write    write;
1341         struct bio_vec          bi_inline_vecs[0]; /* must be last */
1342 };
1343
1344 static const struct rhashtable_params bch_promote_params = {
1345         .head_offset    = offsetof(struct promote_op, hash),
1346         .key_offset     = offsetof(struct promote_op, pos),
1347         .key_len        = sizeof(struct bpos),
1348 };
1349
1350 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1351                                   struct bpos pos,
1352                                   struct bch_io_opts opts,
1353                                   unsigned flags)
1354 {
1355         if (!(flags & BCH_READ_MAY_PROMOTE))
1356                 return false;
1357
1358         if (!opts.promote_target)
1359                 return false;
1360
1361         if (bch2_bkey_has_target(c, k, opts.promote_target))
1362                 return false;
1363
1364         if (bch2_target_congested(c, opts.promote_target)) {
1365                 /* XXX trace this */
1366                 return false;
1367         }
1368
1369         if (rhashtable_lookup_fast(&c->promote_table, &pos,
1370                                    bch_promote_params))
1371                 return false;
1372
1373         return true;
1374 }
1375
1376 static void promote_free(struct bch_fs *c, struct promote_op *op)
1377 {
1378         int ret;
1379
1380         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1381                                      bch_promote_params);
1382         BUG_ON(ret);
1383         percpu_ref_put(&c->writes);
1384         kfree_rcu(op, rcu);
1385 }
1386
1387 static void promote_done(struct closure *cl)
1388 {
1389         struct promote_op *op =
1390                 container_of(cl, struct promote_op, cl);
1391         struct bch_fs *c = op->write.op.c;
1392
1393         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1394                                op->start_time);
1395
1396         bch2_bio_free_pages_pool(c, &op->write.op.wbio.bio);
1397         promote_free(c, op);
1398 }
1399
1400 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1401 {
1402         struct bch_fs *c = rbio->c;
1403         struct closure *cl = &op->cl;
1404         struct bio *bio = &op->write.op.wbio.bio;
1405
1406         trace_promote(&rbio->bio);
1407
1408         /* we now own pages: */
1409         BUG_ON(!rbio->bounce);
1410         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1411
1412         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1413                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1414         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1415
1416         bch2_migrate_read_done(&op->write, rbio);
1417
1418         closure_init(cl, NULL);
1419         closure_call(&op->write.op.cl, bch2_write, c->btree_update_wq, cl);
1420         closure_return_with_destructor(cl, promote_done);
1421 }
1422
1423 static struct promote_op *__promote_alloc(struct bch_fs *c,
1424                                           enum btree_id btree_id,
1425                                           struct bkey_s_c k,
1426                                           struct bpos pos,
1427                                           struct extent_ptr_decoded *pick,
1428                                           struct bch_io_opts opts,
1429                                           unsigned sectors,
1430                                           struct bch_read_bio **rbio)
1431 {
1432         struct promote_op *op = NULL;
1433         struct bio *bio;
1434         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1435         int ret;
1436
1437         if (!percpu_ref_tryget(&c->writes))
1438                 return NULL;
1439
1440         op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1441         if (!op)
1442                 goto err;
1443
1444         op->start_time = local_clock();
1445         op->pos = pos;
1446
1447         /*
1448          * We don't use the mempool here because extents that aren't
1449          * checksummed or compressed can be too big for the mempool:
1450          */
1451         *rbio = kzalloc(sizeof(struct bch_read_bio) +
1452                         sizeof(struct bio_vec) * pages,
1453                         GFP_NOIO);
1454         if (!*rbio)
1455                 goto err;
1456
1457         rbio_init(&(*rbio)->bio, opts);
1458         bio_init(&(*rbio)->bio, (*rbio)->bio.bi_inline_vecs, pages);
1459
1460         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1461                                  GFP_NOIO))
1462                 goto err;
1463
1464         (*rbio)->bounce         = true;
1465         (*rbio)->split          = true;
1466         (*rbio)->kmalloc        = true;
1467
1468         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1469                                           bch_promote_params))
1470                 goto err;
1471
1472         bio = &op->write.op.wbio.bio;
1473         bio_init(bio, bio->bi_inline_vecs, pages);
1474
1475         ret = bch2_migrate_write_init(c, &op->write,
1476                         writepoint_hashed((unsigned long) current),
1477                         opts,
1478                         DATA_PROMOTE,
1479                         (struct data_opts) {
1480                                 .target         = opts.promote_target,
1481                                 .nr_replicas    = 1,
1482                         },
1483                         btree_id, k);
1484         BUG_ON(ret);
1485
1486         return op;
1487 err:
1488         if (*rbio)
1489                 bio_free_pages(&(*rbio)->bio);
1490         kfree(*rbio);
1491         *rbio = NULL;
1492         kfree(op);
1493         percpu_ref_put(&c->writes);
1494         return NULL;
1495 }
1496
1497 noinline
1498 static struct promote_op *promote_alloc(struct bch_fs *c,
1499                                                struct bvec_iter iter,
1500                                                struct bkey_s_c k,
1501                                                struct extent_ptr_decoded *pick,
1502                                                struct bch_io_opts opts,
1503                                                unsigned flags,
1504                                                struct bch_read_bio **rbio,
1505                                                bool *bounce,
1506                                                bool *read_full)
1507 {
1508         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1509         /* data might have to be decompressed in the write path: */
1510         unsigned sectors = promote_full
1511                 ? max(pick->crc.compressed_size, pick->crc.live_size)
1512                 : bvec_iter_sectors(iter);
1513         struct bpos pos = promote_full
1514                 ? bkey_start_pos(k.k)
1515                 : POS(k.k->p.inode, iter.bi_sector);
1516         struct promote_op *promote;
1517
1518         if (!should_promote(c, k, pos, opts, flags))
1519                 return NULL;
1520
1521         promote = __promote_alloc(c,
1522                                   k.k->type == KEY_TYPE_reflink_v
1523                                   ? BTREE_ID_reflink
1524                                   : BTREE_ID_extents,
1525                                   k, pos, pick, opts, sectors, rbio);
1526         if (!promote)
1527                 return NULL;
1528
1529         *bounce         = true;
1530         *read_full      = promote_full;
1531         return promote;
1532 }
1533
1534 /* Read */
1535
1536 #define READ_RETRY_AVOID        1
1537 #define READ_RETRY              2
1538 #define READ_ERR                3
1539
1540 enum rbio_context {
1541         RBIO_CONTEXT_NULL,
1542         RBIO_CONTEXT_HIGHPRI,
1543         RBIO_CONTEXT_UNBOUND,
1544 };
1545
1546 static inline struct bch_read_bio *
1547 bch2_rbio_parent(struct bch_read_bio *rbio)
1548 {
1549         return rbio->split ? rbio->parent : rbio;
1550 }
1551
1552 __always_inline
1553 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1554                            enum rbio_context context,
1555                            struct workqueue_struct *wq)
1556 {
1557         if (context <= rbio->context) {
1558                 fn(&rbio->work);
1559         } else {
1560                 rbio->work.func         = fn;
1561                 rbio->context           = context;
1562                 queue_work(wq, &rbio->work);
1563         }
1564 }
1565
1566 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1567 {
1568         BUG_ON(rbio->bounce && !rbio->split);
1569
1570         if (rbio->promote)
1571                 promote_free(rbio->c, rbio->promote);
1572         rbio->promote = NULL;
1573
1574         if (rbio->bounce)
1575                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1576
1577         if (rbio->split) {
1578                 struct bch_read_bio *parent = rbio->parent;
1579
1580                 if (rbio->kmalloc)
1581                         kfree(rbio);
1582                 else
1583                         bio_put(&rbio->bio);
1584
1585                 rbio = parent;
1586         }
1587
1588         return rbio;
1589 }
1590
1591 /*
1592  * Only called on a top level bch_read_bio to complete an entire read request,
1593  * not a split:
1594  */
1595 static void bch2_rbio_done(struct bch_read_bio *rbio)
1596 {
1597         if (rbio->start_time)
1598                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1599                                        rbio->start_time);
1600         bio_endio(&rbio->bio);
1601 }
1602
1603 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1604                                      struct bvec_iter bvec_iter,
1605                                      struct bch_io_failures *failed,
1606                                      unsigned flags)
1607 {
1608         struct btree_trans trans;
1609         struct btree_iter iter;
1610         struct bkey_buf sk;
1611         struct bkey_s_c k;
1612         int ret;
1613
1614         flags &= ~BCH_READ_LAST_FRAGMENT;
1615         flags |= BCH_READ_MUST_CLONE;
1616
1617         bch2_bkey_buf_init(&sk);
1618         bch2_trans_init(&trans, c, 0, 0);
1619
1620         bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1621                              rbio->read_pos, BTREE_ITER_SLOTS);
1622 retry:
1623         rbio->bio.bi_status = 0;
1624
1625         k = bch2_btree_iter_peek_slot(&iter);
1626         if (bkey_err(k))
1627                 goto err;
1628
1629         bch2_bkey_buf_reassemble(&sk, c, k);
1630         k = bkey_i_to_s_c(sk.k);
1631         bch2_trans_unlock(&trans);
1632
1633         if (!bch2_bkey_matches_ptr(c, k,
1634                                    rbio->pick.ptr,
1635                                    rbio->data_pos.offset -
1636                                    rbio->pick.crc.offset)) {
1637                 /* extent we wanted to read no longer exists: */
1638                 rbio->hole = true;
1639                 goto out;
1640         }
1641
1642         ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1643                                  rbio->read_pos,
1644                                  rbio->data_btree,
1645                                  k, 0, failed, flags);
1646         if (ret == READ_RETRY)
1647                 goto retry;
1648         if (ret)
1649                 goto err;
1650 out:
1651         bch2_rbio_done(rbio);
1652         bch2_trans_iter_exit(&trans, &iter);
1653         bch2_trans_exit(&trans);
1654         bch2_bkey_buf_exit(&sk, c);
1655         return;
1656 err:
1657         rbio->bio.bi_status = BLK_STS_IOERR;
1658         goto out;
1659 }
1660
1661 static void bch2_rbio_retry(struct work_struct *work)
1662 {
1663         struct bch_read_bio *rbio =
1664                 container_of(work, struct bch_read_bio, work);
1665         struct bch_fs *c        = rbio->c;
1666         struct bvec_iter iter   = rbio->bvec_iter;
1667         unsigned flags          = rbio->flags;
1668         subvol_inum inum = {
1669                 .subvol = rbio->subvol,
1670                 .inum   = rbio->read_pos.inode,
1671         };
1672         struct bch_io_failures failed = { .nr = 0 };
1673
1674         trace_read_retry(&rbio->bio);
1675
1676         if (rbio->retry == READ_RETRY_AVOID)
1677                 bch2_mark_io_failure(&failed, &rbio->pick);
1678
1679         rbio->bio.bi_status = 0;
1680
1681         rbio = bch2_rbio_free(rbio);
1682
1683         flags |= BCH_READ_IN_RETRY;
1684         flags &= ~BCH_READ_MAY_PROMOTE;
1685
1686         if (flags & BCH_READ_NODECODE) {
1687                 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1688         } else {
1689                 flags &= ~BCH_READ_LAST_FRAGMENT;
1690                 flags |= BCH_READ_MUST_CLONE;
1691
1692                 __bch2_read(c, rbio, iter, inum, &failed, flags);
1693         }
1694 }
1695
1696 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1697                             blk_status_t error)
1698 {
1699         rbio->retry = retry;
1700
1701         if (rbio->flags & BCH_READ_IN_RETRY)
1702                 return;
1703
1704         if (retry == READ_ERR) {
1705                 rbio = bch2_rbio_free(rbio);
1706
1707                 rbio->bio.bi_status = error;
1708                 bch2_rbio_done(rbio);
1709         } else {
1710                 bch2_rbio_punt(rbio, bch2_rbio_retry,
1711                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1712         }
1713 }
1714
1715 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1716                                    struct bch_read_bio *rbio)
1717 {
1718         struct bch_fs *c = rbio->c;
1719         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1720         struct bch_extent_crc_unpacked new_crc;
1721         struct btree_iter iter;
1722         struct bkey_i *new;
1723         struct bkey_s_c k;
1724         int ret = 0;
1725
1726         if (crc_is_compressed(rbio->pick.crc))
1727                 return 0;
1728
1729         bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1730                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1731         k = bch2_btree_iter_peek_slot(&iter);
1732         if ((ret = bkey_err(k)))
1733                 goto out;
1734
1735         if (bversion_cmp(k.k->version, rbio->version) ||
1736             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1737                 goto out;
1738
1739         /* Extent was merged? */
1740         if (bkey_start_offset(k.k) < data_offset ||
1741             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1742                 goto out;
1743
1744         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1745                         rbio->pick.crc, NULL, &new_crc,
1746                         bkey_start_offset(k.k) - data_offset, k.k->size,
1747                         rbio->pick.crc.csum_type)) {
1748                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1749                 ret = 0;
1750                 goto out;
1751         }
1752
1753         /*
1754          * going to be temporarily appending another checksum entry:
1755          */
1756         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1757                                  sizeof(struct bch_extent_crc128));
1758         if ((ret = PTR_ERR_OR_ZERO(new)))
1759                 goto out;
1760
1761         bkey_reassemble(new, k);
1762
1763         if (!bch2_bkey_narrow_crcs(new, new_crc))
1764                 goto out;
1765
1766         ret = bch2_trans_update(trans, &iter, new,
1767                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1768 out:
1769         bch2_trans_iter_exit(trans, &iter);
1770         return ret;
1771 }
1772
1773 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1774 {
1775         bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1776                       __bch2_rbio_narrow_crcs(&trans, rbio));
1777 }
1778
1779 /* Inner part that may run in process context */
1780 static void __bch2_read_endio(struct work_struct *work)
1781 {
1782         struct bch_read_bio *rbio =
1783                 container_of(work, struct bch_read_bio, work);
1784         struct bch_fs *c        = rbio->c;
1785         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1786         struct bio *src         = &rbio->bio;
1787         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
1788         struct bvec_iter dst_iter = rbio->bvec_iter;
1789         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1790         struct nonce nonce = extent_nonce(rbio->version, crc);
1791         unsigned nofs_flags;
1792         struct bch_csum csum;
1793
1794         nofs_flags = memalloc_nofs_save();
1795
1796         /* Reset iterator for checksumming and copying bounced data: */
1797         if (rbio->bounce) {
1798                 src->bi_iter.bi_size            = crc.compressed_size << 9;
1799                 src->bi_iter.bi_idx             = 0;
1800                 src->bi_iter.bi_bvec_done       = 0;
1801         } else {
1802                 src->bi_iter                    = rbio->bvec_iter;
1803         }
1804
1805         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1806         if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1807                 goto csum_err;
1808
1809         /*
1810          * XXX
1811          * We need to rework the narrow_crcs path to deliver the read completion
1812          * first, and then punt to a different workqueue, otherwise we're
1813          * holding up reads while doing btree updates which is bad for memory
1814          * reclaim.
1815          */
1816         if (unlikely(rbio->narrow_crcs))
1817                 bch2_rbio_narrow_crcs(rbio);
1818
1819         if (rbio->flags & BCH_READ_NODECODE)
1820                 goto nodecode;
1821
1822         /* Adjust crc to point to subset of data we want: */
1823         crc.offset     += rbio->offset_into_extent;
1824         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
1825
1826         if (crc_is_compressed(crc)) {
1827                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1828                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1829                         goto decompression_err;
1830         } else {
1831                 /* don't need to decrypt the entire bio: */
1832                 nonce = nonce_add(nonce, crc.offset << 9);
1833                 bio_advance(src, crc.offset << 9);
1834
1835                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1836                 src->bi_iter.bi_size = dst_iter.bi_size;
1837
1838                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1839
1840                 if (rbio->bounce) {
1841                         struct bvec_iter src_iter = src->bi_iter;
1842                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1843                 }
1844         }
1845
1846         if (rbio->promote) {
1847                 /*
1848                  * Re encrypt data we decrypted, so it's consistent with
1849                  * rbio->crc:
1850                  */
1851                 bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1852                 promote_start(rbio->promote, rbio);
1853                 rbio->promote = NULL;
1854         }
1855 nodecode:
1856         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1857                 rbio = bch2_rbio_free(rbio);
1858                 bch2_rbio_done(rbio);
1859         }
1860 out:
1861         memalloc_nofs_restore(nofs_flags);
1862         return;
1863 csum_err:
1864         /*
1865          * Checksum error: if the bio wasn't bounced, we may have been
1866          * reading into buffers owned by userspace (that userspace can
1867          * scribble over) - retry the read, bouncing it this time:
1868          */
1869         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1870                 rbio->flags |= BCH_READ_MUST_BOUNCE;
1871                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1872                 goto out;
1873         }
1874
1875         bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1876                 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %u)",
1877                 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1878                 csum.hi, csum.lo, crc.csum_type);
1879         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1880         goto out;
1881 decompression_err:
1882         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1883                                  "decompression error");
1884         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1885         goto out;
1886 }
1887
1888 static void bch2_read_endio(struct bio *bio)
1889 {
1890         struct bch_read_bio *rbio =
1891                 container_of(bio, struct bch_read_bio, bio);
1892         struct bch_fs *c        = rbio->c;
1893         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1894         struct workqueue_struct *wq = NULL;
1895         enum rbio_context context = RBIO_CONTEXT_NULL;
1896
1897         if (rbio->have_ioref) {
1898                 bch2_latency_acct(ca, rbio->submit_time, READ);
1899                 percpu_ref_put(&ca->io_ref);
1900         }
1901
1902         if (!rbio->split)
1903                 rbio->bio.bi_end_io = rbio->end_io;
1904
1905         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1906                                     rbio->read_pos.inode,
1907                                     rbio->read_pos.offset,
1908                                     "data read error: %s",
1909                                bch2_blk_status_to_str(bio->bi_status))) {
1910                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1911                 return;
1912         }
1913
1914         if (rbio->pick.ptr.cached &&
1915             (((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 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
1976                        struct bvec_iter iter, struct bpos read_pos,
1977                        enum btree_id data_btree, struct bkey_s_c k,
1978                        unsigned offset_into_extent,
1979                        struct bch_io_failures *failed, unsigned flags)
1980 {
1981         struct bch_fs *c = trans->c;
1982         struct extent_ptr_decoded pick;
1983         struct bch_read_bio *rbio = NULL;
1984         struct bch_dev *ca;
1985         struct promote_op *promote = NULL;
1986         bool bounce = false, read_full = false, narrow_crcs = false;
1987         struct bpos data_pos = bkey_start_pos(k.k);
1988         int pick_ret;
1989
1990         if (bkey_extent_is_inline_data(k.k)) {
1991                 unsigned bytes = min_t(unsigned, iter.bi_size,
1992                                        bkey_inline_data_bytes(k.k));
1993
1994                 swap(iter.bi_size, bytes);
1995                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
1996                 swap(iter.bi_size, bytes);
1997                 bio_advance_iter(&orig->bio, &iter, bytes);
1998                 zero_fill_bio_iter(&orig->bio, iter);
1999                 goto out_read_done;
2000         }
2001
2002         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2003
2004         /* hole or reservation - just zero fill: */
2005         if (!pick_ret)
2006                 goto hole;
2007
2008         if (pick_ret < 0) {
2009                 bch_err_inum_ratelimited(c, k.k->p.inode,
2010                                          "no device to read from");
2011                 goto err;
2012         }
2013
2014         if (pick_ret > 0)
2015                 ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2016
2017         if (flags & BCH_READ_NODECODE) {
2018                 /*
2019                  * can happen if we retry, and the extent we were going to read
2020                  * has been merged in the meantime:
2021                  */
2022                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2023                         goto hole;
2024
2025                 iter.bi_size    = pick.crc.compressed_size << 9;
2026                 goto get_bio;
2027         }
2028
2029         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2030             bio_flagged(&orig->bio, BIO_CHAIN))
2031                 flags |= BCH_READ_MUST_CLONE;
2032
2033         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2034                 bch2_can_narrow_extent_crcs(k, pick.crc);
2035
2036         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2037                 flags |= BCH_READ_MUST_BOUNCE;
2038
2039         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2040
2041         if (crc_is_compressed(pick.crc) ||
2042             (pick.crc.csum_type != BCH_CSUM_none &&
2043              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2044               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2045                (flags & BCH_READ_USER_MAPPED)) ||
2046               (flags & BCH_READ_MUST_BOUNCE)))) {
2047                 read_full = true;
2048                 bounce = true;
2049         }
2050
2051         if (orig->opts.promote_target)
2052                 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2053                                         &rbio, &bounce, &read_full);
2054
2055         if (!read_full) {
2056                 EBUG_ON(crc_is_compressed(pick.crc));
2057                 EBUG_ON(pick.crc.csum_type &&
2058                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2059                          bvec_iter_sectors(iter) != pick.crc.live_size ||
2060                          pick.crc.offset ||
2061                          offset_into_extent));
2062
2063                 data_pos.offset += offset_into_extent;
2064                 pick.ptr.offset += pick.crc.offset +
2065                         offset_into_extent;
2066                 offset_into_extent              = 0;
2067                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
2068                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
2069                 pick.crc.offset                 = 0;
2070                 pick.crc.live_size              = bvec_iter_sectors(iter);
2071                 offset_into_extent              = 0;
2072         }
2073 get_bio:
2074         if (rbio) {
2075                 /*
2076                  * promote already allocated bounce rbio:
2077                  * promote needs to allocate a bio big enough for uncompressing
2078                  * data in the write path, but we're not going to use it all
2079                  * here:
2080                  */
2081                 EBUG_ON(rbio->bio.bi_iter.bi_size <
2082                        pick.crc.compressed_size << 9);
2083                 rbio->bio.bi_iter.bi_size =
2084                         pick.crc.compressed_size << 9;
2085         } else if (bounce) {
2086                 unsigned sectors = pick.crc.compressed_size;
2087
2088                 rbio = rbio_init(bio_alloc_bioset(GFP_NOIO,
2089                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
2090                                                   &c->bio_read_split),
2091                                  orig->opts);
2092
2093                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2094                 rbio->bounce    = true;
2095                 rbio->split     = true;
2096         } else if (flags & BCH_READ_MUST_CLONE) {
2097                 /*
2098                  * Have to clone if there were any splits, due to error
2099                  * reporting issues (if a split errored, and retrying didn't
2100                  * work, when it reports the error to its parent (us) we don't
2101                  * know if the error was from our bio, and we should retry, or
2102                  * from the whole bio, in which case we don't want to retry and
2103                  * lose the error)
2104                  */
2105                 rbio = rbio_init(bio_clone_fast(&orig->bio, GFP_NOIO,
2106                                                 &c->bio_read_split),
2107                                  orig->opts);
2108                 rbio->bio.bi_iter = iter;
2109                 rbio->split     = true;
2110         } else {
2111                 rbio = orig;
2112                 rbio->bio.bi_iter = iter;
2113                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2114         }
2115
2116         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2117
2118         rbio->c                 = c;
2119         rbio->submit_time       = local_clock();
2120         if (rbio->split)
2121                 rbio->parent    = orig;
2122         else
2123                 rbio->end_io    = orig->bio.bi_end_io;
2124         rbio->bvec_iter         = iter;
2125         rbio->offset_into_extent= offset_into_extent;
2126         rbio->flags             = flags;
2127         rbio->have_ioref        = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2128         rbio->narrow_crcs       = narrow_crcs;
2129         rbio->hole              = 0;
2130         rbio->retry             = 0;
2131         rbio->context           = 0;
2132         /* XXX: only initialize this if needed */
2133         rbio->devs_have         = bch2_bkey_devs(k);
2134         rbio->pick              = pick;
2135         rbio->subvol            = orig->subvol;
2136         rbio->read_pos          = read_pos;
2137         rbio->data_btree        = data_btree;
2138         rbio->data_pos          = data_pos;
2139         rbio->version           = k.k->version;
2140         rbio->promote           = promote;
2141         INIT_WORK(&rbio->work, NULL);
2142
2143         rbio->bio.bi_opf        = orig->bio.bi_opf;
2144         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2145         rbio->bio.bi_end_io     = bch2_read_endio;
2146
2147         if (rbio->bounce)
2148                 trace_read_bounce(&rbio->bio);
2149
2150         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2151
2152         /*
2153          * If it's being moved internally, we don't want to flag it as a cache
2154          * hit:
2155          */
2156         if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2157                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2158                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
2159
2160         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2161                 bio_inc_remaining(&orig->bio);
2162                 trace_read_split(&orig->bio);
2163         }
2164
2165         if (!rbio->pick.idx) {
2166                 if (!rbio->have_ioref) {
2167                         bch_err_inum_ratelimited(c, k.k->p.inode,
2168                                                  "no device to read from");
2169                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2170                         goto out;
2171                 }
2172
2173                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2174                              bio_sectors(&rbio->bio));
2175                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2176
2177                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2178                         submit_bio(&rbio->bio);
2179                 else
2180                         submit_bio_wait(&rbio->bio);
2181         } else {
2182                 /* Attempting reconstruct read: */
2183                 if (bch2_ec_read_extent(c, rbio)) {
2184                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2185                         goto out;
2186                 }
2187
2188                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2189                         bio_endio(&rbio->bio);
2190         }
2191 out:
2192         if (likely(!(flags & BCH_READ_IN_RETRY))) {
2193                 return 0;
2194         } else {
2195                 int ret;
2196
2197                 rbio->context = RBIO_CONTEXT_UNBOUND;
2198                 bch2_read_endio(&rbio->bio);
2199
2200                 ret = rbio->retry;
2201                 rbio = bch2_rbio_free(rbio);
2202
2203                 if (ret == READ_RETRY_AVOID) {
2204                         bch2_mark_io_failure(failed, &pick);
2205                         ret = READ_RETRY;
2206                 }
2207
2208                 if (!ret)
2209                         goto out_read_done;
2210
2211                 return ret;
2212         }
2213
2214 err:
2215         if (flags & BCH_READ_IN_RETRY)
2216                 return READ_ERR;
2217
2218         orig->bio.bi_status = BLK_STS_IOERR;
2219         goto out_read_done;
2220
2221 hole:
2222         /*
2223          * won't normally happen in the BCH_READ_NODECODE
2224          * (bch2_move_extent()) path, but if we retry and the extent we wanted
2225          * to read no longer exists we have to signal that:
2226          */
2227         if (flags & BCH_READ_NODECODE)
2228                 orig->hole = true;
2229
2230         zero_fill_bio_iter(&orig->bio, iter);
2231 out_read_done:
2232         if (flags & BCH_READ_LAST_FRAGMENT)
2233                 bch2_rbio_done(orig);
2234         return 0;
2235 }
2236
2237 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2238                  struct bvec_iter bvec_iter, subvol_inum inum,
2239                  struct bch_io_failures *failed, unsigned flags)
2240 {
2241         struct btree_trans trans;
2242         struct btree_iter iter;
2243         struct bkey_buf sk;
2244         struct bkey_s_c k;
2245         u32 snapshot;
2246         int ret;
2247
2248         BUG_ON(flags & BCH_READ_NODECODE);
2249
2250         bch2_bkey_buf_init(&sk);
2251         bch2_trans_init(&trans, c, 0, 0);
2252 retry:
2253         bch2_trans_begin(&trans);
2254         iter = (struct btree_iter) { NULL };
2255
2256         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2257         if (ret)
2258                 goto err;
2259
2260         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2261                              SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2262                              BTREE_ITER_SLOTS|BTREE_ITER_FILTER_SNAPSHOTS);
2263         while (1) {
2264                 unsigned bytes, sectors, offset_into_extent;
2265                 enum btree_id data_btree = BTREE_ID_extents;
2266
2267                 /*
2268                  * read_extent -> io_time_reset may cause a transaction restart
2269                  * without returning an error, we need to check for that here:
2270                  */
2271                 if (!bch2_trans_relock(&trans)) {
2272                         ret = -EINTR;
2273                         break;
2274                 }
2275
2276                 bch2_btree_iter_set_pos(&iter,
2277                                 POS(inum.inum, bvec_iter.bi_sector));
2278
2279                 k = bch2_btree_iter_peek_slot(&iter);
2280                 ret = bkey_err(k);
2281                 if (ret)
2282                         break;
2283
2284                 offset_into_extent = iter.pos.offset -
2285                         bkey_start_offset(k.k);
2286                 sectors = k.k->size - offset_into_extent;
2287
2288                 bch2_bkey_buf_reassemble(&sk, c, k);
2289
2290                 ret = bch2_read_indirect_extent(&trans, &data_btree,
2291                                         &offset_into_extent, &sk);
2292                 if (ret)
2293                         break;
2294
2295                 k = bkey_i_to_s_c(sk.k);
2296
2297                 /*
2298                  * With indirect extents, the amount of data to read is the min
2299                  * of the original extent and the indirect extent:
2300                  */
2301                 sectors = min(sectors, k.k->size - offset_into_extent);
2302
2303                 /*
2304                  * Unlock the iterator while the btree node's lock is still in
2305                  * cache, before doing the IO:
2306                  */
2307                 bch2_trans_unlock(&trans);
2308
2309                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2310                 swap(bvec_iter.bi_size, bytes);
2311
2312                 if (bvec_iter.bi_size == bytes)
2313                         flags |= BCH_READ_LAST_FRAGMENT;
2314
2315                 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2316                                          data_btree, k,
2317                                          offset_into_extent, failed, flags);
2318                 if (ret)
2319                         break;
2320
2321                 if (flags & BCH_READ_LAST_FRAGMENT)
2322                         break;
2323
2324                 swap(bvec_iter.bi_size, bytes);
2325                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2326
2327                 ret = btree_trans_too_many_iters(&trans);
2328                 if (ret)
2329                         break;
2330         }
2331 err:
2332         bch2_trans_iter_exit(&trans, &iter);
2333
2334         if (ret == -EINTR || ret == READ_RETRY || ret == READ_RETRY_AVOID)
2335                 goto retry;
2336
2337         bch2_trans_exit(&trans);
2338         bch2_bkey_buf_exit(&sk, c);
2339
2340         if (ret) {
2341                 bch_err_inum_ratelimited(c, inum.inum,
2342                                          "read error %i from btree lookup", ret);
2343                 rbio->bio.bi_status = BLK_STS_IOERR;
2344                 bch2_rbio_done(rbio);
2345         }
2346 }
2347
2348 void bch2_fs_io_exit(struct bch_fs *c)
2349 {
2350         if (c->promote_table.tbl)
2351                 rhashtable_destroy(&c->promote_table);
2352         mempool_exit(&c->bio_bounce_pages);
2353         bioset_exit(&c->bio_write);
2354         bioset_exit(&c->bio_read_split);
2355         bioset_exit(&c->bio_read);
2356 }
2357
2358 int bch2_fs_io_init(struct bch_fs *c)
2359 {
2360         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2361                         BIOSET_NEED_BVECS) ||
2362             bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2363                         BIOSET_NEED_BVECS) ||
2364             bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2365                         BIOSET_NEED_BVECS) ||
2366             mempool_init_page_pool(&c->bio_bounce_pages,
2367                                    max_t(unsigned,
2368                                          c->opts.btree_node_size,
2369                                          c->opts.encoded_extent_max) /
2370                                    PAGE_SIZE, 0) ||
2371             rhashtable_init(&c->promote_table, &bch_promote_params))
2372                 return -ENOMEM;
2373
2374         return 0;
2375 }