fs/bcachefs/io_write.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2010, 2011 Kent Overstreet <[email protected]>
   4  * Copyright 2012 Google, Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_foreground.h"
   9 #include "bkey_buf.h"
  10 #include "bset.h"
  11 #include "btree_update.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "clock.h"
  15 #include "compress.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "extent_update.h"
  20 #include "inode.h"
  21 #include "io_write.h"
  22 #include "journal.h"
  23 #include "keylist.h"
  24 #include "move.h"
  25 #include "nocow_locking.h"
  26 #include "rebalance.h"
  27 #include "subvolume.h"
  28 #include "super.h"
  29 #include "super-io.h"
  30 #include "trace.h"
  31
  32 #include <linux/blkdev.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/random.h>
  35 #include <linux/sched/mm.h>
  36
  37 #ifndef CONFIG_BCACHEFS_NO_LATENCY_ACCT
  38
  39 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
  40                                        u64 now, int rw)
  41 {
  42         u64 latency_capable =
  43                 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
  44         /* ideally we'd be taking into account the device's variance here: */
  45         u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
  46         s64 latency_over = io_latency - latency_threshold;
  47
  48         if (latency_threshold && latency_over > 0) {
  49                 /*
  50                  * bump up congested by approximately latency_over * 4 /
  51                  * latency_threshold - we don't need much accuracy here so don't
  52                  * bother with the divide:
  53                  */
  54                 if (atomic_read(&ca->congested) < CONGESTED_MAX)
  55                         atomic_add(latency_over >>
  56                                    max_t(int, ilog2(latency_threshold) - 2, 0),
  57                                    &ca->congested);
  58
  59                 ca->congested_last = now;
  60         } else if (atomic_read(&ca->congested) > 0) {
  61                 atomic_dec(&ca->congested);
  62         }
  63 }
  64
  65 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
  66 {
  67         atomic64_t *latency = &ca->cur_latency[rw];
  68         u64 now = local_clock();
  69         u64 io_latency = time_after64(now, submit_time)
  70                 ? now - submit_time
  71                 : 0;
  72         u64 old, new, v = atomic64_read(latency);
  73
  74         do {
  75                 old = v;
  76
  77                 /*
  78                  * If the io latency was reasonably close to the current
  79                  * latency, skip doing the update and atomic operation - most of
  80                  * the time:
  81                  */
  82                 if (abs((int) (old - io_latency)) < (old >> 1) &&
  83                     now & ~(~0U << 5))
  84                         break;
  85
  86                 new = ewma_add(old, io_latency, 5);
  87         } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
  88
  89         bch2_congested_acct(ca, io_latency, now, rw);
  90
  91         __bch2_time_stats_update(&ca->io_latency[rw].stats, submit_time, now);
  92 }
  93
  94 #endif
  95
  96 /* Allocate, free from mempool: */
  97
  98 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
  99 {
 100         struct bvec_iter_all iter;
 101         struct bio_vec *bv;
 102
 103         bio_for_each_segment_all(bv, bio, iter)
 104                 if (bv->bv_page != ZERO_PAGE(0))
 105                         mempool_free(bv->bv_page, &c->bio_bounce_pages);
 106         bio->bi_vcnt = 0;
 107 }
 108
 109 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
 110 {
 111         struct page *page;
 112
 113         if (likely(!*using_mempool)) {
 114                 page = alloc_page(GFP_NOFS);
 115                 if (unlikely(!page)) {
 116                         mutex_lock(&c->bio_bounce_pages_lock);
 117                         *using_mempool = true;
 118                         goto pool_alloc;
 119
 120                 }
 121         } else {
 122 pool_alloc:
 123                 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOFS);
 124         }
 125
 126         return page;
 127 }
 128
 129 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
 130                                size_t size)
 131 {
 132         bool using_mempool = false;
 133
 134         while (size) {
 135                 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
 136                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 137
 138                 BUG_ON(!bio_add_page(bio, page, len, 0));
 139                 size -= len;
 140         }
 141
 142         if (using_mempool)
 143                 mutex_unlock(&c->bio_bounce_pages_lock);
 144 }
 145
 146 /* Extent update path: */
 147
 148 int bch2_sum_sector_overwrites(struct btree_trans *trans,
 149                                struct btree_iter *extent_iter,
 150                                struct bkey_i *new,
 151                                bool *usage_increasing,
 152                                s64 *i_sectors_delta,
 153                                s64 *disk_sectors_delta)
 154 {
 155         struct bch_fs *c = trans->c;
 156         struct btree_iter iter;
 157         struct bkey_s_c old;
 158         unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
 159         bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
 160         int ret = 0;
 161
 162         *usage_increasing       = false;
 163         *i_sectors_delta        = 0;
 164         *disk_sectors_delta     = 0;
 165
 166         bch2_trans_copy_iter(&iter, extent_iter);
 167
 168         for_each_btree_key_upto_continue_norestart(iter,
 169                                 new->k.p, BTREE_ITER_slots, old, ret) {
 170                 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 171                         max(bkey_start_offset(&new->k),
 172                             bkey_start_offset(old.k));
 173
 174                 *i_sectors_delta += sectors *
 175                         (bkey_extent_is_allocation(&new->k) -
 176                          bkey_extent_is_allocation(old.k));
 177
 178                 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
 179                 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
 180                         ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 181                         : 0;
 182
 183                 if (!*usage_increasing &&
 184                     (new->k.p.snapshot != old.k->p.snapshot ||
 185                      new_replicas > bch2_bkey_replicas(c, old) ||
 186                      (!new_compressed && bch2_bkey_sectors_compressed(old))))
 187                         *usage_increasing = true;
 188
 189                 if (bkey_ge(old.k->p, new->k.p))
 190                         break;
 191         }
 192
 193         bch2_trans_iter_exit(trans, &iter);
 194         return ret;
 195 }
 196
 197 static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
 198                                                     struct btree_iter *extent_iter,
 199                                                     u64 new_i_size,
 200                                                     s64 i_sectors_delta)
 201 {
 202         /*
 203          * Crazy performance optimization:
 204          * Every extent update needs to also update the inode: the inode trigger
 205          * will set bi->journal_seq to the journal sequence number of this
 206          * transaction - for fsync.
 207          *
 208          * But if that's the only reason we're updating the inode (we're not
 209          * updating bi_size or bi_sectors), then we don't need the inode update
 210          * to be journalled - if we crash, the bi_journal_seq update will be
 211          * lost, but that's fine.
 212          */
 213         unsigned inode_update_flags = BTREE_UPDATE_nojournal;
 214
 215         struct btree_iter iter;
 216         struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 217                               SPOS(0,
 218                                    extent_iter->pos.inode,
 219                                    extent_iter->snapshot),
 220                               BTREE_ITER_cached);
 221         int ret = bkey_err(k);
 222         if (unlikely(ret))
 223                 return ret;
 224
 225         /*
 226          * varint_decode_fast(), in the inode .invalid method, reads up to 7
 227          * bytes past the end of the buffer:
 228          */
 229         struct bkey_i *k_mut = bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + 8);
 230         ret = PTR_ERR_OR_ZERO(k_mut);
 231         if (unlikely(ret))
 232                 goto err;
 233
 234         bkey_reassemble(k_mut, k);
 235
 236         if (unlikely(k_mut->k.type != KEY_TYPE_inode_v3)) {
 237                 k_mut = bch2_inode_to_v3(trans, k_mut);
 238                 ret = PTR_ERR_OR_ZERO(k_mut);
 239                 if (unlikely(ret))
 240                         goto err;
 241         }
 242
 243         struct bkey_i_inode_v3 *inode = bkey_i_to_inode_v3(k_mut);
 244
 245         if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_i_size_dirty) &&
 246             new_i_size > le64_to_cpu(inode->v.bi_size)) {
 247                 inode->v.bi_size = cpu_to_le64(new_i_size);
 248                 inode_update_flags = 0;
 249         }
 250
 251         if (i_sectors_delta) {
 252                 le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
 253                 inode_update_flags = 0;
 254         }
 255
 256         if (inode->k.p.snapshot != iter.snapshot) {
 257                 inode->k.p.snapshot = iter.snapshot;
 258                 inode_update_flags = 0;
 259         }
 260
 261         ret = bch2_trans_update(trans, &iter, &inode->k_i,
 262                                 BTREE_UPDATE_internal_snapshot_node|
 263                                 inode_update_flags);
 264 err:
 265         bch2_trans_iter_exit(trans, &iter);
 266         return ret;
 267 }
 268
 269 int bch2_extent_update(struct btree_trans *trans,
 270                        subvol_inum inum,
 271                        struct btree_iter *iter,
 272                        struct bkey_i *k,
 273                        struct disk_reservation *disk_res,
 274                        u64 new_i_size,
 275                        s64 *i_sectors_delta_total,
 276                        bool check_enospc)
 277 {
 278         struct bpos next_pos;
 279         bool usage_increasing;
 280         s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 281         int ret;
 282
 283         /*
 284          * This traverses us the iterator without changing iter->path->pos to
 285          * search_key() (which is pos + 1 for extents): we want there to be a
 286          * path already traversed at iter->pos because
 287          * bch2_trans_extent_update() will use it to attempt extent merging
 288          */
 289         ret = __bch2_btree_iter_traverse(iter);
 290         if (ret)
 291                 return ret;
 292
 293         ret = bch2_extent_trim_atomic(trans, iter, k);
 294         if (ret)
 295                 return ret;
 296
 297         next_pos = k->k.p;
 298
 299         ret = bch2_sum_sector_overwrites(trans, iter, k,
 300                         &usage_increasing,
 301                         &i_sectors_delta,
 302                         &disk_sectors_delta);
 303         if (ret)
 304                 return ret;
 305
 306         if (disk_res &&
 307             disk_sectors_delta > (s64) disk_res->sectors) {
 308                 ret = bch2_disk_reservation_add(trans->c, disk_res,
 309                                         disk_sectors_delta - disk_res->sectors,
 310                                         !check_enospc || !usage_increasing
 311                                         ? BCH_DISK_RESERVATION_NOFAIL : 0);
 312                 if (ret)
 313                         return ret;
 314         }
 315
 316         /*
 317          * Note:
 318          * We always have to do an inode update - even when i_size/i_sectors
 319          * aren't changing - for fsync to work properly; fsync relies on
 320          * inode->bi_journal_seq which is updated by the trigger code:
 321          */
 322         ret =   bch2_extent_update_i_size_sectors(trans, iter,
 323                                                   min(k->k.p.offset << 9, new_i_size),
 324                                                   i_sectors_delta) ?:
 325                 bch2_trans_update(trans, iter, k, 0) ?:
 326                 bch2_trans_commit(trans, disk_res, NULL,
 327                                 BCH_TRANS_COMMIT_no_check_rw|
 328                                 BCH_TRANS_COMMIT_no_enospc);
 329         if (unlikely(ret))
 330                 return ret;
 331
 332         if (i_sectors_delta_total)
 333                 *i_sectors_delta_total += i_sectors_delta;
 334         bch2_btree_iter_set_pos(iter, next_pos);
 335         return 0;
 336 }
 337
 338 static int bch2_write_index_default(struct bch_write_op *op)
 339 {
 340         struct bch_fs *c = op->c;
 341         struct bkey_buf sk;
 342         struct keylist *keys = &op->insert_keys;
 343         struct bkey_i *k = bch2_keylist_front(keys);
 344         struct btree_trans *trans = bch2_trans_get(c);
 345         struct btree_iter iter;
 346         subvol_inum inum = {
 347                 .subvol = op->subvol,
 348                 .inum   = k->k.p.inode,
 349         };
 350         int ret;
 351
 352         BUG_ON(!inum.subvol);
 353
 354         bch2_bkey_buf_init(&sk);
 355
 356         do {
 357                 bch2_trans_begin(trans);
 358
 359                 k = bch2_keylist_front(keys);
 360                 bch2_bkey_buf_copy(&sk, c, k);
 361
 362                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol,
 363                                                   &sk.k->k.p.snapshot);
 364                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 365                         continue;
 366                 if (ret)
 367                         break;
 368
 369                 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 370                                      bkey_start_pos(&sk.k->k),
 371                                      BTREE_ITER_slots|BTREE_ITER_intent);
 372
 373                 ret =   bch2_bkey_set_needs_rebalance(c, sk.k, &op->opts) ?:
 374                         bch2_extent_update(trans, inum, &iter, sk.k,
 375                                         &op->res,
 376                                         op->new_i_size, &op->i_sectors_delta,
 377                                         op->flags & BCH_WRITE_CHECK_ENOSPC);
 378                 bch2_trans_iter_exit(trans, &iter);
 379
 380                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 381                         continue;
 382                 if (ret)
 383                         break;
 384
 385                 if (bkey_ge(iter.pos, k->k.p))
 386                         bch2_keylist_pop_front(&op->insert_keys);
 387                 else
 388                         bch2_cut_front(iter.pos, k);
 389         } while (!bch2_keylist_empty(keys));
 390
 391         bch2_trans_put(trans);
 392         bch2_bkey_buf_exit(&sk, c);
 393
 394         return ret;
 395 }
 396
 397 /* Writes */
 398
 399 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 400                                enum bch_data_type type,
 401                                const struct bkey_i *k,
 402                                bool nocow)
 403 {
 404         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 405         struct bch_write_bio *n;
 406
 407         BUG_ON(c->opts.nochanges);
 408
 409         bkey_for_each_ptr(ptrs, ptr) {
 410                 struct bch_dev *ca = nocow
 411                         ? bch2_dev_have_ref(c, ptr->dev)
 412                         : bch2_dev_get_ioref(c, ptr->dev, type == BCH_DATA_btree ? READ : WRITE);
 413
 414                 if (to_entry(ptr + 1) < ptrs.end) {
 415                         n = to_wbio(bio_alloc_clone(NULL, &wbio->bio, GFP_NOFS, &c->replica_set));
 416
 417                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 418                         n->bio.bi_private       = wbio->bio.bi_private;
 419                         n->parent               = wbio;
 420                         n->split                = true;
 421                         n->bounce               = false;
 422                         n->put_bio              = true;
 423                         n->bio.bi_opf           = wbio->bio.bi_opf;
 424                         bio_inc_remaining(&wbio->bio);
 425                 } else {
 426                         n = wbio;
 427                         n->split                = false;
 428                 }
 429
 430                 n->c                    = c;
 431                 n->dev                  = ptr->dev;
 432                 n->have_ioref           = ca != NULL;
 433                 n->nocow                = nocow;
 434                 n->submit_time          = local_clock();
 435                 n->inode_offset         = bkey_start_offset(&k->k);
 436                 if (nocow)
 437                         n->nocow_bucket = PTR_BUCKET_NR(ca, ptr);
 438                 n->bio.bi_iter.bi_sector = ptr->offset;
 439
 440                 if (likely(n->have_ioref)) {
 441                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 442                                      bio_sectors(&n->bio));
 443
 444                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 445
 446                         if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
 447                                 bio_endio(&n->bio);
 448                                 continue;
 449                         }
 450
 451                         submit_bio(&n->bio);
 452                 } else {
 453                         n->bio.bi_status        = BLK_STS_REMOVED;
 454                         bio_endio(&n->bio);
 455                 }
 456         }
 457 }
 458
 459 static void __bch2_write(struct bch_write_op *);
 460
 461 static void bch2_write_done(struct closure *cl)
 462 {
 463         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 464         struct bch_fs *c = op->c;
 465
 466         EBUG_ON(op->open_buckets.nr);
 467
 468         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 469         bch2_disk_reservation_put(c, &op->res);
 470
 471         if (!(op->flags & BCH_WRITE_MOVE))
 472                 bch2_write_ref_put(c, BCH_WRITE_REF_write);
 473         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 474
 475         EBUG_ON(cl->parent);
 476         closure_debug_destroy(cl);
 477         if (op->end_io)
 478                 op->end_io(op);
 479 }
 480
 481 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
 482 {
 483         struct keylist *keys = &op->insert_keys;
 484         struct bkey_i *src, *dst = keys->keys, *n;
 485
 486         for (src = keys->keys; src != keys->top; src = n) {
 487                 n = bkey_next(src);
 488
 489                 if (bkey_extent_is_direct_data(&src->k)) {
 490                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 491                                             test_bit(ptr->dev, op->failed.d));
 492
 493                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
 494                                 return -EIO;
 495                 }
 496
 497                 if (dst != src)
 498                         memmove_u64s_down(dst, src, src->k.u64s);
 499                 dst = bkey_next(dst);
 500         }
 501
 502         keys->top = dst;
 503         return 0;
 504 }
 505
 506 /**
 507  * __bch2_write_index - after a write, update index to point to new data
 508  * @op:         bch_write_op to process
 509  */
 510 static void __bch2_write_index(struct bch_write_op *op)
 511 {
 512         struct bch_fs *c = op->c;
 513         struct keylist *keys = &op->insert_keys;
 514         unsigned dev;
 515         int ret = 0;
 516
 517         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
 518                 ret = bch2_write_drop_io_error_ptrs(op);
 519                 if (ret)
 520                         goto err;
 521         }
 522
 523         if (!bch2_keylist_empty(keys)) {
 524                 u64 sectors_start = keylist_sectors(keys);
 525
 526                 ret = !(op->flags & BCH_WRITE_MOVE)
 527                         ? bch2_write_index_default(op)
 528                         : bch2_data_update_index_update(op);
 529
 530                 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 531                 BUG_ON(keylist_sectors(keys) && !ret);
 532
 533                 op->written += sectors_start - keylist_sectors(keys);
 534
 535                 if (ret && !bch2_err_matches(ret, EROFS)) {
 536                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
 537
 538                         bch_err_inum_offset_ratelimited(c,
 539                                 insert->k.p.inode, insert->k.p.offset << 9,
 540                                 "%s write error while doing btree update: %s",
 541                                 op->flags & BCH_WRITE_MOVE ? "move" : "user",
 542                                 bch2_err_str(ret));
 543                 }
 544
 545                 if (ret)
 546                         goto err;
 547         }
 548 out:
 549         /* If some a bucket wasn't written, we can't erasure code it: */
 550         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 551                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 552
 553         bch2_open_buckets_put(c, &op->open_buckets);
 554         return;
 555 err:
 556         keys->top = keys->keys;
 557         op->error = ret;
 558         op->flags |= BCH_WRITE_DONE;
 559         goto out;
 560 }
 561
 562 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
 563 {
 564         if (state != wp->state) {
 565                 u64 now = ktime_get_ns();
 566
 567                 if (wp->last_state_change &&
 568                     time_after64(now, wp->last_state_change))
 569                         wp->time[wp->state] += now - wp->last_state_change;
 570                 wp->state = state;
 571                 wp->last_state_change = now;
 572         }
 573 }
 574
 575 static inline void wp_update_state(struct write_point *wp, bool running)
 576 {
 577         enum write_point_state state;
 578
 579         state = running                  ? WRITE_POINT_running :
 580                 !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
 581                                          : WRITE_POINT_stopped;
 582
 583         __wp_update_state(wp, state);
 584 }
 585
 586 static CLOSURE_CALLBACK(bch2_write_index)
 587 {
 588         closure_type(op, struct bch_write_op, cl);
 589         struct write_point *wp = op->wp;
 590         struct workqueue_struct *wq = index_update_wq(op);
 591         unsigned long flags;
 592
 593         if ((op->flags & BCH_WRITE_DONE) &&
 594             (op->flags & BCH_WRITE_MOVE))
 595                 bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
 596
 597         spin_lock_irqsave(&wp->writes_lock, flags);
 598         if (wp->state == WRITE_POINT_waiting_io)
 599                 __wp_update_state(wp, WRITE_POINT_waiting_work);
 600         list_add_tail(&op->wp_list, &wp->writes);
 601         spin_unlock_irqrestore (&wp->writes_lock, flags);
 602
 603         queue_work(wq, &wp->index_update_work);
 604 }
 605
 606 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
 607 {
 608         op->wp = wp;
 609
 610         if (wp->state == WRITE_POINT_stopped) {
 611                 spin_lock_irq(&wp->writes_lock);
 612                 __wp_update_state(wp, WRITE_POINT_waiting_io);
 613                 spin_unlock_irq(&wp->writes_lock);
 614         }
 615 }
 616
 617 void bch2_write_point_do_index_updates(struct work_struct *work)
 618 {
 619         struct write_point *wp =
 620                 container_of(work, struct write_point, index_update_work);
 621         struct bch_write_op *op;
 622
 623         while (1) {
 624                 spin_lock_irq(&wp->writes_lock);
 625                 op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
 626                 if (op)
 627                         list_del(&op->wp_list);
 628                 wp_update_state(wp, op != NULL);
 629                 spin_unlock_irq(&wp->writes_lock);
 630
 631                 if (!op)
 632                         break;
 633
 634                 op->flags |= BCH_WRITE_IN_WORKER;
 635
 636                 __bch2_write_index(op);
 637
 638                 if (!(op->flags & BCH_WRITE_DONE))
 639                         __bch2_write(op);
 640                 else
 641                         bch2_write_done(&op->cl);
 642         }
 643 }
 644
 645 static void bch2_write_endio(struct bio *bio)
 646 {
 647         struct closure *cl              = bio->bi_private;
 648         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 649         struct bch_write_bio *wbio      = to_wbio(bio);
 650         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 651         struct bch_fs *c                = wbio->c;
 652         struct bch_dev *ca              = wbio->have_ioref
 653                 ? bch2_dev_have_ref(c, wbio->dev)
 654                 : NULL;
 655
 656         if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
 657                                     op->pos.inode,
 658                                     wbio->inode_offset << 9,
 659                                     "data write error: %s",
 660                                     bch2_blk_status_to_str(bio->bi_status))) {
 661                 set_bit(wbio->dev, op->failed.d);
 662                 op->flags |= BCH_WRITE_IO_ERROR;
 663         }
 664
 665         if (wbio->nocow) {
 666                 bch2_bucket_nocow_unlock(&c->nocow_locks,
 667                                          POS(ca->dev_idx, wbio->nocow_bucket),
 668                                          BUCKET_NOCOW_LOCK_UPDATE);
 669                 set_bit(wbio->dev, op->devs_need_flush->d);
 670         }
 671
 672         if (wbio->have_ioref) {
 673                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 674                 percpu_ref_put(&ca->io_ref);
 675         }
 676
 677         if (wbio->bounce)
 678                 bch2_bio_free_pages_pool(c, bio);
 679
 680         if (wbio->put_bio)
 681                 bio_put(bio);
 682
 683         if (parent)
 684                 bio_endio(&parent->bio);
 685         else
 686                 closure_put(cl);
 687 }
 688
 689 static void init_append_extent(struct bch_write_op *op,
 690                                struct write_point *wp,
 691                                struct bversion version,
 692                                struct bch_extent_crc_unpacked crc)
 693 {
 694         struct bkey_i_extent *e;
 695
 696         op->pos.offset += crc.uncompressed_size;
 697
 698         e = bkey_extent_init(op->insert_keys.top);
 699         e->k.p          = op->pos;
 700         e->k.size       = crc.uncompressed_size;
 701         e->k.version    = version;
 702
 703         if (crc.csum_type ||
 704             crc.compression_type ||
 705             crc.nonce)
 706                 bch2_extent_crc_append(&e->k_i, crc);
 707
 708         bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
 709                                        op->flags & BCH_WRITE_CACHED);
 710
 711         bch2_keylist_push(&op->insert_keys);
 712 }
 713
 714 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 715                                         struct write_point *wp,
 716                                         struct bio *src,
 717                                         bool *page_alloc_failed,
 718                                         void *buf)
 719 {
 720         struct bch_write_bio *wbio;
 721         struct bio *bio;
 722         unsigned output_available =
 723                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 724         unsigned pages = DIV_ROUND_UP(output_available +
 725                                       (buf
 726                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 727                                        : 0), PAGE_SIZE);
 728
 729         pages = min(pages, BIO_MAX_VECS);
 730
 731         bio = bio_alloc_bioset(NULL, pages, 0,
 732                                GFP_NOFS, &c->bio_write);
 733         wbio                    = wbio_init(bio);
 734         wbio->put_bio           = true;
 735         /* copy WRITE_SYNC flag */
 736         wbio->bio.bi_opf        = src->bi_opf;
 737
 738         if (buf) {
 739                 bch2_bio_map(bio, buf, output_available);
 740                 return bio;
 741         }
 742
 743         wbio->bounce            = true;
 744
 745         /*
 746          * We can't use mempool for more than c->sb.encoded_extent_max
 747          * worth of pages, but we'd like to allocate more if we can:
 748          */
 749         bch2_bio_alloc_pages_pool(c, bio,
 750                                   min_t(unsigned, output_available,
 751                                         c->opts.encoded_extent_max));
 752
 753         if (bio->bi_iter.bi_size < output_available)
 754                 *page_alloc_failed =
 755                         bch2_bio_alloc_pages(bio,
 756                                              output_available -
 757                                              bio->bi_iter.bi_size,
 758                                              GFP_NOFS) != 0;
 759
 760         return bio;
 761 }
 762
 763 static int bch2_write_rechecksum(struct bch_fs *c,
 764                                  struct bch_write_op *op,
 765                                  unsigned new_csum_type)
 766 {
 767         struct bio *bio = &op->wbio.bio;
 768         struct bch_extent_crc_unpacked new_crc;
 769         int ret;
 770
 771         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 772
 773         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 774             bch2_csum_type_is_encryption(new_csum_type))
 775                 new_csum_type = op->crc.csum_type;
 776
 777         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 778                                   NULL, &new_crc,
 779                                   op->crc.offset, op->crc.live_size,
 780                                   new_csum_type);
 781         if (ret)
 782                 return ret;
 783
 784         bio_advance(bio, op->crc.offset << 9);
 785         bio->bi_iter.bi_size = op->crc.live_size << 9;
 786         op->crc = new_crc;
 787         return 0;
 788 }
 789
 790 static int bch2_write_decrypt(struct bch_write_op *op)
 791 {
 792         struct bch_fs *c = op->c;
 793         struct nonce nonce = extent_nonce(op->version, op->crc);
 794         struct bch_csum csum;
 795         int ret;
 796
 797         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 798                 return 0;
 799
 800         /*
 801          * If we need to decrypt data in the write path, we'll no longer be able
 802          * to verify the existing checksum (poly1305 mac, in this case) after
 803          * it's decrypted - this is the last point we'll be able to reverify the
 804          * checksum:
 805          */
 806         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 807         if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 808                 return -EIO;
 809
 810         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 811         op->crc.csum_type = 0;
 812         op->crc.csum = (struct bch_csum) { 0, 0 };
 813         return ret;
 814 }
 815
 816 static enum prep_encoded_ret {
 817         PREP_ENCODED_OK,
 818         PREP_ENCODED_ERR,
 819         PREP_ENCODED_CHECKSUM_ERR,
 820         PREP_ENCODED_DO_WRITE,
 821 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 822 {
 823         struct bch_fs *c = op->c;
 824         struct bio *bio = &op->wbio.bio;
 825
 826         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 827                 return PREP_ENCODED_OK;
 828
 829         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 830
 831         /* Can we just write the entire extent as is? */
 832         if (op->crc.uncompressed_size == op->crc.live_size &&
 833             op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
 834             op->crc.compressed_size <= wp->sectors_free &&
 835             (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
 836              op->incompressible)) {
 837                 if (!crc_is_compressed(op->crc) &&
 838                     op->csum_type != op->crc.csum_type &&
 839                     bch2_write_rechecksum(c, op, op->csum_type) &&
 840                     !c->opts.no_data_io)
 841                         return PREP_ENCODED_CHECKSUM_ERR;
 842
 843                 return PREP_ENCODED_DO_WRITE;
 844         }
 845
 846         /*
 847          * If the data is compressed and we couldn't write the entire extent as
 848          * is, we have to decompress it:
 849          */
 850         if (crc_is_compressed(op->crc)) {
 851                 struct bch_csum csum;
 852
 853                 if (bch2_write_decrypt(op))
 854                         return PREP_ENCODED_CHECKSUM_ERR;
 855
 856                 /* Last point we can still verify checksum: */
 857                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 858                                          extent_nonce(op->version, op->crc),
 859                                          bio);
 860                 if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 861                         return PREP_ENCODED_CHECKSUM_ERR;
 862
 863                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 864                         return PREP_ENCODED_ERR;
 865         }
 866
 867         /*
 868          * No longer have compressed data after this point - data might be
 869          * encrypted:
 870          */
 871
 872         /*
 873          * If the data is checksummed and we're only writing a subset,
 874          * rechecksum and adjust bio to point to currently live data:
 875          */
 876         if ((op->crc.live_size != op->crc.uncompressed_size ||
 877              op->crc.csum_type != op->csum_type) &&
 878             bch2_write_rechecksum(c, op, op->csum_type) &&
 879             !c->opts.no_data_io)
 880                 return PREP_ENCODED_CHECKSUM_ERR;
 881
 882         /*
 883          * If we want to compress the data, it has to be decrypted:
 884          */
 885         if ((op->compression_opt ||
 886              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 887              bch2_csum_type_is_encryption(op->csum_type)) &&
 888             bch2_write_decrypt(op))
 889                 return PREP_ENCODED_CHECKSUM_ERR;
 890
 891         return PREP_ENCODED_OK;
 892 }
 893
 894 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 895                              struct bio **_dst)
 896 {
 897         struct bch_fs *c = op->c;
 898         struct bio *src = &op->wbio.bio, *dst = src;
 899         struct bvec_iter saved_iter;
 900         void *ec_buf;
 901         unsigned total_output = 0, total_input = 0;
 902         bool bounce = false;
 903         bool page_alloc_failed = false;
 904         int ret, more = 0;
 905
 906         BUG_ON(!bio_sectors(src));
 907
 908         ec_buf = bch2_writepoint_ec_buf(c, wp);
 909
 910         switch (bch2_write_prep_encoded_data(op, wp)) {
 911         case PREP_ENCODED_OK:
 912                 break;
 913         case PREP_ENCODED_ERR:
 914                 ret = -EIO;
 915                 goto err;
 916         case PREP_ENCODED_CHECKSUM_ERR:
 917                 goto csum_err;
 918         case PREP_ENCODED_DO_WRITE:
 919                 /* XXX look for bug here */
 920                 if (ec_buf) {
 921                         dst = bch2_write_bio_alloc(c, wp, src,
 922                                                    &page_alloc_failed,
 923                                                    ec_buf);
 924                         bio_copy_data(dst, src);
 925                         bounce = true;
 926                 }
 927                 init_append_extent(op, wp, op->version, op->crc);
 928                 goto do_write;
 929         }
 930
 931         if (ec_buf ||
 932             op->compression_opt ||
 933             (op->csum_type &&
 934              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 935             (bch2_csum_type_is_encryption(op->csum_type) &&
 936              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 937                 dst = bch2_write_bio_alloc(c, wp, src,
 938                                            &page_alloc_failed,
 939                                            ec_buf);
 940                 bounce = true;
 941         }
 942
 943         saved_iter = dst->bi_iter;
 944
 945         do {
 946                 struct bch_extent_crc_unpacked crc = { 0 };
 947                 struct bversion version = op->version;
 948                 size_t dst_len = 0, src_len = 0;
 949
 950                 if (page_alloc_failed &&
 951                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 952                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 953                         break;
 954
 955                 BUG_ON(op->compression_opt &&
 956                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 957                        bch2_csum_type_is_encryption(op->crc.csum_type));
 958                 BUG_ON(op->compression_opt && !bounce);
 959
 960                 crc.compression_type = op->incompressible
 961                         ? BCH_COMPRESSION_TYPE_incompressible
 962                         : op->compression_opt
 963                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 964                                             op->compression_opt)
 965                         : 0;
 966                 if (!crc_is_compressed(crc)) {
 967                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 968                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 969
 970                         if (op->csum_type)
 971                                 dst_len = min_t(unsigned, dst_len,
 972                                                 c->opts.encoded_extent_max);
 973
 974                         if (bounce) {
 975                                 swap(dst->bi_iter.bi_size, dst_len);
 976                                 bio_copy_data(dst, src);
 977                                 swap(dst->bi_iter.bi_size, dst_len);
 978                         }
 979
 980                         src_len = dst_len;
 981                 }
 982
 983                 BUG_ON(!src_len || !dst_len);
 984
 985                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 986                         if (bversion_zero(version)) {
 987                                 version.lo = atomic64_inc_return(&c->key_version);
 988                         } else {
 989                                 crc.nonce = op->nonce;
 990                                 op->nonce += src_len >> 9;
 991                         }
 992                 }
 993
 994                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 995                     !crc_is_compressed(crc) &&
 996                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
 997                     bch2_csum_type_is_encryption(op->csum_type)) {
 998                         u8 compression_type = crc.compression_type;
 999                         u16 nonce = crc.nonce;
1000                         /*
1001                          * Note: when we're using rechecksum(), we need to be
1002                          * checksumming @src because it has all the data our
1003                          * existing checksum covers - if we bounced (because we
1004                          * were trying to compress), @dst will only have the
1005                          * part of the data the new checksum will cover.
1006                          *
1007                          * But normally we want to be checksumming post bounce,
1008                          * because part of the reason for bouncing is so the
1009                          * data can't be modified (by userspace) while it's in
1010                          * flight.
1011                          */
1012                         if (bch2_rechecksum_bio(c, src, version, op->crc,
1013                                         &crc, &op->crc,
1014                                         src_len >> 9,
1015                                         bio_sectors(src) - (src_len >> 9),
1016                                         op->csum_type))
1017                                 goto csum_err;
1018                         /*
1019                          * rchecksum_bio sets compression_type on crc from op->crc,
1020                          * this isn't always correct as sometimes we're changing
1021                          * an extent from uncompressed to incompressible.
1022                          */
1023                         crc.compression_type = compression_type;
1024                         crc.nonce = nonce;
1025                 } else {
1026                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1027                             bch2_rechecksum_bio(c, src, version, op->crc,
1028                                         NULL, &op->crc,
1029                                         src_len >> 9,
1030                                         bio_sectors(src) - (src_len >> 9),
1031                                         op->crc.csum_type))
1032                                 goto csum_err;
1033
1034                         crc.compressed_size     = dst_len >> 9;
1035                         crc.uncompressed_size   = src_len >> 9;
1036                         crc.live_size           = src_len >> 9;
1037
1038                         swap(dst->bi_iter.bi_size, dst_len);
1039                         ret = bch2_encrypt_bio(c, op->csum_type,
1040                                                extent_nonce(version, crc), dst);
1041                         if (ret)
1042                                 goto err;
1043
1044                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1045                                          extent_nonce(version, crc), dst);
1046                         crc.csum_type = op->csum_type;
1047                         swap(dst->bi_iter.bi_size, dst_len);
1048                 }
1049
1050                 init_append_extent(op, wp, version, crc);
1051
1052                 if (dst != src)
1053                         bio_advance(dst, dst_len);
1054                 bio_advance(src, src_len);
1055                 total_output    += dst_len;
1056                 total_input     += src_len;
1057         } while (dst->bi_iter.bi_size &&
1058                  src->bi_iter.bi_size &&
1059                  wp->sectors_free &&
1060                  !bch2_keylist_realloc(&op->insert_keys,
1061                                       op->inline_keys,
1062                                       ARRAY_SIZE(op->inline_keys),
1063                                       BKEY_EXTENT_U64s_MAX));
1064
1065         more = src->bi_iter.bi_size != 0;
1066
1067         dst->bi_iter = saved_iter;
1068
1069         if (dst == src && more) {
1070                 BUG_ON(total_output != total_input);
1071
1072                 dst = bio_split(src, total_input >> 9,
1073                                 GFP_NOFS, &c->bio_write);
1074                 wbio_init(dst)->put_bio = true;
1075                 /* copy WRITE_SYNC flag */
1076                 dst->bi_opf             = src->bi_opf;
1077         }
1078
1079         dst->bi_iter.bi_size = total_output;
1080 do_write:
1081         *_dst = dst;
1082         return more;
1083 csum_err:
1084         bch_err(c, "%s writ error: error verifying existing checksum while rewriting existing data (memory corruption?)",
1085                 op->flags & BCH_WRITE_MOVE ? "move" : "user");
1086         ret = -EIO;
1087 err:
1088         if (to_wbio(dst)->bounce)
1089                 bch2_bio_free_pages_pool(c, dst);
1090         if (to_wbio(dst)->put_bio)
1091                 bio_put(dst);
1092
1093         return ret;
1094 }
1095
1096 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1097                                      struct bkey_s_c k)
1098 {
1099         struct bch_fs *c = op->c;
1100         struct bkey_s_c_extent e;
1101         struct extent_ptr_decoded p;
1102         const union bch_extent_entry *entry;
1103         unsigned replicas = 0;
1104
1105         if (k.k->type != KEY_TYPE_extent)
1106                 return false;
1107
1108         e = bkey_s_c_to_extent(k);
1109
1110         rcu_read_lock();
1111         extent_for_each_ptr_decode(e, p, entry) {
1112                 if (crc_is_encoded(p.crc) || p.has_ec) {
1113                         rcu_read_unlock();
1114                         return false;
1115                 }
1116
1117                 replicas += bch2_extent_ptr_durability(c, &p);
1118         }
1119         rcu_read_unlock();
1120
1121         return replicas >= op->opts.data_replicas;
1122 }
1123
1124 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1125                                                   struct btree_iter *iter,
1126                                                   struct bkey_i *orig,
1127                                                   struct bkey_s_c k,
1128                                                   u64 new_i_size)
1129 {
1130         if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1131                 /* trace this */
1132                 return 0;
1133         }
1134
1135         struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1136         int ret = PTR_ERR_OR_ZERO(new);
1137         if (ret)
1138                 return ret;
1139
1140         bch2_cut_front(bkey_start_pos(&orig->k), new);
1141         bch2_cut_back(orig->k.p, new);
1142
1143         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1144         bkey_for_each_ptr(ptrs, ptr)
1145                 ptr->unwritten = 0;
1146
1147         /*
1148          * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1149          * that was done when we kicked off the write, and here it's important
1150          * that we update the extent that we wrote to - even if a snapshot has
1151          * since been created. The write is still outstanding, so we're ok
1152          * w.r.t. snapshot atomicity:
1153          */
1154         return  bch2_extent_update_i_size_sectors(trans, iter,
1155                                         min(new->k.p.offset << 9, new_i_size), 0) ?:
1156                 bch2_trans_update(trans, iter, new,
1157                                   BTREE_UPDATE_internal_snapshot_node);
1158 }
1159
1160 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1161 {
1162         struct bch_fs *c = op->c;
1163         struct btree_trans *trans = bch2_trans_get(c);
1164
1165         for_each_keylist_key(&op->insert_keys, orig) {
1166                 int ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
1167                                      bkey_start_pos(&orig->k), orig->k.p,
1168                                      BTREE_ITER_intent, k,
1169                                      NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1170                         bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1171                 }));
1172
1173                 if (ret && !bch2_err_matches(ret, EROFS)) {
1174                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1175
1176                         bch_err_inum_offset_ratelimited(c,
1177                                 insert->k.p.inode, insert->k.p.offset << 9,
1178                                 "%s write error while doing btree update: %s",
1179                                 op->flags & BCH_WRITE_MOVE ? "move" : "user",
1180                                 bch2_err_str(ret));
1181                 }
1182
1183                 if (ret) {
1184                         op->error = ret;
1185                         break;
1186                 }
1187         }
1188
1189         bch2_trans_put(trans);
1190 }
1191
1192 static void __bch2_nocow_write_done(struct bch_write_op *op)
1193 {
1194         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1195                 op->error = -EIO;
1196         } else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1197                 bch2_nocow_write_convert_unwritten(op);
1198 }
1199
1200 static CLOSURE_CALLBACK(bch2_nocow_write_done)
1201 {
1202         closure_type(op, struct bch_write_op, cl);
1203
1204         __bch2_nocow_write_done(op);
1205         bch2_write_done(cl);
1206 }
1207
1208 struct bucket_to_lock {
1209         struct bpos             b;
1210         unsigned                gen;
1211         struct nocow_lock_bucket *l;
1212 };
1213
1214 static void bch2_nocow_write(struct bch_write_op *op)
1215 {
1216         struct bch_fs *c = op->c;
1217         struct btree_trans *trans;
1218         struct btree_iter iter;
1219         struct bkey_s_c k;
1220         DARRAY_PREALLOCATED(struct bucket_to_lock, 3) buckets;
1221         u32 snapshot;
1222         struct bucket_to_lock *stale_at;
1223         int ret;
1224
1225         if (op->flags & BCH_WRITE_MOVE)
1226                 return;
1227
1228         darray_init(&buckets);
1229         trans = bch2_trans_get(c);
1230 retry:
1231         bch2_trans_begin(trans);
1232
1233         ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1234         if (unlikely(ret))
1235                 goto err;
1236
1237         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1238                              SPOS(op->pos.inode, op->pos.offset, snapshot),
1239                              BTREE_ITER_slots);
1240         while (1) {
1241                 struct bio *bio = &op->wbio.bio;
1242
1243                 buckets.nr = 0;
1244
1245                 ret = bch2_trans_relock(trans);
1246                 if (ret)
1247                         break;
1248
1249                 k = bch2_btree_iter_peek_slot(&iter);
1250                 ret = bkey_err(k);
1251                 if (ret)
1252                         break;
1253
1254                 /* fall back to normal cow write path? */
1255                 if (unlikely(k.k->p.snapshot != snapshot ||
1256                              !bch2_extent_is_writeable(op, k)))
1257                         break;
1258
1259                 if (bch2_keylist_realloc(&op->insert_keys,
1260                                          op->inline_keys,
1261                                          ARRAY_SIZE(op->inline_keys),
1262                                          k.k->u64s))
1263                         break;
1264
1265                 /* Get iorefs before dropping btree locks: */
1266                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1267                 bkey_for_each_ptr(ptrs, ptr) {
1268                         struct bch_dev *ca = bch2_dev_get_ioref(c, ptr->dev, WRITE);
1269                         if (unlikely(!ca))
1270                                 goto err_get_ioref;
1271
1272                         struct bpos b = PTR_BUCKET_POS(ca, ptr);
1273                         struct nocow_lock_bucket *l =
1274                                 bucket_nocow_lock(&c->nocow_locks, bucket_to_u64(b));
1275                         prefetch(l);
1276
1277                         /* XXX allocating memory with btree locks held - rare */
1278                         darray_push_gfp(&buckets, ((struct bucket_to_lock) {
1279                                                    .b = b, .gen = ptr->gen, .l = l,
1280                                                    }), GFP_KERNEL|__GFP_NOFAIL);
1281
1282                         if (ptr->unwritten)
1283                                 op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1284                 }
1285
1286                 /* Unlock before taking nocow locks, doing IO: */
1287                 bkey_reassemble(op->insert_keys.top, k);
1288                 bch2_trans_unlock(trans);
1289
1290                 bch2_cut_front(op->pos, op->insert_keys.top);
1291                 if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1292                         bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1293
1294                 darray_for_each(buckets, i) {
1295                         struct bch_dev *ca = bch2_dev_have_ref(c, i->b.inode);
1296
1297                         __bch2_bucket_nocow_lock(&c->nocow_locks, i->l,
1298                                                  bucket_to_u64(i->b),
1299                                                  BUCKET_NOCOW_LOCK_UPDATE);
1300
1301                         rcu_read_lock();
1302                         bool stale = gen_after(*bucket_gen(ca, i->b.offset), i->gen);
1303                         rcu_read_unlock();
1304
1305                         if (unlikely(stale)) {
1306                                 stale_at = i;
1307                                 goto err_bucket_stale;
1308                         }
1309                 }
1310
1311                 bio = &op->wbio.bio;
1312                 if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1313                         bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1314                                         GFP_KERNEL, &c->bio_write);
1315                         wbio_init(bio)->put_bio = true;
1316                         bio->bi_opf = op->wbio.bio.bi_opf;
1317                 } else {
1318                         op->flags |= BCH_WRITE_DONE;
1319                 }
1320
1321                 op->pos.offset += bio_sectors(bio);
1322                 op->written += bio_sectors(bio);
1323
1324                 bio->bi_end_io  = bch2_write_endio;
1325                 bio->bi_private = &op->cl;
1326                 bio->bi_opf |= REQ_OP_WRITE;
1327                 closure_get(&op->cl);
1328                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1329                                           op->insert_keys.top, true);
1330
1331                 bch2_keylist_push(&op->insert_keys);
1332                 if (op->flags & BCH_WRITE_DONE)
1333                         break;
1334                 bch2_btree_iter_advance(&iter);
1335         }
1336 out:
1337         bch2_trans_iter_exit(trans, &iter);
1338 err:
1339         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1340                 goto retry;
1341
1342         if (ret) {
1343                 bch_err_inum_offset_ratelimited(c,
1344                         op->pos.inode, op->pos.offset << 9,
1345                         "%s: btree lookup error %s", __func__, bch2_err_str(ret));
1346                 op->error = ret;
1347                 op->flags |= BCH_WRITE_DONE;
1348         }
1349
1350         bch2_trans_put(trans);
1351         darray_exit(&buckets);
1352
1353         /* fallback to cow write path? */
1354         if (!(op->flags & BCH_WRITE_DONE)) {
1355                 closure_sync(&op->cl);
1356                 __bch2_nocow_write_done(op);
1357                 op->insert_keys.top = op->insert_keys.keys;
1358         } else if (op->flags & BCH_WRITE_SYNC) {
1359                 closure_sync(&op->cl);
1360                 bch2_nocow_write_done(&op->cl.work);
1361         } else {
1362                 /*
1363                  * XXX
1364                  * needs to run out of process context because ei_quota_lock is
1365                  * a mutex
1366                  */
1367                 continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1368         }
1369         return;
1370 err_get_ioref:
1371         darray_for_each(buckets, i)
1372                 percpu_ref_put(&bch2_dev_have_ref(c, i->b.inode)->io_ref);
1373
1374         /* Fall back to COW path: */
1375         goto out;
1376 err_bucket_stale:
1377         darray_for_each(buckets, i) {
1378                 bch2_bucket_nocow_unlock(&c->nocow_locks, i->b, BUCKET_NOCOW_LOCK_UPDATE);
1379                 if (i == stale_at)
1380                         break;
1381         }
1382
1383         /* We can retry this: */
1384         ret = -BCH_ERR_transaction_restart;
1385         goto err_get_ioref;
1386 }
1387
1388 static void __bch2_write(struct bch_write_op *op)
1389 {
1390         struct bch_fs *c = op->c;
1391         struct write_point *wp = NULL;
1392         struct bio *bio = NULL;
1393         unsigned nofs_flags;
1394         int ret;
1395
1396         nofs_flags = memalloc_nofs_save();
1397
1398         if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1399                 bch2_nocow_write(op);
1400                 if (op->flags & BCH_WRITE_DONE)
1401                         goto out_nofs_restore;
1402         }
1403 again:
1404         memset(&op->failed, 0, sizeof(op->failed));
1405
1406         do {
1407                 struct bkey_i *key_to_write;
1408                 unsigned key_to_write_offset = op->insert_keys.top_p -
1409                         op->insert_keys.keys_p;
1410
1411                 /* +1 for possible cache device: */
1412                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1413                     ARRAY_SIZE(op->open_buckets.v))
1414                         break;
1415
1416                 if (bch2_keylist_realloc(&op->insert_keys,
1417                                         op->inline_keys,
1418                                         ARRAY_SIZE(op->inline_keys),
1419                                         BKEY_EXTENT_U64s_MAX))
1420                         break;
1421
1422                 /*
1423                  * The copygc thread is now global, which means it's no longer
1424                  * freeing up space on specific disks, which means that
1425                  * allocations for specific disks may hang arbitrarily long:
1426                  */
1427                 ret = bch2_trans_do(c, NULL, NULL, 0,
1428                         bch2_alloc_sectors_start_trans(trans,
1429                                 op->target,
1430                                 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1431                                 op->write_point,
1432                                 &op->devs_have,
1433                                 op->nr_replicas,
1434                                 op->nr_replicas_required,
1435                                 op->watermark,
1436                                 op->flags,
1437                                 (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1438                                               BCH_WRITE_ONLY_SPECIFIED_DEVS))
1439                                 ? NULL : &op->cl, &wp));
1440                 if (unlikely(ret)) {
1441                         if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1442                                 break;
1443
1444                         goto err;
1445                 }
1446
1447                 EBUG_ON(!wp);
1448
1449                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1450                 ret = bch2_write_extent(op, wp, &bio);
1451
1452                 bch2_alloc_sectors_done_inlined(c, wp);
1453 err:
1454                 if (ret <= 0) {
1455                         op->flags |= BCH_WRITE_DONE;
1456
1457                         if (ret < 0) {
1458                                 if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT))
1459                                         bch_err_inum_offset_ratelimited(c,
1460                                                 op->pos.inode,
1461                                                 op->pos.offset << 9,
1462                                                 "%s(): %s error: %s", __func__,
1463                                                 op->flags & BCH_WRITE_MOVE ? "move" : "user",
1464                                                 bch2_err_str(ret));
1465                                 op->error = ret;
1466                                 break;
1467                         }
1468                 }
1469
1470                 bio->bi_end_io  = bch2_write_endio;
1471                 bio->bi_private = &op->cl;
1472                 bio->bi_opf |= REQ_OP_WRITE;
1473
1474                 closure_get(bio->bi_private);
1475
1476                 key_to_write = (void *) (op->insert_keys.keys_p +
1477                                          key_to_write_offset);
1478
1479                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1480                                           key_to_write, false);
1481         } while (ret);
1482
1483         /*
1484          * Sync or no?
1485          *
1486          * If we're running asynchronously, wne may still want to block
1487          * synchronously here if we weren't able to submit all of the IO at
1488          * once, as that signals backpressure to the caller.
1489          */
1490         if ((op->flags & BCH_WRITE_SYNC) ||
1491             (!(op->flags & BCH_WRITE_DONE) &&
1492              !(op->flags & BCH_WRITE_IN_WORKER))) {
1493                 if (closure_sync_timeout(&op->cl, HZ * 10)) {
1494                         bch2_print_allocator_stuck(c);
1495                         closure_sync(&op->cl);
1496                 }
1497
1498                 __bch2_write_index(op);
1499
1500                 if (!(op->flags & BCH_WRITE_DONE))
1501                         goto again;
1502                 bch2_write_done(&op->cl);
1503         } else {
1504                 bch2_write_queue(op, wp);
1505                 continue_at(&op->cl, bch2_write_index, NULL);
1506         }
1507 out_nofs_restore:
1508         memalloc_nofs_restore(nofs_flags);
1509 }
1510
1511 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1512 {
1513         struct bio *bio = &op->wbio.bio;
1514         struct bvec_iter iter;
1515         struct bkey_i_inline_data *id;
1516         unsigned sectors;
1517         int ret;
1518
1519         memset(&op->failed, 0, sizeof(op->failed));
1520
1521         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1522         op->flags |= BCH_WRITE_DONE;
1523
1524         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1525
1526         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1527                                    ARRAY_SIZE(op->inline_keys),
1528                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1529         if (ret) {
1530                 op->error = ret;
1531                 goto err;
1532         }
1533
1534         sectors = bio_sectors(bio);
1535         op->pos.offset += sectors;
1536
1537         id = bkey_inline_data_init(op->insert_keys.top);
1538         id->k.p         = op->pos;
1539         id->k.version   = op->version;
1540         id->k.size      = sectors;
1541
1542         iter = bio->bi_iter;
1543         iter.bi_size = data_len;
1544         memcpy_from_bio(id->v.data, bio, iter);
1545
1546         while (data_len & 7)
1547                 id->v.data[data_len++] = '\0';
1548         set_bkey_val_bytes(&id->k, data_len);
1549         bch2_keylist_push(&op->insert_keys);
1550
1551         __bch2_write_index(op);
1552 err:
1553         bch2_write_done(&op->cl);
1554 }
1555
1556 /**
1557  * bch2_write() - handle a write to a cache device or flash only volume
1558  * @cl:         &bch_write_op->cl
1559  *
1560  * This is the starting point for any data to end up in a cache device; it could
1561  * be from a normal write, or a writeback write, or a write to a flash only
1562  * volume - it's also used by the moving garbage collector to compact data in
1563  * mostly empty buckets.
1564  *
1565  * It first writes the data to the cache, creating a list of keys to be inserted
1566  * (if the data won't fit in a single open bucket, there will be multiple keys);
1567  * after the data is written it calls bch_journal, and after the keys have been
1568  * added to the next journal write they're inserted into the btree.
1569  *
1570  * If op->discard is true, instead of inserting the data it invalidates the
1571  * region of the cache represented by op->bio and op->inode.
1572  */
1573 CLOSURE_CALLBACK(bch2_write)
1574 {
1575         closure_type(op, struct bch_write_op, cl);
1576         struct bio *bio = &op->wbio.bio;
1577         struct bch_fs *c = op->c;
1578         unsigned data_len;
1579
1580         EBUG_ON(op->cl.parent);
1581         BUG_ON(!op->nr_replicas);
1582         BUG_ON(!op->write_point.v);
1583         BUG_ON(bkey_eq(op->pos, POS_MAX));
1584
1585         op->nr_replicas_required = min_t(unsigned, op->nr_replicas_required, op->nr_replicas);
1586         op->start_time = local_clock();
1587         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1588         wbio_init(bio)->put_bio = false;
1589
1590         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1591                 bch_err_inum_offset_ratelimited(c,
1592                         op->pos.inode,
1593                         op->pos.offset << 9,
1594                         "%s write error: misaligned write",
1595                         op->flags & BCH_WRITE_MOVE ? "move" : "user");
1596                 op->error = -EIO;
1597                 goto err;
1598         }
1599
1600         if (c->opts.nochanges) {
1601                 op->error = -BCH_ERR_erofs_no_writes;
1602                 goto err;
1603         }
1604
1605         if (!(op->flags & BCH_WRITE_MOVE) &&
1606             !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1607                 op->error = -BCH_ERR_erofs_no_writes;
1608                 goto err;
1609         }
1610
1611         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1612         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1613
1614         data_len = min_t(u64, bio->bi_iter.bi_size,
1615                          op->new_i_size - (op->pos.offset << 9));
1616
1617         if (c->opts.inline_data &&
1618             data_len <= min(block_bytes(c) / 2, 1024U)) {
1619                 bch2_write_data_inline(op, data_len);
1620                 return;
1621         }
1622
1623         __bch2_write(op);
1624         return;
1625 err:
1626         bch2_disk_reservation_put(c, &op->res);
1627
1628         closure_debug_destroy(&op->cl);
1629         if (op->end_io)
1630                 op->end_io(op);
1631 }
1632
1633 static const char * const bch2_write_flags[] = {
1634 #define x(f)    #f,
1635         BCH_WRITE_FLAGS()
1636 #undef x
1637         NULL
1638 };
1639
1640 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1641 {
1642         prt_str(out, "pos: ");
1643         bch2_bpos_to_text(out, op->pos);
1644         prt_newline(out);
1645         printbuf_indent_add(out, 2);
1646
1647         prt_str(out, "started: ");
1648         bch2_pr_time_units(out, local_clock() - op->start_time);
1649         prt_newline(out);
1650
1651         prt_str(out, "flags: ");
1652         prt_bitflags(out, bch2_write_flags, op->flags);
1653         prt_newline(out);
1654
1655         prt_printf(out, "ref: %u\n", closure_nr_remaining(&op->cl));
1656
1657         printbuf_indent_sub(out, 2);
1658 }
1659
1660 void bch2_fs_io_write_exit(struct bch_fs *c)
1661 {
1662         mempool_exit(&c->bio_bounce_pages);
1663         bioset_exit(&c->replica_set);
1664         bioset_exit(&c->bio_write);
1665 }
1666
1667 int bch2_fs_io_write_init(struct bch_fs *c)
1668 {
1669         if (bioset_init(&c->bio_write,   1, offsetof(struct bch_write_bio, bio), BIOSET_NEED_BVECS) ||
1670             bioset_init(&c->replica_set, 4, offsetof(struct bch_write_bio, bio), 0))
1671                 return -BCH_ERR_ENOMEM_bio_write_init;
1672
1673         if (mempool_init_page_pool(&c->bio_bounce_pages,
1674                                    max_t(unsigned,
1675                                          c->opts.btree_node_size,
1676                                          c->opts.encoded_extent_max) /
1677                                    PAGE_SIZE, 0))
1678                 return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1679
1680         return 0;
1681 }