git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/io_write.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   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], 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         struct btree_iter iter;
 203         struct bkey_i *k;
 204         struct bkey_i_inode_v3 *inode;
 205         /*
 206          * Crazy performance optimization:
 207          * Every extent update needs to also update the inode: the inode trigger
 208          * will set bi->journal_seq to the journal sequence number of this
 209          * transaction - for fsync.
 210          *
 211          * But if that's the only reason we're updating the inode (we're not
 212          * updating bi_size or bi_sectors), then we don't need the inode update
 213          * to be journalled - if we crash, the bi_journal_seq update will be
 214          * lost, but that's fine.
 215          */
 216         unsigned inode_update_flags = BTREE_UPDATE_NOJOURNAL;
 217         int ret;
 218
 219         k = bch2_bkey_get_mut_noupdate(trans, &iter, BTREE_ID_inodes,
 220                               SPOS(0,
 221                                    extent_iter->pos.inode,
 222                                    extent_iter->snapshot),
 223                               BTREE_ITER_CACHED);
 224         ret = PTR_ERR_OR_ZERO(k);
 225         if (unlikely(ret))
 226                 return ret;
 227
 228         if (unlikely(k->k.type != KEY_TYPE_inode_v3)) {
 229                 k = bch2_inode_to_v3(trans, k);
 230                 ret = PTR_ERR_OR_ZERO(k);
 231                 if (unlikely(ret))
 232                         goto err;
 233         }
 234
 235         inode = bkey_i_to_inode_v3(k);
 236
 237         if (!(le64_to_cpu(inode->v.bi_flags) & BCH_INODE_i_size_dirty) &&
 238             new_i_size > le64_to_cpu(inode->v.bi_size)) {
 239                 inode->v.bi_size = cpu_to_le64(new_i_size);
 240                 inode_update_flags = 0;
 241         }
 242
 243         if (i_sectors_delta) {
 244                 le64_add_cpu(&inode->v.bi_sectors, i_sectors_delta);
 245                 inode_update_flags = 0;
 246         }
 247
 248         if (inode->k.p.snapshot != iter.snapshot) {
 249                 inode->k.p.snapshot = iter.snapshot;
 250                 inode_update_flags = 0;
 251         }
 252
 253         ret = bch2_trans_update(trans, &iter, &inode->k_i,
 254                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 255                                 inode_update_flags);
 256 err:
 257         bch2_trans_iter_exit(trans, &iter);
 258         return ret;
 259 }
 260
 261 int bch2_extent_update(struct btree_trans *trans,
 262                        subvol_inum inum,
 263                        struct btree_iter *iter,
 264                        struct bkey_i *k,
 265                        struct disk_reservation *disk_res,
 266                        u64 new_i_size,
 267                        s64 *i_sectors_delta_total,
 268                        bool check_enospc)
 269 {
 270         struct bpos next_pos;
 271         bool usage_increasing;
 272         s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 273         int ret;
 274
 275         /*
 276          * This traverses us the iterator without changing iter->path->pos to
 277          * search_key() (which is pos + 1 for extents): we want there to be a
 278          * path already traversed at iter->pos because
 279          * bch2_trans_extent_update() will use it to attempt extent merging
 280          */
 281         ret = __bch2_btree_iter_traverse(iter);
 282         if (ret)
 283                 return ret;
 284
 285         ret = bch2_extent_trim_atomic(trans, iter, k);
 286         if (ret)
 287                 return ret;
 288
 289         next_pos = k->k.p;
 290
 291         ret = bch2_sum_sector_overwrites(trans, iter, k,
 292                         &usage_increasing,
 293                         &i_sectors_delta,
 294                         &disk_sectors_delta);
 295         if (ret)
 296                 return ret;
 297
 298         if (disk_res &&
 299             disk_sectors_delta > (s64) disk_res->sectors) {
 300                 ret = bch2_disk_reservation_add(trans->c, disk_res,
 301                                         disk_sectors_delta - disk_res->sectors,
 302                                         !check_enospc || !usage_increasing
 303                                         ? BCH_DISK_RESERVATION_NOFAIL : 0);
 304                 if (ret)
 305                         return ret;
 306         }
 307
 308         /*
 309          * Note:
 310          * We always have to do an inode update - even when i_size/i_sectors
 311          * aren't changing - for fsync to work properly; fsync relies on
 312          * inode->bi_journal_seq which is updated by the trigger code:
 313          */
 314         ret =   bch2_extent_update_i_size_sectors(trans, iter,
 315                                                   min(k->k.p.offset << 9, new_i_size),
 316                                                   i_sectors_delta) ?:
 317                 bch2_trans_update(trans, iter, k, 0) ?:
 318                 bch2_trans_commit(trans, disk_res, NULL,
 319                                 BCH_TRANS_COMMIT_no_check_rw|
 320                                 BCH_TRANS_COMMIT_no_enospc);
 321         if (unlikely(ret))
 322                 return ret;
 323
 324         if (i_sectors_delta_total)
 325                 *i_sectors_delta_total += i_sectors_delta;
 326         bch2_btree_iter_set_pos(iter, next_pos);
 327         return 0;
 328 }
 329
 330 static int bch2_write_index_default(struct bch_write_op *op)
 331 {
 332         struct bch_fs *c = op->c;
 333         struct bkey_buf sk;
 334         struct keylist *keys = &op->insert_keys;
 335         struct bkey_i *k = bch2_keylist_front(keys);
 336         struct btree_trans *trans = bch2_trans_get(c);
 337         struct btree_iter iter;
 338         subvol_inum inum = {
 339                 .subvol = op->subvol,
 340                 .inum   = k->k.p.inode,
 341         };
 342         int ret;
 343
 344         BUG_ON(!inum.subvol);
 345
 346         bch2_bkey_buf_init(&sk);
 347
 348         do {
 349                 bch2_trans_begin(trans);
 350
 351                 k = bch2_keylist_front(keys);
 352                 bch2_bkey_buf_copy(&sk, c, k);
 353
 354                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol,
 355                                                   &sk.k->k.p.snapshot);
 356                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 357                         continue;
 358                 if (ret)
 359                         break;
 360
 361                 bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
 362                                      bkey_start_pos(&sk.k->k),
 363                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 364
 365                 ret =   bch2_bkey_set_needs_rebalance(c, sk.k,
 366                                         op->opts.background_target,
 367                                         op->opts.background_compression) ?:
 368                         bch2_extent_update(trans, inum, &iter, sk.k,
 369                                         &op->res,
 370                                         op->new_i_size, &op->i_sectors_delta,
 371                                         op->flags & BCH_WRITE_CHECK_ENOSPC);
 372                 bch2_trans_iter_exit(trans, &iter);
 373
 374                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 375                         continue;
 376                 if (ret)
 377                         break;
 378
 379                 if (bkey_ge(iter.pos, k->k.p))
 380                         bch2_keylist_pop_front(&op->insert_keys);
 381                 else
 382                         bch2_cut_front(iter.pos, k);
 383         } while (!bch2_keylist_empty(keys));
 384
 385         bch2_trans_put(trans);
 386         bch2_bkey_buf_exit(&sk, c);
 387
 388         return ret;
 389 }
 390
 391 /* Writes */
 392
 393 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 394                                enum bch_data_type type,
 395                                const struct bkey_i *k,
 396                                bool nocow)
 397 {
 398         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 399         const struct bch_extent_ptr *ptr;
 400         struct bch_write_bio *n;
 401         struct bch_dev *ca;
 402
 403         BUG_ON(c->opts.nochanges);
 404
 405         bkey_for_each_ptr(ptrs, ptr) {
 406                 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
 407                        !c->devs[ptr->dev]);
 408
 409                 ca = bch_dev_bkey_exists(c, ptr->dev);
 410
 411                 if (to_entry(ptr + 1) < ptrs.end) {
 412                         n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
 413                                                 GFP_NOFS, &ca->replica_set));
 414
 415                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 416                         n->bio.bi_private       = wbio->bio.bi_private;
 417                         n->parent               = wbio;
 418                         n->split                = true;
 419                         n->bounce               = false;
 420                         n->put_bio              = true;
 421                         n->bio.bi_opf           = wbio->bio.bi_opf;
 422                         bio_inc_remaining(&wbio->bio);
 423                 } else {
 424                         n = wbio;
 425                         n->split                = false;
 426                 }
 427
 428                 n->c                    = c;
 429                 n->dev                  = ptr->dev;
 430                 n->have_ioref           = nocow || bch2_dev_get_ioref(ca,
 431                                         type == BCH_DATA_btree ? READ : WRITE);
 432                 n->nocow                = nocow;
 433                 n->submit_time          = local_clock();
 434                 n->inode_offset         = bkey_start_offset(&k->k);
 435                 n->bio.bi_iter.bi_sector = ptr->offset;
 436
 437                 if (likely(n->have_ioref)) {
 438                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 439                                      bio_sectors(&n->bio));
 440
 441                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 442
 443                         if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
 444                                 bio_endio(&n->bio);
 445                                 continue;
 446                         }
 447
 448                         submit_bio(&n->bio);
 449                 } else {
 450                         n->bio.bi_status        = BLK_STS_REMOVED;
 451                         bio_endio(&n->bio);
 452                 }
 453         }
 454 }
 455
 456 static void __bch2_write(struct bch_write_op *);
 457
 458 static void bch2_write_done(struct closure *cl)
 459 {
 460         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 461         struct bch_fs *c = op->c;
 462
 463         EBUG_ON(op->open_buckets.nr);
 464
 465         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 466         bch2_disk_reservation_put(c, &op->res);
 467
 468         if (!(op->flags & BCH_WRITE_MOVE))
 469                 bch2_write_ref_put(c, BCH_WRITE_REF_write);
 470         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 471
 472         EBUG_ON(cl->parent);
 473         closure_debug_destroy(cl);
 474         if (op->end_io)
 475                 op->end_io(op);
 476 }
 477
 478 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
 479 {
 480         struct keylist *keys = &op->insert_keys;
 481         struct bch_extent_ptr *ptr;
 482         struct bkey_i *src, *dst = keys->keys, *n;
 483
 484         for (src = keys->keys; src != keys->top; src = n) {
 485                 n = bkey_next(src);
 486
 487                 if (bkey_extent_is_direct_data(&src->k)) {
 488                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 489                                             test_bit(ptr->dev, op->failed.d));
 490
 491                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
 492                                 return -EIO;
 493                 }
 494
 495                 if (dst != src)
 496                         memmove_u64s_down(dst, src, src->k.u64s);
 497                 dst = bkey_next(dst);
 498         }
 499
 500         keys->top = dst;
 501         return 0;
 502 }
 503
 504 /**
 505  * __bch2_write_index - after a write, update index to point to new data
 506  * @op:         bch_write_op to process
 507  */
 508 static void __bch2_write_index(struct bch_write_op *op)
 509 {
 510         struct bch_fs *c = op->c;
 511         struct keylist *keys = &op->insert_keys;
 512         unsigned dev;
 513         int ret = 0;
 514
 515         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
 516                 ret = bch2_write_drop_io_error_ptrs(op);
 517                 if (ret)
 518                         goto err;
 519         }
 520
 521         if (!bch2_keylist_empty(keys)) {
 522                 u64 sectors_start = keylist_sectors(keys);
 523
 524                 ret = !(op->flags & BCH_WRITE_MOVE)
 525                         ? bch2_write_index_default(op)
 526                         : bch2_data_update_index_update(op);
 527
 528                 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 529                 BUG_ON(keylist_sectors(keys) && !ret);
 530
 531                 op->written += sectors_start - keylist_sectors(keys);
 532
 533                 if (ret && !bch2_err_matches(ret, EROFS)) {
 534                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
 535
 536                         bch_err_inum_offset_ratelimited(c,
 537                                 insert->k.p.inode, insert->k.p.offset << 9,
 538                                 "write error while doing btree update: %s",
 539                                 bch2_err_str(ret));
 540                 }
 541
 542                 if (ret)
 543                         goto err;
 544         }
 545 out:
 546         /* If some a bucket wasn't written, we can't erasure code it: */
 547         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 548                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 549
 550         bch2_open_buckets_put(c, &op->open_buckets);
 551         return;
 552 err:
 553         keys->top = keys->keys;
 554         op->error = ret;
 555         op->flags |= BCH_WRITE_DONE;
 556         goto out;
 557 }
 558
 559 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
 560 {
 561         if (state != wp->state) {
 562                 u64 now = ktime_get_ns();
 563
 564                 if (wp->last_state_change &&
 565                     time_after64(now, wp->last_state_change))
 566                         wp->time[wp->state] += now - wp->last_state_change;
 567                 wp->state = state;
 568                 wp->last_state_change = now;
 569         }
 570 }
 571
 572 static inline void wp_update_state(struct write_point *wp, bool running)
 573 {
 574         enum write_point_state state;
 575
 576         state = running                  ? WRITE_POINT_running :
 577                 !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
 578                                          : WRITE_POINT_stopped;
 579
 580         __wp_update_state(wp, state);
 581 }
 582
 583 static void bch2_write_index(struct closure *cl)
 584 {
 585         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 586         struct write_point *wp = op->wp;
 587         struct workqueue_struct *wq = index_update_wq(op);
 588         unsigned long flags;
 589
 590         if ((op->flags & BCH_WRITE_DONE) &&
 591             (op->flags & BCH_WRITE_MOVE))
 592                 bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
 593
 594         spin_lock_irqsave(&wp->writes_lock, flags);
 595         if (wp->state == WRITE_POINT_waiting_io)
 596                 __wp_update_state(wp, WRITE_POINT_waiting_work);
 597         list_add_tail(&op->wp_list, &wp->writes);
 598         spin_unlock_irqrestore (&wp->writes_lock, flags);
 599
 600         queue_work(wq, &wp->index_update_work);
 601 }
 602
 603 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
 604 {
 605         op->wp = wp;
 606
 607         if (wp->state == WRITE_POINT_stopped) {
 608                 spin_lock_irq(&wp->writes_lock);
 609                 __wp_update_state(wp, WRITE_POINT_waiting_io);
 610                 spin_unlock_irq(&wp->writes_lock);
 611         }
 612 }
 613
 614 void bch2_write_point_do_index_updates(struct work_struct *work)
 615 {
 616         struct write_point *wp =
 617                 container_of(work, struct write_point, index_update_work);
 618         struct bch_write_op *op;
 619
 620         while (1) {
 621                 spin_lock_irq(&wp->writes_lock);
 622                 op = list_first_entry_or_null(&wp->writes, struct bch_write_op, wp_list);
 623                 if (op)
 624                         list_del(&op->wp_list);
 625                 wp_update_state(wp, op != NULL);
 626                 spin_unlock_irq(&wp->writes_lock);
 627
 628                 if (!op)
 629                         break;
 630
 631                 op->flags |= BCH_WRITE_IN_WORKER;
 632
 633                 __bch2_write_index(op);
 634
 635                 if (!(op->flags & BCH_WRITE_DONE))
 636                         __bch2_write(op);
 637                 else
 638                         bch2_write_done(&op->cl);
 639         }
 640 }
 641
 642 static void bch2_write_endio(struct bio *bio)
 643 {
 644         struct closure *cl              = bio->bi_private;
 645         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 646         struct bch_write_bio *wbio      = to_wbio(bio);
 647         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 648         struct bch_fs *c                = wbio->c;
 649         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
 650
 651         if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
 652                                     op->pos.inode,
 653                                     wbio->inode_offset << 9,
 654                                     "data write error: %s",
 655                                     bch2_blk_status_to_str(bio->bi_status))) {
 656                 set_bit(wbio->dev, op->failed.d);
 657                 op->flags |= BCH_WRITE_IO_ERROR;
 658         }
 659
 660         if (wbio->nocow)
 661                 set_bit(wbio->dev, op->devs_need_flush->d);
 662
 663         if (wbio->have_ioref) {
 664                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 665                 percpu_ref_put(&ca->io_ref);
 666         }
 667
 668         if (wbio->bounce)
 669                 bch2_bio_free_pages_pool(c, bio);
 670
 671         if (wbio->put_bio)
 672                 bio_put(bio);
 673
 674         if (parent)
 675                 bio_endio(&parent->bio);
 676         else
 677                 closure_put(cl);
 678 }
 679
 680 static void init_append_extent(struct bch_write_op *op,
 681                                struct write_point *wp,
 682                                struct bversion version,
 683                                struct bch_extent_crc_unpacked crc)
 684 {
 685         struct bkey_i_extent *e;
 686
 687         op->pos.offset += crc.uncompressed_size;
 688
 689         e = bkey_extent_init(op->insert_keys.top);
 690         e->k.p          = op->pos;
 691         e->k.size       = crc.uncompressed_size;
 692         e->k.version    = version;
 693
 694         if (crc.csum_type ||
 695             crc.compression_type ||
 696             crc.nonce)
 697                 bch2_extent_crc_append(&e->k_i, crc);
 698
 699         bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
 700                                        op->flags & BCH_WRITE_CACHED);
 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(NULL, pages, 0,
 723                                GFP_NOFS, &c->bio_write);
 724         wbio                    = wbio_init(bio);
 725         wbio->put_bio           = true;
 726         /* copy WRITE_SYNC flag */
 727         wbio->bio.bi_opf        = src->bi_opf;
 728
 729         if (buf) {
 730                 bch2_bio_map(bio, buf, output_available);
 731                 return bio;
 732         }
 733
 734         wbio->bounce            = true;
 735
 736         /*
 737          * We can't use mempool for more than c->sb.encoded_extent_max
 738          * worth of pages, but we'd like to allocate more if we can:
 739          */
 740         bch2_bio_alloc_pages_pool(c, bio,
 741                                   min_t(unsigned, output_available,
 742                                         c->opts.encoded_extent_max));
 743
 744         if (bio->bi_iter.bi_size < output_available)
 745                 *page_alloc_failed =
 746                         bch2_bio_alloc_pages(bio,
 747                                              output_available -
 748                                              bio->bi_iter.bi_size,
 749                                              GFP_NOFS) != 0;
 750
 751         return bio;
 752 }
 753
 754 static int bch2_write_rechecksum(struct bch_fs *c,
 755                                  struct bch_write_op *op,
 756                                  unsigned new_csum_type)
 757 {
 758         struct bio *bio = &op->wbio.bio;
 759         struct bch_extent_crc_unpacked new_crc;
 760         int ret;
 761
 762         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 763
 764         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 765             bch2_csum_type_is_encryption(new_csum_type))
 766                 new_csum_type = op->crc.csum_type;
 767
 768         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 769                                   NULL, &new_crc,
 770                                   op->crc.offset, op->crc.live_size,
 771                                   new_csum_type);
 772         if (ret)
 773                 return ret;
 774
 775         bio_advance(bio, op->crc.offset << 9);
 776         bio->bi_iter.bi_size = op->crc.live_size << 9;
 777         op->crc = new_crc;
 778         return 0;
 779 }
 780
 781 static int bch2_write_decrypt(struct bch_write_op *op)
 782 {
 783         struct bch_fs *c = op->c;
 784         struct nonce nonce = extent_nonce(op->version, op->crc);
 785         struct bch_csum csum;
 786         int ret;
 787
 788         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 789                 return 0;
 790
 791         /*
 792          * If we need to decrypt data in the write path, we'll no longer be able
 793          * to verify the existing checksum (poly1305 mac, in this case) after
 794          * it's decrypted - this is the last point we'll be able to reverify the
 795          * checksum:
 796          */
 797         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 798         if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 799                 return -EIO;
 800
 801         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 802         op->crc.csum_type = 0;
 803         op->crc.csum = (struct bch_csum) { 0, 0 };
 804         return ret;
 805 }
 806
 807 static enum prep_encoded_ret {
 808         PREP_ENCODED_OK,
 809         PREP_ENCODED_ERR,
 810         PREP_ENCODED_CHECKSUM_ERR,
 811         PREP_ENCODED_DO_WRITE,
 812 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 813 {
 814         struct bch_fs *c = op->c;
 815         struct bio *bio = &op->wbio.bio;
 816
 817         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 818                 return PREP_ENCODED_OK;
 819
 820         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 821
 822         /* Can we just write the entire extent as is? */
 823         if (op->crc.uncompressed_size == op->crc.live_size &&
 824             op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
 825             op->crc.compressed_size <= wp->sectors_free &&
 826             (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
 827              op->incompressible)) {
 828                 if (!crc_is_compressed(op->crc) &&
 829                     op->csum_type != op->crc.csum_type &&
 830                     bch2_write_rechecksum(c, op, op->csum_type) &&
 831                     !c->opts.no_data_io)
 832                         return PREP_ENCODED_CHECKSUM_ERR;
 833
 834                 return PREP_ENCODED_DO_WRITE;
 835         }
 836
 837         /*
 838          * If the data is compressed and we couldn't write the entire extent as
 839          * is, we have to decompress it:
 840          */
 841         if (crc_is_compressed(op->crc)) {
 842                 struct bch_csum csum;
 843
 844                 if (bch2_write_decrypt(op))
 845                         return PREP_ENCODED_CHECKSUM_ERR;
 846
 847                 /* Last point we can still verify checksum: */
 848                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 849                                          extent_nonce(op->version, op->crc),
 850                                          bio);
 851                 if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 852                         return PREP_ENCODED_CHECKSUM_ERR;
 853
 854                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 855                         return PREP_ENCODED_ERR;
 856         }
 857
 858         /*
 859          * No longer have compressed data after this point - data might be
 860          * encrypted:
 861          */
 862
 863         /*
 864          * If the data is checksummed and we're only writing a subset,
 865          * rechecksum and adjust bio to point to currently live data:
 866          */
 867         if ((op->crc.live_size != op->crc.uncompressed_size ||
 868              op->crc.csum_type != op->csum_type) &&
 869             bch2_write_rechecksum(c, op, op->csum_type) &&
 870             !c->opts.no_data_io)
 871                 return PREP_ENCODED_CHECKSUM_ERR;
 872
 873         /*
 874          * If we want to compress the data, it has to be decrypted:
 875          */
 876         if ((op->compression_opt ||
 877              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 878              bch2_csum_type_is_encryption(op->csum_type)) &&
 879             bch2_write_decrypt(op))
 880                 return PREP_ENCODED_CHECKSUM_ERR;
 881
 882         return PREP_ENCODED_OK;
 883 }
 884
 885 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 886                              struct bio **_dst)
 887 {
 888         struct bch_fs *c = op->c;
 889         struct bio *src = &op->wbio.bio, *dst = src;
 890         struct bvec_iter saved_iter;
 891         void *ec_buf;
 892         unsigned total_output = 0, total_input = 0;
 893         bool bounce = false;
 894         bool page_alloc_failed = false;
 895         int ret, more = 0;
 896
 897         BUG_ON(!bio_sectors(src));
 898
 899         ec_buf = bch2_writepoint_ec_buf(c, wp);
 900
 901         switch (bch2_write_prep_encoded_data(op, wp)) {
 902         case PREP_ENCODED_OK:
 903                 break;
 904         case PREP_ENCODED_ERR:
 905                 ret = -EIO;
 906                 goto err;
 907         case PREP_ENCODED_CHECKSUM_ERR:
 908                 goto csum_err;
 909         case PREP_ENCODED_DO_WRITE:
 910                 /* XXX look for bug here */
 911                 if (ec_buf) {
 912                         dst = bch2_write_bio_alloc(c, wp, src,
 913                                                    &page_alloc_failed,
 914                                                    ec_buf);
 915                         bio_copy_data(dst, src);
 916                         bounce = true;
 917                 }
 918                 init_append_extent(op, wp, op->version, op->crc);
 919                 goto do_write;
 920         }
 921
 922         if (ec_buf ||
 923             op->compression_opt ||
 924             (op->csum_type &&
 925              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 926             (bch2_csum_type_is_encryption(op->csum_type) &&
 927              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 928                 dst = bch2_write_bio_alloc(c, wp, src,
 929                                            &page_alloc_failed,
 930                                            ec_buf);
 931                 bounce = true;
 932         }
 933
 934         saved_iter = dst->bi_iter;
 935
 936         do {
 937                 struct bch_extent_crc_unpacked crc = { 0 };
 938                 struct bversion version = op->version;
 939                 size_t dst_len = 0, src_len = 0;
 940
 941                 if (page_alloc_failed &&
 942                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 943                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 944                         break;
 945
 946                 BUG_ON(op->compression_opt &&
 947                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 948                        bch2_csum_type_is_encryption(op->crc.csum_type));
 949                 BUG_ON(op->compression_opt && !bounce);
 950
 951                 crc.compression_type = op->incompressible
 952                         ? BCH_COMPRESSION_TYPE_incompressible
 953                         : op->compression_opt
 954                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 955                                             op->compression_opt)
 956                         : 0;
 957                 if (!crc_is_compressed(crc)) {
 958                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 959                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 960
 961                         if (op->csum_type)
 962                                 dst_len = min_t(unsigned, dst_len,
 963                                                 c->opts.encoded_extent_max);
 964
 965                         if (bounce) {
 966                                 swap(dst->bi_iter.bi_size, dst_len);
 967                                 bio_copy_data(dst, src);
 968                                 swap(dst->bi_iter.bi_size, dst_len);
 969                         }
 970
 971                         src_len = dst_len;
 972                 }
 973
 974                 BUG_ON(!src_len || !dst_len);
 975
 976                 if (bch2_csum_type_is_encryption(op->csum_type)) {
 977                         if (bversion_zero(version)) {
 978                                 version.lo = atomic64_inc_return(&c->key_version);
 979                         } else {
 980                                 crc.nonce = op->nonce;
 981                                 op->nonce += src_len >> 9;
 982                         }
 983                 }
 984
 985                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
 986                     !crc_is_compressed(crc) &&
 987                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
 988                     bch2_csum_type_is_encryption(op->csum_type)) {
 989                         u8 compression_type = crc.compression_type;
 990                         u16 nonce = crc.nonce;
 991                         /*
 992                          * Note: when we're using rechecksum(), we need to be
 993                          * checksumming @src because it has all the data our
 994                          * existing checksum covers - if we bounced (because we
 995                          * were trying to compress), @dst will only have the
 996                          * part of the data the new checksum will cover.
 997                          *
 998                          * But normally we want to be checksumming post bounce,
 999                          * because part of the reason for bouncing is so the
1000                          * data can't be modified (by userspace) while it's in
1001                          * flight.
1002                          */
1003                         if (bch2_rechecksum_bio(c, src, version, op->crc,
1004                                         &crc, &op->crc,
1005                                         src_len >> 9,
1006                                         bio_sectors(src) - (src_len >> 9),
1007                                         op->csum_type))
1008                                 goto csum_err;
1009                         /*
1010                          * rchecksum_bio sets compression_type on crc from op->crc,
1011                          * this isn't always correct as sometimes we're changing
1012                          * an extent from uncompressed to incompressible.
1013                          */
1014                         crc.compression_type = compression_type;
1015                         crc.nonce = nonce;
1016                 } else {
1017                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1018                             bch2_rechecksum_bio(c, src, version, op->crc,
1019                                         NULL, &op->crc,
1020                                         src_len >> 9,
1021                                         bio_sectors(src) - (src_len >> 9),
1022                                         op->crc.csum_type))
1023                                 goto csum_err;
1024
1025                         crc.compressed_size     = dst_len >> 9;
1026                         crc.uncompressed_size   = src_len >> 9;
1027                         crc.live_size           = src_len >> 9;
1028
1029                         swap(dst->bi_iter.bi_size, dst_len);
1030                         ret = bch2_encrypt_bio(c, op->csum_type,
1031                                                extent_nonce(version, crc), dst);
1032                         if (ret)
1033                                 goto err;
1034
1035                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1036                                          extent_nonce(version, crc), dst);
1037                         crc.csum_type = op->csum_type;
1038                         swap(dst->bi_iter.bi_size, dst_len);
1039                 }
1040
1041                 init_append_extent(op, wp, version, crc);
1042
1043                 if (dst != src)
1044                         bio_advance(dst, dst_len);
1045                 bio_advance(src, src_len);
1046                 total_output    += dst_len;
1047                 total_input     += src_len;
1048         } while (dst->bi_iter.bi_size &&
1049                  src->bi_iter.bi_size &&
1050                  wp->sectors_free &&
1051                  !bch2_keylist_realloc(&op->insert_keys,
1052                                       op->inline_keys,
1053                                       ARRAY_SIZE(op->inline_keys),
1054                                       BKEY_EXTENT_U64s_MAX));
1055
1056         more = src->bi_iter.bi_size != 0;
1057
1058         dst->bi_iter = saved_iter;
1059
1060         if (dst == src && more) {
1061                 BUG_ON(total_output != total_input);
1062
1063                 dst = bio_split(src, total_input >> 9,
1064                                 GFP_NOFS, &c->bio_write);
1065                 wbio_init(dst)->put_bio = true;
1066                 /* copy WRITE_SYNC flag */
1067                 dst->bi_opf             = src->bi_opf;
1068         }
1069
1070         dst->bi_iter.bi_size = total_output;
1071 do_write:
1072         *_dst = dst;
1073         return more;
1074 csum_err:
1075         bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1076         ret = -EIO;
1077 err:
1078         if (to_wbio(dst)->bounce)
1079                 bch2_bio_free_pages_pool(c, dst);
1080         if (to_wbio(dst)->put_bio)
1081                 bio_put(dst);
1082
1083         return ret;
1084 }
1085
1086 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1087                                      struct bkey_s_c k)
1088 {
1089         struct bch_fs *c = op->c;
1090         struct bkey_s_c_extent e;
1091         struct extent_ptr_decoded p;
1092         const union bch_extent_entry *entry;
1093         unsigned replicas = 0;
1094
1095         if (k.k->type != KEY_TYPE_extent)
1096                 return false;
1097
1098         e = bkey_s_c_to_extent(k);
1099         extent_for_each_ptr_decode(e, p, entry) {
1100                 if (crc_is_encoded(p.crc) || p.has_ec)
1101                         return false;
1102
1103                 replicas += bch2_extent_ptr_durability(c, &p);
1104         }
1105
1106         return replicas >= op->opts.data_replicas;
1107 }
1108
1109 static inline void bch2_nocow_write_unlock(struct bch_write_op *op)
1110 {
1111         struct bch_fs *c = op->c;
1112         const struct bch_extent_ptr *ptr;
1113         struct bkey_i *k;
1114
1115         for_each_keylist_key(&op->insert_keys, k) {
1116                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
1117
1118                 bkey_for_each_ptr(ptrs, ptr)
1119                         bch2_bucket_nocow_unlock(&c->nocow_locks,
1120                                                PTR_BUCKET_POS(c, ptr),
1121                                                BUCKET_NOCOW_LOCK_UPDATE);
1122         }
1123 }
1124
1125 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1126                                                   struct btree_iter *iter,
1127                                                   struct bkey_i *orig,
1128                                                   struct bkey_s_c k,
1129                                                   u64 new_i_size)
1130 {
1131         struct bkey_i *new;
1132         struct bkey_ptrs ptrs;
1133         struct bch_extent_ptr *ptr;
1134         int ret;
1135
1136         if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1137                 /* trace this */
1138                 return 0;
1139         }
1140
1141         new = bch2_bkey_make_mut_noupdate(trans, k);
1142         ret = PTR_ERR_OR_ZERO(new);
1143         if (ret)
1144                 return ret;
1145
1146         bch2_cut_front(bkey_start_pos(&orig->k), new);
1147         bch2_cut_back(orig->k.p, new);
1148
1149         ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1150         bkey_for_each_ptr(ptrs, ptr)
1151                 ptr->unwritten = 0;
1152
1153         /*
1154          * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1155          * that was done when we kicked off the write, and here it's important
1156          * that we update the extent that we wrote to - even if a snapshot has
1157          * since been created. The write is still outstanding, so we're ok
1158          * w.r.t. snapshot atomicity:
1159          */
1160         return  bch2_extent_update_i_size_sectors(trans, iter,
1161                                         min(new->k.p.offset << 9, new_i_size), 0) ?:
1162                 bch2_trans_update(trans, iter, new,
1163                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1164 }
1165
1166 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1167 {
1168         struct bch_fs *c = op->c;
1169         struct btree_trans *trans = bch2_trans_get(c);
1170         struct btree_iter iter;
1171         struct bkey_i *orig;
1172         struct bkey_s_c k;
1173         int ret;
1174
1175         for_each_keylist_key(&op->insert_keys, orig) {
1176                 ret = for_each_btree_key_upto_commit(trans, iter, BTREE_ID_extents,
1177                                      bkey_start_pos(&orig->k), orig->k.p,
1178                                      BTREE_ITER_INTENT, k,
1179                                      NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1180                         bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1181                 }));
1182
1183                 if (ret && !bch2_err_matches(ret, EROFS)) {
1184                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1185
1186                         bch_err_inum_offset_ratelimited(c,
1187                                 insert->k.p.inode, insert->k.p.offset << 9,
1188                                 "write error while doing btree update: %s",
1189                                 bch2_err_str(ret));
1190                 }
1191
1192                 if (ret) {
1193                         op->error = ret;
1194                         break;
1195                 }
1196         }
1197
1198         bch2_trans_put(trans);
1199 }
1200
1201 static void __bch2_nocow_write_done(struct bch_write_op *op)
1202 {
1203         bch2_nocow_write_unlock(op);
1204
1205         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1206                 op->error = -EIO;
1207         } else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1208                 bch2_nocow_write_convert_unwritten(op);
1209 }
1210
1211 static void bch2_nocow_write_done(struct closure *cl)
1212 {
1213         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1214
1215         __bch2_nocow_write_done(op);
1216         bch2_write_done(cl);
1217 }
1218
1219 static void bch2_nocow_write(struct bch_write_op *op)
1220 {
1221         struct bch_fs *c = op->c;
1222         struct btree_trans *trans;
1223         struct btree_iter iter;
1224         struct bkey_s_c k;
1225         struct bkey_ptrs_c ptrs;
1226         const struct bch_extent_ptr *ptr;
1227         struct {
1228                 struct bpos     b;
1229                 unsigned        gen;
1230                 struct nocow_lock_bucket *l;
1231         } buckets[BCH_REPLICAS_MAX];
1232         unsigned nr_buckets = 0;
1233         u32 snapshot;
1234         int ret, i;
1235
1236         if (op->flags & BCH_WRITE_MOVE)
1237                 return;
1238
1239         trans = bch2_trans_get(c);
1240 retry:
1241         bch2_trans_begin(trans);
1242
1243         ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1244         if (unlikely(ret))
1245                 goto err;
1246
1247         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1248                              SPOS(op->pos.inode, op->pos.offset, snapshot),
1249                              BTREE_ITER_SLOTS);
1250         while (1) {
1251                 struct bio *bio = &op->wbio.bio;
1252
1253                 nr_buckets = 0;
1254
1255                 k = bch2_btree_iter_peek_slot(&iter);
1256                 ret = bkey_err(k);
1257                 if (ret)
1258                         break;
1259
1260                 /* fall back to normal cow write path? */
1261                 if (unlikely(k.k->p.snapshot != snapshot ||
1262                              !bch2_extent_is_writeable(op, k)))
1263                         break;
1264
1265                 if (bch2_keylist_realloc(&op->insert_keys,
1266                                         op->inline_keys,
1267                                         ARRAY_SIZE(op->inline_keys),
1268                                         k.k->u64s))
1269                         break;
1270
1271                 /* Get iorefs before dropping btree locks: */
1272                 ptrs = bch2_bkey_ptrs_c(k);
1273                 bkey_for_each_ptr(ptrs, ptr) {
1274                         buckets[nr_buckets].b = PTR_BUCKET_POS(c, ptr);
1275                         buckets[nr_buckets].gen = ptr->gen;
1276                         buckets[nr_buckets].l =
1277                                 bucket_nocow_lock(&c->nocow_locks,
1278                                                   bucket_to_u64(buckets[nr_buckets].b));
1279
1280                         prefetch(buckets[nr_buckets].l);
1281
1282                         if (unlikely(!bch2_dev_get_ioref(bch_dev_bkey_exists(c, ptr->dev), WRITE)))
1283                                 goto err_get_ioref;
1284
1285                         nr_buckets++;
1286
1287                         if (ptr->unwritten)
1288                                 op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1289                 }
1290
1291                 /* Unlock before taking nocow locks, doing IO: */
1292                 bkey_reassemble(op->insert_keys.top, k);
1293                 bch2_trans_unlock(trans);
1294
1295                 bch2_cut_front(op->pos, op->insert_keys.top);
1296                 if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1297                         bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1298
1299                 for (i = 0; i < nr_buckets; i++) {
1300                         struct bch_dev *ca = bch_dev_bkey_exists(c, buckets[i].b.inode);
1301                         struct nocow_lock_bucket *l = buckets[i].l;
1302                         bool stale;
1303
1304                         __bch2_bucket_nocow_lock(&c->nocow_locks, l,
1305                                                  bucket_to_u64(buckets[i].b),
1306                                                  BUCKET_NOCOW_LOCK_UPDATE);
1307
1308                         rcu_read_lock();
1309                         stale = gen_after(*bucket_gen(ca, buckets[i].b.offset), buckets[i].gen);
1310                         rcu_read_unlock();
1311
1312                         if (unlikely(stale))
1313                                 goto err_bucket_stale;
1314                 }
1315
1316                 bio = &op->wbio.bio;
1317                 if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1318                         bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1319                                         GFP_KERNEL, &c->bio_write);
1320                         wbio_init(bio)->put_bio = true;
1321                         bio->bi_opf = op->wbio.bio.bi_opf;
1322                 } else {
1323                         op->flags |= BCH_WRITE_DONE;
1324                 }
1325
1326                 op->pos.offset += bio_sectors(bio);
1327                 op->written += bio_sectors(bio);
1328
1329                 bio->bi_end_io  = bch2_write_endio;
1330                 bio->bi_private = &op->cl;
1331                 bio->bi_opf |= REQ_OP_WRITE;
1332                 closure_get(&op->cl);
1333                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1334                                           op->insert_keys.top, true);
1335
1336                 bch2_keylist_push(&op->insert_keys);
1337                 if (op->flags & BCH_WRITE_DONE)
1338                         break;
1339                 bch2_btree_iter_advance(&iter);
1340         }
1341 out:
1342         bch2_trans_iter_exit(trans, &iter);
1343 err:
1344         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1345                 goto retry;
1346
1347         if (ret) {
1348                 bch_err_inum_offset_ratelimited(c,
1349                                 op->pos.inode,
1350                                 op->pos.offset << 9,
1351                                 "%s: btree lookup error %s",
1352                                 __func__, bch2_err_str(ret));
1353                 op->error = ret;
1354                 op->flags |= BCH_WRITE_DONE;
1355         }
1356
1357         bch2_trans_put(trans);
1358
1359         /* fallback to cow write path? */
1360         if (!(op->flags & BCH_WRITE_DONE)) {
1361                 closure_sync(&op->cl);
1362                 __bch2_nocow_write_done(op);
1363                 op->insert_keys.top = op->insert_keys.keys;
1364         } else if (op->flags & BCH_WRITE_SYNC) {
1365                 closure_sync(&op->cl);
1366                 bch2_nocow_write_done(&op->cl);
1367         } else {
1368                 /*
1369                  * XXX
1370                  * needs to run out of process context because ei_quota_lock is
1371                  * a mutex
1372                  */
1373                 continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1374         }
1375         return;
1376 err_get_ioref:
1377         for (i = 0; i < nr_buckets; i++)
1378                 percpu_ref_put(&bch_dev_bkey_exists(c, buckets[i].b.inode)->io_ref);
1379
1380         /* Fall back to COW path: */
1381         goto out;
1382 err_bucket_stale:
1383         while (i >= 0) {
1384                 bch2_bucket_nocow_unlock(&c->nocow_locks,
1385                                          buckets[i].b,
1386                                          BUCKET_NOCOW_LOCK_UPDATE);
1387                 --i;
1388         }
1389         for (i = 0; i < nr_buckets; i++)
1390                 percpu_ref_put(&bch_dev_bkey_exists(c, buckets[i].b.inode)->io_ref);
1391
1392         /* We can retry this: */
1393         ret = -BCH_ERR_transaction_restart;
1394         goto out;
1395 }
1396
1397 static void __bch2_write(struct bch_write_op *op)
1398 {
1399         struct bch_fs *c = op->c;
1400         struct write_point *wp = NULL;
1401         struct bio *bio = NULL;
1402         unsigned nofs_flags;
1403         int ret;
1404
1405         nofs_flags = memalloc_nofs_save();
1406
1407         if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1408                 bch2_nocow_write(op);
1409                 if (op->flags & BCH_WRITE_DONE)
1410                         goto out_nofs_restore;
1411         }
1412 again:
1413         memset(&op->failed, 0, sizeof(op->failed));
1414
1415         do {
1416                 struct bkey_i *key_to_write;
1417                 unsigned key_to_write_offset = op->insert_keys.top_p -
1418                         op->insert_keys.keys_p;
1419
1420                 /* +1 for possible cache device: */
1421                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1422                     ARRAY_SIZE(op->open_buckets.v))
1423                         break;
1424
1425                 if (bch2_keylist_realloc(&op->insert_keys,
1426                                         op->inline_keys,
1427                                         ARRAY_SIZE(op->inline_keys),
1428                                         BKEY_EXTENT_U64s_MAX))
1429                         break;
1430
1431                 /*
1432                  * The copygc thread is now global, which means it's no longer
1433                  * freeing up space on specific disks, which means that
1434                  * allocations for specific disks may hang arbitrarily long:
1435                  */
1436                 ret = bch2_trans_do(c, NULL, NULL, 0,
1437                         bch2_alloc_sectors_start_trans(trans,
1438                                 op->target,
1439                                 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1440                                 op->write_point,
1441                                 &op->devs_have,
1442                                 op->nr_replicas,
1443                                 op->nr_replicas_required,
1444                                 op->watermark,
1445                                 op->flags,
1446                                 (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1447                                               BCH_WRITE_ONLY_SPECIFIED_DEVS))
1448                                 ? NULL : &op->cl, &wp));
1449                 if (unlikely(ret)) {
1450                         if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1451                                 break;
1452
1453                         goto err;
1454                 }
1455
1456                 EBUG_ON(!wp);
1457
1458                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1459                 ret = bch2_write_extent(op, wp, &bio);
1460
1461                 bch2_alloc_sectors_done_inlined(c, wp);
1462 err:
1463                 if (ret <= 0) {
1464                         op->flags |= BCH_WRITE_DONE;
1465
1466                         if (ret < 0) {
1467                                 op->error = ret;
1468                                 break;
1469                         }
1470                 }
1471
1472                 bio->bi_end_io  = bch2_write_endio;
1473                 bio->bi_private = &op->cl;
1474                 bio->bi_opf |= REQ_OP_WRITE;
1475
1476                 closure_get(bio->bi_private);
1477
1478                 key_to_write = (void *) (op->insert_keys.keys_p +
1479                                          key_to_write_offset);
1480
1481                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1482                                           key_to_write, false);
1483         } while (ret);
1484
1485         /*
1486          * Sync or no?
1487          *
1488          * If we're running asynchronously, wne may still want to block
1489          * synchronously here if we weren't able to submit all of the IO at
1490          * once, as that signals backpressure to the caller.
1491          */
1492         if ((op->flags & BCH_WRITE_SYNC) ||
1493             (!(op->flags & BCH_WRITE_DONE) &&
1494              !(op->flags & BCH_WRITE_IN_WORKER))) {
1495                 closure_sync(&op->cl);
1496                 __bch2_write_index(op);
1497
1498                 if (!(op->flags & BCH_WRITE_DONE))
1499                         goto again;
1500                 bch2_write_done(&op->cl);
1501         } else {
1502                 bch2_write_queue(op, wp);
1503                 continue_at(&op->cl, bch2_write_index, NULL);
1504         }
1505 out_nofs_restore:
1506         memalloc_nofs_restore(nofs_flags);
1507 }
1508
1509 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1510 {
1511         struct bio *bio = &op->wbio.bio;
1512         struct bvec_iter iter;
1513         struct bkey_i_inline_data *id;
1514         unsigned sectors;
1515         int ret;
1516
1517         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1518         op->flags |= BCH_WRITE_DONE;
1519
1520         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1521
1522         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1523                                    ARRAY_SIZE(op->inline_keys),
1524                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1525         if (ret) {
1526                 op->error = ret;
1527                 goto err;
1528         }
1529
1530         sectors = bio_sectors(bio);
1531         op->pos.offset += sectors;
1532
1533         id = bkey_inline_data_init(op->insert_keys.top);
1534         id->k.p         = op->pos;
1535         id->k.version   = op->version;
1536         id->k.size      = sectors;
1537
1538         iter = bio->bi_iter;
1539         iter.bi_size = data_len;
1540         memcpy_from_bio(id->v.data, bio, iter);
1541
1542         while (data_len & 7)
1543                 id->v.data[data_len++] = '\0';
1544         set_bkey_val_bytes(&id->k, data_len);
1545         bch2_keylist_push(&op->insert_keys);
1546
1547         __bch2_write_index(op);
1548 err:
1549         bch2_write_done(&op->cl);
1550 }
1551
1552 /**
1553  * bch2_write() - handle a write to a cache device or flash only volume
1554  * @cl:         &bch_write_op->cl
1555  *
1556  * This is the starting point for any data to end up in a cache device; it could
1557  * be from a normal write, or a writeback write, or a write to a flash only
1558  * volume - it's also used by the moving garbage collector to compact data in
1559  * mostly empty buckets.
1560  *
1561  * It first writes the data to the cache, creating a list of keys to be inserted
1562  * (if the data won't fit in a single open bucket, there will be multiple keys);
1563  * after the data is written it calls bch_journal, and after the keys have been
1564  * added to the next journal write they're inserted into the btree.
1565  *
1566  * If op->discard is true, instead of inserting the data it invalidates the
1567  * region of the cache represented by op->bio and op->inode.
1568  */
1569 void bch2_write(struct closure *cl)
1570 {
1571         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1572         struct bio *bio = &op->wbio.bio;
1573         struct bch_fs *c = op->c;
1574         unsigned data_len;
1575
1576         EBUG_ON(op->cl.parent);
1577         BUG_ON(!op->nr_replicas);
1578         BUG_ON(!op->write_point.v);
1579         BUG_ON(bkey_eq(op->pos, POS_MAX));
1580
1581         op->start_time = local_clock();
1582         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1583         wbio_init(bio)->put_bio = false;
1584
1585         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1586                 bch_err_inum_offset_ratelimited(c,
1587                         op->pos.inode,
1588                         op->pos.offset << 9,
1589                         "misaligned write");
1590                 op->error = -EIO;
1591                 goto err;
1592         }
1593
1594         if (c->opts.nochanges) {
1595                 op->error = -BCH_ERR_erofs_no_writes;
1596                 goto err;
1597         }
1598
1599         if (!(op->flags & BCH_WRITE_MOVE) &&
1600             !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1601                 op->error = -BCH_ERR_erofs_no_writes;
1602                 goto err;
1603         }
1604
1605         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1606         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1607
1608         data_len = min_t(u64, bio->bi_iter.bi_size,
1609                          op->new_i_size - (op->pos.offset << 9));
1610
1611         if (c->opts.inline_data &&
1612             data_len <= min(block_bytes(c) / 2, 1024U)) {
1613                 bch2_write_data_inline(op, data_len);
1614                 return;
1615         }
1616
1617         __bch2_write(op);
1618         return;
1619 err:
1620         bch2_disk_reservation_put(c, &op->res);
1621
1622         closure_debug_destroy(&op->cl);
1623         if (op->end_io)
1624                 op->end_io(op);
1625 }
1626
1627 static const char * const bch2_write_flags[] = {
1628 #define x(f)    #f,
1629         BCH_WRITE_FLAGS()
1630 #undef x
1631         NULL
1632 };
1633
1634 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1635 {
1636         prt_str(out, "pos: ");
1637         bch2_bpos_to_text(out, op->pos);
1638         prt_newline(out);
1639         printbuf_indent_add(out, 2);
1640
1641         prt_str(out, "started: ");
1642         bch2_pr_time_units(out, local_clock() - op->start_time);
1643         prt_newline(out);
1644
1645         prt_str(out, "flags: ");
1646         prt_bitflags(out, bch2_write_flags, op->flags);
1647         prt_newline(out);
1648
1649         prt_printf(out, "ref: %u", closure_nr_remaining(&op->cl));
1650         prt_newline(out);
1651
1652         printbuf_indent_sub(out, 2);
1653 }
1654
1655 void bch2_fs_io_write_exit(struct bch_fs *c)
1656 {
1657         mempool_exit(&c->bio_bounce_pages);
1658         bioset_exit(&c->bio_write);
1659 }
1660
1661 int bch2_fs_io_write_init(struct bch_fs *c)
1662 {
1663         if (bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
1664                         BIOSET_NEED_BVECS))
1665                 return -BCH_ERR_ENOMEM_bio_write_init;
1666
1667         if (mempool_init_page_pool(&c->bio_bounce_pages,
1668                                    max_t(unsigned,
1669                                          c->opts.btree_node_size,
1670                                          c->opts.encoded_extent_max) /
1671                                    PAGE_SIZE, 0))
1672                 return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1673
1674         return 0;
1675 }