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;
  73
  74         old = atomic64_read(latency);
  75         do {
  76                 /*
  77                  * If the io latency was reasonably close to the current
  78                  * latency, skip doing the update and atomic operation - most of
  79                  * the time:
  80                  */
  81                 if (abs((int) (old - io_latency)) < (old >> 1) &&
  82                     now & ~(~0U << 5))
  83                         break;
  84
  85                 new = ewma_add(old, io_latency, 5);
  86         } while (!atomic64_try_cmpxchg(latency, &old, new));
  87
  88         bch2_congested_acct(ca, io_latency, now, rw);
  89
  90         __bch2_time_stats_update(&ca->io_latency[rw].stats, submit_time, now);
  91 }
  92
  93 #endif
  94
  95 /* Allocate, free from mempool: */
  96
  97 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
  98 {
  99         struct bvec_iter_all iter;
 100         struct bio_vec *bv;
 101
 102         bio_for_each_segment_all(bv, bio, iter)
 103                 if (bv->bv_page != ZERO_PAGE(0))
 104                         mempool_free(bv->bv_page, &c->bio_bounce_pages);
 105         bio->bi_vcnt = 0;
 106 }
 107
 108 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
 109 {
 110         struct page *page;
 111
 112         if (likely(!*using_mempool)) {
 113                 page = alloc_page(GFP_NOFS);
 114                 if (unlikely(!page)) {
 115                         mutex_lock(&c->bio_bounce_pages_lock);
 116                         *using_mempool = true;
 117                         goto pool_alloc;
 118
 119                 }
 120         } else {
 121 pool_alloc:
 122                 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOFS);
 123         }
 124
 125         return page;
 126 }
 127
 128 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
 129                                size_t size)
 130 {
 131         bool using_mempool = false;
 132
 133         while (size) {
 134                 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
 135                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 136
 137                 BUG_ON(!bio_add_page(bio, page, len, 0));
 138                 size -= len;
 139         }
 140
 141         if (using_mempool)
 142                 mutex_unlock(&c->bio_bounce_pages_lock);
 143 }
 144
 145 /* Extent update path: */
 146
 147 int bch2_sum_sector_overwrites(struct btree_trans *trans,
 148                                struct btree_iter *extent_iter,
 149                                struct bkey_i *new,
 150                                bool *usage_increasing,
 151                                s64 *i_sectors_delta,
 152                                s64 *disk_sectors_delta)
 153 {
 154         struct bch_fs *c = trans->c;
 155         struct btree_iter iter;
 156         struct bkey_s_c old;
 157         unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
 158         bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
 159         int ret = 0;
 160
 161         *usage_increasing       = false;
 162         *i_sectors_delta        = 0;
 163         *disk_sectors_delta     = 0;
 164
 165         bch2_trans_copy_iter(&iter, extent_iter);
 166
 167         for_each_btree_key_max_continue_norestart(iter,
 168                                 new->k.p, BTREE_ITER_slots, old, ret) {
 169                 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 170                         max(bkey_start_offset(&new->k),
 171                             bkey_start_offset(old.k));
 172
 173                 *i_sectors_delta += sectors *
 174                         (bkey_extent_is_allocation(&new->k) -
 175                          bkey_extent_is_allocation(old.k));
 176
 177                 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
 178                 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
 179                         ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 180                         : 0;
 181
 182                 if (!*usage_increasing &&
 183                     (new->k.p.snapshot != old.k->p.snapshot ||
 184                      new_replicas > bch2_bkey_replicas(c, old) ||
 185                      (!new_compressed && bch2_bkey_sectors_compressed(old))))
 186                         *usage_increasing = true;
 187
 188                 if (bkey_ge(old.k->p, new->k.p))
 189                         break;
 190         }
 191
 192         bch2_trans_iter_exit(trans, &iter);
 193         return ret;
 194 }
 195
 196 static inline int bch2_extent_update_i_size_sectors(struct btree_trans *trans,
 197                                                     struct btree_iter *extent_iter,
 198                                                     u64 new_i_size,
 199                                                     s64 i_sectors_delta)
 200 {
 201         /*
 202          * Crazy performance optimization:
 203          * Every extent update needs to also update the inode: the inode trigger
 204          * will set bi->journal_seq to the journal sequence number of this
 205          * transaction - for fsync.
 206          *
 207          * But if that's the only reason we're updating the inode (we're not
 208          * updating bi_size or bi_sectors), then we don't need the inode update
 209          * to be journalled - if we crash, the bi_journal_seq update will be
 210          * lost, but that's fine.
 211          */
 212         unsigned inode_update_flags = BTREE_UPDATE_nojournal;
 213
 214         struct btree_iter iter;
 215         struct bkey_s_c k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_inodes,
 216                               SPOS(0,
 217                                    extent_iter->pos.inode,
 218                                    extent_iter->snapshot),
 219                               BTREE_ITER_intent|
 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, &op->opts, sk.k) ?:
 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 static void __bch2_write_op_error(struct printbuf *out, struct bch_write_op *op,
 400                                   u64 offset)
 401 {
 402         bch2_inum_offset_err_msg(op->c, out,
 403                                  (subvol_inum) { op->subvol, op->pos.inode, },
 404                                  offset << 9);
 405         prt_printf(out, "write error%s: ",
 406                    op->flags & BCH_WRITE_MOVE ? "(internal move)" : "");
 407 }
 408
 409 void bch2_write_op_error(struct printbuf *out, struct bch_write_op *op)
 410 {
 411         __bch2_write_op_error(out, op, op->pos.offset);
 412 }
 413
 414 static void bch2_write_op_error_trans(struct btree_trans *trans, struct printbuf *out,
 415                                       struct bch_write_op *op, u64 offset)
 416 {
 417         bch2_inum_offset_err_msg_trans(trans, out,
 418                                        (subvol_inum) { op->subvol, op->pos.inode, },
 419                                        offset << 9);
 420         prt_printf(out, "write error%s: ",
 421                    op->flags & BCH_WRITE_MOVE ? "(internal move)" : "");
 422 }
 423
 424 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 425                                enum bch_data_type type,
 426                                const struct bkey_i *k,
 427                                bool nocow)
 428 {
 429         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 430         struct bch_write_bio *n;
 431
 432         BUG_ON(c->opts.nochanges);
 433
 434         bkey_for_each_ptr(ptrs, ptr) {
 435                 struct bch_dev *ca = nocow
 436                         ? bch2_dev_have_ref(c, ptr->dev)
 437                         : bch2_dev_get_ioref(c, ptr->dev, type == BCH_DATA_btree ? READ : WRITE);
 438
 439                 if (to_entry(ptr + 1) < ptrs.end) {
 440                         n = to_wbio(bio_alloc_clone(NULL, &wbio->bio, GFP_NOFS, &c->replica_set));
 441
 442                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 443                         n->bio.bi_private       = wbio->bio.bi_private;
 444                         n->parent               = wbio;
 445                         n->split                = true;
 446                         n->bounce               = false;
 447                         n->put_bio              = true;
 448                         n->bio.bi_opf           = wbio->bio.bi_opf;
 449                         bio_inc_remaining(&wbio->bio);
 450                 } else {
 451                         n = wbio;
 452                         n->split                = false;
 453                 }
 454
 455                 n->c                    = c;
 456                 n->dev                  = ptr->dev;
 457                 n->have_ioref           = ca != NULL;
 458                 n->nocow                = nocow;
 459                 n->submit_time          = local_clock();
 460                 n->inode_offset         = bkey_start_offset(&k->k);
 461                 if (nocow)
 462                         n->nocow_bucket = PTR_BUCKET_NR(ca, ptr);
 463                 n->bio.bi_iter.bi_sector = ptr->offset;
 464
 465                 if (likely(n->have_ioref)) {
 466                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 467                                      bio_sectors(&n->bio));
 468
 469                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 470
 471                         if (type != BCH_DATA_btree && unlikely(c->opts.no_data_io)) {
 472                                 bio_endio(&n->bio);
 473                                 continue;
 474                         }
 475
 476                         submit_bio(&n->bio);
 477                 } else {
 478                         n->bio.bi_status        = BLK_STS_REMOVED;
 479                         bio_endio(&n->bio);
 480                 }
 481         }
 482 }
 483
 484 static void __bch2_write(struct bch_write_op *);
 485
 486 static void bch2_write_done(struct closure *cl)
 487 {
 488         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 489         struct bch_fs *c = op->c;
 490
 491         EBUG_ON(op->open_buckets.nr);
 492
 493         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 494         bch2_disk_reservation_put(c, &op->res);
 495
 496         if (!(op->flags & BCH_WRITE_MOVE))
 497                 bch2_write_ref_put(c, BCH_WRITE_REF_write);
 498         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 499
 500         EBUG_ON(cl->parent);
 501         closure_debug_destroy(cl);
 502         if (op->end_io)
 503                 op->end_io(op);
 504 }
 505
 506 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
 507 {
 508         struct keylist *keys = &op->insert_keys;
 509         struct bkey_i *src, *dst = keys->keys, *n;
 510
 511         for (src = keys->keys; src != keys->top; src = n) {
 512                 n = bkey_next(src);
 513
 514                 if (bkey_extent_is_direct_data(&src->k)) {
 515                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 516                                             test_bit(ptr->dev, op->failed.d));
 517
 518                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
 519                                 return -EIO;
 520                 }
 521
 522                 if (dst != src)
 523                         memmove_u64s_down(dst, src, src->k.u64s);
 524                 dst = bkey_next(dst);
 525         }
 526
 527         keys->top = dst;
 528         return 0;
 529 }
 530
 531 /**
 532  * __bch2_write_index - after a write, update index to point to new data
 533  * @op:         bch_write_op to process
 534  */
 535 static void __bch2_write_index(struct bch_write_op *op)
 536 {
 537         struct bch_fs *c = op->c;
 538         struct keylist *keys = &op->insert_keys;
 539         unsigned dev;
 540         int ret = 0;
 541
 542         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
 543                 ret = bch2_write_drop_io_error_ptrs(op);
 544                 if (ret)
 545                         goto err;
 546         }
 547
 548         if (!bch2_keylist_empty(keys)) {
 549                 u64 sectors_start = keylist_sectors(keys);
 550
 551                 ret = !(op->flags & BCH_WRITE_MOVE)
 552                         ? bch2_write_index_default(op)
 553                         : bch2_data_update_index_update(op);
 554
 555                 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 556                 BUG_ON(keylist_sectors(keys) && !ret);
 557
 558                 op->written += sectors_start - keylist_sectors(keys);
 559
 560                 if (unlikely(ret && !bch2_err_matches(ret, EROFS))) {
 561                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
 562
 563                         struct printbuf buf = PRINTBUF;
 564                         __bch2_write_op_error(&buf, op, bkey_start_offset(&insert->k));
 565                         prt_printf(&buf, "btree update error: %s", bch2_err_str(ret));
 566                         bch_err_ratelimited(c, "%s", buf.buf);
 567                         printbuf_exit(&buf);
 568                 }
 569
 570                 if (ret)
 571                         goto err;
 572         }
 573 out:
 574         /* If some a bucket wasn't written, we can't erasure code it: */
 575         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 576                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 577
 578         bch2_open_buckets_put(c, &op->open_buckets);
 579         return;
 580 err:
 581         keys->top = keys->keys;
 582         op->error = ret;
 583         op->flags |= BCH_WRITE_SUBMITTED;
 584         goto out;
 585 }
 586
 587 static inline void __wp_update_state(struct write_point *wp, enum write_point_state state)
 588 {
 589         if (state != wp->state) {
 590                 u64 now = ktime_get_ns();
 591
 592                 if (wp->last_state_change &&
 593                     time_after64(now, wp->last_state_change))
 594                         wp->time[wp->state] += now - wp->last_state_change;
 595                 wp->state = state;
 596                 wp->last_state_change = now;
 597         }
 598 }
 599
 600 static inline void wp_update_state(struct write_point *wp, bool running)
 601 {
 602         enum write_point_state state;
 603
 604         state = running                  ? WRITE_POINT_running :
 605                 !list_empty(&wp->writes) ? WRITE_POINT_waiting_io
 606                                          : WRITE_POINT_stopped;
 607
 608         __wp_update_state(wp, state);
 609 }
 610
 611 static CLOSURE_CALLBACK(bch2_write_index)
 612 {
 613         closure_type(op, struct bch_write_op, cl);
 614         struct write_point *wp = op->wp;
 615         struct workqueue_struct *wq = index_update_wq(op);
 616         unsigned long flags;
 617
 618         if ((op->flags & BCH_WRITE_SUBMITTED) &&
 619             (op->flags & BCH_WRITE_MOVE))
 620                 bch2_bio_free_pages_pool(op->c, &op->wbio.bio);
 621
 622         spin_lock_irqsave(&wp->writes_lock, flags);
 623         if (wp->state == WRITE_POINT_waiting_io)
 624                 __wp_update_state(wp, WRITE_POINT_waiting_work);
 625         list_add_tail(&op->wp_list, &wp->writes);
 626         spin_unlock_irqrestore (&wp->writes_lock, flags);
 627
 628         queue_work(wq, &wp->index_update_work);
 629 }
 630
 631 static inline void bch2_write_queue(struct bch_write_op *op, struct write_point *wp)
 632 {
 633         op->wp = wp;
 634
 635         if (wp->state == WRITE_POINT_stopped) {
 636                 spin_lock_irq(&wp->writes_lock);
 637                 __wp_update_state(wp, WRITE_POINT_waiting_io);
 638                 spin_unlock_irq(&wp->writes_lock);
 639         }
 640 }
 641
 642 void bch2_write_point_do_index_updates(struct work_struct *work)
 643 {
 644         struct write_point *wp =
 645                 container_of(work, struct write_point, index_update_work);
 646         struct bch_write_op *op;
 647
 648         while (1) {
 649                 spin_lock_irq(&wp->writes_lock);
 650                 op = list_pop_entry(&wp->writes, struct bch_write_op, wp_list);
 651                 wp_update_state(wp, op != NULL);
 652                 spin_unlock_irq(&wp->writes_lock);
 653
 654                 if (!op)
 655                         break;
 656
 657                 op->flags |= BCH_WRITE_IN_WORKER;
 658
 659                 __bch2_write_index(op);
 660
 661                 if (!(op->flags & BCH_WRITE_SUBMITTED))
 662                         __bch2_write(op);
 663                 else
 664                         bch2_write_done(&op->cl);
 665         }
 666 }
 667
 668 static void bch2_write_endio(struct bio *bio)
 669 {
 670         struct closure *cl              = bio->bi_private;
 671         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 672         struct bch_write_bio *wbio      = to_wbio(bio);
 673         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 674         struct bch_fs *c                = wbio->c;
 675         struct bch_dev *ca              = wbio->have_ioref
 676                 ? bch2_dev_have_ref(c, wbio->dev)
 677                 : NULL;
 678
 679         if (bch2_dev_inum_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
 680                                     op->pos.inode,
 681                                     wbio->inode_offset << 9,
 682                                     "data write error: %s",
 683                                     bch2_blk_status_to_str(bio->bi_status))) {
 684                 set_bit(wbio->dev, op->failed.d);
 685                 op->flags |= BCH_WRITE_IO_ERROR;
 686         }
 687
 688         if (wbio->nocow) {
 689                 bch2_bucket_nocow_unlock(&c->nocow_locks,
 690                                          POS(ca->dev_idx, wbio->nocow_bucket),
 691                                          BUCKET_NOCOW_LOCK_UPDATE);
 692                 set_bit(wbio->dev, op->devs_need_flush->d);
 693         }
 694
 695         if (wbio->have_ioref) {
 696                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 697                 percpu_ref_put(&ca->io_ref);
 698         }
 699
 700         if (wbio->bounce)
 701                 bch2_bio_free_pages_pool(c, bio);
 702
 703         if (wbio->put_bio)
 704                 bio_put(bio);
 705
 706         if (parent)
 707                 bio_endio(&parent->bio);
 708         else
 709                 closure_put(cl);
 710 }
 711
 712 static void init_append_extent(struct bch_write_op *op,
 713                                struct write_point *wp,
 714                                struct bversion version,
 715                                struct bch_extent_crc_unpacked crc)
 716 {
 717         struct bkey_i_extent *e;
 718
 719         op->pos.offset += crc.uncompressed_size;
 720
 721         e = bkey_extent_init(op->insert_keys.top);
 722         e->k.p          = op->pos;
 723         e->k.size       = crc.uncompressed_size;
 724         e->k.bversion   = version;
 725
 726         if (crc.csum_type ||
 727             crc.compression_type ||
 728             crc.nonce)
 729                 bch2_extent_crc_append(&e->k_i, crc);
 730
 731         bch2_alloc_sectors_append_ptrs_inlined(op->c, wp, &e->k_i, crc.compressed_size,
 732                                        op->flags & BCH_WRITE_CACHED);
 733
 734         bch2_keylist_push(&op->insert_keys);
 735 }
 736
 737 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 738                                         struct write_point *wp,
 739                                         struct bio *src,
 740                                         bool *page_alloc_failed,
 741                                         void *buf)
 742 {
 743         struct bch_write_bio *wbio;
 744         struct bio *bio;
 745         unsigned output_available =
 746                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 747         unsigned pages = DIV_ROUND_UP(output_available +
 748                                       (buf
 749                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 750                                        : 0), PAGE_SIZE);
 751
 752         pages = min(pages, BIO_MAX_VECS);
 753
 754         bio = bio_alloc_bioset(NULL, pages, 0,
 755                                GFP_NOFS, &c->bio_write);
 756         wbio                    = wbio_init(bio);
 757         wbio->put_bio           = true;
 758         /* copy WRITE_SYNC flag */
 759         wbio->bio.bi_opf        = src->bi_opf;
 760
 761         if (buf) {
 762                 bch2_bio_map(bio, buf, output_available);
 763                 return bio;
 764         }
 765
 766         wbio->bounce            = true;
 767
 768         /*
 769          * We can't use mempool for more than c->sb.encoded_extent_max
 770          * worth of pages, but we'd like to allocate more if we can:
 771          */
 772         bch2_bio_alloc_pages_pool(c, bio,
 773                                   min_t(unsigned, output_available,
 774                                         c->opts.encoded_extent_max));
 775
 776         if (bio->bi_iter.bi_size < output_available)
 777                 *page_alloc_failed =
 778                         bch2_bio_alloc_pages(bio,
 779                                              output_available -
 780                                              bio->bi_iter.bi_size,
 781                                              GFP_NOFS) != 0;
 782
 783         return bio;
 784 }
 785
 786 static int bch2_write_rechecksum(struct bch_fs *c,
 787                                  struct bch_write_op *op,
 788                                  unsigned new_csum_type)
 789 {
 790         struct bio *bio = &op->wbio.bio;
 791         struct bch_extent_crc_unpacked new_crc;
 792         int ret;
 793
 794         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 795
 796         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 797             bch2_csum_type_is_encryption(new_csum_type))
 798                 new_csum_type = op->crc.csum_type;
 799
 800         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 801                                   NULL, &new_crc,
 802                                   op->crc.offset, op->crc.live_size,
 803                                   new_csum_type);
 804         if (ret)
 805                 return ret;
 806
 807         bio_advance(bio, op->crc.offset << 9);
 808         bio->bi_iter.bi_size = op->crc.live_size << 9;
 809         op->crc = new_crc;
 810         return 0;
 811 }
 812
 813 static int bch2_write_decrypt(struct bch_write_op *op)
 814 {
 815         struct bch_fs *c = op->c;
 816         struct nonce nonce = extent_nonce(op->version, op->crc);
 817         struct bch_csum csum;
 818         int ret;
 819
 820         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 821                 return 0;
 822
 823         /*
 824          * If we need to decrypt data in the write path, we'll no longer be able
 825          * to verify the existing checksum (poly1305 mac, in this case) after
 826          * it's decrypted - this is the last point we'll be able to reverify the
 827          * checksum:
 828          */
 829         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 830         if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 831                 return -EIO;
 832
 833         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 834         op->crc.csum_type = 0;
 835         op->crc.csum = (struct bch_csum) { 0, 0 };
 836         return ret;
 837 }
 838
 839 static enum prep_encoded_ret {
 840         PREP_ENCODED_OK,
 841         PREP_ENCODED_ERR,
 842         PREP_ENCODED_CHECKSUM_ERR,
 843         PREP_ENCODED_DO_WRITE,
 844 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 845 {
 846         struct bch_fs *c = op->c;
 847         struct bio *bio = &op->wbio.bio;
 848
 849         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 850                 return PREP_ENCODED_OK;
 851
 852         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 853
 854         /* Can we just write the entire extent as is? */
 855         if (op->crc.uncompressed_size == op->crc.live_size &&
 856             op->crc.uncompressed_size <= c->opts.encoded_extent_max >> 9 &&
 857             op->crc.compressed_size <= wp->sectors_free &&
 858             (op->crc.compression_type == bch2_compression_opt_to_type(op->compression_opt) ||
 859              op->incompressible)) {
 860                 if (!crc_is_compressed(op->crc) &&
 861                     op->csum_type != op->crc.csum_type &&
 862                     bch2_write_rechecksum(c, op, op->csum_type) &&
 863                     !c->opts.no_data_io)
 864                         return PREP_ENCODED_CHECKSUM_ERR;
 865
 866                 return PREP_ENCODED_DO_WRITE;
 867         }
 868
 869         /*
 870          * If the data is compressed and we couldn't write the entire extent as
 871          * is, we have to decompress it:
 872          */
 873         if (crc_is_compressed(op->crc)) {
 874                 struct bch_csum csum;
 875
 876                 if (bch2_write_decrypt(op))
 877                         return PREP_ENCODED_CHECKSUM_ERR;
 878
 879                 /* Last point we can still verify checksum: */
 880                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 881                                          extent_nonce(op->version, op->crc),
 882                                          bio);
 883                 if (bch2_crc_cmp(op->crc.csum, csum) && !c->opts.no_data_io)
 884                         return PREP_ENCODED_CHECKSUM_ERR;
 885
 886                 if (bch2_bio_uncompress_inplace(op, bio))
 887                         return PREP_ENCODED_ERR;
 888         }
 889
 890         /*
 891          * No longer have compressed data after this point - data might be
 892          * encrypted:
 893          */
 894
 895         /*
 896          * If the data is checksummed and we're only writing a subset,
 897          * rechecksum and adjust bio to point to currently live data:
 898          */
 899         if ((op->crc.live_size != op->crc.uncompressed_size ||
 900              op->crc.csum_type != op->csum_type) &&
 901             bch2_write_rechecksum(c, op, op->csum_type) &&
 902             !c->opts.no_data_io)
 903                 return PREP_ENCODED_CHECKSUM_ERR;
 904
 905         /*
 906          * If we want to compress the data, it has to be decrypted:
 907          */
 908         if ((op->compression_opt ||
 909              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 910              bch2_csum_type_is_encryption(op->csum_type)) &&
 911             bch2_write_decrypt(op))
 912                 return PREP_ENCODED_CHECKSUM_ERR;
 913
 914         return PREP_ENCODED_OK;
 915 }
 916
 917 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 918                              struct bio **_dst)
 919 {
 920         struct bch_fs *c = op->c;
 921         struct bio *src = &op->wbio.bio, *dst = src;
 922         struct bvec_iter saved_iter;
 923         void *ec_buf;
 924         unsigned total_output = 0, total_input = 0;
 925         bool bounce = false;
 926         bool page_alloc_failed = false;
 927         int ret, more = 0;
 928
 929         BUG_ON(!bio_sectors(src));
 930
 931         ec_buf = bch2_writepoint_ec_buf(c, wp);
 932
 933         switch (bch2_write_prep_encoded_data(op, wp)) {
 934         case PREP_ENCODED_OK:
 935                 break;
 936         case PREP_ENCODED_ERR:
 937                 ret = -EIO;
 938                 goto err;
 939         case PREP_ENCODED_CHECKSUM_ERR:
 940                 goto csum_err;
 941         case PREP_ENCODED_DO_WRITE:
 942                 /* XXX look for bug here */
 943                 if (ec_buf) {
 944                         dst = bch2_write_bio_alloc(c, wp, src,
 945                                                    &page_alloc_failed,
 946                                                    ec_buf);
 947                         bio_copy_data(dst, src);
 948                         bounce = true;
 949                 }
 950                 init_append_extent(op, wp, op->version, op->crc);
 951                 goto do_write;
 952         }
 953
 954         if (ec_buf ||
 955             op->compression_opt ||
 956             (op->csum_type &&
 957              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 958             (bch2_csum_type_is_encryption(op->csum_type) &&
 959              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 960                 dst = bch2_write_bio_alloc(c, wp, src,
 961                                            &page_alloc_failed,
 962                                            ec_buf);
 963                 bounce = true;
 964         }
 965
 966         saved_iter = dst->bi_iter;
 967
 968         do {
 969                 struct bch_extent_crc_unpacked crc = { 0 };
 970                 struct bversion version = op->version;
 971                 size_t dst_len = 0, src_len = 0;
 972
 973                 if (page_alloc_failed &&
 974                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 975                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 976                         break;
 977
 978                 BUG_ON(op->compression_opt &&
 979                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 980                        bch2_csum_type_is_encryption(op->crc.csum_type));
 981                 BUG_ON(op->compression_opt && !bounce);
 982
 983                 crc.compression_type = op->incompressible
 984                         ? BCH_COMPRESSION_TYPE_incompressible
 985                         : op->compression_opt
 986                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 987                                             op->compression_opt)
 988                         : 0;
 989                 if (!crc_is_compressed(crc)) {
 990                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 991                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 992
 993                         if (op->csum_type)
 994                                 dst_len = min_t(unsigned, dst_len,
 995                                                 c->opts.encoded_extent_max);
 996
 997                         if (bounce) {
 998                                 swap(dst->bi_iter.bi_size, dst_len);
 999                                 bio_copy_data(dst, src);
1000                                 swap(dst->bi_iter.bi_size, dst_len);
1001                         }
1002
1003                         src_len = dst_len;
1004                 }
1005
1006                 BUG_ON(!src_len || !dst_len);
1007
1008                 if (bch2_csum_type_is_encryption(op->csum_type)) {
1009                         if (bversion_zero(version)) {
1010                                 version.lo = atomic64_inc_return(&c->key_version);
1011                         } else {
1012                                 crc.nonce = op->nonce;
1013                                 op->nonce += src_len >> 9;
1014                         }
1015                 }
1016
1017                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1018                     !crc_is_compressed(crc) &&
1019                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
1020                     bch2_csum_type_is_encryption(op->csum_type)) {
1021                         u8 compression_type = crc.compression_type;
1022                         u16 nonce = crc.nonce;
1023                         /*
1024                          * Note: when we're using rechecksum(), we need to be
1025                          * checksumming @src because it has all the data our
1026                          * existing checksum covers - if we bounced (because we
1027                          * were trying to compress), @dst will only have the
1028                          * part of the data the new checksum will cover.
1029                          *
1030                          * But normally we want to be checksumming post bounce,
1031                          * because part of the reason for bouncing is so the
1032                          * data can't be modified (by userspace) while it's in
1033                          * flight.
1034                          */
1035                         if (bch2_rechecksum_bio(c, src, version, op->crc,
1036                                         &crc, &op->crc,
1037                                         src_len >> 9,
1038                                         bio_sectors(src) - (src_len >> 9),
1039                                         op->csum_type))
1040                                 goto csum_err;
1041                         /*
1042                          * rchecksum_bio sets compression_type on crc from op->crc,
1043                          * this isn't always correct as sometimes we're changing
1044                          * an extent from uncompressed to incompressible.
1045                          */
1046                         crc.compression_type = compression_type;
1047                         crc.nonce = nonce;
1048                 } else {
1049                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1050                             bch2_rechecksum_bio(c, src, version, op->crc,
1051                                         NULL, &op->crc,
1052                                         src_len >> 9,
1053                                         bio_sectors(src) - (src_len >> 9),
1054                                         op->crc.csum_type))
1055                                 goto csum_err;
1056
1057                         crc.compressed_size     = dst_len >> 9;
1058                         crc.uncompressed_size   = src_len >> 9;
1059                         crc.live_size           = src_len >> 9;
1060
1061                         swap(dst->bi_iter.bi_size, dst_len);
1062                         ret = bch2_encrypt_bio(c, op->csum_type,
1063                                                extent_nonce(version, crc), dst);
1064                         if (ret)
1065                                 goto err;
1066
1067                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1068                                          extent_nonce(version, crc), dst);
1069                         crc.csum_type = op->csum_type;
1070                         swap(dst->bi_iter.bi_size, dst_len);
1071                 }
1072
1073                 init_append_extent(op, wp, version, crc);
1074
1075                 if (dst != src)
1076                         bio_advance(dst, dst_len);
1077                 bio_advance(src, src_len);
1078                 total_output    += dst_len;
1079                 total_input     += src_len;
1080         } while (dst->bi_iter.bi_size &&
1081                  src->bi_iter.bi_size &&
1082                  wp->sectors_free &&
1083                  !bch2_keylist_realloc(&op->insert_keys,
1084                                       op->inline_keys,
1085                                       ARRAY_SIZE(op->inline_keys),
1086                                       BKEY_EXTENT_U64s_MAX));
1087
1088         more = src->bi_iter.bi_size != 0;
1089
1090         dst->bi_iter = saved_iter;
1091
1092         if (dst == src && more) {
1093                 BUG_ON(total_output != total_input);
1094
1095                 dst = bio_split(src, total_input >> 9,
1096                                 GFP_NOFS, &c->bio_write);
1097                 wbio_init(dst)->put_bio = true;
1098                 /* copy WRITE_SYNC flag */
1099                 dst->bi_opf             = src->bi_opf;
1100         }
1101
1102         dst->bi_iter.bi_size = total_output;
1103 do_write:
1104         *_dst = dst;
1105         return more;
1106 csum_err:
1107         {
1108                 struct printbuf buf = PRINTBUF;
1109                 bch2_write_op_error(&buf, op);
1110                 prt_printf(&buf, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1111                 bch_err_ratelimited(c, "%s", buf.buf);
1112                 printbuf_exit(&buf);
1113         }
1114
1115         ret = -EIO;
1116 err:
1117         if (to_wbio(dst)->bounce)
1118                 bch2_bio_free_pages_pool(c, dst);
1119         if (to_wbio(dst)->put_bio)
1120                 bio_put(dst);
1121
1122         return ret;
1123 }
1124
1125 static bool bch2_extent_is_writeable(struct bch_write_op *op,
1126                                      struct bkey_s_c k)
1127 {
1128         struct bch_fs *c = op->c;
1129         struct bkey_s_c_extent e;
1130         struct extent_ptr_decoded p;
1131         const union bch_extent_entry *entry;
1132         unsigned replicas = 0;
1133
1134         if (k.k->type != KEY_TYPE_extent)
1135                 return false;
1136
1137         e = bkey_s_c_to_extent(k);
1138
1139         rcu_read_lock();
1140         extent_for_each_ptr_decode(e, p, entry) {
1141                 if (crc_is_encoded(p.crc) || p.has_ec) {
1142                         rcu_read_unlock();
1143                         return false;
1144                 }
1145
1146                 replicas += bch2_extent_ptr_durability(c, &p);
1147         }
1148         rcu_read_unlock();
1149
1150         return replicas >= op->opts.data_replicas;
1151 }
1152
1153 static int bch2_nocow_write_convert_one_unwritten(struct btree_trans *trans,
1154                                                   struct btree_iter *iter,
1155                                                   struct bkey_i *orig,
1156                                                   struct bkey_s_c k,
1157                                                   u64 new_i_size)
1158 {
1159         if (!bch2_extents_match(bkey_i_to_s_c(orig), k)) {
1160                 /* trace this */
1161                 return 0;
1162         }
1163
1164         struct bkey_i *new = bch2_bkey_make_mut_noupdate(trans, k);
1165         int ret = PTR_ERR_OR_ZERO(new);
1166         if (ret)
1167                 return ret;
1168
1169         bch2_cut_front(bkey_start_pos(&orig->k), new);
1170         bch2_cut_back(orig->k.p, new);
1171
1172         struct bkey_ptrs ptrs = bch2_bkey_ptrs(bkey_i_to_s(new));
1173         bkey_for_each_ptr(ptrs, ptr)
1174                 ptr->unwritten = 0;
1175
1176         /*
1177          * Note that we're not calling bch2_subvol_get_snapshot() in this path -
1178          * that was done when we kicked off the write, and here it's important
1179          * that we update the extent that we wrote to - even if a snapshot has
1180          * since been created. The write is still outstanding, so we're ok
1181          * w.r.t. snapshot atomicity:
1182          */
1183         return  bch2_extent_update_i_size_sectors(trans, iter,
1184                                         min(new->k.p.offset << 9, new_i_size), 0) ?:
1185                 bch2_trans_update(trans, iter, new,
1186                                   BTREE_UPDATE_internal_snapshot_node);
1187 }
1188
1189 static void bch2_nocow_write_convert_unwritten(struct bch_write_op *op)
1190 {
1191         struct bch_fs *c = op->c;
1192         struct btree_trans *trans = bch2_trans_get(c);
1193
1194         for_each_keylist_key(&op->insert_keys, orig) {
1195                 int ret = for_each_btree_key_max_commit(trans, iter, BTREE_ID_extents,
1196                                      bkey_start_pos(&orig->k), orig->k.p,
1197                                      BTREE_ITER_intent, k,
1198                                      NULL, NULL, BCH_TRANS_COMMIT_no_enospc, ({
1199                         bch2_nocow_write_convert_one_unwritten(trans, &iter, orig, k, op->new_i_size);
1200                 }));
1201
1202                 if (ret && !bch2_err_matches(ret, EROFS)) {
1203                         struct bkey_i *insert = bch2_keylist_front(&op->insert_keys);
1204
1205                         struct printbuf buf = PRINTBUF;
1206                         bch2_write_op_error_trans(trans, &buf, op, bkey_start_offset(&insert->k));
1207                         prt_printf(&buf, "btree update error: %s", bch2_err_str(ret));
1208                         bch_err_ratelimited(c, "%s", buf.buf);
1209                         printbuf_exit(&buf);
1210                 }
1211
1212                 if (ret) {
1213                         op->error = ret;
1214                         break;
1215                 }
1216         }
1217
1218         bch2_trans_put(trans);
1219 }
1220
1221 static void __bch2_nocow_write_done(struct bch_write_op *op)
1222 {
1223         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
1224                 op->error = -EIO;
1225         } else if (unlikely(op->flags & BCH_WRITE_CONVERT_UNWRITTEN))
1226                 bch2_nocow_write_convert_unwritten(op);
1227 }
1228
1229 static CLOSURE_CALLBACK(bch2_nocow_write_done)
1230 {
1231         closure_type(op, struct bch_write_op, cl);
1232
1233         __bch2_nocow_write_done(op);
1234         bch2_write_done(cl);
1235 }
1236
1237 struct bucket_to_lock {
1238         struct bpos             b;
1239         unsigned                gen;
1240         struct nocow_lock_bucket *l;
1241 };
1242
1243 static void bch2_nocow_write(struct bch_write_op *op)
1244 {
1245         struct bch_fs *c = op->c;
1246         struct btree_trans *trans;
1247         struct btree_iter iter;
1248         struct bkey_s_c k;
1249         DARRAY_PREALLOCATED(struct bucket_to_lock, 3) buckets;
1250         u32 snapshot;
1251         struct bucket_to_lock *stale_at;
1252         int stale, ret;
1253
1254         if (op->flags & BCH_WRITE_MOVE)
1255                 return;
1256
1257         darray_init(&buckets);
1258         trans = bch2_trans_get(c);
1259 retry:
1260         bch2_trans_begin(trans);
1261
1262         ret = bch2_subvolume_get_snapshot(trans, op->subvol, &snapshot);
1263         if (unlikely(ret))
1264                 goto err;
1265
1266         bch2_trans_iter_init(trans, &iter, BTREE_ID_extents,
1267                              SPOS(op->pos.inode, op->pos.offset, snapshot),
1268                              BTREE_ITER_slots);
1269         while (1) {
1270                 struct bio *bio = &op->wbio.bio;
1271
1272                 buckets.nr = 0;
1273
1274                 ret = bch2_trans_relock(trans);
1275                 if (ret)
1276                         break;
1277
1278                 k = bch2_btree_iter_peek_slot(&iter);
1279                 ret = bkey_err(k);
1280                 if (ret)
1281                         break;
1282
1283                 /* fall back to normal cow write path? */
1284                 if (unlikely(k.k->p.snapshot != snapshot ||
1285                              !bch2_extent_is_writeable(op, k)))
1286                         break;
1287
1288                 if (bch2_keylist_realloc(&op->insert_keys,
1289                                          op->inline_keys,
1290                                          ARRAY_SIZE(op->inline_keys),
1291                                          k.k->u64s))
1292                         break;
1293
1294                 /* Get iorefs before dropping btree locks: */
1295                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1296                 bkey_for_each_ptr(ptrs, ptr) {
1297                         struct bch_dev *ca = bch2_dev_get_ioref(c, ptr->dev, WRITE);
1298                         if (unlikely(!ca))
1299                                 goto err_get_ioref;
1300
1301                         struct bpos b = PTR_BUCKET_POS(ca, ptr);
1302                         struct nocow_lock_bucket *l =
1303                                 bucket_nocow_lock(&c->nocow_locks, bucket_to_u64(b));
1304                         prefetch(l);
1305
1306                         /* XXX allocating memory with btree locks held - rare */
1307                         darray_push_gfp(&buckets, ((struct bucket_to_lock) {
1308                                                    .b = b, .gen = ptr->gen, .l = l,
1309                                                    }), GFP_KERNEL|__GFP_NOFAIL);
1310
1311                         if (ptr->unwritten)
1312                                 op->flags |= BCH_WRITE_CONVERT_UNWRITTEN;
1313                 }
1314
1315                 /* Unlock before taking nocow locks, doing IO: */
1316                 bkey_reassemble(op->insert_keys.top, k);
1317                 bch2_trans_unlock(trans);
1318
1319                 bch2_cut_front(op->pos, op->insert_keys.top);
1320                 if (op->flags & BCH_WRITE_CONVERT_UNWRITTEN)
1321                         bch2_cut_back(POS(op->pos.inode, op->pos.offset + bio_sectors(bio)), op->insert_keys.top);
1322
1323                 darray_for_each(buckets, i) {
1324                         struct bch_dev *ca = bch2_dev_have_ref(c, i->b.inode);
1325
1326                         __bch2_bucket_nocow_lock(&c->nocow_locks, i->l,
1327                                                  bucket_to_u64(i->b),
1328                                                  BUCKET_NOCOW_LOCK_UPDATE);
1329
1330                         int gen = bucket_gen_get(ca, i->b.offset);
1331                         stale = gen < 0 ? gen : gen_after(gen, i->gen);
1332                         if (unlikely(stale)) {
1333                                 stale_at = i;
1334                                 goto err_bucket_stale;
1335                         }
1336                 }
1337
1338                 bio = &op->wbio.bio;
1339                 if (k.k->p.offset < op->pos.offset + bio_sectors(bio)) {
1340                         bio = bio_split(bio, k.k->p.offset - op->pos.offset,
1341                                         GFP_KERNEL, &c->bio_write);
1342                         wbio_init(bio)->put_bio = true;
1343                         bio->bi_opf = op->wbio.bio.bi_opf;
1344                 } else {
1345                         op->flags |= BCH_WRITE_SUBMITTED;
1346                 }
1347
1348                 op->pos.offset += bio_sectors(bio);
1349                 op->written += bio_sectors(bio);
1350
1351                 bio->bi_end_io  = bch2_write_endio;
1352                 bio->bi_private = &op->cl;
1353                 bio->bi_opf |= REQ_OP_WRITE;
1354                 closure_get(&op->cl);
1355                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1356                                           op->insert_keys.top, true);
1357
1358                 bch2_keylist_push(&op->insert_keys);
1359                 if (op->flags & BCH_WRITE_SUBMITTED)
1360                         break;
1361                 bch2_btree_iter_advance(&iter);
1362         }
1363 out:
1364         bch2_trans_iter_exit(trans, &iter);
1365 err:
1366         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1367                 goto retry;
1368
1369         bch2_trans_put(trans);
1370         darray_exit(&buckets);
1371
1372         if (ret) {
1373                 struct printbuf buf = PRINTBUF;
1374                 bch2_write_op_error(&buf, op);
1375                 prt_printf(&buf, "%s(): btree lookup error: %s", __func__, bch2_err_str(ret));
1376                 bch_err_ratelimited(c, "%s", buf.buf);
1377                 printbuf_exit(&buf);
1378                 op->error = ret;
1379                 op->flags |= BCH_WRITE_SUBMITTED;
1380         }
1381
1382         /* fallback to cow write path? */
1383         if (!(op->flags & BCH_WRITE_SUBMITTED)) {
1384                 closure_sync(&op->cl);
1385                 __bch2_nocow_write_done(op);
1386                 op->insert_keys.top = op->insert_keys.keys;
1387         } else if (op->flags & BCH_WRITE_SYNC) {
1388                 closure_sync(&op->cl);
1389                 bch2_nocow_write_done(&op->cl.work);
1390         } else {
1391                 /*
1392                  * XXX
1393                  * needs to run out of process context because ei_quota_lock is
1394                  * a mutex
1395                  */
1396                 continue_at(&op->cl, bch2_nocow_write_done, index_update_wq(op));
1397         }
1398         return;
1399 err_get_ioref:
1400         darray_for_each(buckets, i)
1401                 percpu_ref_put(&bch2_dev_have_ref(c, i->b.inode)->io_ref);
1402
1403         /* Fall back to COW path: */
1404         goto out;
1405 err_bucket_stale:
1406         darray_for_each(buckets, i) {
1407                 bch2_bucket_nocow_unlock(&c->nocow_locks, i->b, BUCKET_NOCOW_LOCK_UPDATE);
1408                 if (i == stale_at)
1409                         break;
1410         }
1411
1412         struct printbuf buf = PRINTBUF;
1413         if (bch2_fs_inconsistent_on(stale < 0, c,
1414                                     "pointer to invalid bucket in nocow path on device %llu\n  %s",
1415                                     stale_at->b.inode,
1416                                     (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1417                 ret = -EIO;
1418         } else {
1419                 /* We can retry this: */
1420                 ret = -BCH_ERR_transaction_restart;
1421         }
1422         printbuf_exit(&buf);
1423
1424         goto err_get_ioref;
1425 }
1426
1427 static void __bch2_write(struct bch_write_op *op)
1428 {
1429         struct bch_fs *c = op->c;
1430         struct write_point *wp = NULL;
1431         struct bio *bio = NULL;
1432         unsigned nofs_flags;
1433         int ret;
1434
1435         nofs_flags = memalloc_nofs_save();
1436
1437         if (unlikely(op->opts.nocow && c->opts.nocow_enabled)) {
1438                 bch2_nocow_write(op);
1439                 if (op->flags & BCH_WRITE_SUBMITTED)
1440                         goto out_nofs_restore;
1441         }
1442 again:
1443         memset(&op->failed, 0, sizeof(op->failed));
1444
1445         do {
1446                 struct bkey_i *key_to_write;
1447                 unsigned key_to_write_offset = op->insert_keys.top_p -
1448                         op->insert_keys.keys_p;
1449
1450                 /* +1 for possible cache device: */
1451                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1452                     ARRAY_SIZE(op->open_buckets.v))
1453                         break;
1454
1455                 if (bch2_keylist_realloc(&op->insert_keys,
1456                                         op->inline_keys,
1457                                         ARRAY_SIZE(op->inline_keys),
1458                                         BKEY_EXTENT_U64s_MAX))
1459                         break;
1460
1461                 /*
1462                  * The copygc thread is now global, which means it's no longer
1463                  * freeing up space on specific disks, which means that
1464                  * allocations for specific disks may hang arbitrarily long:
1465                  */
1466                 ret = bch2_trans_run(c, lockrestart_do(trans,
1467                         bch2_alloc_sectors_start_trans(trans,
1468                                 op->target,
1469                                 op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1470                                 op->write_point,
1471                                 &op->devs_have,
1472                                 op->nr_replicas,
1473                                 op->nr_replicas_required,
1474                                 op->watermark,
1475                                 op->flags,
1476                                 &op->cl, &wp)));
1477                 if (unlikely(ret)) {
1478                         if (bch2_err_matches(ret, BCH_ERR_operation_blocked))
1479                                 break;
1480
1481                         goto err;
1482                 }
1483
1484                 EBUG_ON(!wp);
1485
1486                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1487                 ret = bch2_write_extent(op, wp, &bio);
1488
1489                 bch2_alloc_sectors_done_inlined(c, wp);
1490 err:
1491                 if (ret <= 0) {
1492                         op->flags |= BCH_WRITE_SUBMITTED;
1493
1494                         if (unlikely(ret < 0)) {
1495                                 if (!(op->flags & BCH_WRITE_ALLOC_NOWAIT)) {
1496                                         struct printbuf buf = PRINTBUF;
1497                                         bch2_write_op_error(&buf, op);
1498                                         prt_printf(&buf, "%s(): %s", __func__, bch2_err_str(ret));
1499                                         bch_err_ratelimited(c, "%s", buf.buf);
1500                                         printbuf_exit(&buf);
1501                                 }
1502                                 op->error = ret;
1503                                 break;
1504                         }
1505                 }
1506
1507                 bio->bi_end_io  = bch2_write_endio;
1508                 bio->bi_private = &op->cl;
1509                 bio->bi_opf |= REQ_OP_WRITE;
1510
1511                 closure_get(bio->bi_private);
1512
1513                 key_to_write = (void *) (op->insert_keys.keys_p +
1514                                          key_to_write_offset);
1515
1516                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1517                                           key_to_write, false);
1518         } while (ret);
1519
1520         /*
1521          * Sync or no?
1522          *
1523          * If we're running asynchronously, wne may still want to block
1524          * synchronously here if we weren't able to submit all of the IO at
1525          * once, as that signals backpressure to the caller.
1526          */
1527         if ((op->flags & BCH_WRITE_SYNC) ||
1528             (!(op->flags & BCH_WRITE_SUBMITTED) &&
1529              !(op->flags & BCH_WRITE_IN_WORKER))) {
1530                 bch2_wait_on_allocator(c, &op->cl);
1531
1532                 __bch2_write_index(op);
1533
1534                 if (!(op->flags & BCH_WRITE_SUBMITTED))
1535                         goto again;
1536                 bch2_write_done(&op->cl);
1537         } else {
1538                 bch2_write_queue(op, wp);
1539                 continue_at(&op->cl, bch2_write_index, NULL);
1540         }
1541 out_nofs_restore:
1542         memalloc_nofs_restore(nofs_flags);
1543 }
1544
1545 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1546 {
1547         struct bio *bio = &op->wbio.bio;
1548         struct bvec_iter iter;
1549         struct bkey_i_inline_data *id;
1550         unsigned sectors;
1551         int ret;
1552
1553         memset(&op->failed, 0, sizeof(op->failed));
1554
1555         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1556         op->flags |= BCH_WRITE_SUBMITTED;
1557
1558         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1559
1560         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1561                                    ARRAY_SIZE(op->inline_keys),
1562                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1563         if (ret) {
1564                 op->error = ret;
1565                 goto err;
1566         }
1567
1568         sectors = bio_sectors(bio);
1569         op->pos.offset += sectors;
1570
1571         id = bkey_inline_data_init(op->insert_keys.top);
1572         id->k.p         = op->pos;
1573         id->k.bversion  = op->version;
1574         id->k.size      = sectors;
1575
1576         iter = bio->bi_iter;
1577         iter.bi_size = data_len;
1578         memcpy_from_bio(id->v.data, bio, iter);
1579
1580         while (data_len & 7)
1581                 id->v.data[data_len++] = '\0';
1582         set_bkey_val_bytes(&id->k, data_len);
1583         bch2_keylist_push(&op->insert_keys);
1584
1585         __bch2_write_index(op);
1586 err:
1587         bch2_write_done(&op->cl);
1588 }
1589
1590 /**
1591  * bch2_write() - handle a write to a cache device or flash only volume
1592  * @cl:         &bch_write_op->cl
1593  *
1594  * This is the starting point for any data to end up in a cache device; it could
1595  * be from a normal write, or a writeback write, or a write to a flash only
1596  * volume - it's also used by the moving garbage collector to compact data in
1597  * mostly empty buckets.
1598  *
1599  * It first writes the data to the cache, creating a list of keys to be inserted
1600  * (if the data won't fit in a single open bucket, there will be multiple keys);
1601  * after the data is written it calls bch_journal, and after the keys have been
1602  * added to the next journal write they're inserted into the btree.
1603  *
1604  * If op->discard is true, instead of inserting the data it invalidates the
1605  * region of the cache represented by op->bio and op->inode.
1606  */
1607 CLOSURE_CALLBACK(bch2_write)
1608 {
1609         closure_type(op, struct bch_write_op, cl);
1610         struct bio *bio = &op->wbio.bio;
1611         struct bch_fs *c = op->c;
1612         unsigned data_len;
1613
1614         EBUG_ON(op->cl.parent);
1615         BUG_ON(!op->nr_replicas);
1616         BUG_ON(!op->write_point.v);
1617         BUG_ON(bkey_eq(op->pos, POS_MAX));
1618
1619         if (op->flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
1620                 op->flags |= BCH_WRITE_ALLOC_NOWAIT;
1621
1622         op->nr_replicas_required = min_t(unsigned, op->nr_replicas_required, op->nr_replicas);
1623         op->start_time = local_clock();
1624         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1625         wbio_init(bio)->put_bio = false;
1626
1627         if (unlikely(bio->bi_iter.bi_size & (c->opts.block_size - 1))) {
1628                 struct printbuf buf = PRINTBUF;
1629                 bch2_write_op_error(&buf, op);
1630                 prt_printf(&buf, "misaligned write");
1631                 printbuf_exit(&buf);
1632                 op->error = -EIO;
1633                 goto err;
1634         }
1635
1636         if (c->opts.nochanges) {
1637                 op->error = -BCH_ERR_erofs_no_writes;
1638                 goto err;
1639         }
1640
1641         if (!(op->flags & BCH_WRITE_MOVE) &&
1642             !bch2_write_ref_tryget(c, BCH_WRITE_REF_write)) {
1643                 op->error = -BCH_ERR_erofs_no_writes;
1644                 goto err;
1645         }
1646
1647         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1648         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1649
1650         data_len = min_t(u64, bio->bi_iter.bi_size,
1651                          op->new_i_size - (op->pos.offset << 9));
1652
1653         if (c->opts.inline_data &&
1654             data_len <= min(block_bytes(c) / 2, 1024U)) {
1655                 bch2_write_data_inline(op, data_len);
1656                 return;
1657         }
1658
1659         __bch2_write(op);
1660         return;
1661 err:
1662         bch2_disk_reservation_put(c, &op->res);
1663
1664         closure_debug_destroy(&op->cl);
1665         if (op->end_io)
1666                 op->end_io(op);
1667 }
1668
1669 static const char * const bch2_write_flags[] = {
1670 #define x(f)    #f,
1671         BCH_WRITE_FLAGS()
1672 #undef x
1673         NULL
1674 };
1675
1676 void bch2_write_op_to_text(struct printbuf *out, struct bch_write_op *op)
1677 {
1678         prt_str(out, "pos: ");
1679         bch2_bpos_to_text(out, op->pos);
1680         prt_newline(out);
1681         printbuf_indent_add(out, 2);
1682
1683         prt_str(out, "started: ");
1684         bch2_pr_time_units(out, local_clock() - op->start_time);
1685         prt_newline(out);
1686
1687         prt_str(out, "flags: ");
1688         prt_bitflags(out, bch2_write_flags, op->flags);
1689         prt_newline(out);
1690
1691         prt_printf(out, "ref: %u\n", closure_nr_remaining(&op->cl));
1692
1693         printbuf_indent_sub(out, 2);
1694 }
1695
1696 void bch2_fs_io_write_exit(struct bch_fs *c)
1697 {
1698         mempool_exit(&c->bio_bounce_pages);
1699         bioset_exit(&c->replica_set);
1700         bioset_exit(&c->bio_write);
1701 }
1702
1703 int bch2_fs_io_write_init(struct bch_fs *c)
1704 {
1705         if (bioset_init(&c->bio_write,   1, offsetof(struct bch_write_bio, bio), BIOSET_NEED_BVECS) ||
1706             bioset_init(&c->replica_set, 4, offsetof(struct bch_write_bio, bio), 0))
1707                 return -BCH_ERR_ENOMEM_bio_write_init;
1708
1709         if (mempool_init_page_pool(&c->bio_bounce_pages,
1710                                    max_t(unsigned,
1711                                          c->opts.btree_node_size,
1712                                          c->opts.encoded_extent_max) /
1713                                    PAGE_SIZE, 0))
1714                 return -BCH_ERR_ENOMEM_bio_bounce_pages_init;
1715
1716         return 0;
1717 }