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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Some low level IO code, and hacks for various block layer limitations
   4  *
   5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   6  * Copyright 2012 Google, Inc.
   7  */
   8
   9 #include "bcachefs.h"
  10 #include "alloc_background.h"
  11 #include "alloc_foreground.h"
  12 #include "bkey_buf.h"
  13 #include "bset.h"
  14 #include "btree_update.h"
  15 #include "buckets.h"
  16 #include "checksum.h"
  17 #include "compress.h"
  18 #include "clock.h"
  19 #include "debug.h"
  20 #include "disk_groups.h"
  21 #include "ec.h"
  22 #include "error.h"
  23 #include "extent_update.h"
  24 #include "inode.h"
  25 #include "io.h"
  26 #include "journal.h"
  27 #include "keylist.h"
  28 #include "move.h"
  29 #include "rebalance.h"
  30 #include "subvolume.h"
  31 #include "super.h"
  32 #include "super-io.h"
  33
  34 #include <linux/blkdev.h>
  35 #include <linux/random.h>
  36 #include <linux/sched/mm.h>
  37
  38 #include <trace/events/bcachefs.h>
  39
  40 const char *bch2_blk_status_to_str(blk_status_t status)
  41 {
  42         if (status == BLK_STS_REMOVED)
  43                 return "device removed";
  44         return blk_status_to_str(status);
  45 }
  46
  47 static bool bch2_target_congested(struct bch_fs *c, u16 target)
  48 {
  49         const struct bch_devs_mask *devs;
  50         unsigned d, nr = 0, total = 0;
  51         u64 now = local_clock(), last;
  52         s64 congested;
  53         struct bch_dev *ca;
  54
  55         if (!target)
  56                 return false;
  57
  58         rcu_read_lock();
  59         devs = bch2_target_to_mask(c, target) ?:
  60                 &c->rw_devs[BCH_DATA_user];
  61
  62         for_each_set_bit(d, devs->d, BCH_SB_MEMBERS_MAX) {
  63                 ca = rcu_dereference(c->devs[d]);
  64                 if (!ca)
  65                         continue;
  66
  67                 congested = atomic_read(&ca->congested);
  68                 last = READ_ONCE(ca->congested_last);
  69                 if (time_after64(now, last))
  70                         congested -= (now - last) >> 12;
  71
  72                 total += max(congested, 0LL);
  73                 nr++;
  74         }
  75         rcu_read_unlock();
  76
  77         return bch2_rand_range(nr * CONGESTED_MAX) < total;
  78 }
  79
  80 static inline void bch2_congested_acct(struct bch_dev *ca, u64 io_latency,
  81                                        u64 now, int rw)
  82 {
  83         u64 latency_capable =
  84                 ca->io_latency[rw].quantiles.entries[QUANTILE_IDX(1)].m;
  85         /* ideally we'd be taking into account the device's variance here: */
  86         u64 latency_threshold = latency_capable << (rw == READ ? 2 : 3);
  87         s64 latency_over = io_latency - latency_threshold;
  88
  89         if (latency_threshold && latency_over > 0) {
  90                 /*
  91                  * bump up congested by approximately latency_over * 4 /
  92                  * latency_threshold - we don't need much accuracy here so don't
  93                  * bother with the divide:
  94                  */
  95                 if (atomic_read(&ca->congested) < CONGESTED_MAX)
  96                         atomic_add(latency_over >>
  97                                    max_t(int, ilog2(latency_threshold) - 2, 0),
  98                                    &ca->congested);
  99
 100                 ca->congested_last = now;
 101         } else if (atomic_read(&ca->congested) > 0) {
 102                 atomic_dec(&ca->congested);
 103         }
 104 }
 105
 106 void bch2_latency_acct(struct bch_dev *ca, u64 submit_time, int rw)
 107 {
 108         atomic64_t *latency = &ca->cur_latency[rw];
 109         u64 now = local_clock();
 110         u64 io_latency = time_after64(now, submit_time)
 111                 ? now - submit_time
 112                 : 0;
 113         u64 old, new, v = atomic64_read(latency);
 114
 115         do {
 116                 old = v;
 117
 118                 /*
 119                  * If the io latency was reasonably close to the current
 120                  * latency, skip doing the update and atomic operation - most of
 121                  * the time:
 122                  */
 123                 if (abs((int) (old - io_latency)) < (old >> 1) &&
 124                     now & ~(~0U << 5))
 125                         break;
 126
 127                 new = ewma_add(old, io_latency, 5);
 128         } while ((v = atomic64_cmpxchg(latency, old, new)) != old);
 129
 130         bch2_congested_acct(ca, io_latency, now, rw);
 131
 132         __bch2_time_stats_update(&ca->io_latency[rw], submit_time, now);
 133 }
 134
 135 /* Allocate, free from mempool: */
 136
 137 void bch2_bio_free_pages_pool(struct bch_fs *c, struct bio *bio)
 138 {
 139         struct bvec_iter_all iter;
 140         struct bio_vec *bv;
 141
 142         bio_for_each_segment_all(bv, bio, iter)
 143                 if (bv->bv_page != ZERO_PAGE(0))
 144                         mempool_free(bv->bv_page, &c->bio_bounce_pages);
 145         bio->bi_vcnt = 0;
 146 }
 147
 148 static struct page *__bio_alloc_page_pool(struct bch_fs *c, bool *using_mempool)
 149 {
 150         struct page *page;
 151
 152         if (likely(!*using_mempool)) {
 153                 page = alloc_page(GFP_NOIO);
 154                 if (unlikely(!page)) {
 155                         mutex_lock(&c->bio_bounce_pages_lock);
 156                         *using_mempool = true;
 157                         goto pool_alloc;
 158
 159                 }
 160         } else {
 161 pool_alloc:
 162                 page = mempool_alloc(&c->bio_bounce_pages, GFP_NOIO);
 163         }
 164
 165         return page;
 166 }
 167
 168 void bch2_bio_alloc_pages_pool(struct bch_fs *c, struct bio *bio,
 169                                size_t size)
 170 {
 171         bool using_mempool = false;
 172
 173         while (size) {
 174                 struct page *page = __bio_alloc_page_pool(c, &using_mempool);
 175                 unsigned len = min_t(size_t, PAGE_SIZE, size);
 176
 177                 BUG_ON(!bio_add_page(bio, page, len, 0));
 178                 size -= len;
 179         }
 180
 181         if (using_mempool)
 182                 mutex_unlock(&c->bio_bounce_pages_lock);
 183 }
 184
 185 /* Extent update path: */
 186
 187 int bch2_sum_sector_overwrites(struct btree_trans *trans,
 188                                struct btree_iter *extent_iter,
 189                                struct bkey_i *new,
 190                                bool *usage_increasing,
 191                                s64 *i_sectors_delta,
 192                                s64 *disk_sectors_delta)
 193 {
 194         struct bch_fs *c = trans->c;
 195         struct btree_iter iter;
 196         struct bkey_s_c old;
 197         unsigned new_replicas = bch2_bkey_replicas(c, bkey_i_to_s_c(new));
 198         bool new_compressed = bch2_bkey_sectors_compressed(bkey_i_to_s_c(new));
 199         int ret = 0;
 200
 201         *usage_increasing       = false;
 202         *i_sectors_delta        = 0;
 203         *disk_sectors_delta     = 0;
 204
 205         bch2_trans_copy_iter(&iter, extent_iter);
 206
 207         for_each_btree_key_continue_norestart(iter, BTREE_ITER_SLOTS, old, ret) {
 208                 s64 sectors = min(new->k.p.offset, old.k->p.offset) -
 209                         max(bkey_start_offset(&new->k),
 210                             bkey_start_offset(old.k));
 211
 212                 *i_sectors_delta += sectors *
 213                         (bkey_extent_is_allocation(&new->k) -
 214                          bkey_extent_is_allocation(old.k));
 215
 216                 *disk_sectors_delta += sectors * bch2_bkey_nr_ptrs_allocated(bkey_i_to_s_c(new));
 217                 *disk_sectors_delta -= new->k.p.snapshot == old.k->p.snapshot
 218                         ? sectors * bch2_bkey_nr_ptrs_fully_allocated(old)
 219                         : 0;
 220
 221                 if (!*usage_increasing &&
 222                     (new->k.p.snapshot != old.k->p.snapshot ||
 223                      new_replicas > bch2_bkey_replicas(c, old) ||
 224                      (!new_compressed && bch2_bkey_sectors_compressed(old))))
 225                         *usage_increasing = true;
 226
 227                 if (bkey_cmp(old.k->p, new->k.p) >= 0)
 228                         break;
 229         }
 230
 231         bch2_trans_iter_exit(trans, &iter);
 232         return ret;
 233 }
 234
 235 int bch2_extent_update(struct btree_trans *trans,
 236                        subvol_inum inum,
 237                        struct btree_iter *iter,
 238                        struct bkey_i *k,
 239                        struct disk_reservation *disk_res,
 240                        u64 *journal_seq,
 241                        u64 new_i_size,
 242                        s64 *i_sectors_delta_total,
 243                        bool check_enospc)
 244 {
 245         struct btree_iter inode_iter = { NULL };
 246         struct bpos next_pos;
 247         bool usage_increasing;
 248         s64 i_sectors_delta = 0, disk_sectors_delta = 0;
 249         int ret;
 250
 251         /*
 252          * This traverses us the iterator without changing iter->path->pos to
 253          * search_key() (which is pos + 1 for extents): we want there to be a
 254          * path already traversed at iter->pos because
 255          * bch2_trans_extent_update() will use it to attempt extent merging
 256          */
 257         ret = __bch2_btree_iter_traverse(iter);
 258         if (ret)
 259                 return ret;
 260
 261         ret = bch2_extent_trim_atomic(trans, iter, k);
 262         if (ret)
 263                 return ret;
 264
 265         new_i_size = min(k->k.p.offset << 9, new_i_size);
 266         next_pos = k->k.p;
 267
 268         ret = bch2_sum_sector_overwrites(trans, iter, k,
 269                         &usage_increasing,
 270                         &i_sectors_delta,
 271                         &disk_sectors_delta);
 272         if (ret)
 273                 return ret;
 274
 275         if (disk_res &&
 276             disk_sectors_delta > (s64) disk_res->sectors) {
 277                 ret = bch2_disk_reservation_add(trans->c, disk_res,
 278                                         disk_sectors_delta - disk_res->sectors,
 279                                         !check_enospc || !usage_increasing
 280                                         ? BCH_DISK_RESERVATION_NOFAIL : 0);
 281                 if (ret)
 282                         return ret;
 283         }
 284
 285         if (new_i_size || i_sectors_delta) {
 286                 struct bkey_s_c k;
 287                 struct bkey_s_c_inode_v3 inode;
 288                 struct bkey_i_inode_v3 *new_inode;
 289                 bool i_size_update;
 290
 291                 bch2_trans_iter_init(trans, &inode_iter, BTREE_ID_inodes,
 292                                      SPOS(0, inum.inum, iter->snapshot),
 293                                      BTREE_ITER_INTENT|BTREE_ITER_CACHED);
 294                 k = bch2_btree_iter_peek_slot(&inode_iter);
 295                 ret = bkey_err(k);
 296                 if (unlikely(ret))
 297                         goto err;
 298
 299                 ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 300                 if (unlikely(ret))
 301                         goto err;
 302
 303                 if (unlikely(k.k->type != KEY_TYPE_inode_v3)) {
 304                         k = bch2_inode_to_v3(trans, k);
 305                         ret = bkey_err(k);
 306                         if (unlikely(ret))
 307                                 goto err;
 308                 }
 309
 310                 inode = bkey_s_c_to_inode_v3(k);
 311                 i_size_update = !(le64_to_cpu(inode.v->bi_flags) & BCH_INODE_I_SIZE_DIRTY) &&
 312                         new_i_size > le64_to_cpu(inode.v->bi_size);
 313
 314                 if (!i_sectors_delta && !i_size_update)
 315                         goto no_inode_update;
 316
 317                 new_inode = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 318                 ret = PTR_ERR_OR_ZERO(new_inode);
 319                 if (unlikely(ret))
 320                         goto err;
 321
 322                 bkey_reassemble(&new_inode->k_i, k);
 323
 324                 if (i_size_update)
 325                         new_inode->v.bi_size = cpu_to_le64(new_i_size);
 326
 327                 le64_add_cpu(&new_inode->v.bi_sectors, i_sectors_delta);
 328                 ret = bch2_trans_update(trans, &inode_iter, &new_inode->k_i, 0);
 329                 if (unlikely(ret))
 330                         goto err;
 331         }
 332 no_inode_update:
 333         ret =   bch2_trans_update(trans, iter, k, 0) ?:
 334                 bch2_trans_commit(trans, disk_res, journal_seq,
 335                                 BTREE_INSERT_NOCHECK_RW|
 336                                 BTREE_INSERT_NOFAIL);
 337         if (unlikely(ret))
 338                 goto err;
 339
 340         if (i_sectors_delta_total)
 341                 *i_sectors_delta_total += i_sectors_delta;
 342         bch2_btree_iter_set_pos(iter, next_pos);
 343 err:
 344         bch2_trans_iter_exit(trans, &inode_iter);
 345         return ret;
 346 }
 347
 348 /*
 349  * Returns -BCH_ERR_transacton_restart if we had to drop locks:
 350  */
 351 int bch2_fpunch_at(struct btree_trans *trans, struct btree_iter *iter,
 352                    subvol_inum inum, u64 end,
 353                    s64 *i_sectors_delta)
 354 {
 355         struct bch_fs *c        = trans->c;
 356         unsigned max_sectors    = KEY_SIZE_MAX & (~0 << c->block_bits);
 357         struct bpos end_pos = POS(inum.inum, end);
 358         struct bkey_s_c k;
 359         int ret = 0, ret2 = 0;
 360         u32 snapshot;
 361
 362         while (!ret ||
 363                bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
 364                 struct disk_reservation disk_res =
 365                         bch2_disk_reservation_init(c, 0);
 366                 struct bkey_i delete;
 367
 368                 if (ret)
 369                         ret2 = ret;
 370
 371                 bch2_trans_begin(trans);
 372
 373                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 374                 if (ret)
 375                         continue;
 376
 377                 bch2_btree_iter_set_snapshot(iter, snapshot);
 378
 379                 k = bch2_btree_iter_peek(iter);
 380                 if (bkey_cmp(iter->pos, end_pos) >= 0) {
 381                         bch2_btree_iter_set_pos(iter, end_pos);
 382                         break;
 383                 }
 384
 385                 ret = bkey_err(k);
 386                 if (ret)
 387                         continue;
 388
 389                 bkey_init(&delete.k);
 390                 delete.k.p = iter->pos;
 391
 392                 /* create the biggest key we can */
 393                 bch2_key_resize(&delete.k, max_sectors);
 394                 bch2_cut_back(end_pos, &delete);
 395
 396                 ret = bch2_extent_update(trans, inum, iter, &delete,
 397                                 &disk_res, NULL,
 398                                 0, i_sectors_delta, false);
 399                 bch2_disk_reservation_put(c, &disk_res);
 400         }
 401
 402         return ret ?: ret2;
 403 }
 404
 405 int bch2_fpunch(struct bch_fs *c, subvol_inum inum, u64 start, u64 end,
 406                 s64 *i_sectors_delta)
 407 {
 408         struct btree_trans trans;
 409         struct btree_iter iter;
 410         int ret;
 411
 412         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 413         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 414                              POS(inum.inum, start),
 415                              BTREE_ITER_INTENT);
 416
 417         ret = bch2_fpunch_at(&trans, &iter, inum, end, i_sectors_delta);
 418
 419         bch2_trans_iter_exit(&trans, &iter);
 420         bch2_trans_exit(&trans);
 421
 422         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 423                 ret = 0;
 424
 425         return ret;
 426 }
 427
 428 int bch2_write_index_default(struct bch_write_op *op)
 429 {
 430         struct bch_fs *c = op->c;
 431         struct bkey_buf sk;
 432         struct keylist *keys = &op->insert_keys;
 433         struct bkey_i *k = bch2_keylist_front(keys);
 434         struct btree_trans trans;
 435         struct btree_iter iter;
 436         subvol_inum inum = {
 437                 .subvol = op->subvol,
 438                 .inum   = k->k.p.inode,
 439         };
 440         int ret;
 441
 442         BUG_ON(!inum.subvol);
 443
 444         bch2_bkey_buf_init(&sk);
 445         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 1024);
 446
 447         do {
 448                 bch2_trans_begin(&trans);
 449
 450                 k = bch2_keylist_front(keys);
 451                 bch2_bkey_buf_copy(&sk, c, k);
 452
 453                 ret = bch2_subvolume_get_snapshot(&trans, inum.subvol,
 454                                                   &sk.k->k.p.snapshot);
 455                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 456                         continue;
 457                 if (ret)
 458                         break;
 459
 460                 bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
 461                                      bkey_start_pos(&sk.k->k),
 462                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 463
 464                 ret = bch2_extent_update(&trans, inum, &iter, sk.k,
 465                                          &op->res, op_journal_seq(op),
 466                                          op->new_i_size, &op->i_sectors_delta,
 467                                          op->flags & BCH_WRITE_CHECK_ENOSPC);
 468                 bch2_trans_iter_exit(&trans, &iter);
 469
 470                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 471                         continue;
 472                 if (ret)
 473                         break;
 474
 475                 if (bkey_cmp(iter.pos, k->k.p) >= 0)
 476                         bch2_keylist_pop_front(&op->insert_keys);
 477                 else
 478                         bch2_cut_front(iter.pos, k);
 479         } while (!bch2_keylist_empty(keys));
 480
 481         bch2_trans_exit(&trans);
 482         bch2_bkey_buf_exit(&sk, c);
 483
 484         return ret;
 485 }
 486
 487 /* Writes */
 488
 489 void bch2_submit_wbio_replicas(struct bch_write_bio *wbio, struct bch_fs *c,
 490                                enum bch_data_type type,
 491                                const struct bkey_i *k)
 492 {
 493         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(bkey_i_to_s_c(k));
 494         const struct bch_extent_ptr *ptr;
 495         struct bch_write_bio *n;
 496         struct bch_dev *ca;
 497
 498         BUG_ON(c->opts.nochanges);
 499
 500         bkey_for_each_ptr(ptrs, ptr) {
 501                 BUG_ON(ptr->dev >= BCH_SB_MEMBERS_MAX ||
 502                        !c->devs[ptr->dev]);
 503
 504                 ca = bch_dev_bkey_exists(c, ptr->dev);
 505
 506                 if (to_entry(ptr + 1) < ptrs.end) {
 507                         n = to_wbio(bio_alloc_clone(NULL, &wbio->bio,
 508                                                 GFP_NOIO, &ca->replica_set));
 509
 510                         n->bio.bi_end_io        = wbio->bio.bi_end_io;
 511                         n->bio.bi_private       = wbio->bio.bi_private;
 512                         n->parent               = wbio;
 513                         n->split                = true;
 514                         n->bounce               = false;
 515                         n->put_bio              = true;
 516                         n->bio.bi_opf           = wbio->bio.bi_opf;
 517                         bio_inc_remaining(&wbio->bio);
 518                 } else {
 519                         n = wbio;
 520                         n->split                = false;
 521                 }
 522
 523                 n->c                    = c;
 524                 n->dev                  = ptr->dev;
 525                 n->have_ioref           = bch2_dev_get_ioref(ca,
 526                                         type == BCH_DATA_btree ? READ : WRITE);
 527                 n->submit_time          = local_clock();
 528                 n->bio.bi_iter.bi_sector = ptr->offset;
 529
 530                 if (likely(n->have_ioref)) {
 531                         this_cpu_add(ca->io_done->sectors[WRITE][type],
 532                                      bio_sectors(&n->bio));
 533
 534                         bio_set_dev(&n->bio, ca->disk_sb.bdev);
 535                         submit_bio(&n->bio);
 536                 } else {
 537                         n->bio.bi_status        = BLK_STS_REMOVED;
 538                         bio_endio(&n->bio);
 539                 }
 540         }
 541 }
 542
 543 static void __bch2_write(struct closure *);
 544
 545 static void bch2_write_done(struct closure *cl)
 546 {
 547         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 548         struct bch_fs *c = op->c;
 549
 550         if (!op->error && (op->flags & BCH_WRITE_FLUSH))
 551                 op->error = bch2_journal_error(&c->journal);
 552
 553         bch2_disk_reservation_put(c, &op->res);
 554         percpu_ref_put(&c->writes);
 555         bch2_keylist_free(&op->insert_keys, op->inline_keys);
 556
 557         bch2_time_stats_update(&c->times[BCH_TIME_data_write], op->start_time);
 558
 559         if (op->end_io) {
 560                 EBUG_ON(cl->parent);
 561                 closure_debug_destroy(cl);
 562                 op->end_io(op);
 563         } else {
 564                 closure_return(cl);
 565         }
 566 }
 567
 568 static noinline int bch2_write_drop_io_error_ptrs(struct bch_write_op *op)
 569 {
 570         struct keylist *keys = &op->insert_keys;
 571         struct bch_extent_ptr *ptr;
 572         struct bkey_i *src, *dst = keys->keys, *n;
 573
 574         for (src = keys->keys; src != keys->top; src = n) {
 575                 n = bkey_next(src);
 576
 577                 if (bkey_extent_is_direct_data(&src->k)) {
 578                         bch2_bkey_drop_ptrs(bkey_i_to_s(src), ptr,
 579                                             test_bit(ptr->dev, op->failed.d));
 580
 581                         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(src)))
 582                                 return -EIO;
 583                 }
 584
 585                 if (dst != src)
 586                         memmove_u64s_down(dst, src, src->u64s);
 587                 dst = bkey_next(dst);
 588         }
 589
 590         keys->top = dst;
 591         return 0;
 592 }
 593
 594 /**
 595  * bch_write_index - after a write, update index to point to new data
 596  */
 597 static void __bch2_write_index(struct bch_write_op *op)
 598 {
 599         struct bch_fs *c = op->c;
 600         struct keylist *keys = &op->insert_keys;
 601         struct bkey_i *k;
 602         unsigned dev;
 603         int ret;
 604
 605         if (unlikely(op->flags & BCH_WRITE_IO_ERROR)) {
 606                 ret = bch2_write_drop_io_error_ptrs(op);
 607                 if (ret)
 608                         goto err;
 609         }
 610
 611         /*
 612          * probably not the ideal place to hook this in, but I don't
 613          * particularly want to plumb io_opts all the way through the btree
 614          * update stack right now
 615          */
 616         for_each_keylist_key(keys, k) {
 617                 bch2_rebalance_add_key(c, bkey_i_to_s_c(k), &op->opts);
 618
 619                 if (bch2_bkey_is_incompressible(bkey_i_to_s_c(k)))
 620                         bch2_check_set_feature(op->c, BCH_FEATURE_incompressible);
 621
 622         }
 623
 624         if (!bch2_keylist_empty(keys)) {
 625                 u64 sectors_start = keylist_sectors(keys);
 626                 int ret = op->index_update_fn(op);
 627
 628                 BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart));
 629                 BUG_ON(keylist_sectors(keys) && !ret);
 630
 631                 op->written += sectors_start - keylist_sectors(keys);
 632
 633                 if (ret) {
 634                         bch_err_inum_ratelimited(c, op->pos.inode,
 635                                 "write error while doing btree update: %s", bch2_err_str(ret));
 636                         op->error = ret;
 637                 }
 638         }
 639 out:
 640         /* If some a bucket wasn't written, we can't erasure code it: */
 641         for_each_set_bit(dev, op->failed.d, BCH_SB_MEMBERS_MAX)
 642                 bch2_open_bucket_write_error(c, &op->open_buckets, dev);
 643
 644         bch2_open_buckets_put(c, &op->open_buckets);
 645         return;
 646 err:
 647         keys->top = keys->keys;
 648         op->error = ret;
 649         goto out;
 650 }
 651
 652 static void bch2_write_index(struct closure *cl)
 653 {
 654         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
 655         struct bch_fs *c = op->c;
 656
 657         __bch2_write_index(op);
 658
 659         if (!(op->flags & BCH_WRITE_DONE)) {
 660                 continue_at(cl, __bch2_write, index_update_wq(op));
 661         } else if (!op->error && (op->flags & BCH_WRITE_FLUSH)) {
 662                 bch2_journal_flush_seq_async(&c->journal,
 663                                              *op_journal_seq(op),
 664                                              cl);
 665                 continue_at(cl, bch2_write_done, index_update_wq(op));
 666         } else {
 667                 continue_at_nobarrier(cl, bch2_write_done, NULL);
 668         }
 669 }
 670
 671 static void bch2_write_endio(struct bio *bio)
 672 {
 673         struct closure *cl              = bio->bi_private;
 674         struct bch_write_op *op         = container_of(cl, struct bch_write_op, cl);
 675         struct bch_write_bio *wbio      = to_wbio(bio);
 676         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
 677         struct bch_fs *c                = wbio->c;
 678         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
 679
 680         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
 681                                     op->pos.inode,
 682                                     op->pos.offset - bio_sectors(bio), /* XXX definitely wrong */
 683                                     "data write error: %s",
 684                                     bch2_blk_status_to_str(bio->bi_status))) {
 685                 set_bit(wbio->dev, op->failed.d);
 686                 op->flags |= BCH_WRITE_IO_ERROR;
 687         }
 688
 689         if (wbio->have_ioref) {
 690                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
 691                 percpu_ref_put(&ca->io_ref);
 692         }
 693
 694         if (wbio->bounce)
 695                 bch2_bio_free_pages_pool(c, bio);
 696
 697         if (wbio->put_bio)
 698                 bio_put(bio);
 699
 700         if (parent)
 701                 bio_endio(&parent->bio);
 702         else if (!(op->flags & BCH_WRITE_SKIP_CLOSURE_PUT))
 703                 closure_put(cl);
 704         else
 705                 continue_at_nobarrier(cl, bch2_write_index, index_update_wq(op));
 706 }
 707
 708 static void init_append_extent(struct bch_write_op *op,
 709                                struct write_point *wp,
 710                                struct bversion version,
 711                                struct bch_extent_crc_unpacked crc)
 712 {
 713         struct bch_fs *c = op->c;
 714         struct bkey_i_extent *e;
 715
 716         op->pos.offset += crc.uncompressed_size;
 717
 718         e = bkey_extent_init(op->insert_keys.top);
 719         e->k.p          = op->pos;
 720         e->k.size       = crc.uncompressed_size;
 721         e->k.version    = version;
 722
 723         if (crc.csum_type ||
 724             crc.compression_type ||
 725             crc.nonce)
 726                 bch2_extent_crc_append(&e->k_i, crc);
 727
 728         bch2_alloc_sectors_append_ptrs(c, wp, &e->k_i, crc.compressed_size,
 729                                        op->flags & BCH_WRITE_CACHED);
 730
 731         bch2_keylist_push(&op->insert_keys);
 732 }
 733
 734 static struct bio *bch2_write_bio_alloc(struct bch_fs *c,
 735                                         struct write_point *wp,
 736                                         struct bio *src,
 737                                         bool *page_alloc_failed,
 738                                         void *buf)
 739 {
 740         struct bch_write_bio *wbio;
 741         struct bio *bio;
 742         unsigned output_available =
 743                 min(wp->sectors_free << 9, src->bi_iter.bi_size);
 744         unsigned pages = DIV_ROUND_UP(output_available +
 745                                       (buf
 746                                        ? ((unsigned long) buf & (PAGE_SIZE - 1))
 747                                        : 0), PAGE_SIZE);
 748
 749         pages = min(pages, BIO_MAX_VECS);
 750
 751         bio = bio_alloc_bioset(NULL, pages, 0,
 752                                GFP_NOIO, &c->bio_write);
 753         wbio                    = wbio_init(bio);
 754         wbio->put_bio           = true;
 755         /* copy WRITE_SYNC flag */
 756         wbio->bio.bi_opf        = src->bi_opf;
 757
 758         if (buf) {
 759                 bch2_bio_map(bio, buf, output_available);
 760                 return bio;
 761         }
 762
 763         wbio->bounce            = true;
 764
 765         /*
 766          * We can't use mempool for more than c->sb.encoded_extent_max
 767          * worth of pages, but we'd like to allocate more if we can:
 768          */
 769         bch2_bio_alloc_pages_pool(c, bio,
 770                                   min_t(unsigned, output_available,
 771                                         c->opts.encoded_extent_max));
 772
 773         if (bio->bi_iter.bi_size < output_available)
 774                 *page_alloc_failed =
 775                         bch2_bio_alloc_pages(bio,
 776                                              output_available -
 777                                              bio->bi_iter.bi_size,
 778                                              GFP_NOFS) != 0;
 779
 780         return bio;
 781 }
 782
 783 static int bch2_write_rechecksum(struct bch_fs *c,
 784                                  struct bch_write_op *op,
 785                                  unsigned new_csum_type)
 786 {
 787         struct bio *bio = &op->wbio.bio;
 788         struct bch_extent_crc_unpacked new_crc;
 789         int ret;
 790
 791         /* bch2_rechecksum_bio() can't encrypt or decrypt data: */
 792
 793         if (bch2_csum_type_is_encryption(op->crc.csum_type) !=
 794             bch2_csum_type_is_encryption(new_csum_type))
 795                 new_csum_type = op->crc.csum_type;
 796
 797         ret = bch2_rechecksum_bio(c, bio, op->version, op->crc,
 798                                   NULL, &new_crc,
 799                                   op->crc.offset, op->crc.live_size,
 800                                   new_csum_type);
 801         if (ret)
 802                 return ret;
 803
 804         bio_advance(bio, op->crc.offset << 9);
 805         bio->bi_iter.bi_size = op->crc.live_size << 9;
 806         op->crc = new_crc;
 807         return 0;
 808 }
 809
 810 static int bch2_write_decrypt(struct bch_write_op *op)
 811 {
 812         struct bch_fs *c = op->c;
 813         struct nonce nonce = extent_nonce(op->version, op->crc);
 814         struct bch_csum csum;
 815         int ret;
 816
 817         if (!bch2_csum_type_is_encryption(op->crc.csum_type))
 818                 return 0;
 819
 820         /*
 821          * If we need to decrypt data in the write path, we'll no longer be able
 822          * to verify the existing checksum (poly1305 mac, in this case) after
 823          * it's decrypted - this is the last point we'll be able to reverify the
 824          * checksum:
 825          */
 826         csum = bch2_checksum_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 827         if (bch2_crc_cmp(op->crc.csum, csum))
 828                 return -EIO;
 829
 830         ret = bch2_encrypt_bio(c, op->crc.csum_type, nonce, &op->wbio.bio);
 831         op->crc.csum_type = 0;
 832         op->crc.csum = (struct bch_csum) { 0, 0 };
 833         return ret;
 834 }
 835
 836 static enum prep_encoded_ret {
 837         PREP_ENCODED_OK,
 838         PREP_ENCODED_ERR,
 839         PREP_ENCODED_CHECKSUM_ERR,
 840         PREP_ENCODED_DO_WRITE,
 841 } bch2_write_prep_encoded_data(struct bch_write_op *op, struct write_point *wp)
 842 {
 843         struct bch_fs *c = op->c;
 844         struct bio *bio = &op->wbio.bio;
 845
 846         if (!(op->flags & BCH_WRITE_DATA_ENCODED))
 847                 return PREP_ENCODED_OK;
 848
 849         BUG_ON(bio_sectors(bio) != op->crc.compressed_size);
 850
 851         /* Can we just write the entire extent as is? */
 852         if (op->crc.uncompressed_size == op->crc.live_size &&
 853             op->crc.compressed_size <= wp->sectors_free &&
 854             (op->crc.compression_type == op->compression_type ||
 855              op->incompressible)) {
 856                 if (!crc_is_compressed(op->crc) &&
 857                     op->csum_type != op->crc.csum_type &&
 858                     bch2_write_rechecksum(c, op, op->csum_type))
 859                         return PREP_ENCODED_CHECKSUM_ERR;
 860
 861                 return PREP_ENCODED_DO_WRITE;
 862         }
 863
 864         /*
 865          * If the data is compressed and we couldn't write the entire extent as
 866          * is, we have to decompress it:
 867          */
 868         if (crc_is_compressed(op->crc)) {
 869                 struct bch_csum csum;
 870
 871                 if (bch2_write_decrypt(op))
 872                         return PREP_ENCODED_CHECKSUM_ERR;
 873
 874                 /* Last point we can still verify checksum: */
 875                 csum = bch2_checksum_bio(c, op->crc.csum_type,
 876                                          extent_nonce(op->version, op->crc),
 877                                          bio);
 878                 if (bch2_crc_cmp(op->crc.csum, csum))
 879                         return PREP_ENCODED_CHECKSUM_ERR;
 880
 881                 if (bch2_bio_uncompress_inplace(c, bio, &op->crc))
 882                         return PREP_ENCODED_ERR;
 883         }
 884
 885         /*
 886          * No longer have compressed data after this point - data might be
 887          * encrypted:
 888          */
 889
 890         /*
 891          * If the data is checksummed and we're only writing a subset,
 892          * rechecksum and adjust bio to point to currently live data:
 893          */
 894         if ((op->crc.live_size != op->crc.uncompressed_size ||
 895              op->crc.csum_type != op->csum_type) &&
 896             bch2_write_rechecksum(c, op, op->csum_type))
 897                 return PREP_ENCODED_CHECKSUM_ERR;
 898
 899         /*
 900          * If we want to compress the data, it has to be decrypted:
 901          */
 902         if ((op->compression_type ||
 903              bch2_csum_type_is_encryption(op->crc.csum_type) !=
 904              bch2_csum_type_is_encryption(op->csum_type)) &&
 905             bch2_write_decrypt(op))
 906                 return PREP_ENCODED_CHECKSUM_ERR;
 907
 908         return PREP_ENCODED_OK;
 909 }
 910
 911 static int bch2_write_extent(struct bch_write_op *op, struct write_point *wp,
 912                              struct bio **_dst)
 913 {
 914         struct bch_fs *c = op->c;
 915         struct bio *src = &op->wbio.bio, *dst = src;
 916         struct bvec_iter saved_iter;
 917         void *ec_buf;
 918         unsigned total_output = 0, total_input = 0;
 919         bool bounce = false;
 920         bool page_alloc_failed = false;
 921         int ret, more = 0;
 922
 923         BUG_ON(!bio_sectors(src));
 924
 925         ec_buf = bch2_writepoint_ec_buf(c, wp);
 926
 927         switch (bch2_write_prep_encoded_data(op, wp)) {
 928         case PREP_ENCODED_OK:
 929                 break;
 930         case PREP_ENCODED_ERR:
 931                 ret = -EIO;
 932                 goto err;
 933         case PREP_ENCODED_CHECKSUM_ERR:
 934                 goto csum_err;
 935         case PREP_ENCODED_DO_WRITE:
 936                 /* XXX look for bug here */
 937                 if (ec_buf) {
 938                         dst = bch2_write_bio_alloc(c, wp, src,
 939                                                    &page_alloc_failed,
 940                                                    ec_buf);
 941                         bio_copy_data(dst, src);
 942                         bounce = true;
 943                 }
 944                 init_append_extent(op, wp, op->version, op->crc);
 945                 goto do_write;
 946         }
 947
 948         if (ec_buf ||
 949             op->compression_type ||
 950             (op->csum_type &&
 951              !(op->flags & BCH_WRITE_PAGES_STABLE)) ||
 952             (bch2_csum_type_is_encryption(op->csum_type) &&
 953              !(op->flags & BCH_WRITE_PAGES_OWNED))) {
 954                 dst = bch2_write_bio_alloc(c, wp, src,
 955                                            &page_alloc_failed,
 956                                            ec_buf);
 957                 bounce = true;
 958         }
 959
 960         saved_iter = dst->bi_iter;
 961
 962         do {
 963                 struct bch_extent_crc_unpacked crc = { 0 };
 964                 struct bversion version = op->version;
 965                 size_t dst_len, src_len;
 966
 967                 if (page_alloc_failed &&
 968                     dst->bi_iter.bi_size  < (wp->sectors_free << 9) &&
 969                     dst->bi_iter.bi_size < c->opts.encoded_extent_max)
 970                         break;
 971
 972                 BUG_ON(op->compression_type &&
 973                        (op->flags & BCH_WRITE_DATA_ENCODED) &&
 974                        bch2_csum_type_is_encryption(op->crc.csum_type));
 975                 BUG_ON(op->compression_type && !bounce);
 976
 977                 crc.compression_type = op->incompressible
 978                         ? BCH_COMPRESSION_TYPE_incompressible
 979                         : op->compression_type
 980                         ? bch2_bio_compress(c, dst, &dst_len, src, &src_len,
 981                                             op->compression_type)
 982                         : 0;
 983                 if (!crc_is_compressed(crc)) {
 984                         dst_len = min(dst->bi_iter.bi_size, src->bi_iter.bi_size);
 985                         dst_len = min_t(unsigned, dst_len, wp->sectors_free << 9);
 986
 987                         if (op->csum_type)
 988                                 dst_len = min_t(unsigned, dst_len,
 989                                                 c->opts.encoded_extent_max);
 990
 991                         if (bounce) {
 992                                 swap(dst->bi_iter.bi_size, dst_len);
 993                                 bio_copy_data(dst, src);
 994                                 swap(dst->bi_iter.bi_size, dst_len);
 995                         }
 996
 997                         src_len = dst_len;
 998                 }
 999
1000                 BUG_ON(!src_len || !dst_len);
1001
1002                 if (bch2_csum_type_is_encryption(op->csum_type)) {
1003                         if (bversion_zero(version)) {
1004                                 version.lo = atomic64_inc_return(&c->key_version);
1005                         } else {
1006                                 crc.nonce = op->nonce;
1007                                 op->nonce += src_len >> 9;
1008                         }
1009                 }
1010
1011                 if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1012                     !crc_is_compressed(crc) &&
1013                     bch2_csum_type_is_encryption(op->crc.csum_type) ==
1014                     bch2_csum_type_is_encryption(op->csum_type)) {
1015                         u8 compression_type = crc.compression_type;
1016                         u16 nonce = crc.nonce;
1017                         /*
1018                          * Note: when we're using rechecksum(), we need to be
1019                          * checksumming @src because it has all the data our
1020                          * existing checksum covers - if we bounced (because we
1021                          * were trying to compress), @dst will only have the
1022                          * part of the data the new checksum will cover.
1023                          *
1024                          * But normally we want to be checksumming post bounce,
1025                          * because part of the reason for bouncing is so the
1026                          * data can't be modified (by userspace) while it's in
1027                          * flight.
1028                          */
1029                         if (bch2_rechecksum_bio(c, src, version, op->crc,
1030                                         &crc, &op->crc,
1031                                         src_len >> 9,
1032                                         bio_sectors(src) - (src_len >> 9),
1033                                         op->csum_type))
1034                                 goto csum_err;
1035                         /*
1036                          * rchecksum_bio sets compression_type on crc from op->crc,
1037                          * this isn't always correct as sometimes we're changing
1038                          * an extent from uncompressed to incompressible.
1039                          */
1040                         crc.compression_type = compression_type;
1041                         crc.nonce = nonce;
1042                 } else {
1043                         if ((op->flags & BCH_WRITE_DATA_ENCODED) &&
1044                             bch2_rechecksum_bio(c, src, version, op->crc,
1045                                         NULL, &op->crc,
1046                                         src_len >> 9,
1047                                         bio_sectors(src) - (src_len >> 9),
1048                                         op->crc.csum_type))
1049                                 goto csum_err;
1050
1051                         crc.compressed_size     = dst_len >> 9;
1052                         crc.uncompressed_size   = src_len >> 9;
1053                         crc.live_size           = src_len >> 9;
1054
1055                         swap(dst->bi_iter.bi_size, dst_len);
1056                         ret = bch2_encrypt_bio(c, op->csum_type,
1057                                                extent_nonce(version, crc), dst);
1058                         if (ret)
1059                                 goto err;
1060
1061                         crc.csum = bch2_checksum_bio(c, op->csum_type,
1062                                          extent_nonce(version, crc), dst);
1063                         crc.csum_type = op->csum_type;
1064                         swap(dst->bi_iter.bi_size, dst_len);
1065                 }
1066
1067                 init_append_extent(op, wp, version, crc);
1068
1069                 if (dst != src)
1070                         bio_advance(dst, dst_len);
1071                 bio_advance(src, src_len);
1072                 total_output    += dst_len;
1073                 total_input     += src_len;
1074         } while (dst->bi_iter.bi_size &&
1075                  src->bi_iter.bi_size &&
1076                  wp->sectors_free &&
1077                  !bch2_keylist_realloc(&op->insert_keys,
1078                                       op->inline_keys,
1079                                       ARRAY_SIZE(op->inline_keys),
1080                                       BKEY_EXTENT_U64s_MAX));
1081
1082         more = src->bi_iter.bi_size != 0;
1083
1084         dst->bi_iter = saved_iter;
1085
1086         if (dst == src && more) {
1087                 BUG_ON(total_output != total_input);
1088
1089                 dst = bio_split(src, total_input >> 9,
1090                                 GFP_NOIO, &c->bio_write);
1091                 wbio_init(dst)->put_bio = true;
1092                 /* copy WRITE_SYNC flag */
1093                 dst->bi_opf             = src->bi_opf;
1094         }
1095
1096         dst->bi_iter.bi_size = total_output;
1097 do_write:
1098         *_dst = dst;
1099         return more;
1100 csum_err:
1101         bch_err(c, "error verifying existing checksum while rewriting existing data (memory corruption?)");
1102         ret = -EIO;
1103 err:
1104         if (to_wbio(dst)->bounce)
1105                 bch2_bio_free_pages_pool(c, dst);
1106         if (to_wbio(dst)->put_bio)
1107                 bio_put(dst);
1108
1109         return ret;
1110 }
1111
1112 static void __bch2_write(struct closure *cl)
1113 {
1114         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1115         struct bch_fs *c = op->c;
1116         struct write_point *wp;
1117         struct bio *bio = NULL;
1118         bool skip_put = true;
1119         unsigned nofs_flags;
1120         int ret;
1121
1122         nofs_flags = memalloc_nofs_save();
1123 again:
1124         memset(&op->failed, 0, sizeof(op->failed));
1125
1126         do {
1127                 struct bkey_i *key_to_write;
1128                 unsigned key_to_write_offset = op->insert_keys.top_p -
1129                         op->insert_keys.keys_p;
1130
1131                 /* +1 for possible cache device: */
1132                 if (op->open_buckets.nr + op->nr_replicas + 1 >
1133                     ARRAY_SIZE(op->open_buckets.v))
1134                         goto flush_io;
1135
1136                 if (bch2_keylist_realloc(&op->insert_keys,
1137                                         op->inline_keys,
1138                                         ARRAY_SIZE(op->inline_keys),
1139                                         BKEY_EXTENT_U64s_MAX))
1140                         goto flush_io;
1141
1142                 /*
1143                  * The copygc thread is now global, which means it's no longer
1144                  * freeing up space on specific disks, which means that
1145                  * allocations for specific disks may hang arbitrarily long:
1146                  */
1147                 wp = bch2_alloc_sectors_start(c,
1148                         op->target,
1149                         op->opts.erasure_code && !(op->flags & BCH_WRITE_CACHED),
1150                         op->write_point,
1151                         &op->devs_have,
1152                         op->nr_replicas,
1153                         op->nr_replicas_required,
1154                         op->alloc_reserve,
1155                         op->flags,
1156                         (op->flags & (BCH_WRITE_ALLOC_NOWAIT|
1157                                       BCH_WRITE_ONLY_SPECIFIED_DEVS)) ? NULL : cl);
1158                 EBUG_ON(!wp);
1159
1160                 if (IS_ERR(wp)) {
1161                         if (unlikely(wp != ERR_PTR(-EAGAIN))) {
1162                                 ret = PTR_ERR(wp);
1163                                 goto err;
1164                         }
1165
1166                         goto flush_io;
1167                 }
1168
1169                 /*
1170                  * It's possible for the allocator to fail, put us on the
1171                  * freelist waitlist, and then succeed in one of various retry
1172                  * paths: if that happens, we need to disable the skip_put
1173                  * optimization because otherwise there won't necessarily be a
1174                  * barrier before we free the bch_write_op:
1175                  */
1176                 if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
1177                         skip_put = false;
1178
1179                 bch2_open_bucket_get(c, wp, &op->open_buckets);
1180                 ret = bch2_write_extent(op, wp, &bio);
1181                 bch2_alloc_sectors_done(c, wp);
1182
1183                 if (ret < 0)
1184                         goto err;
1185
1186                 if (ret) {
1187                         skip_put = false;
1188                 } else {
1189                         /*
1190                          * for the skip_put optimization this has to be set
1191                          * before we submit the bio:
1192                          */
1193                         op->flags |= BCH_WRITE_DONE;
1194                 }
1195
1196                 bio->bi_end_io  = bch2_write_endio;
1197                 bio->bi_private = &op->cl;
1198                 bio->bi_opf |= REQ_OP_WRITE;
1199
1200                 if (!skip_put)
1201                         closure_get(bio->bi_private);
1202                 else
1203                         op->flags |= BCH_WRITE_SKIP_CLOSURE_PUT;
1204
1205                 key_to_write = (void *) (op->insert_keys.keys_p +
1206                                          key_to_write_offset);
1207
1208                 bch2_submit_wbio_replicas(to_wbio(bio), c, BCH_DATA_user,
1209                                           key_to_write);
1210         } while (ret);
1211
1212         if (!skip_put)
1213                 continue_at(cl, bch2_write_index, index_update_wq(op));
1214 out:
1215         memalloc_nofs_restore(nofs_flags);
1216         return;
1217 err:
1218         op->error = ret;
1219         op->flags |= BCH_WRITE_DONE;
1220
1221         continue_at(cl, bch2_write_index, index_update_wq(op));
1222         goto out;
1223 flush_io:
1224         /*
1225          * If the write can't all be submitted at once, we generally want to
1226          * block synchronously as that signals backpressure to the caller.
1227          *
1228          * However, if we're running out of a workqueue, we can't block here
1229          * because we'll be blocking other work items from completing:
1230          */
1231         if (current->flags & PF_WQ_WORKER) {
1232                 continue_at(cl, bch2_write_index, index_update_wq(op));
1233                 goto out;
1234         }
1235
1236         closure_sync(cl);
1237
1238         if (!bch2_keylist_empty(&op->insert_keys)) {
1239                 __bch2_write_index(op);
1240
1241                 if (op->error) {
1242                         op->flags |= BCH_WRITE_DONE;
1243                         continue_at_nobarrier(cl, bch2_write_done, NULL);
1244                         goto out;
1245                 }
1246         }
1247
1248         goto again;
1249 }
1250
1251 static void bch2_write_data_inline(struct bch_write_op *op, unsigned data_len)
1252 {
1253         struct closure *cl = &op->cl;
1254         struct bio *bio = &op->wbio.bio;
1255         struct bvec_iter iter;
1256         struct bkey_i_inline_data *id;
1257         unsigned sectors;
1258         int ret;
1259
1260         bch2_check_set_feature(op->c, BCH_FEATURE_inline_data);
1261
1262         ret = bch2_keylist_realloc(&op->insert_keys, op->inline_keys,
1263                                    ARRAY_SIZE(op->inline_keys),
1264                                    BKEY_U64s + DIV_ROUND_UP(data_len, 8));
1265         if (ret) {
1266                 op->error = ret;
1267                 goto err;
1268         }
1269
1270         sectors = bio_sectors(bio);
1271         op->pos.offset += sectors;
1272
1273         id = bkey_inline_data_init(op->insert_keys.top);
1274         id->k.p         = op->pos;
1275         id->k.version   = op->version;
1276         id->k.size      = sectors;
1277
1278         iter = bio->bi_iter;
1279         iter.bi_size = data_len;
1280         memcpy_from_bio(id->v.data, bio, iter);
1281
1282         while (data_len & 7)
1283                 id->v.data[data_len++] = '\0';
1284         set_bkey_val_bytes(&id->k, data_len);
1285         bch2_keylist_push(&op->insert_keys);
1286
1287         op->flags |= BCH_WRITE_WROTE_DATA_INLINE;
1288         op->flags |= BCH_WRITE_DONE;
1289
1290         continue_at_nobarrier(cl, bch2_write_index, NULL);
1291         return;
1292 err:
1293         bch2_write_done(&op->cl);
1294 }
1295
1296 /**
1297  * bch_write - handle a write to a cache device or flash only volume
1298  *
1299  * This is the starting point for any data to end up in a cache device; it could
1300  * be from a normal write, or a writeback write, or a write to a flash only
1301  * volume - it's also used by the moving garbage collector to compact data in
1302  * mostly empty buckets.
1303  *
1304  * It first writes the data to the cache, creating a list of keys to be inserted
1305  * (if the data won't fit in a single open bucket, there will be multiple keys);
1306  * after the data is written it calls bch_journal, and after the keys have been
1307  * added to the next journal write they're inserted into the btree.
1308  *
1309  * If op->discard is true, instead of inserting the data it invalidates the
1310  * region of the cache represented by op->bio and op->inode.
1311  */
1312 void bch2_write(struct closure *cl)
1313 {
1314         struct bch_write_op *op = container_of(cl, struct bch_write_op, cl);
1315         struct bio *bio = &op->wbio.bio;
1316         struct bch_fs *c = op->c;
1317         unsigned data_len;
1318
1319         BUG_ON(!op->nr_replicas);
1320         BUG_ON(!op->write_point.v);
1321         BUG_ON(!bkey_cmp(op->pos, POS_MAX));
1322
1323         op->start_time = local_clock();
1324         bch2_keylist_init(&op->insert_keys, op->inline_keys);
1325         wbio_init(bio)->put_bio = false;
1326
1327         if (bio->bi_iter.bi_size & (c->opts.block_size - 1)) {
1328                 bch_err_inum_ratelimited(c, op->pos.inode,
1329                                          "misaligned write");
1330                 op->error = -EIO;
1331                 goto err;
1332         }
1333
1334         if (c->opts.nochanges ||
1335             !percpu_ref_tryget_live(&c->writes)) {
1336                 op->error = -EROFS;
1337                 goto err;
1338         }
1339
1340         this_cpu_add(c->counters[BCH_COUNTER_io_write], bio_sectors(bio));
1341         bch2_increment_clock(c, bio_sectors(bio), WRITE);
1342
1343         data_len = min_t(u64, bio->bi_iter.bi_size,
1344                          op->new_i_size - (op->pos.offset << 9));
1345
1346         if (c->opts.inline_data &&
1347             data_len <= min(block_bytes(c) / 2, 1024U)) {
1348                 bch2_write_data_inline(op, data_len);
1349                 return;
1350         }
1351
1352         continue_at_nobarrier(cl, __bch2_write, NULL);
1353         return;
1354 err:
1355         bch2_disk_reservation_put(c, &op->res);
1356
1357         if (op->end_io) {
1358                 EBUG_ON(cl->parent);
1359                 closure_debug_destroy(cl);
1360                 op->end_io(op);
1361         } else {
1362                 closure_return(cl);
1363         }
1364 }
1365
1366 /* Cache promotion on read */
1367
1368 struct promote_op {
1369         struct closure          cl;
1370         struct rcu_head         rcu;
1371         u64                     start_time;
1372
1373         struct rhash_head       hash;
1374         struct bpos             pos;
1375
1376         struct data_update      write;
1377         struct bio_vec          bi_inline_vecs[0]; /* must be last */
1378 };
1379
1380 static const struct rhashtable_params bch_promote_params = {
1381         .head_offset    = offsetof(struct promote_op, hash),
1382         .key_offset     = offsetof(struct promote_op, pos),
1383         .key_len        = sizeof(struct bpos),
1384 };
1385
1386 static inline bool should_promote(struct bch_fs *c, struct bkey_s_c k,
1387                                   struct bpos pos,
1388                                   struct bch_io_opts opts,
1389                                   unsigned flags)
1390 {
1391         if (!(flags & BCH_READ_MAY_PROMOTE))
1392                 return false;
1393
1394         if (!opts.promote_target)
1395                 return false;
1396
1397         if (bch2_bkey_has_target(c, k, opts.promote_target))
1398                 return false;
1399
1400         if (bch2_target_congested(c, opts.promote_target)) {
1401                 /* XXX trace this */
1402                 return false;
1403         }
1404
1405         if (rhashtable_lookup_fast(&c->promote_table, &pos,
1406                                    bch_promote_params))
1407                 return false;
1408
1409         return true;
1410 }
1411
1412 static void promote_free(struct bch_fs *c, struct promote_op *op)
1413 {
1414         int ret;
1415
1416         ret = rhashtable_remove_fast(&c->promote_table, &op->hash,
1417                                      bch_promote_params);
1418         BUG_ON(ret);
1419         percpu_ref_put(&c->writes);
1420         kfree_rcu(op, rcu);
1421 }
1422
1423 static void promote_done(struct closure *cl)
1424 {
1425         struct promote_op *op =
1426                 container_of(cl, struct promote_op, cl);
1427         struct bch_fs *c = op->write.op.c;
1428
1429         bch2_time_stats_update(&c->times[BCH_TIME_data_promote],
1430                                op->start_time);
1431
1432         bch2_data_update_exit(&op->write);
1433         promote_free(c, op);
1434 }
1435
1436 static void promote_start(struct promote_op *op, struct bch_read_bio *rbio)
1437 {
1438         struct closure *cl = &op->cl;
1439         struct bio *bio = &op->write.op.wbio.bio;
1440
1441         trace_and_count(op->write.op.c, read_promote, &rbio->bio);
1442
1443         /* we now own pages: */
1444         BUG_ON(!rbio->bounce);
1445         BUG_ON(rbio->bio.bi_vcnt > bio->bi_max_vecs);
1446
1447         memcpy(bio->bi_io_vec, rbio->bio.bi_io_vec,
1448                sizeof(struct bio_vec) * rbio->bio.bi_vcnt);
1449         swap(bio->bi_vcnt, rbio->bio.bi_vcnt);
1450
1451         closure_init(cl, NULL);
1452         bch2_data_update_read_done(&op->write, rbio->pick.crc, cl);
1453         closure_return_with_destructor(cl, promote_done);
1454 }
1455
1456 static struct promote_op *__promote_alloc(struct bch_fs *c,
1457                                           enum btree_id btree_id,
1458                                           struct bkey_s_c k,
1459                                           struct bpos pos,
1460                                           struct extent_ptr_decoded *pick,
1461                                           struct bch_io_opts opts,
1462                                           unsigned sectors,
1463                                           struct bch_read_bio **rbio)
1464 {
1465         struct promote_op *op = NULL;
1466         struct bio *bio;
1467         unsigned pages = DIV_ROUND_UP(sectors, PAGE_SECTORS);
1468         int ret;
1469
1470         if (!percpu_ref_tryget_live(&c->writes))
1471                 return NULL;
1472
1473         op = kzalloc(sizeof(*op) + sizeof(struct bio_vec) * pages, GFP_NOIO);
1474         if (!op)
1475                 goto err;
1476
1477         op->start_time = local_clock();
1478         op->pos = pos;
1479
1480         /*
1481          * We don't use the mempool here because extents that aren't
1482          * checksummed or compressed can be too big for the mempool:
1483          */
1484         *rbio = kzalloc(sizeof(struct bch_read_bio) +
1485                         sizeof(struct bio_vec) * pages,
1486                         GFP_NOIO);
1487         if (!*rbio)
1488                 goto err;
1489
1490         rbio_init(&(*rbio)->bio, opts);
1491         bio_init(&(*rbio)->bio, NULL, (*rbio)->bio.bi_inline_vecs, pages, 0);
1492
1493         if (bch2_bio_alloc_pages(&(*rbio)->bio, sectors << 9,
1494                                  GFP_NOIO))
1495                 goto err;
1496
1497         (*rbio)->bounce         = true;
1498         (*rbio)->split          = true;
1499         (*rbio)->kmalloc        = true;
1500
1501         if (rhashtable_lookup_insert_fast(&c->promote_table, &op->hash,
1502                                           bch_promote_params))
1503                 goto err;
1504
1505         bio = &op->write.op.wbio.bio;
1506         bio_init(bio, NULL, bio->bi_inline_vecs, pages, 0);
1507
1508         ret = bch2_data_update_init(c, &op->write,
1509                         writepoint_hashed((unsigned long) current),
1510                         opts,
1511                         (struct data_update_opts) {
1512                                 .target         = opts.promote_target,
1513                                 .extra_replicas = 1,
1514                                 .write_flags    = BCH_WRITE_ALLOC_NOWAIT|BCH_WRITE_CACHED,
1515                         },
1516                         btree_id, k);
1517         BUG_ON(ret);
1518
1519         return op;
1520 err:
1521         if (*rbio)
1522                 bio_free_pages(&(*rbio)->bio);
1523         kfree(*rbio);
1524         *rbio = NULL;
1525         kfree(op);
1526         percpu_ref_put(&c->writes);
1527         return NULL;
1528 }
1529
1530 noinline
1531 static struct promote_op *promote_alloc(struct bch_fs *c,
1532                                                struct bvec_iter iter,
1533                                                struct bkey_s_c k,
1534                                                struct extent_ptr_decoded *pick,
1535                                                struct bch_io_opts opts,
1536                                                unsigned flags,
1537                                                struct bch_read_bio **rbio,
1538                                                bool *bounce,
1539                                                bool *read_full)
1540 {
1541         bool promote_full = *read_full || READ_ONCE(c->promote_whole_extents);
1542         /* data might have to be decompressed in the write path: */
1543         unsigned sectors = promote_full
1544                 ? max(pick->crc.compressed_size, pick->crc.live_size)
1545                 : bvec_iter_sectors(iter);
1546         struct bpos pos = promote_full
1547                 ? bkey_start_pos(k.k)
1548                 : POS(k.k->p.inode, iter.bi_sector);
1549         struct promote_op *promote;
1550
1551         if (!should_promote(c, k, pos, opts, flags))
1552                 return NULL;
1553
1554         promote = __promote_alloc(c,
1555                                   k.k->type == KEY_TYPE_reflink_v
1556                                   ? BTREE_ID_reflink
1557                                   : BTREE_ID_extents,
1558                                   k, pos, pick, opts, sectors, rbio);
1559         if (!promote)
1560                 return NULL;
1561
1562         *bounce         = true;
1563         *read_full      = promote_full;
1564         return promote;
1565 }
1566
1567 /* Read */
1568
1569 #define READ_RETRY_AVOID        1
1570 #define READ_RETRY              2
1571 #define READ_ERR                3
1572
1573 enum rbio_context {
1574         RBIO_CONTEXT_NULL,
1575         RBIO_CONTEXT_HIGHPRI,
1576         RBIO_CONTEXT_UNBOUND,
1577 };
1578
1579 static inline struct bch_read_bio *
1580 bch2_rbio_parent(struct bch_read_bio *rbio)
1581 {
1582         return rbio->split ? rbio->parent : rbio;
1583 }
1584
1585 __always_inline
1586 static void bch2_rbio_punt(struct bch_read_bio *rbio, work_func_t fn,
1587                            enum rbio_context context,
1588                            struct workqueue_struct *wq)
1589 {
1590         if (context <= rbio->context) {
1591                 fn(&rbio->work);
1592         } else {
1593                 rbio->work.func         = fn;
1594                 rbio->context           = context;
1595                 queue_work(wq, &rbio->work);
1596         }
1597 }
1598
1599 static inline struct bch_read_bio *bch2_rbio_free(struct bch_read_bio *rbio)
1600 {
1601         BUG_ON(rbio->bounce && !rbio->split);
1602
1603         if (rbio->promote)
1604                 promote_free(rbio->c, rbio->promote);
1605         rbio->promote = NULL;
1606
1607         if (rbio->bounce)
1608                 bch2_bio_free_pages_pool(rbio->c, &rbio->bio);
1609
1610         if (rbio->split) {
1611                 struct bch_read_bio *parent = rbio->parent;
1612
1613                 if (rbio->kmalloc)
1614                         kfree(rbio);
1615                 else
1616                         bio_put(&rbio->bio);
1617
1618                 rbio = parent;
1619         }
1620
1621         return rbio;
1622 }
1623
1624 /*
1625  * Only called on a top level bch_read_bio to complete an entire read request,
1626  * not a split:
1627  */
1628 static void bch2_rbio_done(struct bch_read_bio *rbio)
1629 {
1630         if (rbio->start_time)
1631                 bch2_time_stats_update(&rbio->c->times[BCH_TIME_data_read],
1632                                        rbio->start_time);
1633         bio_endio(&rbio->bio);
1634 }
1635
1636 static void bch2_read_retry_nodecode(struct bch_fs *c, struct bch_read_bio *rbio,
1637                                      struct bvec_iter bvec_iter,
1638                                      struct bch_io_failures *failed,
1639                                      unsigned flags)
1640 {
1641         struct btree_trans trans;
1642         struct btree_iter iter;
1643         struct bkey_buf sk;
1644         struct bkey_s_c k;
1645         int ret;
1646
1647         flags &= ~BCH_READ_LAST_FRAGMENT;
1648         flags |= BCH_READ_MUST_CLONE;
1649
1650         bch2_bkey_buf_init(&sk);
1651         bch2_trans_init(&trans, c, 0, 0);
1652
1653         bch2_trans_iter_init(&trans, &iter, rbio->data_btree,
1654                              rbio->read_pos, BTREE_ITER_SLOTS);
1655 retry:
1656         rbio->bio.bi_status = 0;
1657
1658         k = bch2_btree_iter_peek_slot(&iter);
1659         if (bkey_err(k))
1660                 goto err;
1661
1662         bch2_bkey_buf_reassemble(&sk, c, k);
1663         k = bkey_i_to_s_c(sk.k);
1664         bch2_trans_unlock(&trans);
1665
1666         if (!bch2_bkey_matches_ptr(c, k,
1667                                    rbio->pick.ptr,
1668                                    rbio->data_pos.offset -
1669                                    rbio->pick.crc.offset)) {
1670                 /* extent we wanted to read no longer exists: */
1671                 rbio->hole = true;
1672                 goto out;
1673         }
1674
1675         ret = __bch2_read_extent(&trans, rbio, bvec_iter,
1676                                  rbio->read_pos,
1677                                  rbio->data_btree,
1678                                  k, 0, failed, flags);
1679         if (ret == READ_RETRY)
1680                 goto retry;
1681         if (ret)
1682                 goto err;
1683 out:
1684         bch2_rbio_done(rbio);
1685         bch2_trans_iter_exit(&trans, &iter);
1686         bch2_trans_exit(&trans);
1687         bch2_bkey_buf_exit(&sk, c);
1688         return;
1689 err:
1690         rbio->bio.bi_status = BLK_STS_IOERR;
1691         goto out;
1692 }
1693
1694 static void bch2_rbio_retry(struct work_struct *work)
1695 {
1696         struct bch_read_bio *rbio =
1697                 container_of(work, struct bch_read_bio, work);
1698         struct bch_fs *c        = rbio->c;
1699         struct bvec_iter iter   = rbio->bvec_iter;
1700         unsigned flags          = rbio->flags;
1701         subvol_inum inum = {
1702                 .subvol = rbio->subvol,
1703                 .inum   = rbio->read_pos.inode,
1704         };
1705         struct bch_io_failures failed = { .nr = 0 };
1706
1707         trace_and_count(c, read_retry, &rbio->bio);
1708
1709         if (rbio->retry == READ_RETRY_AVOID)
1710                 bch2_mark_io_failure(&failed, &rbio->pick);
1711
1712         rbio->bio.bi_status = 0;
1713
1714         rbio = bch2_rbio_free(rbio);
1715
1716         flags |= BCH_READ_IN_RETRY;
1717         flags &= ~BCH_READ_MAY_PROMOTE;
1718
1719         if (flags & BCH_READ_NODECODE) {
1720                 bch2_read_retry_nodecode(c, rbio, iter, &failed, flags);
1721         } else {
1722                 flags &= ~BCH_READ_LAST_FRAGMENT;
1723                 flags |= BCH_READ_MUST_CLONE;
1724
1725                 __bch2_read(c, rbio, iter, inum, &failed, flags);
1726         }
1727 }
1728
1729 static void bch2_rbio_error(struct bch_read_bio *rbio, int retry,
1730                             blk_status_t error)
1731 {
1732         rbio->retry = retry;
1733
1734         if (rbio->flags & BCH_READ_IN_RETRY)
1735                 return;
1736
1737         if (retry == READ_ERR) {
1738                 rbio = bch2_rbio_free(rbio);
1739
1740                 rbio->bio.bi_status = error;
1741                 bch2_rbio_done(rbio);
1742         } else {
1743                 bch2_rbio_punt(rbio, bch2_rbio_retry,
1744                                RBIO_CONTEXT_UNBOUND, system_unbound_wq);
1745         }
1746 }
1747
1748 static int __bch2_rbio_narrow_crcs(struct btree_trans *trans,
1749                                    struct bch_read_bio *rbio)
1750 {
1751         struct bch_fs *c = rbio->c;
1752         u64 data_offset = rbio->data_pos.offset - rbio->pick.crc.offset;
1753         struct bch_extent_crc_unpacked new_crc;
1754         struct btree_iter iter;
1755         struct bkey_i *new;
1756         struct bkey_s_c k;
1757         int ret = 0;
1758
1759         if (crc_is_compressed(rbio->pick.crc))
1760                 return 0;
1761
1762         bch2_trans_iter_init(trans, &iter, rbio->data_btree, rbio->data_pos,
1763                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1764         k = bch2_btree_iter_peek_slot(&iter);
1765         if ((ret = bkey_err(k)))
1766                 goto out;
1767
1768         if (bversion_cmp(k.k->version, rbio->version) ||
1769             !bch2_bkey_matches_ptr(c, k, rbio->pick.ptr, data_offset))
1770                 goto out;
1771
1772         /* Extent was merged? */
1773         if (bkey_start_offset(k.k) < data_offset ||
1774             k.k->p.offset > data_offset + rbio->pick.crc.uncompressed_size)
1775                 goto out;
1776
1777         if (bch2_rechecksum_bio(c, &rbio->bio, rbio->version,
1778                         rbio->pick.crc, NULL, &new_crc,
1779                         bkey_start_offset(k.k) - data_offset, k.k->size,
1780                         rbio->pick.crc.csum_type)) {
1781                 bch_err(c, "error verifying existing checksum while narrowing checksum (memory corruption?)");
1782                 ret = 0;
1783                 goto out;
1784         }
1785
1786         /*
1787          * going to be temporarily appending another checksum entry:
1788          */
1789         new = bch2_trans_kmalloc(trans, bkey_bytes(k.k) +
1790                                  sizeof(struct bch_extent_crc128));
1791         if ((ret = PTR_ERR_OR_ZERO(new)))
1792                 goto out;
1793
1794         bkey_reassemble(new, k);
1795
1796         if (!bch2_bkey_narrow_crcs(new, new_crc))
1797                 goto out;
1798
1799         ret = bch2_trans_update(trans, &iter, new,
1800                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1801 out:
1802         bch2_trans_iter_exit(trans, &iter);
1803         return ret;
1804 }
1805
1806 static noinline void bch2_rbio_narrow_crcs(struct bch_read_bio *rbio)
1807 {
1808         bch2_trans_do(rbio->c, NULL, NULL, BTREE_INSERT_NOFAIL,
1809                       __bch2_rbio_narrow_crcs(&trans, rbio));
1810 }
1811
1812 /* Inner part that may run in process context */
1813 static void __bch2_read_endio(struct work_struct *work)
1814 {
1815         struct bch_read_bio *rbio =
1816                 container_of(work, struct bch_read_bio, work);
1817         struct bch_fs *c        = rbio->c;
1818         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1819         struct bio *src         = &rbio->bio;
1820         struct bio *dst         = &bch2_rbio_parent(rbio)->bio;
1821         struct bvec_iter dst_iter = rbio->bvec_iter;
1822         struct bch_extent_crc_unpacked crc = rbio->pick.crc;
1823         struct nonce nonce = extent_nonce(rbio->version, crc);
1824         unsigned nofs_flags;
1825         struct bch_csum csum;
1826         int ret;
1827
1828         nofs_flags = memalloc_nofs_save();
1829
1830         /* Reset iterator for checksumming and copying bounced data: */
1831         if (rbio->bounce) {
1832                 src->bi_iter.bi_size            = crc.compressed_size << 9;
1833                 src->bi_iter.bi_idx             = 0;
1834                 src->bi_iter.bi_bvec_done       = 0;
1835         } else {
1836                 src->bi_iter                    = rbio->bvec_iter;
1837         }
1838
1839         csum = bch2_checksum_bio(c, crc.csum_type, nonce, src);
1840         if (bch2_crc_cmp(csum, rbio->pick.crc.csum))
1841                 goto csum_err;
1842
1843         /*
1844          * XXX
1845          * We need to rework the narrow_crcs path to deliver the read completion
1846          * first, and then punt to a different workqueue, otherwise we're
1847          * holding up reads while doing btree updates which is bad for memory
1848          * reclaim.
1849          */
1850         if (unlikely(rbio->narrow_crcs))
1851                 bch2_rbio_narrow_crcs(rbio);
1852
1853         if (rbio->flags & BCH_READ_NODECODE)
1854                 goto nodecode;
1855
1856         /* Adjust crc to point to subset of data we want: */
1857         crc.offset     += rbio->offset_into_extent;
1858         crc.live_size   = bvec_iter_sectors(rbio->bvec_iter);
1859
1860         if (crc_is_compressed(crc)) {
1861                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1862                 if (ret)
1863                         goto decrypt_err;
1864
1865                 if (bch2_bio_uncompress(c, src, dst, dst_iter, crc))
1866                         goto decompression_err;
1867         } else {
1868                 /* don't need to decrypt the entire bio: */
1869                 nonce = nonce_add(nonce, crc.offset << 9);
1870                 bio_advance(src, crc.offset << 9);
1871
1872                 BUG_ON(src->bi_iter.bi_size < dst_iter.bi_size);
1873                 src->bi_iter.bi_size = dst_iter.bi_size;
1874
1875                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1876                 if (ret)
1877                         goto decrypt_err;
1878
1879                 if (rbio->bounce) {
1880                         struct bvec_iter src_iter = src->bi_iter;
1881                         bio_copy_data_iter(dst, &dst_iter, src, &src_iter);
1882                 }
1883         }
1884
1885         if (rbio->promote) {
1886                 /*
1887                  * Re encrypt data we decrypted, so it's consistent with
1888                  * rbio->crc:
1889                  */
1890                 ret = bch2_encrypt_bio(c, crc.csum_type, nonce, src);
1891                 if (ret)
1892                         goto decrypt_err;
1893
1894                 promote_start(rbio->promote, rbio);
1895                 rbio->promote = NULL;
1896         }
1897 nodecode:
1898         if (likely(!(rbio->flags & BCH_READ_IN_RETRY))) {
1899                 rbio = bch2_rbio_free(rbio);
1900                 bch2_rbio_done(rbio);
1901         }
1902 out:
1903         memalloc_nofs_restore(nofs_flags);
1904         return;
1905 csum_err:
1906         /*
1907          * Checksum error: if the bio wasn't bounced, we may have been
1908          * reading into buffers owned by userspace (that userspace can
1909          * scribble over) - retry the read, bouncing it this time:
1910          */
1911         if (!rbio->bounce && (rbio->flags & BCH_READ_USER_MAPPED)) {
1912                 rbio->flags |= BCH_READ_MUST_BOUNCE;
1913                 bch2_rbio_error(rbio, READ_RETRY, BLK_STS_IOERR);
1914                 goto out;
1915         }
1916
1917         bch2_dev_inum_io_error(ca, rbio->read_pos.inode, (u64) rbio->bvec_iter.bi_sector,
1918                 "data checksum error: expected %0llx:%0llx got %0llx:%0llx (type %s)",
1919                 rbio->pick.crc.csum.hi, rbio->pick.crc.csum.lo,
1920                 csum.hi, csum.lo, bch2_csum_types[crc.csum_type]);
1921         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
1922         goto out;
1923 decompression_err:
1924         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1925                                  "decompression error");
1926         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1927         goto out;
1928 decrypt_err:
1929         bch_err_inum_ratelimited(c, rbio->read_pos.inode,
1930                                  "decrypt error");
1931         bch2_rbio_error(rbio, READ_ERR, BLK_STS_IOERR);
1932         goto out;
1933 }
1934
1935 static void bch2_read_endio(struct bio *bio)
1936 {
1937         struct bch_read_bio *rbio =
1938                 container_of(bio, struct bch_read_bio, bio);
1939         struct bch_fs *c        = rbio->c;
1940         struct bch_dev *ca      = bch_dev_bkey_exists(c, rbio->pick.ptr.dev);
1941         struct workqueue_struct *wq = NULL;
1942         enum rbio_context context = RBIO_CONTEXT_NULL;
1943
1944         if (rbio->have_ioref) {
1945                 bch2_latency_acct(ca, rbio->submit_time, READ);
1946                 percpu_ref_put(&ca->io_ref);
1947         }
1948
1949         if (!rbio->split)
1950                 rbio->bio.bi_end_io = rbio->end_io;
1951
1952         if (bch2_dev_inum_io_err_on(bio->bi_status, ca,
1953                                     rbio->read_pos.inode,
1954                                     rbio->read_pos.offset,
1955                                     "data read error: %s",
1956                                bch2_blk_status_to_str(bio->bi_status))) {
1957                 bch2_rbio_error(rbio, READ_RETRY_AVOID, bio->bi_status);
1958                 return;
1959         }
1960
1961         if (((rbio->flags & BCH_READ_RETRY_IF_STALE) && race_fault()) ||
1962             ptr_stale(ca, &rbio->pick.ptr)) {
1963                 trace_and_count(c, read_reuse_race, &rbio->bio);
1964
1965                 if (rbio->flags & BCH_READ_RETRY_IF_STALE)
1966                         bch2_rbio_error(rbio, READ_RETRY, BLK_STS_AGAIN);
1967                 else
1968                         bch2_rbio_error(rbio, READ_ERR, BLK_STS_AGAIN);
1969                 return;
1970         }
1971
1972         if (rbio->narrow_crcs ||
1973             rbio->promote ||
1974             crc_is_compressed(rbio->pick.crc) ||
1975             bch2_csum_type_is_encryption(rbio->pick.crc.csum_type))
1976                 context = RBIO_CONTEXT_UNBOUND, wq = system_unbound_wq;
1977         else if (rbio->pick.crc.csum_type)
1978                 context = RBIO_CONTEXT_HIGHPRI, wq = system_highpri_wq;
1979
1980         bch2_rbio_punt(rbio, __bch2_read_endio, context, wq);
1981 }
1982
1983 int __bch2_read_indirect_extent(struct btree_trans *trans,
1984                                 unsigned *offset_into_extent,
1985                                 struct bkey_buf *orig_k)
1986 {
1987         struct btree_iter iter;
1988         struct bkey_s_c k;
1989         u64 reflink_offset;
1990         int ret;
1991
1992         reflink_offset = le64_to_cpu(bkey_i_to_reflink_p(orig_k->k)->v.idx) +
1993                 *offset_into_extent;
1994
1995         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink,
1996                              POS(0, reflink_offset),
1997                              BTREE_ITER_SLOTS);
1998         k = bch2_btree_iter_peek_slot(&iter);
1999         ret = bkey_err(k);
2000         if (ret)
2001                 goto err;
2002
2003         if (k.k->type != KEY_TYPE_reflink_v &&
2004             k.k->type != KEY_TYPE_indirect_inline_data) {
2005                 bch_err_inum_ratelimited(trans->c, orig_k->k->k.p.inode,
2006                         "%llu len %u points to nonexistent indirect extent %llu",
2007                         orig_k->k->k.p.offset,
2008                         orig_k->k->k.size,
2009                         reflink_offset);
2010                 bch2_inconsistent_error(trans->c);
2011                 ret = -EIO;
2012                 goto err;
2013         }
2014
2015         *offset_into_extent = iter.pos.offset - bkey_start_offset(k.k);
2016         bch2_bkey_buf_reassemble(orig_k, trans->c, k);
2017 err:
2018         bch2_trans_iter_exit(trans, &iter);
2019         return ret;
2020 }
2021
2022 static noinline void read_from_stale_dirty_pointer(struct btree_trans *trans,
2023                                                    struct bkey_s_c k,
2024                                                    struct bch_extent_ptr ptr)
2025 {
2026         struct bch_fs *c = trans->c;
2027         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr.dev);
2028         struct btree_iter iter;
2029         struct printbuf buf = PRINTBUF;
2030         int ret;
2031
2032         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
2033                              PTR_BUCKET_POS(c, &ptr),
2034                              BTREE_ITER_CACHED);
2035
2036         prt_printf(&buf, "Attempting to read from stale dirty pointer:");
2037         printbuf_indent_add(&buf, 2);
2038         prt_newline(&buf);
2039
2040         bch2_bkey_val_to_text(&buf, c, k);
2041         prt_newline(&buf);
2042
2043         prt_printf(&buf, "memory gen: %u", *bucket_gen(ca, iter.pos.offset));
2044
2045         ret = lockrestart_do(trans, bkey_err(k = bch2_btree_iter_peek_slot(&iter)));
2046         if (!ret) {
2047                 prt_newline(&buf);
2048                 bch2_bkey_val_to_text(&buf, c, k);
2049         }
2050
2051         bch2_fs_inconsistent(c, "%s", buf.buf);
2052
2053         bch2_trans_iter_exit(trans, &iter);
2054         printbuf_exit(&buf);
2055 }
2056
2057 int __bch2_read_extent(struct btree_trans *trans, struct bch_read_bio *orig,
2058                        struct bvec_iter iter, struct bpos read_pos,
2059                        enum btree_id data_btree, struct bkey_s_c k,
2060                        unsigned offset_into_extent,
2061                        struct bch_io_failures *failed, unsigned flags)
2062 {
2063         struct bch_fs *c = trans->c;
2064         struct extent_ptr_decoded pick;
2065         struct bch_read_bio *rbio = NULL;
2066         struct bch_dev *ca = NULL;
2067         struct promote_op *promote = NULL;
2068         bool bounce = false, read_full = false, narrow_crcs = false;
2069         struct bpos data_pos = bkey_start_pos(k.k);
2070         int pick_ret;
2071
2072         if (bkey_extent_is_inline_data(k.k)) {
2073                 unsigned bytes = min_t(unsigned, iter.bi_size,
2074                                        bkey_inline_data_bytes(k.k));
2075
2076                 swap(iter.bi_size, bytes);
2077                 memcpy_to_bio(&orig->bio, iter, bkey_inline_data_p(k));
2078                 swap(iter.bi_size, bytes);
2079                 bio_advance_iter(&orig->bio, &iter, bytes);
2080                 zero_fill_bio_iter(&orig->bio, iter);
2081                 goto out_read_done;
2082         }
2083 retry_pick:
2084         pick_ret = bch2_bkey_pick_read_device(c, k, failed, &pick);
2085
2086         /* hole or reservation - just zero fill: */
2087         if (!pick_ret)
2088                 goto hole;
2089
2090         if (pick_ret < 0) {
2091                 bch_err_inum_ratelimited(c, k.k->p.inode,
2092                                          "no device to read from");
2093                 goto err;
2094         }
2095
2096         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
2097
2098         /*
2099          * Stale dirty pointers are treated as IO errors, but @failed isn't
2100          * allocated unless we're in the retry path - so if we're not in the
2101          * retry path, don't check here, it'll be caught in bch2_read_endio()
2102          * and we'll end up in the retry path:
2103          */
2104         if ((flags & BCH_READ_IN_RETRY) &&
2105             !pick.ptr.cached &&
2106             unlikely(ptr_stale(ca, &pick.ptr))) {
2107                 read_from_stale_dirty_pointer(trans, k, pick.ptr);
2108                 bch2_mark_io_failure(failed, &pick);
2109                 goto retry_pick;
2110         }
2111
2112         /*
2113          * Unlock the iterator while the btree node's lock is still in
2114          * cache, before doing the IO:
2115          */
2116         bch2_trans_unlock(trans);
2117
2118         if (flags & BCH_READ_NODECODE) {
2119                 /*
2120                  * can happen if we retry, and the extent we were going to read
2121                  * has been merged in the meantime:
2122                  */
2123                 if (pick.crc.compressed_size > orig->bio.bi_vcnt * PAGE_SECTORS)
2124                         goto hole;
2125
2126                 iter.bi_size    = pick.crc.compressed_size << 9;
2127                 goto get_bio;
2128         }
2129
2130         if (!(flags & BCH_READ_LAST_FRAGMENT) ||
2131             bio_flagged(&orig->bio, BIO_CHAIN))
2132                 flags |= BCH_READ_MUST_CLONE;
2133
2134         narrow_crcs = !(flags & BCH_READ_IN_RETRY) &&
2135                 bch2_can_narrow_extent_crcs(k, pick.crc);
2136
2137         if (narrow_crcs && (flags & BCH_READ_USER_MAPPED))
2138                 flags |= BCH_READ_MUST_BOUNCE;
2139
2140         EBUG_ON(offset_into_extent + bvec_iter_sectors(iter) > k.k->size);
2141
2142         if (crc_is_compressed(pick.crc) ||
2143             (pick.crc.csum_type != BCH_CSUM_none &&
2144              (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2145               (bch2_csum_type_is_encryption(pick.crc.csum_type) &&
2146                (flags & BCH_READ_USER_MAPPED)) ||
2147               (flags & BCH_READ_MUST_BOUNCE)))) {
2148                 read_full = true;
2149                 bounce = true;
2150         }
2151
2152         if (orig->opts.promote_target)
2153                 promote = promote_alloc(c, iter, k, &pick, orig->opts, flags,
2154                                         &rbio, &bounce, &read_full);
2155
2156         if (!read_full) {
2157                 EBUG_ON(crc_is_compressed(pick.crc));
2158                 EBUG_ON(pick.crc.csum_type &&
2159                         (bvec_iter_sectors(iter) != pick.crc.uncompressed_size ||
2160                          bvec_iter_sectors(iter) != pick.crc.live_size ||
2161                          pick.crc.offset ||
2162                          offset_into_extent));
2163
2164                 data_pos.offset += offset_into_extent;
2165                 pick.ptr.offset += pick.crc.offset +
2166                         offset_into_extent;
2167                 offset_into_extent              = 0;
2168                 pick.crc.compressed_size        = bvec_iter_sectors(iter);
2169                 pick.crc.uncompressed_size      = bvec_iter_sectors(iter);
2170                 pick.crc.offset                 = 0;
2171                 pick.crc.live_size              = bvec_iter_sectors(iter);
2172                 offset_into_extent              = 0;
2173         }
2174 get_bio:
2175         if (rbio) {
2176                 /*
2177                  * promote already allocated bounce rbio:
2178                  * promote needs to allocate a bio big enough for uncompressing
2179                  * data in the write path, but we're not going to use it all
2180                  * here:
2181                  */
2182                 EBUG_ON(rbio->bio.bi_iter.bi_size <
2183                        pick.crc.compressed_size << 9);
2184                 rbio->bio.bi_iter.bi_size =
2185                         pick.crc.compressed_size << 9;
2186         } else if (bounce) {
2187                 unsigned sectors = pick.crc.compressed_size;
2188
2189                 rbio = rbio_init(bio_alloc_bioset(NULL,
2190                                                   DIV_ROUND_UP(sectors, PAGE_SECTORS),
2191                                                   0,
2192                                                   GFP_NOIO,
2193                                                   &c->bio_read_split),
2194                                  orig->opts);
2195
2196                 bch2_bio_alloc_pages_pool(c, &rbio->bio, sectors << 9);
2197                 rbio->bounce    = true;
2198                 rbio->split     = true;
2199         } else if (flags & BCH_READ_MUST_CLONE) {
2200                 /*
2201                  * Have to clone if there were any splits, due to error
2202                  * reporting issues (if a split errored, and retrying didn't
2203                  * work, when it reports the error to its parent (us) we don't
2204                  * know if the error was from our bio, and we should retry, or
2205                  * from the whole bio, in which case we don't want to retry and
2206                  * lose the error)
2207                  */
2208                 rbio = rbio_init(bio_alloc_clone(NULL, &orig->bio, GFP_NOIO,
2209                                                  &c->bio_read_split),
2210                                  orig->opts);
2211                 rbio->bio.bi_iter = iter;
2212                 rbio->split     = true;
2213         } else {
2214                 rbio = orig;
2215                 rbio->bio.bi_iter = iter;
2216                 EBUG_ON(bio_flagged(&rbio->bio, BIO_CHAIN));
2217         }
2218
2219         EBUG_ON(bio_sectors(&rbio->bio) != pick.crc.compressed_size);
2220
2221         rbio->c                 = c;
2222         rbio->submit_time       = local_clock();
2223         if (rbio->split)
2224                 rbio->parent    = orig;
2225         else
2226                 rbio->end_io    = orig->bio.bi_end_io;
2227         rbio->bvec_iter         = iter;
2228         rbio->offset_into_extent= offset_into_extent;
2229         rbio->flags             = flags;
2230         rbio->have_ioref        = pick_ret > 0 && bch2_dev_get_ioref(ca, READ);
2231         rbio->narrow_crcs       = narrow_crcs;
2232         rbio->hole              = 0;
2233         rbio->retry             = 0;
2234         rbio->context           = 0;
2235         /* XXX: only initialize this if needed */
2236         rbio->devs_have         = bch2_bkey_devs(k);
2237         rbio->pick              = pick;
2238         rbio->subvol            = orig->subvol;
2239         rbio->read_pos          = read_pos;
2240         rbio->data_btree        = data_btree;
2241         rbio->data_pos          = data_pos;
2242         rbio->version           = k.k->version;
2243         rbio->promote           = promote;
2244         INIT_WORK(&rbio->work, NULL);
2245
2246         rbio->bio.bi_opf        = orig->bio.bi_opf;
2247         rbio->bio.bi_iter.bi_sector = pick.ptr.offset;
2248         rbio->bio.bi_end_io     = bch2_read_endio;
2249
2250         if (rbio->bounce)
2251                 trace_and_count(c, read_bounce, &rbio->bio);
2252
2253         this_cpu_add(c->counters[BCH_COUNTER_io_read], bio_sectors(&rbio->bio));
2254         bch2_increment_clock(c, bio_sectors(&rbio->bio), READ);
2255
2256         /*
2257          * If it's being moved internally, we don't want to flag it as a cache
2258          * hit:
2259          */
2260         if (pick.ptr.cached && !(flags & BCH_READ_NODECODE))
2261                 bch2_bucket_io_time_reset(trans, pick.ptr.dev,
2262                         PTR_BUCKET_NR(ca, &pick.ptr), READ);
2263
2264         if (!(flags & (BCH_READ_IN_RETRY|BCH_READ_LAST_FRAGMENT))) {
2265                 bio_inc_remaining(&orig->bio);
2266                 trace_and_count(c, read_split, &orig->bio);
2267         }
2268
2269         if (!rbio->pick.idx) {
2270                 if (!rbio->have_ioref) {
2271                         bch_err_inum_ratelimited(c, k.k->p.inode,
2272                                                  "no device to read from");
2273                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2274                         goto out;
2275                 }
2276
2277                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_user],
2278                              bio_sectors(&rbio->bio));
2279                 bio_set_dev(&rbio->bio, ca->disk_sb.bdev);
2280
2281                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2282                         submit_bio(&rbio->bio);
2283                 else
2284                         submit_bio_wait(&rbio->bio);
2285         } else {
2286                 /* Attempting reconstruct read: */
2287                 if (bch2_ec_read_extent(c, rbio)) {
2288                         bch2_rbio_error(rbio, READ_RETRY_AVOID, BLK_STS_IOERR);
2289                         goto out;
2290                 }
2291
2292                 if (likely(!(flags & BCH_READ_IN_RETRY)))
2293                         bio_endio(&rbio->bio);
2294         }
2295 out:
2296         if (likely(!(flags & BCH_READ_IN_RETRY))) {
2297                 return 0;
2298         } else {
2299                 int ret;
2300
2301                 rbio->context = RBIO_CONTEXT_UNBOUND;
2302                 bch2_read_endio(&rbio->bio);
2303
2304                 ret = rbio->retry;
2305                 rbio = bch2_rbio_free(rbio);
2306
2307                 if (ret == READ_RETRY_AVOID) {
2308                         bch2_mark_io_failure(failed, &pick);
2309                         ret = READ_RETRY;
2310                 }
2311
2312                 if (!ret)
2313                         goto out_read_done;
2314
2315                 return ret;
2316         }
2317
2318 err:
2319         if (flags & BCH_READ_IN_RETRY)
2320                 return READ_ERR;
2321
2322         orig->bio.bi_status = BLK_STS_IOERR;
2323         goto out_read_done;
2324
2325 hole:
2326         /*
2327          * won't normally happen in the BCH_READ_NODECODE
2328          * (bch2_move_extent()) path, but if we retry and the extent we wanted
2329          * to read no longer exists we have to signal that:
2330          */
2331         if (flags & BCH_READ_NODECODE)
2332                 orig->hole = true;
2333
2334         zero_fill_bio_iter(&orig->bio, iter);
2335 out_read_done:
2336         if (flags & BCH_READ_LAST_FRAGMENT)
2337                 bch2_rbio_done(orig);
2338         return 0;
2339 }
2340
2341 void __bch2_read(struct bch_fs *c, struct bch_read_bio *rbio,
2342                  struct bvec_iter bvec_iter, subvol_inum inum,
2343                  struct bch_io_failures *failed, unsigned flags)
2344 {
2345         struct btree_trans trans;
2346         struct btree_iter iter;
2347         struct bkey_buf sk;
2348         struct bkey_s_c k;
2349         u32 snapshot;
2350         int ret;
2351
2352         BUG_ON(flags & BCH_READ_NODECODE);
2353
2354         bch2_bkey_buf_init(&sk);
2355         bch2_trans_init(&trans, c, 0, 0);
2356 retry:
2357         bch2_trans_begin(&trans);
2358         iter = (struct btree_iter) { NULL };
2359
2360         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
2361         if (ret)
2362                 goto err;
2363
2364         bch2_trans_iter_init(&trans, &iter, BTREE_ID_extents,
2365                              SPOS(inum.inum, bvec_iter.bi_sector, snapshot),
2366                              BTREE_ITER_SLOTS);
2367         while (1) {
2368                 unsigned bytes, sectors, offset_into_extent;
2369                 enum btree_id data_btree = BTREE_ID_extents;
2370
2371                 /*
2372                  * read_extent -> io_time_reset may cause a transaction restart
2373                  * without returning an error, we need to check for that here:
2374                  */
2375                 ret = bch2_trans_relock(&trans);
2376                 if (ret)
2377                         break;
2378
2379                 bch2_btree_iter_set_pos(&iter,
2380                                 POS(inum.inum, bvec_iter.bi_sector));
2381
2382                 k = bch2_btree_iter_peek_slot(&iter);
2383                 ret = bkey_err(k);
2384                 if (ret)
2385                         break;
2386
2387                 offset_into_extent = iter.pos.offset -
2388                         bkey_start_offset(k.k);
2389                 sectors = k.k->size - offset_into_extent;
2390
2391                 bch2_bkey_buf_reassemble(&sk, c, k);
2392
2393                 ret = bch2_read_indirect_extent(&trans, &data_btree,
2394                                         &offset_into_extent, &sk);
2395                 if (ret)
2396                         break;
2397
2398                 k = bkey_i_to_s_c(sk.k);
2399
2400                 /*
2401                  * With indirect extents, the amount of data to read is the min
2402                  * of the original extent and the indirect extent:
2403                  */
2404                 sectors = min(sectors, k.k->size - offset_into_extent);
2405
2406                 bytes = min(sectors, bvec_iter_sectors(bvec_iter)) << 9;
2407                 swap(bvec_iter.bi_size, bytes);
2408
2409                 if (bvec_iter.bi_size == bytes)
2410                         flags |= BCH_READ_LAST_FRAGMENT;
2411
2412                 ret = __bch2_read_extent(&trans, rbio, bvec_iter, iter.pos,
2413                                          data_btree, k,
2414                                          offset_into_extent, failed, flags);
2415                 if (ret)
2416                         break;
2417
2418                 if (flags & BCH_READ_LAST_FRAGMENT)
2419                         break;
2420
2421                 swap(bvec_iter.bi_size, bytes);
2422                 bio_advance_iter(&rbio->bio, &bvec_iter, bytes);
2423
2424                 ret = btree_trans_too_many_iters(&trans);
2425                 if (ret)
2426                         break;
2427         }
2428 err:
2429         bch2_trans_iter_exit(&trans, &iter);
2430
2431         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
2432             ret == READ_RETRY ||
2433             ret == READ_RETRY_AVOID)
2434                 goto retry;
2435
2436         bch2_trans_exit(&trans);
2437         bch2_bkey_buf_exit(&sk, c);
2438
2439         if (ret) {
2440                 bch_err_inum_ratelimited(c, inum.inum,
2441                                          "read error %i from btree lookup", ret);
2442                 rbio->bio.bi_status = BLK_STS_IOERR;
2443                 bch2_rbio_done(rbio);
2444         }
2445 }
2446
2447 void bch2_fs_io_exit(struct bch_fs *c)
2448 {
2449         if (c->promote_table.tbl)
2450                 rhashtable_destroy(&c->promote_table);
2451         mempool_exit(&c->bio_bounce_pages);
2452         bioset_exit(&c->bio_write);
2453         bioset_exit(&c->bio_read_split);
2454         bioset_exit(&c->bio_read);
2455 }
2456
2457 int bch2_fs_io_init(struct bch_fs *c)
2458 {
2459         if (bioset_init(&c->bio_read, 1, offsetof(struct bch_read_bio, bio),
2460                         BIOSET_NEED_BVECS) ||
2461             bioset_init(&c->bio_read_split, 1, offsetof(struct bch_read_bio, bio),
2462                         BIOSET_NEED_BVECS) ||
2463             bioset_init(&c->bio_write, 1, offsetof(struct bch_write_bio, bio),
2464                         BIOSET_NEED_BVECS) ||
2465             mempool_init_page_pool(&c->bio_bounce_pages,
2466                                    max_t(unsigned,
2467                                          c->opts.btree_node_size,
2468                                          c->opts.encoded_extent_max) /
2469                                    PAGE_SIZE, 0) ||
2470             rhashtable_init(&c->promote_table, &bch_promote_params))
2471                 return -ENOMEM;
2472
2473         return 0;
2474 }